content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
A derivation from a Banach algebra $A$ to a
Banach $A$-bimodule $X$ with the continuous module operations is a bounded linear mapping $d:A\rightarrow X$
such that $$d(ab)=d(a)b+ad(b)\ \ \ (a,b\in A).$$ For each $x\in
X$ the mapping $\delta_x:a\rightarrow ax-xa$, $(a \in A)$ is a
bounded derivation, called an inner derivation.\\
Let $X$ be a Banach $A$-bimodule. Then $X^*$ is a dual Banach $A$-bimodule, by defining $a.f$ and $f.a$, for each $a\in A$ and $f\in X^*$ by$$a.f(x)=f(xa) \ \ \ ,\ \ \ \ f.a(x)=f(ax)\hspace{1cm}(x\in X).$$
Similarly, the higher duals $X^{(n)}$ can be made into Banach $A$-bimodules in a natural fashion.
A Banach algebra $A$ is called amenable if for each Banach $A$-bimodule $X$, the only derivations from $A$ to $X^*$ are inner derivations. For more details about this notion see \cite{Ru}.
Let $A$ be a Banach algebra and $X$ be a Banach $A$-bimodule, a bounded bilinear mapping $D:A\times A \rightarrow X$ is called a biderivation if $D$ is a derivation with respect to both arguments. That is the mappings $_aD:A\rightarrow X$ and $D_b:A\rightarrow X$ are derivations. Where $$_aD(b)=D(a,b)=D_b(a) \hspace{2cm}(a, b\in A).$$ We denote the space of such biderivations by $BZ^1(A, X)$.
Consider the subspace $Z(A,X) =\{x\in X: ax=xa \ \ \forall a\in A\}$ of $X$. Then for each $x\in Z(A,X)$,
the mapping $\Delta_x:A \times A\rightarrow X$ defined by
$$\Delta_x(a,b)=x[a,b]=x(ab-ba) \ \ \ \ \ \ (a, b\in A)$$
is a basic example of a biderivation and called an inner biderivation. We denote the space of such inner biderivations by $BN^1(A,X)$.
For more applications of biderivations, see the survey article {\cite[Section 3]{Bre}}. Some algebraic aspects of biderivations on certain algebras were investigated by many authors; see for example \cite{Ben, DW}, where the structures of biderivations on triangular algebras and generalized matrix algebras are discussed, and particularly the question of whether biderivations on these algebras are inner, was considered.
We define the first bicohomology group $BH^1(A,X)$ as follows, $$BH^1(A,X)=\frac{BZ^1(A,X)}{BN^1(A,X)}.$$
Obviously $BH^1(A,X)=0$ if and only if every biderivation from $A\times A$ to $X$ is an inner biderivation. Now we are motivated to define the concept of biamenability of Banach algebras as follows.\\
A Banach algebra $A$ is biamenable if for each Banach $A$-bimodule $X$ we have $BH^1(A, X^*) = 0.$
Although one might expect that biderivations and biamenability must run parallel to derivations and amenable Banach algebras what is true is that although there are some external similarities between them but they lead to very different, and
somewhat opposed, theories. Indeed we show that commutative Banach algebras tend
to lack biamenability, while highly noncommutative Banach algebras tend
to be biamenable. Thus for instance, the ground algebras $\mathbb{C}$ and $\mathbb{R}$ are not
biamenable (while they are trivially amenable) and $B(H)$, the algebra of all bounded operators
on an infinite dimensional Hilbert space $H$, is biamenable, but not amenable. Moreover, if $H$ is
finite dimensional, it turns out that $B(H)$ is amenable, but it fails to be biamenable.
\section{biamenable Banach algebras}
For an example of a biamenable Banach algebra we commence with the next lemmas.
The following lemma is similar to Corollary 2.4 of \cite{b}, where it is introduced for a biderivation $D:A\times A\rightarrow A$.
\begin{lemma}\label{lem1}
For each Banach $A$-bimodule $X$ and each biderivation $D:A\times A\rightarrow X$, $$D(a,b)[c,d]=[a,b]D(c,d)\hspace{1cm}(a, b, c, d\in A).$$
\end{lemma}
In Proposition 2.1.3 of \cite{Ru} it is shown that if $A$ has a bounded approximate identity,
and one of module actions is trivial, then the only derivations from $A$ to $X^*$ are inner derivations. The following lemma introduces a condition that not only implies the innerness of biderivations but it also forces them to be zero.
\begin{lemma}\label{lem2}
If a Banach algebra $A$ has a bounded left approximate identity and $span\{ab-ba: a,b\in A\}$ is dense in $A$, then for every Banach $A$-bimodule $X$ such that $XA=0$ we have $BZ^1(A,X)=0$.
\end{lemma}
\begin{proof}
Let $(e_\alpha)$ be a left approximate identity of $A$. Lemma \ref{lem1} says that for each $D\in BZ^1(A,X)$,$$[a,b]D(c,d)=0\hspace{1cm}(a, b, c, d\in A).$$
So by density we have $aD(b,c)=0$, for each $a, b, c\in A$.
Now since $XA=0$ we have $$\begin{array}{rcl}
D(a,b)&=&\lim_\alpha D(e_\alpha a,b)\\
&=&\lim_\alpha[e_\alpha D(a,b)+D(e_\alpha,b)a]\\
&=&0.
\end{array}$$
That is $BZ^1(A,X)=0$.
\end{proof}
A very similar proof may be applied if $A$ has a right approximate identity and the left module action is trivial.\\
The following lemma introduces a condition that under which some biderivations are inner. This result will lead to a condition implying biamenability of a Banach algebra. We shall see that $B(H)$, for each infinite dimensional Hilbert space $H$, satisfies this condition.
\begin{lemma}\label{lem3}
If $A$ is unital and $A=span\{ab-ba: a,b\in A\}$, then for every unital Banach $A$-bimodule $X$, every biderivation $D:A\times A\rightarrow X$ is an inner biderivation.
\end{lemma}
\begin{proof}
Let $D$ be a biderivation and $e$ be the identity of $A$. Since $A=span\{ab-ba: a,b\in A\}$, there exist $a_i$ and $b_i$ in $A$ such that $e=\sum_{i}[a_i,b_i]$. So by Lemma \ref{lem1}, for every $a, b\in A$, we have
$$\begin{array}{rcl}
D(a,b)&=& D(a,b)e\\
&=& D(a,b)\sum_{i}[a_i,b_i]\\
&=& \sum_{i}D(a,b)[a_i,b_i]\\
&=&\sum_i[a,b]D(a_i,b_i)\\
&=&[a,b]\lambda.
\end{array}$$
Where $\lambda= \sum_{i}D(a_i,b_i)$. Similarly we have $D(a,b)=\lambda [a,b],$ and so $$\lambda [a,b]=[a,b]\lambda\hspace{1cm}(a\in A, b\in B).$$
Now since $A=span\{ab-ba: a,b\in A\}$, $\lambda\in Z(A,X)$. So $D=\Delta_\lambda$ is an inner biderivation.
\end{proof}
\begin{lemma}\label{lem4}
If a Banach algebra $A$ has a bounded approximate identity and $span\{ab-ba: a,b\in A\}$ is dense in $A$,
then the following statements are equivalent.
\begin{itemize}
\item[(i)] $A$ is biamenable.
\item[(ii)] $BH^1(A,X^*)=0$, for every left approximately unital Banach $A$-bimodule $X$.
\item[(iii)] $BH^1(A,X^*)=0$, for every right approximately unital Banach $A$-bimodule $X$.
\item[(iv)] $BH^1(A,X^*)=0$, for every approximately unital Banach $A$-bimodule $X$.
\end{itemize}
\end{lemma}
\begin{proof}
We only prove (i) is equivalent to (ii). The equivalence of (i) and (iii) is similar and then the equivalence of (i) and (iv) is obvious.\\
Clearly if $A$ is biamenable then (ii) is true. Now let $BH^1(A,Y^*)=0$, for every left approximately unital Banach $A$-bimodule $Y$ and let $X$ be a Banach $A$-bimodule. Then Corollary 2.9.26 of \cite{D} implies that $X_0=AX$ is a left approximately unital closed submodule of $X$. Also $A(\frac{X}{X_0})=0$ and so $(\frac{X}{X_0})^*A=0$. Therefore Lemma \ref{lem2} says that $BZ^1(A,X_0^\perp)= BZ^1(A,(\frac{X}{X_0})^*)=0$.\\
Let $D\in BZ^1(A,X^*)$ and $J:X_0\rightarrow X$ be the inclusion mapping. Then $J^*\circ D\in BZ^1(A,X_0^*)$ and by assumption $J^*\circ D=\Delta_{\phi_0}$, for some $\phi_0\in Z(A,X_0^*)$. Now the equation $X^*=X_0^*\oplus X_0^\perp$, which is implied from Theorem 4.9 of \cite{R}, shows that there exists an extension $\phi$ of $\phi_0$ such that $\phi\in Z(A,X^*)$. \\
Define $D_0=D-\Delta_\phi$. Then $D_0\in BZ^1(A, X_0^\perp)=0$ and so $D=\Delta_\phi$.
\end{proof}
A similar result of the previous lemma in the area of amenability is given in Proposition 2.1.5 of \cite{Ru}.\\
Now combination of the Lemmas \ref{lem3} and \ref{lem4} gives the following theorem that leads to a condition for biamenability of Banach algebras and then we can find some examples of biamenable Banach algebras which are not amenable.
\begin{theorem}\label{cor}
Each unital Banach algebra $A$ with $A=span\{ab-ba: a,b\in A\}$, is biamenable.
\end{theorem}
\begin{corollary}
If $A=span\{ab-ba: a, b\in A\}$ and $A$ has an identity, then the only biderivation $D:A\times A\rightarrow A^*$ is zero.
\end{corollary}
\begin{proof}
Let $D:A\times A\rightarrow A^*$ be a biderivation. By Theorem \ref{cor} $A$ is biamenable, so $D$ is inner. That is there exists an $f\in Z(A,A^*)$ such that for every $a, b\in A$, $D(a, b)=f[a, b]$.
Now $$\langle f,[a,b]\rangle=\langle fa-af,b\rangle=0.$$
Hence our assumption implies that $f=0$ and so $D=0$.
\end{proof}
Every bounded bilinear mapping $f: X\times Y\rightarrow Z$ on normed spaces $X, Y$ and $Z$, has two natural extensions $f^{***}$ and $f^{t***t}$ from $X^{**}\times Y^{**}$ to $Z^{**}$ as follows. \\
We define the adjoint $ f^*:{Z}^*\times {X} \rightarrow {Y}^*$ of $f$ by \[\langle f^*(z^*, x),y\rangle=\langle z^*,f(x,y)\rangle,\] where $x\in {X}, y\in {Y}\ \ {\rm and}\ \ z^*\in {Z}^*.$
We then define
$f^{**}=(f^*)^*$ and $f^{***}=(f^{**})^*$.
Let $f^{t}: {Y} \times {X}
\longrightarrow {Z}$ be the flip map of $f$ which is defined by $f^{t}(y,x)= f(x,y)\ ( x \in {X}, y
\in {Y}).$ If we continue the latter process with $f^t$ instead of $f$, we come to the bounded bilinear mapping $f^{t***t}: {X}^{**}\times {Y}^{**}\rightarrow Z^{**}$.\\
Where $\pi$ is the multiplication of a Banach algebra $A$, $\pi^{***}$ and $\pi^{t***t}$ are actually the first and second Arens products, which are denoted by
$\square$ and $\lozenge$, respectively. For more detailes see \cite{A} and \cite{Ar}.
Now we give some examples of biamenable Banach algebras.
\begin{example}\label{ex}
According to Lemma 5.8 of \cite{Z}, since for each infinite dimensional Hilbert space $H$ and every integer $n\geq 0$,
$$B(H)^{(2n)}=span\{au-ua: a\in B(H), u\in B(H)^{(2n)}\}.$$
So Theorem \ref{cor} help us to find some biamenable Banach algebras such as the Banach algebra
$B(H)^{(2n)}$ and the module extension Banach algebra $B(H)\oplus B(H)^{(2n)}$ with the product and norm as follows. $$(a,u)(b,v)=(ab,av+ub),\ \ \ \ \|(a,u)\|=\|a\|+\|u\|, \ \ \ \ \ (a\in B(H), u\in B(H)^{(2n)}).$$ Although $B(H)$ is not amenable in general.
Note that since $\{au-ua: a\in B(H), u\in B(H)^{(2n)}\}$ is a subset of $\{uv-vu: u, v\in B(H)^{(2n)}\}$ and $\{[a,v]-[b,u]: a,b\in B(H), u,v\in B(H)^{(2n)}\}$.
Therefore the commutators span the whole of $B(H)^{(2n)}$ and $B(H)\oplus B(H)^{(2n)}$. \\
Also similar to last corollary we can show that the only biderivation from $B(H)\times B(H)$ to $B(H)^{(2n+1)}$ is zero.
A similar method as Lemma 5.7 of \cite{Z}, can show that for the Banach algebra $K(H)$ of compact operators on $H$, $$K(H)=span\{ku-uk: k\in K(H): u\in B(H)\}.$$ So similarly $B(H)\oplus K(H)$ is biamenable. Although Remark 5.10 of \cite{Z} says that it is not amenable.
\end{example}
Let $G$ be a locally compact group. $m\in L^\infty(G)^*$ is a mean on $L^\infty(G)$ if $m(1)=\|m\|=1$. A mean $m$ on $L^\infty(G)$ is called a left invariant mean if for each $x\in G$ and $g\in L^\infty(G)$, $m(\delta_x*g)=m(g)$.
$G$ is called amenable if there is a left invariant mean on $L^\infty(G)$.\\
Consider $L^\infty(G)$ as an $L^1(G)$-bimodule with the left and right module actions
$$\pi_\ell:L^1(G)\times L^\infty(G) \rightarrow L^\infty(G)\quad \pi_r:L^\infty(G)\times L^1(G)\rightarrow L^\infty(G),$$ defined by $\pi_\ell(f,g)=f*g$ and $\pi_r(g,f)=(\int_G f) g$. Then we have the following proposition.
\begin{proposition}
Let $G$ be a locally compact group such that $Z(L^1(G),L^\infty(G)^*)$ containes an element $n$ such that $n(1)\neq 0$. Then $G$ is amenable.
\end{proposition}
\begin{proof}
Define $|n|$ by $|n|(\phi)=|n(g)|$, for each $g\in L^\infty(G)$. Then $m=\frac{|n|}{|n(1)|}$ is a positive element of $Z(L^1(G),L^\infty(G)^*)$ such that $m(1)=1$. Therefore by {\cite[Proposition 1.1.2]{Ru}}, $m$ is a mean on $L^\infty(G)$.\\
Now we have
\begin{align*}
\langle m,f*g\rangle&=\langle m,\pi_\ell(f,g)\rangle\\
&=\langle \pi_\ell^*(m,f),g\rangle\\
&=\langle\pi_r^{t*t}(f,m),g\rangle\\
&=\langle m,\pi_r(g,f)\rangle\\
&=\langle m,g\rangle\int_Gf\\
&=\langle m,g\rangle\quad (f\in P(G), g\in L^\infty(G)).\end{align*}
Where $P(G)=\{f\in L^1(G); \|f\|_1=1, f\geq 0\}$. So {\cite[Lemma 1.1.7]{Ru}} implies that $G$ is amenable.
\end{proof}
A big class of Banach algebras are not biamenable, although they may be amenable.
For example if there exists a non-zero derivation $d:A\rightarrow A^{**}$ on a commutative Banach algebra $A$ such that for some $a, b\in A$, $d(a) \square d(b)\neq 0$, then $A$ is not biamenable.
Since the map
$$\begin{array}{rcl}
D:A \times A&\longrightarrow& A^{**}\\
(a,b)&\mapsto& d(a)\square d(b)
\end{array}$$
where $\square $ denotes the first Arens product of $A^{**}$, defines a biderivation which is not inner. Also
if there is a Banach $A$-bimodule $X$ such that $Z(A,X)=\{0\}$ and there is a non zero biderivation from $A\times A$ into $X^*$, then $A$ is not biamenable.
Since in this case $Z(A,X^*)=\{0\}$ and so the only inner biderivation $D:A\times A\rightarrow X^*$ is zero and therefore we have the following proposition.
\begin{proposition}\label{pr3}
Let $\sigma(A)$ be the spectrum of $A$. If $\sigma(A)\neq\emptyset$, then $A$ is not biamenable.
\end{proposition}
\begin{proof}
Let $A$ be a Banach algebra such that $\sigma(A)\neq \emptyset$, $f$ be an element in $\sigma(A)$ and $X$ be a non zero Banach $A$-bimodule with module actions $$ax=0, \ \ \ xa=f(a) x\hspace{1cm}(a\in A, x\in X).$$ Then $Z(A,X)=\{0\}$. But for a non-zero element $h\in X^*$
$$\begin{array}{rcl}
D:A\times A&\rightarrow& X^*\\
(a,b)&\mapsto&f(a) (bh-hb)
\end{array} $$
is a non-zero biderivation.
\end{proof}
Now by applying Proposition \ref{pr3} and Theorem 1.3.3 of \cite{M}, we conclude that every unital commutative Banach algebra is not biamenable. For example $\mathbb{C}$, $\mathbb{R}$, $C(\Omega)$, for each Hausdorff space $\Omega$ and the group algebra $M(G)$, for each locally compact abelian group $G$ are not biamenable. In the next section we extend this result
to arbitrary commutative Banach algebras. Also a combination of Example \ref{ex} and Proposition
\ref{pr3} implies the following.
\begin{corollary}
For each integer $n\geq 0$ and each infinite dimensional Hilbert space $H$, $\sigma(B(H)^{(2n)})=\emptyset$ and $\sigma(B(H)\oplus B(H)^{(2n)})=\emptyset$.
\end{corollary}
\begin{example}
Let $A$ be a Banach space and $\theta$ be a non zero element of $A^*$. Then $A$ is a Banach algebra with the multiplication
$$ab=\theta(a) b,\hspace{1cm}(a, b\in A).$$
Now since $\theta\in \sigma(A)$, $A$ with this multiplication is not biamenable.
\end{example}
\section{Some properties}
In this section we study some properties of biamenable Banach algebras and we tend to some another examples of non biamenable Banach algebras.
\begin{theorem}
Let $A$ be a Banach algebra and consider $\mathbb{C}$ as a Banach $A$-bimodule. If there is a nonzero derivation $d : A \rightarrow\mathbb{C}$, then biamenability of $A$ implies amenability of $A$.
\end{theorem}
\begin{proof}
Let $X$ be a Banach $A$-bimodule and $d' : A \rightarrow X^*$ be a bounded derivation. Then
$$\begin{array}{rcl}
D : A\times A &\rightarrow & X^*\\ (a,b)&\mapsto & d(a)d'(b)
\end{array}$$
is a bounded biderivation and so there is $f\in Z(A,X^*)$ such that $$d(a)d'(b) = D(a,b) = f[a,b]\quad (a,b\in A).$$ Therefore for every $b\in A$ and for some $a \in A$ such that $d(a)\neq 0$ we have $d'(b) = \delta_{-\frac{fa}{d(a)}}(b)$.
\end{proof}
\begin{example}
Let $\mathbb{D} =\{z \in\mathbb{C}; |z|\leq 1\}$ be the unit disc, and $A(\mathbb{D})$ be the disc algebra. We can consider $\mathbb{C}$ as an $A(\mathbb{D})$-bimodule with module actions $\alpha f =\alpha f(0) = f\alpha$ and
$$\begin{array}{rcl}
d : A(\mathbb{D}) &\rightarrow & \mathbb{C}\\ f &\mapsto & f'(0)
\end{array}$$ is a nonzero derivation. Therefore since $A(\mathbb{D})$ is not amenable so it is not biamenable.
\end{example}
We know that every amenable Banach algebra has an approximate identity (See Proposition 2.2.1 of \cite{Ru}). A similar result is given in the following.
\begin{proposition}\label{pr1}
If $A=span\{ab-ba: a,b\in A\}$ and $A$ is biamenable, then $A$ has a bounded approximate identity.
\end{proposition}
\begin{proof}
If $A$ is biamenable, then for the biderivation
$$\begin{array}{rcl}
D:A\times A&\rightarrow &A^{**}\\
(a,b)&\mapsto &[a,b]
\end{array}$$
there is $E\in Z(A,A^{**})$ such that for each $a, b\in A$, $E[a,b]=[a,b]$. Now let $(e_\alpha)$ be a bounded net in $A$ which is $w^*$-convergent to $E$. Then we have
$$\begin{array}{rcl}
\lim_\alpha e_\alpha [a,b]&=&w-\lim_\alpha e_\alpha [a,b] \\
&=&E[a,b]\\
&=&[a,b]\\
&=&[a,b]E\\
&=&w-\lim_\alpha [a,b]e_\alpha \\
&=&\lim_\alpha [a,b]e_\alpha ,
\end{array}$$
and by assumption $A$ has an approximate identity $(e_\alpha)$.
\end{proof}
Note that the converse of this proposition is not true in general. For example in the sequel we see that every commutative Banach algebra is not biamenable. Although it may be unital or approximately unital.
For each integer $n\geq 0$ put $$AA^{(2n)}+A^{(2n)}A=\{aa^{(2n)}+b^{(2n)}b: a, b \in A, a^{(2n)}, b^{(2n)}\in A^{(2n)}\}.$$ Then we have the following proposition.
\begin{proposition}\label{n}
If $A$ is biamenable, then for each integer $n\geq 0$, $span (AA^{(2n)}+A^{(2n)}A)$ is dense in $A^{(2n)}$.
\end{proposition}
\begin{proof}
If $span (AA^{(2n)}+A^{(2n)}A)$ is not dense in $A^{(2n)}$, then there exists a non-zero linear functional $f\in A^{(2n+1)}$ such that it is zero on $span (AA^{2n}+A^{(2n)}A)$. Now the bilinear mapping
$$\begin{array}{rcl}
D:A\times A&\rightarrow &A^{(2n+1)}\\
(a,b)&\mapsto &f(a)f(b)f
\end{array}$$
is a biderivation which is not inner. So $A$ is not biamenable, which is a contradiction.
\end{proof}
Let $A$ be a Banach algebra and $$A^{n}=span\{a_1...a_n: a_i\in A\}\hspace{1cm}(n\in \mathbb{N}).$$
As a corollary of the latter proposition we have:
\begin{corollary}\label{cor3}
If $A$ is biamenable then for each $n\in \mathbb{N},$ $A^{n}$ is dense in $A$.
\end{corollary}
\begin{proof}
By Proposition \ref{n} $A^{2}$ is dense in $A$. Now by applying the density of $A^{2}$ in $A$ we can prove that $A^{3}$ is dense in $A$ and also by an inductive method we can prove that for each $n$, $A^{n}$ is dense in $A$.
\end{proof}
For a Banach algebra $A$, put $$[A,A]=\{[a,b]: a,b\in A\} \ \ \ \mbox{and} \ \ \ [A,A]A=\{[a,b]c: a, b, c\in A\}.$$ The following proposition gives a big class of non-biamenable Banach algebras.
\begin{proposition}
Let $A$ be a biamenable Banach algebra. Then $span([A,A]\cup [A,A]A)$ is dense in $A$.
\end{proposition}
\begin{proof}
Suppose $S=span([A,A]\cup [A,A]A)$ is not dense in $A$. Then there exists a nonzero element $f\in A^*$ such that $f|_S=0$. In particular for each $a, b, c\in A$, we have $f(ab)=f(ba)$ and $c.f(ab)=c.f(ba)$. Consider $X=\overline{f.A}$ as an $A$-bimodule with module actions
$$(f.a).b=f.ab, \ \ \ \mbox{and} \ \ \ b.(f.a)=0\hspace{1cm}(a, b\in A).$$ Then $Z(A,X^*)=\{0\}$ and so the only inner biderivation from $A\times A$ to $X^*$ is zero. Now by Corollary \ref{cor3} the bilinear mapping $D:A\times A\rightarrow X^*$ defined by $$\langle D(a,b),f.c\rangle=f(abc), \hspace{1cm}(a, b, c\in A)$$
is nonzero. Also for each $a, b, c, d\in A$ we have
$$\begin{array}{rcl}
\langle D(ab,c),f.d\rangle &=&f(abcd)\\
&=&f(bcda)\\
&=&\langle D(b,c),f.da\rangle \\
&=&\langle D(b,c),(f.d).a)\rangle \\
&=&\langle aD(b,c),f.d\rangle \\
&=&\langle aD(b,c)+D(a,c)b,f.d\rangle ,
\end{array}$$ and similarly
$$\begin{array}{rcl}
\langle D(a,bc),f.d\rangle &=&f(abcd)\\
&=&f(cdab)\\
&=&(b.f)(cda)\\
&=&(b.f)(acd)\\
&=&f(acdb)\\
&=&\langle D(a,c),f.db\rangle \\
&=&\langle bD(a,c)+D(a,b)c, f.d\rangle .
\end{array}$$
So $D$ is a nonzero biderivation and so it is not inner. That is a contradiction.
\end{proof}
Note that if a biamenable Banach algebra $A$ has a right approximate identity, then $[A,A]\subseteq \overline{[A,A]A}$ and therefore $span([A,A]A)$ is dense in $A$. This may be compared with the converse of Proposition \ref{pr1}, for each biamenable Banach algebra.
\begin{corollary}\label{com}
Every non zero commutative Banach algebra is not biamenable.
\end{corollary}
This corollary implies that for every locally compact abelian group $G$, $L^1(G)$ is not biamenable but it is amenable by applying Johnson's theorem and the Example 1.1.5 of \cite{Ru}.
\begin{theorem}
Let $A$ be a Banach algebra, $X$ be a Banach $A$-bimodule and $I$ be a closed ideal of it such that $Z(A,X^*)=Z(I,X^*)$. Then if $BH^1(I,X^*)=\{0\}$ and $\frac{A}{I}$ is biamenable, then $BH^1(A,X^*)=\{0\}$.
\end{theorem}
\begin{proof}
Let $D:A\times A\rightarrow X^*$ be a biderivation. Then $D_0=D|_{I\times I}\in BZ^1(I,X^*)$ and so $D_0=\Delta_E$, for some $E\in Z(A,X^*)$. Put $\tilde{D}=D-\Delta_E$. Then $\tilde{D}(I\times I)=0$ and so $\mathfrak{D}:\frac{A}{I}\times\frac{A}{I}\rightarrow X^*$, defined by $\mathfrak{D}((a+I,b+I))=\tilde{D}((a,b))$ is a well defined map. \\
Put $X_0=IX+XI$. Then $$(\frac{X}{X_0})^*=X_0^\perp=\{\phi\in X^*; \phi i=0=i\phi, \text{ for all } \ i\in I\}.$$
and so we can consider $X_0^\perp$ as an $\frac{A}{I}$-bimodule with the module actions $(a+I).\phi=a.\phi$ and $\phi.(a+I)=\phi.a$, for each $a\in A$ and $\phi\in X_0^\perp$. On the other hand we have
\begin{align*}
\tilde{D}(a,b)ij&=(\tilde{D}(ai,b)-a\tilde{D}(i,b))j\\
&=\tilde{D}(ai,bj)-b\tilde{D}(ai,j)-a\tilde{D}(i,bj)+ab\tilde{D}(i,j)\\
&=0, \quad (i, j\in I, a, b\in A).
\end{align*}
Similarly we can show that $ij\tilde{D}(a,b)=0$. Therefore by density of $I^2$ in $I$ (Proposition \ref{n}) we conclude that $\mathfrak{D}(\frac{A}{I}\times\frac{A}{I})\subseteq X_0^\perp$ and then we can coclude that $\mathfrak{D}:\frac{A}{I}\times\frac{A}{I}\rightarrow X_0^\perp$ is a biderivation. So there is $\psi\in Z(\frac{A}{I},X_0^\perp)\subseteq Z(A,X^*)$ such that $D-\Delta_E=\tilde{D}=\Delta_\psi$. Hence $D=\Delta_{E+\psi}$ and $E+\psi\in Z(A,X^*)$.
\end{proof}
We know that a Banach algebra $A$ is amenable if and only if the unconditional unitization $A^\flat$ of $A$ (see {\cite[Definition 1.3.3]{D}}) is amenable {\cite[Corollary 2.3.11]{Ru}}. But it is not true for biamenability of Banach algebras. Indeed we show that the unconditional unitization of each Banach algebra is not biamenable.
\begin{lemma}
If $\theta: A\rightarrow B$ is a continuous homomorphism of Banach algebras with dense range and $A$ is biamenable, then so is $B$.
\end{lemma}
\begin{proof}
Let $X$ be a Banach $B$-bimodule. Consider $X$ as an $A$-bimodule with module actions $ax=\theta(a) x$ and $xa=x\theta(a)$. Now for each $D\in BZ^1(B,X^*)$, $D\circ (\theta\times\theta)\in BZ^1(A,X^*)$ and biamenability of $A$ implies that $D\circ (\theta\times\theta)=\Delta_\phi$ for some $\phi\in Z(A,X^*)$. Now by density we conclude that $\phi\in Z(B,X^*)$ and $D=\Delta_\phi$ .
\end{proof}
\begin{corollary}\label{co}
If $A$ is biamenable then for each closed ideal $I$ of $A$, $\frac{A}{I}$ is biamenable.
\end{corollary}
For analogues of the above two results in the area
of amenability see Proposition 2.3.1 and Corollary 2.3.2 of \cite{Ru}.
Now we have the following theorem for the unconditional unitization $A^\flat$ of a Banach algebra $A$.
\begin{theorem}
For each Banach algebra $A$, the unconditional unitization $A^\flat$ is not biamenable.
\end{theorem}
\begin{proof}
If $A^\flat$ is biamenable then $\mathbb{C}=\frac{A^\flat}{A}$ is biamenable by
Corollary \ref{co} (recall that A is a closed ideal in $A^\flat$). A contradiction.
\end{proof}
\section{biamenability of a pair of Banach algebras}
Let $A$ and $B$ be Banach algebras and $X$ be an $A-B$-bimodule that is $X$ is an $A$-bimodule and $B$-bimodule and we have $$a(xb)=(ax)b \ \ , \ \ b(xa)=(bx)a\hspace{2cm}(a\in A, b\in B , x\in X).$$ The bounded bilinear mapping $D:A\times B \rightarrow X$ is called a biderivation if $D$ is a derivation with respect to both arguments. That is the mappings $_aD:B\rightarrow X$ and $D_b:A\rightarrow X$ where $$_aD(b)=D(a,b)=D_b(a) \hspace{2cm}(a\in A , b\in B)$$ are derivations. We denote the space of such biderivations by $BZ^1(A\times B, X)$.
Let $x\in Z(A,X)\cap Z(B,X)$, where $$Z(A,X) =\{x\in X ; ax=xa \ \ \forall a\in A\}.$$
The map $D_x:A \times B\rightarrow X$ that
$$D_x(a,b)=x[a,b]=(xa)b-(xb)a \ \ \ \ \ \ (a\in A, b\in B )$$
is a basic example of biderivations and called an inner biderivation. We denote the space of such inner biderivations by $BN^1(A\times B,X)$. Also we define the first bicohomology group $BH^1(A\times B,X)$ as follows, $$BH^1(A\times B,X)=\frac{BZ^1(A\times B,X)}{BN^1(A\times B,X)}.$$
Let $A$ and $B$ be Banach algebras. We say that $A$ and $B$ commute with respect to an $A-B$-bimodule $X$ if for each $a\in A, b\in B$ and $x\in X$ we have
$a(bx)=b(ax)$ and $(xb)a=(xa)b$.\\
Note that if $A$ and $B$ commute with respect to $X$ then they commute with respect to $A-B$-bimodule $X^*$.
For example if we consider $X$ as an $A-B$-bimodule with module actions zero on $A$, then $A$ and $B$ commute with respect to this $A-B$-bimodule. Also if $A$ is commutative then $A$ commutes with itself with respect to $A$.
\begin{proposition}
If $A$ and $B$ commute with respect to an $A-B$-bimodule $X$ and there are nonzero derivations $d:A\rightarrow X$ and $d':B\rightarrow X$, then there is a nonzero biderivation from $A\times B$ into $X$.
\end{proposition}
\begin{proof}
Define $D:A\times B\rightarrow X$ by $D(a,b)=\delta_{d(a)}(b)+\delta_{d'(b)}(a)$.
\end{proof}
\begin{proposition}
For each biderivation $D:A\times B\rightarrow X$, $$D(a,b)[c,d]=[a,b]D(c,d)\hspace{1cm}(a,c\in A , b,d\in B).$$
\end{proposition}
\begin{proof}
Since $X$ is an $A-B$-bimodule, we have for each $a,c\in A, b,d\in B$,
$$\begin{array}{rcl}
a(bD(c,d))+a(D(c,b)d)&+&(D(a,b)d)c+(bD(a,d))c\\&=&aD(c,bd)+D(a,bd)c\\
&=&D(ac,bd)\\&=&bD(ac,d)+D(ac,b)d\\
&=&b(aD(c,d))+b(D(a,d)c)+(aD(c,b))d+(D(a,b)c)d.
\end{array}$$
So $$[a,b]D(c,d)-D(a,b)[c,d]=a(bD(c,d))+b(aD(c,d)+(D(a,b)d)c-(D(a,b)c)d=0.$$
\end{proof}
\begin{definition}
We say that the pair $(A,B)$ is biamenable if for each $A-B$-bimodule $X$, $BH^1(A\times B, X^*)=\{0\}$.
\end{definition}
Obviously biamenability of a Banach algebra $A$ is nothing else than some refinement of biamenability of the pair $(A,A)$ with the same module actions on each argument, Or briefly biamenability of the pair $(A,A)$ implies biamenability of the Banach algebra $A$.
\begin{remark}
If $Z(A,X)=\{0\}$ or $Z(B,X)=\{0\}$ then the only inner biderivation $D:A\times B\rightarrow X$ is zero. Therefore if $X$ is an $A-B$-bimodule such that there is a non zero biderivation from $A\times B$ into $X^*$ and $Z(A,X)=\{0\}$, then $(A, B)$ is not biamenable.
\end{remark}
The following theorem says that amenability of Banach algebras $A$ and $B$ are necessary conditions for biamenability of the pair $(A,B)$.
\begin{theorem}
\begin{itemize}
\item[(i)] If $\sigma(B)\neq\emptyset$ and $(A,B)$ is biamenable then $A$ is amenable.
\item[(ii)] If $\sigma(A)\neq\emptyset$ and $(A,B)$ is biamenable then $B$ is amenable.
\end{itemize}
\end{theorem}
\begin{proof}
We only prove (i).\\
Let $X$ be a Banach $A$-bimodule and $d:A\rightarrow X^*$ be a derivation. consider $B$-bimodule actions $bx=0$ and $xb=\varphi(b)x$, for some $\varphi\in \sigma(B)$.
Then it is easy to see that $X$ is a Banach $A-B$-bimodule.
Now $D:A \times B\rightarrow X^*$ with $D(a,b)=\varphi(b)d(a)$ is a biderivation and so $D(a,b)=x^*[a,b]=\delta_{x^*b}(a)$, for some $x^*\in Z(A,X^*)\cap Z(B,X^*)$. so $d(a)=\frac{D(a,b)}{\varphi(b)}=\delta_{\frac{x^*b}{\varphi(b)}}(a)$. for some $b\in B$ such that $\varphi(b)\neq 0.$ i.e. $A$ is amenable.
\end{proof}
Note that the converse of this theorem is not true, in general. For example $\mathbb{C}$ is not biamenable by Corollary \ref{com} and so $(\mathbb{C},\mathbb{C})$ is not biamenable. Also $\sigma(\mathbb{C})\neq \emptyset$, but it is amenable.
\section{a condition for preserving biamenability}
In this section we investigate the relation between biamenability of two pairs $(A,B)$ and $(E,F)$ of Banach algebras, where there exist homomorphisms $T:E\rightarrow A$ and $S:F\rightarrow B$.
Let $A, B, E$ and $F$ be Banach algebras and $T:E\rightarrow A$ and $S:F\rightarrow B$ be bounded homomorphism of Banach algebras. If $X$ is a Banach $A-B$-bimodule, then we may consider $X$ as a Banach $E-F$-bimodule with module actions $e.x=T(e)x$, $x.e=xT(e)$, $f.x=S(f)x$ and $x.f=xS(f)$, which $x\in X$, $e\in E$ and $f\in F$.
\begin{theorem}
If $D:A\times B\rightarrow X$ is a biderivation, then $D\circ(T\times S):E\times F\rightarrow X$ is a biderivation. Also if $D$ is inner so is $D\circ(T\times S)$.
The converse is true if $T$ and $S$ are onto.
\end{theorem}
\begin{proof}
It is easy to check that for biderivation $D$, $D\circ(T\times S)$ is a biderivation. Also if $D$ is inner, then there is $x\in Z(A,X)\cap Z(B,X)$ such that $$D(a,b)=(xa)b-(xb)a\hspace{1cm}(a\in A, b\in B).$$ Therefore
$$D\circ(T\times S)(e,f)=(xT(e))S(f)-(xS(f))T(e)=(x.e).f-(x.f).e\hspace{1cm}(e\in E, f\in F).$$
Similarly we can show that $x\in Z(E,X)\cap Z(F,X)$.
For the next part note that if $T$ and $S$ are onto, then for each $a\in A$ and $b\in B$, $D(a,b)=D\circ(T\times S)(e,f)$ for some $e\in E$ and $f\in F$.
So for each $a, a'\in A$ and $b\in B$ there are $e, e'\in E$ and $f\in F$ such that
$$\begin{array}{rcl}
D(aa',b)&=&D\circ(T\times S)(ee',f)\\
&=&D\circ(T\times S)(e,f).e'+e.D\circ(T\times S)(e',f)\\
&=&D(a,b)T(e')+T(e)D(a',b)\\
&=&D(a,b)a'+aD(a',b).
\end{array}$$
Similarly we have for each $a\in A$ and $b, b'\in B$, $D(a,bb')=D(a,b)b'+bD(a,b').$
By a similar argument we may prove that if $T$ and $S$ are onto and $D\circ(T\times S)$ is inner so is $D$.
\end{proof}
\begin{corollary}
Consider $A, B, E$ and $F$ as above. If $T$ and $S$ are onto and $(E,F)$ is biamenable, then so is $(A,B)$.
\end{corollary}
|
\section*{Acknowledgment}
\noindent The authors would like to thank the anonymous reviewers for their comments and suggestions which helped improve this paper, as well as suggestions of potential interesting examples of Cartesian difference categories to study in the future (which we discuss in the conclusion).
\section{Introduction}
In the early 2000's, Ehrhard and Regnier introduced the differential $\lambda$-calculus~\cite{ehrhard2003differential}, an extension of the $\lambda$-calculus equipped with a differential combinator capable of taking the derivative of arbitrary higher-order functions. This development, based on models of linear logic equipped with a natural notion of ``derivative''~\cite{ehrhard2018introduction}, sparked a wave of research into categorical models of differentiation.
One of the most notable developments in the area are Cartesian differential categories~\cite{blute2009cartesian}, introduced by Blute, Cockett and Seely, which provide an abstract categorical axiomatization of the directional derivative from differential calculus. The relevance of Cartesian differential categories lies in their ability to model both ``classical'' differential calculus (with the canonical example being the category of Euclidean spaces and smooth functions between) and the differential $\lambda$-calculus (as every categorical model for it gives rise to a Cartesian differential category~\cite{manzonetto2012categorical}). However, while Cartesian differential categories have proven to be an immensely successful formalism, they have, by design, some limitations. Firstly, they cannot account for certain ``exotic'' notions of derivative, such as the difference operator from the calculus of finite differences~\cite{richardson1954introduction}. This is because the axioms of a Cartesian differential category stipulate that derivatives should be linear in their second argument (in the same way that the directional derivative is), whereas these aforementioned discrete sorts of derivatives need not be. Additionally, every Cartesian differential category is equipped with a tangent bundle monad~\cite{cockett2014differential,manzyuk2012tangent} whose Kleisli category can be intuitively understood as a category of generalized vector fields.
More recently, discrete derivatives have been suggested as a semantic framework for understanding incremental computation. This led to the development of change structures~\cite{cai2014theory} and change actions~\cite{alvarez2019change}. Change action models have been successfully used to provide a model for incrementalizing Datalog programs~\cite{alvarez2019fixing}, and have also been shown to model the calculus of finite differences. Change action models, however, are very general, lacking many of the nice properties of Cartesian differential categories (for example, addition in a change action model is not required to be commutative), even though they are verified in most change action models. As a consequence of this generality, the tangent bundle endofunctor in a change action model can fail to be a monad.
In this work, we introduce Cartesian difference categories (Section~\ref{CdCsec}), whose key ingredients are an infinitesimal extension operator and a difference combinator, whose axioms are a generalization of the differential combinator axioms of a Cartesian differential category. In Section~\ref{CDCisCdCsec}, we show that every Cartesian differential category is a Cartesian difference category whose infinitesimal extension operator is zero, and conversely how every Cartesian difference category admits a full subcategory which is a Cartesian differential category. In Section~\ref{CdCisCAsec}, we show that every Cartesian difference category is a change action model, and conversely how a full subcategory of suitably well-behaved objects and maps of a change action model is a Cartesian difference category. In Section~\ref{EXsec} we provide some examples of Cartesian difference categories; notably, the calculus of finite differences and the stream calculus. In Section~\ref{monadsec}, we show that every Cartesian difference category comes equipped with a monad whose Kleisli category (Section~\ref{Kleislisec}) and a certain full sub-category of its Eilenberg-Moore category (Section~\ref{EMsec}) are again both Cartesian difference categories. Lastly, in Section~\ref{sec:differential-lambda-categories}, we briefly discuss difference $\lambda$-categories, which are Cartesian difference categories that are Cartesian closed. We conclude with a discussion on future work regarding Cartesian difference categories.
This paper is an extended version of a conference paper~\cite{alvarez2020cartesian} for Foundations of Software Science and Computation Structures: 23rd International Conference (FOSSACS 2020). While the overall structure and story of both versions are very similar, this version includes additional details, proofs, and results, as well as some important corrections. We highlight the main differences between this paper and the conference paper. Probably the most important change to point out is the correction made in Section~\ref{Kleislisec}. In this version, we correct the Cartesian difference structure of the Kleisli category of the tangent bundle monad. The proposed infinitesimal extension and difference combinator for said Kleisli category in the conference paper~\cite{alvarez2020cartesian} was based on a result from another paper~\cite{alvarez2019change}. Unfortunately, we have found that the result in said other paper is incorrect, and therefore so were the proposed infinitesimal extension and difference combinator for the Kleisli category in the conference paper. Section~\ref{CDCsec} and Section~\ref{CAsec}, the background sections, are mostly the same with some added background details. In Section~\ref{sec:CdC}, more detailed proofs and computations are provided throughout, and both Lemma~\ref{lem:d-epsilon} and Lemma~\ref{epsilon-nilponent} are new results. In Section~\ref{CDCisCdCsec}, we give a specific name to objects whose infinitesimal extensions is zero, which in Definition~\ref{def:ep-vanishing} we call $\varepsilon$-vanishing. In Section~\ref{CdCisCAsec} we introduce flat objects, whose definition (Definition~\ref{def:flat}) is slightly more general than the original one. Section~\ref{EXsec}, the example section, is mostly unchanged with some added details and proofs. In Section~\ref{monadsec}, we provide full detailed computations for all the necessary structure relating to the tangent bundle monad. Section~\ref{EMsec} and Section~\ref{sec:differential-lambda-categories} are both new sections.
\section*{Conventions:} In an arbitrary category, we use the classical notation for composition as opposed to diagrammatic order which was used in~\cite{blute2009cartesian}: this means that the composite map $g \circ f: A \to C$ is the map which first does $f: A\to B$ then $g: B \to C$. By a Cartesian category we mean a category $\mathbb{X}$ with chosen finite products where we denote the binary product of objects $A$ and $B$ by $A \times B$ with projection maps $\pi_0: A \times B \to A$ and $\pi_1: A \times B \to B$ and pairing operation $\langle -, - \rangle$, so the product of maps is $f \times g = \langle f \circ \pi_0, g \circ \pi_1 \rangle$, and the chosen terminal object as $\top$ with unique terminal maps $!_A: A \to \top$.
\section{Cartesian Differential Categories}\label{CDCsec}
In this section, we very briefly review Cartesian differential categories, so that the reader may compare Cartesian differential categories with the new notion of Cartesian \emph{difference} categories which we introduce in Section~\ref{CdCsec}. For a full detailed introduction on Cartesian differential categories, we refer the reader to the original paper~\cite{blute2009cartesian}.
\subsection{Cartesian Left Additive Categories}
Here we recall the definition of Cartesian left additive categories~\cite{blute2009cartesian}, which is the underlying structure of both Cartesian differential categories and Cartesian difference categories. Here ``left additive'' is meant being skew enriched~\cite{Campbell2018} over commutative monoids, so in particular, this means that we do not assume the existence of additive inverses, i.e., ``negative elements''. This left additive structure allows one to sum maps and have zero maps while allowing for maps that do not preserve the sum or zero. Maps that do preserve the additive structure are called additive maps.
\begin{defiC}[{\cite[Definition 1.1.1]{blute2009cartesian}}] A \textbf{left additive category} is a category $\mathbb{X}$ such that each hom-set $\mathbb{X}(A,B)$ is a commutative monoid with addition operation $+: \mathbb{X}(A,B) \times \mathbb{X}(A,B) \to \mathbb{X}(A,B)$ and zero element (called the zero map) $0 \in \mathbb{X}(A,B)$, such that pre-composition preserves the additive structure:
\begin{align*}
(f + g) \circ h = f \circ h + g \circ h && 0 \circ f = 0
\end{align*}
A map $k$ is \textbf{additive} if post-composition by $k$ preserves the additive structure:
\begin{align*}
k \circ (f +g) = k \circ f + k \circ g && k \circ 0=0
\end{align*}
\end{defiC}
\begin{defiC}[{\cite[Definition 2.3]{lemay2018tangent}}]\label{CLACdef} A \textbf{Cartesian left additive category} is a Cartesian category $\mathbb{X}$ which is also a left additive category such all projection maps $\pi_0: A \times B \to A$ and $\pi_1: A \times B \to B$ are additive.
\end{defiC}
Examples of Cartesian left additive categories can be found in Section~\ref{EXsec}. We note that the definition given here of a Cartesian left additive category is slightly different from the one found in~\cite[Definition 1.2.1]{blute2009cartesian}, which also assumed that the pairing of additive maps be additive, but it is indeed equivalent as explained in~\cite{lemay2018tangent}. Furthermore, by~\cite[Proposition 1.2.2]{blute2009cartesian}, an equivalent axiomatization of a Cartesian left additive category is that of a Cartesian category where every object comes equipped with a commutative monoid structure such that the projection maps are monoid morphisms, which we discuss further in Lemma~\ref{caelem}. Here are some useful identities on how the additive and product structures are compatible:
\begin{lemC}[{\cite[Lemma 2.4]{lemay2018tangent}}]\label{claclemma} In a Cartesian left additive category $\mathbb{X}$:
\begin{enumerate}[\emph{(i)}]
\item $\langle f ,g \rangle + \langle h, k \rangle = \langle f + h, g+k \rangle$ and $\langle 0, 0 \rangle = 0$;
\item If $f: C \to A$ and $g: C \to B$ are additive then $\langle f,g \rangle: C \to A \times B$ is additive;
\item If $h: A \to B$ and $k: C \to D$ are additive then $h \times k: A \times C \to B \times D$ is additive;
\item For any object $A$, the unique map to the terminal object $\top$ is the zero map, $!_A = 0$.
\end{enumerate}
\end{lemC}
\subsection{Cartesian Differential Categories}
Cartesian differential categories are Cartesian left additive categories that come equipped with a differential combinator, which in turn is axiomatized by the basic properties of the directional derivative from multivariable differential calculus. In the following definition, it is important to note that here, following the more recent work on Cartesian differential categories, we've flipped the convention found in~\cite{blute2009cartesian}, so that the linear argument is in the second argument rather than in the first argument.
\begin{defiC}[{\cite[Definition 2.1.1]{blute2009cartesian}}]\label{cartdiffdef} A \textbf{Cartesian differential category} is a Cartesian left additive category equipped with a \textbf{differential combinator} $\mathsf{D}$ of the form
\[ \frac{f : A \to B}{\mathsf{D}[f]: A \times A \to B} \]
verifying the following coherence conditions:
\begin{enumerate}[{\CDCax{\arabic*}},ref={\CDCax{\arabic*}},align=left]
\item\label{CDCax1} $\mathsf{D}[f+g] = \mathsf{D}[f] + \mathsf{D}[g]$ and $\mathsf{D}[0]=0$
\item\label{CDCax2} $\mathsf{D}[f] \circ \langle x, y+z \rangle= \mathsf{D}[f] \circ \langle x, y \rangle + \mathsf{D}[f] \circ \langle x, z \rangle$ and $\mathsf{D}[f] \circ \langle x, 0 \rangle=0$
\item\label{CDCax3} $\mathsf{D}[1_A]=\pi_1$ and $\mathsf{D}[\pi_0] = \pi_0 \circ \pi_1$ and $\mathsf{D}[\pi_1] = \pi_1 \circ \pi_1$
\item\label{CDCax4} $\mathsf{D}[\langle f, g \rangle] = \langle \mathsf{D}[f], \mathsf{D}[g] \rangle$ and $\mathsf{D}[!_A]=!_{A \times A}$
\item\label{CDCax5} $\mathsf{D}[g \circ f] = \mathsf{D}[g] \circ \langle f \circ \pi_0, \mathsf{D}[f] \rangle$
\item\label{CDCax6} $\mathsf{D}\left[\mathsf{D}[f] \right] \circ \left \langle \langle x, y \rangle, \langle 0, z \rangle \right \rangle= \mathsf{D}[f] \circ \langle x, z \rangle$
\item\label{CDCax7} $\mathsf{D}\left[\mathsf{D}[f] \right] \circ \left \langle \langle x, y \rangle, \langle z, 0 \rangle \right \rangle = \mathsf{D}\left[\mathsf{D}[f] \right] \circ \left \langle \langle x, z \rangle, \langle y, 0 \rangle \right \rangle$
\end{enumerate}
\end{defiC}
\noindent
As we will see in Section~\ref{CDCisCdCsec}, a Cartesian difference category is a generalization of a Cartesian differential category. Therefore, we leave the discussion of the intuition of these axioms for later in Section~\ref{CdCsec} below. We also refer to~\cite[Section 4]{blute2009cartesian} for a term calculus which may help better understand the axioms of a Cartesian differential category. We highlight that the last two axioms~\ref{CDCax6} and~\ref{CDCax7} can equivalently be expressed as follows:
\begin{lemC}[{\cite[Proposition 4.2]{cockett2014differential}}] In the presence of the other axioms,~\ref{CDCax6} and~\ref{CDCax7} are equivalent to the following two axioms:
\begin{enumerate}[\CDCax{\arabic*.{a}},ref={\CDCax{\arabic*.{a}}},align=left]
\setcounter{enumi}{5}
\item\label{CDCax6a} $\mathsf{D}\left[\mathsf{D}[f] \right] \circ \left \langle \langle x, 0 \rangle, \langle 0, y \rangle \right \rangle= \mathsf{D}[f] \circ \langle x, y \rangle$
\item\label{CDCax7a} $\mathsf{D}\left[\mathsf{D}[f] \right] \circ \left \langle \langle x, y \rangle, \langle z, w \rangle \right \rangle= \mathsf{D}\left[\mathsf{D}[f] \right] \circ \left \langle \langle x, z \rangle, \langle y, w \rangle \right \rangle$
\end{enumerate}
\end{lemC}
\noindent
The canonical example of a Cartesian differential category is the category of real smooth functions, which we will discuss in Section~\ref{smoothex}. Other interesting examples of Cartesian differential categories can be found throughout the literature such as categorical models of the differential $\lambda$-calculus~\cite{ehrhard2003differential,manzonetto2012categorical}, the subcategory of differential objects of a tangent category~\cite{cockett2014differential}, and the coKleisli category of a differential category~\cite{blute2006differential,blute2009cartesian}.
\section{Change Action Models}\label{CAsec}
Change actions~\cite{alvarez2019fixing,alvarez2019change} have recently been proposed as a setting for reasoning about higher-order incremental computation, based on a discrete notion of differentiation. Together with Cartesian differential categories, they provide the core ideas behind Cartesian difference categories. In this section, we quickly review change actions and change action models, in particular, to highlight where some of the axioms of a Cartesian difference category come from. For more details on change actions, we invite readers to see the original paper~\cite{alvarez2019change}.
\subsection{Change Actions}
The basic intuition for a change action is an object $A$ equipped with a monoid $\Delta A$ which represents the possible ``changes'' or ``updates'' that might be applied to $A$, with the monoid structure on $\Delta A$ representing the capability to compose updates.
\begin{defiC}[{\cite[Section 2]{alvarez2019change}}]
A \textbf{change action} $\cact{A}$ in a Cartesian category $\mathbb{X}$ is a quintuple:
\[\cact{A} \equiv (A, \Delta A, \oplus_A, +_A, 0_A)\]
consisting of two objects $A$ and $\Delta A$, and three maps:
\begin{align*}
\oplus_A: A \times \Delta A \to A && +_A: \Delta A \times \Delta A \to \Delta A && 0_A: \top \to \Delta A
\end{align*}
such that:
\begin{enumerate}[{\CAax{\arabic*}},ref={\CAax{\arabic*}},align=left]
\item\label{CAax1} $(\Delta A, +_A, 0_A)$ is a monoid, that is, the following equalities hold:
\[+_A \circ \langle 1_{\Delta A}, 0_{\Delta A} \circ{} !_{\Delta A} \rangle = 1_{\Delta A} = +_{\Delta A} \circ \langle 0_{\Delta A} \circ{} !_{\Delta_A}, 1_{\Delta A} \rangle\]
\begin{align*}
+_A \circ (1_{\Delta A} \times +_A) = +_A \circ \pair{+_A \circ (1_{\Delta A} \times \pi_0)}{\pi_1 \circ \pi_1}
\end{align*}
\item\label{CAax2} $\oplus_A : A \times \Delta A \to A$ is an action of $\Delta A$ on $A$, that is, the following equalities hold:
\begin{align*}
\oplus_A \circ \langle 1_A, 0_A \circ{} !_A \rangle = 1_A
\end{align*}
\begin{align*}
\oplus_A \circ (1_A \times +_A) = \oplus_A \circ \pair{\oplus_A \circ (1_A \times \pi_0)}{\pi_1 \circ \pi_1}
\end{align*}
\end{enumerate}
\end{defiC}
\noindent
In~\ref{CAax1}, the first line says that $0_A$ is a unit for $+_A$ while the second line is the associativity of $+_A$. In~\ref{CAax2}, the first line is the unit law of the module action, while the second line is the associativity law of the module action. These are highlighted in Lemma~\ref{lem:CAadd} below. If $\cact{A} \equiv (A, \Delta A, \oplus_A, +_A, 0_A)$ is a change action, then for a pair of parallel maps $h: C \to \Delta A$ and $k: C \to \Delta A$, we define their $\Delta A$-sum $h +_{\cact{A}} k: C \to \Delta A$ as follows:
\[ h +_{\cact{A}} k = +_A \circ \langle h, k \rangle\]
Similarly, for a pair of maps $f: C \to A$ and $g: C \to \Delta A$, we define $f \oplus_{\cact{A}} g: C \to A$ as:
\[f \oplus_{\cact{A}} g = \oplus_A \circ \langle f, g \rangle \]
Here, $f \oplus_{\cact{A}} g$ should be thought of as $g$ acting on $f$.
\begin{lem}\label{lem:CAadd}
In a Cartesian category $\mathbb{X}$, let $\cact{A} \equiv (A, \Delta A, \oplus_A, +_A, 0_A)$ be a change action. Then for any suitable maps:
\begin{enumerate}[(\roman{enumi}),ref={\thelem.\roman{enumi}}]
\item $f+_{\cact{A}} (0_A \circ{} !_C) = f = (0_A \circ{} !_C) +_{\cact{A}} f$
\item $f +_{\cact{A}} (g +_{\cact{A}} h) = (f +_{\cact{A}} g) +_{\cact{A}} h$
\item\label{lem:CAadd.0} $f \oplus_{\cact{A}} (0_A \circ{} !_C) = f$
\item\label{lem:CAadd.+} $f\oplus_{\cact{A}} (g +_{\cact{A}} h) = (f \oplus_{\cact{A}} g) \oplus_{\cact{A}} h$
\item\label{lem:CAadd.circ} $(f \oplus_{\cact{A}} g) \circ h = (f \circ h) \oplus_{\cact{A}} (g \circ h)$
\end{enumerate}
\end{lem}
\begin{proof}We leave these as a simple exercises for the reader. Briefly, $(i)$ and $(ii)$ will follow from~\ref{CAax1}, while $(iii)$ and $(iv)$ will follow from~\ref{CAax2}, and $(v)$ follows directly from the definition.
\end{proof}
Change actions give rise to a notion of derivative, with a distinctly ``discrete'' flavour. Given change actions structure on $A$ and $B$, a map $f : A \to B$ can be said to be differentiable when changes to the input (in the sense of elements of $\Delta A$) are mapped to changes to the output (that is, elements of $\Delta B$). In the setting of incremental computation~\cite{alvarez2019fixing,cai2014theory,kelly2016evolving}, this is precisely what it means for $f$ to be incrementalizable, with the derivative of $f$ corresponding to an incremental version of $f$.
\begin{defiC}[{\cite[Definition 1]{alvarez2019change}}]\label{def:derivative}
Let
\[\cact{A} \!\equiv\! (A, \Delta A, \oplus_A, +_A, 0_A) \qquad\text{and}\qquad \cact{B} \!\equiv\! (B, \Delta B, \oplus_B, +_B, 0_B)\]
be change actions in a Cartesian category $\mathbb{X}$. For a map $f: A \to B$, a map of type $\dd[f] : A \times \Delta A \to \Delta B$ is a \textbf{derivative} of $f$ whenever the following equalities hold:
\begin{enumerate}[{\CADax{\arabic*}},ref={\CADax{\arabic*}},align=left]
\item\label{CADax1} $f \circ (x \oplus_{\cact{A}} y) = (f \circ x) \oplus_{\cact{B}} \left(\dd[f] \circ \pair{x}{y}\right)$
\item\label{CADax2} $\dd[f] \circ \pair{x}{y +_{\cact{A}} z} = (\dd[f] \circ \pair{x}{y}) +_{\cact{B}} (\dd[f] \circ \pair{x \oplus_{\cact{A}} y}{z})$ and \\ \noindent$\dd[f] \circ \langle x, 0_B \circ{} !_B \rangle = 0_B \circ{} !_{A \times \Delta A}$
\end{enumerate}
\end{defiC}
\noindent
The second axiom~\ref{CADax2} is also known as regularity~\cite[Definition 2]{alvarez2019change}. The intuition for these axioms will be explained in more detail in Section~\ref{CdCsec} when we explain the axioms of a Cartesian difference category. It is important to note that there is nothing in the definition that says that derivatives are necessarily unique, and therefore a map $f$ could have multiple possible derivatives.
\subsection{Change Action Models}
The chain rule for derivatives still holds for change actions, or in other words, differentiation is compositional. Indeed by~\cite[Lemma 1]{alvarez2019change}, whenever $\dd[f]$ and $\dd[g]$ are derivatives for composable maps $f$ and $g$ respectively, then $\dd[g] \circ \langle f \circ \pi_0, \dd[f] \rangle$ is a derivative for $g \circ f$. As a corollary, change actions in $\mathbb{X}$ form a category where the maps are pairs consisting a map $f$ and a derivative of $f$. For more details on this category of change actions, see~\cite[Section 3]{alvarez2019change}.
\begin{defiC}[{\cite[Section 3]{alvarez2019change}}]
Given a Cartesian category $\mathbb{X}$, define its change actions category $\mathsf{CAct}(\mathbb{X})$ as the category
whose objects are change actions $\cact{A}$ in $\mathbb{X}$ and whose maps $\cact{f} : \cact{A} \to
\cact{B}$ are the pairs $(f, \dd[f])$, where $f : A \to B$ is a map in $\mathbb{X}$ and $\dd[f] :
A \times \Delta A \to \Delta B$ is a derivative for $f$. The identity is $(1_A, \pi_1)$, while composition of $(f, \dd[f])$ and $(g, \dd[g])$ is $(g \circ f, \dd[g] \circ \langle f \circ \pi_0, \dd[f] \rangle)$.
\end{defiC}
For a Cartesian category $\mathbb{X}$, it is straightforward to see that $\mathsf{CAct}(\mathbb{X})$ is also a Cartesian category where the terminal object and the product of objects is the same as in $\mathbb{X}$, where the projection maps are the pairs $(\pi_i, \pi_i \circ \pi_1)$, and where the pairing of maps is given point-wise $\pair{(f, \dd[f])}{(g, \dd[g])} = (\pair{f}{g}, \pair{\dd[f]}{\dd[g]})$. There is an obvious product-preserving forgetful functor $\mathcal{E} : \mathsf{CAct}(\mathbb{X}) \to \mathbb{X}$ sending every change action $(A, \Delta A, \oplus, +, 0)$ to its base object $A$ and every map $(f, \dd[f])$ to the underlying map $f$. Here is a useful lemma which describes the compatibility between change action structure and the product structure.
\begin{lem}\label{lem:CACTprod} Let $\cact{A} \!\equiv\! (A, \Delta A, \oplus_A, +_A, 0_A)$ and $\cact{B} \!\equiv\! (B, \Delta B, \oplus_B, +_B, 0_B)$ be change actions in a Cartesian category $\mathbb{X}$. Then for any suitable maps:
\begin{enumerate}[(\roman{enumi}),ref={\thelem.\roman{enumi}}]
\item\label{lem:CACTprod.oplus} $\pair{f}{g} \oplus_{\cact{A \times B}} \pair{h}{k} = \pair{f \oplus_{\cact{A}} h}{ g \oplus_{\cact{B}} k}$
\item\label{lem:CACTprod.plus} $\pair{f}{g} +_{\cact{A \times B}} \pair{h}{k} = \pair{f +_{\cact{A}} h}{ g +_{\cact{B}} k}$
\item\label{lem:CACTprod.0} $\pair{0_A}{0_B} = 0_{A \times B}$
\end{enumerate}
\end{lem}
\begin{proof} These are immediate from the definition of product structure in $\mathsf{CAct}(\mathbb{X})$, and so we leave this as excercise for the reader.
\end{proof}
As a setting for studying differentiation, the category $\mathsf{CAct}(\mathbb{X})$ is rather lacklustre, since there is no notion of higher derivatives. Instead, one works with change action models. Informally, a change action model consists of a rule which for every object $A$ of $\mathbb{X}$ associates a change action over it, and for every map a choice of a derivative.
\begin{defiC}[{\cite[Definition 5]{alvarez2019change}}]
A \textbf{change action model} is a Cartesian category $\mathbb{X}$ with a product-preserving functor
$\alpha : \mathbb{X} \to \mathsf{CAct}(\mathbb{X})$ that is a section of the forgetful
functor $\mathcal{E}$, that is, $\mathcal{E} \circ \alpha = 1_{\mathbb{X}}$.
\end{defiC}
For brevity, when $A$ is an object of a change action model, we will simply write its associate change action as $\alpha(A) = (A, \Delta A, \oplus_A, +_A)$. Examples of change action models can be found in~\cite{alvarez2019change}. In particular, as was shown in~\cite[Theorem 6]{alvarez2019change}, every Cartesian differential category provides a change model action. We will generalize this result, and show in Section~\ref{CdCisCAsec} that a Cartesian difference category also always provides a change action model.
\section{Cartesian Difference Categories}\label{sec:CdC}
In this section, we introduce \emph{Cartesian difference categories}, which are generalizations of Cartesian differential categories. Examples of Cartesian difference categories can be found in Section~\ref{EXsec}.
\subsection{Infinitesimal Extensions in Left Additive Categories}
We first introduce infinitesimal extensions, which is an operator that turns a map into an ``infinitesimal'' version of itself, in the sense that every map coincides with its Taylor approximation on infinitesimal elements.
\begin{defi} A Cartesian left additive category $\mathbb{X}$ is said to have an \textbf{infinitesimal extension} $\varepsilon$ if every pair of objects $A$ and $B$, there is a function $\varepsilon: \mathbb{X}(A,B) \to \mathbb{X}(A,B)$ such that:
\begin{enumerate}[{\Eax{\arabic*}},ref={\Eax{\arabic*}},align=left]
\item\label{Eax1} $\varepsilon$ is a monoid morphism, that is, $\varepsilon(f+g) = \varepsilon(f) + \varepsilon(g)$ and $\varepsilon(0)=0$
\item\label{Eax2} $\varepsilon(g \circ f) = \varepsilon(g) \circ f$
\item\label{Eax3} $\varepsilon(\pi_0) = \pi_0 \circ \varepsilon(1_{A\times B})$ and $ \varepsilon(\pi_1) = \pi_1 \circ \varepsilon(1_{A\times B})$
\end{enumerate}
\end{defi}
\noindent
By~\ref{Eax1}, $\varepsilon(1_A)$ is an additive map, while by~\ref{Eax2}, it follows that $\varepsilon(f) = \varepsilon(1_B) \circ f$. In light of this, it turns out that infinitesimal extensions can equivalently be described as a class of additive maps $\varepsilon_A: A \to A$. The equivalence is given by setting $\varepsilon(f) = \varepsilon_B \circ f$ and $\varepsilon_A = \varepsilon(1_A)$.
\begin{lem}\label{lem:epbij} For a Cartesian left additive category $\mathbb{X}$, the following are in bijective correspondence:
\begin{enumerate}[(\roman{enumi}.),ref={\thelem.\roman{enumi}}]
\item An infinitesimal extension $\varepsilon$ on $\mathbb{X}$;
\item A family of maps $\varepsilon_A: A \to A$ indexed by objects $A$, such that $\varepsilon_A$ is additive and $\varepsilon_{A \times B}=\varepsilon_A \times \varepsilon_B$.
\end{enumerate}
Therefore, for any map $f: A \to B$, $\varepsilon(f) = \varepsilon_A \circ f$.
\end{lem}
\begin{proof} Let $\varepsilon$ be an infinitesimal extension. Then for each object $A$, define $\varepsilon_A: A \to A$ as $\varepsilon_A = \varepsilon(1_A)$. Since $\varepsilon$ preserves the additive structure and $\varepsilon(g \circ f) = \varepsilon(g) \circ f$, it follows that:
\begin{align*}
\varepsilon_A \circ (f + g) &=~ \varepsilon(1_A) \circ (f + g) \\
&=~ \varepsilon\left(1_A \circ (f+g) \right) \tag*{\ref{Eax2}} \\
&=~ \varepsilon(f+g) \\
&=~\varepsilon(f) + \varepsilon(g) \tag*{\ref{Eax1}}\\
&=~\varepsilon\left(1_A \circ f \right) + \varepsilon\left(1_A \circ g \right) \\
&=~\left( \varepsilon(1_A) \circ f \right) \circ \left(\varepsilon(1_A) \circ g \right) \tag*{\ref{Eax2}} \\
&=~\left( \varepsilon_A \circ f \right) + \left( \varepsilon_A \circ g \right) \\\\
\varepsilon_A \circ 0 &=~\varepsilon(1_A) \circ 0 \\
&=~ \varepsilon\left(1_A \circ 0 \right) \tag*{\ref{Eax2}} \\
&=~\varepsilon(0) \\
&=~ 0 \tag*{\ref{Eax1}}
\end{align*}
So therefore, $\varepsilon_A$ is additive. Now for a pair of objects $A$ and $B$, we have the following:
\begin{align*}
\pi_0 \circ \varepsilon_{A \times B} &=~\pi_0 \circ \varepsilon(1_{A\times B})\\
&=~ \varepsilon(\pi_0) \tag*{\ref{Eax3}} \\
&=~\varepsilon(1_A \circ \pi_0) \\
&=~ \varepsilon(1_A) \circ \pi_0 \tag*{\ref{Eax2}} \\
&=~ \varepsilon_A \circ \pi_0
\end{align*}
So $\pi_0 \circ \varepsilon_{A \times B} = \varepsilon_A \circ \pi_0$, and similarly, $\pi_1 \circ \varepsilon_{A \times B} = \varepsilon_B \circ \pi_1$. Then by the universal property of the product, it follows that $\varepsilon_{A \times B}=\varepsilon_A \times \varepsilon_B$.
Conversely, suppose that each object $A$ comes equipped with an additive map $\varepsilon_A: A \to A$ such that $\varepsilon_{A \times B}=\varepsilon_A \times \varepsilon_B$. Define $\varepsilon: \mathbb{X}(A,B) \to \mathbb{X}(A,B)$ as $\varepsilon(f) = \varepsilon_B \circ f$. Since $\varepsilon_B$ is additive, it follows that:
\begin{align*}
\varepsilon(f + g) &=~ \varepsilon_B \circ (f + g) =~ \left( \varepsilon_B \circ f \right) + \left( \varepsilon_B \circ g \right) =~ \varepsilon(f) + \varepsilon(g) \end{align*}
\begin{align*}
\varepsilon(0) &=~\varepsilon_B \circ 0 =~0
\end{align*}
So $\varepsilon$ is a monoid morphism, and so~\ref{Eax1} holds. Next, it is straightforward by definition that $\varepsilon(g \circ f) = \varepsilon(g) \circ f$, so~\ref{Eax2} holds. So it remains to show that $\varepsilon$ is compatible with the projections. Note that $\varepsilon_{A \times B}=\varepsilon_A \times \varepsilon_B$ implies that $\pi_0 \circ \varepsilon_{A \times B} = \varepsilon_A \circ \pi_0$ and $\pi_1 \circ \varepsilon_{A \times B} = \varepsilon_B \circ \pi_1$. Therefore, we have that:
\begin{align*}
\varepsilon(\pi_0) &=~ \varepsilon_{A} \circ \pi_0 =~ \pi_0 \circ \varepsilon_{A \times B} \\
\varepsilon(\pi_1) &=~ \varepsilon_{B} \circ \pi_1 =~ \pi_1 \circ \varepsilon_{A \times B}
\end{align*}
So we have that~\ref{Eax3} holds. Therefore, $\varepsilon$ is an infinitesimal extension.
Lastly, we need to show that these constructions are inverses of each other. Starting with an infinitesimal extension $\varepsilon$, we have that:
\begin{align*}
\varepsilon(f) &=~\varepsilon(1_B \circ f) =~ \varepsilon(1_B) \circ f =~\varepsilon_B \circ f
\end{align*}
While in the other direction, it is automatic that $\varepsilon(1_A) = \varepsilon_A$. So we conclude that infinitesimal extensions are bijective correspondence with a family of additive maps $\varepsilon_A: A \to A$ such that $\varepsilon_{A \times B}=\varepsilon_A \times \varepsilon_B$.
\end{proof}
As an immediate consequence of the previous lemma, it follows that infinitesimal extensions are compatible with the product structure.
\begin{lem}\label{lem:ep-pair} Let $\mathbb{X}$ be a Cartesian left additive category with an infinitesimal extension $\varepsilon$. Then $\varepsilon(\pair{f}{g}) = \pair{\varepsilon(f)}{\varepsilon(g)}$ and $\varepsilon(h \times k) = \varepsilon(h) \times \varepsilon(k)$.
\end{lem}
\begin{proof} We compute that:
\begin{align*}
\varepsilon(\pair{f}{g}) &=~ \varepsilon_{A \times B} \circ \pair{f}{g} \tag{Lemma~\ref{lem:epbij}} \\
&=~ (\varepsilon_A \times \varepsilon_B) \circ \pair{f}{g} \tag{Lemma~\ref{lem:epbij}} \\
&=~ \pair{\varepsilon_A \circ f}{\varepsilon_B \circ g} \\
&=~ \pair{\varepsilon(f)}{\varepsilon(g)} \tag{Lemma~\ref{lem:epbij}}
\end{align*}
So $\varepsilon(\pair{f}{g}) = \pair{\varepsilon(f)}{\varepsilon(g)}$. By similar calculations, we also have that:
\begin{align*}
\varepsilon(h \times k) &=~ \varepsilon_{A \times B} \circ (h \times k) \tag{Lemma~\ref{lem:epbij}} \\
&=~ (\varepsilon_A \times \varepsilon_B) \circ (h \times k) \tag{Lemma~\ref{lem:epbij}} \\
&=~ (\varepsilon_A \circ h) \times (\varepsilon_B \circ k) \\
&=~ \varepsilon(h) \times \varepsilon(k) \tag{Lemma~\ref{lem:epbij}}
\end{align*}
So $\varepsilon(h \times k) = \varepsilon(h) \times \varepsilon(k)$.
\end{proof}
Infinitesimal extensions equip each object with a canonical change action structure:
\begin{lem}\label{caelem} Let $\mathbb{X}$ be a Cartesian left additive category with infinitesimal extension $\varepsilon$. For every object $A$, define the three maps $\oplus_A: A \times A \to A$, $+_A: A \times A \to A$, and $0_A: \top \to A$ respectively as follows:
\begin{align*}
\oplus_A = \pi_0 + \varepsilon(\pi_1) && +_A = \pi_0 + \pi_1 && 0_A = 0
\end{align*}
Then $(A, A, \oplus_A, +_A, 0_A)$ is a change action in $\mathbb{X}$.
\end{lem}
\begin{proof} That $(A, +_A, 0_A)$ is a commutative monoid was shown in~\cite[Proposition 1.2.2]{blute2009cartesian}. Thus~\ref{CAax1} holds. It remains to show that $\oplus_A$ is an action, which follows directly from the fact that $\varepsilon$ preserves the additive structure. So we first that compute:
\begin{align*}
\oplus_A \circ \langle 1_A, 0_A \circ{} !_A \rangle &=~\left( \pi_0 + \varepsilon(\pi_1) \right) \circ \langle 1_A, 0 \rangle \\
&=~ \pi_0 \circ \langle 1_A, 0_A \rangle + \varepsilon(\pi_1) \circ \langle 1_A, 0 \rangle \\
&=~ \pi_0 \circ \langle 1_A, 0_A \rangle + \varepsilon\left(\pi_1 \circ \langle 1_A, 0 \rangle \right) \tag*{\ref{Eax2}} \\
&=~ 1_A + \varepsilon(0) \\
&=~ 1_A + 0 \tag*{\ref{Eax1}} \\
&=~ 1_A
\end{align*}
So $ \oplus_A \circ \langle 1_A, 0_A \circ{} !_A \rangle = 1_A$. Next we compute that:
\begin{align*}
&\oplus_A \circ \pair{\oplus_A \circ (1_A \times \pi_0)}{\pi_1 \circ \pi_1} =~ \left( \pi_0 + \varepsilon(\pi_1) \right) \circ \pair{\oplus_A \circ (1_A \times \pi_0)}{\pi_1 \circ \pi_1} \\
&=~ \pi_0 \circ \pair{\oplus_A \circ (1_A \times \pi_0)}{\pi_1 \circ \pi_1} + \varepsilon(\pi_1) \circ \pair{\oplus_A \circ (1_A \times \pi_0)}{\pi_1 \circ \pi_1} \\
&=~ \pi_0 \circ \pair{\oplus_A \circ (1_A \times \pi_0)}{\pi_1 \circ \pi_1} + \varepsilon\left(\pi_1 \circ \pair{\oplus_A \circ (1_A \times \pi_0)}{\pi_1 \circ \pi_1} \right)\tag*{\ref{Eax2}} \\
&=~ \oplus_A \circ (1_A \times \pi_0) + \varepsilon\left(\pi_1 \circ \pi_1 \right) \\
&=~ \left( \pi_0 + \varepsilon(\pi_1) \right) \circ (1_A \times \pi_0) + \varepsilon\left(\pi_1 \circ \pi_1 \right) \\
&=~ \pi_0 \circ (1_A \times \pi_0) + \varepsilon(\pi_1) \circ (1_A \times \pi_0) + \varepsilon\left(\pi_1 \circ \pi_1 \right) \\
&=~ \pi_0 \circ (1_A \times \pi_0) + \varepsilon\left(\pi_1 \circ (1_A \times \pi_0) \right) + \varepsilon\left(\pi_1 \circ \pi_1 \right) \tag*{\ref{Eax2}} \\
&=~ \pi_0 + \varepsilon(\pi_1 \circ \pi_0) + \varepsilon\left(\pi_1 \circ \pi_1 \right) \\
&=~ \pi_0 + \varepsilon\left( \pi_1 \circ \pi_0 + \pi_1 \circ \pi_1 \right) \tag*{\ref{Eax1}} \\
&=~ \pi_0 + \varepsilon\left( \pi_1 \circ \left( \pi_0 + \pi_1 \right) \right) \tag{$\pi_1$ is additive} \\
&=~ \pi_0 + \varepsilon\left( \pi_1 \circ +_A \right) \\
&=~ \pi_0 \circ (1_A \times +_A) + \varepsilon\left( \pi_1 \circ (1_A \times +_A) \right) \\
&=~ \pi_0 \circ (1_A \times +_A) + \varepsilon\left( \pi_1 \right) \circ (1_A \times +_A) \tag*{\ref{Eax2}} \\
&=~ \left( \pi_0 + \varepsilon(\pi_1) \right) \circ (1_A \times +_A) \\
&=~ \oplus_A \circ (1_A \times +_A)
\end{align*}
So $\oplus_A \circ (1_A \times +_A) = \oplus_A \circ \pair{\oplus_A \circ (1_A \times \pi_0)}{\pi_1 \circ \pi_1}$. Thus~\ref{CAax2} holds. So we conclude that $(A, A, \oplus_A, +_A, 0_A)$ is a change action. \hfill
\end{proof}
It is important to note that in an arbitrary Cartesian left additive category, $\oplus_A$, $+_A$, and $0_A$ are not necessarily natural transformations. That said, $+_A$ and $0_A$ are natural with respect to additive maps, that is, if $f: A \to B$ is additive then $f \circ 0_A = 0_B$ and $f \circ +_A = +_B \circ (f \times f)$. Setting $\cact{A} \equiv (A, A, \oplus_A, +_A, 0_A)$, we note that $f \oplus_{\cact{A}} g = f + \varepsilon(g)$ and $f +_{\cact{A}} g = f + g$, and so in particular $+_{\cact{A}} = +$. Therefore, from now on we will omit the subscripts and simply write $\oplus$ and $+$.
For every Cartesian left additive category, there are always at least two possible infinitesimal extensions: one given by setting the infinitesimal extension of a map to be zero and another given by setting the infinitesimal extension of a map to be itself.
\begin{lem} For any Cartesian left additive category $\mathbb{X}$,
\begin{enumerate}[(\roman{enumi}).,ref={\thelem.\roman{enumi}}]
\item Setting $\varepsilon(f) = 0$ defines an infinitesimal extension on $\mathbb{X}$ and therefore in this case, $\oplus_A = \pi_0$ and $f \oplus g = f$.
\item Setting $\varepsilon(f) = f$ defines an infinitesimal extension on $\mathbb{X}$ and therefore in this case, $\oplus_A = +_A$ and $f \oplus g = f + g$.
\end{enumerate}
\end{lem}
\begin{proof} These are both straightforward to check and we leave it as an exercise. \hfill
\end{proof}
We note that while these examples of infinitesimal extensions may seem trivial, they are both very important as they will give rise to key examples of Cartesian difference categories.
\subsection{Cartesian Difference Categories}\label{CdCsec} Here we introduce Cartesian difference categories, the main novel contribution of this paper.
\begin{defi} A \textbf{Cartesian difference category} is a Cartesian left additive category with an infinitesimal extension $\varepsilon$ which is equipped with a \textbf{difference combinator} $\dd$ of the form:
\[ \frac{f : A \to B}{\dd[f]: A \times A \to B} \]
verifying the following coherence conditions:
\begin{enumerate}[{\CdCax{\arabic*}},ref={\CdCax{\arabic*}},align=left]
\setcounter{enumi}{-1}
\item $f \circ (x + \varepsilon(y)) = f \circ x + \varepsilon\left( \dd[f] \circ \langle x, y \rangle \right)$\label{CdCax0}
\item $\dd[f+g] = \dd[f] + \dd[g]$, $\dd[0] = 0$, and $\dd[\varepsilon(f)] = \varepsilon(\dd[f])$\label{CdCax1}
\item $\dd[f] \circ \langle x, y + z \rangle = \dd[f] \circ \langle x, y \rangle + \dd[f] \circ \langle x + \varepsilon(y), z \rangle$ and $\dd[f] \circ \langle x, 0 \rangle = 0$\label{CdCax2}
\item $\dd[1_A] = \pi_1$ and $\dd[\pi_0] = \pi_0 \circ \pi_1$ and $\dd[\pi_1] = \pi_1 \circ \pi_1$\label{CdCax3}
\item $\dd[\langle f, g \rangle] = \langle \dd[f], \dd[g] \rangle$\label{CdCax4}
\item $\dd[g \circ f] = \dd[g] \circ \langle f \circ \pi_0, \dd[f] \rangle$\label{CdCax5}
\item $\dd\left[\dd[f] \right] \circ \left \langle \langle x, y \rangle, \langle 0, z \rangle \right \rangle= \dd[f] \circ \langle x + \varepsilon(y), z \rangle$\label{CdCax6}
\item $ \dd\left[\dd[f] \right] \circ \left \langle \langle x, y \rangle, \langle z, 0 \rangle \right \rangle= \dd\left[\dd[f] \right] \circ \left \langle \langle x, z \rangle, \langle y, 0 \rangle \right \rangle$\label{CdCax7}
\end{enumerate}
We say that $\dd[f]$ is the derivative of $f$.
\end{defi}
Before giving some intuition on the axioms~\ref{CdCax0} to~\ref{CdCax7}, we first observe that one could have used change action notation $\oplus$ to express~\ref{CdCax0},~\ref{CdCax2}, and~\ref{CdCax6}:
\begin{description}[leftmargin=*, widest=a]
\item[\ref{CdCax0}] $f \circ (x \oplus y) = f \circ x \oplus \dd[f] \circ \langle x, y \rangle$
\item[\ref{CdCax2}] $\dd[f] \circ \langle x, y + z \rangle = \dd[f] \circ \langle x, y \rangle + \dd[f] \circ \langle x \oplus y, z \rangle$ and $\dd[f] \circ \langle x, 0 \rangle = 0$
\item[\ref{CdCax6}] $\dd\left[\dd[f] \right] \circ \left \langle \langle x, y \rangle, \langle 0, z \rangle \right \rangle= \dd[f] \circ \langle x \oplus y, z \rangle$
\end{description}
And also, just like Cartesian differential categories,~\ref{CdCax6} and~\ref{CdCax7} have alternative equivalent expressions.
\begin{lem}%
\label{lem:cdc6a}
In the presence of the other axioms,~\ref{CdCax6} and~\ref{CdCax7} are equivalent to:
\begin{enumerate}[\CdCax{\arabic*.{a}},ref={\CdCax{\arabic*.{a}}},align=left]
\setcounter{enumi}{5}
\item $\dd\left[\dd[f] \right] \circ \left \langle \langle x, 0 \rangle, \langle 0, y \rangle \right \rangle = \dd[f] \circ \langle x, y \rangle$\label{CdCax6a}
\item $\dd\left[\dd[f] \right] \circ \left \langle \langle x, y \rangle, \langle z, w \rangle \right \rangle= \dd\left[\dd[f] \right] \circ \left \langle \langle x, z \rangle, \langle y, w \rangle \right \rangle$\label{CdCax7a}
\end{enumerate}
\end{lem}
\begin{proof} The proof is essentially the same as~\cite[Proposition 4.2]{cockett2014differential}. Assume that the other axioms~\ref{CdCax0} to~\ref{CdCax5} hold. Suppose that~\ref{CdCax6} and~\ref{CdCax7} also hold. We first compute~\ref{CdCax6a} using~\ref{CdCax6}:
\begin{align*}
\dd\left[\dd[f] \right] \circ \left \langle \langle x, 0 \rangle, \langle 0, y \rangle \right \rangle &=~ \dd[f] \circ \langle x + \varepsilon(0), y \rangle \tag*{\ref{CdCax6}} \\
&=~ \dd[f] \circ \langle x + 0, y \rangle \tag*{\ref{Eax1}} \\
&=~ \dd[f] \circ \langle x, y \rangle
\end{align*}
Next we compute~\ref{CdCax7a} using~\ref{CdCax2},~\ref{CdCax6}, and~\ref{CdCax7}:
\begin{align*} &\dd\left[\dd[f] \right] \circ \left \langle \langle x, y \rangle, \langle z, w \rangle \right \rangle =~ \dd\left[\dd[f] \right] \circ \pair{\pair{x}{y}}{\pair{z}{0} + \pair{0}{w}} \\
&=~ \dd\left[\dd[f] \right] \circ \left \langle \langle x, y \rangle, \langle z, 0 \rangle\right \rangle + \dd\left[\dd[f] \right] \circ \left \langle \langle x, y \rangle + \varepsilon(\pair{z}{0}), \langle 0, w \rangle\right \rangle \tag*{\ref{CdCax2}} \\
&=~ \dd\left[\dd[f] \right] \circ \left \langle \langle x, 0 \rangle, \langle y, 0 \rangle\right \rangle + \dd\left[\dd[f] \right] \circ \left \langle \langle x, y \rangle + \pair{\varepsilon(z)}{\varepsilon(0)} , \langle 0, w \rangle\right \rangle \tag{Lemma~\ref{lem:ep-pair}} \\
&=~ \dd\left[\dd[f] \right] \circ \left \langle \langle x, 0 \rangle, \langle y, 0 \rangle\right \rangle + \dd\left[\dd[f] \right] \circ \left \langle \langle x, y \rangle + \pair{\varepsilon(z)}{0} , \langle 0, w \rangle\right \rangle \tag*{\ref{Eax1}} \\
&=~ \dd\left[\dd[f] \right] \circ \left \langle \langle x, 0 \rangle, \langle y, 0 \rangle\right \rangle + \dd\left[\dd[f] \right] \circ \left \langle \langle x + \varepsilon(z), y \rangle , \langle 0, w \rangle\right \rangle \\
&=~ \dd\left[\dd[f] \right] \circ \left \langle \langle x, z \rangle, \langle y, 0 \rangle\right \rangle + \dd[f] \circ \pair{x + \varepsilon(z) + \varepsilon(y)}{w} \tag*{\ref{CdCax7} +~\ref{CdCax6}} \\
&=~ \dd\left[\dd[f] \right] \circ \left \langle \langle x, z \rangle, \langle y, 0 \rangle\right \rangle + \dd[f] \circ \pair{x + \varepsilon(y) + \varepsilon(z)}{w} \\
&=~ \dd\left[\dd[f] \right] \circ \left \langle \langle x, z \rangle, \langle y, 0 \rangle\right \rangle + \dd\left[\dd[f] \right] \circ \left \langle \langle x + \varepsilon(y), z \rangle, \langle 0, w \rangle\right \rangle \tag*{\ref{CdCax6}} \\
&=~ \dd\left[\dd[f] \right] \circ \left \langle \langle x, z \rangle, \langle y, 0 \rangle\right \rangle + \dd\left[\dd[f] \right] \circ \left \langle \langle x + \varepsilon(y), z + 0 \rangle, \langle 0, w \rangle\right \rangle \\
&=~ \dd\left[\dd[f] \right] \circ \left \langle \langle x, z \rangle, \langle y, 0 \rangle\right \rangle + \dd\left[\dd[f] \right] \circ \left \langle \langle x + \varepsilon(y), z + \varepsilon(0) \rangle, \langle 0, w \rangle\right \rangle \tag*{\ref{Eax1}} \\
&=~ \dd\left[\dd[f] \right] \circ \left \langle \langle x, z \rangle, \langle y, 0 \rangle\right \rangle + \dd\left[\dd[f] \right] \circ \left \langle \langle x , z \rangle + \langle \varepsilon(y), \varepsilon(0) \rangle, \langle 0, w \rangle\right \rangle \\
&=~ \dd\left[\dd[f] \right] \circ \left \langle \langle x, z \rangle, \langle y, 0 \rangle\right \rangle + \dd\left[\dd[f] \right] \circ \left \langle \langle x , z \rangle + \varepsilon(\langle y, 0 \rangle), \langle 0, w \rangle\right \rangle \tag{Lemma~\ref{lem:ep-pair}} \\
&=~ \dd\left[\dd[f] \right] \circ \pair{\pair{x}{z}}{\pair{y}{0} + \pair{0}{w}} \tag*{\ref{CdCax2}} \\
&=~ \dd\left[\dd[f] \right] \circ \left \langle \langle x, y \rangle, \langle z, w \rangle \right \rangle
\tag*{\ref{CdCax7}}
\end{align*}
On the other, suppose instead that~\ref{CdCax6a} and~\ref{CdCax7a} hold. Then~\ref{CdCax7} follows immediately from~\ref{CdCax7a} by setting the last term $w=0$.
\begin{align*}
\dd\left[\dd[f] \right] \circ \left \langle \langle x, y \rangle, \langle z, 0 \rangle \right \rangle= \dd\left[\dd[f] \right] \circ \left \langle \langle x, z \rangle, \langle y, 0 \rangle \right \rangle \tag*{\ref{CdCax7a}}
\end{align*}
Next we compute~\ref{CdCax6} using~\ref{CdCax2},~\ref{CdCax6a}, and~\ref{CdCax7a}.
\begin{align*}
& \dd\left[\dd[f] \right] \circ \left \langle \langle x, y \rangle, \langle 0, z \rangle \right \rangle =~ \dd\left[\dd[f] \right] \circ \left \langle \langle x, 0 \rangle, \langle y, z \rangle \right \rangle \tag*{\ref{CdCax7a}} \\
&=~ \dd\left[\dd[f] \right] \circ \pair{\pair{x}{0}}{\pair{y}{0} + \pair{0}{z}} \\
&=~ \dd\left[\dd[f] \right] \circ \left \langle \langle x, 0 \rangle, \langle y, 0 \rangle\right \rangle + \dd\left[\dd[f] \right] \circ \left \langle \langle x, 0 \rangle + \varepsilon(\pair{y}{0}) , \langle 0, z \rangle\right \rangle \tag*{\ref{CdCax2}} \\
&=~ \dd\left[\dd[f] \right] \circ \left \langle \langle x, y \rangle, \langle 0, 0 \rangle\right \rangle + \dd\left[\dd[f] \right] \circ \left \langle \langle x, 0 \rangle + \varepsilon(\pair{y}{0}) , \langle 0, z \rangle\right \rangle \tag*{\ref{CdCax7a}} \\
&=~ 0 + \dd\left[\dd[f] \right] \circ \left \langle \langle x, 0 \rangle + \varepsilon(\pair{y}{0}) , \langle 0, z \rangle\right \rangle \tag*{\ref{CdCax2}} \\
&=~ \dd\left[\dd[f] \right] \circ \left \langle \langle x, 0 \rangle + \pair{\varepsilon(y)}{\varepsilon(0)} , \langle 0, z \rangle\right \rangle \tag{Lemma~\ref{lem:ep-pair}} \\
&=~ \dd\left[\dd[f] \right] \circ \left \langle \langle x, 0 \rangle + \pair{\varepsilon(y)}{0} , \langle 0, z \rangle\right \rangle \tag*{\ref{Eax1}} \\
&=~ \dd\left[\dd[f] \right] \circ \left \langle \langle x + \varepsilon(y), 0 \rangle , \langle 0, z \rangle\right \rangle \\
&=~ \dd[f] \circ \langle x + \varepsilon(y), z \rangle \tag*{\ref{CdCax6a}}
\end{align*}
So we conclude that~\ref{CdCax6} and~\ref{CdCax7} are equivalent to~\ref{CdCax6a} and~\ref{CdCax7a}.
\hfill
\end{proof}
The keen eyed reader will notice that the axioms of a Cartesian difference category are very similar to the axioms of a Cartesian differential category. Indeed,~\ref{CdCax1},~\ref{CdCax3},~\ref{CdCax4}, \ref{CdCax5}, and~\ref{CdCax7} are the same as their Cartesian differential category counterpart. The axioms which are different are~\ref{CdCax2} and~\ref{CdCax6} where the infinitesimal extension $\varepsilon$ is now included, and also there is the new extra axiom~\ref{CdCax0}. On the other hand, interestingly enough,~\ref{CDCax6a} is the same as~\ref{CdCax6a}. We also point out that writing out~\ref{CdCax0} and~\ref{CdCax2} using change action notion, we see that these axioms are precisely~\ref{CADax1} and~\ref{CADax2} respectively. To better understand~\ref{CdCax0} to~\ref{CdCax7} it may be useful to write them out using element-like notation. In element-like notation,~\ref{CdCax0} is written as:
\[
f(x + \varepsilon(y)) = f(x) + \varepsilon\left(\dd[f](x,y) \right)
\]
This condition can be read as a generalization of the Kock-Lawvere axiom that characterizes the derivative from synthetic differential geometry~\cite{kock2006synthetic}. Broadly speaking, the Kock-Lawvere axiom states that, for any map $f : \mathcal{R} \to \mathcal{R}$ and any $x \in \mathcal{R}$ and $d \in \mathcal{D}$, there exists a unique $f'(x) \in \mathcal{R}$ verifying $f(x + d) = f(x) + d \cdot f'(x)$, where $\mathcal{D}$ is the subset of $\mathcal{R}$ consisting of infinitesimal elements. It is by analogy with the Kock-Lawvere axiom that we refer to $\varepsilon$ as an ``infinitesimal extension'' as it can be thought of as embedding the space $A$ into a subspace $\varepsilon(A)$ of infinitesimal elements.
\ref{CdCax1} states that the differential of a sum of maps is the sum of differentials, and similarly for zero maps and the infinitesimal extension of a map.
\ref{CdCax2} is the first crucial difference between a Cartesian difference category and a Cartesian differential category. In a Cartesian differential category, the differential of a map is assumed to be additive in its second argument. In a Cartesian difference category, just as derivatives for change actions, while the differential is still required to preserve zeros in its second argument, it is only additive ``up to a small perturbation'', that is:
\[\dd[f](x,y+z) = \dd[f](x,y) + \dd[f](x + \varepsilon(y), z)\]
\ref{CdCax3} tells us what the differential of the identity and projection maps are, while~\ref{CdCax4} says that the differential of a pairing of maps is the pairing of their differentials.~\ref{CdCax5} is the chain rule which expresses what the differential of a composition of maps is:
\[ \dd[g \circ f](x,y) = \dd[g](f(x), \dd[f](x,y))\]
The last two axioms,~\ref{CdCax6} and~\ref{CdCax7} tell us how to work with second order differentials.~\ref{CdCax6} is expressed as follows:
\[\dd \left[ \dd[f] \right]\left( x,y, 0,z \right) = \dd[f](x + \varepsilon(y), z) \]
and finally~\ref{CdCax7} is expressed as:
\[\dd \left[ \dd[f] \right]\left( x,y, z, 0 \right) = \dd \left[ \dd[f] \right]\left( x, z, y, 0 \right) \]
It is interesting to note that while~\ref{CdCax6} is different from~\ref{CDCax6}, its alternative version~\ref{CdCax6a} is the same as~\ref{CDCax6a}.
\[\dd \left[ \dd[f] \right]\left( (x,0), (0,y) \right) = \dd[f](x, z) \]
The interplay between~\ref{CdCax0},~\ref{CdCax2}, and~\ref{CdCax6} gives rise to some remarkable and
counter-intuitive consequences.
\begin{lem}%
\label{lem:d-epsilon} In a Cartesian difference category $\mathbb{X}$, for any map $f : A \to B$, the following equalities hold for any suitable maps:
\begin{enumerate}[(\roman{enumi}).,ref={\thelem.\roman{enumi}}]
\item $\dd[f] \circ \pair{x}{\varepsilon(y)} = \varepsilon(\dd[f]) \circ
\pair{x}{y}$\label{lem:d-epsilon-i}
\item $\varepsilon(\dd[f]) \circ \pair{x + \varepsilon(y)}{z}
= \varepsilon(\dd[f]) \circ \pair{x + \varepsilon^2(y)}{z}$\label{lem:d-epsilon-ii}
\item $\varepsilon^2\left(\dd\left[ \dd[f] \right] \right) \circ \four{x}{y}{z}{0}
= \varepsilon^3\left(\dd\left[ \dd[f] \right] \right) \circ \four{x}{y}{z}{0}$\label{lem:d-epsilon-iii}
\end{enumerate}
\end{lem}
\begin{proof} These are mostly straightforward calculations. We start by computing (i):
\begin{align*}
\dd[f] \circ \pair{x}{\varepsilon(y)} &=~ \dd[f] \circ \left( \pair{x}{0} + \pair{0}{\varepsilon(y)} \right) \\
&=~ \dd[f] \circ \left( \pair{x}{0} + \pair{\varepsilon(0)}{\varepsilon(y)} \right) \tag*{\ref{Eax1}} \\
&=~ \dd[f] \circ \left( \pair{x}{0} + \varepsilon(\pair{0}{y}) \right) \tag{Lemma~\ref{lem:ep-pair}} \\
&=~ \dd[f]\circ \pair{x}{0} + \varepsilon\left(\dd^2[f] \circ \four{x}{0}{0}{y} \right) \tag*{\ref{CdCax0}} \\
&=~0 + \varepsilon(\dd[f] \circ \pair{x}{y}) \tag*{\ref{CdCax2} +~\ref{CdCax6a}} \\
&=~ \varepsilon(\dd[f]) \circ \pair{x}{y} \tag*{\ref{Eax2}} \\
\end{align*}
Next, we use (i) to compute (ii):
\begin{align*}
\varepsilon(\dd[f]) \circ \pair{x + \varepsilon(y)}{z} &=~ \varepsilon\left( \dd[f] \circ \pair{x + \varepsilon(y)}{z} \right) \tag*{\ref{Eax2}} \\
&=~ \varepsilon\left( \dd\left[ \dd[f] \right] \circ \four{x}{y}{0}{z} \right) \tag*{\ref{CdCax6}} \\
&=~ \varepsilon\left( \dd\left[ \dd[f] \right] \circ \four{x}{0}{y}{z} \right) \tag*{\ref{CdCax7a}} \\
&=~ \varepsilon\left( \dd\left[ \dd[f] \right] \right) \circ \four{x}{0}{y}{z} \tag*{\ref{Eax2}} \\
&=~ \dd\left[ \dd[f] \right] \circ \pair{\pair{x}{0}}{\varepsilon(\pair{y}{z})} \tag{Lemma~\ref{lem:d-epsilon-i}} \\
&=~ \dd\left[ \dd[f] \right] \circ \four{x}{0}{\varepsilon(y)}{\varepsilon(z)} \tag{Lemma~\ref{lem:ep-pair}} \\
&=~ \dd\left[ \dd[f] \right] \circ \four{x}{\varepsilon(y)}{0}{\varepsilon(z)} \tag*{\ref{CdCax7a}} \\
&=~ \dd[f] \circ \pair{x + \varepsilon^2(y)}{\varepsilon(z)} \tag*{\ref{CdCax6}} \\
&=~ \varepsilon(\dd[f]) \circ \pair{x + \varepsilon^2(y)}{z}
\tag{Lemma~\ref{lem:d-epsilon-i}}
\end{align*}
Lastly for (iii), we use (i) and (ii):
\begin{align*}
\varepsilon^2\left(\dd\left[ \dd[f] \right] \right) \circ \four{x}{y}{z}{0} &=~ \dd\left[ \dd[\varepsilon^2(f)] \right] \circ \four{x}{y}{z}{0} \tag*{\ref{CdCax1}} \\
&=~ \dd\left[ \dd[\varepsilon^2(f)] \right] \circ \four{x}{z}{y}{0} \tag*{\ref{CdCax7a}} \\
&=~ \varepsilon \left( \dd\left[ \dd[\varepsilon(f)] \right] \right) \circ \four{x}{z}{y}{0} \tag*{\ref{CdCax1}} \\
&=~ \dd\left[ \dd[\varepsilon(f)] \right] \circ \pair{\pair{x}{z}}{\varepsilon(\pair{y}{0})} \tag{Lemma~\ref{lem:d-epsilon-i}} \\
&=~ \dd\left[ \dd[\varepsilon(f)] \right] \circ \four{x}{z}{\varepsilon(y)}{\varepsilon(0)} \tag{Lemma~\ref{lem:ep-pair}} \\
&=~ \dd\left[ \dd[\varepsilon(f)] \right] \circ \four{x}{z}{\varepsilon(y)}{0} \tag*{\ref{Eax2}} \\
&=~ \dd\left[ \dd[\varepsilon(f)] \right] \circ \four{x}{\varepsilon(y)}{z}{0} \tag*{\ref{CdCax7a}} \\
&=~ \varepsilon\left(\dd\left[ \dd[f] \right] \right) \circ \four{x}{\varepsilon(y)}{z}{0} \tag*{\ref{CdCax1}} \\
&=~ \varepsilon\left(\dd\left[ \dd[f] \right] \right) \circ \pair{\pair{x}{0} + \pair{0}{\varepsilon(y)}}{\pair{z}{0}} \\
&=~ \varepsilon\left(\dd\left[ \dd[f] \right] \right) \circ \pair{\pair{x}{0} + \pair{\varepsilon(0)}{\varepsilon(y)}}{\pair{z}{0}} \tag*{\ref{Eax1}} \\
&=~ \varepsilon\left(\dd\left[ \dd[f] \right] \right) \circ \pair{\pair{x}{0} + \varepsilon(\pair{0}{y})}{\pair{z}{0}} \tag{Lemma~\ref{lem:ep-pair}} \\
&=~ \varepsilon\left(\dd\left[ \dd[f] \right] \right) \circ \pair{\pair{x}{0} + \varepsilon^2(\pair{0}{y})}{\pair{z}{0}} \tag{Lemma~\ref{lem:d-epsilon-ii}} \\
&=~ \varepsilon\left(\dd\left[ \dd[f] \right] \right) \circ \pair{\pair{x}{0} + \pair{\varepsilon^2(0)}{\varepsilon^2(y)}}{\pair{z}{0}} \tag{Lemma~\ref{lem:ep-pair}} \\
&=~ \varepsilon\left(\dd\left[ \dd[f] \right] \right) \circ \pair{\pair{x}{0} + \pair{0}{\varepsilon^2(y)}}{\pair{z}{0}}\tag*{\ref{Eax1}} \\
&=~ \varepsilon\left(\dd\left[ \dd[f] \right] \right) \circ \four{x}{\varepsilon^2(y)}{z}{0} \\
&=~ \dd\left[ \dd[\varepsilon(f)] \right] \circ \four{x}{\varepsilon^2(y)}{z}{0} \tag*{\ref{CdCax1}} \\
&=~ \dd\left[ \dd[\varepsilon(f)] \right] \circ \four{x}{z}{\varepsilon^2(y)}{0} \tag*{\ref{CdCax7a}} \\
&=~ \dd\left[ \dd[\varepsilon(f)] \right] \circ \four{x}{z}{\varepsilon^2(y)}{\varepsilon^2(0)} \tag*{\ref{Eax1}} \\
&=~ \dd\left[ \dd[\varepsilon(f)] \right] \circ \pair{\pair{x}{z}}{\varepsilon^2(\pair{y}{0})} \tag{Lemma~\ref{lem:ep-pair}} \\
&=~ \varepsilon^2\left(\dd\left[ \dd[\varepsilon(f)] \right] \right) \circ \four{x}{z}{y}{0} \tag{Lemma~\ref{lem:d-epsilon-i}} \\
&=~ \varepsilon^2\left(\dd\left[ \dd[\varepsilon(f)] \right] \circ \four{x}{z}{y}{0} \right) \tag*{\ref{Eax2}} \\
&=~ \varepsilon^2\left(\dd\left[ \dd[\varepsilon(f)] \right] \circ \four{x}{y}{z}{0} \right) \tag*{\ref{CdCax7a}} \\
&=~ \varepsilon^2\left(\dd\left[ \dd[\varepsilon(f)] \right] \right) \circ \four{x}{y}{z}{0} \tag*{\ref{Eax2}} \\
&=~ \varepsilon^3\left(\dd\left[ \dd[f] \right] \right) \circ \four{x}{y}{z}{0} \tag*{\ref{CdCax1} \qedhere}
\end{align*}
\end{proof}
A desirable identity we would like to hold is a slightly stronger version of Lemma~\ref{lem:d-epsilon-iii}:
\begin{equation}\label{lem:d-epsilon-iii'}\begin{gathered} \varepsilon(\dd\left[ \dd[f] \right]) \circ \four{x}{y}{z}{0}
= \varepsilon^2(\dd\left[ \dd[f] \right]) \circ \four{x}{y}{z}{0}
\end{gathered}\end{equation}
Note that if $\varepsilon$ is injective (i.e $\varepsilon(f) = \varepsilon(g)$ implies that $f=g$) or if $\varepsilon$ is idempotent (i.e. $\varepsilon^2(f)=\varepsilon(f)$), then (\ref{lem:d-epsilon-iii'}) holds. In fact, in all of our examples of Cartesian difference categories in Section~\ref{EXsec}, (\ref{lem:d-epsilon-iii'}) holds (even for those where $\varepsilon$ is neither injective or idempotent). It is currently unclear if (\ref{lem:d-epsilon-iii'}) holds in any Cartesian difference category, as we do not have a counterexample. If (\ref{lem:d-epsilon-iii'}) holds and the infinitesimal extension is nilpotent, then it turns out that~\ref{CdCax2} is the same as~\ref{CDCax2}.
\begin{lem}\label{epsilon-nilponent}
Let $\mathbb{X}$ be a Cartesian difference category with a nilpotent
infinitesimal extension, that is, for every map $f: A \to B$ there is some $k \in \mathbb{N}$ such that
$\varepsilon^k(f) = 0$, and suppose that (\ref{lem:d-epsilon-iii'}) holds.
Then~\ref{CDCax2} holds, or in other words, every derivative $\dd[f]$ is additive in its second argument, that is, $\dd[f] \circ \langle x, y + z \rangle = \dd[f] \circ \langle x, y \rangle + \dd[f] \circ \langle x, z \rangle$ and $\dd[f] \circ \langle x, 0 \rangle = 0$.
\end{lem}
\begin{proof} By~\ref{CdCax2}, it already holds that $\dd[f] \circ \langle x, 0 \rangle = 0$. Therefore it remains to show that $\dd[f] \circ \langle x, y + z \rangle = \dd[f] \circ \langle x, y \rangle + \dd[f] \circ \langle x, z \rangle$. So suppose that $\varepsilon^k(f) = 0$ for some $k \in \mathbb{N}$. Then using Lemma~\ref{lem:d-epsilon-iii'}, we compute that:
\begin{align*}
&\dd[f] \circ \langle x, y + z \rangle =~\dd[f] \circ \langle x, y \rangle + \dd[f] \circ \langle x + \varepsilon(y), z \rangle \tag*{\ref{CdCax2}} \\
&=~ \dd[f] \circ \langle x, y \rangle + \dd[f] \circ \left( \langle x, z \rangle + \langle \varepsilon(y), 0 \rangle \right) \\
&=~ \dd[f] \circ \langle x, y \rangle + \dd[f] \circ \left( \langle x, z \rangle + \langle \varepsilon(y), \varepsilon(0) \rangle \right) \tag*{\ref{Eax1}} \\
&=~ \dd[f] \circ \langle x, y \rangle + \dd[f] \circ \left( \langle x, z \rangle + \varepsilon(\langle y, 0 \rangle) \right) \tag{Lemma~\ref{lem:ep-pair}} \\
&=~ \dd[f] \circ \langle x, y \rangle + \dd[f] \circ \langle x , z \rangle + \varepsilon \left( \dd\left[ \dd[f] \right] \circ \four{x}{z}{y}{0} \right) \tag*{\ref{CdCax0}} \\
&=~ \dd[f] \circ \langle x, y \rangle + \dd[f] \circ \langle x , z \rangle + \varepsilon \left( \dd\left[ \dd[f] \right] \right) \circ \four{x}{z}{y}{0} \tag*{\ref{Eax2}} \\
&=~ \dd[f] \circ \langle x, y \rangle + \dd[f] \circ \langle x , z \rangle + \varepsilon^k \left( \dd\left[ \dd[f] \right] \right) \circ \four{x}{z}{y}{0} \tag{Iterating (\ref{lem:d-epsilon-iii'})} \\
&=~ \dd[f] \circ \langle x, y \rangle + \dd[f] \circ \langle x , z \rangle + \dd\left[ \dd[\varepsilon^k(f)] \right] \circ \four{x}{z}{y}{0} \tag*{\ref{CdCax1}} \\
&=~ \dd[f] \circ \langle x, y \rangle + \dd[f] \circ \langle x , z \rangle + \dd\left[ \dd[0] \right] \circ \four{x}{z}{y}{0} \tag{$\varepsilon$ is nilpotent} \\
&=~ \dd[f] \circ \langle x, y \rangle + \dd[f] \circ \langle x , z \rangle + 0 \circ \four{x}{z}{y}{0} \tag*{\ref{CdCax1}} \\
&=~ \dd[f] \circ \langle x, y \rangle + \dd[f] \circ \langle x , z \rangle + 0 \\
&=~ \dd[f] \circ \langle x, y \rangle + \dd[f] \circ \langle x , z \rangle
\end{align*}
We conclude that $\dd[f]$ is additive in its second argument, and so~\ref{CDCax2} holds.
\end{proof}
\subsection{Another look at Cartesian Differential Categories}\label{CDCisCdCsec}
Here we explain how every Cartesian differential category is a Cartesian difference category where the infinitesimal extension is given by zero.
\begin{prop}\label{CDtoCd} Every Cartesian differential category $\mathbb{X}$ with differential combinator $\mathsf{D}$ is a Cartesian difference category where the infinitesimal extension is defined as $\varepsilon(f) = 0$ and the difference combinator is defined to be the differential combinator, $\dd = \mathsf{D}$.
\end{prop}
\begin{proof} As noted before, the first two parts of the~\ref{CdCax1}, the second part of~\ref{CdCax2},~\ref{CdCax3}, \ref{CdCax4},~\ref{CdCax5}, and~\ref{CdCax7} are precisely the same as their Cartesian differential axiom counterparts. On the other hand, since $\varepsilon(f) =0$,~\ref{CdCax0} and the third part of~\ref{CdCax1} trivially state that $0=0$, while the first part of~\ref{CdCax2} and~\ref{CdCax6} end up being precisely the first part of~\ref{CDCax2} and~\ref{CDCax6}. Therefore, the differential combinator satisfies the Cartesian difference axioms and we conclude that a Cartesian differential category is a Cartesian difference category.
\hfill
\end{proof}
Conversely, one can always build a Cartesian differential category from a Cartesian difference category by considering the objects for which the infinitesimal extension of the identity map is the zero map. We call such objects $\varepsilon$-vanishing.
\begin{defi}\label{def:ep-vanishing} In a Cartesian left additive category $\mathbb{X}$ with infinitesimal extension $\varepsilon$, an object $A$ is said to be \textbf{$\varepsilon$-vanishing} if $\varepsilon_A = \varepsilon(1_A) = 0$. We denote $\mathbb{X}_{\varepsilon\text{-van}}$ to be the full subcategory of $\varepsilon$-vanishing objects of $\mathbb{X}$.
\end{defi}
The full subcategory of $\varepsilon$-vanishing objects always forms a Cartesian left additive category.
\begin{lem}\label{lem:ep-van-prod} In a Cartesian left additive category $\mathbb{X}$ with infinitesimal extension $\varepsilon$,
\begin{enumerate}[(\roman{enumi}).,ref={\thelem.\roman{enumi}}]
\item The terminal object $\top$ is $\varepsilon$-vanishing;
\item If $A$ and $B$ are $\varepsilon$-vanishing then their product $A \times B$ is $\varepsilon$-vanishing;
\end{enumerate}
Therefore, $\mathbb{X}_{\varepsilon\text{-van}}$ is a Cartesian left additive category with the same Cartesian left additive structure as $\mathbb{X}$. Furthermore:
\begin{enumerate}[(\roman{enumi}).,ref={\thelem.\roman{enumi}}]
\setcounter{enumi}{2}
\item If $B$ is $\varepsilon$-vanishing, then for any map $f: A \to B$, $\varepsilon(f) = 0$.
\end{enumerate}
\end{lem}
\begin{proof} For (i), recall that for the terminal object, $1_\top = 0$. Therefore by~\ref{Eax1} we easily see that $\varepsilon(1_\top) = \varepsilon(0) = 0$. So the terminal object $\top$ is $\varepsilon$-vanishing. For (ii), first recall that by Lemma~\ref{lem:epbij}, for any pair of objects $A$ and $B$, $\varepsilon_{A \times B} = \varepsilon_A \times \varepsilon_B$. Therefore, if both $A$ and $B$ are $\varepsilon$-vanishing, we have that $\varepsilon_{A \times B} = \varepsilon_A \times \varepsilon_B = 0 \times 0 = 0$. So there product of $\varepsilon$-vanishing objects is again $\varepsilon$-vanishing. Thus $\mathbb{X}_{\varepsilon\text{-van}}$ is closed under finite products and therefore we conclude that $\mathbb{X}_{\varepsilon\text{-van}}$ is also a Cartesian left additive category. For (iii), suppose that $B$ is $\varepsilon$-vanishing. Then by~\ref{Eax2}, we have that $\varepsilon(f) = \varepsilon(1_B) \circ f = 0 \circ f = 0$. \end{proof}
For a Cartesian difference category, its subcategory of $\varepsilon$-vanishing objects is a Cartesian differential category.
\begin{prop}\label{CdtoCD} For a Cartesian difference category $\mathbb{X}$ with infinitesimal extension $\varepsilon$ and difference combinator $\dd$, then $\mathbb{X}_{\varepsilon\text{-van}}$ is a Cartesian differential category where the differential combinator is defined to be the difference combinator, $\mathsf{D} = \dd$.
\end{prop}
\begin{proof} In Lemma~\ref{lem:ep-van-prod} we already explained why $\mathbb{X}_{\varepsilon\text{-van}}$ is a Cartesian left additive category. So it remains to explain why $\dd$ is a differential combinator. However, by Lemma~\ref{lem:ep-van-prod}.iii, every map in $\mathbb{X}_{\varepsilon\text{-van}}$ satisfies $\varepsilon(f) =0$. Therefore, for maps between objects in $\mathbb{X}_{\varepsilon\text{-van}}$, the Cartesian difference axioms are precisely the Cartesian differential axioms. So we conclude that the difference combinator is a differential combinator for this subcategory, and thus $\mathbb{X}_{\varepsilon\text{-van}}$ is a Cartesian differential category. \hfill
\end{proof}
Note that by Lemma~\ref{lem:ep-van-prod}.i, for any Cartesian difference category $\mathbb{X}$, the terminal object $\top$ is always $\varepsilon$-vanishing, and so therefore, $\mathbb{X}_{\varepsilon\text{-van}}$ is never empty. On the other hand, applying Proposition~\ref{CdtoCD} to a Cartesian differential category results in the entire category, since every object is $\varepsilon$-vanishing by definition. It is also important to note that the above two propositions do not imply that if a difference combinator is a differential combinator then the infinitesimal extension must be zero. In Section~\ref{moduleex}, we provide such an example of a Cartesian differential category that comes equipped with a non-zero infinitesimal extension such that the differential combinator is a difference combinator with respect to this non-zero infinitesimal extension.
\subsection{Cartesian Difference Categories as Change Action Models}\label{CdCisCAsec}
In this section, we show how every Cartesian difference category is a particularly well-behaved change action model, and conversely how every change action model contains a Cartesian difference category.
\begin{prop}\label{CdtoCA} Let $\mathbb{X}$ be a Cartesian difference category with infinitesimal extension $\varepsilon$ and difference combinator $\dd$. Define the functor $\alpha: \mathbb{X} \to \mathsf{CAct}(\mathbb{X})$ on objects as $\alpha(A) = (A, A, \oplus_A, +_A, 0_A)$ (as defined in Lemma~\ref{caelem}) and on maps as $\alpha(f) = (f, \dd[f])$. Then $(\mathbb{X}, \alpha : \mathbb{X} \to \mathsf{CAct}(\mathbb{X}))$ is a change action model.
\end{prop}
\begin{proof} By Lemma~\ref{caelem}, $(A, A, \oplus_A, +_A, 0_A)$ is a change action and so $\alpha$ is well-defined on objects. While for a map $f$, $\dd[f]$ is a derivative of $f$ in the change action sense since~\ref{CdCax0} and~\ref{CdCax2} are precisely~\ref{CADax1} and~\ref{CADax2}, and so $\alpha$ is well-defined on maps.
Next we show that $\alpha$ preserves identities and composition, which follows from~\ref{CdCax3} and~\ref{CdCax5} respectively:
\begin{align*}
\alpha(1_A) &=~ (1_A, \dd[1_A]) \\
&=~ (1_A, \pi_1) \tag*{\ref{CdCax3}} \\ \\
\alpha(g \circ f) &=~(g \circ f, \dd[g \circ f]) \\
&=~ \left( g \circ f, \dd[g] \circ \langle f \circ \pi_0, \dd[f] \rangle \right) \tag*{\ref{CdCax5}} \\
&=~(g, \dd[g]) \circ (f, \dd[g]) \tag{by def. of comp. in $\mathsf{CAct}(\mathbb{X})$} \\
&=~\alpha(g) \circ \alpha(f)
\end{align*}
So $\alpha$ is a functor. Next we check that $\alpha$ preserves projections and pairings, which will follow from~\ref{CdCax3} and~\ref{CdCax4}:
\begin{align*}
\alpha(\pi_i) &=~ (\pi_i, \dd[\pi_i]) \\
&=~ (\pi_i, \pi_i \circ \pi_1) \tag*{\ref{CdCax3}} \\ \\
\alpha(\pair{f}{g}) &=~ (\pair{f}{g}, \dd[\pair{f}{g}]) \\
&=~ (\pair{f}{g}, \pair{\dd[f]}{\dd[g]} ) \tag*{\ref{CdCax4}} \\
&=~ \pair{(f,\dd[f])}{(g, \dd[g])} \\
&=~ \pair{\alpha(f)}{\alpha(g)}
\end{align*}
So $\alpha$ preserves finite products. Lastly, it is clear that $\alpha$ is a section of the forgetful functor, and therefore we conclude that $(\mathbb{X}, \alpha)$ is a change action model.
\hfill
\end{proof}
Not every change action model is a Cartesian difference category. For example, change action models do not require the addition to be commutative. On the other hand, it can be shown that every change action model contains a Cartesian difference category as a full subcategory. We call the objects and maps in this subcategory \emph{flat}. It is important to note that the definition here is a slight generalisation of the one in~\cite{alvarez2020cartesian}. The original suggested definition implied that $\oplus = +$ and as a result $\varepsilon(f) = f$. While this indeed resulted in a Cartesian difference category (specifically one that generalized the Abelian group example from Section~\ref{discreteex}), we found that it was a bit too restrictive of a construction. We correct this by generalizing the definition slightly, while still keeping the construction's overall core ideas the same.
\begin{defi}\label{def:flat}
Let $(\mathbb{X}, \alpha : \mathbb{X} \to \mathsf{CAct}(\mathbb{X}))$ be a change action model. An object $A$ is \textbf{flat} whenever the following equations hold:
\begin{enumerate}[{\bf [F.1]},ref={{\bf [F.{\arabic*}]}},align=left]
\item $\Delta A = A$\label{F1}
\item $\alpha(\oplus_A) = (\oplus_A, \oplus_A \circ \pi_1)$ and $\alpha(+_A) = (+_A, +_A \circ \pi_1)$\label{F2}
\item $\oplus_A$ is right-injective, that is, if $\oplus_A \circ \langle f, g \rangle = \oplus_A \circ \langle f, h \rangle$ then $g=h$.\label{F4}
\end{enumerate}
Let $f : A \to B$ be a map between flat objects in the above change action model, with $\alpha(f) = (f, \dd[f])$. Then $f$ is \textbf{flat}
whenever the following condition holds:
\begin{enumerate}[{\bf [F.1]},ref={{\bf [F.{\arabic*}]}},align=left]
\setcounter{enumi}{3}
\item $0_B \oplus_B \partial[f] \circ \pair{x}{y} = \partial[f] \circ \pair{x}{0_A \oplus_A y}$\label{Fd}
\end{enumerate}
We define $\mathsf{Flat}_\alpha$ to be the subcategory of $\mathbb{X}$ containing all flat objects and all flat morphisms
between them.
\end{defi}
Intuitively, the condition~\ref{F1} simply states that a flat object is similar to an Euclidean space
in the sense that its ``tangent space'' is equal to itself.~\ref{F2} can be read as stating that
the action $\oplus_A$ and the addition of changes $+_A$ are linear. In this sense, flat objects behave like Euclidean spaces.
\ref{F4} tells us that we may cancel the left argument of $\oplus_A$, which will imply that the result infinitesimal extension is injective. This cancellation ability that~\ref{F4} provides will be used throughout most of the proofs in this section. Lastly, as we will see below,~\ref{Fd} is a generalization of Lemma~\ref{lem:d-epsilon-i}.
We would like to show that for any change action model $(\mathbb{X}, \alpha)$, $\mathsf{Flat}_\alpha$ is a Cartesian difference category. We first explain the finite product structure of $\mathsf{Flat}_\alpha$.
\begin{lem} Let $(\mathbb{X}, \alpha : \mathbb{X} \to \mathsf{CAct}(\mathbb{X}))$ be a change action model. Then:
\begin{enumerate}[(\roman{enumi}).,ref={\thelem.\roman{enumi}}]
\item The terminal object $\top$ is flat;
\item If $A$ and $B$ are flat then their product $A \times B$ is flat;
\item If $A$ is a flat object, then the identity $1_A: A \to A$ is flat;
\item If $A$, $B$, and $C$ are flat objects, and $f: A \to B$ and $g: B \to C$ are flat maps, then their composite $g \circ f: A \to C$ is flat;
\item If $A$ and $B$ are flat objects, then the projection maps $\pi_0: A \times B \to A$ and ${\pi_1: A \times B \to B}$ are flat;
\item If $A$, $B$, and $C$ are flat objects, and $f: C \to A$ and $g: C \to B$ are flat maps, then their pairing $\langle f, g \rangle: C \to A \times B$ is flat.
\end{enumerate}
Furthermore, $\mathsf{Flat}_\alpha$ is a Cartesian category with the same Cartesian structure as $\mathbb{X}$.
\end{lem}
\begin{proof} These are straightforward and we leave it as an exercise for the reader to check for themselves. (i), (ii), (v), and (vi) follows from the fact that $\alpha$ preserves finite products, while (iii) and (v) follows from the functoriality of $\alpha$.
\end{proof}
As an immediate consequence, we note that for any change action model $(\mathbb{X}, \alpha)$, since the terminal object is always flat, $\mathsf{Flat}_\alpha$ is never empty. Next we discuss the additive structure of $\mathsf{Flat}_\alpha$. The sum of maps $f: A \to B$ and $g: A \to B$ in $\mathsf{Flat}_\alpha$ is defined using the change action structure $f+g := f +_B g$, while the zero map $0: A \to B$ is $0 := 0_B \circ{} !_A$. And so we obtain that:
\begin{lem}\label{EUCLCLAC} Let $(\mathbb{X}, \alpha : \mathbb{X} \to \mathsf{CAct}(\mathbb{X}))$ be a change action model. Then:
\begin{enumerate}[(\roman{enumi}).,ref={\thelem.\roman{enumi}}]
\item If $A$ and $B$ are flat objects, then the zero map $0: A \to B$ is flat;
\item If $A$ and $B$ are flat objects, and $f: A \to B$ and $g: A \to B$ are flat morphisms, then their sum $f + g: A \to B$ is flat.
\end{enumerate}
Furthermore, $\mathsf{Flat}_\alpha$ is a Cartesian left additive category.
\end{lem}
\begin{proof} Most of the Cartesian left additive structure is straightforward. However, since the addition is not required to be commutative for arbitrary change actions, we will show that the addition is commutative for flat objects. Using that $\oplus_B$ is an action as in Lemma~\ref{lem:CAadd}, that by~\ref{F2} we have that $\oplus_B \circ \pi_1$ is a derivative for $\oplus_B$, and~\ref{CADax1}, we obtain that:
\begin{align*}
0_B \oplus_B (f + g)
&= (0_B \oplus_B f) \oplus_B g
\tag{Lemma~\ref{lem:CAadd.+}} \\
&= (0_B \oplus_B f) \oplus_B (g \oplus_B 0)
\tag{Lemma~\ref{lem:CAadd.0}} \\
&= \oplus_B \circ \pair{0_B \oplus_B f}{g \oplus_B 0} \\
&= \oplus_B \circ \left( \pair{0_B}{g} \oplus_B \pair{f}{0} \right) \tag{Lemma~\ref{lem:CACTprod.oplus}} \\
&= \left( \oplus_B \circ \pair{0_B}{g} \right) \oplus_B \left(\dd[\oplus_B]\four{0_B}{g}{f}{0} \right) \tag*{\ref{CADax1}} \\
&= (0_B \oplus_B g) \oplus_B \dd[\oplus_B]\four{0_B}{g}{f}{0} \\
&= (0_B \oplus_B g) \oplus_B \left( \oplus_B \circ \pi_1 \circ \four{0_B}{g}{f}{0} \right) \tag*{\ref{F2}} \\
&= (0_B \oplus_B g) \oplus_B \left( \oplus_B \circ \pair{f}{0} \right) \\
&= (0_B \oplus_B g) \oplus_B (f \oplus_B 0)
\\
&= (0_B \oplus_B g) \oplus_B f
\tag{Lemma~\ref{lem:CAadd.0}} \\
&= 0_B \oplus_B (g + f)
\tag{Lemma~\ref{lem:CAadd.+}}
\end{align*}
By~\ref{F4}, $\oplus_B$ is right-injective and we conclude that $f + g = g + f$. \hfill
\end{proof}
We use the action of the change action structure to define the infinitesimal extension. So for a map $f: A \to B$ in $\mathsf{Flat}_\alpha$, define $\varepsilon(f): A \to B$ as follows:
\[ \varepsilon(f) = \oplus_B \circ \pair{0_B \circ{} !_A}{f} = 0 \oplus_B f \]
As such, we may rewrite~\ref{Fd} as follows:
\[ \dd[f] \circ \pair{x}{\varepsilon(y)} = \varepsilon(\dd[f]) \circ
\pair{x}{y} \]
which is Lemma~\ref{lem:d-epsilon-i}.
\begin{lem} $\varepsilon$ is an infinitesimal extension for $\mathsf{Flat}_\alpha$.
\end{lem}
\begin{proof} We show that $\varepsilon$ satisfies~\ref{Eax1},~\ref{Eax2},~\ref{Eax3}. Starting with~\ref{Eax1}, we compute that:
\begin{align*}
\varepsilon(0) &=~ 0 \oplus_B 0 \\
&=~ 0 \tag{Lemma~\ref{lem:CAadd.0}}
\end{align*}
So $\varepsilon(0)= 0$. Next following the same idea as in the proof of Lemma~\ref{EUCLCLAC}, we obtain the following:
\begin{align*}
0_B \oplus_B \varepsilon(f + g) &= 0_B \oplus_B (0 \oplus_B (f + g))\\
&= 0_B \oplus_B ((0 \oplus_B f) \oplus_B g) \tag{Lemma~\ref{lem:CAadd.+}} \\
&= 0_B \oplus_B (\varepsilon(f) \oplus_B g) \tag{Lemma~\ref{lem:CAadd.+}} \\
&= (0_B \oplus_B 0) \oplus_B (\varepsilon(f) \oplus_B g) \tag{Lemma~\ref{lem:CAadd.0}} \\
&= \oplus_B \circ \pair{0_B \oplus_B 0}{\varepsilon(f) \oplus_B g} \\
&= \oplus_B \circ \left( \pair{0_B}{\varepsilon(f)} \oplus_B \pair{0}{g} \right) \tag{Lemma~\ref{lem:CACTprod.oplus}} \\
&= \left( \oplus_B \circ \pair{0_B}{\varepsilon(f)} \right) \oplus_B \left( \partial[\oplus_B] \circ \four{0_B}{\varepsilon(f)}{0}{g} \right) \tag*{\ref{CADax1}} \\
&= \left( 0_B \oplus \varepsilon(f) \right) \oplus_B \left( \partial[\oplus_B] \circ \four{0_B}{\varepsilon(f)}{0}{g} \right) \\
&= \left( 0_B \oplus \varepsilon(f) \right) \oplus_B \left( \oplus_B \circ \pi_1 \circ \four{0_B}{\varepsilon(f)}{0}{g} \right) \tag*{\ref{F2}}\\
&= \left( 0_B \oplus \varepsilon(f) \right) \oplus_B \left( \oplus_B \circ \pair{0}{g} \right) \tag*{\ref{F2}}\\
&= \left( 0_B \oplus \varepsilon(f) \right) \oplus_B (0 \oplus_B g) \\
&= (0_B \oplus_B \varepsilon(f)) \oplus_B \varepsilon(g)\\
&= 0_B \oplus_B (\varepsilon(f) +_B \varepsilon(g)) \tag{Lemma~\ref{lem:CAadd.+}}
\end{align*}
Then by~\ref{F4}, it follows that $\varepsilon(f + g) = \varepsilon(f) + \varepsilon(g)$. So~\ref{Eax1} holds. Next, it is easy to show that $\varepsilon$ is compatible with composition:
\begin{align*}
\varepsilon(g \circ f) &= 0 \oplus_C (g \circ f) \\
&= (0 \circ f) \oplus_C (g \circ f) \\
&= (0 \oplus_C g) \circ f \tag{Lemma~\ref{lem:CAadd.circ}} \\
&= \varepsilon(g) \circ f
\end{align*}
So~\ref{Eax2} holds. Finally, for the projections we compute:
\begin{align*}
\varepsilon(\pi_i) &= 0 \oplus \pi_i \\
&= \oplus_A \circ \pair{0_A \circ !_{A \times B}}{\pi_i}\\
&= \oplus_A \circ \pair{\pi_i \circ 0_{A \times B} \circ !_{A \times B}}{\pi_i} \tag{Lemma~\ref{lem:CACTprod.0}} \\
&= \oplus_A \circ (\pi_i \times \pi_i) \circ \pair{0_{A \times B} \circ !_{A \times B}}{1_{A \times B}}\\
&= \pi_i \circ \oplus_{A \times B} \circ \pair{0_{A \times B} \circ !_{A \times B}}{1_{A \times B}}\\
&= \pi_i \circ \left( 0 \oplus 1_{A \times B} \right) \\
&= \pi_i \circ \varepsilon(1_{A \times B})
\end{align*}
So~\ref{Eax3} holds. Thus we conclude that $\varepsilon$ is an infinitesimal extension. \hfill
\end{proof}
Next, we observe that for a flat object, the action can be expressed in terms of the infinitesimal extension.
\begin{lem}%
\label{oplus-epsilon} For any maps $f, g : A \to B$ in $\mathsf{Flat}_\alpha$, the following equality holds:
\begin{align*}
f \oplus_B g = f + \varepsilon(g)
\end{align*}
\end{lem}
\begin{proof}
The proof is a straightforward consequence of linearity of addition:
\begin{align*}
f + \varepsilon(g) &= f + (0 \oplus_B g) \\
&= (f \oplus_B 0) + (0 \oplus_B g)
\tag{Lemma~\ref{lem:CAadd.0}}\\
&= +_B \circ \left( \pair{f \oplus_B 0}{0 \oplus_B g} \right) \\
&= +_B \circ \left( \pair{f}{0} \oplus_{B \times B} \pair{0}{g} \right) \tag{Lemma~\ref{lem:CACTprod.oplus}} \\
&= \left(+_B \circ \pair{f}{0} \right) \oplus_B \left( \dd[+_B] \circ \four{f}{0}{0}{g} \right)
\tag*{\ref{CADax1}} \\
&= (f + 0) \oplus_B \left( \dd[+_B] \circ \four{f}{0}{0}{g} \right)
\\
&= f \oplus_B \left( \dd[+_B] \circ \four{f}{0}{0}{g} \right) \\
&= f \oplus_B \left( +_B \circ \pi_1 \circ \four{f}{0}{0}{g} \right)
\tag*{\ref{F2}}
\\
&= f \oplus_B \left( +_B \circ \pair{0}{g} \right)
\\
&= f \oplus_B (0 + g) \\
&= f \oplus_B g
\end{align*}
Thus we conclude that the desired equality holds. \end{proof}
The difference combinator for $\mathsf{Flat}_\alpha$ is defined in the obvious way, that is, $\dd[f]$ is defined as the second component of $\alpha(f)$.
\begin{prop}\label{CAtoCd} Let $(\mathbb{X}, \alpha : \mathbb{X} \to \mathsf{CAct}(\mathbb{X}))$ be a change action model. Then $\mathsf{Flat}_\alpha$ is a Cartesian difference category.
\end{prop}
\begin{proof}
\ref{CdCax0} and~\ref{CdCax2} are simply a restatement of~\ref{CADax1} and~\ref{CADax2}.~\ref{CdCax3} and~\ref{CdCax4} follow immediately from the fact that
$\alpha$ preserves finite products and from the structure of products in
$\mathsf{CAct}(\mathbb{X})$, while~\ref{CdCax5}
follows from the definition of composition in $\mathsf{CAct}(\mathbb{X})$. So it remains to show that~\ref{CdCax1},~\ref{CdCax6}, and~\ref{CdCax7} hold. We start by proving~\ref{CdCax1}, for which it suffices to calculate and apply~\ref{F2}:
\begin{align*}
\dd[f + g] &=~ \dd[+ \circ \pair{f}{g}] \\
&=~ \dd[+_{B}] \circ \pair{\pair{f}{g} \circ \pi_0}{ \pair{\dd[f]}{\dd[g]}} \tag*{\ref{CdCax5}} \\
&=~ + \circ \pi_1 \circ \pair{\pair{f}{g} \circ \pi_0}{ \pair{\dd[f]}{\dd[g]}} \tag*{\ref{CdCax2}} \\
&=~ + \circ \pair{\dd[f]}{\dd[g]} \\
&=~ \dd[f] + \dd[g]
\end{align*}
It is not hard to show that $\dd[0] = 0$ and $\dd[\varepsilon(f)] = \varepsilon(\dd[f])$. For the first property, on one hand we have:
\begin{align*}
0 \circ \oplus \circ \pair{x}{y} &= 0
= \oplus \circ \pair{0 \circ x}{0 \circ \pair{x}{y}}
\end{align*}
On the other hand, applying~\ref{CADax1}:
\begin{align*}
0 \circ \oplus \circ \pair{x}{y}
= \oplus \circ \pair{0 \circ x}{\dd[0] \circ \pair{x}{y}}
\end{align*}
Hence by~\ref{F4} we have $\dd[0] \circ \pair{x}{y} = 0 \circ \pair{x}{y}$ for any choice of $x$ and $y$.
In particular, $\dd[0] = \dd[0] \circ \pair{\pi_0}{\pi_1} = 0 \circ \pair{\pi_0}{\pi_1} = 0$ as desired.
For the infinitesimal extension, we simply apply the chain rule and the equation $\dd[0] = 0$:
\begin{align*}
\dd[\varepsilon(f)]
&= \dd[\oplus_B \circ \pair{0}{f}]\\
&= \dd[\oplus_B] \circ \pair{\pair{0}{f} \circ \pi_0}{\dd[\pair{0}{f}]} \tag*{\ref{CdCax5}}\\
&= \oplus_B \circ \dd[\pair{0}{f}]
\tag*{\ref{F2}}
\\
&= \oplus_B \circ \pair{\dd[0]}{\dd[f]} \tag*{\ref{CdCax4}} \\
&= \oplus_B \circ \pair{0}{\dd[f]} \tag{$\dd[0] = 0$}
\\
&= \varepsilon(\dd[f])
\end{align*}
So we conclude that~\ref{CdCax1} holds.
We proceed to prove axioms~\ref{CdCax6a} and~\ref{CdCax7a} --- which, per Lemma~\ref{lem:cdc6a},
is equivalent to, and easier to prove than~\ref{CdCax6} and~\ref{CdCax7}. Starting with~\ref{CdCax6a}, as before, we pick arbitrary
$x, y : A \to B$ and calculate:
\begin{align*}
0 \oplus \dd^2[f] \circ \four{x}{0}{0}{y}
&= \left( \dd[f] \circ \pair{x}{0} \right) \oplus \left( \dd^2[f] \circ \four{x}{0}{0}{y} \right)
\tag*{\ref{CdCax2}}
\\
&= \dd[f] \circ \left( \pair{x}{0} \oplus \pair{0}{y} \right) \tag*{\ref{CADax1}} \\
&= \dd[f] \circ \left( \pair{x \oplus 0}{0 \oplus y} \right) \tag{Lemma~\ref{lem:CACTprod.oplus}} \\
&= \dd[f] \circ \pair{x}{0 \oplus y} \tag{Lemma~\ref{lem:CAadd.0}}\\
&= 0 \oplus \dd[f] \circ \pair{x}{y}
\tag*{\ref{Fd}}
\end{align*}
Hence by~\ref{F4}, $\dd^2[f]\circ \four{x}{0}{0}{y} = \dd[f] \circ \pair{x}{y}$ as desired.
Finally, for~\ref{CdCax7a}, first observe that by Lemma~\ref{oplus-epsilon} we can compute the following equality for any suitable maps:
\begin{align*}
(f \oplus g) \oplus (h \oplus k) &=~ (f +\varepsilon(g) ) \oplus (h + \varepsilon(k) ) \tag{Lemma~\ref{oplus-epsilon}} \\
&= f +\varepsilon(g) + \varepsilon\left(h + \varepsilon(k) \right) \tag{Lemma~\ref{oplus-epsilon}} \\
&= f +\varepsilon(g) + \varepsilon(h) + \varepsilon^2(k) \tag*{\ref{Eax1}} \\
&=f +\varepsilon(h) + \varepsilon(g) + \varepsilon^2(k) \tag{by commutativity of $+$} \\
&= f +\varepsilon(h) + \varepsilon\left(g + \varepsilon(k) \right) \tag*{\ref{Eax1}} \\
&= (f +\varepsilon(h) ) \oplus (g + \varepsilon(k) ) \tag{Lemma~\ref{oplus-epsilon}} \\
&= (f \oplus h) \oplus (g \oplus k)
\end{align*}
Thus we have the following identity:
\begin{equation}\label{swapoplus}\begin{gathered} (f \oplus g) \oplus (h \oplus k) = (f \oplus h) \oplus (g \oplus k)
\end{gathered}\end{equation}
Alternatively, we could have computed the following for any suitable maps:
\begin{align*}
(f \oplus g) \oplus (h \oplus k) &=~ (f +\varepsilon(g) ) \oplus (h + \varepsilon(k) ) \tag{Lemma~\ref{oplus-epsilon}} \\
&= f +\varepsilon(g) + \varepsilon\left(h + \varepsilon(k) \right) \tag{Lemma~\ref{oplus-epsilon}} \\
&= f +\varepsilon(g) + \varepsilon(h) + \varepsilon^2(k) \tag*{\ref{Eax1}} \\
&= f + \varepsilon\left( g +h + \varepsilon(k) \right) \tag*{\ref{Eax1}} \\
&= f \oplus \left( g +h + \varepsilon(k) \right) \tag{Lemma~\ref{oplus-epsilon}} \\
&= f \oplus \left( (g+h) \oplus k \right) \tag{Lemma~\ref{oplus-epsilon}}
\end{align*}
Thus we also have the following identity:
\begin{equation}\label{swapoplus2}\begin{gathered} (f \oplus g) \oplus (h \oplus k) = f \oplus \left( (g+h) \oplus k \right)
\end{gathered}\end{equation}
Then we compute the following for suitable maps:
\begin{align*}
&f \circ \left((x \oplus y) \oplus (z \oplus w) \right) = \left( f \circ \left( x \oplus y \right) \right) \oplus \left( \dd[f]\circ \langle x \oplus y, z \oplus w \rangle \right)
\tag*{\ref{CADax1}} \\
&= \left( (f \circ x) \oplus \left( \dd[f]\circ \langle x, y \rangle \right) \right) \oplus \left( \dd[f] \circ \langle x \oplus y, z \oplus w \rangle \right)
\tag*{\ref{CADax1}}
\\
&= \left( (f \circ x) \oplus \left( \dd[f]\circ \langle x, y \rangle \right) \right) \oplus \left( \dd[f] \circ \left( \pair{x}{z} \oplus \pair{y}{w} \right) \right)
\\
&= \left( (f \circ x) \oplus \left( \dd[f]\circ \langle x, y \rangle \right) \right) \oplus \left( \left( \dd[f] \circ \langle x, z \rangle \right) \oplus \left( \dd^2[f] \circ \four{x}{z}{y}{w} \right) \right) \tag*{\ref{CADax1}} \\
&= (f \circ x) \oplus \left( \left( \left( \dd[f]\circ \langle x, y \rangle \right) + \left( \dd[f] \circ \langle x, z \rangle \right) \right) \oplus \left( \dd^2[f] \circ \four{x}{z}{y}{w} \right) \right) \tag{\ref{swapoplus2}}
\end{align*}
So we have the following identity:
\begin{equation}\label{swapoplus3}\begin{gathered} f \circ \left((x \oplus y) \oplus (z \oplus w) \right) = \\ (f \circ x) \oplus \left( \left( \left( \dd[f]\circ \langle x, y \rangle \right) + \left( \dd[f] \circ \langle x, z \rangle \right) \right) \oplus \left( \dd^2[f] \circ \four{x}{z}{y}{w} \right) \right)
\end{gathered}\end{equation}
However, by swapping the middle two arguments on the left hand side using (\ref{swapoplus}) we compute:
\begin{align*}
&(f \circ x) \oplus \left( \left( \left( \dd[f]\circ \langle x, y \rangle \right) + \left( \dd[f] \circ \langle x, z \rangle \right) \right) \oplus \left( \dd^2[f] \circ \four{x}{z}{y}{w} \right) \right) \\
&= f \circ \left((x \oplus y) \oplus (z \oplus w) \right) \tag{\ref{swapoplus3}} \\
&= f \circ \left((x \oplus z) \oplus (y \oplus w) \right) \tag{\ref{swapoplus}} \\
&=(f \circ x) \oplus \left( \left( \left( \dd[f]\circ \langle x, z \rangle \right) + \left( \dd[f] \circ \langle x, y \rangle \right) \right) \oplus \left( \dd^2[f] \circ \four{x}{y}{z}{w} \right) \right)\tag{\ref{swapoplus3}} \\
&= (f \circ x) \oplus \left( \left( \left( \dd[f]\circ \langle x, y \rangle \right) + \left( \dd[f] \circ \langle x, z \rangle \right) \right) \oplus \left( \dd^2[f] \circ \four{x}{y}{z}{w} \right) \right) \tag{by com. of $+$}
\end{align*}
So we have that:
\begin{align*}
&(f \circ x) \oplus \left( \left( \left( \dd[f]\circ \langle x, y \rangle \right) + \left( \dd[f] \circ \langle x, z \rangle \right) \right) \oplus \left( \dd^2[f] \circ \four{x}{z}{y}{w} \right) \right) \\
&= (f \circ x) \oplus \left( \left( \left( \dd[f]\circ \langle x, y \rangle \right) + \left( \dd[f] \circ \langle x, z \rangle \right) \right) \oplus \left( \dd^2[f] \circ \four{x}{y}{z}{w} \right) \right)
\end{align*}
By applying~\ref{F4}, we obtain that:
\begin{align*} &\left( \left( \dd[f]\circ \langle x, y \rangle \right) + \left( \dd[f] \circ \langle x, z \rangle \right) \right) \oplus \left( \dd^2[f] \circ \four{x}{z}{y}{w} \right) \\& = \left( \left( \dd[f]\circ \langle x, y \rangle \right) + \left( \dd[f] \circ \langle x, z \rangle \right) \right) \oplus \left( \dd^2[f] \circ \four{x}{y}{z}{w} \right) \end{align*}
By applying~\ref{F4} again we finally obtain $\dd^2[f]\left \langle\langle x, y \rangle, \langle z, w \rangle \right \rangle= \dd^2[f]\left \langle\langle x, z \rangle, \langle y, w \rangle \right \rangle$ as desired. So we conclude that $\mathsf{Flat}_\alpha$ is a Cartesian difference category.
\hfill
\end{proof}
\subsection{Linear Maps and \texorpdfstring{$\varepsilon$}{Epsilon}-Linear Maps}
An important subclass of maps in a Cartesian differential category is the subclass of \emph{linear maps}~\cite[Definition 2.2.1]{blute2009cartesian}. One can also define linear maps in a Cartesian difference category by using the same definition.
\begin{defi}%
\label{def:linearity} In a Cartesian difference category, a map $f$ is \textbf{linear} if the following equality holds: $\dd[f] = f \circ \pi_1$.
\end{defi}
Using element-like notation, a map $f$ is linear if $\dd[f](x,y) = f(y)$. Linear maps in a Cartesian difference category satisfy many of the same properties found in~\cite[Lemma 2.2.2]{blute2009cartesian}.
\begin{lem}\label{lem:linear} In a Cartesian difference category,
\begin{enumerate}[(\roman{enumi}),ref={\thelem.\roman{enumi}}]
\item\label{epsilon-linear} If $f: A \to B$ is linear then $\varepsilon(f) = f \circ
\varepsilon(1_{A})$.
\item\label{add-linear} If $f: A \to B$ is linear, then $f$ is additive.
\item\label{id-linear} Identity maps, projection maps, and zero maps are linear.
\item\label{comp-linear} The composite, sum, pairing, and product of linear maps are linear.
\item\label{chain-linear} If $f: A \to B$ and $k: C \to D$ are linear, then for any map $g: B \to
C$:
\[ \dd[k \circ g \circ f] = k \circ \dd[g] \circ (f \times f) \]
\item If an isomorphism is linear, then its inverse is linear.
\item For any object $A$, $\oplus_A$ and $+_A$ are linear.\label{struct-linear}
\end{enumerate}
\end{lem}
\begin{proof}
Most of the above results are either immediate or admit a similar proof to the
ones in~\cite[Lemma~2.2.2]{blute2009cartesian}. We will prove
i, as it differs from the differential setting:
\begin{align*}
f \circ \varepsilon(1_{A}) &=~ f \circ (0 + \varepsilon(1_{A})) \\
&=~ f \circ 0 + \varepsilon(\dd[f] \circ \pair{0}{1_A}) \tag*{\ref{CdCax0}}\\
&=~ f \circ \pi_1 \circ \pair{0}{0} + \varepsilon(\dd[f] \circ \pair{0}{1_A}) \\
&=~ \dd[f] \circ \pair{0}{0} + \varepsilon(f \circ \pi_1 \circ \pair{0}{1_A}) \tag{$f$ is linear} \\
&=~ 0 + \varepsilon(f \circ 1_A) \tag*{\ref{CdCax2}} \\
&=~ \varepsilon(f)
\end{align*}
So $\varepsilon(f) = f \circ \varepsilon(1_{A})$.
\end{proof}
Using element-like notation, the first point of the above lemma says that if $f$ is linear then $f(\varepsilon(x)) = \varepsilon(f(x))$. It is important to note that while all linear maps are additive, the converse is not necessarily true, see~\cite[Corollary 2.3.4]{blute2009cartesian}. That said, an immediate consequence of the above lemma is that the subcategory of linear maps of a Cartesian difference category has finite biproducts.
Another interesting subclass of maps is the subclass of $\varepsilon$-linear maps, which are maps whose infinitesimal extension is linear.
\begin{defi} In a Cartesian difference category, a map $f$ is \textbf{$\varepsilon$-linear} if $\varepsilon(f)$ is linear.
\end{defi}
\begin{lem} In a Cartesian difference category,
\begin{enumerate}[(\roman{enumi}),ref={\thelem.\roman{enumi}}]
\item If $f: A \to B$ is $\varepsilon$-linear then $f \circ (x + \varepsilon(y)) = f \circ x + \varepsilon(f) \circ y$;
\item Every linear map is $\varepsilon$-linear;
\item The composite, sum, and pairing of $\varepsilon$-linear maps is $\varepsilon$-linear;
\item If an isomorphism is $\varepsilon$-linear, then its inverse is again $\varepsilon$-linear.
\item If $\varepsilon$ is nilpotent (.i.e for every $f$ there exists a $k \in \mathbb{N}$ such that $\varepsilon^k(f)= 0$), then for every map $f$, $\dd[f] \circ \pair{0}{1}$ is linear.\label{nilpotent-linear}
\end{enumerate}
\end{lem}
\begin{proof} The first four (i) through (iv) are straightforward and so we leave them as an exercise for the reader.
For (v), suppose that $\varepsilon$ is nilpotent. We need to show that $\dd[\varepsilon(\dd[f] \circ \pair{0}{1})]
= \varepsilon(\dd[f] \circ \pair{0}{1}) \circ \pi_1$. So we compute:
\begin{align*}
\dd[\varepsilon(\dd[f] \circ \pair{0}{1})] &=~ \varepsilon\left( \dd\left[ \dd[f] \circ \pair{0}{1} \right] \right) \tag*{\ref{CdCax1}} \\
&=~ \varepsilon\left( \dd\left[\dd[f] \right] \circ \left( \pair{0}{1} \times \pair{0}{1} \right) \right)\tag{Lemma~\ref{chain-linear}} \\
&=~ \varepsilon\left( \dd\left[\dd[f] \right] \circ \four{0}{\pi_0}{0}{\pi_1} \right) \\
&=~ \varepsilon\left(\dd[f] \circ \pair{0 + \varepsilon(\pi_0)}{\pi_1}\right) \tag*{\ref{CdCax6}} \\
&=~ \varepsilon\left(\dd[f] \circ \pair{\varepsilon(\pi_0)}{\pi_1}\right) \\
&=~ \varepsilon\left(\dd[f] \circ \left( \pair{0}{\pi_1}+ \pair{\varepsilon(\pi_0)}{0} \right)\right) \\
&=~ \varepsilon\left(\dd[f] \circ \left( \pair{0}{\pi_1}+ \pair{\varepsilon(\pi_0)}{\varepsilon(0)} \right)\right) \tag*{\ref{Eax1}} \\
&=~ \varepsilon\left(\dd[f] \circ \left( \pair{0}{\pi_1}+ \varepsilon(\pair{\pi_0}{0}) \right)\right) \tag{Lemma~\ref{lem:ep-pair}} \\
&=~ \varepsilon\left(\dd[f] \circ \pair{0}{\pi_1} + \varepsilon\left( \dd[\dd[f]] \circ \four{0}{\pi_1}{\pi_0}{0} \right) \right) \tag*{\ref{CdCax0}} \\
&=~ \varepsilon\left(\dd[f] \circ \pair{0}{\pi_1} \right) + \varepsilon^2\left( \dd[\dd[f]] \circ \four{0}{\pi_1}{\pi_0}{0} \right) \tag*{\ref{Eax1}} \\
&=~ \varepsilon\left(\dd[f] \right) \circ \pair{0}{\pi_1} + \varepsilon^2\left( \dd[\dd[f]] \right) \circ \four{0}{\pi_1}{\pi_0}{0} \tag*{\ref{Eax2}} \\
&=~ \varepsilon\left(\dd[f] \right) \circ \pair{0}{\pi_1} + \varepsilon^k\left( \dd[\dd[f]] \right) \circ \four{0}{\pi_1}{\pi_0}{0} \tag{Iterating Lemma~\ref{lem:d-epsilon-iii}} \\
&=~ \varepsilon\left(\dd[f] \right) \circ \pair{0}{\pi_1} + \dd[\dd[\varepsilon^k\left( f \right) ]] \circ \four{0}{\pi_1}{\pi_0}{0} \tag*{(\ref{CdCax1})} \\
&=~ \varepsilon\left(\dd[f] \right) \circ \pair{0}{\pi_1} + \dd[\dd[0]] \circ \four{0}{\pi_1}{\pi_0}{0} \tag{$\varepsilon$ is nilpotent} \\
&=~ \varepsilon\left(\dd[f] \right) \circ \pair{0}{\pi_1} + 0 \circ \four{0}{\pi_1}{\pi_0}{0} \tag*{\ref{CdCax1}} \\
&=~ \varepsilon\left(\dd[f] \right) \circ \pair{0}{\pi_1} + 0 \\
&=~ \varepsilon(\dd[f]) \circ \pair{0}{1} \circ \pi_1 \\
&=~ \varepsilon(\dd[f] \circ \pair{0}{1}) \circ \pi_1 \tag*{\ref{Eax2}}
\end{align*}
So we conclude that $\varepsilon(\dd[f]) \circ \pair{0}{1}$ is linear and hence $\dd[f] \circ \pair{0}{1}$ is $\varepsilon$-linear.
\end{proof}
Using element-like notation, the first point of the above lemma says that if $f$ is $\varepsilon$-linear then $f(x + \varepsilon(y)) = f(x) + \varepsilon(f(y))$. So $\varepsilon$-linear maps are additive on ``infinitesimal'' elements (i.e.\ those of the form $\varepsilon(y)$). For a Cartesian differential category, linear maps in the Cartesian difference category sense are precisely the same as the Cartesian differential category sense~\cite[Definition 2.2.1]{blute2009cartesian}, while every map is $\varepsilon$-linear since $\varepsilon =0$.
\section{Examples of Cartesian Difference Categories}\label{EXsec}
\subsection{Smooth Functions}\label{smoothex}
As we have shown in Proposition~\ref{CDtoCd}, every Cartesian differential category is a Cartesian difference category where the infinitesimal extension is zero. As a particular example, we consider the category of real smooth functions
which, as mentioned above, can be considered to be the canonical (and motivating) example of a Cartesian differential category.
\begin{defi}
Let $\mathbb{R}$ denote the set of real numbers. The category $\mathsf{SMOOTH}$ is the category whose objects are Euclidean
spaces $\mathbb{R}^n$ (including the point $\mathbb{R}^0 = \lbrace \ast \rbrace$), and whose maps are smooth functions $F:
\mathbb{R}^n \to \mathbb{R}^m$.
\end{defi}
$\mathsf{SMOOTH}$ is a Cartesian left additive category where the product structure is given by the standard Cartesian product of Euclidean spaces and where the additive structure is defined by point-wise addition, $(F+G)(\vec x) = F(\vec x) + G(\vec x)$ and $0(\vec x) = (0, \hdots, 0)$, where $\vec x \in \mathbb{R}^n$. $\mathsf{SMOOTH}$ is a Cartesian differential category where the differential combinator is defined by the directional derivative of smooth functions. Explicitly, for a smooth function $F: \mathbb{R}^n \to \mathbb{R}^m$, which is in fact a tuple of smooth functions $F= (f_1, \hdots, f_n)$ where $f_i: \mathbb{R}^n \to \mathbb{R}$, $\mathsf{D}[F]: \mathbb{R}^n \times \mathbb{R}^n \to \mathbb{R}^m$ is defined as follows:
\[\mathsf{D}[F]\left(\vec x, \vec y \right) := \left( \sum \limits^n_{i=1} \frac{\partial f_1}{\partial u_i}(\vec x) y_i, \hdots, \sum \limits^n_{i=1} \frac{\partial f_n}{\partial u_i}(\vec x) y_i \right)\]
where $\vec x = (x_1, \hdots, x_n), \vec y = (y_1, \hdots, y_n) \in \mathbb{R}^n$. Alternatively, $\mathsf{D}[F]$ can also be defined in terms of the Jacobian matrix of $F$. Therefore $\mathsf{SMOOTH}$ is a Cartesian difference category with infinitesimal extesion $\varepsilon =0$ and with difference combinator $\mathsf{D}$. Since $\varepsilon = 0$, the induced action is simply $\vec x \oplus \vec y = \vec x$. A smooth function is linear in the Cartesian difference category sense precisely if it is $\mathbb{R}$-linear in the classical sense, and every smooth function is $\varepsilon$-linear since $\varepsilon(f)=0$ is linear.
\subsection{Calculus of Finite Differences}\label{discreteex}
The calculus of finite differences~\cite{jordan1965calculus, gleich2005finite} is a field which aims to apply methods from
differential calculus to discrete settings. It does so by introducing a ``discrete derivative'' or ``finite difference''
operator $\Delta$ which, when applied to an integer-valued function $f : \mathbb{Z} \to \mathbb{Z}$, gives the function
$\Delta(f)$ defined by
\[
\Delta(f) (x) = f(x + 1) - f(x)
\]
Here we show that a generalization of the finite difference operator gives an example of a Cartesian
difference category (but \emph{not} a Cartesian differential category). This example was the main motivating example for
developing Cartesian difference categories. The behaviour of the calculus of finite differences, as well as some of its
generalizations (notably, the Boolean differential calculus~\cite{thayse1981boolean, steinbach2009boolean}) is captured by
the category of abelian groups and arbitrary set functions between them, equipped with a suitable Cartesian difference
structure.
\begin{defi}
$\overline{\mathsf{Ab}}$ is the category whose objects are abelian groups $G$ (where we use additive notation for group structure) and where a map $f: G \to H$ is simply an arbitrary function between them (and therefore does not necessarily preserve the group structure).
\end{defi}
\begin{prop}
$\overline{\mathsf{Ab}}$ is a Cartesian left additive category where the product structure is given by the standard Cartesian product of sets and where the additive structure is again given by point-wise addition, $(f+g)(x)=f(x) + g(x)$ and $0(x)=0$. Furthermore, $\overline{\mathsf{Ab}}$ is a Cartesian difference category where the infinitesimal extension is given by the identity, that is, $\varepsilon(f)=f$, and and where the difference combinator $\dd$ is defined as follows for a map $f: G \to H$:
\[\dd[f](x,y) = f(x + y) - f(x)\]
\end{prop}
On the other hand, note that $\dd$ is not a differential combinator for $\overline{\mathsf{Ab}}$ since it does not satisfy~\ref{CDCax6} and part of~\ref{CDCax2}. Indeed, since $f$ is not necessarily a group homomorphism,~\ref{CDCax2} fails to
hold as:
\[\dd[f](x,y +z) = f(x + y +z) - f(x)\]
is not necessarily equal to:
\[\dd[f](x,y) + \dd[f](x,z)= f(x+y) - f(x) + f(x+z) - f(x)\]
$\dd$ does satisfy~\ref{CdCax2} and~\ref{CdCax6}, as well as~\ref{CdCax0}.\ which in this case are respectively:
\[\dd[f](x,y+z) = \dd[f](x,y) + \dd[f](x + y, z)\]
\[\dd \left[ \dd[f] \right]\left( (x,y), (0,z) \right) = \dd[f](x +y, z) \]
\[ f(x + y) = f(x) + \dd[f](x,y) \]
However, as noted in~\cite{FMCS2018}, it is interesting to note that $\dd$ does satisfy~\ref{CDCax1}, the second part of~\ref{CDCax2},~\ref{CDCax3},~\ref{CDCax4},~\ref{CDCax5},~\ref{CDCax7}, and~\ref{CdCax6a}. It is worth emphasizing that in~\cite{FMCS2018}, the goal was to drop the addition and develop a ``non-additive'' version of Cartesian differential
categories, whereas the current presentation keeps the additive structure while suitably relaxing the differential
combinator to a difference combinator.
In $\overline{\mathsf{Ab}}$, since the infinitesimal operator is given by the identity, the induced action is simply the
addition, $x \oplus y = x + y$. On the other hand, the linear maps in $\overline{\mathsf{Ab}}$ are precisely the group
homomorphisms. Indeed, $f$ is linear if $\dd[f](x,y) = f(y)$. But by~\ref{CdCax0} and~\ref{CdCax2}, we get that:
\[f(x + y) = f(x) + \dd[f](x,y)= f(x) + f(y) \quad \quad \quad f(0) = \dd[f](x,0) = 0 \]
So $f$ is a group homomorphism. Conversely, whenever $f$ is a group homomorphism:
\[\dd[f](x,y) = f(x+y) - f(x) = f(x) + f(y) - f(x) = f(y)\]
So $f$ is linear. Since $\varepsilon(f)=f$, the $\varepsilon$-linear maps are precisely the linear maps.
\subsection{Module Morphisms}\label{moduleex}
Here we provide a simple example of a Cartesian difference category whose difference combinator is also a differential combinator, but where the infinitesimal extension is neither zero nor the identity.
\begin{defi}
Let $R$ be a commutative semiring. We define the category $\mathsf{MOD}_R$ as the category whose objects are all
$R$-modules and whose maps are all the $R$-linear maps between them.
\end{defi}
$\mathsf{MOD}_R$ has finite biproducts and is, therefore, a Cartesian additive (in particular, left additive) category.
Every $r \in R$ induces an infinitesimal extension $\varepsilon^r$ defined by scalar multiplication,
$\varepsilon^r(f)(m) = r f(m)$. For any choice of $r$, the category $\mathsf{MOD}_R$ is a Cartesian difference category
with the infinitesimal extension $\varepsilon^r$ and its difference combinator $\dd$ defined as:
\[\dd[f](m,n)=f(n)\]
$R$-linearity of $f$ assures that~\ref{CdCax0} holds, while the remaining Cartesian difference axioms hold trivially. In
fact, $\dd$ is also a differential combinator and therefore $\mathsf{MOD}_R$ is also a Cartesian differential category,
but note that the Cartesian difference structure of $\mathsf{MOD}_R$ is (whenever $R$ is non-zero) different than the
Cartesian difference structure that corresponds to this differential structure. The
induced action is given by $m \oplus n = m + rn$. By definition of $\dd$, every map in $\mathsf{MOD}_R$ is linear, and by
definition of $\varepsilon^r$ and $R$-linearity, every map is also $\varepsilon$-linear.
\newcommand{\seq}[1]{\left[ {#1} \right]}
\newcommand{\Ab}[0]{\overline{\mathsf{Ab}}}
\newcommand{\z}[0]{\mathbf{z}}
\subsection{Stream calculus}\label{streamex}
It is common knowledge that streams, i.e.\ infinite sequences of values, can be studied using methods from differential
calculus. For example, in~\cite{rutten2005coinductive}, a notion of \emph{stream derivative} operator is introduced,
and streams are characterized as the solutions of stream differential equations involving stream derivatives. More recent work in the setting of causal functions between streams of real numbers~\cite{sprunger2019differential, sprunger2019differentiable} has focused on extending the ``classical'' notion of the derivative of a real-valued
function to stream-valued functions.
We seek now to show that causal functions between streams are indeed endowed with the structure of a Cartesian
difference category, the corresponding difference combinator capturing the change of the stream over time. For this we will introduce an idempotent infinitesimal extension on streams that plays a similar role in this setting as Rutten's stream derivative operator~\cite{rutten2005coinductive}, which is given by discarding the head of the stream. On the other hand, our work is more closely related to Sprunger et al.'s work~\cite{sprunger2019differential, sprunger2019differentiable} as it focuses on the differentiation of functions between streams, rather than the description of single streams in terms of differential equations.
For a set $A$, let $A^\omega$ denote the set of infinite sequences of elements of $A$. We write $\seq{a_i}$ for the infinite sequence $\seq{a_i} = (a_0, a_1, a_2, \hdots)$ and $a_{i:\omega}$ for the (infinite) subsequence $(a_i, a_{i + 1}, \hdots)$. A function $f : A^\omega \to B^\omega$ is \textbf{causal} whenever the
$n$-th element ${f\left(\seq{a_i}\right)}_n$ of the output sequence only depends on the first $n$ elements of $\seq{a_i}$. More formally, the function $f$ is causal if and only if, whenever $a_{0:n} = b_{0:n}$,
then ${f\left(\seq{a_i}\right)}_{0:n} = {f\left(\seq{b_i}\right)}_{0:n}$.
We will restrict ourselves to considering streams over abelian groups\footnote{
A similar approach to the one in~\cite{sprunger2019differentiable} is possible where we consider streams on arbitrary difference categories, and lift the difference operator of the underlying category to its category of streams, although it would complicate the presentation of this section without gaining clarity.
}, so let $\Ab^\omega$ be the category whose objects are all the abelian groups and where a morphism from $G$ to $H$ in $\Ab^\omega$ is a causal map from $G^\omega$ to $H^\omega$. $\Ab^\omega$ is a Cartesian left-additive category, where the product is given by the standard product of abelian groups and where the additive structure is lifted point-wise from the structure of $\Ab$.
More concretely, whenever $G, H$ are abelian groups, the hom-set $\Ab^\omega(G, H)$ is endowed with the additive structure that comes from setting
\begin{align*}
{(f+g)\left(\seq{a_i}\right)}_n = {f\left(\seq{a_i}\right)}_n + {g\left(\seq{a_i}\right)}_n && {0\left(\seq{a_i}\right)}_n = 0
\end{align*}
At this point, one might think that the ``natural'' choice of an infinitesimal extension in this setting would
be something akin to Rutten's stream derivative operator, which is given by dropping the first element of the stream, that is, $(\seq{a_i})' = {(\seq{a_i})}_{1:\omega}$. This, however, is not a causal function and so it
does not exist in the category $\Ab^\omega$. That said, it is possible to construct a larger category where this operator is used for differentiation, it however fails to satisfy~\ref{CdCax6}. We instead an define infinitesimal extension for $\Ab^\omega$ as the \emph{truncation} operator $\z$.
\begin{defi} For an abelian group $A$, define the \textbf{truncation operator} $\z_A : A^\omega \to A^\omega$ as follows:
\begin{align*}
{(\z \seq{a_i})}_0 = 0 && {(\z \seq{a_i})}_{j + 1} = a_{j + 1}
\end{align*}
\end{defi}
Note that $\z_A$ is a monoid homomorphism according to the pointwise monoid structure on $A^\omega$. Thus it is straightforward to see that we obtain an infinitesimal extension $\z$ for $\Ab^\omega$.
\begin{thm}
The category $\Ab^\omega$ is a Cartesian difference category, with the infinitesimal extension $\varepsilon(f) = \z \circ f$ and a difference operator defined as:
\begin{align*}
{\dd[f]\left(\seq{a_i}, \seq{b_i}\right)}_0 &= {f\left(\seq{a_i} + \seq{b_i}\right)}_0 - {f\left(\seq{a_i}\right)}_0 \\
{\dd[f]\left(\seq{a_i}, \seq{b_i}\right)}_{n + 1} &= {f\left(\seq{a_i} + \z(\seq{b_i}) \right)}_{n + 1} - {f\left(\seq{a_i}\right)}_{n+1}
\end{align*}
\end{thm}
\begin{proof} We leave it to the reader to check for themselves that $\z$ is an infinitesimal extension.~\ref{CdCax0} is satisfied because of causality. Indeed, since $f$ is causal, $f(\seq{a_i} + \z(\seq{b_i}))$ depends
only on ${\left(\seq{a_i} + \z(\seq{b_i})\right)}_0$, but note that, by definition of $\z$, this term is precisely
equal to $a_0$. Hence we obtain the following:
\begin{align*}
{\left(f(\seq{a_i}) + \z\left(\dd[f](\seq{a_i}, \seq{b_i})\right)\right)}_0
= {f\left(\seq{a_i}\right)}_0
= {f\left(\seq{a_i} + \z(\seq{b_i})\right)}_0
\end{align*}
For any other index, we have that:
\begin{align*}
{\left(f(\seq{a_i}) + \z\left(\dd[f](\seq{a_i}, \seq{b_i})\right)\right)}_{i+1}
&= {f(\seq{a_i})}_{i+1} + {\z\left(\dd[f](\seq{a_i}, \seq{b_i})\right)}_{i+1}\\
&= {f(\seq{a_i})}_{i+1} + {\dd[f]\left(\seq{a_i}, \seq{b_i}\right)}_{i+1}\\
&= {f(\seq{a_i})}_{i+1} + {f\left(\seq{a_i} + \z(\seq{b_i})\right)}_{i+1} - {f(\seq{a_i})}_{i+1}\\
&= {f\left(\seq{a_i} + \z(\seq{b_i})\right)}_{i+1}
\end{align*}
We also explicitly prove the part of~\ref{CdCax1} concerning the infinitesimal extension $\z$:
\begin{align*}
{\dd[\z]\left(\seq{a_i}, \seq{b_i}\right)}_0 &= {\z\left(\seq{a_i} + \seq{b_i}\right)}_0 - {\z(\seq{a_i})}_0 = 0 = {\z(\seq{b_i})}_0\\
{\dd[\z]\left(\seq{a_i}, \seq{b_i}\right)}_{n+1} &= {\z\left(\seq{a_i} + \z(\seq{b_i})\right)}_{n+1} - {\z(\seq{a_i})}_{n+1} = b_{n+1} = {\z(\seq{b_i})}_{n+1}
\end{align*}
Finally, we show that~\ref{CdCax3} holds for the identity map, which is a matter of simple
calculation as well
(linearity of the projection maps follows by an almost identical argument). However, as we remark
later, this property would fail to hold if we had chosen a more ``natural'' candidate for the
difference operator.
\begin{align*}
{\dd[1]\left(\seq{a_i}, \seq{b_i}\right)}_0 &= a_0 + b_0 - a_0 = b_0\\
{\dd[1]\left(\seq{a_i}, \seq{b_i}\right)}_{n+1} &= {\left(\seq{a_i} + \z(\seq{b_i})\right)}_{n+1} - a_{n+1} = b_{n+1}
\end{align*}
The remaining axioms can be shown to hold by similar pointwise reasoning; the corresponding calculations are very similar to the case of $\Ab$.
\end{proof}
\begin{rem}
One might expect the difference operator in $\Ab^\omega$ to be given by the simpler expression
\[
\dd[f](\seq{a_i}, \seq{b_i}) = f(\seq{a_i} + \z(\seq{b_i})) - f(\seq{a_i})
\]
While this satisfies some of the Cartesian difference axioms (notably~\ref{CdCax0})
it does not satisfy all of them: for example,~\ref{CdCax3} fails to hold since
$\dd[1]\left( \seq{a_i}, \seq{b_i}\right) = \z(\seq{b_i}) \neq \seq{b_i}$.
\end{rem}
Note the similarities between the difference combinator on $\Ab$ and that on $\Ab^\omega$. The induced action can be computed out to be:
\begin{align*}
{\left(\seq{a_i} \oplus \seq{b_i}\right)}_0 = a_0 && {\left(\seq{a_i} \oplus \seq{b_i}\right)}_{n+1} = a_{n+1} + b_{n+1}~(\equiv~a_{n+1} \oplus b_{n+1})
\end{align*}
The linear maps (in the Cartesian difference category sense) in $\Ab^\omega$ are precisely those maps $f$ that are group
homomorphisms (when the set of streams $G^\omega$ is equipped with the structure lifted pointwise from the group $G$)
satisfying the additional property that whenever $\seq{a_i}_{1:\omega} = \seq{b_i}_{1:\omega}$ then ${f(\seq{a_i})}_{n+1}
= {f(\seq{b_i})}_{n+1}$, but this is far from evident. The reader can easily verify that any such homomorphism is linear,
we prove here the converse.
\begin{prop}
Any linear map $f$ in $\Ab^\omega$ is a group homomorphism. Furthermore, whenever $\seq{a_i}_{1:\omega} =
\seq{b_i}_{1:\omega}$, the map $f$ satisfies ${f(\seq{a_i})}_{n+1} = {f(\seq{b_i})}_{n+1}$.
\end{prop}
\begin{proof}
The first part of the proposition is simply a corollary of Lemma~\ref{add-linear}, according to which
every linear map in $\Ab^\omega$ is additive and, therefore, a group homomorphism. For the second property, since $f$ is linear, we have:
\begin{align*}
{f\left(\seq{b_i}\right)}_{n+1}
= {\dd[f]\left(\seq{a_i}, \seq{b_i}\right)}_{n+1}
= {f\left(\seq{a_i} + \z(\seq{b_i})\right)}_{n+1} - {f(\seq{a_i})}_{n+1}
\end{align*}
Therefore ${f\left(\seq{a_i} + \z(\seq{b_i})\right)}_{n+1} = {f\left(\seq{a_i}\right)}_{n+1} + {f\left(\seq{b_i}\right)}_{n+1}$.
By setting $\seq{a_i} = \seq{0}$ in the above equation, and since $f$ preserves the identity element,
we establish ${f\left(\z(\seq{b_i})\right)}_{n+1} = {f\left(\seq{b_i}\right)}_{n+1}$, from which the
second part of the desired property follows as an immediate corollary.
\end{proof}
On the other hand, since every map of the form $f \circ \z$ verifies this second property of ``insensitivity'' to the initial element of the stream, it follows that for a map $f$ to be $\varepsilon$-linear it is sufficient (and necessary)
that $f \circ \varepsilon$ be a group homomorphism.
\newcommand{\T}[0]{\mathsf{T}}
\section{Tangent Bundles in Cartesian Difference Categories}%
\label{monadsec}
In this section, we show that the difference combinator of a Cartesian difference category induces a monad, called the \emph{tangent bundle monad}. This construction is a generalization of the tangent bundle monad for Cartesian differential categories~\cite{cockett2014differential,manzyuk2012tangent}. Furthermore, we show that a full subcategory of the Eilenberg-Moore category and the Kleisli category of the tangent bundle monad is again a Cartesian difference category. The general intuition of these categories for the tangent bundle monad are more or less the same as explained in~\cite[Section 3.2]{cockett2014differential}. The maps of the Kleisli category are generalized vector fields, while the considered full subcategory of the Eilenberg-Moore category consists of objects equipped with a linear map which associates tangent vectors to points.
\subsection{The Tangent Bundle Monad} If only to introduce notation, recall that a monad on a category $\mathbb{X}$ is a triple $(\T, \mu, \eta)$ consisting of an endofunctor $\T: \mathbb{X} \to \mathbb{X}$, and two natural transformations $\mu: \mathsf{T}^2 \Rightarrow \mathsf{T}$ and $\eta: \mathsf{1}_\mathbb{X} \Rightarrow \mathsf{T}$ (where $\mathsf{1}_\mathbb{X}: \mathbb{X} \to \mathbb{X}$ is the identity functor), such that the following equalities hold for all objects $A \in \mathbb{X}$:
\begin{equation}\label{monaddef}\begin{gathered} \mu_A \circ \eta_{\T(A)} = 1_{\T(A)} = \mu_A \circ \T(\eta_A) \quad \quad \quad \mu_A \circ \mu_{\T(A)} = \mu_A \circ \T(\mu_A)
\end{gathered}\end{equation}
Now let $\mathbb{X}$ be a Cartesian difference category with infinitesimal extension $\varepsilon$ and difference combinator $\dd$. Define the functor $\mathsf{T}: \mathbb{X} \to \mathbb{X}$ as follows:
\[ \mathsf{T}(A) = A \times A \quad \quad \quad \mathsf{T}(f) = \langle f \circ \pi_0, \dd[f] \rangle \]
and define the natural transformations $\eta: \mathsf{1}_\mathbb{X} \Rightarrow \mathsf{T}$ and $\mu: \mathsf{T}^2 \Rightarrow \mathsf{T}$ as follows:
\[\eta_A := \langle 1_A, 0 \rangle \quad \quad \quad \mu_A := \left \langle \pi_{00}, \pi_{10} + \pi_{01} + \varepsilon(\pi_{11}) \right \rangle \]
where $\pi_{ij}= \pi_i \circ \pi_j$. Before providing the proof that $(\mathsf{T}, \mu, \eta)$ is indeed a monad for any Cartesian difference category, let us provide some examples for intuition.
\begin{exa}\label{ex:tsmooth}
For a Cartesian differential category, since $\varepsilon = 0$, the induced monad is precisely the monad induced by its tangent category structure~\cite{cockett2014differential,manzyuk2012tangent}. For example, in the Cartesian differential category $\mathsf{SMOOTH}$ (as defined in Section~\ref{smoothex}), one has that $\mathsf{T}(\mathbb{R}^n) = \mathbb{R}^n \times \mathbb{R}^n$, which is in fact the classical tangent bundle over $\mathbb{R}^n$ from differential geometry, and also that $\mathsf{T}(F)(\vec x, \vec y) = (F(\vec x), \mathsf{D}[F](\vec x, \vec y))$, $\eta_{\mathbb{R}^n}(\vec x) = (\vec x,\vec 0)$, and $\mu_{\mathbb{R}^n}((\vec x,\vec y), (\vec z, \vec w)) = (\vec x, \vec y + \vec z)$.
\end{exa}
\begin{exa}\label{ex:tab}
In the Cartesian difference category $\overline{\mathsf{Ab}}$ (as defined in Section~\ref{discreteex}), the monad is given by $\mathsf{T}(G) = G \times G$ $\mathsf{T}(f)(x,y) = (f(x), f(x+y) - f(x))$, $\eta_G(x) = (x,0)$, and $\mu_G((x,y),(z,w)) = (x, y + z + w)$.
\end{exa}
We now show that the above construction is indeed a monad.
\begin{prop} Let $\mathbb{X}$ be a Cartesian difference category with infinitesimal extension $\varepsilon$ and difference combinator $\dd$. Then $(\mathsf{T}, \mu, \eta)$, as defined above, is a monad on $\mathbb{X}$.
\end{prop}
\begin{proof} That $\mathsf{T}$ is a functor follows from the fact that the change action model $\alpha : \mathbb{X} \to \mathsf{CAct}(\mathbb{X})$ from Proposition~\ref{CdtoCA} is a functor. Specifically, $\mathsf{T}$ is the second component of $\alpha$. Next we show the naturality of $\eta$ and $\mu$. For $\eta$ we have:
\begin{align*}
\T(f) \circ \eta_A &=~ \pair{f \circ \pi_0}{\dd[f]} \circ \eta_A \\
&=~ \pair{f \circ \pi_0 \circ \eta_A}{
\dd[f] \circ \eta_A} \\
&=~ \pair{f \circ \pi_0 \circ \pair{1_{A}}{0}}{
\dd[f] \circ \pair{1_{A}}{0}} \\
&=~ \pair{f}{0} \tag*{\ref{CdCax2}} \\
&=~ \pair{1_{B}}{0} \circ f \\
&=~ \eta_B \circ f
\end{align*}
For the naturality of $\mu$, first it is straightforward to check that:
\begin{align*}
\T^2(f) = \four{f \circ \pi_{00}}
{\dd[f] \circ \pi_0}
{\dd[f] \circ (\pi_0 \times \pi_0)}
{\dd^2[f]}
\end{align*}
and therefore we compute:
\begin{align*}
&\mu_B \circ \T^2(f) =~ \left \langle \pi_{00}, \pi_{10} + \pi_{01} + \varepsilon(\pi_{11}) \right \rangle \circ \T^2(f) \\
&=~ \left \langle \pi_{00} \circ \T^2(f) , \left(\pi_{10} + \pi_{01} + \varepsilon(\pi_{11}) \right) \circ \T^2(f) \right \rangle \\
&=~ \left \langle \pi_{00} \circ \T^2(f) , \pi_{10} \circ \T^2(f) + \pi_{01} \circ \T^2(f) + \varepsilon(\pi_{11}) \circ \T^2(f) \right \rangle \\
&=~ \left \langle \pi_{00} \circ \T^2(f) , \pi_{10} \circ \T^2(f) + \pi_{01} \circ \T^2(f) + \varepsilon(\pi_{11} \circ \T^2(f) ) \right \rangle \tag*{\ref{Eax2}} \\
&=~ \left \langle f \circ \pi_{00} , \dd[f] \circ \pi_0 + \dd[f] \circ (\pi_0 \times \pi_0)+ \varepsilon(\dd^2[f] ) \right \rangle \\
& =~
\pair{f \circ \pi_{00}}{
\dd[f] \circ (\pi_0 \times \pi_0)
+ \left(
\dd[f] \circ \pi_0 + \varepsilon(\dd^2[f])
\right)
} \\
&=~ \pair{
f \circ \pi_{00}
}{
\dd[f] \circ \pair{\pi_{00}}{\pi_{01}}
+ \left(\dd[f] \circ \pair{\pi_{00}}{\pi_{10}}
+ \varepsilon\left(\dd^2[f] \circ \four{\pi_{00}}{\pi_{10}}{\pi_{01}}{\pi_{11}} \right) \right)
}\\
&=~ \pair{
f \circ \pi_{00}
}{
\dd[f] \circ \pair{\pi_{00}}{\pi_{01}}
+ \dd[f] \circ \pair{\pi_{00} + \varepsilon(\pi_{01})}{\pi_{10} + \varepsilon(\pi_{11})}
}
\tag*{\ref{CdCax0}} \\
&=~ \pair{
f \circ \pi_{00}
}{
\dd[f] \circ \pair{\pi_{00}}{\pi_{01} + \pi_{10} + \varepsilon(\pi_{11})}
}
\tag*{\ref{CdCax2}} \\
&=~ \pair{f \circ \pi_0}{
\dd[f]} \circ \pair{\pi_{00}}{\pi_{01} + \pi_{10} + \varepsilon(\pi_{11})
}
\\
&=~ \T(f) \circ \mu_A
\end{align*}
So $\eta$ and $\mu$ are natural transformations. Now we show that $\mu$ and $\eta$ satisfy the monad identities (\ref{monaddef}). Starting with $\mu_{A} \circ \eta_A = 1_{\T(A)}$:
\begin{align*}
\mu_A \circ \eta_{\T(A)}
&=~ \left \langle \pi_{00}, \pi_{10} + \pi_{01} + \varepsilon(\pi_{11}) \right \rangle \circ \eta_{\T(A)} \\
&=~ \left \langle \pi_{00} \circ \eta_{\T(A)} , \left(\pi_{10} + \pi_{01} + \varepsilon(\pi_{11}) \right) \circ \eta_{\T(A)} \right \rangle \\
&=~ \left \langle \pi_{00} \circ \eta_{\T(A)} , \pi_{10} \circ \eta_{\T(A)} + \pi_{01} \circ \eta_{\T(A)} + \varepsilon(\pi_{11}) \circ \eta_{\T(A)} \right \rangle \\
&=~ \left \langle \pi_{00} \circ \eta_{\T(A)} , \pi_{10} \circ \eta_{\T(A)} + \pi_{01} \circ \eta_{\T(A)} + \varepsilon(\pi_{11} \circ \eta_{\T(A)} ) \right \rangle \tag*{\ref{Eax2}} \\
&=~ \left \langle \pi_{00} \circ \pair{1_{T(A)}}{0} , \pi_{10} \circ \pair{1_{T(A)}}{0} + \pi_{01} \circ \pair{1_{T(A)}}{0} + \varepsilon\left(\pi_{11} \circ \pair{1_{T(A)}}{0} \right) \right \rangle \\
&=~ \pair{\pi_0}{\pi_1 + 0 + \varepsilon(0)}
\\
&=~ \pair{\pi_0}{\pi_1 + 0} \tag*{\ref{Eax1}} \\
&=~ \pair{\pi_0}{\pi_1}\\
&= 1_{\T(A)}
\end{align*}
Next we check that $\mu_{A} \circ \T(\eta_A) = 1_{\T(A)}$. First note that $\T(\eta_A) = \four{\pi_0}{0}{\pi_1}{0}$. Then we compute:
\begin{align*}
\mu_A \circ\T(\eta_{A}) &=~ \pair{\pi_{00}}{\pi_{01} + \pi_{10} + \varepsilon(\pi_{11})}
\circ \T(\eta_{A})
\\
&=~ \pair{\pi_{00} \circ\T(\eta_{A}) }{\left(\pi_{01} + \pi_{10} + \varepsilon(\pi_{11}) \right) \circ\T(\eta_{A}) } \\
&=~ \pair{\pi_{00} \circ\T(\eta_{A})}{\pi_{01}\circ\T(\eta_{A}) + \pi_{10}\circ\T(\eta_{A}) + \varepsilon(\pi_{11}) \circ\T(\eta_{A}) } \\
&=~ \pair{\pi_{00} \circ\T(\eta_{A}) }{\pi_{01}\circ\T(\eta_{A})+ \pi_{10}\circ\T(\eta_{A}) + \varepsilon\left(\pi_{11} \circ\T(\eta_{A})\right)} \tag*{\ref{Eax2}} \\
&=~ \pair{\pi_0}{\pi_1 + 0 + \varepsilon(0)}
\\
&=~ \pair{\pi_0}{\pi_1 + 0} \tag*{\ref{Eax1}} \\
&=~ \pair{\pi_0}{\pi_1} \\
&=~ 1_{\T(A)}
\end{align*}
For the last of the monad laws, we first note that, since $\mu$ is linear,
it follows that $\T(\mu) = \mu \times \mu$. Then it suffices to compute:
\begin{align*}
& \mu_A \circ \T(\mu_A)
= \mu_A \circ (\mu_A \times \mu_A)
\\
&= \pair{\pi_{00}}{\pi_{10} + \pi_{01} + \varepsilon(\pi_{11})}
\circ (\mu_A \times \mu_A)
\\
&= \pair{\pi_{00} \circ (\mu_A \times \mu_A)}{\left( \pi_{10} + \pi_{01} + \varepsilon(\pi_{11}) \right) \circ (\mu_A \times \mu_A)}
\\
&= \pair{\pi_{00} \circ (\mu_A \times \mu_A)}{\pi_{10} \circ (\mu_A \times \mu_A)+ \pi_{01} \circ (\mu_A \times \mu_A)+ \varepsilon(\pi_{11}) \circ (\mu_A \times \mu_A)}
\\
&= \pair{\pi_{00} \circ (\mu_A \times \mu_A)}{\pi_{10} \circ (\mu_A \times \mu_A)+ \pi_{01} \circ (\mu_A \times \mu_A)+ \varepsilon(\pi_{11} \circ (\mu_A \times \mu_A)) }
\tag*{\ref{Eax2}} \\
&=
\pair{\pi_0 \circ \mu_A \circ \pi_0}{\pi_1 \circ \mu_A \circ \pi_0
+ \pi_0 \circ \mu_A \circ \pi_1 + \varepsilon(\pi_1 \circ \mu_A \circ \pi_1)}
\\
&=
\pair{\pi_{000}}{
\pi_{100} + \pi_{010} + \varepsilon(\pi_{110})
+ \pi_{001}
+ \varepsilon \pa{\pi_{101} + \pi_{011} + \varepsilon(\pi_{111})}
}
\\
&=
\pair{\pi_0 \circ \pi_{00}}{
\pi_{1} \circ \pi_{00}
+ \pi_0 \circ \pa{\pi_{10} + \pi_{01} + \varepsilon(\pi_{11})}
+ \varepsilon\pa{\pi_1 \circ \pa{\pi_{10} + \pi_{01} + \varepsilon(\pi_{11})}}
}
\\
&=
\pair{\pi_{00}}{\pi_{10} + \pi_{01} + \varepsilon(\pi_{11}) }
\circ \pair{\pi_{00}}{\pi_{10} + \pi_{01} + \varepsilon(\pi_{11}) }
\\
&= \mu_A \circ \mu_{\T(A)}
\end{align*}
So we conclude that $(\T, \mu, \eta)$ is a monad.
\end{proof}
Those familiar with monads may wonder if the Kleisli triple approach might have simplified the above proof. Recall that for a category $\mathbb{X}$, a Kleisli triple is a triple $(\T, \eta, {(\_)}^\sharp)$ consisting of a function on object $\T$, $A \mapsto \T(A)$, a family of maps indexed by the objects of $\mathbb{X}$, $\eta = \lbrace \eta_A: A \to \T(A) \vert~ A \in \mathbb{X} \rbrace$, and an operator ${(\_)}^\sharp$ which for any map $f: A \to \T(B)$ results in a map $f^\sharp: \T(A) \to \T(B)$, and such that the following equalities hold:
\[ f^\sharp \circ \eta_A = f \quad \quad \quad \eta_A^\sharp = 1_{\T(A)} \quad \quad \quad {(g^\sharp \circ f)}^\sharp = g^\sharp \circ f^\sharp \]
There is a bijective correspondence between monads and Kleisli triples. So for the tangent bundle monad, we could have instead defined its associated Kleisli triple and prove 3 equations instead of the functoriality of $\T$, the naturality of $\mu$ and $\eta$, and the 3 monad identities. However, we find the operator ${(\_)}^\sharp$ slightly more complicated to work with. As such, we elected to work out the monad identities directly, since while there are more identities to prove, we find the computations simpler and the proof easier to follow. That said, it is still interesting to work out the ${(\_)}^\sharp$ for the tangent bundle monad. In general, for a map $f: A \to \T(B)$, $f^\sharp: \T(A) \to \T(B)$ is equal to the composite $\T(f) = \mu_B \circ \T(f)$. In the case of the tangent bundle monad, note that a $f: A \to \T(B)$ would be a pair $f = \langle f_0, f_1 \rangle$, and so we obtain that (which we leave an excercise for the reader to compute for themselves):
\begin{align*}
f^\sharp = \pair{f_0 \circ \pi_0}{ f_1 \circ \pi_0 + \partial[f_0]+ \varepsilon\left( \partial[f_1] \right) }
\end{align*}
We next observe that the tangent bundle functor $\T$ preserves finite products up to isomorphism. Indeed, note that $\T(A \times B) \cong \T(A) \times \T(B)$ via the canonical natural isomorphism:
\begin{equation}\label{phidef1}\begin{gathered} \phi_{A,B}: \T(A \times B) \to \T(A) \times \T(B) \quad \quad \quad \quad \phi_{A,B} := \pair{T(\pi_0)}{T(\pi_1)}
\end{gathered}\end{equation}
By~\ref{CdCax3}, it follows that $\phi_{A,B} = \pair{\pi_0 \times \pi_0}{\pi_1 \times \pi_1}$ and its inverse $\phi^{-1}_{A,B}: \T(A) \times \T(B) \to \T(A \times B)$ is defined in the same way, that is, $\phi^{-1}_{A,B} = \pair{\pi_0 \times \pi_0}{\pi_1 \times \pi_1}$. Expanding this out, we can see that $\phi_{A,B}: (A \times B) \times (A \times B) \to (A \times A) \times (B \times B)$ swaps the middle two arguments, and so does $\phi^{-1}_{A,B}$. The following lemma will be extremely useful in many proofs throughout the remainder of this section.
\begin{lem}\label{lem:t-prod} In a Cartesian difference category:
\begin{enumerate}[(\roman{enumi}),ref={\thelem.\roman{enumi}}]
\item $\mu$, $\eta$, and $\phi$ are linear;
\item\label{T-sum} $\T(f + g) = \T(f) + \T(g)$ and $\T(0) = 0$;
\item If $f$ is linear then $\T(f)$ is linear and $\T(f) = f \times f$;
\item $\T(\pi_i) = \pi_i \circ \phi$ and $\phi \circ \T(\pair{f}{g}) = \pair{\T(f)}{\T(g)}$ and $\phi \circ \T(h \times k) = (\T(h) \times \T(k)) \circ \phi$;
\item\label{T-diff} $\T(\dd[f]) = \dd[\T(f)] \circ \phi$ and $\T(\varepsilon(f)) = \varepsilon(\T(f))$;
\item\label{lem:mu-ep} $\varepsilon(\mu_A) = \mu_A \circ \varepsilon(1_{\T(A)})$, $\varepsilon(\eta_A) = \eta_A \circ \varepsilon(1_A)$, and $\varepsilon(\phi_{A,B}) = \phi_{A,B} \circ \varepsilon(1_{\T(A \times B)})$
\end{enumerate}
\end{lem}
\begin{proof} The first three are mostly straightforward. Indeed, (i) follows immediately from Lemma~\ref{lem:linear} and by construction of $\mu$, $\eta$, and $\phi$, while (ii) follows from~\ref{CdCax1}, and (iii) follows from the definition of $\T$ and linear maps. Next, we compute the identities of (iv), the first two follow from~\ref{CdCax3} and~\ref{CdCax4}. We first show $\T(\pi_i) = \pi_i \circ \phi$:
\begin{align*}
\T(\pi_i) &=~ \pair{\pi_i \circ \pi_0}{\dd[\pi_i]} \\
&=~ \pair{\pi_i \circ \pi_0}{\pi_i \circ \pi_1} \tag*{\ref{CdCax3}} \\
&=~ \pi_i \times \pi_i \\
&=~ \pi_i \circ \phi \end{align*}
Next we compute that $\phi \circ \T(\pair{f}{g}) = \pair{\T(f)}{\T(g)}$:
\begin{align*}
&\phi \circ \T(\pair{f}{g}) =~ \phi \circ \pair{\pair{f}{g} \circ \pi_0}{\dd[\pair{f}{g}]} \\
&=~ \phi \circ \four{f \circ \pi_0}{g \circ \pi_0}{\dd[f]}{\dd[g]} \tag*{\ref{CdCax4}} \\
&=~ \pair{\pi_0 \times \pi_0}{\pi_1 \times \pi_1} \circ \four{f \circ \pi_0}{g \circ \pi_0}{\dd[f]}{\dd[g]} \\
&=~ \pair{(\pi_0 \times \pi_0) \circ \four{f \circ \pi_0}{g \circ \pi_0}{\dd[f]}{\dd[g]}}{(\pi_1 \times \pi_1) \circ \four{f \circ \pi_0}{g \circ \pi_0}{\dd[f]}{\dd[g]}} \\
&=~ \pair{\pair{\pi_0 \circ \pair{f \circ \pi_0}{g \circ \pi_0}}{\pi_0 \circ \pair{\dd[f]}{\dd[g]}}}{\pair{\pi_1 \circ \pair{f \circ \pi_0}{g \circ \pi_0}}{\pi_1 \circ\pair{ \dd[f]}{\dd[g]}}} \\
&=~ \four{f \circ\pi_0}{\dd[f]}{g \circ \pi_0}{\dd[g]} \\
&=~ \pair{\T(f)}{\T(g)}
\end{align*}
For the remaining identity, $\phi \circ \T(h \times k) = (\T(h) \times \T(k)) \circ \phi$, we use the previous one:
\begin{align*}
\phi \circ \T(h \times k) &=~ \phi \circ \T(\pair{h \circ \pi_0}{k \circ \pi_1}) \\
&=~ \pair{\T(h \circ \pi_0)}{\T(k \circ \pi_1)} \tag{Lem.\ref{lem:t-prod}.iv} \\
&=~ \pair{\T(h) \circ \T(\pi_0)}{\T(k) \circ \T(\pi_1)} \tag{$\T$ is a functor} \\
&=~ \pair{\T(h) \circ (\pi_0 \times \pi_0)}{\T(k) \circ (\pi_1 \times \pi_1)} \tag{Lem.\ref{lem:t-prod}.iii} \\
&=~ \pair{\T(h) \circ \pi_0 \circ \phi}{\T(k) \circ \pi_1 \circ \phi} \\
&=~ \pair{\T(h) \circ \pi_0}{ \T(k) \circ \pi_1} \circ \phi \\
&=~ (\T(h) \times \T(k)) \circ \phi
\end{align*}
Next we compute the two identities of (v). First we show $\T(\dd[f]) = \dd[\T(f)] \circ \phi$:
\begin{align*}
&\dd[\T(f)] \circ \phi =~ \dd[\pair{f \circ \pi_0}{\dd[f]}] \circ \phi \\
&=~ \pair{\dd[f \circ \pi_0]}{\dd^2[f]} \circ \phi \tag*{\ref{CdCax4}} \\
&=~ \pair{\dd[f] \circ (\pi_0 \times \pi_0)}{\dd^2[f]} \circ \phi \tag{Lem.\ref{lem:linear}} \\
&=~ \pair{\dd[f] \circ (\pi_0 \times \pi_0) \circ \phi}{\dd^2[f] \circ \phi} \\
&=~ \pair{\dd[f] \circ (\pi_0 \times \pi_0) \circ \pair{\pi_0 \times \pi_0}{\pi_1 \times \pi_1}}{\dd^2[f] \circ \pair{\pi_0 \times \pi_0}{\pi_1 \times \pi_1}} \\
&=~ \pair{\dd[f] \circ \pair{\pi_0 (\pi_0 \times \pi_0)}{\pi_0 \circ (\pi_1 \times \pi_1)}}{\dd^2[f] \circ \four{\pi_0 \circ \pi_0}{\pi_0 \circ \pi_1}{\pi_1 \circ \pi_0}{\pi_1 \circ \pi_1}} \\
&=~ \pair{\dd[f] \circ \pair{\pi_1 \circ \pi_0}{\pi_1 \circ \pi_0}}{\dd^2[f] \circ \four{\pi_0 \circ \pi_0}{\pi_0 \circ \pi_1}{\pi_1 \circ \pi_0}{\pi_1 \circ \pi_1}} \\
&=~ \pair{\dd[f] \circ \pair{\pi_0 \circ \pi_0}{\pi_1 \circ \pi_0}}{\dd^2[f] \circ \four{\pi_0 \circ \pi_0}{\pi_1 \circ \pi_0}{\pi_0 \circ \pi_1}{\pi_1 \circ \pi_1}} \tag*{\ref{CdCax7a}}\\
&=~ \pair{\dd[f] \circ \pi_0}{\dd^2[f]} \\
&=~ \T(\dd[f])
\end{align*}
Next we show $\T(\varepsilon(f)) = \varepsilon(\T(f))$:
\begin{align*}
\T(\varepsilon(f)) &=~ \pair{\varepsilon(f) \circ \pi_0}{\dd[\varepsilon(f)]} \\
&=~ \pair{\varepsilon(f \circ \pi_0)}{\varepsilon(\dd[f])} \tag*{\ref{Eax2} +~\ref{CdCax1}} \\
&=~\varepsilon(\pair{f \circ \pi_0}{\dd[f]}) \tag{Lem.\ref{lem:ep-pair}} \\
&=~ \varepsilon(\T(f))
\end{align*}
Lastly (vi) follows from the fact we have shown that $\mu$, $\eta$, and $\phi$ are linear and so by Lemma~\ref{epsilon-linear}, the desired equalities hold.
\end{proof}
\subsection{The Eilenberg-Moore Category of \texorpdfstring{$\mathsf{T}$}{T}}%
\label{EMsec}
Recall that a $\T$-algebra of the monad $(\mathsf{T}, \mu, \eta)$ is a pair $(A, \nu)$ consisting of an object $A$ and a map $\nu: \T(A) \to A$ such that $\nu \circ \eta_A = 1_A$ and $\nu \circ \T(\nu) = \nu \circ \mu_A$, and that a $\T$-algebra morphism $f: (A, \nu) \to (B, \omega)$ is a map $f: A \to B$ such that $\omega \circ\T(f) = f \circ \nu$. The Eilenberg-Moore category of $(\mathsf{T}, \mu, \eta)$ is the category of $\T$-algebras, that is, the category $\mathbb{X}^\mathsf{T}$ whose objects are $\T$-algebras and whose maps are $\T$-algebra morphisms. It is well known that for any monad on a category with finite products, the Eilenberg-Moore category also has finite products. Indeed, in this case, for $\T$-algebras $(A, \nu)$ and $(B, \omega)$, their product is defined as:
\[ (A, \nu) \times (B, \omega) := (A \times B, (\nu \times \omega) \circ \phi_{A,B}) \]
while the projection maps and the pairing of maps are the same as in the base category. The terminal object in $\mathbb{X}^\mathsf{T}$ is defined as $(\top, 0)$.
Therefore, it may be tempting to think that the Eilenberg-Moore category of the tangent bundle monad is also a Cartesian difference category. Unfortunately, there are two problems with this: the sum of a $\T$-algebra morphism is not necessarily a $\T$-algebra morphism and the differential of a $\T$-algebra may not necessarily be a $\T$-algebra. Indeed, let $f,g: (A, \nu) \to (B, \omega)$ be $\T$-algebra morphism. On the one hand, since $\T(f + g) = \T(f) + \T(g)$ we have that:
\[ \omega \circ \T(f+ g) = \omega \circ (\T(f) + \T(g)) = \omega \circ \T(f) + \omega \circ \T(g) = f \circ \nu + g \circ \nu \]
However, if $\nu$ is not additive then $f \circ \nu + g \circ \nu$ may not be equal to $(f + g) \circ \nu$. As such, one solution could be to consider the subcategory of additive $\T$-algebras. While this full subcategory of additive $\T$-algebras will be a Cartesian left additive category, this does not solve the problem of the differential of a $\T$-algebra morphism. For a $\T$-algebra morphism $f: (A, \nu) \to (B, \omega)$, its derivative $\dd[f]: A \times A \to B$ should be a $\T$-algebra morphism of type $(A, \nu) \times (A, \nu) \to (B, \omega)$. However one cannot get very far in trying to show that $\dd[f]$ is a $\T$-algebra morphism. The solution to this problem is instead to consider the full subcategory of \emph{linear} $\T$-algebras.
Define $\mathbb{X}^{\T}_{lin}$ as the category of \emph{linear} $\T$-algebra, that is, the category whose objects are $\T$-algebras $(A, \nu)$ such that $\nu$ is linear and whose maps are arbitrary $\T$-algebra morphisms between them. In particular, note that a map in $\mathbb{X}^{\T}_{lin}$ need not be linear. Like the Eilenberg-Moore category, there is an obvious forgetful functor $\mathsf{U}: \mathbb{X}^{\T}_{lin} \to \mathbb{X}$ defined as $\mathsf{U}(A, \nu) = A$ and $\mathsf{U}(f) = f$. A linear $\T$-algebra structure should be interpreted as a map which linearly modifies a point by a tangent vector. Unfortunately, as mentioned in~\cite{cockett2014differential}, the Eilenberg-Moore category of the tangent bundle monad has yet to be studied in full, even in classical differential geometry.
\begin{exa}
For the tangent bundle monad on the Cartesian differential category $\mathsf{SMOOTH}$ (Example~\ref{ex:tsmooth}), a linear $\T$-algebra is a pair $(\mathbb{R}^n, \nu)$ such that $\nu(\vec x, \vec y) = \vec x + t \vec y$ for some fixed $t \in \mathbb{R}$. Therefore, $\mathsf{SMOOTH}^{\T}_{lin}$ is equivalent to the category whose objects are pairs $(\mathbb{R}^n, t)$, with $t \in \mathbb{R}$, where a map $F: (\mathbb{R}^n, t) \to (\mathbb{R}^m, s)$ is a smooth function $F: \mathbb{R}^n \to \mathbb{R}^m$ such that $F(\vec x + t \vec y) = F(\vec x) + s \mathsf{D}[F](\vec x, \vec y)$.
\end{exa}
\begin{exa}
For the tangent bundle monad on the Cartesian differential category $\overline{\mathsf{Ab}}$ (Example~\ref{ex:tab}), a linear $\T$-algebra is a pair $(G, \nu)$ such that $\nu(x,y) = x + e(y)$ for some group endomorphism $e: G \to G$. Therefore, $\overline{\mathsf{Ab}}$ is equivalent to the category whose objects are pairs $(G,e)$, with $G$ an abelian group and $e: G \to G$ a group endomorphism, where a map $f: (G, e) \to (H, e^\prime)$ is a function $f: G \to H$ such that \[f( x + e(y) ) = f(x) + e^\prime \left( f(x+y) \right) - e^\prime(f(x)).\]
\end{exa}
We will now explain how $\mathbb{X}^{\T}_{lin}$ is a Cartesian difference category. Simply put, the Cartesian difference structure of $\mathbb{X}^{\T}_{lin}$ is the same as $\mathbb{X}$, and therefore we say that the forgetful functor preserves the Cartesian difference structure strictly. Starting with the Cartesian left additive structure: the finite product structure of $\mathbb{X}^{\T}_{lin}$ is defined in the same way as $\mathbb{X}^{\T}$, while the sum of $\T$-algebra morphisms and zero $\T$-algebra morphisms are defined as the sum and zero maps in $\mathbb{X}$. Of course, we must check that this is well-defined.
\begin{lem} $\mathbb{X}^{\T}_{lin}$ is a Cartesian left additive category.
\end{lem}
\begin{proof} We must first check that if $(A, \nu)$ and $(B, \omega)$ are linear $\T$-algebras, then so is their product $(A, \nu) \times (B, \omega)$, that is, we must show that $(\nu \times \omega) \circ \phi_{A,B}$ is linear. However, since $\nu$, $\omega$, and $\phi_{A,B}$ are linear, it follows from Lemma~\ref{comp-linear} that $(\nu \times \omega) \circ \phi_{A,B}$ is also linear. Therefore, the product of linear $\T$-algebras is also a linear $\T$-algebra. Next, since zero maps are linear, $(\top, 0)$ is also a linear $\T$-algebra. Therefore $\mathbb{X}^{\T}_{lin}$ has finite products since it is a full subcategory of $\mathbb{X}^\mathsf{T}$. Next, we show that if $f: (A, \nu) \to (B, \omega)$ and $g: (A, \nu) \to (B, \omega)$ are $\T$-algebra morphisms between linear $\T$-algebras, then their sum $f +g: (A, \nu) \to (B, \omega)$ is also a $\T$-algebra morphism. This follows from the fact that linear maps are additive (Lemma~\ref{add-linear}):
\begin{align*}
\omega \circ \T(f+ g) &=~ \omega \circ (\T(f) + \T(g)) \tag{by Prop.\ref{T-sum}} \\
&=~ \omega \circ \T(f) + \omega \circ \T(g)\\
&=~ f \circ \nu + g \circ \nu \tag{by $\T$-alg. morph. def.} \\
&=~ (f + g) \circ \nu \tag{$\nu$ is additive}
\end{align*}
So we have that $f+g$ is a $\T$-algebra morphism. Similarly, we must show that zero maps between linear $\T$-algebras $(A, \nu)$ and $(B, \omega)$ are $\T$-algebra morphisms:
\begin{align*}
\omega \circ \T(0) &=~ \omega \circ 0 \tag{by Prop.\ref{T-sum}} \\
&=~ 0 \\
&=~ 0 \circ \nu \tag{$\nu$ is additive}
\end{align*}
Therefore $0: (A, \nu) \to (B, \omega)$ is a $\T$-algebra morphism. And clearly, the remaining Cartesian left additive structure axioms hold since composition in $\mathbb{X}^{\T}_{lin}$ is the same as in $\mathbb{X}$. So we conclude that $\mathbb{X}^{\T}_{lin}$ is a Cartesian left additive category.
\end{proof}
Similarly, the infinitesimal extension and the difference combinator of $\mathbb{X}^{\T}_{lin}$ are defined in the same way as in $\mathbb{X}$. Once again, while the infinitesimal extension and difference combinator axioms will automatically hold, we will have to check that this is all well-defined.
\begin{prop} For a Cartesian difference category $\mathbb{X}$, the category of linear $\T$-algebras $\mathbb{X}^{\T}_{lin}$ is a Cartesian difference category such that the obvious forgetful functor $\mathsf{U}: \mathbb{X}^{\T}_{lin} \to \mathbb{X}$ preserves the Cartesian difference structure strictly.
\end{prop}
\begin{proof} We first show that if $f: (A, \nu) \to (B, \omega)$ is a $\T$-algebra morphism between linear $\T$-algebras, then so is $\varepsilon(f)$:
\begin{align*}
\omega \circ \T(\varepsilon(f)) &=~\omega \circ \varepsilon(\T(f)) \tag{by Prop.\ref{T-diff}} \\
&=~ \varepsilon(\omega \circ \T(f)) \tag{by Lem.\ref{epsilon-linear}} \\
&=~ \varepsilon(f \circ \nu) \tag{by $\T$-alg. morph. def.} \\
&=~ \varepsilon(f) \circ \nu \tag{by $\varepsilon$ def.}
\end{align*}
So we have that $\varepsilon(f): (A, \nu) \to (B, \omega)$ is a $\T$-algebra morphism. Next, we check that the derivative $\dd[f]: A \times A \to B$ is a $\T$-algebra morphism of type $(A, \nu) \times (A, \nu) \to (B, \omega)$.
\begin{align*}
\omega \circ \T(\dd[f]) &=~\omega \circ \dd[\T(f)] \circ \phi_{A,A} \tag{by Prop.\ref{T-diff}} \\
&=~ \dd[\omega \circ \T(f)] \circ \phi_{A,A} \tag{by Lem.\ref{chain-linear} since $\omega$ is linear} \\
&=~ \dd[f \circ \nu] \circ \phi_{A,A} \tag{by $\T$-alg. morph. def.} \\
&=~ \dd[f] \circ (\nu \times \nu) \circ \phi_{A,A} \tag{by Lem.\ref{chain-linear} since $\nu$ is linear}
\end{align*}
So we have that $\dd[f]: (A, \nu) \times (A, \nu) \to (B, \omega)$ is a $\T$-algebra morphism. Since composition and the Cartesian left additive structure of $\mathbb{X}^{\T}_{lin}$ is the same as $\mathbb{X}$, it automatically follows that $\varepsilon$ is an infinitesimal extension and $\dd$ is a difference combinator on $\mathbb{X}^{\T}_{lin}$. Therefore, we conclude that $\mathbb{X}^{\T}_{lin}$ is a Cartesian difference category, and clearly the forgetful functor preserves the Cartesian difference structure strictly.
\end{proof}
\subsection{The Kleisli Category of \texorpdfstring{$\mathsf{T}$}{T}}%
\label{Kleislisec}
The construction found here is different from the one found in the conference paper~\cite{alvarez2020cartesian}. Indeed, the proposed infinitesimal extension and difference combinator in~\cite{alvarez2020cartesian} were based on the ones that appeared in~\cite{alvarez2019change}. Unfortunately, we have found that said proposed infinitesimal extension and difference combinator failed to satisfy {\bf [C$\dd$.2]} and therefore both of the aforementioned results in~\cite{alvarez2020cartesian,alvarez2019change} are incorrect. We rectify this mistake here by changing the infinitesimal extension to the correct one, while keeping the difference combinator the same.
Recall that the Kleisli category of the monad $(\mathsf{T}, \mu, \eta)$ is defined as the category $\mathbb{X}_\mathsf{T}$ whose objects are the objects of $\mathbb{X}$, and where a map $A \to B$ in $\mathbb{X}_\mathsf{T}$ is a map $f: A \to \mathsf{T}(B)$ in $\mathbb{X}$, which would be a pair $f = \langle f_0, f_1 \rangle$ where $f_j: A \to B$. The identity map in $\mathbb{X}_\mathsf{T}$ is the monad unit $\eta_A: A \to \mathsf{T}(A)$, while composition of Kleisli maps $f: A \to \mathsf{T}(B)$ and $g: B \to \mathsf{T}(C)$ is defined as the composite $\mu_C \circ \mathsf{T}(g) \circ f$. To distinguish between composition in $\mathbb{X}$ and $\mathbb{X}_\mathsf{T}$, we denote the Kleisli composition as follows:
\begin{equation}\label{Kcomp1}\begin{gathered} g \circ^\T f = \mu_C \circ \mathsf{T}(g) \circ f
\end{gathered}\end{equation}
If $f=\langle f_0, f_1 \rangle$ and $g=\langle g_0, g_1 \rangle$, then their Kleisli composition can be worked out to be:
\begin{equation}\label{Kcomp2}\begin{gathered} g \circ^\T f = \langle g_0, g_1 \rangle \circ^\T \langle f_0, f_1 \rangle = \left \langle g_0 \circ f_0, \dd[g_0] \circ \langle f_0, f_1 \rangle + g_1 \circ (f_0 \oplus f_1) \right \rangle \end{gathered}\end{equation}
Kleisli maps can be understood as ``generalized'' vector fields. Indeed, $\T(A)$ should be thought of as the tangent bundle over $A$, and therefore a vector field would be a map $\langle 1, f \rangle: A \to \T(A)$, which is of course also a Kleisli map. For more details on the intuition behind this Kleisli category see~\cite{cockett2014differential}. Furthermore, in general, the Kleisli category of a monad is equivalently to a full subcategory of the Eilenberg-Moore consisting of the free algebras. In this case, note that every free $\T$-algebra $(\T(A), \mu_A)$ is also an object in $\mathbb{X}^{\T}_{lin}$, since $\mu_A$ is linear. Therefore, $\mathbb{X}_\mathsf{T}$ is equivalent to the full subcategory of free $\T$-algebras of $\mathbb{X}^{\T}_{lin}$.
\begin{exa}
For the tangent bundle monad on the Cartesian differential category $\mathsf{SMOOTH}$ (Example~\ref{ex:tsmooth}), a Kleisli map is a smooth function $F: \mathbb{R}^n \to \mathbb{R}^m \times \mathbb{R}^m$, which is interpreted as a pair of smooth functions $F = \langle F_0, F_1 \rangle$ where $F_i: \mathbb{R}^n \to \mathbb{R}^m$. The composition of Kleisli maps $F = \langle F_0, F_1 \rangle$ and $G = \langle G_0, G_1 \rangle$ is computed out to be $(G \circ^\T F)(\vec x) = ( G_0(F_0(\vec x)), \mathsf{D}[G_0](F_0(\vec x), F_1(\vec x)) + G_1(F_0(\vec x))$. Among these Kleisli maps are the vector fields on $\mathbb{R}^n$, which are precisely the Kleisli maps of the form $F = \langle 1_{\mathbb{R}^n}, f \rangle$, for some smooth function $f: \mathbb{R}^n \to \mathbb{R}^n$.
\end{exa}
\begin{exa}
For the tangent bundle monad on the Cartesian differential category $\overline{\mathsf{Ab}}$ (Example~\ref{ex:tab}), a Kleisli map is a function $f: G \to H \times H$, which is interpreted as a pair of functions $f = \langle f_0, f_1 \rangle$ where $f_i: G \to H$. In this case, the composition of Kleisli maps $f = \langle f_0, f_1 \rangle$ and $g = \langle g_0, g_1 \rangle$ is $(g \circ^\T g)(x) = ( g_0(g_0(x)), g_0\left( f_0(x) + f_1(x) \right) - g_0(f_0(x)) + g_1(f_0(x))$.
\end{exa}
We now wish to explain how the Kleisli category $\mathbb{X}_\mathsf{T}$ is again a Cartesian difference category. We begin by exhibiting the Cartesian left additive structure of the Kleisli category $\mathbb{X}_\mathsf{T}$. Generally, the Kleisli category does not automatically inherit the product structure of the base category, even if the Eilenberg-Moore category does. However, since $\T$ preserves finite products up to isomorphism, it follows that its Kleisli category has finite products. As such, the product of objects in $\mathbb{X}_\mathsf{T}$ is defined as $A \times B$ with projections $\pi^{\mathsf{T}}_0: A \times B \to \mathsf{T}(A)$ and $\pi^{\mathsf{T}}_1: A \times B \to \mathsf{T}(B)$ defined respectively as $\pi^{\mathsf{T}}_0 = \langle \pi_0, 0 \rangle$ and $\pi^{\mathsf{T}}_1 = \langle \pi_1, 0 \rangle$, and the pairing of Kleisli maps $f=\langle f_0, f_1 \rangle$ and $g=\langle g_0, g_1 \rangle$ is defined as:
\begin{equation}\label{Kpair}\begin{gathered} \langle f, g \rangle^\mathsf{T} = \phi^{-1} \circ \pair{f}{g} = \four{f_0}{g_0}{f_1}{g_1} \end{gathered}\end{equation}
where recall $\phi^{-1}:\T(A) \times\T(B) \to\T(A \times B)$ is the inverse of $\phi$ as defined in (\ref{phidef1}). The terminal object is again $\top$ and where the unique map to the terminal object is $!^{\mathsf{T}}_A = 0$. The sum of Kleisli maps $f=\langle f_0, f_1 \rangle$ and $g=\langle g_0, g_1 \rangle$ is defined as:
\[f +^\mathsf{T} g = f + g = \langle f_0 + g_0, f_1 + g_1 \rangle\]
and the zero Kleisli maps is simply $0^\T = 0 = \langle 0, 0 \rangle$. Therefore we conclude that the Kleisli category of the tangent monad is a Cartesian left additive category.
\begin{lem} $\mathbb{X}_\mathsf{T}$ is a Cartesian left additive category.
\end{lem}
\begin{proof} As explained above, by Lemma~\ref{lem:t-prod}, $\T$ preserves the finite product structure, and thus it follows that the Kleisli category $\mathbb{X}_\mathsf{T}$ is a Cartesian category. So it remains to show that $\mathbb{X}_\mathsf{T}$ is a left additive category and that the projection maps are additive. We start by showing that the proposed additive structure is compatible with Kleisli composition. This follows from the fact that $\T$ preserves the additive structure by Lemma~\ref{lem:t-prod}.(ii) and Lemma~\ref{add-linear}, that $\mu$ is linear and therefore also additive.
\begin{align*}
(f +^\T g) \circ^\T x &=~ (f+g) \circ^\T x \\
&=~ \mu \circ \T(f +g) \circ x \\
&=~ \mu \circ (\T(f) + \T(g) ) \circ x \tag{Lem\ref{lem:t-prod}.ii} \\
&=~ \mu \circ (\T(f) \circ x + \T(g) \circ x) \\
&=~ \mu \circ \T(f) \circ x + \mu \circ \T(g) \circ x \tag{$\mu$ is additive} \\
&=~ f \circ^\T x + g \circ^\T x \\
&=~ f \circ^\T x +^\T g \circ^\T x \\ \\
0^\T \circ x &=~ 0 \circ^\T x \\
&=~ \mu \circ\T(0) \circ x \\
&=~ \mu \circ 0 \circ x \tag{Lem\ref{lem:t-prod}.ii} \\
&=~0 \circ x \tag{$\mu$ is additive} \\
&=~ 0
\end{align*}
So we have that $\mathbb{X}_\mathsf{T}$ is a left additive category. Next we show that the projection maps are additive. Note that $\pi_i^\T = \eta \circ \pi_i$ and that by Lemma\ref{lem:t-prod}.(iii), $\T(\pi_i) = \pi_i \times \pi_i$ is linear and therefore also additive. So we have that:
\begin{align*}
\pi^\T_i \circ^\T (x + y) &=~ (\eta \circ \pi_i) \circ^\T (x+y) \\
&=~ \mu \circ \T(\eta \circ \pi_i) \circ (x+y) \\
&=~ \mu \circ \T(\eta) \circ \T(\pi_i) \circ (x+y) \tag{$\T$ is a functor} \\
&=~ \T(\pi_i) \circ (x+y) \tag{Monad identities} \\
&=~ \T(\pi_i) \circ x + \T(\pi_i) \circ y \tag{$\T(\pi_i)$ is additive} \\
&=~\mu \circ \T(\eta) \circ \T(\pi_i) \circ x + \mu \circ \T(\eta) \circ \T(\pi_i) \circ y \tag{Monad identities} \\
&=~ \mu \circ \T(\eta \circ \pi_i) \circ x + \mu \circ \T(\eta \circ \pi_i) \circ y \tag{$\T$ is a functor} \\
&=~ (\eta \circ \pi_i) \circ^\T x + (\eta \circ \pi_i) \circ^\T y \\
&=~ \pi^\T_i \circ^\T x + \pi^\T_i \circ^\T y \\ \\
\pi^\T_i \circ^\T 0^\T &=~ (\eta \circ \pi_i) \circ^\T 0 \\
&=~ \mu \circ \T(\eta \circ \pi_i) \circ 0 \\
&=~ \mu \circ \T(\eta) \circ \T(\pi_i) \circ 0 \tag{$\T$ is a functor} \\
&=~ \T(\pi_i) \circ 0 \tag{Monad identities} \\
&=~ 0 \tag{$\T(\pi_i)$ is additive}
\end{align*}
So we conclude that $\mathbb{X}_\mathsf{T}$ is a Cartesian left additive category.
\end{proof}
The infinitesimal extension $\varepsilon^\T$ for the Kleisli category is the same as the infinitesimal extension of the base category, that is, for a Kleisli map $f= \langle f_0, f_1 \rangle$:
\[ \varepsilon^\T(f) = \varepsilon(f) = \varepsilon(\pair{f_0}{f_1}) = \pair{\varepsilon(f_0)}{\varepsilon(f_1)}\]
where the last equality is due to Lemma~\ref{lem:ep-pair}. It is important to note that this is the major correction made from the conference paper version. Indeed, the infinitesimal extension suggested in~\cite{alvarez2020cartesian} was based on the change action structure suggested in~\cite{alvarez2019change}. However, it unfortunately turns out that a result in~\cite{alvarez2019change} was incorrect, since~\ref{CADax2} actually fails, and therefore the infinitesimal extension and difference combinator suggested in~\cite{alvarez2020cartesian} do not satisfy~\ref{CdCax2}. Luckily, the infinitesimal extension provide in this paper does work, as we carefully prove below.
\begin{lem} $\varepsilon^\T$ is an infinitesimal extension on $\mathbb{X}_\mathsf{T}$.
\end{lem}
\begin{proof} We check that $\varepsilon^\T$ satisfies~\ref{Eax1},~\ref{Eax2}, and~\ref{Eax3}. For~\ref{Eax1}, we compute:
\begin{align*}
\varepsilon^\T(f +^\T g) &=~ \varepsilon^\T(f+g) \\
&=~ \varepsilon(f+g) \\
&=~ \varepsilon(f) + \varepsilon(g) \tag*{\ref{Eax1}} \\
&=~ \varepsilon(f) +^\T \varepsilon(g) \\
&=~ \varepsilon^\T(f) +^\T \varepsilon^\T(g) \\ \\
\varepsilon^\T(0^\T) &=~ \varepsilon(0^\T) \\
&=~ \varepsilon(0) \\
&=~ 0 \tag*{\ref{Eax1}} \\
&=~ 0^\T
\end{align*}
Next for~\ref{Eax2}, we compute:
\begin{align*}
\varepsilon^\T(g \circ^\T f) &=~ \varepsilon(g \circ^\T f) \\
&=~ \varepsilon(\mu \circ \T(g) \circ f) \\
&=~ \varepsilon(\mu) \circ \T(g) \circ f \tag*{\ref{Eax2}} \\
&=~ \mu \circ \varepsilon(1) \circ \T(g) \circ f \tag{Lem.\ref{lem:t-prod}.(vi)}\\
&=~ \mu \circ \varepsilon(\T(g)) \circ f \\
&=~ \mu \circ \T(\varepsilon(g)) \circ f \tag{Lem.\ref{lem:t-prod}.(v)} \\
&=~ \varepsilon(g) \circ^\T f \\
&=~ \varepsilon^\T(g) \circ^\T f
\end{align*}
Lastly for~\ref{Eax3}, recall again that $\pi_i^\T = \eta \circ \pi_i$, so we can compute:
\begin{align*}
\varepsilon^\T(\pi_i^\T) &=~ \varepsilon(\pi_i^\T) \\
&=~ \varepsilon(\eta \circ \pi) \\
&=~ \varepsilon(\T(\pi_i) \circ \eta) \tag{Nat. of $\eta$} \\
&=~ \varepsilon(\T(\pi_i)) \circ \eta \tag*{\ref{Eax2}} \\
&=~ \T(\varepsilon(\pi_i)) \circ \eta \tag{Lem.\ref{lem:t-prod}.(v)} \\
&=~ \T(\pi_i \circ \varepsilon(1)) \circ \eta\tag*{\ref{Eax3}} \\
&=~ \T(\pi_i) \circ \T(\varepsilon(1)) \circ \eta \tag{$\T$ is a functor} \\
&=~ \T(\pi_i) \circ \eta \circ \varepsilon(1) \tag{Nat. of $\eta$} \\
&=~ \T(\pi_i) \circ \varepsilon(\eta) \tag{Lem.\ref{lem:t-prod}.(vi)}\\
&=~ \mu \circ \T(\eta) \circ \T(\pi_i) \circ \varepsilon(\eta) \tag{Monad identities} \\
&=~ \mu \circ \T(\eta \circ \pi_i) \circ \varepsilon(\eta) \tag{$\T$ is a functor} \\
&=~ (\eta \circ \pi_i) \circ^\T \varepsilon(\eta) \\
&=~ (\eta \circ \pi_i) \circ^\T \varepsilon^\T(\eta) \\
&=~ \pi_i^\T \circ^\T \varepsilon^\T(\eta)
\end{align*}
So we conclude that $\varepsilon^\T$ is an infinitesimal extension on $\mathbb{X}_\mathsf{T}$.
\end{proof}
To define the difference combinator for the Kleisli category, first note that difference combinators by definition do not change the codomain. That is, if $f : A \to \T(B)$ is a Kleisli map, then the type of its derivative should be $A \times A \to \T(B)$, which coincides with the type of its derivative in $\mathbb{X}$. Therefore, the difference combinator $\dd^\T$ for the Kleisli category is defined to be the difference combinator of the base category, that is, for a Kleisli map $f= \langle f_0, f_1 \rangle$, its derivative is defined as:
\begin{equation}\label{Kd}\begin{gathered} \dd^\T[f] = \dd[f] = \langle \dd[f_0], \dd[f_1] \rangle
\end{gathered}\end{equation}
where the last equality is due to~\ref{CdCax4}. We note that this difference combinator was the one suggested in~\cite{alvarez2020cartesian} and is the derivative underlying the change action model structure suggested in~\cite{alvarez2019change}. However, as mentioned above, the difference combinator with the infinitesimal extension or change action suggested in those papers does not satisfy~\ref{CdCax2} or~\ref{CADax2}. Luckily, we are able to correct this and still obtain a positive result by carefully proving that the proposed infinitesimal extension and difference combinator in this paper does provide a Cartesian difference structure on the Kleisli category.
\begin{prop} For a Cartesian difference category $\mathbb{X}$, the Kleisli category $\mathbb{X}_\mathsf{T}$ is a Cartesian difference category with infinitesimal extension $\varepsilon^\T$ and difference combinator $\dd^\T$.
\end{prop}
\begin{proof}
We first note that we can easily compute the following:
\begin{align*}
\dd^T[f] \circ^\T \pair{x}{y}^\T &=~ \mu \circ \T(\dd^T[f]) \circ \pair{x}{y}^\T \\
&=~ \mu \circ \T(\dd[f]) \circ \phi^{-1} \circ \pair{x}{y} \\
&=~ \mu \circ \dd[{\T(f)}] \circ \phi \circ \phi^{-1} \circ \pair{x}{y} \tag{by Prop.\ref{T-diff}} \\
&=~ \mu \circ \dd[{\T(f)}] \circ \pair{x}{y}
\end{align*}
So we have that $\dd^T[f] \circ^\T \pair{x}{y}^\T = \mu \circ \dd[{\T(f)}] \circ \pair{x}{y}$.
This will help simplify many of the calculations to follow, since
$\T(\dd[f])$ appears everywhere due to the definition of Kleisli composition. We now prove the Cartesian difference category axioms.
\\\\
\noindent~\ref{CdCax0} $f \circ^\T (x +^\T \varepsilon^\T(y)) = f \circ^\T x +^\T \varepsilon^\T\left( \dd^\T[f] \circ^\T \langle x, y \rangle^\T \right)$ \\\\
First note that $\mu$ is linear and therefore additive. Then we compute that:
\begin{align*}
&f \circ^\T x +^\T \varepsilon^\T\left( \dd^\T[f] \circ^\T \langle x, y \rangle^\T \right) =~ f \circ^\T x + \varepsilon\left( \dd^\T[f] \circ^\T \langle x, y \rangle^\T \right) \\
&=~ \mu \circ \T(f) \circ x + \varepsilon\left( \mu \circ \dd[{\T(f)}] \circ \pair{x}{y} \right) \\
&=~ \mu \circ \T(f) \circ x + \varepsilon\left( \mu \right) \circ \dd[{\T(f)}] \circ \pair{x}{y} \tag*{\ref{Eax2}} \\
&=~ \mu \circ \T(f) \circ x + \mu \circ \varepsilon(1_{\T(A)}) \circ \dd[{\T(f)}] \circ \pair{x}{y} \tag{Lemma~\ref{lem:mu-ep}}\\
&=~ \mu \circ \T(f) \circ x + \mu \circ \varepsilon\left( \dd[{\T(f)}] \circ \pair{x}{y} \right) \tag{Lemma~\ref{lem:epbij}} \\
&=~ \mu \circ \left(\T(f) \circ x + \varepsilon\left( \dd[{\T(f)}] \circ \pair{x}{y} \right) \right) \tag{$\mu$ is additive} \\
&=~\mu \circ \T(f) \circ (x + \varepsilon(y)) \tag*{\ref{CdCax0}} \\
&=~ f \circ^\T (x + \varepsilon(y)) \\
&=~f \circ^\T (x +^\T \varepsilon^\T(y))
\end{align*}
\\
\noindent~\ref{CdCax1} $\dd^\T[f +^\T g] = \dd^\T[f] +^\T \dd^\T[g]$, $[\dd^\T[0^\T] = \dd[0]$, and $\dd^\T[\varepsilon^\T(f)] = \varepsilon^\T\left( \dd^\T[f] \right)$ \\\\
Since both the sum, zero maps, infinitesimal extension, and differential combinator in the Kleisli category are the same as in the base category, by~\ref{CdCax1} it easily follows that:
\begin{align*}
\dd^\T[f +^\T g] = \dd[f+g] = \dd[f] + \dd[g] = \dd^\T[f] +^\T \dd^\T[g]
\end{align*}
\[\dd^\T[0^\T] = \dd[0] = 0 = 0^\T\]
\[\dd^\T[\varepsilon^\T(f)] = \dd[\varepsilon(f)] = \varepsilon\left( \dd[f] \right) = \varepsilon^\T\left( \dd^\T[f] \right) \]
\\
\noindent~\ref{CdCax2} $\dd^\T[f] \circ^\T \langle x, y +^\T z \rangle^\T = \dd^\T[f] \circ^\T \langle x, y \rangle^\T + \dd^\T[f] \circ^\T \langle x + \varepsilon^\T(y), z \rangle^\T$\\
and ${\dd^\T[f] \circ^\T \langle x, 0^\T \rangle^\T = 0^\T}$
\begin{align*}
\dd^\T[f] \circ^\T \pair{x}{y +^\T z}^\T &=~\mu \circ \dd\left[ \T(f) \right] \circ \pair{x}{y +^\T z} \\
&=~\mu \circ \dd\left[ \T(f) \right] \circ \pair{x}{y + z} \\
&=~\mu \circ \left( \dd\left[ \T(f) \right] \circ \pair{x}{y} + \dd\left[ \T(f) \right] \circ \pair{x + \varepsilon(y)}{z} \right) \tag*{\ref{CdCax2}} \\
&=~\mu \circ \dd\left[ \T(f) \right] \circ \pair{x}{y} + \mu \circ \dd\left[ \T(f) \right] \circ \pair{x + \varepsilon(y)}{z} \tag{$\mu$ is additive} \\
&=~ \mu \circ \dd\left[ \T(f) \right] \circ \pair{x}{y} + \mu \circ \dd\left[ \T(f) \right] \circ \pair{x +^\T \varepsilon^\T(y)}{z} \\
&=~ \dd^\T[f] \circ^\T \langle x, y \rangle^\T + \dd^\T[f] \circ^\T \langle x + \varepsilon^\T(y), z \rangle^\T \\\\
\dd^\T[f] \circ^\T \pair{x}{0^\T}^\T &=~ \dd^\T[f] \circ^\T \pair{x}{0}^\T \\
&=~\mu \circ \dd\left[ \T(f) \right] \circ \pair{x}{0} \\
&=~ 0 \tag*{\ref{CdCax2}} \\
&=~ 0^\T
\end{align*}
\\
\noindent~\ref{CdCax3} $\dd^\T[\eta] = \pi_1^\T$ and $\dd^\T[\pi^\T_i] =\pi_i^\T\circ^\T \pi_1^\T$ \\\\
Recall that $\pi_i^\T = \eta \circ \pi_i$ and so $\pi_i^\T\circ^\T \pi_1^\T = \eta \circ \pi_i \circ \pi_1$. Then since $\eta$ is linear, we have that:
\begin{align*}
\dd^\T[\eta] &=~ \dd[\eta] \\
&=~ \eta \circ \pi_1 \tag{$\eta$ is linear} \\
&=~ \pi_1^\T \\\\
\dd^\T[\pi^\T_i] &=~ \dd[\pi^\T_i] \\
&=~ \dd[\eta \circ \pi_i] \\
&=~\eta \circ \dd[\pi_i] \tag{$\eta$ linear and Lem.\ref{chain-linear}} \\
&=~\eta \circ \pi_i \circ \pi_1 \tag*{\ref{CdCax3}}
\end{align*}
\\
\noindent~\ref{CdCax4} $\dd^\T[\langle f, g \rangle^\T] = \langle \dd^\T[f], \dd^\T[g] \rangle^\T$ \\\\
First note that since $\phi$ is linear, so it its inverse $\phi^{-1}$. Then we have that:
\begin{align*}
\dd^\T[\langle f, g \rangle^\T] &=~ \dd[\langle f, g \rangle^\T] \\
&=~ \dd\left[\phi^{-1} \circ \langle f, g \rangle \right] \\
&=~ \phi^{-1} \circ \dd\left[ \langle f, g \rangle \right] \tag{$\phi^{-1}$ linear and Lem.\ref{chain-linear}} \\
&=~ \phi^{-1} \circ \langle \dd[f], \dd[g] \rangle \tag*{\ref{CdCax4}} \\
&=~ \langle \dd[f], \dd[g] \rangle^\T \\
&=~ \langle \dd^\T[f], \dd^\T[g] \rangle^\T
\end{align*}
\\
\noindent~\ref{CdCax5} $\dd^\T[g \circ^\T f] = \dd^\T[g] \circ^\T \langle f \circ^\T \pi_0^\T, \dd^\T[f] \rangle^\T$ \\\\
First note that since $\pi_0^\T = \eta \circ \pi_0$, it easily follows that $f \circ^\T \pi_0^\T = f \circ \pi_0$ (using the monad identities and the naturality of $\eta$). Therefore, we compute that:
\begin{align*}
\dd^\T[g] \circ^\T \langle f \circ^\T \pi_0^\T, \dd^\T[f] \rangle^\T &=~ \mu \circ \dd\left[ \T(g) \right] \circ \langle f \circ^\T \pi_0^\T, \dd^\T[f] \rangle \\
&=~\mu \circ \dd\left[ \T(g) \right] \circ \langle f \circ \pi_0, \dd[f] \rangle \\
&=~ \mu \circ \dd[\T(g) \circ f] \tag*{\ref{CdCax5}} \\
&=~ \dd\left[\mu \circ\T(g) \circ f \right] \tag{$\mu$ linear and Lem.\ref{chain-linear}} \\
&=~\dd[g \circ^\T f] \\
&=~ \dd^\T[g \circ^\T f]
\end{align*}
For the remaining two axioms, we will instead prove~\ref{CdCax6a} and~\ref{CdCax7a}. Before we do so, we first compute the following:
\begin{align*}
&\dd^\T\left[\dd^\T[f] \right] \circ^\T \left \langle \langle x, y \rangle^\T, \langle z,w \rangle^\T \right \rangle^\T =~ \mu \circ \dd\left[ \T(\dd^\T[f]) \right] \circ \left \langle \langle x, y \rangle^\T, \langle z,w \rangle^\T \right \rangle \\
&=~ \mu \circ \dd\left[ \T(\dd[f]) \right] \circ \left \langle \phi^{-1} \circ \langle x, y \rangle, \phi^{-1} \circ \langle z,w \rangle \right \rangle \\
&=~ \mu \circ \dd\left[ \T(\dd[f]) \right] \circ (\phi^{-1} \times \phi^{-1}) \circ \four{x}{y}{z}{w} \\
&=~ \mu \circ \dd\left[ \T(\dd[f]) \circ \phi^{-1} \right]\circ \four{x}{y}{z}{w} \tag{$\phi^{-1}$ linear and Lem.\ref{chain-linear}} \\
&=~ \mu \circ \dd\left[ \dd[\T(f)] \circ \phi \circ \phi^{-1} \right]\circ \four{x}{y}{z}{w} \tag{by Prop.\ref{T-diff}} \\
&=~\mu \circ \dd^2\left[ \T(f) \right] \circ \four{x}{y}{z}{w}
\end{align*}
Therefore we have that $\dd^\T\left[\dd^\T[f] \right] \circ^\T \left \langle \langle x, y \rangle^\T, \langle z,w \rangle^\T \right \rangle^\T = \mu \circ \dd^2\left[ \T(f) \right] \circ \four{x}{y}{z}{w}$. \\\\
\noindent~\ref{CdCax6a} $\dd^\T\left[\dd^\T[f] \right] \circ^\T \left \langle \langle x, 0^\T \rangle^\T, \langle 0^\T, y \rangle^\T \right \rangle^\T = \dd^\T[f] \circ^\T \langle x, y \rangle^\T$
\begin{align*}
\dd^\T\left[\dd^\T[f] \right] \circ^\T \left \langle \langle x, 0^\T \rangle^\T, \langle 0^\T, y \rangle^\T \right \rangle^\T &=~ \mu \circ \dd^2 \left[\T(f) \right] \circ \left \langle \langle x, 0^\T \rangle, \langle 0^\T,y \rangle \right \rangle \\
&=~ \mu \circ \dd^2 \left[\T(f) \right] \circ \left \langle \langle x, 0 \rangle, \langle 0,y \rangle \right \rangle \\
&=~\mu \circ \dd[\T(f)] \circ \pair{x}{y} \tag*{\ref{CdCax6a}} \\
&=~ \dd^\T[f] \circ^\T \langle x, y \rangle^\T
\end{align*}
\\
\noindent~\ref{CdCax7a} $\dd^\T\left[\dd^\T[f] \right] \circ^\T \left \langle \langle x, y \rangle^\T, \langle z, w \rangle^\T \right \rangle^\T= \dd^\T\left[\dd^\T[f] \right] \circ \left \langle \langle x, z \rangle^\T, \langle y, w \rangle^\T \right \rangle^\T$
\begin{align*}
\dd^\T\left[\dd^\T[f] \right] \circ^\T \left \langle \langle x, y \rangle^\T, \langle z, w \rangle^\T \right \rangle^\T &=~\mu \circ \dd^2\left[ \T(f) \right] \circ \four{x}{y}{z}{w} \\
&=~ \mu \circ \dd \left[ \dd[\T(f)] \right] \circ \left \langle \langle x, z \rangle, \langle y,w \rangle \right \rangle \tag*{\ref{CdCax7a}} \\
&=~ \dd^\T\left[\dd^\T[f] \right] \circ \left \langle \langle x, z \rangle^\T, \langle y, w \rangle^\T \right \rangle^\T
\end{align*}
So we conclude that the Kleisli category is a Cartesian difference category.
\end{proof}
We also point that in the case of a Cartesian differential category, since $\varepsilon =0$, it follows that $\varepsilon^T = 0$. Therefore we have that the Kleisli category of the tangent bundle monad of a Cartesian differential category is again a Cartesian differential category. To the knowledge of the authors, this is a novel observation.
\begin{cor} For a Cartesian differential category $\mathbb{X}$, the Kleisli category $\mathbb{X}_\mathsf{T}$ is a Cartesian differential category with differential combinator $\D^\T = \D$.
\end{cor}
We conclude this section by briefly taking a look at the linear maps and the $\varepsilon^\T$-linear maps in the Kleisli category. A Kleisli map $f=\langle f_0, f_1 \rangle$ is linear in the Kleisli category if $\dd^\T[f] = f \circ^\T \pi^\T_1$, which amounts to requiring that:
\[ \langle \dd[f_0], \dd[f_1] \rangle = \langle f_0 \circ \pi_1, f_1 \circ \pi_1 \rangle \]
Therefore a Kleisli map is linear in the Kleisli category if and only if it is the pairing of maps which are linear in the base category. Similarly, a Kleisli map is $\varepsilon^\T$-linear if and only if is the pairing of $\varepsilon$-linear maps.
\section{Difference \texorpdfstring{$\lambda$}{Lambda}-Categories}%
\label{sec:differential-lambda-categories}
Categorical models of the differential $\lambda$-calculus~\cite{ehrhard2003differential} are known as differential $\lambda$-cate\-gories~\cite{bucciarelli2010categorical}. However, a differential $\lambda$-category is not simply a Cartesian differential category which is Cartesian closed. In a differential $\lambda$-category, both the additive structure and the differential structure must be compatible with the curry operator. The same is true of Cartesian difference categories. In this section, we introduce difference $\lambda$-categories. Briefly, a difference $\lambda$-category is a Cartesian difference category which is Cartesian closed and such that the Cartesian difference structure and the curry operations are compatible.
For a Cartesian closed category $\mathbb{X}$, we denote the exponential as $A \Rightarrow B$, the evaluation map as $\A{ev}: (A \Rightarrow B) \times A \to B$, and the curry of a map $f: A \times B \to C$ as $\Lambda(f): A \to (B \Rightarrow C)$, that is, $\Lambda(f)$ is the unique map such that $\A{ev} \circ (\Lambda(f) \times 1_B) = f$. Conversly, given a map of type $g: A \to (B \Rightarrow C)$, define $\Lambda^{-1}(g): A \times B \to C$ as $\Lambda^{-1}(g) = \A{ev} \circ (g \times 1_B)$. Of course, $\Lambda$ and $\Lambda^{-1}$ are inverses of each other, that is, $\Lambda\left( \Lambda^{-1}(g) \right) = g$ and $\Lambda^{-1}\left( \Lambda(f) \right) = f$. Another useful map will be the canonical natural isomorphism $\A{sw}: (A \times B) \times C \to (A \times C) \times B$ which swaps the last arguments, that is, $\A{sw}$ is defined as follows:
\[
\A{sw} := \pair{\pair{\pi_{00}}{\pi_1}}{\pi_{10}}
\]
where recall that $\pi_{ij} = \pi_i \circ \pi_j$. Note that $\A{sw}$ is its own inverse, that is, $\A{sw} \circ \A{sw} = 1$.
\begin{defi} A \textbf{Cartesian closed left additive category} is a Cartesian left additive category $\mathbb{X}$ such that $\mathbb{X}$ is Cartesian closed and such that the curry operator preserves the additive structure, that is, $\Lambda(f + g) = \Lambda(f) + \Lambda(g)$ and $\Lambda(0) = 0$.
\end{defi}
\begin{defi}%
\label{def:difference-lambda-category} A \textbf{difference $\lambda$-category} is a Cartesian difference category $\mathbb{X}$, with difference combinator $\dd$ and infinitesimal extension $\varepsilon$, such that $\mathbb{X}$ is a Cartesian closed left additive category and the following additional axioms hold:
\begin{enumerate}[{\CdlCax{\arabic*}},ref={\CdlCax{\arabic*}},align=left]
\item $\dd[\Lambda(f)] = \Lambda\pa{\dd[f]
\circ \pair{\pa{\pi_0 \times 1_{}}}{\pa{\pi_1 \times 0}}
}
$\label{CdlCax1}
\item $\Lambda(\varepsilon(f)) = \varepsilon\pa{\Lambda(f)}$\label{CdlCax2}
\end{enumerate}
\end{defi}
The first axiom~\ref{CdlCax1} is identical to its differential combinator analogue in a differential $\lambda$-category. As such, it follows the same broad intuition. First note that~\ref{CdlCax1} can also equivalently written in terms of $\A{sw}$ as follows:
\[
\dd[\Lambda(f)] = \Lambda\pa{
\dd[f]
\circ \pa{(1 \times 1) \times \pair{1}{0}}
\circ \A{sw}
}
\]
which will be useful for many calculations in this section. Now given a map
$f : A \times B \to C$, we usually understand the composite:
\[\dd[f] \circ (1_{A \times B} \times \pair{1_A}{0}) \circ \A{sw} : (A \times A) \times B \to C\]
as the partial derivative of $f$ with respect to its first argument $A$. Hence,~\ref{CdlCax1} states that the
derivative of a curried function is precisely the curry of the partial derivative of the function with respect to its first argument. On the other hand,~\ref{CdlCax2} simply says that the curry of an infinitesimal extension of a function is the infinitesimal extension of the curry of the function. Using $\lambda$-calculus notation, this implies that $\lambda x . \varepsilon(f(t,y)) = \varepsilon (\lambda x . f(t,y))$.
\begin{exa}
Every differential $\lambda$-category~\cite{bucciarelli2010categorical} is precisely a difference $\lambda$-category such that $\varepsilon = 0$.
\end{exa}
\begin{exa}
The category $\Ab$ (as defined in Section~\ref{discreteex}) is a difference $\lambda$-category. Given abelian groups $G, H$, the exponential $G \Rightarrow H$ is defined as the abelian group of all functions between $G$ and $H$, where the group structure is defined point-wise. The remaining Cartesian closed structure is defined in the standard way, that is, for an arbitrary function $f: G \times H \to K$, $\Lambda(f)(x)(y) = f(x,y)$. We leave it to the reader to check for themselves that $\Ab$ is indeed a Cartesian closed left additive category. Since $\varepsilon =1$, clearly~\ref{CdlCax2} holds automatically. So it remains to verify that the other axiom~\ref{CdlCax1} also holds:
\begin{align*}
\dd[\Lambda(f)](x, u)(y) &=~ \Lambda(f)(x + u)(y) - \Lambda(f)(x)(y)
\\
&=~ f(x + u, y) - f(x, y) \\
&=~ f\left( (x,y) + (u,0) \right) - f(x,y) \\
&=~\dd[f](x,y, u, 0) \\
&=~ (\dd[f] \circ \left( (1 \times 1) \times \pair{1_{}}{0} \right) \circ \A{sw})(x,u,y) \\
&=~ \Lambda(\dd[f] \circ \left( (1 \times 1) \times \pair{1_{}}{0} \right) \circ \A{sw})(x, u)(y)
\end{align*}
Therefore, $\Ab$ is indeed a difference $\lambda$-category.
\end{exa}
A central property of differential $\lambda$-categories is a deep correspondence
between differentiation and the evaluation map. As one would expect, the partial
derivative of the evaluation map gives a first-class derivative operator, see for example~\cite[Lemma~4.5]{bucciarelli2010categorical}, which provides
an interpretation for the differential substitution operator in the differential
$\lambda$-calculus. This property
still holds in difference $\lambda$-categories, although its formulation is somewhat more
involved.
\begin{lem}%
\label{lem:lambda-d-ev} In a difference $\lambda$-category, for maps $g: A \times B \to C$ and $f: A \to B$, the following identities hold:
\begin{enumerate}[(\roman{enumi}),ref={\thelem.\roman{enumi}}]
\item $
\dd[\A{ev} \circ \pair{\Lambda(g)}{f}]
= \A{ev} \circ \pair{\dd[\Lambda(g)]}{f \circ \pi_0}
+\dd[g] \circ \pair{\pair{\pi_0 + \varepsilon(\pi_1)}{f \circ
\pi_0}}{\pair{0}{\dd[f]}}$
\item $
\dd[\A{ev} \circ \pair{\Lambda(g)}{f}]
=
\A{ev} \circ \pair{\dd[\Lambda(g)]}{f \circ \pi_0 + \varepsilon(\dd[f])}
+\dd[g] \circ \pair{\pair{\pi_0}{f \circ \pi_0}}{\pair{0}{\dd[f]}}
$
\end{enumerate}
\end{lem}
\begin{proof} For (i), we compute:
\begin{align*}
&\dd[\A{ev} \circ \pair{\Lambda(g)}{f}] =~ \dd[g \circ \pair{1_A}{f}] \\
&=~ \dd[g] \circ \pair{\pair{1_A}{f} \circ \pi_0}{\dd[\pair{1_A}{f}]} \tag*{\ref{CdCax5}} \\
&=~ \dd[g] \circ \four{\pi_0}{f \circ \pi_0}{\pi_1}{\dd[f]} \tag*{\ref{CdCax3}+\ref{CdCax4}} \\
&=~ \dd[g] \circ \pair{\pair{\pi_0}{f \circ \pi_0}}{ \pair{\pi_1}{0} + \pair{0}{\dd[f]} } \\
&=~\dd[g] \circ \four{\pi_0}{f \circ \pi_0}{\pi_1}{0} + \dd[g] \circ \pair{\pair{\pi_0}{f \circ \pi_0} + \varepsilon(\pair{\pi_1}{0} )}{\pair{0}{\dd[f]} } \tag*{\ref{CdCax2}} \\
&=~\dd[g] \circ \four{\pi_0}{f \circ \pi_0}{\pi_1}{0} + \dd[g] \circ \pair{\pair{\pi_0}{f \circ \pi_0} + \pair{ \varepsilon(\pi_1)}{0}}{\pair{0}{\dd[f]} } \tag{Lem~\ref{lem:ep-pair} +~\ref{Eax1}} \\
&=~\dd[g] \circ \four{\pi_0}{f \circ \pi_0}{\pi_1}{0} + \dd[g] \circ \four{\pi_0 + \varepsilon(\pi_1)}{f \circ \pi_0}{0}{\dd[f]} \\
&=~\dd[g] \circ \pair{\pi_0 \times 1_B}{\pi_1 \times 0} \circ \pair{1}{f \circ \pi_0} + \dd[g] \circ \four{\pi_0 + \varepsilon(\pi_1)}{f \circ \pi_0}{0}{\dd[f]} \\
&=~ \Lambda^{-1}\left( \Lambda\left( \dd[g] \circ \pair{\pi_0 \times 1_B}{\pi_1 \times 0} \right) \right) \circ \pair{1}{f \circ \pi_0} + \dd[g] \circ \four{\pi_0 + \varepsilon(\pi_1)}{f \circ \pi_0}{0}{\dd[f]} \\
&=~ \Lambda^{-1}\left( \dd[\Lambda[g]]\right) \circ \pair{1}{f \circ \pi_0} + \dd[g] \circ \four{\pi_0 + \varepsilon(\pi_1)}{f \circ \pi_0}{0}{\dd[f]} \tag{by~\ref{CdlCax1}} \\
&=~ \A{ev} \circ ( \dd[\Lambda[g]] \times 1_B) \circ \pair{1}{f \circ \pi_0} + \dd[g] \circ \four{\pi_0 + \varepsilon(\pi_1)}{f \circ \pi_0}{0}{\dd[f]} \\
&=~ \A{ev} \circ \pair{ \dd[\Lambda[g]]}{f \circ \pi_0} + \dd[g] \circ \four{\pi_0 + \varepsilon(\pi_1)}{f \circ \pi_0}{0}{\dd[f]}
\end{align*}
On the other hand, for (ii), we compute:
\begin{align*}
& \dd[\A{ev} \circ \pair{\Lambda(g)}{f}] =~ \dd[g \circ \pair{1_A}{f}] \\
&=~\dd[g] \circ \pair{\pair{1_A}{f} \circ \pi_0}{\dd[\pair{1_A}{f}]} \tag*{\ref{CdCax5}} \\
&=~ \dd[g] \circ \four{\pi_0}{f \circ \pi_0}{\pi_1}{\dd[f]} \tag*{\ref{CdCax3}+\ref{CdCax4}} \\
&=~ \dd[g] \circ \pair{\pair{\pi_0}{f \circ \pi_0}}{\pair{0}{\dd[f]} + \pair{\pi_1}{0} } \\
&=~\dd[g] \circ \four{\pi_0}{f \circ \pi_0}{0}{\dd[f]} + \dd[g] \circ \pair{\pair{\pi_0}{f \circ \pi_0} + \varepsilon(\pair{0}{\dd[f]})}{\pair{\pi_1}{0} } \tag*{\ref{CdCax2}} \\
&=~\dd[g] \circ \four{\pi_0}{f \circ \pi_0}{0}{\dd[f]} + \dd[g] \circ \pair{\pair{\pi_0}{f \circ \pi_0} + \pair{0}{\varepsilon(\dd[f])} }{\pair{\pi_1}{0} } \tag{Lem~\ref{lem:ep-pair} +~\ref{Eax1}} \\
&=~\dd[g] \circ \four{\pi_0}{f \circ \pi_0}{0}{\dd[f]} + \dd[g] \circ \four{\pi_0}{f \circ \pi_0 + \varepsilon(\dd[f])}{\pi_1}{0} \\
&=~\dd[g] \circ \four{\pi_0}{f \circ \pi_0}{0}{\dd[f]} + \dd[g] \circ \pair{\pi_0 \times 1_B}{\pi_1 \times 0} \circ \pair{1}{f \circ \pi_0 + \varepsilon(\dd[f])}\\
&=~\dd[g] \circ \four{\pi_0}{f \circ \pi_0}{0}{\dd[f]} + \Lambda^{-1}\left( \Lambda\left( \dd[g] \circ \pair{\pi_0 \times 1_B}{\pi_1 \times 0} \right) \right) \circ \pair{1}{f \circ \pi_0 + \varepsilon(\dd[f])}\\
&=~\dd[g] \circ \four{\pi_0}{f \circ \pi_0}{0}{\dd[f]} + \Lambda^{-1}\left( \dd[\Lambda[g]]\right) \circ \pair{1}{f \circ \pi_0 + \varepsilon(\dd[f])} \tag{by~\ref{CdlCax1}} \\
&=~\dd[g] \circ \four{\pi_0}{f \circ \pi_0}{0}{\dd[f]} + \A{ev} \circ ( \dd[\Lambda[g]] \times 1_B) \circ \pair{1}{f \circ \pi_0 + \varepsilon(\dd[f])}\\
&=~ \dd[g] \circ \four{\pi_0}{f \circ \pi_0}{0}{\dd[f]} + \A{ev} \circ \pair{\dd[\Lambda[g]]}{f \circ \pi_0 + \varepsilon(\dd[f])}\\
&=~ \A{ev} \circ \pair{\dd[\Lambda[g]]}{f \circ \pi_0 + \varepsilon(\dd[f])} + \dd[g] \circ \four{\pi_0}{f \circ \pi_0}{0}{\dd[f]}
\end{align*}
Thus the desired identities hold. \end{proof}
\section{Conclusions and Future Work}
We have presented Cartesian difference categories, which generalize Cartesian differential categories to account for more discrete definitions of derivatives while providing additional structure that is absent in change action models. We have also exhibited important examples and shown that Cartesian difference categories arise quite naturally from considering tangent bundles in any Cartesian differential category. We claim that Cartesian difference categories can facilitate the exploration of differentiation in discrete spaces, by generalizing techniques and ideas from the study of their differential counterparts. For example, Cartesian differential categories can be extended to allow objects whose tangent space is not necessarily isomorphic to the object itself~\cite{cruttwell2017cartesian}. The same generalization could be applied to Cartesian difference categories --- with some caveats: for example, the equation defining a linear map (Definition~\ref{def:linearity}) becomes ill-typed, but the notion of $\varepsilon$-linear map remains meaningful.
Perhaps the most important topic for further research is identifying and cataloguing more instances of Cartesian difference
categories. There is a number of very natural candidates for such; among those, we would like to single out synthetic differential
geometry~\cite{kock2006synthetic} and non-standard analysis~\cite{keisler2013elementary}, both of which feature a notion
of ``infinitesimal elements''. In either case, multiplying by a suitably chosen infinitesimal $\varepsilon$ would be an obvious
choice of an infinitesimal extension. As suggested by one of the anonymous reviewers, the case of non-standard analysis based on hyperreal numbers might be of particular interest:
since they contain invertible infinitesimals (as opposed to the nilpotent infinitesimals from SDG), one could define the derivative
of a function $f$ as $\dd[f](x, y) = \varepsilon^{-1}\left( f(x + \varepsilon y) - f(x)\right)$ (assuming a choice of an
``infinitesimal unit'' $\varepsilon$). With an infinitesimal extension being defined by multiplication by the above infinitesimal
unit $\varepsilon$, most of the axioms of a Cartesian difference category can be proven by a suitable generalisation of the proofs
for the calculus of finite differences (Section~\ref{discreteex}), although more work would be needed to prove \CdCax{7}. Lie
groups, and Lie algebras, have also been suggested by one of the anonymous reviewers as a potential source of models, but work in this direction is at a less
mature stage.
Another relevant path to consider is developing the analogue of the ``tensor'' story for Cartesian difference categories. Indeed, an important source of examples of Cartesian differential categories are the coKleisli categories of a tensor differential category~\cite{blute2006differential,blute2009cartesian}. A similar result likely holds for a hypothetical ``tensor difference category'', but it is not immediately clear how these should be defined: \CdCax{2} implies that derivatives in the difference sense are non-linear and therefore their interplay with the tensor structure will be much different. As suggested by one of the anonymous reviewers, a potential leading example of a ``tensor difference category'' is $\mathsf{Ab}$, the category of abelian groups and group homomorphisms between them, which is a model of linear logic via a linear-non-linear adjunction between $\mathsf{Ab}$ and $\mathsf{SET}$, the category of sets and arbitrary set functions between them~\cite{benton1994mixed}. The induced comonad (which models the exponential modality of linear logic) on $\mathsf{Ab}$ is given by mapping an abelian group $G$ to the free abelian group over the underlying set of $G$. Even more interesting is the fact that the coKleisli category of this comonad is isomorphic to $\overline{\mathsf{Ab}}$, as defined in Section~\ref{discreteex}, which is one of our main examples of a Cartesian difference category. As such, this is a pretty convincing argument that $\mathsf{Ab}$ will indeed be a prime candidate for an example of a ``tensor difference category'' and help guide us in constructing the correct definition.
A further generalization of Cartesian differential categories, categories with tangent structure~\cite{cockett2014differential} are defined directly in terms of a tangent bundle functor rather than requiring that every tangent bundle be trivial (that is, in a tangent category it may not be the case that $\T A = A \times A$). Some preliminary research on change actions has already shown that, when generalized in this way, change actions are precisely internal categories, but the consequences of this for change action models (and, \emph{a fortiori}, Cartesian difference categories) are not understood. More recently, some work has emerged about differential equations using the language of tangent categories~\cite{cockett2017connections}. We believe similar techniques can be applied in a straightforward way to Cartesian difference categories, where they might be of use to give an abstract formalization of discrete dynamical systems and difference equations.
\bibliographystyle{alpha}
|
\section{Introduction}\label{intro}
How magnetism fits into the picture of galaxy evolution over cosmic time is a subject of debate among astronomers. The link between magnetic fields and galaxy evolution is theorized \citep[e.g.][]{Lou1998} but
observational constraints have been lacking, due to technical limitations
\citep[e.g.][]{Chamandy2016}.
Understanding the evolution of magnetic
fields and their relation to galaxy evolution is one of the key science objectives of next generation radio telescopes such as the Square Kilometre Array
\citep{Johnston-hollitt2015,Heald2020}. The polarization of radio emission is a powerful probe of cosmic magnetic fields.
\par
Some observational studies in the radio \citep{Eichendorf1980,Beck2002,Beck2004,Beck2012d,Farnes2014} have provided insights into the workings of magnetism in galaxies.
Highly sensitive and resolved imaging observations of nearby disk galaxies \citep{Beck2005}, show large scale magnetic fields related to spiral structure.
For more distant galaxies that remain unresolved by current radio imaging telescopes
we must rely on the frequency dependence of polarization, instead of direct imaging, as our primary probe of
magnetic properties \citep{Stil2009a}.
However, the relatively narrow instantaneous bandwidths typical of most past studies of
radio polarization limits both the resolution of Faraday spectral features and the ability
to characterize the frequency dependent effects of internal depolarization.
\par
Recent studies have begun to probe polarized radio emission over broader bands. For example, \citet{OSullivan2012} presented a broad-band spectropolarimetric study of four
radio bright and highly polarized AGN sources, for which Faraday rotation features could be resolved.
\cite{Anderson2015} report a wide band study of polarimetry in more than 500 sources,
\cite{Anderson2016} observed 36 strongly polarized
sources over a very wide band from $\sim 1 - 10$ GHz, and find Faraday complexity in most of their sources, asserting that complexity is thus common among strong polarized sources. \black{A more recent study, \citet{Schnitzeler2019}, provides a wide-band (1.3 - 3.1 GHz) polarimetry survey of unresolved (angular resolution $\sim 2^{\prime}\times1^{\prime}$) southern sources. The authors argue in favour of the Galactic foreground being a dominant component in terms of Faraday rotation of background sources, implying simple rather than complex intrinsic distributions of Faraday rotation.}
\par
In this paper we extend the studies of broad-band polarimetry of extragalactic radio sources
by studying a sample of six radio sources, including two disk galaxies and four AGN
dominated sources.
The study was carried out with observations from the Karoo Array Telescope (KAT\ 7), which was
constructed as an engineering prototype array at the South African SKA site as part of the
technical development of the MeerKAT SKA precursor telescope. \black{The 64 dish MeerKAT array will form part of the SKA mid-frequency component.}
The KAT 7 wavelength coverage from 1.2 to 1.9 GHz provides us with a \black{good resolution of the Faraday spectrum of polarized emission (see Section \ref{subsec:RMsynth} )}.
The rotation measure synthesis technique (RM Synthesis), first described by \citet{Burn1966}, is employed to explore the intrinsic polarization properties of the sources.
\par
The spatial resolution limitations of KAT\,7 at low redshift are comparable to those of the more sensitive MeerKAT (and the future full SKA array) at high redshift. MeerKAT has the sensitivity to detect polarization in faint distant objects but does not resolve them, so we explore the peak flux properties of nearby objects to probe the spectropolarimetric behaviour in a fashion that can be extended to more distant and fainter objects with much more powerful instruments. This article is structured as follows: section \ref{Observations+Reductions} discusses our observations and data reductions; section \ref{PolarimetrySect} presents the results of our polarimetric analysis; \black{we discuss our findings in section \ref{Discussion}}; and we summarise, discuss and also make conclusions based on this work in section \ref{Summary+conclusions}.
\section{Observations and Data Reduction}\label{Observations+Reductions}
\subsection{Sample Selection}
We selected our sample of disk galaxies from the source list of the MeerKAT HI
Observations of Nearby Galactic Objects: Observing Southern Emitters survey\footnote{PI: Erwin de Blok, \url{http://mhongoose.astron.nl/}}, ``MHONGOOSE" \citep{MHONGOOSE}. The survey will secure very high sensitivity atomic hydrogen images of a sample of nearby galaxies. A further criterion was that our sources be sufficiently radio bright, with a reported integrated total intensity \black{greater than 200 mJy at L band. The flux limit was chosen so as to have sources with sufficient flux density to be detectable by KAT\,7 at a few percent polarization}. This resulted in a sample of two disk galaxies. We also observed four AGN dominated sources, with three of them serving as calibrators for the remaining targets. \black{The two disk galaxies both have unique features in the context of the overall disk galaxy population, but the analysis done here will shed some light on the polarization resulting from the large scale magnetic field in disk galaxies that are unresolved.}
\subsubsection*{Disk Galaxies} \label{subsubsec:disk gals}
{\bf NGC 1097} is identified as a Sy1 galaxy \citep{StorchiBergmann1997} with a circumnuclear star-burst ring \citep{Gerin1988}. It is a barred spiral, SBb, with its radio emission dominated by the circumnuclear ring and nucleus \citep{Beck2005}. The magnetic properties of NGC 1097 have been studied using narrow band ($\sim 50$ MHz) high resolution radio data ($2^{\prime\prime} - 15^{\prime\prime}$ resolutions) at 4.8 and 8.4 GHz, with the NRAO Very Large Array, VLA \citep{Beck2002,Beck2005}. The galactic magnetic field was observed to control the ISM gas flow at kpc scales and found to have a regular component that follows the shape of the spiral arms \citep{Beck2005}. The galaxy has a redshift of 0.0042.
\par
{\bf NGC 1808} is identified as a Sy2 galaxy \citep{Tacconi-Garman1996,Galliano2008} with a low luminosity AGN \citep{Veron-Cetty1985,Dahlem1990,Gonzalez-Martin2013,Esquej2014}. It was classified as a peculiar Sbc galaxy \citep{Veron-Cetty1985}. It also has a star-burst region in the inner $\sim 750$ pc dominating the infra-red emission \citep{Sengupta2006}. From observations at 6 and 20\,cm wavelengths with the VLA, \citet{Dahlem1990} found linear polarization degrees of up to $30\%$ and a steep ($\alpha=-0.88$) radio spectrum from outside the central region of the galaxy. The galaxy has a redshift of 0.0033.
\subsubsection*{AGN}
{\bf J0240$-$231} has been identified to be a QSO and also a Gigahertz-peaked spectrum (GPS) source \citep{Kuehr1981,ODea1998,Veron-Cetty2006}. It has a redshift of z = 2.22 \citep{Hewitt1989}. The source has been observed to show significant linear polarization at 4.8 GHz where $p = 3.47 \%$ \citep{Edwards2004}. VLBI studies show structure resembling two lobes that are separated by 12 mas \citep{Dallacasa1998}.
\par
{\bf J0538$-$440} is reported to be a BL Lac, GPS object and also a variable $\gamma$-ray source located at redshift 0.896 \citep{Impey1988,Romero2000,Tornikoski2001,Romero2002,Andruchow2005,Torniainen2005}. It has also been identified as a potential gravitational lens \citep{Surpi1996}. \black{This source presents a time variable total intensity spectrum, at cm and mm wavelengths, as can be seen in calibrator databases such as the ATCA Calibrator Database\footnote{\url{https://www.narrabri.atnf.csiro.au/calibrators/calibrator_database_viewcal.html?source=0537-441} }. It has also been observed to be highly polarized and variable in the optical \citep{Impey1988,Romero2000,Romero2002,Andruchow2005}.}
\par
\textbf{PKS B$1934-638$} is a quasar and GPS source at redshift 0.18 while \textbf{PKS B$0407-658$} is a quasar and possible compact steep-spectrum (CSS) source \citep{ODea1991,Labiano2006} at a redshift of 0.96. VLBI imaging of PKS B$1934-638$ shows that it has a double lobed structure \citep{Tzioumis1989}. \black{Henceforth, we use the name "PKS $1934-638$" for PKS B$1934-638$ in this work as this name is most commonly used for the source.}
\begin{table*}
\caption{Summary of the KAT\,7 observations. The table lists the radio sources (some aliases in brackets) observed during each run, their J2000 coordinates, the observation dates (and times given in UTC), central frequency ($\nu_{c}$), number of antennas available ($N_{\rm ant}$), and the total recovered bandwidth ($\Delta \nu$).}
\begin{tabular}{lllcccc}
\hline\hline\\
[-1.25ex]
Source (aliases) & RA (J2000) & DEC (J2000) & Date (UTC time) & $\nu_{c}$ & $N_{\rm ant}$ & $\Delta \nu$ \\
& HH:MM:SS & DD:MM:SS &&(GHz)&& (MHz) \\
[1.25ex]
\hline
NGC 1808 & 05:07:42.3000 & -37:30:47.0000 & 29/07/2016 (01:07:20.1 - 07:12:30.4) & 1.894 & 6 & 152 \\
PKS $1934-638$ (J1939-6342) & 19:39:25.0267 & -63:42:45.6255 & & & & \\
J$0538-440$ (J0538-4405, 0537-441) & 05:38:50.1800 & -44:05:10.3000 & & & & \\
PKS B$0407-658$ (J0408-6545) & 04:08:20.3788 &-65:45:09.0806 & & & & \\
-- & -- & -- & -- & -- & -- & --\\
&&&&&&\\
NGC 1808 &&& 31/07/2016 (01:00:19.5 - 07:03:29.8) & 1.394 & 6 & 148 \\
PKS $1934-638$ &&&& & & \\
J$0538-440$ &&&& & & \\
PKS B$0407-658$ &&&& & & \\
-- & -- & -- & -- & -- & -- & --\\
&&&&&&\\
NGC 1097 &02:46:19.0000 & -30:16:30.0000 & 09/06/2015 (02:32:11.9 - 08:34:42.2) & 1.850 & 7 & 223 \\
PKS $1934-638$ &&&& & & \\
J$0240-231$ (J0240-2309, 0237-233) & 02:40:08.1400 & -23:09:16.0000 && & & \\
PKS B$0407-658$ &&&& & & \\
&&&&&&\\
-- & -- & -- & -- & -- & -- & --\\
NGC 1097 &&& 15/06/2015 (02:30:22.9 - 08:35:13.2) & 1.350 & 7 & 211 \\
PKS $1934-638$ &&&& & & \\
J$0240-231$ &&&& & & \\
PKS B$0407-658$ &&&& & & \\
\hline
\end{tabular}
\label{Obstable1}
\end{table*}
\subsection{Observations}
KAT\,7 is a synthesis array telescope located adjacent to the
MeerKAT and SKA phase 1 central site in the Karoo plateau of South Africa.
It was built as an engineering test-bed preceding the 64 antenna SKA mid-frequency precursor array, MeerKAT. KAT\,7 consists of seven 12-metre, centre-fed parabolic antennas in a fixed configuration with baselines ranging from 26 to 185 metres \citep{Carignan2013}.
It has a maximum instantaneous bandwidth of 256 MHz within the RF range 1200 - 1950 MHz.
Two components (X and Y) of linear polarization are detected with prime-focus feeds on each antenna.
We observed in wide-band, full-polarization mode, providing raw data consisting of four polarization correlation products sampling a
256 MHz band in 1024 spectral channels, each of width 390.625 kHz.
Our observations covered the entire 750\,MHz RF band using three separate observing runs for each disk galaxy target with the 256 MHz centred near 1350, 1600, and 1850 MHz (see Table \ref{Obstable1}).
\par
Each observation consisted of 12 hour tracks during which on-source observations of 4 minutes of either NGC\,1097 or NGC\,1808 were alternated with 2 minute scans of a phase calibrator, either J$0240-231$ (for NGC\,1097) or J$0538-440$ (for NGC\,1808). Flux and bandpass were calibrated with several 3 minute scans on PKS $1934-638$ and PKS B$0407-658$ depending on which calibrator was visible. The cadence on these flux calibrator scans was approximately an hour. The 12 hour tracks provided a good range of parallactic angle variation for polarization calibration.
\subsection{Data reduction} \label{Calibration and Imaging}
Processing of the raw visibility data was carried out using the Common Astronomy Software Applications, {\sc casa} \citep{Mcmullin2007} package. After removal of the channels at the upper and lower ends of the band due to roll-off of the bandpass shape, the visibility data were inspected visually and manually flagged to remove radio frequency interference (RFI) signals. The mid-band, centred at 1600 MHz, was heavily populated with RFI. Analysis in this band was severely compromised as a result and we did not include the mid-band in our subsequent analysis.
The objects observed, recovered band pass, and number of operating antennas for each observing run are listed in Table~\ref{Obstable1}.
\par
PKS $1934-638$ was used for absolute flux calibration, and to measure the complex bandpass shape and
cross-hand delay for all observations. The sources J$0538-440$ and J$0240-231$ were used as secondary calibrators to track time-dependent amplitude and phase calibration over the course of each observation and for on-axis polarization leakage calibration given the large parallactic angle covered.
The time-dependent total intensity gain is coupled to any detectable
polarization signal from the astronomical sources and telescope. The broad parallactic angle coverage of the observations of the secondary calibrators allows us to solve for the polarization of the calibrators from the initial antenna based time-dependent gain solutions as well as the frequency-dependent XY-phase calibration. The source polarization solutions are then used to correct the time-dependent gains and to solve for frequency-dependent polarization leakage. {\sc casa} polarization calibration tasks do not account for rotation measure across a wide band and thus the polarization solution from the gains is the average across the band. This is not ideal for a non-zero rotation measure. This effect is not mitigated during data reduction in this work and thus may affect polarization measurements. We assume that the effect will not be significant at the bandwidths of our data, $\sim 200$ MHz, and that none of the targets have large RM values. For example, \cite{Taylor2009} report an RM of $14.5\pm 2.2$ rad\,m$^{-2}$ with polarization degree of $0.8 +/- 0.1\%$ at 1.4 GHz for J$0240-231$.
\par
We do not observe a source with known absolute polarization position angle. Therefore, while the relative polarization position angles (with respect to frequency) are calibrated, the absolute polarization position angles for the polarization solutions are not known. The instrumental polarization is accounted for by solving for leakages using a source with a large parallactic angle coverage \citep[e.g.][]{Jagannathan2017}.
\par
After applying the calibration solutions, Stokes $I$, $Q$, and $U$ image cubes were made for each source. The imaging and deconvolution was done using the Clarke-Stokes algorithm implemented in {\sc casa}'s {\sc tclean} task. \black{We averaged the visibility channels into 3.9 MHz channels to construct the image cubes which resulted in the maximum detectable Faraday depth decreasing by an order of magnitude (see Section \ref{subsec:RMsynth} and Table \ref{RMresolution}).} The typical synthesized beam at the low band is $\sim 4^{\prime}.9$ and $\sim 3^{\prime}.6$ at the high band. The observed angular size of our sources is consistent with unresolved sources.
\begin{table*}
\caption{Band averaged values of Stokes \myavg{I}$_{\nu_c}$, band averaged linear polarization intensities (\myavg{P}$_{\nu_c}$ and bias corrected \myavg{P_o}$_{\nu_c}$), and also Q, U noise ($\sigma_{QU}$). The error in the averaged values are $\sigma/\sqrt{N}$, where $N$ is the number of channels, $\sigma$ is the standard deviation of the values and $\nu_c$ is the frequency associated with the band. The low and high bands are indicated by sub-/superscripts $1.35$ and $1.85$, respectively indicating $\nu_c$ in GHz. All values of intensity have units of mJy per beam, abbreviated as mJy/bm. The last column displays the mean spectral index, -\myavg{\alpha}$_{1.85}^{1.35}$, over the two KAT 7 bands}
\begin{tabular}{lrrrcrrrcc}
\hline \hline\\
[-1.25ex]
Source & \myavg{I}$_{1.35}$\ \ \ \hfill & \myavg{P}$_{1.35}$\ \ \hfill & \myavg{P_o}$_{1.35}$\ \hfil & \myavg{\sigma}$_{QU}^{1.35}$\hfill & \myavg{I}$_{1.85}$\ \ \ \hfill & \myavg{P}$_{1.85}$\ \ \hfill & \myavg{P_o}$_{1.85}$\ \hfill & \myavg{\sigma}$_{QU}^{1.85}$ & $\alpha_{1.85}^{1.35}$\\
& ($\mathrm{\frac{mJy}{bm}}$)\ \ \ \ & ($\mathrm{\frac{mJy}{bm}}$)\ \ \ & ($\mathrm{\frac{mJy}{bm}}$)\ \ \ & ($\mathrm{\frac{mJy}{bm}}$) & ($\mathrm{\frac{mJy}{bm}}$)\ \ \ \ & ($\mathrm{\frac{mJy}{bm}}$)\ \ \ & ($\mathrm{\frac{mJy}{bm}}$)\ \ \ & ($\mathrm{\frac{mJy}{bm}}$)& \\
[1.25ex]
\hline
NGC 1808 & $514.6 \pm 0.2$ & $2.4\pm 0.5$ & $2.4\pm 0.5$ & $0.3$& $400.3 \pm 0.2$ & $2.2\pm 0.4$ & $2.2\pm 0.4$ &$0.3$ &$-0.82\pm0.01$\\
NGC 1097 & $340.8 \pm 0.3$ & $2.6\pm 0.4$ & $2.6\pm 0.4$ & $0.4$& $248.0 \pm 0.1 $& $3.6\pm 0.3$ & $3.5\pm 0.3$ &$0.2$ &$-1.06\pm0.02$\\
J$0240-231$ & $5727.4\pm1.2$ & $58.9\pm 3.6$& $58.9\pm 3.6$& $0.6$& $4861.1\pm 0.8$ &$101.5\pm 3.1$&$101.5\pm3.1$ &$0.5$ &$-0.56\pm0.01$\\
J$0538-440$ & $3923.0\pm 0.3$& $27.6\pm 2.8$& $27.6\pm 2.8$& $0.6$ &$3806.5\pm 1.3$ & $10.9\pm 2.5$& $10.9\pm2.5$ &$0.4$ &$-0.10\pm0.01$\\
PKS $1934-638$& $14876.9\pm 0.5$& -- & -- & $1.5$ &$13315.6\pm 0.3$& -- & -- &$1.0$ &$-0.36\pm0.01$\\
PKS B$0407-658$&$15130.7\pm 1.0$& $18.7\pm 9.6$& $18.7\pm 9.6$&$1.5$ & $10848.2\pm 1.0$&$21.7\pm 6.7$ &$21.7\pm 6.7$ &$1.3$ &$-1.12\pm0.01$\\
\hline
\end{tabular}
\label{StokesI+pTable}
\end{table*}
\begin{figure*}
\includegraphics[width=\textwidth]{Iplot-All-freq-v-I-allbands-NED+KAT7band-v7.3.png}
\caption{Total intensity across KAT-7 band (blue points), some corresponding literature values (from NED, \url{https://ned.ipac.caltech.edu/}) at frequencies near the KAT-7 band. \textbf{Top panels}: \black{Left: The primary flux calibrator, PKS $1934-638$, with the Reynolds 1994 model plotted as a red curve \citep{Reynolds1994}}. Middle: NGC 1808. Right: NGC 1097. \textbf{Bottom panels}: Left: The polarization calibrators, J$0538-440$ and J$0240-231$ (middle). Right: PKS B$0407-658$.}
\label{Tot_flux_plots}
\end{figure*}
\begin{figure*}
\begin{subfigure}{1.0\hsize
\includegraphics[width=\textwidth]{NGC1097-pqu-model+data_fits-rmclean-v4.pdf}\\
\includegraphics[width=\textwidth]{0240-231-pqu-model+data_fits-rmclean-v4.pdf}
\subcaption{\textbf{Top panels}: $q(\lambda^2)$ and $u(\lambda^2)$, $p(\lambda^2)$, and $q$ vs. $u$ for NGC 1097. \textbf{Bottom panels}: Same measurements for J$0240-231$. See main caption below.}
\label{pquX_fullband-ngc1097}
\end{subfigure}
\begin{subfigure}{1.0\hsize
\includegraphics[width=\textwidth]{NGC1808-pqu-model+data_fits-rmclean-v4.pdf}\\
\includegraphics[width=\textwidth]{0538-440-pqu-model+data_fits-rmclean-v4.pdf}
\subcaption{Same as in Figure \ref{pquX_fullband-ngc1097}, \textbf{top panels}: Measurements for NGC 1808. \textbf{Bottom panels}: Same measurements for J$0538-440$. See main caption below.}
\label{pquX_fullband-ngc1808}
\end{subfigure}
\end{figure*}
\begin{figure*}\ContinuedFloat
\begin{subfigure}{1.0\hsize
\includegraphics[width=\textwidth]{pks1934-pqu-model+data_fits-rmclean-v4.pdf}\\
\includegraphics[width=\textwidth]{PKSB0407-pqu-model+data_fits-rmclean-v4.pdf}
\subcaption{Same as in Figure \ref{pquX_fullband-ngc1097}, \textbf{top panels}: PKS $1934-638$. \textbf{Bottom panels}: Same measurements for PKS B$0407-658$. See main caption below.}
\label{pquX_fullband-fluxcals}
\end{subfigure}
\caption{Figures \ref{pquX_fullband-ngc1097} - \ref{pquX_fullband-fluxcals}: Polarization measurements, fractional Stokes $Q$ and $U$ \black{(with errors, error bars omitted, of the order $\sim 10^{-5}$)} and also total linear polarization $p_0$, across the KAT-7 band along with the RM ``cleaned" versions of each quantity -- in each panel, shown as curves in the leading two plots and black points in the trailing plot.}
\end{figure*}
\section{Results}
\label{PolarimetrySect}
\subsection{Spectropolarimetry}
We extracted spectra in Stokes $I$, $Q$ and $U$ for each source at the peak total intensity
in each cube channel. The errors for each of the peak intensities are estimated by taking the standard deviation of intensities at source free/off-source regions in each cube channel. We then propagate these errors for derived quantities, such as linear polarization, calculated from $I$, $Q$ and $U$. The peak intensity values are measures of the integrated intensity, as these sources are unresolved.
\par
Per-channel Stokes $I$ intensities for each source are shown as blue dots in each panel of Figure \ref{Tot_flux_plots}.
For comparison we also show flux densities for each source over a broader frequency range obtained from the NED database.
There is very good agreement with previously reported values for PKS $1934-638$,
NGC 1808 and PKS B$0407-658$. The spectrum, for NGC 1097, while consistently steep, is very poorly defined by previous data.
However, our data fall within the general trend of past observations.
\black{The data for J$0538-440$ suggest variability. Our measured spectrum differs significantly from the literature values. The discrepancy is brought about most likely by time variability in the J$0538-440$ spectrum as can be seen on the ATCA database}. The source has a total intensity spectrum that is flat near $\sim 1$ GHz with a standard deviation of $\sim 1.0$ Jy in the region while showing a bump towards, and also reaching a peak near, $\sim 10$ GHz. The results for the flat spectrum source J$0240-231$ also indicate variability.
The total flux of J$0240-231$ in the 1.36 - 2.0 GHz frequency range has been reported to vary over the range 4.4 - 7.1 Jy
\citep{Kuehr1981, Condon1998, Stanghellini1998a, Tingay2003a, Stanghellini2005}.
\begin{table*}
\centering
\caption{Degrees of linear polarization (in percentage) from \citet{Eichendorf1980} along with corresponding values form the KAT-7 bands in this work; \myavg{p_o}$_{1.35}$ and \myavg{p_o}$_{1.85}$. Subscripts are as in Table \ref{StokesI+pTable}. }
\begin{tabular}{lllllll}
\hline\hline\\
[-1.25ex]
Source & $p_{5.00}$ & $p_{2.73}$ & $p_{1.67}$& $p_{1.43}$ & \myavg{p_o}$_{1.35}$& \myavg{p_o}$_{1.85}$ \\
& (\%) & (\%) & (\%) & (\%) & (\%) & (\%) \\
[1.25ex]
\hline
NGC 1808 & -- & -- & -- & -- & $0.5\pm0.09$ & $0.5\pm0.10$\\
NGC 1097 & -- & -- & -- & -- & $0.8\pm0.11$ & $1.4\pm0.11$\\
PKS $1934-638$ & 0.1 & 0.3 & 0.4 & 0.2 & -- & --\\
PKS B$0407-658$ & 0.5 & 0.4 & 0.6 & 0.4 & $0.2\pm0.07$ & $0.2\pm0.07$\\
J$0240-231$ & 5.4 & 3.4 & 1.7 & 1.8 & $1.0\pm0.07$ & $2.1\pm0.07$ \\
J$0538-440$ & -- & -- & -- & -- & $0.7\pm0.08$ & $0.3\pm0.07$\\
\hline
\end{tabular}
\label{Eichendorf_1980Table}
\end{table*}
We derive the bias-corrected \citep{Simmons1985} linear polarized flux density as
\begin{equation}
P_{\rm o} = \sqrt{P^2- \sigma_{QU}^2}, \\
\end{equation}
where $P^2=Q^2+U^2$ and $\sigma_{QU}$ is the noise in the peak $Q$ and $U$ values derived as the rms of the amplitudes of off-source positions in each channel. For channel-averaged quantities \myavg{Q} and \myavg{U}, the noise is reduced by $1/\sqrt{N}$, where $N$ is the number of channels. Band averaged polarization is given by \myavg{P_o}. Similarly, the band averaged fractional polarization is given by
\begin{equation}
\langle p_o\rangle = \frac{\langle P_o \rangle}{\langle I_{pow}\rangle}.
\end{equation}
with fractional $Q$ and $U$ given by $u = U/I_{pow}$ and $q = Q/I_{pow}$. We compare band-averaged polarization values with previous studies. $I_{pow}$ is the simple power law fit to the measured peak total intensity:
\begin{equation}
I_{pow}= I_o\big(\frac{\nu}{\nu_o}\big)^{\alpha}.
\end{equation}
The bias correction method applied best approximates $p_{\rm o}$ when $p/\sigma_{qu}\gtrsim 3$. Here $\sigma_{qu}$ is the rms noise in $q$ and $u$. Table~\ref{StokesI+pTable} summarizes the band averaged total and polarized flux densities we derived for each source in each band, as well as the total intensity spectral index between the two bands.
\par
The recent study by \cite{Jagannathan2017} characterized the effect of instrumental polarization leakage in interferometric imaging. Such effects are much reduced for long integrations with large coverage in parallactic angle and also for sources nearer to the pointing centre of the observed field. We use the secondary calibrators to calibrate on-axis leakages. We find leakage terms at the level of $\sim 0.06\%$. PKS $1934-638$ was observed during each observing run as the primary calibrator. \black{This source has no detectable polarization at L-band and is a well known Southern hemisphere flux calibrator \citep[e.g.][]{Reynolds1994}.} Our average observed percent polarization for PKS $1934-638$ from the two frequency bands is $p = 0.06 \pm 0.07$ \%, indistinguishable from the on-axis instrumental polarization. The most conservative estimate of instrumental polarization then, is to assume that PKS $1934-638$ is unpolarized at this level of sensitivity and that the observed polarization is entirely instrumental. We have thus included in Table \ref{StokesI+pTable} an additional uncertainty of $0.06 \%$ in quadrature to the formal noise derived error on the linear polarization degree.
\par
The fractional polarization degree of each source in each band are listed in Table \ref{Eichendorf_1980Table}.
The table also lists some polarization fractions at similar frequencies found in the literature from \citet{Eichendorf1980}, who presented a
collection of polarization properties (position angles, flux densities, and rotation measures) that were measured
between 1965 and the middle of 1974 and these include data at 5.00, 2.73, 1.67, and 1.43 GHz -- listed in Table~\ref{Eichendorf_1980Table}. The typical observational errors in their data are in the range of 1 to 2 percent in 1965 while being better at 0.5 to 1 percent in 1974. Their data set contains observations for only three of our sources. \\
Linear polarization for just one of our targets, NGC 1097, was measured \citep{Beck2002,Beck2005} at 22, 18, 6, and 3 cm (corresponding to $\sim 1.36,\ 1.66,\ 5.00 $ and $\sim 10.00$ GHz, respectively) and also at 4.80 GHz \citep{Stil2009a}. The integrated linear polarization degrees for NGC 1097 were reported at low significance where $p_{\rm 22cm}=1.5\pm0.9\%$ and $p_{\rm 18cm}=1.3\pm1.0\%$ at resolutions of $2^{\prime\prime}-15^{\prime\prime}$, but still consistent with our results calculated from the peak polarized flux: $p_{\rm 22cm}=0.8\pm0.1\%$ and $p_{\rm 16cm}=1.4\pm 0.1\%$ (see Table \ref{Eichendorf_1980Table}). The source was undetected at 4.80 GHz where $p_o = 0.7\pm0.8\%$.
\subsection{RM Synthesis} \label{subsec:RMsynth}
The electric field vector of linearly polarized electro-magnetic waves is rotated through the Faraday process, as the waves pass through a magneto-ionic medium. The component of the magnetic field along our line of sight (LoS) can be probed through observations of this emission. \cite{Burn1966} presented a technique that can resolve the different components of magnetic fields that alter the polarization of the incident radiation along the LoS. The technique was later expanded and dubbed ``Rotation Measure Synthesis" (RM-Synthesis, for short) by \cite{Brentjens2005}.
\par
We apply the Rotation Measure Synthesis technique to analyze the intrinsic polarization properties of our sources. The technique transforms $q(\lambda^2)$ and $u(\lambda^2)$ to $q(\phi)$ and $u(\phi)$ through a Fourier transform -- $\phi$ is the Faraday depth, the line of sight integral of the thermal electron number density (denoted as $n_e$) and magnetic field ($\vec{B}$) component along the observer's line of sight:
\begin{equation}
\phi = K_c \int^0_r n_e\vec{B}\cdot d\vec{r}.
\end{equation}
$K_c$ is a constant. Faraday rotation is quantified by the Rotation Measure (or RM) in units of rad\,m$^{-2}$. The RM is defined by the change in polarization angle $\chi(\lambda^2)$ as a function of the square of the wavelength, $\lambda^2$:
\begin{equation}
{\rm RM} = \frac{d\chi(\lambda^2)}{\ d\lambda^2},
\end{equation}
with the polarization angle given by
\begin{equation}\label{eq:Xi-linear}
\chi = \frac{1}{2}\arctan\frac{u}{q}.
\end{equation}
The Faraday depth and the rotation measure are the same in the simplest case where there is only a single source of non-varying Faraday rotation. This is the so called Faraday simple case \citep[e.g.][]{Brown2017}, where polarization angle, $\chi$, is related to the Faraday depth by
\begin{equation}\label{eq:PA}
\chi=\chi_0 + {\rm RM}\lambda^2.
\label{eq:simpleFaraday}
\end{equation}
Equation \ref{eq:simpleFaraday} may be valid at modest bandwidths, where $\chi(\lambda^2)$ is linear, but this is not a guarantee.
In the case that the Faraday rotation plasma and the emitting plasma are mixed, this simple relationship cannot be assumed \citep{Farnsworth2011}. Observations over broad bandwidths can
display more complex Faraday rotation.
To characterize such Faraday complexity we use the Faraday spectrum, denoted as $F(\phi)$ \citep{Sun2015}, referred to as the Faraday dispersion function \citep{Brentjens2005}. Such spectra can reveal
Faraday components in polarized emission at different Faraday depths, $\phi$. $F(\phi)$ is related to linear polarized intensity, $P$ by:
\begin{eqnarray}
P(\lambda^2)= \int^{\infty}_{-\infty}F(\phi)e^{2i\phi\lambda^2}d\phi .
\label{PFtransform}
\end{eqnarray}
\begin{table*}
\centering
\caption{Key parameters in our RM Synthesis analysis. f$_{c}$ is the frequency range covered, ``Target'' refers to the disk galaxy target's observation run from which the quantities in the table are derived, $\delta\lambda^2$ is the spacing in $\lambda^2$, $\lambda^2_{min}$ and $\lambda^2_{max}$ are the minimum and maximum values, respectively, of $\lambda^2$ corresponding to each band. $\phi_{max-scale}$ is the largest Faraday depth scale to which our RM Synthesis analysis is sensitive and $\vert\phi_{max}\vert$ is the largest depth we can detect. $\delta\phi^{\prime}$ is the RMTF FWHM due to the non-continuous $\lambda^2$ coverage in this work while $\delta\phi$ is the same for the continuous case.}
\footnotesize{\begin{tabular}{ccccccccc}
\hline\hline\\
[-1.25ex]
f$_{c}$ & Target & $\delta\lambda^2$ & $\lambda^2_{min}$ & $\lambda^2_{max}$ & $\phi_{max-scale}$ & $\delta\phi$ & $\delta\phi^{\prime}$ & $|\phi_{max}|$ \\
& & [m$^2$] & [m$^2$] & [m$^2$] & [rad\,m$^{-2}$]& [rad\,m$^{-2}$] & [rad\,m$^{-2}$] & [rad\,m$^{-2}$]\\
[1.25ex]
\hline\\
[-1.25ex]
1.251 - 1.962 GHz & NGC 1097 & $9.29\times10^{-5}$ & $2.33\times10^{-2}$ & $5.74\times10^{-2}$ & 134.6 & 53.1 & $101.0\pm 0.8$ & $1.86\times10^{4}$\\
1.318 - 1.970 GHz & NGC 1808 & $9.18\times10^{-5}$ & $2.32\times10^{-2}$ & $5.17\times10^{-2}$ & 135.6 & 67.7 & $98.0\pm 0.8$ & $1.89\times10^{4}$\\
\hline
\end{tabular}}
\label{RMresolution}
\end{table*}
To perform RM Synthesis, we used a grid of trial $\phi$ values, spaced by 0.1 rad/m$^2$ ($\phi \in [-1000, 1000]$ rad\,m$^{-2}$), to measure $F(\phi)$. The inverse Fourier transform to derive $F(\phi)$ is approximated by a summation as observations do not cover a continuous wavelength range and also $\lambda > 0$ at all times:
\begin{equation}
F(\phi) \approx K\sum_{k} P(\lambda_{k}^2)e^{-2i\phi(\lambda^{2}_{k}-\lambda^{2}_{0})}
\label{eq:FsumAprox2}
\end{equation}
$K$ is the inverse of the $\lambda^2$ integral of the sampling function, $W(\lambda^2)$ (which is zero everywhere beyond the limits of the observed $\lambda^2$ range but non-zero where the range is sampled):
\begin{equation}
K = \left(\int^{\infty}_{-\infty} W(\lambda^2)\ d\lambda^2\right)^{-1},
\end{equation}
We progress through the following steps in applying RM Synthesis and deconvolving the intrinsic Faraday spectrum of each source:
\begin{itemize}
\item[1.] Calculate the Faraday spectrum for an input $q, u$ array to avoid the effects of a varying spectral index across a broad band which affects $Q$ and $U$.
\item[2.] Perform a deconvolution on the Faraday spectrum (i.e perform an ``RM clean" \citep{Heald2009} with a loop gain of 0.1. This deconvolution is iterated in a computational loop set to terminate at the RM clean threshold, explained below. \black{The RM synthesis analysis is done in the range $-1000$ to $+1000$ rad m$^{-2}$. RM ranges up to $\vert\phi_{max}\vert$ (see Table \ref{RMresolution}) were also tested and no emission components were detected beyond $\pm 1000$ rad m$^{-2}$.}
\item[3.] The result of the RM clean loop is then a multi-component model with each $\phi$-component characterized by a position angle, peak amplitude of the Faraday spectrum ($|F(\phi)|$), and a $\phi$ value.
\item[4.] We then create RM-cleaned $F(\phi)$, $q(\phi)$, $u(\phi)$, $q(\lambda^2)$, and also $u(\lambda^2)$ data sets based on the multi-component model.
\end{itemize}
Table \ref{RMresolution} lists the approximate values of parameters quantifying the limits of our KAT\,7 data with regard to RM Synthesis. The predicted resolution in Faraday space, $\delta\phi$, is slightly worse (67.7 rad\,m$^{-2}$) in the observations of NGC 1808, J$0538-440$, PKS B$0407-658$, and PKS $1934-638$ than for NGC 1097 and J$0240-231$ (where $\delta\phi = 53.1$ rad\,m$^{-2}$) due to a larger fraction of the band being flagged in the former case. \black{The $\lambda^2$ sampling of KAT 7 allows detection of RM components up to $\vert\phi_{max}\vert \sim 1.86 - 1.89\times10^{4} $ rad\,m$^{-2}$, where $\vert\phi_{max}\vert \approx \sqrt{3}/\delta\lambda^2$, \citep{Brentjens2005}}. \black{$\vert\phi_{max}\vert$ is calculated at the $3.9$ MHz channel width where $\delta\lambda^2\sim 9\times10^{-5}$ m$^2$ as listed in Table \ref{RMresolution}. The worst-case unaccounted rotation (occurring at the lowest wavelength where $\lambda^2_{max} \approx 0.05$ m$^{2}$) would therefore be $\sim 196^{\circ}$}. Note that the values listed in Table~\ref{RMresolution} are strictly correct for only continuous frequency coverage between the high and low frequencies of the
observations. The effect of the absence of the central band is to create strong ``side-lobes'' of the ``dirty'' Faraday transfer function. The quantity $\delta\phi^{\prime}$ lists the measured FWHM of the main lobe of the ``dirty'' RM spread function
(RMTF), and represents the actual Faraday resolution of our RM Synthesis spectra.
\par
We quantify the statistical significance of $\phi$-component detections by imposing an amplitude threshold in the RM clean algorithm. Our target confidence is set at $\gtrsim 99\%$. To this end, we select 1000 off-source points in each of our image cubes, extract the $Q, U$ and $I$ fluxes in a similar fashion as described previously. We then apply our RM clean algorithm and for each science target image cube, we determine the maximum amplitude $|F|_{max}$ from the set of off-source positions, \black{above which the Probability$(|F|_{max,i}\geqslant |F|_{max}) \leqslant 0.001$. This gives us a $0.999$} confidence of individual RM components for each of our science targets. Finally, we combine this limit with the instrumental polarization leakage error estimate of $0.06$ \% by adding them in quadrature and using the result as the lower threshold for
detection of Faraday components.
Table \ref{rm-components} lists the thresholds and detected Faraday components for each source. We compared the RM Synthesis results with and without a weighting of $\sigma^{-2}_{qu}$ applied and found that the results do not differ. We thus apply no weighting in our final RM Synthesis analysis.
\par
We characterize Faraday complexity according to the second moment about the mean, $\sigma_{\phi}$, of the RM clean model spectrum \citep{Brown2011,Anderson2015} given by:
\begin{equation}
\sigma_{\phi} = \sqrt{\frac{\sum_{i}{(\phi_i - \mu_{\phi})^2|F(\phi_i)|}}{\sum_{i}{|F(\phi_i)|}}},
\end{equation}
with
\begin{equation}
\mu_{\phi} = \frac{\sum_{i}{\phi_i|F(\phi_i)|}}{\sum_{i}{|F(\phi_i)|}}.
\label{eq:mu_phi}
\end{equation}
A Faraday simple spectrum will have $\sigma_{\phi} \approx 0$ rad\,m$^{-2}$ while the complex counterpart will have $\sigma_{\phi}$ very different from zero. We could not reliably resolve the complexity in individual components given that $\delta\phi^{\prime} \sim 100$ rad\,m$^{-2}$ but we classify a source spectrum as Faraday complex (denoted by C in Table \ref{rm-components}) or simple (denoted as S) according to $\sigma_{\phi}$ calculated from all components that are detected above the noise. Table \ref{rm-components} summarizes the $\phi$-components we found near the RM clean threshold level for each source. The {\it major/main/prominent} (highest signal-to-noise ratio, S/N) components show some width manifesting as a number of $minor$ (low S/N) components separated by less than $\delta\phi^{\prime}$ from the main component. A Faraday emission component, with $\phi = \phi_i$, is called {\it main/most prominent} if $|F(\phi_i)| = $ maximum$(|F(\phi)|)$, and {\it minor} otherwise. The $\phi_i$ of minor components within $\delta\phi^{\prime}$ of each other are combined to give $\mu_{\phi}$, of detections as listed in Table \ref{rm-components}. The $1\times\sigma$ deviations of these minor components is reported as the error on $\mu_{\phi}$.
\begin{table*}
\centering
\caption{RM clean components for each of our sources. $|F|_{comp}$ is the peak of the Faraday spectrum derived during the RM clean algorithm, $\mu_{\phi}$ (equation \ref{eq:mu_phi}) represents the Faraday depth of each component, ${\sigma_{\phi}}$ is the second moment of the RM clean model spectrum \citep{Brown2011,Anderson2015}, and the RM clean threshold. The column that follows, ``Class", shows classification of each source as Faraday complex (C) or Faraday simple (S) according to the value of ${\sigma_{\phi}}$. The last three columns show galactic coordinates ($l,b$) and also estimates of the smooth Galactic foreground Faraday depths, $\phi_{gal}$, at the locations of each our targets \citep{Oppermann2012,Oppermann2014}. }
\begin{tabular}{lcccccccc}
\hline\hline\\
[-1.25ex]
Source& $|F|_{comp}$ & $\mu_{\phi}$ & ${\sigma_{\phi}}$ & Threshold & Class & $l$ & $b$ & $\phi_{gal}$ \\
& [$\%$] & [rad\,m$^{-2}$]& [rad\,m$^{-2}$] & [$\%$] & & [deg]&[deg]& [rad\,m$^{-2}$] \\
[1.25ex]
\hline\\
[-1.25ex]
J$0240-231$ & 1.62 & $8.0\pm 0.1$ & $37.7$ & 0.061 & C & 39.415 & -23.224 & 9.1 \\
& 1.48 & $33.9\pm 1.4$ & & & & & & \\
& 1.29 & $-38.1\pm 4.8$ & & & & & & \\
& 0.69 & $-110.8\pm 2.0$ & & & & & & \\
& & & & & & & \\
J$0538-440$ & 0.45 & $91.2\pm 9.6$ & $40.8$ & 0.061 & C & 84.491 & -43.997 & 42.8 \\
& 0.17 & $9.6\pm 1.4$ & & & & & & \\
& & & & & & & \\
NGC 1097 & 1.09 & $-21.5\pm 8.9$ & $120.2$ & 0.072 & C & 41.579 & -30.275 & 9.1 \\
& 0.81 & $44.6\pm 3.5$ & & & & & & \\
& 0.47 & $147.3\pm 1.9$ & & & & & & \\
& 0.23 &$-520.5\pm 0.3$ & & & & & & \\
& 0.23 & $793.0\pm 0.3$ & & & & & & \\
& 0.22 & $619.9\pm 0.8$ & & & & & & \\
& & & & & & & & \\
NGC 1808 & 0.43 & $-0.5\pm 1.3$ & $49.4$ & 0.065 & C & 76.926 & -37.513 & 35.3 \\
& 0.24 & $144.7\pm 0.7$ & & & & & & \\
& 0.12 & $-291.6\pm 0.1$ & & & & & & \\
& & & & & & & \\
PKS $1934-638$ & -- & -- & -- & 0.061 & -- & 65.146 & -63.713 & 44.7 \\
& & & & & \\
PKS B$0407-658$ & 0.15 & $8.0\pm 0.1$ & $0.0$ & 0.061 & S & 62.085 & -65.753 & 23.9 \\
\hline
\end{tabular}
\label{rm-components}
\end{table*}
\begin{figure*}
\begin{subfigure}[t]{1.0\hsize
\centering
\includegraphics[width=0.8\textwidth]{0240-231-F_phi+residuals-v4.2.pdf}
\subcaption{The clean Faraday spectrum, $|F(\phi)|$, of J$0240-231$ resulting from our RM clean algorithm. See main caption below.}
\label{0240-rmclean}
\end{subfigure}
\begin{subfigure}[t]{1.0\hsize
\centering
\includegraphics[width=0.8\textwidth]{0538-440-F_phi+residuals-v4.2.pdf}
\subcaption{ The clean Faraday spectrum, $|F(\phi)|$, of J$0538-440$ resulting from our RM clean algorithm. See main caption below.}
\label{0538-rmclean}
\end{subfigure}
\end{figure*}
\begin{figure*}\ContinuedFloat
\begin{subfigure}{1.0\hsize
\centering
\includegraphics[width=0.8\textwidth]{NGC1097-F_phi+residuals-v4.2.pdf}
\subcaption{ The clean Faraday spectrum, $|F(\phi)|$ (green), of NGC 1097 resulting from our RM clean algorithm. See main caption below.}
\label{NGC1097-rmclean}
\end{subfigure}
\begin{subfigure}{1.0\hsize
\centering
\includegraphics[width=0.8\textwidth]{NGC1808-F_phi+residuals-v4.2.pdf}
\subcaption{ The clean Faraday spectrum, $|F(\phi)|$, of NGC 1808 resulting from our RM clean algorithm. See main caption below. }
\label{NGC1808-rmclean}
\end{subfigure}
\end{figure*}
\begin{figure*}\ContinuedFloat
\begin{subfigure}{.9\textwidth}
\centering
\includegraphics[width=0.8\textwidth]{pks1934-F_phi+residuals-v4.2.pdf}
\subcaption{ The clean Faraday spectrum, $|F(\phi)|$, of PKS $1934-638$ resulting from our RM clean algorithm. See main caption below. }
\label{pks1934-rmclean}
\end{subfigure}
\begin{subfigure}{.9\textwidth}
\centering
\includegraphics[width=0.8\textwidth]{PKSB0407-F_phi+residuals-v4.2.pdf}
\subcaption{ The clean Faraday spectrum, $|F(\phi)|$, of PKS B$0407-658$ resulting from our RM clean algorithm. See main caption below. }
\label{PKSB0407-rmclean}
\end{subfigure}
\caption{Figures \ref{0240-rmclean} - \ref{PKSB0407-rmclean}: The clean Faraday spectrum, $|F(\phi)|$ resulting from our RM clean algorithm (green solid curve) with residuals added back in. Top left: Zoom-in of the main plot showing deconvolved Faraday emission components (blue solid vertical lines show the peak $\gamma|F(\phi_i)|$ at the corresponding $\phi_i$ of each clean component in the vicinity of the brightest components) -- $\gamma$ is the loopgain parameter set to 0.1. The red dash-dotted horizontal line is the RM clean threshold as quoted in the legend. The dashed magenta curve is the ``dirty" Faraday spectrum while the dotted cyan curve is the RMTF scaled to the amplitude of the deconvolved $|F(\phi)|$ and with peak located at the Faraday depth corresponding to the peak of $|F(\phi)|$. Blue vertical dashed lines in the unzoomed plot display the value of $\mu_{\phi}$ of each detected Faraday component as quoted in Table \ref{rm-components}.}
\end{figure*}
Figures~\ref{pquX_fullband-ngc1097} - \ref{pquX_fullband-fluxcals} show the $\lambda^2$ dependence of $q$, $u$ and $p_o$ along with the model spectra reconstructed from the RM Synthesis clean components for each source. Figures~\ref{0240-rmclean} - \ref{PKSB0407-rmclean} show the ``dirty'' $F(\phi)$ for each source along with the RM Synthesis clean components. The strong side-lobes of the RM transfer function from the missing mid-band data set are visible in the dirty spectra. A ``cleaned'' $F(\phi)$ is also plotted, constructed by convolving the RM clean components with an idealized window function corresponding to full frequency coverage over the frequency range of the observations. The results for each source are discussed below.
\par
{\bf J$0240-231$}: This is the most polarized source in our sample with several Faraday components that are above the detection threshold -- see Figure \ref{0240-rmclean}. The most prominent component, as listed in Table \ref{rm-components}, is roughly consistent with the value of 14.4 $\pm$ 2.2 rad\,m$^{-2}$ measured by \cite{Taylor2009} based
on a simple two-point measure of polarization angle. Several Faraday components are seen in Figure \ref{0240-rmclean} and indicate that the fainter components are Faraday thick. Generally, the Faraday spectrum is complex with $\sigma_{\phi} = 37.7$ rad\,m$^{-2}$.
\par
{\bf J$0538-440$}: This source shows two relatively thick components, each composed of a number of closely adjacent minor components (Figure \ref{0538-rmclean}). The peaks secondary in prominence to the two most prominent, seen as the lower level clean components in Figure \ref{0538-rmclean}, have been assumed to form part of the thick components. The source is classified as Faraday complex with $\sigma_{\phi}= 40.8$ rad\,m$^{-2}$.
\par
{\bf NGC 1097}: Displays the most complex Faraday depth profile amongst the two disk galaxy sources -- Figure \ref{NGC1097-rmclean}, with thick components at low $\phi$ and some faint components just above the threshold at large $|\phi| > 500$ rad\,m$^{-2}$ (particularly near $-520,\ 590,$ and $790$ rad\,m$^{-2}$). The most prominent component seems Faraday thick, with minor components spanning the range $\phi = -11$ to $-32$ rad\,m$^{-2}$. This source is classified as Faraday complex and displays the largest dispersion in Faraday space with $\sigma_{\phi} = 120.2$ rad\,m$^{-2}$. The component detected at $147.3\pm 1.9$ rad\,m$^{-2}$ most likely has some sidelobe emission from the first sidelobe, at $\phi = +129.0$ rad\,m$^{-2}$, but cannot be solely due to the sidelobe as the same location reflected about 0.0 rad\,m$^{-2}$ does not show the same detectable signal.
\par
{\bf NGC 1808}: Displays a Faraday spectrum with three Faraday-thin components -- Figure \ref{NGC1808-rmclean}. We detect three prominent $\phi$-components. The component with $\mu_{\phi} = 144.7\pm 0.7$ rad\,m$^{-2}$ is close to the first RMTF sidelobe located at $\phi = +140.5$ rad\,m$^{-2}$, however $|F(\phi)|$ is more than $3\times$ the detection threshold, and its presence is clearly seen as an asymmetry in the dirty Faraday spectrum. The component located at $\mu_{\phi} = -291.6\pm 0.1$ rad\,m$^{-2}$ also appears to coincide with the second RMTF sidelobe. The dirty Faraday spectrum hints at a component near the positive (i.e the RMTF in the $\phi > 0.0$ rad\,m$^{-2}$ region) second sidelobe but there are no detectable clean components in the region. The source is Faraday complex with $\sigma_{\phi} = 49.4$ rad\,m$^{-2}$.
\par
{\bf PKS $1934-638$}: This source is undetected in polarization. The peak emission in the Faraday spectrum is near $\phi = 50.0$ rad\,m$^{-2}$ with a peak $|F| \approx 0.061\%$ ($\approx 0.064\%$ when the residuals are added in), which is set as the instrumental polarization level. Figure \ref{pks1934-rmclean} displays the Faraday spectrum.
\par
{\bf PKS B$0407-658$}: The Faraday spectrum of this source displays a single Faraday-thin component and does not show any other components above the threshold (Figure \ref{PKSB0407-rmclean}. The lone component is detected at $\mu_{\phi} = 8.0 \pm 0.1$ rad\,m$^{-2}$ with a peak $|F| = 0.15\%$.
\par
\subsubsection{Galactic Foreground and Ionospheric Effects}\label{Ionosphere-Subsect}
\black{Foreground RM contributions, such as from the ISM in the Galaxy and the Earth's ionosphere, can also be significant along the line of sight. This is especially so at frequencies below 1 GHz \citep[e.g.][]{Erickson2001}. However, most of the Galactic foreground emission is resolved out by an interferometer because of missing short spacings. There can be small scale structure in Stokes $Q$ and $U$ that the interferometer detects, and that can in principle become visible as a component in the Faraday depth spectrum. These are sometimes called Faraday Ghosts and are independent of position on the sky \citep[e.g.][]{Shukurov2003}.}
\par
\black{\cite{Oppermann2012} and \cite{Oppermann2015} have created a map of the smoothly varying Galactic Faraday component, $\phi_{gal}$, by utilising all-sky Faraday rotation measures of compact background sources from the literature. We list $\phi_{gal}$ at the locations of each of our targets in Table \ref{rm-components}. The effect of the Galactic ISM on an extragalactic source is mainly as a foreground Faraday-rotating screen. The correction for this effect is to subtract $\phi_{gal}$ from detected components. In this work, we note that this effect does exist and has possibly affected our measurements but we do not correct for it as it is not adequately constrained.}
\par
\black{In an effort to analyse Faraday complexity introduced by the ionosphere at the different observation epochs, we look for rotation in the polarization angle ($\chi$) as a function of time. This we do by producing full Stokes images for each scan of the phase calibrator for each epoch. We then restrict our analysis to those scans that have Stokes $V$ flux and $p_0$ that are below the leakage, and also Stokes I that is above $3\sigma$. We then analyse the resulting $q$, $u$ and $\chi$ as functions of scan i.e time. However, the resulting images prove to be much too noisy to lead to a conclusion on ionospheric effects.}
\par
\black{Typical Faraday rotation measure variations from the ionosphere are up to $\sim 1$ rad m$^{-2}$. Typical variations are of order one radian at 1m wavelength as $\chi(\lambda^2)\sim \lambda^2$. Variations of this order are seen at 327 MHz at the VLA \citep{Erickson2001}. At our lowest frequency even a 90 degree change at 327 MHz scales to only about 6 degrees at 1300 MHz, and 2.5 degrees at 1900 MHz. This is small enough to ignore. Note also that it is only the time variation during an epoch that will affect the observations. A constant offset from the ionosphere will be accounted for by the calibration. In the absence of an absolute polarization calibrator, the solution is referenced to the mean feed position angle which is not expected to change as it is hardwired into the instrument. We therefore conclude that the expected ionospheric rotation is negligible and thus not corrected.}
\section{Discussion}\label{Discussion}
Our RM Synthesis analysis shows Faraday complexity, of varying degrees, in four of the six sources in our KAT 7 sample. The most polarized source, J$0240-231$, displays the most complex Faraday spectrum with several emission components (see Figure \ref{0240-rmclean} and Table \ref{rm-components}). The presence of Faraday thick components can be seen in the Faraday spectrum as well. Our analysis clearly resolves this complex spectrum both near the main peak and the smaller peaks. This source also shows the most rapidly changing polarized intensity that depolarizes toward larger wavelengths (see the bottom panel of Figure \ref{pquX_fullband-ngc1097}). The behaviour of $q(\lambda^2)$ and $u(\lambda^2)$ is also far from the simple sinusoidal behaviour, displayed by a single Faraday component, and this deviation is best illustrated in Figure \ref{pquX_fullband-ngc1097}. This plot resembles a circle of constant radius in the Faraday simple case, but we see quite a substantial deviation from this. Our RM clean model can however, reproduce the behaviour we see in the spectra. The $q(\lambda^2)$ and $u(\lambda^2)$ profiles have little scatter and are in good agreement with our RM clean model adding confidence to the derived $\phi$-components. \cite{Anderson2016} show that Faraday-complexity is common in highly polarized sources and our observation in the case of J$0240-231$ is in accord. \citet{Schnitzeler2019}, \black{however, fit one Faraday component to the J$0240-231$ spectrum through a QU-fitting analysis. The source was found to be significantly polarized at a reference frequency of 2.1 GHz with $p = 2.28\%$ suggesting agreement with the measurement made in this work at the high, 1.85 GHz, band.}
\par
Faraday complex sources are also found to more likely have steep spectra in the optical and radio \citep[e.g.][]{Scarpa1997,Anderson2015}. We find J$0240-231$ to have only a mildly steep total intensity spectrum of \myavg{\alpha}$_{1.85}^{1.35}=-0.562 \pm 0.006$. Our results for J$0240-231$ agree with previous findings that the source is significantly polarized with $p_{4.9GHz}=3.47\%$ \citep{Perley1982} and $p_{1.4GHz} = 0.8\%$ \citep{Taylor2009}. The object is a bright compact radio source \citep{Fanti1990,Fanti2000}, however, the double lobed structure reported by \cite{Dallacasa1998} may contribute to the observed Faraday complexity as radio lobes can be sites of complex Faraday structures \citep[e.g.][]{Sebokolodi2019,Banfield2019}.
\par
The case of J$0538-440$ is another where we find evidence of Faraday complexity, although not to the level of J$0240-231$ (contrast Figure \ref{0538-rmclean} with Figure \ref{0240-rmclean}). The behaviour of $q(\lambda^2)$ and $u(\lambda^2)$ profiles for this source (bottom panels of Figure \ref{pquX_fullband-ngc1808}) resemble that of a simple one-component Faraday spectrum with somewhat larger scatter than in J$0240-231$, indicating that perhaps higher sensitivity observations with a larger $\lambda^2$ coverage and thus Faraday depth resolution are needed to further explore the behaviour and further resolve the $\phi-$components. This source is highly polarized, $\sim 18\%$, in the optical \citep{Giommi1995,Scarpa1997} and has possibly suffered significant depolarization at L-band. Our RM clean implementation is able to model the behaviour well at the $\sim 2.5\sigma$ level imposed. \black{The J$0538-440$ total intensity spectrum is more time variable but appears flatter than in the case of J$0240-231$ with \myavg{\alpha}$_{1.85}^{1.35}= -0.098 \pm 0.003$ and a lower degree of polarization of less than $1\%$ at L-band.} \black{This is another source that is also in the \citet{Schnitzeler2019} catalog where it is found to be Faraday simple and highly polarized, $p=3.32\%$ (at $RM \approx 60 $ rad m$^{-2}$), at 2.1 GHz in contrast with $p=0.3 \pm 0.07\%$ measured at the high band in this work. The total intensity of this source is a highly time variable and we suspect that the significant discrepancy we observe is due to this variability as sources that are variable in total intensity may also be variable in polarization \citep{Anderson2019}. We observed this source in late July 2016 as compared to the 2011 observations of \citet{Schnitzeler2019}. The authors obtained their results from fitting simple Faraday components to the sources spectrum which differs significantly from our open ended approach in which our model can contain as many (clean) components as is necessary to fit the data. With QU-fitting, a much more constrained model is fit to the data while minimizing errors. We do not apply constraints to this work. Another contributing factor to observed discrepancies is the difference in frequency bands, L and S-band. At L-band, it is typical that polarization is lower than at S-band, especially for compact sources.}
\par
NGC 1097 displays the most complex Faraday spectrum in the two disk galaxies. The Faraday spectrum of the source shows the largest spread in $\phi$ (see Figure \ref{NGC1097-rmclean}). Low S/N emission at comparatively large $\phi$ could be contaminated by sidelobes of the RMTF but may also be due to small regions of highly magnetised media which may emit at large Faraday depths. These small regions may coincide with highly turbulent regions near the core of a galaxy \citep[e.g][]{Pasetto2018}. The low S/N of these components may also point to depolarization that may be due to turbulence of the magnetised plasma in the bar and circumnuclear ring regions of the galaxy and/or frequency dependent depolarization that diminishes the polarized signal at L-band. Higher resolution measurements by \cite{Beck2005} do not show evidence of high rotation measure regions in NGC 1097 (the authors report $\phi$ in the range -200 to +200 rad\,m$^{-2}$ from 50 MHz bandwidth observations at 4.86 and 8.46 GHz) and so we may be detecting a noise contribution to the signal that is unaccounted for. More sensitive observations with a wider bandwidth coverage will be required to confirm these faint components. Polarimetry at higher frequency may also be able reveal such components in the absence of internal depolarization effects \citep[e.g.][]{Anderson2016}.
\par
NGC 1808 is also classified as Faraday complex in our RM clean analysis with the detection of three Faraday-thin components -- see Figure \ref{NGC1808-rmclean}. The complexity classification in this disk galaxy is due to the fairly large spread of Faraday depths found. Our RM-clean implementation shows subtle structure of the polarized spectra of the galaxy while the data shows a low S/N and significant scatter at the wavelengths observed. A larger $\lambda^2$ coverage with higher sensitivity is required as we will be able to better constrain the polarized spectra of such Faraday complex and low polarized S/N sources \citep[e.g][]{OSullivan2012,Anderson2015,Anderson2016}.
\par
\black{The apparent spectral indices we derive suggest agreement with those expected for spiral galaxies; $\alpha=-0.8 $ to $-1.0$, \citep{Beck2015}. We find \myavg{\alpha}$_{1.85}^{1.35} = -0.815 \pm 0.011$ for NGC 1808 and \myavg{\alpha}$_{1.85}^{1.35} = -1.060 \pm 0.020$ for NGC 1097. NGC 1808 is weakly polarized, with $p<1\%$ in accordance with models of the integrated polarization of galaxies at 1.4 GHz \citep{Stil2009,Sun2012}. The source also shows evidence of possible polarized emission components emitted at high Faraday depths as suggested by low level (undetected) large $\phi$ emission in the deconvolved Faraday spectrum.}
\par
PKS $1934-638$ \black{is not detected in polarization}. This source shows polarization levels that are consistent with noise across the L-band. \cite{Eichendorf1980} quote polarization levels that are of order double what we have found with observational errors typically being in the range of 1 to 2 percent in 1965 and lowering to 0.5 to 1 percent in 1974. We interpret their result as upper limits. Our results are comparable to the earlier data (Table \ref{Eichendorf_1980Table}) showing that PKS $1934-638$ is undetected in polarization.
\par
PKS B$0407-658$, however, is found to have one detected component -- Figure \ref{PKSB0407-rmclean}. The $q(\lambda^2)$, $u(\lambda^2)$, and $p(\lambda^2)$ profiles of this source are not as featureless as in PKS $1934-638$ and may require a larger $\lambda^2$ coverage to further resolve the Faraday spectrum. More sensitive MeerKAT commissioning observations show that PKS B$0407-658$ is polarized at $\sim 0.03 - 0.05\%$ levels in the L-band \citep{Hugo2018}.
\par
\black{There exists no known relation between Faraday complexity and morphological complexity at the angular scales probed here \citep[e.g.][]{Anderson2015}} so even the most complex sources may be much simpler in morphology. More sensitive and higher resolution broad-band spectropolarimetry studies are needed to further resolve Faraday components in Faraday depth, across wide bandwidths, at higher frequencies, and also spatially with high S/N as will be the case with the next generation of radio telescopes such as MeerKAT.
\section{Summary and conclusions}\label{Summary+conclusions}
We have explored the properties of broadband linear polarization in two galaxies and four AGN powered sources in direct pointing mode with the KAT 7 array. The observations were made at the KAT 7 low band ($\nu_c \sim 1350$ MHz) and high band ($\nu_c \sim 1850$ MHz) with the mid band ($\nu_c \sim 1550$ MHz) discarded due to RFI. The synthesized beam at the low band was $\sim 4.9^{\prime}$ and the same at the high band was $\sim 3.9^{\prime}$. We detect linear polarization down to $\sim 2$ mJy/bm. We employed the open-ended method of RM Synthesis and RM clean \citep{Burn1966,Brentjens2005,Heald2009} which enabled us to resolve the Faraday depth components down to a Faraday depth resolution $\delta\phi^{\prime}\sim 100$ rad\,m$^{-2}$ along our line of sight. Our conclusions are summarized as follows:
\begin{itemize}
\item[({\bf i})] The total intensity spectra that we were able to measure are in general agreement with previous studies.
\item[({\bf ii})] Our spectropolarimetric \black{analysis} shows Faraday simple, PKS B$0407-658$, \black{unpolarized PKS $1934-638$}, and Faraday complex sources \black{-- J$0538-440$, J$0240-231$, NGC 1097 and NGC 1808}. The most polarized objects display more complex Faraday spectra which is indicative of multiple emitting and/or diminishing components at different Faraday depths along the line of sight.
\item[({\bf iii})] \black{Our RM Synthesis analysis did not detect Faraday complexity in the least polarized source (PKS B$0407-658$) but such sources may still have some level of underlying Faraday complexity that is beyond detection}. The two disk galaxy sources are classified as complex and have been found to have emission components over larger Faraday depth ranges -- suggesting complex magneto-ionic emission environments along the line of sight.
\item[({\bf iv})] The BL Lac source, J$0538-440$, and QSO, J$0240-231$, are the most highly polarized in the literature and most Faraday complex of our direct pointing sources. Our measurements support earlier findings that Faraday complexity is expected in most polarized sources. The lower Faraday complexity found in the other less polarized sources may require more sensitive wide band and thus higher $\phi$ resolution studies to further tease out the properties that may be beyond our current data. Low polarization sources may also have a significant degree of spectropolarimetric complexity that is beyond detection in studies with similar sensitivity and resolution limitations as ours. \black{}.
\item[({\bf v})] We determine the on-axis instrumental polarization level in KAT 7 to have an upper limit of $0.06\%$. This is based on the polarization level measured for the hitherto unpolarized GPS source PKS $1934-638$ spectrum.
\item[\black{({\bf vi})}] \black{Some BL Lac sources, such as J$0538-440$, are particularly variable with time suggesting more complexity which should be accounted for when such sources are used as calibrators especially over broad bandwidths}
\end{itemize}
\section*{Acknowledgements}
This work is funded by the South African National Research Foundation (SA NRF) and the South African Radio Astronomy Observatory (SARAO), formerly known as the Square Kilometre Array South Africa (SKA SA) project, through their doctoral student scholarship. The analysis is done on the Inter-University Institute for Data Intensive Astronomy (IDIA) cloud computing facility.
\section*{Data availability}
The data underlying this article can be shared upon reasonable request to the corresponding author.
\bibliographystyle{mn2e}
|
\section{Introduction: the controversy}
\label{sec:intro}
Due to its cleanness, it is hard to overstate the experimental
importance of the decay of the Higgs particle into two photons. It
goes mainly via virtual $W$-bosons, the heavier charged particles of
flavourdynamics. The amplitude of this contribution was calculated to
the first non-vanishing order (one-loop, cubic in the couplings) long
ago in the light-higgs limit~\cite{Eliseo} -- and then ``exactly''
in~\cite{RussianCharge2}. The accepted result was confirmed many times
-- see~\cite{elBelen} for a particularly clever calculation. It does
\textit{not} vanish in the heavy-higgs limit -- which seems to fly in
the face of the ``decoupling theorem'' (DT) in~\cite{Loscarrozas}, as
often understood.
Much more recently, those calculations were questioned in
\cite{GastmansWuWu1, GastmansWuWu2}. The ensuing debate highlights the
\textit{theoretical} relevance of this decay. The authors of these
papers made the point that, since the higgs cannot couple directly to
the photons, the one-loop contribution must be finite: there are no
couplings requiring ``renormalization''. The roundabout procedures
through ``renormalizable gauges'', they concluded, were unnecessary.
Eschewing dimensional regularization, they recomputed the amplitude in
the unitary gauge of electroweak (EW) theory. They did obtain a result
differing from the standard one by an additive constant, which shows
up for instance in the heavy-higgs limit -- whereby their result is
equal to zero.
There was no shortage of rejoinders~\cite{TresChinosCutoff,
MartianCharge, RussianCharge3, TresChinosAgudos, Jaguarlehner,
PichininiPP, DedesS, Weinzierl14} to \cite{GastmansWuWu1,
GastmansWuWu2}. The authors of~\cite{RussianCharge3} are the ones of
the original calculation~\cite{RussianCharge2}. Those papers made
several points, some rather implausibly arguing that at a given point
in the calculation in \cite{GastmansWuWu2} electromagnetic gauge
invariance is lost, and criticizing the interpretation of the DT made
in~\cite{GastmansWuWu1, GastmansWuWu2}. There was in some of the the
rejoinders an explanatory reliance on the heuristics of the
Brout--Englert--Higgs mechanism, throwing back the so-called
``equivalence theorem'' (GBET).
The criticisms received a rejoinder in turn in~\cite{WusStrikeBack}.
This later paper argues by the example that two computations of the
same process in different gauges ($R_\xi$ versus unitary gauge) may
yield different results. This goes against the grain, although of
course no theorem contradicts such an assertion. Meanwhile, a
dispersion relation calculation carried out in~\cite{ChristovaI}
appeared to support the contentions of~\cite{GastmansWuWu1,
GastmansWuWu2}, and got in turn a -- quite thoughtful -- rejoinder
in~\cite{MelnikovVainshtein}. More recent papers dealing with the same
or related issues are \cite{GegeliaM18,BoraChristovaEberl}.
By and large, the majority's opinion and the experimental results
\cite{Jacobs} support the first tally. On the other hand, from the
theoretical point of view the situation is still obscure: it had to be
so, since both parties draw strength from different casuistics of the
calculations in perturbative quantum field theory.
\smallskip
The debate about the uses and abuses of the unitary gauge and the role
of the decoupling and equivalence ``theorems'' is to be saluted as
salutary. And it is safe to admit that up to now we lack a full
conceptual understanding of the problem. The cleanest way to address
this lack is surely to renounce \textit{all} the heuristics of
mathematically ill-defined quantities, in favour of a method in which
there can be no argument on the meaning of infinite terms. Such is the
truly (perturbatively) stringent scheme by Bogoliubov, Epstein and
Glaser (BEG) of ``renormalization'' without regularization, by
extension of distributions.
In the BEG construction, governed by causality, there is no such thing
as a ``divergent diagram'': one never encounters infinities. There
may, however, remain in the extension procedures some additive
\textit{ambiguity}, that can be restricted (but not always completely
removed) by physical principles. This is rather to be regarded as a
strength of the BEG paradigm, because those ambiguities express
precisely how, and to what extent, the theory is determined by the
fundamental principles of perturbative QFT.
A particular advantage of the inductive BEG construction
\cite{EpsteinGlaser73} of the (functional) $\bS$-matrix is that in
principle one is allowed to stay on configuration space, which makes
more transparent the physics under examination. For examples of
calculations within the BEG scheme explicitly carried out in
configuration space, see \cite{Elara} or \cite[Sect.~3.5]{Duetsch19}.
It is only for computational convenience that we switch at some moment
to momentum space.
Since we do not deal in infinities, we refer as \textit{normalization}
to the processes taking the place of regularization and
renormalization in the BEG framework. For its relative paucity of
diagrams, in our context the underlying argument is made clearer by
working mostly in the unitary gauge -- whereupon only the
physical particles' data are brought to bear.%
\footnote{The paper~\cite{IrgesKoutroulis17} dwells usefully on the
subject of the $R_\xi$-versus-unitary gauges, leaning to demonstrate
the validity of the latter at the quantum level.}
To summarize, so far: we were motivated to tackle this subject by
wondering why most knowledgeable people, borrowing different (but all
apparently sound) methods to work on such a basic process, were
divided on the outcome. It all turns around a subtlety uncovered by
use of the BEG normalization. That condenses the purpose of the
present paper.
\subsection{Main results and plan of the article}
\label{ssc:facile-est}
In App.~\ref{app:spin-one-basics} we introduce our conventions and
notations, recalling a few well-known formulae of QFT needed in the
body of the paper, in particular the propagators for the EW theory in
the unitary gauge. Let $m_h$ denote the mass of the higgs~$h$. The
amplitude coming from the one-loop calculations may be quoted as
\cite{HungryHunters, BardinP99, GranLev}:
$$
\sA = \frac{g\al}{2\pi M} F_1(\rho) P_{\mu\nu},
$$
with $\al$ the fine structure constant, $g$ the EW coupling constant,
$M$ the mass of the intermediate $W$-boson and $\rho := m^2_h/4M^2$.
The polarization factor $P_{\mu\nu}$, reflecting electromagnetic gauge
invariance (EGI) of~$\sA$,%
\footnote{That is, transversality of the outgoing photons.}
is written in this paper as
\begin{equation}
P_{\mu\nu} := (k_1 k_2) g_{\mu\nu} - k_{1\nu} k_{2\mu}; \quad
(P_{\8\nu} k_1) = (P_{\mu\8} k_2) = 0,
\label{eq:in-tractu-temporis}
\end{equation}
with $k_1$, $k_2$ the outgoing photons' momenta. Finally, for the
dimensionless factor:
\begin{equation}
F_1(\rho) := 2 + \frac{3}{\rho}
+ \frac{3}{\rho} \biggl( 2 - \frac{1}{\rho} \biggr) f(\rho).
\label{eq:bone-of-contention}
\end{equation}
Now that we are at that, we quote as well the comparable result for a
charged scalar particle of mass~$M$ at the place of the $W$-boson:
\begin{equation}
F_0(\rho) = \frac{1}{\rho} \biggl( 1 - \frac{f(\rho)}{\rho} \biggr);
\word{so that} F_1(\rho) = 3F_0(\rho) + \frac{6f(\rho)}{\rho} + 2.
\label{eq:magister-dixit}
\end{equation}
For the benefit of the reader coming to the subject of this paper for
the first time, App.~\ref{app:curioser} introduces the distribution
$f(\rho)$ appearing in both $F_1$~\eqref{eq:bone-of-contention} and
$F_0$~\eqref{eq:magister-dixit} -- as well as in the amplitude of
diphoton decay of~$h$ via virtual fermions.
The bone of contention is that the first summand~$2$ in
\eqref{eq:bone-of-contention} should not be there, according to
\cite{GastmansWuWu1, GastmansWuWu2, ChristovaI}. Relations
\eqref{eq:f1} and \eqref{eq:arcsin-2} tell us that, as $\rho\downto0$:
$$
F_1 = 2 + \frac{3}{\rho}
+ \biggl( \frac{6}{\rho} - \frac{3}{\rho^2} \biggr)
\biggl( \rho + \frac{\rho^2}{3} + \frac{8\rho^3}{45} +\cdots \biggr)
= 7 + \frac{22}{15}\,\rho + O(\rho^2);
$$
so $F_1(0) = 7$ and $F_1(\infty) = 2$ from
\eqref{eq:bone-of-contention}. Precisely the former figure is what was
calculated in the paper~\cite{Eliseo}. The result argued by the
``heretics'' in the controversy is $F_1 - 2$, so their respective
assertions are instead $F_1(0) = 5$ and $F_1(\infty) = 0$. Also, from
\eqref{eq:magister-dixit}: $F_0(0) = -1/3$ and $F_0(\infty) = 0$.
\smallskip
Appendices A and B of this paper deal with conventions and
mathematical prerequisites. The basics of the BEG scheme are recalled
in Appendix~\ref{sec:Streu}. Understanding of the BEG method is
indispensable in what follows, and even readers familiar with it are
advised not to miss our review. The relation between the normalization
problem by extension of distributions (or by ``distribution
splitting'') and \textit{dispersion integrals} is treated in its
subsection~\ref{sec:ipso-facto}. New results in this respect are
required, announced in the short Section \ref{sec:quam-scriptum} and
proved in subsections \ref{ssc:do-ut-des} and
\ref{ssc:hasta-ahi-podiamos-llegar} of this paper. So for
\textit{aficionados} of BEG normalization there is novelty here --
whose interest goes beyond the particular problem that motivated~it.
Sections~\ref{sec:argumentum} and ~\ref{sec:soberbia-pagana}
constitute the heart of the paper. The scalar model leading to $F_0$
is worked out in Section~\ref{sec:argumentum}. One is able to perform
the ``adiabatic limit'' of Epstein and Glaser at an intermediate step,
which simplifies computations -- this is rigorously justified. This
``toy model'' allows the reader to familiarize with the BEG
construction of time-ordered products in a relatively simple case. For
it, the ambiguity in the Epstein--Glaser result can be disposed of,
and the unique outcome happens to coincide with the result of a
``naive'' on-shell calculation, of the kind performed
in~\cite{ChristovaI}.
Finally, in Section~\ref{sec:soberbia-pagana}, we compute the EW
amplitude, working first in the unitary gauge. We start in earnest by
illustrating in this relevant instance the machinery of the BEG
formalism in constructing time-ordered products, at the lowest
non-trivial order: from cubic interaction vertices, identified to
time-ordered products at first order in the couplings, we
\textit{derive} the quartic, second-order $AAWW^\7$-vertex.
It is time to aver why the ``no-renormalization'' argument in
\cite{GastmansWuWu2} is not watertight. A direct $h\ga\ga$~coupling in
flavourdynamics is forbidden also because of EGI. Thus to obtain the
general amplitude, which lives off-shell, one must add to the naive
calculations a polynomial in the external momenta, of degree given by
the singular order of that amplitude. Computing the $1$-loop
contribution in the unitary gauge by the Epstein-Glaser method, we
ratify this fact. To find the coefficients of that polynomial, beyond
EGI here we call upon gauge-fixing independence of the on-shell
amplitude. This locks in the indetermination; and in the end we do
obtain~$F_1(\rho)$. Within the unitary gauge, a different argument to
the same purpose is discussed at the end of this
Section~\ref{sec:soberbia-pagana}. Section~\ref{sec:die-Eule} is the
conclusion.
\section{The obstruction to distribution splitting for null momenta}
\label{sec:quam-scriptum}
Formula \eqref{eq:pro-reo} in App.~\ref{sec:Streu} is our main
workhorse: in momentum space the Epstein--Glaser distribution
splitting amounts to a dispersion integral. But it pertains to remark
that, by construction, prescriptions \eqref{eq:hoist-with-retard}
and~\eqref{eq:pro-reo} are in principle valid \textit{only for
timelike}~$k$. Thus, in order to solve the problem in this paper, one
has to run an extra~mile. The explicit splitting procedure introduced
here exhibits relevant novel features: we have to compute the central
solution $a^c(k_1,k_2)$ for \textit{null momenta}. Hence, one cannot
immediately use the dispersion integrals \eqref{eq:hoist-with-retard}
or~\eqref{eq:pro-reo}. On trying to work instead with the convolution
integral \eqref{eq:ac}, there appears the problem that, in spite of
$k_j^2 = 0$, it generally holds that $(k_j - v_j)^2 \neq 0$ because
$v_j \in V_+$; it \textit{does not suffice} to know the causal
distribution $d(k_1,k_2)$ only for $k_1^2 = 0 = k_2^2$.
The next section solves this problem for models such that $0 < (k_1 +
k_2)^2 < 4M^2$ and $k_1^0 k_2^0 > 0$. The proof's strategy is as
follows: starting from the dispersion
integral~\eqref{eq:hoist-with-retard} for $k_1^2 > 0$, $k_2^2 > 0$ and
$k_1^0 k_2^0 > 0$, we intend to show that $d(k_1,k_2)$ is regular
enough that this integral commutes with the limit $(k_1^2 \downto 0
\wedge k_2^2 \downto 0)$. Therefore the dispersion
integrals~\eqref{eq:hoist-with-retard} and~\eqref{eq:pro-reo} keep
their usefulness for $k_1^2 = 0 = k_2^2$: indeed, for computing
$a^c(k_1,k_2)|_{k_1^2=0=k_2^2}$ it suffices to know $d(k_1,k_2)$ only
for $k_1^2 = k_2^2 = 0$, because $k_1^2 = k_2^2 = 0$ implies $(tk_1)^2
= (tk_2)^2 = 0$ for all~$t$.
Crucially, in the resulting dispersion integrals
\eqref{eq:hoist-with-retard} and \eqref{eq:pro-reo} for $k_1^2 = 0 =
k_2^2$, the parameter $\om$ is the singular order of the
\textit{off-shell} $d(k_1,k_2)$. As a consequence, the general
solution (prior to imposition of other invariance rules) of the
distribution splitting is obtained by adding to
$a^c(k_1,k_2)|_{k_1^2=0=k_2^2}$ a polynomial in $k_1,k_2$, in
principle arbitrary, whose degree is given by the singular order of
the off-shell amplitude $d(k_1,k_2)$. Now, it frequently happens that
the singular order of $d(k_1,k_2)|_{k_1^2=0=k_2^2}$ has a
\textit{smaller} value. Consequently, it may happen that the required
dispersion integral appears to be ``oversubtracted'' -- i.e., it would
be convergent also for a smaller value of~$\om$. Examples for this are
the ``toy model'' in the next section and the EW diphoton decay of the
higgs in the unitary gauge (subsections
\ref{ssc:hasta-ahi-podiamos-llegar} and \ref{ssc:quod-feceris},
respectively).
These issues were realized by Raymond Stora, who, referring to the
very subject process of this paper, pointed out to one of us that the
good behaviour of the absorptive part of the form factor involving
Compton scattering of the $W$-bosons should not make one forget that
BEG-generated dispersion integrals, just as perturbative
renormalization theory in general, applies off-shell.%
\footnote{Private communication, early 2013.}
\section{Higgs to diphoton decay via a charged scalar field}
\label{sec:argumentum}
The scalar electrodynamics computation leading to~$F_0$ works like a
kind of toy model, allowing the reader to familiarize with our methods
in a less complicated, although non-trivial case. We develop it in the
present section. Notice the following: in the Epstein--Glaser scheme
the ``seagull'' $e^2 AA\vf\vf^\7$-vertex \textit{is derived} by
implementing EGI within the construction rules of the method -- as any
other part of $T_2$~\cite{DKS93}. We give full details on how this
comes about for the quartic vertex in the EW theory in subsection
\ref{ssc:per-angusta}. The game here would be similar, only simpler.
The reader is advised to keep in mind the methods and standard
notations recalled in subsection~\ref{ssc:lost-in-translation}.
\subsection{A causal distribution on-shell}
\label{ssc:d-scalar}
The starting point is given by the lower order time-ordered products
(TOPs):
\begin{align*}
&T_1(x_3) = g M\,h(x_3)\, \vf(x_3)\, \vf^\7(x_3); \quad
T_1(x_j) = -ie A^\la(x_j)\, \vf^\7(x_j) \vecvec{\del_\la} \vf(x_j),
\;\; j = 1,2;
\\
& T_2(x_1,x_2)
\\
&\quad = -e^2 A^\mu(x_1) A^\nu(x_2) \bigl[ \vf^\7(x_1)
\,\del_\mu \Delta^F(x_1 - x_2) \,\del_\nu\vf(x_2) -
\del_\mu\vf^\7(x_1) \Delta^F(x_1 - x_2) \,\del_\nu\vf(x_2)
\\
&\quad
+ \vf^\7(x_1)\, \bigl( \del_\nu\del_\mu \Delta^F(x_1 - x_2)
+ i g_{\mu\nu} \,\dl(x_1 - x_2) \bigr) \vf(x_2)
- \del_\mu\vf^\7(x_1) \,\del_\nu \Delta^F(x_1 - x_2)\, \vf(x_2) \bigr]
\\
&\quad + (x_1 \otto x_2) + T_1(x_1)\,T_1(x_2) + [\text{irrelevant loop
diagram terms}],
\end{align*}
where $\Dl^F$ denotes the Feynman propagator~\eqref{eq:post-factum}.
From our formulas~\eqref{eq:quod-non-speratur1} and
\eqref{eq:quod-non-speratur2}:%
\footnote{The $D_n$ are always linear combinations of commutators.}
\begin{equation}
D_3(x_1,x_2,x_3) = -[\ovl{T}_1(x_1), T_2(x_2, x_3)]
- [\ovl{T}_1(x_2), T_2(x_1, x_3)] + [\ovl{T}_2(x_1, x_2), T_1(x_3)].
\label{eq:para-llorar}
\end{equation}
Because the photons emitted at $x_1,x_2$ are on-shell, only the third
commutator is relevant here -- in the language of Cutkosky rules, one
needs only the triangle cut separating the higgs vertex from the
propagator connecting the photons. We give the explanation further on.
From the general formula for the antichronological product
\eqref{eq:full-house}, we particularly know that
\begin{equation}
\ovl{T}_1(x_1) = T_1(x_1); \quad
\ovl{T}_2(x_1, x_2)
= - T_2(x_1, x_2) + T_1(x_1) T_1(x_2) + T_1(x_2) T_1(x_1).
\label{eq:ovlT}
\end{equation}
For the same reasons just argued, only the connected tree diagram part
of the $T_2(x_1, x_2)$ summand in $\ovl{T}_2(x_1, x_2)$ contributes.
A most convenient parallel for the coming calculation is the treatment
of the vertex function in QED in the first edition of the finite QED
book by Scharf~\cite[Sect.~3.8]{Scharf89}. Going to the contractions,
bringing in the vertices and the propagators
\eqref{eq:accidit-in-puncto}, \eqref{eq:nullius-in-verba}, apart from
a factor $4ge^2M$ we obtain:
\begin{align*}
& A_\mu(x_1) A_\nu(x_2) h(x_3) \bigl[
\Dl^-(1)\,\del^\mu \Delta^F(1 - 2)\,\del^\nu\Dl^-(2)
- \del^\mu\Dl^-(1)\,\del^\nu \Delta^F(1 - 2)\,\Dl^-(2)
\\
&\quad - \del^\mu\Dl^-(1)\,\Delta^F(1 - 2)\,\del^\nu\Dl^-(2)
+ \Dl^-(1)\bigl( \del^\mu\del^\nu \Delta^F(1 - 2)
+ i g^{\mu\nu}\,\dl(1 - 2) \bigr) \Dl^-(2)
\\
&\quad -[\text{the same four terms with $\Delta^-$ replaced by
$\Delta^+$}] +\cdots \bigr]
\\
&=: A_\mu(x_1) A_\nu(x_2) h(x_3)\, d^{\mu\nu}(1,2),
\end{align*}
where $1 \equiv y_1 := x_1 - x_3$, $2 \equiv y_2 := x_2 - x_3$. Here
and further down, the dots stand for the terms coming from the other
two cuts and further terms not contributing to the on-shell amplitude.
Note the advertised additional $+i g^{\mu\nu}\,\dl$ to
$\del^\mu\del^\nu \Delta^F$, corresponding to the ``closed seagull''
or fish-like diagram contribution to the $h \to 2\ga$ decay in this
model.
\smallskip
We now proceed to momentum space, where computations are carried out
more simply. For Fourier transformations, consult the convention
\eqref{eq:def_Fourier}. In this section and the next, in keeping with
physicists' notation, we indicate the transforms by just exhibiting
the variables, namely: $d^\mu(k_1,k_2) \equiv \hat d^\mu(k_1,k_2)$. We
obtain
\begin{align}
d^{\mu\nu}(k_1,k_2)
= & \frac{1}{(2\pi)^2} \bigl[
4(I_+^{\mu\nu} - I_-^{\mu\nu}) + 2 k_2^\nu(I_+^\mu - I_-^\mu)
- 2 k_1^\mu(I_+^\nu - I_-^\nu) - k_1^\mu k_2^\nu(I_+ - I_-)
\notag
\\
&\quad
- \frac{i}{(2\pi)^2}\, g^{\mu\nu}\, (J_+ - J_-) \bigr] +\cdots
\label{eq:suerte-o-verdad}
\end{align}
with the integrals
\begin{align}
I_\pm^{\{\cdot|\mu|\mu\nu\}}(k_1,k_2)
&:= \int d^4k\, \{1|k^\mu|k^\mu k^\nu\} \,\Dl^\pm(k_1 - k) \Delta^F(k)
\,\Dl^\pm(k + k_2),
\nonumber \\
J_\pm(k_1,k_2) &:= \int d^4k\, \Dl^\pm(k_1 - k)\,\Dl^\pm(k + k_2),
\label{eq:def-J}
\end{align}
where the $J_\pm$-term is the contribution of the fish-like diagram.
Keep in mind that the terms belonging to~$A'_3:=A_3-T_3$ are those
coming from the integrals $I^{\cdot|\mu|\mu\nu}_-$ and $J_-$, whereas
the contribution of~$R'_3:=R_3-T_3$ is given by the integrals
$I^{\cdot|\mu|\mu\nu}_+$ and~$J_+$.
For our purposes one may perform the adiabatic limit already at this
stage. Since all internal lines of the diagrams correspond to massive
fields, this limit can be done here in the naive way by just setting
the switching function $g(x)$ in~\eqref{EG-summacumlaude} to~$1$:
\begin{align}
& \int dx_1\,dx_2\,dx_3\, A^\mu(x_1) A^\nu(x_2) h(x_3)\,
d^{\mu\nu}(x_1 - x_3, x_2 - x_3)
\notag
\\
&\enspace = (2\pi)^2 \int dk_1\,dk_2\, h(k_1 + k_2) A^\mu(-k_1)
A^\nu(-k_2) \,d^{\mu\nu}(k_1,k_2).
\label{eq:dAAh}
\end{align}
In this limit the momenta $k_1$ and $k_2$ become the momenta of the
external photons: $k_1^2 = k_2^2 = 0$.
From now on, we compute $d^{\mu\nu}(k_1,k_2)|_{k_1^2=0=k_2^2}$. Were
we to have included the other cuts in \eqref{eq:para-llorar} or
$T_1T_1T_1$-terms, there would appear $\Dl^\pm$-type propagators at
the place of the Feynman propagators above. The former are $\sim
\dl(k^2 - M^2)$, with $k$ denoting the internal momentum variable in
the loop: so to speak, in contrast with the Feynman
propagators, the $\Dl^\pm$ are ``always on-shell'', even within loops.%
\footnote{This point is made in \cite[Sect.~6.4]{Diag}.}
Thus no further internal momenta can be on-shell: assuming $k^2 = M^2$
one obtains $(k_1 - k)^2 = M^2 - 2(k_1k) \neq M^2$;
similarly for $(k + k_2)$.%
\footnote{Compare the discussion after \cite[Eq.~(3.8.24)]{Scharf89}.}
\paragraph{Scalar integrals $I_\pm$.}
We have to compute
\begin{align*}
I_\mp(k_1,k_2) := \frac{i}{(2\pi)^4}
& \int d^4k\, \th(\mp(k_1^0 - k^0)) \,\dl((k_1 - k)^2 - M^2)
\\
&\quad \x \frac{1}{k^2 - M^2 + i0}\, \th(\mp(k^0 + k_2^0))
\,\dl((k + k_2)^2 - M^2).
\end{align*}
Let us make a change of variable $q := k + k_2$, and introduce
$P := k_1 + k_2$, noting for later purposes that $P^2 = 2(k_1k_2)$.
One obtains the integral:
\begin{align}
\int d^4q\, \th(\mp(P^0 - q^0))\,\dl((P - q)^2 - M^2)\,
\frac{1}{(q - k_2)^2 - M^2 + i0}\, \th(\mp q^0) \,\dl(q^2 - M^2).
\label{eq:I-supp}
\end{align}
It follows that $I_\mp(k_1,k_2) \propto\th(\mp P^0) \,\th(P^2 - 4M^2)$,
and that $\sgn k_1^0 = \sgn k_2^0$ for $P^2 \geq 4M^2$.
Performing the $q^0$-integration and using the notation
$E_q := \sqrt{|\qq|^2 + M^2}$, we extract
\begin{align*}
I_\mp(k_1,k_2) &= \frac{i}{(2\pi)^4} \,\th(\mp P^0) \,\th(P^2 - 4M^2)
\\
&\quad \x
\int \frac{d^3q}{2E_q}\, \th(\mp(P^0 - q^0)) \,\dl((P - q)^2 - M^2)\,
\frac{1}{(q - k_2)^2 - M^2 + i0}\biggr|_{q^0=\mp E_q}.
\end{align*}
Since $P^2 > 0$, one may choose a particular Lorentz frame such that
\begin{equation}
P = (P^0,\zero); \word{hence} \kk_1 = - \kk_2, \quad
k_1^0 = \mp|\kk_1| = \mp|\kk_2| = k_2^0 = {\mathchoice{\thalf}{\thalf}{\shalf}{\shalf}} P^0.
\label{eq:P=(P0,0)}
\end{equation}
Taking into account $q^2 = M^2$, we observe that
$(P - q)^2 - M^2 = 2P^0({\mathchoice{\thalf}{\thalf}{\shalf}{\shalf}} P^0 - q^0)$, which yields
$$
\dl((P - q)^2 - M^2) = \frac{\dl(q^0 - {\mathchoice{\thalf}{\thalf}{\shalf}{\shalf}} P^0)}{2|P^0|}
= \frac{\dl(E_q - {\mathchoice{\thalf}{\thalf}{\shalf}{\shalf}}|P^0|)}{2|P^0|}\,,
$$
by using $q^0 = \mp E_q$. For later aims, we point out that in the
chosen frame this distribution implies $q^0 = k_2^0$; hence
\begin{equation}
kP = (q - k_2)P = (q^0 - k_2^0)P^0= 0.
\label{eq:q0=k20}
\end{equation}
From $\mp q^0 = \mp{\mathchoice{\thalf}{\thalf}{\shalf}{\shalf}} P^0$ comes $\mp(P^0 - q^0) = \mp{\mathchoice{\thalf}{\thalf}{\shalf}{\shalf}} P^0 >
0$. Therefore the factor $\th(\mp(P^0 - q^0))$ is redundant. Changing
the integration variables,
$$
\int d^3q \cdots = \int_M^\infty dE_q\, E_q\sqrt{E_q^2 - M^2} \int
d\Om_q \cdots\,,
$$
the $E_q$-integration can trivially be done, and we are left with:
\begin{align}
I_\mp(k_1,k_2) = i\,\th(\mp P^0)\, \th(P^2 - 4M^2)\,
\frac{\sqrt{(P^0)^2 - 4M^2}}{(2\pi)^4\,8|P^0|} \int\!\frac{d\Om_q}{(q
- k_2)^2 - M^2 + i0}\biggr|_{q^0=P^0/2}.
\label{eq:I1}
\end{align}
Let $\al$ be the angle between $\kk_2$ and $\qq$, and let $z :=
\cos\al$. Due to $q^2 = M^2$, $k_2^2 = 0$, $|\qq| = \sqrt{E_q^2 - M^2}
= {\mathchoice{\thalf}{\thalf}{\shalf}{\shalf}}\,\sqrt{P_0^2 - 4M^2}$ and relations \eqref{eq:P=(P0,0)}
and~\eqref{eq:q0=k20}, we obtain
\begin{equation}
(q - k_2)^2 - M^2 = -2(k_2 q) = -2(k_2^0 q^0 - |\qq|\.|\kk_2|\,z)
= \tfrac{a}{2}(-a + bz),
\label{eq:denominator}
\end{equation}
where
$$
a := |P^0| > 0, \quad 0 \leq b := \sqrt{(P^0)^2 - 4M^2} < a.
$$
We point out that $(-a + bz) < 0$ for all $z \in [-1,1]$: there is no
infrared problem in our triangle graph. The remaining $\Om_q$-integral
can be easily computed:
\begin{equation}
\frac{4\pi}{a} \int_{-1}^1 \frac{dz}{-a + bz} =
\frac{4\pi}{|P^0|\,\sqrt{(P^0)^2 - 4M^2}}\, \log\frac{(P^0)^2 -
|P^0|\,\sqrt{(P^0)^2 - 4M^2} - 2M^2}{2M^2}\,.
\label{eq:in-aeternum}
\end{equation}
To obtain the result in a generic Lorentz frame, replace
$(P^0)^2$ by $s := P^2 = 2(k_1k_2)$, so
\begin{align}
I_\mp(k_1,k_2) &= \frac{i\th(\mp P^0)\,\th(s - 4M^2)}{4(2\pi)^3\,s}
\log\biggl[ \frac{s - \sqrt{s(s - 4M^2)}}{2M^2} - 1 \biggr]
\notag
\\
&=: \th(\mp P^0)\,\th(s - 4M^2)F(s).
\label{eq:scalar-I}
\end{align}
The result for $J_\pm(k_1,k_2)$ can be read off from \eqref{eq:I1} by
omitting the Feynman propagator $i(2\pi)^{-2}\,((q - k_2)^2 - M^2 +
i0)^{-1}$. One obtains for the contribution of the $J$-integrals:
\begin{align*}
J_\pm(k_1,k_2)
= \frac{1}{8\pi}\,\th(\pm P^0)\,\th(s - 4M^2)\, \sqrt{1 - 4M^2/s}\,.
\end{align*}
\paragraph{Vector integrals $I^\mu_\mp\,$.}
For the same reasons as for the scalar integral, it must hold that
$I^\mu_\mp(k_1,k_2) \propto \th(\mp P^0)\,\th(s - 4M^2)$. From Lorentz
covariance and $I^\mu_\pm(k_1,k_2) = - I^\mu_\pm(k_2,k_1)$ it follows
$$
I^\mu_\mp(k_1,k_2)
= \th(\mp P^0)\, \th(s - 4M^2)\, (k_1^\mu - k_2^\mu)\,G(s)
$$
for appropriate $G(s)$. An immediate consequence is $I^\mu P_\mu = 0$.
To procure $G(s)$, compute
$$
k_{2,\mu}I^\mu_\mp(k_1,k_2) = \frac{1}{2} \th(\mp P^0)\, \th(s -
4M^2)\, s\,G(s) =\bigl(-i/8\,(2\pi)^3 \bigr)\,\th(\mp P^0)\, \th(s -
4M^2)\, \sqrt{1 - 4M^2/s}
$$
The second equality is obtained by comparing with the scalar integral:
there is an extra factor $(k_2 k) = (k_2 q) = -a(-a + bz)/4$, where
\eqref{eq:denominator} is used. Then the $\Om_q$-integral becomes
trivial. Thus we glean
\begin{align}
G(s) = \frac{-i}{32\,\pi^3\,s}\sqrt{1 - 4M^2/s}\,.
\label{eq:G}
\end{align}
\paragraph{Tensor integrals $I^{\mu\nu}_\mp$.}
Proceeding analogously to the vector integrals, one argues that
$$
I^{\mu\nu}_\mp(k_1,k_2) = \th(\mp P^0)\,\th(s - 4M^2)\,
\bigl[ (k_1^\mu k_1^\nu + k_2^\mu k_2^\nu)\,A(s)
+ (k_1^\mu k_2^\nu + k_2^\mu k_1^\nu)\,B(s) + g^{\mu\nu}\,C(s) \bigr].
$$
We need three independent identities to compute $A(s)$, $B(s)$
and~$C(s)$. A first one is:
\begin{align}
I^{\mu\nu}_\mp k_{2\mu} k_{2\nu}
&= \th(\mp P^0)\, \th(s - 4M^2)\, A(s)\, s^2/4
\notag \\
&= \th(\mp P^0)\, \th(s - 4M^2)\, \frac{-i}{2^5\,(2\pi)^3}\,
s\,\sqrt{1 - 4M^2/s}\,.
\label{eq:Ikk}
\end{align}
The second equality is obtained by a modification of the computation
of the scalar integral: there is the extra factor $(k_2 k)^2 = a^2(-a
+ bz)^2/16$. This yields $A(s)=G(s)/2$. A second identity is
given by the \textit{trace}. The result is again obtained by comparing
with the computation of the scalar integral: there is an additional
factor $k^2 = (q - k_2)^2 = M^2 - 2(k_2q) = M^2 - 2(k k_2)$, hence
$$
I^\mu_{\mp,\mu} = \th(\mp P^0)\,\th(s - 4M^2)\,(sB + 4C)
= M^2 I_\mp - 2k_{2,\mu} I^\mu_\mp.
$$
A third identity following from \eqref{eq:q0=k20} reads:
$$
I^{\mu\nu}_\mp P_\nu = \th(\mp P^0)\, \th(s - 4M^2)\, P^\mu\bigl((A +
B)s/2 + C\bigr) = 0.
$$
Pulling together these results, one arrives~at
$$
B(s) = - M^2 F(s)/s \word{and} C(s) = M^2 F(s)/2 - s\,G(s)/4.
$$
\marker
At this point we are able to show that the triangle plus fish-like
parts constitute a gauge-invariant quantity. For that, insert the
results already known for the integrals
into~\eqref{eq:suerte-o-verdad}, obtaining:
\begin{align}
& d^{\mu\nu}(k_1,k_2)\biggr|_{k_1^2=0=k_2^2}
= \frac{\sgn(P^0)\,\th(s - 4M^2)}{(2\pi)^2}\,\bigl[
k_1^\mu k_2^\nu [4G(s) - (1 + 4M^2/s)F(s)]
\notag
\\
&\quad + 2M^2 g^{\mu\nu} F(s)
- k_1^\nu k_2^\mu \frac{4M^2}{s}\, F(s) \bigr]
= \sgn(P^0)\,\th(s - 4M^2) \frac{4M^2}{(2\pi)^2}\,
P^{\mu\nu}\,\frac{F(s)}{s}.
\label{eq:d-onshell-0}
\end{align}
The $k_1^\mu k_2^\nu$-terms
have been dropped in the last identity, due to $k_\mu A^\mu(-k)=0$. The
remainder is electromagnetically gauge-invariant. Introducing the
dimensionless variable
$$
\tilde\rho := \frac{s}{4M^2} = \frac{P^2}{4M^2}\,,
$$
keeping in mind formula \eqref{eq:scalar-I}, and on use of
\eqref{eq:consilia-non-sentis}, equation \eqref{eq:d-onshell-0} can be
rewritten as
\begin{gather}
d^{\mu\nu}_\gi(k_1,k_2)\biggr|_{k_1^2=0=k_2^2}
:= \frac{i\,\sgn(P^0)\,\th(\tilde\rho - 1)}{(2\pi)^5}\,
P^{\mu\nu}\, b(\tilde\rho)
\label{eq:d0-h-h1}
\\
\shortintertext{with}
b(\tilde\rho)
:= \frac{1}{16\,M^2\,\tilde\rho^2}\,
\log\bigl( 2\tilde\rho - 2\sqrt{\tilde\rho(\tilde\rho - 1)} - 1 \bigr)
= - \frac{1}{16\,M^2\,\tilde\rho^2}\,
\log\frac{1 + \sqrt{1 - \tilde\rho^{-1}}}
{1 - \sqrt{1 - \tilde\rho^{-1}}} \,,
\notag
\end{gather}
where `gi' stands for the gauge invariant part. The singular order of
$d^{\mu\nu}_{\gi}\bigr|_{k_1^2=0=k_2^2}$ is $\om = -2$ by power
counting; whereas for the off-shell $d^{\mu\nu}(k_1,k_2)$ the value is
$\om = 0$.
\subsection{Regularity of absorptive parts in momentum space}
\label{ssc:do-ut-des}
This subsection is devoted to prove essential regularity properties of
the \textit{off-shell} $d$-distribution, more precisely of
$d^{\mu\nu}(k_1,k_2)$, for $(k_1,k_2) \in \sV := \bigl( \ovl V_+ \less
\{0\} \bigr)^{\x 2} \cup \bigl( \ovl V_- \less \{0\} \bigr)^{\x 2}$.
We look at the terms coming from \eqref{eq:suerte-o-verdad} by means
of \eqref{eq:def-J}. Introducing the new integration variable $q := -
k + {\mathchoice{\thalf}{\thalf}{\shalf}{\shalf}}(k_1 - k_2)$, the internal lines' momenta are
\begin{equation}
q_1 = q + {\mathchoice{\thalf}{\thalf}{\shalf}{\shalf}} P, \quad
q_2 = q - {\mathchoice{\thalf}{\thalf}{\shalf}{\shalf}} P, \quad
q_3 = q - {\mathchoice{\thalf}{\thalf}{\shalf}{\shalf}}(k_1 - k_2),
\label{eq:q123}
\end{equation}
and one sees that the considered terms are all of the type
\begin{align}
& H^{\mu\nu}(k_1,k_2)
\notag \\
&:= \int d^4 q\, \bigl(
\th(q_1^0)\, \th(-q_2^0) - \th(-q_1^0)\,\th(q_2^0) \bigr)
\,\dl(q_1^2 - M^2) \,\dl(q_2^2 - M^2)\,
\frac{h^{\mu\nu}(k_1,k_2,q)}{M^2 - q_3^2}
\label{eq:F_absorptive_gen}
\end{align}
for $(k_1,k_2) \in \sV_1 := \bigl(\ovl V \less \{0\}\bigr)^{\x 2}$
with $\ovl V := \ovl V_+ \cup \ovl V_-$, and where $h^{\mu\nu}\colon
\bR^{4\x 3} \to \bC\,$ is a polynomial of degree~$2$. We have used that
for $(k_1,k_2) \in \sV_1$ it holds true that
\begin{equation}
\int d^4 q\, \bigl( \th(q_1^0)\,\th(-q_2^0)
- \th(-q_1^0)\,\th(q_2^0) \bigr)
\,\dl(q_1^2 - M^2) \,\dl(q_2^2 - M^2) \,\dl(q_3^2 - M^2) = 0.
\label{eq:dldldl=0}
\end{equation}
This last relation can be argued as follows:%
\footnote{We borrow the standard notation for the mass shell:
$H_M^\pm := \set{p \in \bR^4 : p^2 = M^2,\ \pm p_0 > 0}$.}
$\!\!$the various $\th$- and $\dl$-distributions yield the
restrictions $(q_1,q_2) \in (H_M^+ \x H_M^-) \cup (H_M^- \x H_M^+)$
and $q_3 \in H_M^+ \cup H_M^-$; taking moreover into account that $q_3
= q_2 + k_2$ and $q_3 = q_1 - k_1$, it ensues that the various
restrictions on $q_3$ are not compatible.
The same identity implies that terms of the kind
$T_1(x_{\pi 1})\,T_1(x_{\pi 2})\,T_1(x_{\pi 3})$ do not contribute to
the third commutator in formula \eqref{eq:para-llorar} for~$D_3$ when
$(k_1,k_2) \in \sV_1$, for all permutations~$\pi$: the \textit{whole}
contribution to $d^{\mu\nu}(k_1,k_2)|_{(k_1,k_2)\in\sV_1}$ coming from
this commutator is of the kind~\eqref{eq:F_absorptive_gen}.
The contributions to $d^{\mu\nu}(k_1,k_2)|_{(k_1,k_2)\in \sV}$ coming
from the other two commutators in \eqref{eq:para-llorar} are of the
same form up to cyclic permutations
$k_1 \mapsto k_2 \mapsto -(k_1 + k_2) \mapsto k_1$ of the external
momenta. Here we use that $(k_1, k_2)\in \sV$ implies
$(k_2, -k_1 - k_2) \in \sV_1$ and $(-k_1 - k_2, k_1) \in \sV_1$, hence
we may apply the identity \eqref{eq:dldldl=0} also for the permuted
momenta. However, note that the polynomials $h_j^{\mu\nu}$, $j = 2,3$,
belonging to these other two cuts are not obtained by cyclic
permutations of the external momenta in the original polynomial
$h_1^{\mu\nu}$, meant in \eqref{eq:F_absorptive_gen}. This is due to
the difference between the higgs vertex and the photon vertices; in
particular, these other two cuts contain no term giving rise to a
fish-like diagram. Summing up, it holds that
\begin{equation}
d^{\mu\nu}(k_1, k_2)\bigr|_{(k_1,k_2)\in\sV}
= H^{\mu\nu}_1(k_1, k_2) + H^{\mu\nu}_2(k_2, -k_1 - k_2)
+ H^{\mu\nu}_3(-k_1 - k_2, k_1),
\label{eq:dV1}
\end{equation}
for some $H^{\mu\nu}_j$ $(j = 1,2,3)$ of the
form~\eqref{eq:F_absorptive_gen}, the pertinent polynomials
$h_j^{\mu\nu}$ being of degree~$2$.
\begin{lema}
\label{lem:continuity_absorptive_part}
Let $q_1,q_2,q_3$ and $\sV_1$ be defined as above in~\eqref{eq:q123}
and after~\eqref{eq:F_absorptive_gen}, and let $H^{\mu\nu} \colon
\bR^{4\x2} \to \bC$ be given in terms of a generic polynomial
$h^{\mu\nu} \colon \bR^{4\x3} \to \bC$ of degree $\zeta \in \bN_0$, as
in~\eqref{eq:F_absorptive_gen}. Then for all $(k_1,k_2) \in \sV_1$ and
for some $C > 0$ the function $H^{\mu\nu}$ is \textbf{continuous} in
the region~$\sV_1$, and can be bounded as follows:
\begin{equation}
|H^{\mu\nu}(k_1,k_2)|
\leq C\,\frac{(1 + |(k_1,k_2)|)^\zeta}{|(k_1k_2)|}\,
\th((k_1 + k_2)^2 - 4M^2)\, \log((k_1 + k_2)^2/M^2).
\label{eq:function_F_lemma_bound}
\end{equation}
Note that $|(k_1 k_2)| > 0$ if $(k_1,k_2) \in \sV_1$ and
$(k_1 + k_2)^2 \geq 4M^2$.
\end{lema}
\begin{proof}
Let $P := k_1 + k_2$ and $k := k_1 - k_2$. We first observe, on the
strength of
\begin{gather*}
q_1^2 - q_2^2 = 2(Pq), \quad
q_1^2 + q_2^2 - 2M^2 = 2(q^2 + \quarter P^2 - M^2)
\\
\word{and of\!\!} M^2 - q^3 = (M^2 - q^2 - \quarter P^2) + \quarter
P^2 - \quarter k^2 + (k q) \word{\!\!that}
\\
H^{\mu\nu}(k_1,k_2) \sim \sgn(P^0) \int d^4q\, \dl(q^2 + \quarter P^2
- M^2) \,\dl((Pq))\, \frac{h^{\mu\nu}(k_1,k_2,q)}{P^2/4 - k^2/4 +
(kq)}\,,
\end{gather*}
omitting irrelevant prefactors. Since $q_1 - q_2 = P$ and
$(q_1,q_2) \in (H_M^+ \x H_M^-) \cup (H_M^- \x H_M^+)$, we know that
$H^{\mu\nu}(k_1,k_2)$ vanishes for $P^2 < 4M^2$. Hence, to perform the
integrals in $q^0$ and~$|\qq|$ using the Dirac deltas, we may work in
the frame in which $\PP = 0$. There the $\dl$-distributions yield
$q^0 = 0$ and $|\qq| = \sqrt{P_0^2/4 - M^2}$.
With the notation $\hat p := \pp/|\pp|$ for $p \in \set{\!q,k\!}$, it
follows that
$$
\quarter P^2 - \quarter k^2 + (kq)
= (k_1k_2) \biggl( 1 - (\hat q\,\hat k) \sqrt{1 - 4M^2/P^2}\,
\frac{|P^0|\,|\kk|}{2(k_1 k_2)} \biggr),
$$
and one verifies that
\begin{equation}
0 \leq P_0^2\,|\kk/2|^2 = (k_1 k_2)^2 - k_1^2 k_2^2,
\label{eq:N}
\end{equation}
with $k_1 = (k_1^0,\kk_1)$ and $k_2 = (P^0 - k_1^0, -\kk_1)$. With the
help of these results we obtain
\begin{align}
(k_1 k_2)\, H^{\mu\nu}(k_1,k_2)
&\sim \sgn(P^0)\, \th(P^2 - 4M^2)\, \sqrt{1 - 4M^2/P^2}
\notag \\
&\enspace \x \int_{\bS^2} d\Om(\hat q)\,
\frac{h^{\mu\nu}\bigl(k_1, k_2, (0,\sqrt{P^2/4 - M^2}\,\hat q)\bigr)}
{1 - (\hat q\,\hat k)
\sqrt{(1 - 4M^2/P^2)\,(1 - k_1^2 k_2^2/(k_1 k_2)^2)}} \,,
\label{eq:H}
\end{align}
valid in the frame in which $\PP = \zero$. Let moreover $\sV_1^M :=
\set{(k_1,k_2) \in \sV_1 : (k_1 + k_2)^2 \geq 4M^2}$. We know that
\begin{gather*}
4M^2/P^2 \in (0,1] \word{and} k_1^2 k_2^2/(k_1 k_2)^2 \in [0,1]
\word{for} (k_1,k_2) \in \sV_1^M;
\\
\word{hence}
a := \sqrt{(1 - 4M^2/P^2)\,(1 - k_1^2 k_2^2/(k_1 k_2)^2)} \in [0,1).
\end{gather*}
In particular, the denominator in the integrand of~\eqref{eq:H} does
not vanish for $(k_1,k_2) \in \sV_1^M$. Since $\th(P^2 - 4M^2)\,
\sqrt{1 - 4M^2/P^2}$ is continuous, $H^{\mu\nu}$ \textit{is
continuous} on~$\sV_1$.
Observe now that for all $\hat q \in \bS^2$ the inequality
$$
\bigl| h^{\mu\nu}\bigl(
k_1, k_2, (0,\sqrt{P^2/4 - M^2} \hat q) \bigr) \bigr|
\leq \const (1 + |(k_1, k_2)|)^\zeta
$$
holds, with $|(k_1,k_2)|^2 := \sum_{j=0}^3 (k_{1j}^2 + k_{2j}^2)$.
Setting $z := \bigl(\hat q\hat k\bigr)$, the remaining integral is of
the type
$$
\int_{-1}^1 \frac{dz}{1 - az}
= \frac{1}{a} \log\biggl( \frac{1 + a}{1 - a} \biggr)
\leq 2(1 - \log(1 - a)),
$$
valid for $a \in [0,1)$. Using that $a \leq \sqrt{1 - 4M^2/P^2} \leq
(1 - 2M^2/P^2)$ and monotonicity of the logarithm, we see that
$$
-\log(1 - a) = \log \frac{1}{1 - a} \leq \log \frac{P^2}{2M^2} \,.
$$
Putting together the estimates, we end up with
\begin{equation}
|(k_1 k_2)\, H^{\mu\nu}(k_1,k_2)|
\leq \const \. \th(P^2 - 4M^2)\, (1 + |(k_1,k_2)|)^\zeta
\bigl( 1 + \log(P^2/2M^2) \bigr),
\label{eq:prove_continuity_absorptive}
\end{equation}
impliying \eqref{eq:function_F_lemma_bound}, since $1 + \log(P^2/2M^2)
< 2\,\log(P^2/M^2)$ for $P^2 \geq 4M^2$.
\end{proof}
The reader should keep in mind that $d^{\mu\nu}(k_1,k_2)$ is supported
outside a certain neighbourhood of the origin on momentum space --
have a look back at Eq.~\eqref{eq:H}.
\begin{corl}
\label{cor:d-regular}
The off-shell $d$-distribution $d^{\mu\nu}(k_1,k_2)$ given in
\eqref{eq:dV1} is continuous on~$\sV$ and fulfills the bound:
\begin{equation}
|d^{\mu\nu}(k_1,k_2)|
\leq \const\, \frac{(1 + |(k_1,k_2)|)^{\om+2}}{|(k_1k_2)|}\,
\log(2 + |(k_1,k_2)|/M) \text{ for all } (k_1,k_2) \in \sV.
\label{eq:bound_d}
\end{equation}
\end{corl}
\begin{proof}
Continuity follows immediately from Lemma
\ref{lem:continuity_absorptive_part}. For the bound \eqref{eq:bound_d}
we have substituted $\om + 2 \equiv \om(d) + 2$ for $\zeta$ of the
Lemma, since the singular order of $H_j^{\mu\nu}$ ($j = 1,2,3$) is
$\zeta - 2$ by power counting in \eqref{eq:F_absorptive_gen}. In
addition, for $H^{\mu\nu}_1(k_1,k_2)$ we have used that $(k_1 + k_2)^2
\leq 4\,|(k_1,k_2)|^2$, and in order to omit the $\th$-distribution we
have replaced $\log(2\,|(k_1,k_2)|/M)$ by $2\log(2 + |(k_1,k_2)|/M)$.
One deals analogously with $H^{\mu\nu}_2(k_2, -k_1 - k_2)$ and
$H^{\mu\nu}_3(-k_1 - k_2, k_1)$.
\end{proof}
\subsection{Distribution splitting by the dispersion integral for null
momenta}
\label{ssc:hasta-ahi-podiamos-llegar}
Recall that for $(k_1,k_2)\in V_\eta \x V_\eta$ the advanced part
$a^{\mu\nu}$ of $d^{\mu\nu}$ can be computed by the dispersion
integral \eqref{eq:hoist-with-retard}. Using the regularity properties
of $d^{\mu\nu}$ given in Corollary \ref{cor:d-regular}, we finally aim
to show that the limit $k_1^2 \downto 0$, $k_2^2 \downto 0$ in
\eqref{eq:hoist-with-retard} commutes with integration; that is, the
dispersion integral is also valid for $k_1^2 = 0 = k_2^2$. To
formulate the assertion, let
\begin{equation}
\sK := \set{(k_1,k_2) \in (\bR^4)^{\x 2}
: k_1^2, k_2^2 < 4M^2, \ (k_1 + k_2)^2 < 4M^2, \ (k_1 k_2)\neq 0}.
\label{eq:valet}
\end{equation}
Bearing in mind the factors $\th(q^2 - 4M^2)$ for $q \in \{k_1, k_2,
k_1 + k_2\}$ appearing in each term of $d^{\mu\nu}(k_1,k_2)$, we see
that for $(k_1,k_2) \in (V_\eta \x V_\eta) \cap \sK$, formula
\eqref{eq:hoist-with-retard} can be rewritten as:
\begin{equation}
a^{\mu\nu}(k_1,k_2) = \frac{i\eta}{2\pi} \int_{|t|\geq t_{\min}} dt\,
\frac{d^{\mu\nu}(tk_1, tk_2)}{t^{\om+1}(1 - t)},
\label{eq:dispersion_simplified}
\end{equation}
for some $t_{\min} > 1$ depending on $k_1,k_2$. Now, as discussed in
subsection \ref{sec:ipso-facto}, one knows $a^{\mu\nu}(k_1,k_2)$ to be
analytic on the region $\sK$. The Lebesgue dominated convergence
theorem~\cite[Th.~4.6.3]{RealKippa} with the bound~\eqref{eq:bound_d}
allows us conclude that \eqref{eq:dispersion_simplified} is a valid
identity for $(k_1,k_2) \in \sV \cap \sK$. Indeed, introducing the set
of limit points
$$
\sM := \sV \cap \sK \cap \{(k_1,k_2) \in \bR^8 : k_i^2 = 0\}
= \{(k_1,k_2) \in \bR^8 : k_i^2 = 0, \ 0 < (k_1 + k_2)^2 < 4M^2\},
$$
it is enough to observe that for any $(\tilde k_1,\tilde k_2) \in \sM$
-- implying $(\tilde k_1\,\tilde k_2) > 0$ and
$\tilde k_1^0 \tilde k_2^0 > 0$ -- there is a neighbourhood
$\sU_{(\tilde k_1,\tilde k_2)}$ such that
\begin{align*}
\biggl| \th(|t|-t_{\min}) \frac{d(tk_1,tk_2)}{t^{\om+1}(1-t)} \biggr|
&\leq \const \. \frac{\th(|t| - t_{\min})}{|t(1 - t)|}\,
\frac{\bigl( 1 + |(k_1,k_2)| \bigr)^{\om+2}}{|(k_1\,k_2)|}\,
\log(2 + |t|\,|(k_1,k_2)|/M)
\\
&\leq C\,\frac{\th(|t| - t_1)}{|t(1 - t)|}\, \log(2 + C_1\,|t|),
\end{align*}
for all $(k_1,k_2) \in (V_\eta \x V_\eta) \cap \sK
\cap \sU_{(\tilde k_1,\tilde k_2)}\,$, for some $C,C_1 > 0$ and some
$t_1 > 1$ independent of $(k_1,k_2)$. The function on the right hand
side is absolutely integrable in~$t$ -- here we see the reason for the
condition $(k_1k_2) \neq 0$ in~\eqref{eq:valet}.
\subsection{Normalization of the scalar model by distribution
splitting}
\label{ssc:splitting}
We must finally compute the gauge invariant part
$t^{\mu\nu}_\gi(k_1,k_2)$ for momenta lying on the set~$\sM$.
Considering the formula $T_{3} = A_{3} - A'_{3}$ and reckoning that
$a'^{\mu\nu}(k_1,k_2)|_{k_1^2=0=k_2^2}$ contains the factor $\th(P^2 -
4M^2)$ where $P := k_1 + k_2$, we see that on~$\sM$ its contribution
vanishes, that is $t^{\mu\nu} = a^{\mu\nu}$ there.
The upshot of the preceding two subsections is that we may compute
valid terms of the central solution $a^{\mu\nu}|_\sM \equiv
a^{c\,\mu\nu}|_\sM$ by inserting the on-shell amplitude
\eqref{eq:d-onshell-0} into the dispersion integral, with $\om$ the
singular order \textit{of the off-shell} $d^{\mu\nu}$, equal to~$0$ in
the present case.
Looking at \eqref{eq:d-onshell-0}, observe that a $k_r^\mu k_s^\nu$-
or $g^{\mu\nu}$-term of~$d^{\mu\nu}$ goes over to a $k_r^\mu k_s^\nu$-
or $g^{\mu\nu}$-term of~$a^{\mu\nu}$, respectively. Therefore, such
factors may be taken out of the dispersion integral. Since moreover
$P^{\mu\nu}(tk) = t^2\,P^{\mu\nu}(k)$, we see that the gauge invariant
part $a^{\mu\nu}_\gi$ can be obtained by inserting just the gauge
invariant part $d^{\mu\nu}_\gi$ in~\eqref{eq:d0-h-h1} into the
dispersion integral. The latter is of the form \eqref{eq:d-form}. So
we may use the version \eqref{eq:pro-reo} of the dispersion integral.
Lastly, $t^{\mu\nu}_\gi|_\sM = a^{\mu\nu}_\gi|_\sM$ is obtained from
\eqref{eq:pro-reo} by setting $\om = 0$ and substituting there
$b(u\tilde\rho)$ as given in~\eqref{eq:d0-h-h1} for
$f\bigl(u(k_1^2, k_2^2, (k_1 + k_2)^2)\bigr)$ -- in our case only
$(k_1 + k_2)^2 = s$ is present. Allowing for the dilation factor in
$P^{\mu\nu}$ this leads, for $(k_1,k_2) \in \sM$, to
\begin{align}
t^{\mu\nu}_\gi(k_1,k_2)
&= - \frac{P^{\mu\nu}}{(2\pi)^6} \int_{\tilde\rho^{-1}}^\infty du\,
\frac{u\,b(u\tilde\rho)}{u(1 - u)}
\label{eq:socratic}
\\
&= \frac{P^{\mu\nu}}{16 M^2\,(2\pi)^6} \int_{\tilde\rho^{-1}}^\infty
\frac{du}{\tilde\rho^2 u^2(1 - u)}\,
\log \frac{1 + \sqrt{1 - u^{-1}\tilde\rho^{-1}}}
{1 - \sqrt{1 - u^{-1}\tilde\rho^{-1}}}
= - \frac{P^{\mu\nu}}{8\,(2\pi)^6}\, \frac{J_2(\tilde\rho)}{M^2}\,,
\notag \\
\shortintertext{where}
2\,J_2(\tilde\rho)
&:= \int_1^\infty dv\, \frac{1}{(v - \tilde\rho)v^2}\,
\log \frac{1 + \sqrt{1 - v^{-1}}}{1 - \sqrt{1 - v^{-1}}},
\notag
\end{align}
after the change of integration variable $v := u \tilde\rho$.
Integrals like $J_2$ have been computed in \cite{BoraChristovaEberl}.
From Appendix~C of that reference:
\begin{equation}
J_1(\tilde\rho,a)
:= \frac{1}{2} \int_1^\infty dv\, \frac{1}{(v - \tilde\rho)(v - a)}\,
\log \frac{1 + \sqrt{1 - v^{-1}}}{1 - \sqrt{1 - v^{-1}}}
= \frac{f(\tilde\rho) - f(a)}{\tilde\rho - a}
\label{eq:J1}
\end{equation}
for $0 \leq \tilde\rho \leq 1$, $0 \leq a \leq 1$, where $f$ is the
distribution \eqref{eq:f1}. We infer that
\begin{align}
J_2(\tilde\rho) = \frac{\del}{\del a}\biggr|_{a=0}\, J_1(\tilde\rho,a)
= \frac{f(\tilde\rho)}{\tilde\rho^2} - \frac{1}{\tilde\rho}
\word{for} 0 \leq \tilde\rho \leq 1,
\label{eq:J2}
\end{align}
by bringing in the values $f(0) = 0$ and $f'(0) = 1$, which can be
read off from~\eqref{eq:arcsin-2}.
Summing up, the final result reads, as expected:
\begin{align}
t^{\mu\nu}_\gi(k_1,k_2)
= \frac{P^{\mu\nu}}{8\,(2\pi)^6}\, \frac{1}{M^2} \biggl(
\frac{1}{\tilde\rho} - \frac{f(\tilde\rho)}{\tilde\rho^2} \biggr)
= \frac{P^{\mu\nu}}{8\,(2\pi)^6}\, \frac{F_0(\tilde\rho)}{M^2}
\quad\text{for } (k_1,k_2) \in \sM,
\label{eq:F0}
\end{align}
where $F_0$ was given in~\eqref{eq:magister-dixit}. We conjecture that
this formula holds true for all $(k_1,k_2)$ satisfying
$k_1^2 = 0 = k_2^2$ and $(k_1 + k_2)^2 > 0$.
The reader should remember that \eqref{eq:F0} stands in principle for
just a member of a solution set. Since $\om = 0$, the general
Lorentz-invariant Epstein--Glaser solution is obtained by adding to
expression~\eqref{eq:F0} a term of the type $Cg^{\mu\nu}$ with
$C \in \bC$ arbitrary. But such a term with $C \neq 0$ would violate
EGI. Therefore we regard the above result as unique.
Recovering formula \eqref{eq:dAAh} and the factor $4ge^2M$, one ends
up with
\begin{align*}
\int dx_1\,dx_2\,dx_3\, T_3(x_1,x_2,x_3)
= \frac{g\al}{(2\pi)^3 M} \int dk_1\,dk_2\,
h(k_1 + k_2) A^\mu(-k_1) A^\rho(-k_2)\, P_{\mu\nu}\, F_0(\tilde\rho),
\end{align*}
which, on substituting $\rho$ for~$\tilde\rho$, that is, $m_h^2$ for
$s \equiv (k_1 + k_2)^2$, agrees with the
literature~\cite{HungryHunters}.
\begin{remk}
In the occasion an (unsubtracted) dispersion integral applied to
$b(u)$, performed in \cite[Eq.~(3.2)]{ChristovaI}, leads to the
\textit{same} integral \eqref{eq:socratic} and so the same correct
result. As the next section shows, this does not hold for the
higgs to diphoton decay via EW vector bosons.
\end{remk}
\section{Higgs to diphoton decay via EW vector bosons}
\label{sec:soberbia-pagana}
\subsection{Derivation of the quartic $AAWW^\7$-vertex in the unitary
gauge}
\label{ssc:per-angusta}
The amplitude in question in this paper describes an EW decay process
at third order in the coupling constant. Its structure is given by the
\textit{cubic} vertices in the first TOP~$T_1$ -- that is the
\textit{sole} ``empirical'' input. Here in going from $T_1$ to~$T_2$
we \textit{derive} the $AAWW^\7$-vertex which contributes by a
``fish-like'' diagram to the amplitude to be computed, see
Figure~\ref{fg:diagrams}.
The general idea is to examine the propagator which is to become the
internal line linking the di-photon in the one-loop, three-vertex
graph, and to obtain the one-loop, two-vertex graph from a
modification of that propagator, demanded by EGI -- by which here we
precisely understand invariance of the $\bS$-matrix under the
variations $A^\mu(x) \to A^\mu(x) + \del^\mu\La(x)$: interaction
dictates symmetry. The method is similar to the derivation of the
$AA\vf\vf^\7$ ``seagull'' vertex from the cubic coupling in scalar
QED, first performed in this way in~\cite{DKS93}.
The concept works on configuration space, as follows. Recall the
pertinent Hermitian vertex -- see for instance
\cite[Sect.~7.2.2]{Langacker}, explicitly referring to the unitary
gauge. With $G_{\mu\nu} := \del_\mu W_\nu - \del_\nu W_\mu$, one has:
\begin{align}
T_1(x_1) = ie[(W^\mu G^\7_{\mu\nu} - W^{\7\mu}G_{\mu\nu}) A^\nu
- W^\mu W_\nu^\7 F_\mu^{\,\nu}](x_1).
\label{eq:hic-iacet}
\end{align}
All indicated operator products are Wick products. We copy a second
vertex similar to~\eqref{eq:hic-iacet}:
\begin{align}
T_1(x_2) = ie[(W^\rho G^\7_{\rho\la} - W^{\7\rho} G_{\rho\la}) A^\la
- W^\rho W_\la^\7 F_\rho^{\,\la}](x_2),
\label{eq:lepus}
\end{align}
and proceed to make the contractions, according to the BEG method to
construct (the relevant sector of) the second-order $T_2(x_1,x_2)$ out
of $T_1(x_1),T_1(x_2)$. Now $T_2(x_1,x_2) = T_1(x_1)T_1(x_2)$ for
$x_1$ not to the past of~$x_2$, and $T_2(x_1,x_2) = T_1(x_2)T_1(x_1)$
for $x_2$ not to the past of~$x_1$.
In view of the triangle diagram given in Figure~\ref{fg:diagrams}, we
are only interested in contractions yielding a connected tree diagram
with the photons uncontracted. Once they are made, we analyze the
terms for which the resulting distributions are not uniquely defined
on the diagonal $x_1 = x_2$. The ambiguities in this extension are
eliminated by EGI as sole selection criterion.
From the last pair of terms in \eqref{eq:hic-iacet} and
\eqref{eq:lepus}, and with $\Dl^{\mu\bt}$ standing for the
Feynman--Proca propagator~\eqref{eq:cum-grano-salis} for the
$W$-bosons, there comes:
\begin{align}
& -e^2\, \bigl( F_{\mu\nu}(x_1) W^{\7\nu}(x_1)
\vev{\T W^\mu(x_1)\, W^{\7\bt}(x_2)} W^\al(x_2) F_{\al\bt}(x_2)
\notag \\
&\quad + F_{\mu\nu}(x_1) W^{\mu}(x_1)
\vev{\T W^{\7\nu}(x_1)\, W^\al(x_2)} W^{\7\bt}(x_2) F_{\al\bt}(x_2)
\bigr)
\notag \\
&= -e^2\, F_{\mu\nu}(x_1) F_{\al\bt}(x_2) W^{\7\nu}(x_1)
\,\Dl^{\mu\bt}(x_1 - x_2) W^\al(x_2) + (x_1 \otto x_2),
\notag \\
&\qquad \text{denoted $T_2^D(x_1,x_2)$ for later use;}
\label{eq:TD(1,2)}
\end{align}
where explicitly $\Dl_{\bt}^\mu = -\bigl( g_\bt^\mu +
\del^\mu\del_\bt/M^2 \bigr)\Delta^F$ with $\Delta^F$ the scalar
Feynman propagator~\eqref{eq:accidit-in-puncto}. Electromagnetic gauge
variations obviously do not affect~\eqref{eq:TD(1,2)}.
By pairing the last term in \eqref{eq:lepus} with $A$-type
terms in \eqref{eq:hic-iacet} and the last term in
\eqref{eq:hic-iacet} with $A$-type terms in~\eqref{eq:lepus} we obtain
eight terms, that we organize as follows:
\begin{align}
& e^2\, A^\mu(x_1) F_{\rho\bt}(x_2) \bigl[
G^\7_{\al\mu}(x_1) \,\Dl^{\al\bt}(x_1 - x_2)\, W^\rho(x_2)
- G_{\al\mu}(x_1) \,\Dl^{\al\rho}(x_1 - x_2)\, W^{\bt\7}(x_2) +{}
\notag \\
& W^\al(x_1) (-g^\rho_\mu \del_\al + g^\rho_\al \del_\mu)
\Delta^F(x_1{-}x_2) W^{\bt\7}(x_2)
- W^{\al\7}(x_1) (-g^\bt_\mu \del_\al + g^\bt_\al \del_\mu)
\Delta^F(x_1{-}x_2) W^{\rho}(x_2) \bigr]
\notag \\
& + (x_1 \otto x_2) =: T_2^B(x_1,x_2) + T_2^C(x_1,x_2),
\label{eq:TB(1,2)}
\end{align}
where $T_2^C$ denotes the $(x_1 \otto x_2)$-term. As noted in
App.~\ref{app:spin-one-basics}, third-order derivatives of~$\Delta^F$
cancel here of their own accord. In order to verify EGI in the above
expression, note first that it can only be violated on the diagonal
$x_1 = x_2$, that is by contact terms -- see the discussion on this in
Sect.~\ref{ssc:lost-in-translation}. Therefore in computing the
divergence of expressions like the above one selects only such terms
that under $\del_1^\mu$ produce local expressions. For instance, the
third term in~\eqref{eq:TB(1,2)} brings forth:
$$
W^\al(x_1)\, g^\rho_\al\,\del_1^\mu\del_\mu \,\Delta^F(x_1 - x_2)\,
W^{\bt\,\7}(x_2)
= -i W^\rho(x_1) W^{\7\bt}(x_1) \,\dl(x_1 - x_2) +\cdots
$$
where the dots stand for a term ${}\sim M^2 W \Delta^F W^\7$; but it
is not hard to see that it is cancelled by a similar one coming from
the next term. In conclusion: $T_2^B$ is individually
electromagnetically gauge-invariant, and $T_2^D$, $T_2^B$, $T_2^C$
calculated up to now have no bearing on the generation of the
$AAWW^\7$-vertex in constructing~$T_2$.
Still, we are left with the most interesting contractions to
calculate. By pairing the two first terms in~\eqref{eq:hic-iacet} with
the two first in~\eqref{eq:lepus} we get:
\begin{align}
& -e^2\,A^\mu(x_1) A^\rho(x_2) \bigl[
- W^\al(x_1) \wt D_{\al\mu\;\bt\rho}(x_1 - x_2) W^{\bt\7}(x_2)
- G^\7_{\al\mu}(x_1) \,\Dl^{\al\bt}(x_1 - x_2)\, G_{\bt\rho}(x_2) +{}
\notag \\
& G^\7_{\al\mu}(x_1) (g^\al_\rho \del_\bt - g^\al_\bt \del_\rho)
\Delta^F(x_1 - x_2) W^\bt(x_2)
+ W^{\al\7}(x_1) (-g^\bt_\mu \del_\al + g^\bt_\al \del_\mu)
\Delta^F(x_1 - x_2) G_{\bt\rho}(x_2) \bigr]
\notag \\
& + (x_1 \otto x_2) =: T_2^A(x_1,x_2).
\label{eq:TA(1,2)}
\end{align}
Outside the diagonal the distribution $\wt D_{\al\mu\;\bt\rho}$
necessarily coincides with $D_{\al\mu\;\bt\rho}$, defined in
Eq.~\eqref{eq:rarior-albo} as the propagator for the $G$-fields.
Following the Epstein--Glaser method, we now look for the most general
Lorentz-covariant extension of this distribution having the same
scaling degree. The solution reads:
\begin{equation}
\wt D_{\al\mu,\bt\rho}(y) = D_{\al\mu,\bt\rho}(y)
+ i(a\,g_{\al\bt} g_{\mu\rho} - b\,g_{\al\rho} g_{\mu\bt}) \,\dl(y)
\label{eq:ansatz-tildeD}
\end{equation}
with as yet unknown numbers $a,b \in \bC$. Note that the second term
in the above propagator generates a contact term eventually yielding
the $AAWW^\7$-vertex
\begin{equation}
T_F(x_1,x_2) := e^2\,A^\mu(x_1) A^\rho(x_2) W^\al(x_1) W^{\7\bt}(x_2)
\, i(a\,g_{\al\bt} g_{\mu\rho} - b\,g_{\al\rho} g_{\mu\bt})
\,\dl(x_1-x_2).
\label{eq:sic-transit}
\end{equation}
A third Lorentz tensor might appear in the Ansatz
\eqref{eq:ansatz-tildeD}, namely $g_{\al\mu} g_{\rho\bt} \,\dl(y)$.
However, on insertion into \eqref{eq:TA(1,2)}, one obtains the same
contribution as $g_{\al\rho}g_{\mu\bt}\,\dl(y)$, namely
$$
e^2\,A^\mu(x_1) A^\rho(x_1) W_\mu(x_1) W^{\7}_\rho(x_1)
\,\dl(x_1 - x_2).
$$
Since EGI of $T_2$ can be violated only by local terms, it will
suffice to select the terms which are $\sim (\del)\dl(x_1 - x_2)$
after taking the divergence $\del_1^\mu$. Those can be of two types:
\begin{enumerate}
\item
Either the contact term already contains a $\dl(x_1 - x_2)$; or
\item
Such terms are generated when the $\del^\mu_{x_1}$ acts on
$\del_\mu \Delta^F(x_1 - x_2)$ or on
$\del\del_\mu \Delta^F(x_1 - x_2)$, due to
$(\square + m^2)\Delta^F = -i\dl$.
\end{enumerate}
From the $W(x_1)\,W^\7(x_2)$ part in \eqref{eq:TA(1,2)}, we find:
\begin{itemize}
\item
The type~(a) contributions:
\begin{align}
& i \del^\mu W^\al(x_1)\,(-a\,g_{\al\bt} g_{\mu\rho}
+ b\,g_{\al\rho} g_{\mu\bt}) \,\dl(x_1 - x_2)\, W^{\bt\,\7}(x_2)
\notag \\
&\quad
+ i W^\al(x_1)\,\bigl( (-a\,g_{\al\bt} \del_\rho
+ b\,g_{\al\rho} \del_\bt)\dl \bigr)(x_1 - x_2)\, W^{\bt\,\7}(x_2)
\notag \\
&= i \bigl( -a\,\del_\rho W^\al(x_1) W^{\7}_\al(x_1)
+ b\,\del_\bt W_\rho(x_1) W^{\bt\7}(x_1) \bigr) \dl(x_1 - x_2)
\notag \\
&\quad
-ia\, W^\al(x_1) W^\7_\al(x_2) \,\del_\rho\dl(x_1 - x_2)
+ ib\, W_\rho(x_1 )W^{\bt\7}(x_2) \,\del_\bt\dl(x_1 - x_2),
\label{eq:WWa}
\end{align}
and the type~(b) contribution:
\begin{align}
& i W^\al(x_1)\,\bigl(
(g_{\al\bt} \del_\rho - g_{\al\rho} \del_\bt)\dl \bigr)(x_1 - x_2)\,
W^{\bt\,\7}(x_2)
\notag \\
&= i W^\al(x_1) W^\7_\al(x_2) \,\del_\rho\dl(x_1 - x_2)
- i W_\rho(x_1) W^{\bt\7}(x_2) \,\del_\bt\dl(x_1 - x_2).
\label{eq:WWb}
\end{align}
\item
From the $W^\7(x_1)\,W(x_2)$-part of the exchange term, we obtain
terms \eqref{eq:WWa} and~\eqref{eq:WWb} with $W \otto W^\7$
interchanged.
\item
From the $W^\7(x_1)\,G(x_2)$-part of the original term, we get the
type~(b) contribution:
\begin{align}
& i W^{\al\7}(x_1)\, g^\bt_\al \,\dl(x_1 - x_2)\, \bigl(
-\del_\bt W_\rho(x_2) + \del_\rho W_\bt(x_2) \bigr)
\notag \\
&= i \bigl( -W^{\al\7}(x_1)\,\del_\al W_\rho(x_1)
+ W^{\al\7}(x_1)\,\del_\rho W_\al(x_1) \bigr) \,\dl(x_1 - x_2).
\label{eq:WGb
\end{align}
\item
From the $W(x_1)\,G^\7(x_2)$-part of the exchange term, we obtain the
term~\eqref{eq:WGb} with $W \otto W^\7$ interchanged.
\end{itemize}
Summing up, the requirement is:
\begin{align}
0 &\overset{!}{=}
W^\al(x_1) W^{\7}_\al(x_2) \,\del_\rho\dl(x_1 - x_2)\, (1 - a)
+ W_\rho(x_1) W^{\bt\7}(x_2) \,\del_\bt\dl(x_1 - x_2)\, (b - 1)
\notag \\
&\quad
+ \del_\bt W_\rho(x_1) W^{\bt\7}(x_1) \,\dl(x_1 - x_2)\, (b - 1)
+ \del_\rho W^\al(x_1) W^\7_\al(x_1) \,\dl(x_1 - x_2)\, (1 - a)
\notag \\
&\quad + [W \otto W^\7].
\label{eq:gauinv1}
\end{align}
Generally, for two free fields $B(x)$, $C(x)$ it holds that the three
terms
$$
\del B(x_1)\,C(x_1) \,\dl(x_1 - x_2),\quad
B(x_1)\,\del C(x_1) \,\dl(x_1 - x_2) \word{and}
B(x_1)\,C(x_2) \,\del\dl(x_1 - x_2)
$$
are linearly independent. We conclude that condition
\eqref{eq:gauinv1} has a unique solution, namely $a = 1 = b$.
\textit{In fine}, the resulting contact term may be written as
\begin{equation}
\wt D_{\al\mu,\bt\rho}(y) - D_{\al\mu,\bt\rho}(y)
= \ihalf (2g_{\al\bt} g_{\mu\rho} - g_{\al\rho} g_{\mu\bt}
- g_{\mu\al} g_{\rho\bt}) \,\dl(y),
\label{eq:nisi-dominus-frustra}
\end{equation}
which reproduces the seagull $e^2 AAWW^\7$-vertex in the Feynman rules
for the EW interaction.
\begin{figure}[htb]
\begin{tikzpicture}[>=Stealth, scale=1.3]
\begin{scope}[xshift=-3cm]
\coordinate (V3) at (0,0) ; \coordinate (H) at (-1.3,0) ;
\coordinate (V1) at (1.7,1) ; \coordinate (V2) at (1.7,-1) ;
\coordinate (G1) at (3,1) ; \coordinate (G2) at (3,-1) ;
\draw[dashed] (H) node[left=3pt] {$H$}
-- (V3) node[above=2pt] {$x_3$} ;
\draw[photon] (V1) -- (G1) node[right=3pt] {$\ga$} ;
\draw[photon] (V2) -- (G2) node[right=3pt] {$\ga$} ;
\draw[wplus] (V3) -- (V1) node[pos=0.5, below right=-4pt] {$W^+$}
node[above left] {$x_1$} ;
\draw[wplus] (V1) -- (V2) node[pos=0.5, left=-3pt] {$W^+$}
node[below left] {$x_2$} ;
\draw[wplus] (V2) -- (V3) node[pos=0.5, above right=-4pt] {$W^+$} ;
\foreach \pt in {V1,V2,V3} \fill (\pt) circle (1.6pt) ;
\end{scope}
\begin{scope}[xshift=3cm]
\coordinate (V3) at (0,0) ; \coordinate (H) at (-1.3,0) ;
\coordinate (V4) at (2,0) ;
\coordinate (G1) at (3,1) ; \coordinate (G2) at (3,-1) ;
\draw[dashed] (H) node[left=3pt] {$H$} -- (V3) ;
\draw[photon] (V4) -- (G1) node[right=3pt] {$\ga$} ;
\draw[photon] (V4) -- (G2) node[right=3pt] {$\ga$} ;
\draw[wplus] (V3) arc[radius=1cm, start angle=180, end angle=0]
node[pos=0.5, below] {$W^+$} ;
\draw[wplus] (V4) arc[radius=1cm, start angle=0, end angle=-180]
node[pos=0.5, above] {$W^+$} ;
\foreach \pt in {V3,V4} \fill (\pt) circle (1.6pt) ;
\end{scope}
\end{tikzpicture}
\caption{Diagrams contributing to the amplitude to be computed}
\label{fg:diagrams}
\end{figure}
\marker
\textit{Recapitulation}: assembling $T_3$ from $T_1$ and $T_2$ by the
Epstein--Glaser method, the above derived $AAWW^\7$-vertex (which is
part of $T_2$) generates the fish-like diagram in
Figure~\ref{fg:diagrams}. In the Feynman gauge, the $AAWW^\7$-vertex
was already derived in \cite{DuetschS99} by the same method. The fact
that we had to modify only $D^{\8\8}_{\8\8}$ (i.e., the
$G$-propagator), and not $\Dl^\8_\8$ (i.e., the $W$-propagator),
although here both are of the same singular order $\om = 0$, is in
accordance with the corresponding modification in the Feynman gauge
or, more generally, in a $R_\xi$ gauge: in such gauges the $AAWW^\7$
term \textit{can be derived in the same way}, but only the
$G$-propagator may be modified, because only this propagator has
$\om = 0$. The $WG$ and $GW$-propagators have singular order
$\om = -1$ (see Eqs.~\eqref{eq:extrema-exquisita}
and~\eqref{eq:remedia-optima-sunt}) and are not to be modified.
\marker
\textit{In conclusion}: almost the first thing to learn when working
in the BEG scheme is that what is usually taken from a top-down,
mysterious prevalence in particle physics of classical non-Abelian
gauge theories, with their quartic, second-order couplings, is here
\textit{derived} by pure quantum field theory operations. To wit, the
inductive construction of the time-ordered products respecting
``gauge invariance'', in the sense of \cite{DKS93, DHKS94,
DuetschS99, AsteDS99, Scharf01}, order by order in the couplings. In
fact, \textbf{all} of the reductive Lie algebra structure of the
Standard Model interactions, up to including at least one scalar
particle, comes \textit{automatically} in the BEG formalism from
respecting first principles of quantum field theory in building
the $\bS$-matrix%
\footnote{The derivation of the reductive Lie algebra structure was
annunciated by Stora~\cite{Stora-ESI}. His kind of principled
bottom-up approach has long been neglected in textbooks. A refreshing
exception is \cite[Problem~9.3]{MDSchwartz}.}
-- without invoking unobservable mechanisms. To go further into this
subject here would take us too far afield. We have merely dealt with
the example necessary for our purposes.
\subsection{Computation of the causal distribution at third order}
\label{ssc:eius-est-nolle}
We wish to mention here that triangular graphs in Epstein--Glaser
theory have been examined, for instance, in \cite[Chap.~3.8]{Scharf14}
-- the vertex correction in QED; in \cite{Axial} -- anomalies; as well
as in~\cite{BDF09} and \cite[Sect.~3.2]{Duetsch19}.%
\footnote{Some of these works are very instructive, in that they show
that cherished invariance properties cannot always be preserved
under distribution splitting.}
Of course, here we shall have more terms and a forest of indices.
Proceeding similarly to Sect.~\ref{sec:argumentum}, we perform the
adiabatic limit already at this stage. Since all the propagators in
the loop are massive, we may use \eqref{eq:dAAh}; hence, we only have
to compute the restricted distribution
$d^{\mu\nu}(k_1,k_2)|_{k_1^2=0=k_2^2}$. Again, because of the
kinematic constraints, only the first cut in formula
\eqref{eq:para-llorar} for $D_3$ counts (including the ``fish graph''
as derived in the previous subsection). Moreover, the discussions and
general conclusions of
Sects.~\ref{sec:quam-scriptum},~\ref{ssc:do-ut-des} and
\ref{ssc:hasta-ahi-podiamos-llegar} hold here, and will be assumed
without further ado.
\medskip
The two $WWA$-vertices are given in \eqref{eq:hic-iacet} and
\eqref{eq:lepus}, and the $WWh$-vertex reads
$$
T_1(x_3) = gM\, h(x_3) W^\nu(x_3) W^\7_\nu(x_3).
$$
The first step is to compute $R'^{A|B|C|D|F}_3(x_1, x_2; x_3)
:= T_2^{A|B|C|D|F}(x_1, x_2)\, T_1(x_3) \bigr|_\triangle$, where the
lower-order TOPs $T_2^\8$ have been respectively given by
lexicographical order in~\eqref{eq:TA(1,2)}, \eqref{eq:TB(1,2)},
\eqref{eq:TD(1,2)}, and finally \eqref{eq:sic-transit} together
with~\eqref{eq:nisi-dominus-frustra}. As it happens, their sum is
symmetric in~$(x_1,x_2)$, so we just introduce a general factor~$2$
and need not mention exchange anymore.
We start with
\begin{align}
R'^D_3(x_1,x_2;x_3)
&= -2e^2gM\, h(x_3) F^{\mu\nu}(x_1) F^{\al\bt}(x_2)\,
r'^D_{\mu\nu,\al\bt}(y_1,y_2), \word{where}
\notag
\\
r'^D_{\mu\nu,\al\bt}(y_1,y_2)
&:= \Dl^{\ga\,+}_\nu(y_1) \,\Dl_{\mu\bt}(y_1 - y_2)
\,\Dl^+_{\al\ga}(y_2), \word{and} y_k := x_k - x_3.
\label{eq:R'D_3-x}
\end{align}
The ``fish'' diagram contribution reads:
\begin{align}
R'^F_3(x_1,x_2;x_3) &= -2e^2gM\, h(x_3) A^\mu(x_1) A^\nu(x_2)\,
r'^F_{\mu\nu}(y_1,y_2) \word{with}
\notag \\
r'^F_{\mu\nu}(y_1,y_2) &:= i\,\dl(y_1 - y_2)\, \bigl[
-g_{\mu\nu} \,\Dl^{\ga\,+}_\bt(y_1) \,\Dl^{\bt\,+}_\ga(y_1)
+ \Dl^{\ga\,+}_\nu(y_1) \,\Dl^+_{\mu\ga}(y_1) \bigr].
\label{eq:R'F_3-x}
\end{align}
Next we compute $R'^B_3$:
\begin{align}
R'^B_3(x_1,x_2;x_3) &= -2e^2gM\, h(x_3) A^\mu(x_1) F_{\nu\bt}(x_2)\,
{r'^B}_\mu^{\nu\bt}(y_1,y_2)
\notag \\
\text{with}\quad {r'^B}_\mu^{\nu\bt}(y_1,y_2)
&:= (g_\mu^\ga \del_\al - g^\ga_\al \del_\mu)\Dl^+(y_1)
\,\Dl^{\al\bt}(y_1 - y_2) \,\Dl^{\nu\,+}_\ga(y_2)
\notag \\
&\quad + \Dl^{\al\ga\,+}(y_1)\,
(g_\mu^\nu \del_\al - g^\nu_\al \del_\mu) \Delta^F(y_1 - y_2)
\,\Dl^{\bt\,+}_\ga(y_2).
\label{eq:R'B_3-x}
\end{align}
The relation $R'^C_3(x_1,x_2;x_3) := R'^B_3(x_2,x_1;x_3)$ obviously
holds. Finally, for $R'^A_3$, we collect
\begin{align}
R'^A_3(x_1,x_2;x_3)
&= -2e^2gM\, h(x_3) A^\mu(x_1) A^\nu(x_2)\, r'^A_{\mu\nu}(y_1,y_2),
\word{with}
\notag \\
r'^A_{\mu\nu}(y_1,y_2)
&:= - (g_\mu^\ga \del_\al - g^\ga_\al \del_\mu)\Dl^+(y_1)
\,\Dl^{\al\bt}(y_1 - y_2)\,
(g_{\nu\ga} \del_\bt - g_{\bt\ga} \del_\nu)\Dl^+(y_2)
\notag \\
&\quad - \Dl^{\al\ga\,+}(y_1) \,D_{\al\mu,\bt\nu}(y_1 - y_2)
\,\Dl^{\bt\,+}_\ga(y_2)
\notag \\
&\quad - (g_\mu^\ga \del_\al - g^\ga_\al \del_\mu)\Dl^+(y_1)\,
(g^\al_\nu \del_\bt - g^\al_\bt \del_\nu)\Delta^F(y_1 - y_2)
\,\Dl^{\bt\,+}_\ga(y_2)
\notag \\
&\quad + \Dl^{\al\ga\,+}(y_1)\,
(g_{\mu\bt} \del_\al - g_{\al\bt} \del_\mu)\Delta^F(y_1 - y_2)\,
(g_{\ga\nu} \del^\bt - g^\bt_\ga \del_\nu)\Dl^+(y_2).
\label{eq:R'A_3-x}
\end{align}
Next we express the resulting integrals by momentum space integrals.
By using $F^{\mu\nu}(k) = -i \bigl( k^\mu A^\nu(k) - k^\nu A^\mu(k)
\bigr)$ and $R'_3:=R'^A_3+R'^B_3+R'^C_3+R'^D_3+R'^F_3$ we gather
\begin{align}
&\int dx_1\,dx_2\,dx_3\, R'_3(x_1,x_2;x_3)
\label{eq:R'3-p}
\\
&\quad = -2e^2gM (2\pi)^2\int dk_1\,dk_2\, h(k_1 + k_2)
A^\mu(-k_1) {A^\nu}(-k_2)\, r'_{\mu\nu}(k_1,k_2),
\notag
\\
&\word{where} r'_{\mu\nu}(k_1,k_2) := r'^A_{\mu\nu}(k_1,k_2) +
r'^F_{\mu\nu}(k_1,k_2)
\label{eq:r'(k)}
\\
&+ \bigl[ ik_2^\bt (r'^B_{\mu,\bt\nu}(k_1,k_2) -
r'^B_{\mu,\nu\bt}(k_1,k_2)) + (k_1,\mu) \otto (k_2,\nu) \bigr]
\notag
\\
&- k_1^\bt k_2^\al \bigl[ r'^D_{\bt\mu,\al\nu}(k_1,k_2) -
r'^D_{\mu\bt,\al\nu}(k_1,k_2) - r'^D_{\bt\mu,\nu\al}(k_1,k_2) +
r'^D_{\mu\bt,\nu\al}(k_1,k_2) \bigr].
\notag
\end{align}
To compute the (combinations of) Fourier transformed
${r'}^{\dots}_{\dots}$-distributions appearing on the right hand side
of \eqref{eq:r'(k)}, we bring in $k_1^2 = 0 = k_2^2$, and omit all
terms containing a factor $k_{1\mu}$ or $k_{2\nu}$: this is allowed in
view of $k^\la A_\la(-k) = 0$. Similarly to the analogous computation
for the scalar model in the previous section, let us work with the
integration variable $q := k + k_2$, and introduce $P := k_1 + k_2$,
hence $k_1 - k = P - q$ and $s := P^2 = 2k_1k_2$ and $Pk_1 = k_1k_2 =
Pk_2$. Due to the factors $\Dl^+(q)$ and $\Dl^+(P - q)$, we may use
the relations
\begin{gather*}
q^2 = M^2, \quad (P - q)^2 = M^2,
\word{hence} 2Pq = s = 2(k_1k_2), \word{implying}
\\
(P - q)q = (k_1k_2) - M^2, \quad 0 = (P(q - k_2)) = (Pk), \quad
(q - k_2)^2 = M^2 - 2(qk_2),
\end{gather*}
and $((P - q)\,k_1) = (q k_2)$, that is, $(qk_1) + (qk_2) = (k_1k_2)$.
Remember also that we may replace $P_\mu \to k_{2\mu}$ and
$P_\nu \to k_{1\nu}$.
To tally the fish diagram contribution, we introduce the integrals
\begin{equation}
J^{\{\.|\mu|\mu\nu\}}(P)
:= \int d^4q\, \{1|q^\mu|q^\mu q^\nu\} \,\Dl^+(P - q)\,\Dl^+(q).
\label{eq:J-def}
\end{equation}
These symbols $J^\8(P)$ generalize that of the scalar $J(P)$, defined
in~\eqref{eq:def-J}. We obtain:
\begin{align}
r'^F_{\mu\nu}(k_1,k_2) &= \frac{i}{(2\pi)^4} \biggl( g_{\mu\nu} \Bigl(
-2 + \frac{s}{M^2} - \frac{s^2}{4M^4} \Bigr) J(P) - \frac{k_{2\mu}
k_{1\nu}}{M^2} J(P)
\notag
\\
&\hspace{3em} + \frac{1}{M^2} \Bigl( k_{2\mu} J_\nu(P) + k_{1\nu}
J_\mu(P)\,\frac{s}{2M^2} \Bigr) - J_{\mu\nu}(P)\Bigl( \frac{1}{M^2} +
\frac{s}{2M^4} \Bigr) \biggr).
\label{eq:r'F-p}
\end{align}
To figure out $J^\mu$ and $J^{\mu\nu}$ one first argues that
\begin{align}
J^\mu(P) &= \th(P_0)\,\th(s - 4M^2)\,P^\mu\,g(s),
\notag \\
J^{\mu\nu}(P)
&= \th(P_0)\,\th(s - 4M^2)\, (P^\mu P^\nu\,a(s) + g^{\mu\nu} c(s)),
\label{eq:J-1}
\end{align}
for appropriate $g(s), \,a(s)$ and $c(s)$. The latter can be obtained
from the identites
$$
J^\mu(P) P_\mu = {\mathchoice{\thalf}{\thalf}{\shalf}{\shalf}} sJ(P), \quad
J^\mu_\mu(P) = M^2 J(P), \quad
J^{\mu\nu}(P) P_\mu P_\nu = \quarter s^2J(P).
$$
and from $J(P)=\th(P^0)\, \th(s - 4M^2)\,4i\pi^2\,s\, G(s)$ -- see
\eqref{eq:G}. So we arrive at
$$
g(s) = 2i\pi^2 s\,G(s); \quad
a(s) = \frac{4i\pi^2}{3}\,(s - M^2) G(s); \quad
c(s) = \frac{4i\pi^2}{3}\, (M^2 - s/4) s\,G(s),
$$
yielding
\begin{align}
r'^F_{\mu\nu}(k_1,k_2) &= \frac{1}{4\pi^2}\, \th(P^0)\, \th(s -
4M^2)\, G(s) \biggl[ \Bigl( \frac{14}{3} - \frac{11}{6}\,\frac{s}{M^2}
+ \frac{5}{12}\,\frac{s^2}{M^4} \Bigr) g_{\mu\nu}(k_1 k_2)
\notag
\\
&\quad + k_{1\nu} k_{2\mu} \Bigl( -\frac{1}{3} +
\frac{2}{3}\,\frac{s}{M^2} - \frac{1}{12}\,\frac{s^2}{M^4} \Bigr)
\biggr].
\label{eq:r'F-final}
\end{align}
\paragraph{Other new integrals.}
We introduce already all the new integrals required in this section.
By means of $(q - k_2)^2 - M^2 = - 2(qk_2)$, implying
$(qk_2)\,\Delta^F(q - k_2) = - i/8\pi^2$, some of the integrals to
appear are calculated:
\begin{align*}
K^{\{\.|\mu|\mu\nu\}}(P)& := \int \! d^4q\, \{1|q^\mu|q^\mu q^\nu\}\,
(qk_2)\,\Dl^+(P - q) \Delta^F(q - k_2)\,\Dl^+(q)
= -\frac{i}{8\pi^2}\, J^{\{\.|\mu|\mu\nu\}}(P),\\
L(P) &:= \int \! d^4q\, (qk_2)^2 \,\Dl^+(P - q) \Delta^F(q - k_2)\,\Dl^+(q)
= k_{2\mu}\,K^\mu .
\end{align*}
The following integrals will also be needed:
\begin{align*}
N^{\{\.|\mu|\mu\nu\}}(P) &:= \int d^4q\,
\{1|q^\mu|q^\mu q^\nu\} \,\Dl^+(P - q) \Delta^F(q - k_2) \,\Dl^+(q).
\end{align*}
By using $k_2^\nu=0$, they are easily be expressed in terms of the
integrals computed in Sect.~\ref{ssc:d-scalar}:
$$
N(P) = I(P)\,,\qquad
N^\mu(P) = I^\mu(P) + k_2^\mu I(P)\,,\qquad
N^{\mu\nu}(P) = I^{\mu\nu}(P) + k_2^\mu I^\nu(P)\,,
$$
\paragraph{Electromagnetic gauge invariance.}
The computation of the individual terms in \eqref{eq:r'(k)} is
lengthy, but straightforward. For the $r'^B$-terms we reap
\begin{align*}
& ik_2^\bt \bigl(
r'^B_{\mu,\bt\nu}(k_1,k_2) - r'^B_{\mu,\nu\bt}(k_1,k_2) \bigr)
= \frac{1}{(2\pi)^2} \biggl\{
- 2N(P)(g_{\mu\nu} (k_1k_2) - k_{1\nu} k_{2\mu})
\notag\\
&\quad + g_{\mu\nu} (k_1k_2) \biggl(
2\frac{K(P)}{M^2} - \frac{L(P)}{M^4} \biggr)
- k_{1\nu} k_{2\mu} (k_1k_2)\,\frac{K(P)}{M^4}
+ N_{\mu\nu}(P) \biggl(
2\frac{(k_1k_2)}{M^2} - \frac{(k_1k_2)^2}{M^4} \biggr)
\notag\\
&\quad + k_{2\mu} \biggl( N_\nu(P) \Bigl( - 2\frac{(k_1k_2)}{M^2}
+ \frac{(k_1k_2)^2}{M^4}\, \Bigr)
+ K_\nu(P)\,\frac{(k_1k_2)}{M^4} \biggr)
+ k_{1\nu}\,K_\mu(P) \biggl( - \frac{2}{M^2}
+ \frac{k_1k_2}{M^4} \biggr) \biggr\}.
\end{align*}
Inserting the integrals calculated above, one reaches for $ik_2^\bt
\bigl( r'^B_{\mu,\bt\nu}(k_1,k_2) - r'^B_{\mu,\nu\bt}(k_1,k_2) \bigr)
+ (k_1,\mu) \otto (k_2,\nu)$ the following result:
\begin{align}
\frac{\th(P^0)\,\th(s - 4M^2)}{(2\pi)^2}\,
\bigl( g_{\mu\nu} (k_1k_2) - k_{1\nu} k_{2\mu} \bigr)\, \biggl(
F(s) \Bigl( -2 - \frac{s}{2M^2} \Bigr) + G(s)\,\frac{s}{M^2} \biggr).
\label{eq:rB-final}
\end{align}
Note that $ik_2^\bt \bigl( r'^B_{\mu,\bt\nu}(k_1,k_2)
- r'^B_{\mu,\nu\bt}(k_1,k_2) \bigr)$ is individually gauge invariant.
This reflects the known fact that $T_2^B$ given by~\eqref{eq:TB(1,2)}
is individually gauge invariant.
Analogously, since $T_2^D$ \eqref{eq:TD(1,2)} is trivially gauge
invariant, we expect the combination
$$
- k_1^\bt k_2^\al \bigl( r'^D_{\bt\mu,\al\nu}(k_1,k_2)
- r'^D_{\mu\bt,\al\nu}(k_1,k_2) - r'^D_{\bt\mu,\nu\al}(k_1,k_2)
+ r'^D_{\mu\bt,\nu\al}(k_1,k_2) \bigr),
$$
to contain the factor $g_{\mu\nu} (k_1k_2) - k_{1\nu} k_{2\mu}$.
Indeed, we obtain
\begin{align}
& \frac{1}{(2\pi)^2} \biggl\{
-2(g_{\mu\nu} (k_1k_2) - k_{1\nu} k_{2\mu})\, N(P)
+ g_{\mu\nu} \biggl( -2 \frac{L(P)}{M^2}
+ (k_1k_2) \Bigl( 2\frac{K(P)}{M^2} - \frac{L(P)}{M^4} \Bigr) \biggr)
\notag \\
&\quad + k_{1\nu} k_{2\mu}\, (k_1k_2) \,\frac{K(P)}{M^4}
+ N_{\mu\nu} \biggl( 2\frac{(k_1k_2)}{M^2}
+ \frac{(k_1k_2)^2}{M^4} \biggr)
+ k_{1\nu} \biggl( -2 \frac{K_{1\mu}(P)}{M^2}
- (k_1k_2)\,\frac{K_{\mu}(P)}{M^4} \biggr)
\notag
\\
&\quad + k_{2\mu} \biggl( 2\frac{K_{\nu}(P)}{M^2}
- 2\frac{(k_1k_2)}{M^2}\,N_\nu(P) - \frac{(k_1k_2)^2}{M^4}\,N_\nu(P)
+ (k_1k_2)\,\frac{K_{1\nu}(P)}{M^4} \biggr) \biggr\}
\notag \\
&= \frac{\th(P^0)\,\th(s - 4M^2)}{(2\pi)^2}\,
(g_{\mu\nu} (k_1k_2) - k_{1\nu} k_{2\mu}) \biggl[
\Bigl(-1 + \frac{s}{4M^2} \Bigr)F(s) - \frac{s^2}{4M^4}\,G(s) \biggr].
\label{eq:rD-final}
\end{align}
Finally, for~$r'^A$ we get
\begin{align*}
r'^A_{\mu\nu}(k_1,k_2)
&= \frac{1}{(2\pi)^2} \biggl\{
-2(g_{\mu\nu} (k_1k_2) - k_{1\nu} k_{2\mu})\,N(P)
\notag \\
&\quad + g_{\mu\nu} \biggl(
\Bigl( 2 + \frac{2(k_1k_2)}{M^2} \Bigr)\,K(P)
- \Bigl( \frac{2}{M^2} + \frac{(k_1k_2)}{M^4} \Bigr)\,L(P) \biggr)
\notag \\
&\quad + k_{1\nu} k_{2\mu} \Bigl( -\frac{(k_1k_2)}{M^4} \Bigr)\,K(P)
+ k_{1\nu} \Bigl( -\frac{2}{M^2} + \frac{(k_1k_2)}{M^4} \Bigr)\,K_\mu(P)
\notag \\
&\quad + k_{2\mu} \biggl( \Bigl( -12 + 2\frac{(k_1k_2)}{M^2}
- \frac{(k_1k_2)^2}{M^4} \Bigr)\,N_\nu(P)
+ \Bigl( \frac{4}{M^2} + \frac{(k_1k_2)}{M^4} \Bigr)\,K_\nu(P) \biggr)
\notag \\
&\quad + \Bigl( 12 - 2\frac{k_1k_2}{M^2}
+ \frac{(k_1k_2)^2}{M^4} \Bigr)\,N_{\mu\nu}(P)
- \Bigl( \frac{2}{M^2} + 2\frac{(k_1k_2)}{M^4} \Bigr)\,K_{\mu\nu}(P)
\biggr\}.
\end{align*}
After insertion of the integrals, this is equal to
\begin{align*}
&\frac{\th(P^0)\,\th(s - 4M^2)}{(2\pi)^2} \biggl\{
(g_{\mu\nu} (k_1k_2) - k_{1\nu} k_{2\mu})\,F(s)\,
\Bigl(-3 + 12\,\frac{M^2}{s} + \frac{s}{4M^2} \Bigr)
\\
&+ G(s) \biggl[ g_{\mu\nu} (k_1k_2) \Bigl( -\frac{14}{3}
+ \frac{5}{6}\,\frac{s}{M^2} - \frac{1}{6}\,\frac{s^2}{M^4} \Bigr)
+ k_{1\nu} k_{2\mu} \Bigl( \frac{1}{3} + \frac{s}{3M^2}
- \frac{s^2}{6M^4} \Bigr) \biggr] \biggr\}.
\end{align*}
The sum $r'^A + r'^F$ is indeed gauge invariant:
\begin{align*}
r'^A_{\mu\nu}(k_1,k_2) + r'^F_{\mu\nu}(k_1,k_2)
&= \frac{\th(P^0)\,\th(s - 4M^2)}{(2\pi)^2}\,
(g_{\mu\nu} (k_1k_2) - k_{1\nu} k_{2\mu})
\\
&\quad \x \biggl( F(s)\,\Bigl( -3 + 12\,\frac{M^2}{s}
+ \frac{s}{4M^2} \Bigr) + G(s)\,\Bigl( - \frac{s}{M^2}
+ \frac{s^2}{4M^4} \Bigr) \biggr),
\end{align*}
as expected from the outcome of the discussion in subsection
\ref{ssc:per-angusta}.
The $G(s)$-terms are seen to cancel out, and for the total $r'$ we
obtain the following gauge-invariant result:
\begin{equation}
r'_{\mu\nu}(k_1,k_2) = \frac{\th(P^0)\,\th(s - 4M^2)}{(2\pi)^2}\,
(g_{\mu\nu} (k_1k_2) - k_{1\nu} k_{2\mu})\,6 F(s)\,(-1 + 2M^2/s).
\label{eq:r'-final}
\end{equation}
Like for the scalar model, $a'_{\mu\nu}(k_1,k_2)$ is obtained from
$r'_{\mu\nu}(k_1,k_2)$ in~\eqref{eq:r'-final} by replacing $\th(P^0)$
by $\th(-P^0)$, and hence $d_{\mu\nu} = a'_{\mu\nu} - r'_{\mu\nu}$ by
replacing $\th(P^0)$ by~$-\sgn(P^0)$.
Hence $d^{\mu\nu}_\gi(k_1, k_2)|_{k_1^2=0=k_2^2}$ is of the form
\begin{align}
&d^{\mu\nu}_\gi(k_1,k_2)\bigr|_{k_1^2=0=k_2^2}
= P^{\mu\nu}(k_1,k_2)\,d_0(P), \word{where}
\label{eq:d-onshell-a}
\\
&d_0(P) := \frac{i\,\sgn(P^0)\,\th(\tilde\rho - 1)}{(2\pi)^5}\,
b_1(\tilde\rho), \quad
b_1(\tilde\rho) := -\frac{3}{16 M^2\,\tilde\rho}\,
\biggl( \frac{1}{\tilde\rho} - 2 \biggr)\,
\log\frac{1 + \sqrt{1 - \tilde\rho^{-1}}}
{1 - \sqrt{1 - \tilde\rho^{-1}}}\,.
\notag
\end{align}
The factor $3$ above was to be expected, since the Proca field has
three components. The singular order of the on-shell distribution
$d^{\mu\nu}_\gi(k_1,k_2)|_{k_1^2=0=k_2^2}$ is clearly equal to zero.
It is also easy to show that power counting rules imply that the
singular order of the \textit{off-shell} $d$-distribution%
\footnote{This notation is badly abused in this paper. But that is
hardly avoidable.}
$d^{\mu\nu}$ satisfies the bounds
$6 \geq \om := \om(d^{\mu\nu}) \geq 2$.
\subsection{Distribution splitting in EW theory}
\label{ssc:quod-feceris}
The off-shell $d^{\mu\nu}$-distribution for the diphoton decay of the
higgs via EW vector bosons is also of the genre~\eqref{eq:dV1}, the
difference being that the pertinent polynomials $h_j^{\mu\nu}$ are of
a higher degree, i.e., $\om := \om(d^{\mu\nu})$ has a greater value.
Therefore, the distribution splitting method for null momenta
developed in Section~\ref{sec:quam-scriptum} does apply:
$t^{\mu\nu}_\gi|_\sM = a^{c\,\mu\nu}_\gi|_\sM$ can be computed by
inserting the light-cone restriction of $d^{\mu\nu}_\gi(k_1,k_2)$ into
the dispersion integral~\eqref{eq:dispersion_simplified}. Again, this
restricted $d$-distribution is of the genre~\eqref{eq:d0-h-h1} for the
given~$b_1$. Hence, the \textit{central solution} compatible with EGI
is obtained by:
\begin{align}
& t^{c\,\mu\nu}_\gi(k_1,k_2) = P^{\mu\nu}(k_1,k_2)\,t^c_\gi(P),
\word{where}
\notag \\
& t^c_\gi(P) = \frac{i\eta}{2\pi} \int_{|t|\geq\sqrt{1/\tilde\rho}}\,
\frac{t^2\,d_0(tP)\,dt}{(1 - t)\,t^{\max\{\om+1,\,0\}}}
= -\frac{1}{(2\pi)^6} \int_{1/\tilde\rho}^\infty
\frac{u\,b_1(u\tilde\rho)\,du} {u^{\max\{\piso{\om/2}+1,\,0\}}(1-u)},
\label{eq:t-dispint}
\end{align}
for $(k_1,k_2) \in \sM$. Let us first assume that $\om =2$. Before
computing, we recall that the \textit{ambiguity} of the result for
$t^{\mu\nu}_\gi$ will be given in that case by a polynomial of
\textit{second} degree in~$(k_1,k_2)$ containing the factor
$P^{\mu\nu}$; that is, in the occasion, a constant multiple
of~$P^{\mu\nu}$ itself.
Besides a global factor $-3(16M^2 (2\pi)^6)^{-1}$, in view of
\eqref{eq:d-onshell-a} and making the customary change of variable
$v=u\tilde\rho$, to obtain $t^c_\gi(\tilde\rho)$ we ought to compute:
\begin{align*}
&\tilde\rho\int_1^\infty\!dv\,\Biggl[
\frac{\log\bigl( 1 + \sqrt{1 - v^{-1}} \bigr)
/ \bigl( 1 - \sqrt{1 - v^{-1}} \bigr)}{v^3(\tilde\rho - v)}
- \frac{2 \log\bigl( 1 + \sqrt{1 - v^{-1}} \bigr)
/ \bigl( 1 - \sqrt{1 - v^{-1}} \bigr)}{v^2(\tilde\rho - v)} \Biggr]
\\
&= \int_1^\infty\! dv\, \Biggl[
\frac{\log\bigl( 1 + \sqrt{1 - v^{-1}} \bigr)
/ \bigl( 1 - \sqrt{1 - v^{-1}} \bigr)}{v^2(\tilde\rho - v)}
- \frac{2 \log\bigl( 1 + \sqrt{1 - v^{-1}} \bigr)
/ \bigl( 1 - \sqrt{1 - v^{-1}} \bigr)}{v(\tilde\rho - v)} \Biggr]
\\
&+ \int_1^\infty\! dv\, \Biggl[
\frac{\log\bigl( 1 + \sqrt{1 - v^{-1}} \bigr)
/ \bigl( 1 - \sqrt{1 - v^{-1}} \bigr)}{v^3}
- \frac{2 \log\bigl( 1 + \sqrt{1 - v^{-1}} \bigr)
/ \bigl( 1 - \sqrt{1 - v^{-1}} \bigr)}{v^2} \Biggr].
\end{align*}
Notice that the last two integral terms just yield a number, equal to
minus the sum of the values of the two previous ones at
$\tilde\rho = 0$. The first integral term is already known from the
discussion in subsection~\ref{ssc:splitting}, and in all yields
$$
\frac{3}{8M^2(2\pi)^6} J_2(\tilde\rho)
= \frac{-3}{8M^2(2\pi)^6} \biggl(
\frac{1}{\tilde\rho} - \frac{f(\tilde\rho)}{\tilde\rho^2} \biggr).
$$
This is essentially $3F_0(\tilde\rho)$, that clearly contributes to
the expected final result for $F_1(\tilde\rho)$ -- recall
\eqref{eq:magister-dixit}. According to Eq.~\eqref{eq:J1} -- also in
subsection~\ref{ssc:splitting} -- in the case $a=0$, the second
integral term yields
$$
- \frac{2\cdot 3}{8M^2(2\pi)^6}\,J_1(\tilde\rho,0)=
- \frac{6}{8M^2(2\pi)^6}\, \frac{f(\tilde\rho)}{\tilde\rho}\,.
$$
Including the last two integrals, we recover
$t^{c}_{\gi}(\tilde\rho)\sim F_1(\tilde\rho) - 7$. Taking into account
the Epstein--Glaser ambiguity as limited by EGI, and bringing in
constant prefactor appearing in~\eqref{eq:R'3-p}, we end up with the
amplitude
\begin{equation}
- 2e^2gM (2\pi)^2\,t^{\mu\nu}_{\gi}(k_1,k_2)
= \frac{e^2g}{4(2\pi)^4\,M}\, P^{\mu\nu}(k_1,k_2)\,
\bigl( F_1(\tilde\rho) + C \bigr),
\label{eq:latapadelperol}
\end{equation}
with $C$ an arbitrary constant.
If $\om > 2$, the central solution within the unitary gauge is
obtained by adding the following expression to the above obtained
result, more precisely, to $-(8M^2(2\pi)^6)^{-1} \bigl(
F_1(\tilde\rho) - 7 \bigr)$:
\begin{align}
& -\frac{1}{(2\pi)^6} \int_{\tilde\rho^{-1}}^\infty du\,
\frac{b_1(u\tilde\rho)}{1 - u}
\biggl( \frac{1}{u^{\piso{\om/2}}} - \frac{1}{u} \biggr)
= -\frac{1}{(2\pi)^6} \int_{\tilde\rho^{-1}}^\infty du\,
b_1(u\tilde\rho) \sum_{k=0}^{\piso{\om/2}-2}
\frac{1}{u^{\piso{\om/2} - k}}
\nonumber
\\
&\quad = \sum_{k=0}^{\piso{\om/2}-2} \tilde c_k\,
\tilde\rho^{\piso{\om/2}-1-k}; \word{where the}
\tilde c_k := -\frac{1}{2\pi} \int_1^\infty dv\,
\frac{b_1(v)}{v^{\piso{\om/2}-k}}
\label{eq:t-t0}
\end{align}
do not depend on $\tilde\rho$; in the first step we have used the
relation
$$
\frac{1}{u^{\piso{\om/2}}} - \frac{1}{u}
= (1 - u) \sum_{k=0}^{\piso{\om/2}-2} \frac{1}{u^{\piso{\om/2}-k}}\,,
$$
and in the last step we have made again the substitution $v :=
u\tilde\rho$. The point is that \eqref{eq:t-t0} is a polynomial in
$\tilde\rho$ allowed by the surviving Epstein--Glaser ambiguity.
Therefore, without knowing the precise value of $\om$, the general
Epstein--Glaser solution respecting EGI can be written as
\begin{equation}
- 2e^2gM (2\pi)^2\,t^{\mu\nu}_\gi(k_1,k_2)
= \frac{e^2g}{4(2\pi)^4\,M}\, P^{\mu\nu}(k_1,k_2)\, \biggl(
F_1(\tilde\rho) + \sum_{k=0}^{\piso{\om/2}-1} C_k\tilde\rho^k \biggr),
\label{eq:gratatio-capitis}
\end{equation}
where the constants $C_k$ are all arbitrary. However, terms
corresponding to $\om\geq 4$ can be discarded on grounds of
perturbative unitarity.
\medskip
Let us now to come back to reference~\cite{ChristovaI}. It is
argued there that the convergent integral
$$
t_0(\tilde\rho) := \int_{\tilde\rho^{-1}}^\infty du\,
\frac{b_1(u\tilde\rho)}{1 - u}
$$
leads to the correct result. From the standpoint of this reference,
formula~\eqref{eq:t-dispint} is ``oversubtracted''. One obtains there,
yet again
\begin{align}
P^{\mu\nu} t_0(\tilde\rho)
&= P^{\mu\nu} \! \int_{\tilde\rho^{-1}}^\infty du\,
\frac{b_1(u\tilde\rho)}{1 - u}
= - \frac{3\,P^{\mu\nu}}{8M^2(2\pi)^6}\,
(2 J_1(\tilde\rho,0) - J_2(\tilde\rho))
\nonumber \\
&= - \frac{P^{\mu\nu}}{8M^2(2\pi)^6} \,\biggl[
\frac{3}{\tilde\rho} + \frac{6f(\tilde\rho)}{\tilde\rho}
- \frac{3f(\tilde\rho)}{\tilde\rho^2} \biggr]
= - \frac{P^{\mu\nu}}{8M^2(2\pi)^6}\,(F_1(\tilde\rho) - 2).
\label{eq:disp-int-solu}
\end{align}
So the naive on-shell computation yields a \textit{particular}
Epstein--Glaser solution. In the present case, however,
equation~\eqref{eq:gratatio-capitis} tells us that we are
\textit{forced} to add (at least) a polynomial of degree two
respecting EGI, that is, a term $C P^{\mu\nu}$ for~$C$ an
indeterminate constant -- with which our result for the amplitude is
compatible with the generally accepted one.
\begin{remk}
From our viewpoint, the expression in \eqref{eq:disp-int-solu} is the
unique Epstein--Glaser solution respecting EGI, corresponding to the
following causal $d$-distribution: let the result~\eqref{eq:d-onshell-a} for
$d^{\mu\nu}_\gi(k_1,k_2)$ (obtained by light-cone restriction of the
photon momenta) be \textit{interpreted as an unrestricted element of}
$\sS'(\bR^8)$, that is, all values $(k_1,k_2) \in \bR^8$ are admitted.
One easily verifies that this $d$-distribution has causal support, so
the splitting problem is well defined, and since its singular order is
zero, the EGI requirement selects a unique splitting solution. Writing
the latter suitably as a dispersion integral in momentum space, one
verifies the claim. This procedure strongly simplifies explicit
computations, but it is not conceptually correct.%
\footnote{Actually, in the first edition of the book by Scharf on
quantum electrodynamics (i.e.,~\cite{Scharf89} rather than
\cite{Scharf14}), the vertex function in QED at third order was
computed by such a method.}
\end{remk}
\subsection{Fixing the normalization polynomial by agreement with the
Feynman gauge}
\label{ssc:ita}
In order to determine the normalization polynomial we may as well
invoke the computation of the $h \to \ga\ga$ decay in the Feynman
gauge and \textit{gauge-fixing independence}, namely, the requirement
that observable quantities should not depend on the choice of gauge.%
\footnote{The equivalence or inequivalence of calculations performed
in different gauges was a nagging worry of Raymond Stora in his last
years. The classic paper~\cite{ChristTDLee} illustrates the
difficulties lurking here.}
Motivated by results of \cite{AsteSD98},%
\footnote{This reference works with a formulation of gauge invariance
suitable for the BEG scheme. In that framework it was shown for
the various $R_\xi$-gauges that the $T$-products can be normalized in
such a way that the physical $S$-matrix (i.e., for in- and out-states
being on-shell) does not depend on the gauge-fixing parameter~$\xi$ in
the formal adiabatic limit; and that this normalization is compatible with
gauge invariance in the mentioned sense.}
we contend that the ``entirely on-shell'' amplitude coming out of our
previous computation should coincide with that of an Epstein--Glaser
computation in the Feynman gauge. By ``entirely on-shell'' we mean
that not only the photons, but \textit{also the higgs} is on-shell,
that is, $\tilde\rho = \rho := m_h^2/4M^2$.
Denote the Epstein--Glaser result for the $d$-distribution in the
Feynman gauge by $d^1_{\mu\nu}$. In contrast with the unitary gauge,
there additionally contribute diagrams with St\"uckelberg fields and
Faddeev--Popov ghosts (as inner lines) to $d^1_{\mu\nu}$, see
e.g.~\cite{WusStrikeBack}. We spare the reader the details of the
construction of the TOPs, and in particular the derivation of the
$AAWW^\7$-vertex in this context. For photons on-shell with physical
polarizations (setting $k_1^2 = 0 = k_2^2$ and omitting pure gauge
terms $\sim k_{1\mu}$ or $\sim k_{2\nu}$), our result reads:
\begin{align}
&d_{\mu\nu}^1(k_1,k_2) = -\frac{1}{2^3\,(2\pi)^6\,M^2}
\label{eq:absorptive_r_xi}\x
\\
&\bigg[ (k_1 k_2)g_{\mu\nu} \left(-\frac{3}{\tilde\rho^2}
+\frac{7}{\tilde\rho} -\frac{\rho}{\tilde\rho^2} \right) -k_{2\mu}
k_{1\nu} \left(-\frac{3}{\tilde\rho^2}+\frac{8}{\tilde\rho} -
\frac{2\rho}{\tilde\rho^2} \right) \bigg]\Im f(\tilde\rho).
\notag
\end{align}
The tedious computation of the above absorptive part was done with the
aid of the Mathematica package FeynCalc~\cite{feyncalc}. The
computation proceeds along the lines of the computations of related
absorptive parts in scalar electrodynamics and electroweak theory in
the unitary gauge presented in full detail in Sec.~\ref{ssc:d-scalar}
and \ref{ssc:eius-est-nolle}, respectively. As before, all terms
contributing to the distribution $d_{\mu\nu}^1$ can be represented by
Feynman diagrams with cuts -- for the complete list see e.g.
\cite{WusStrikeBack}. Just like in subsections.~\ref{ssc:d-scalar}
and~\ref{ssc:eius-est-nolle}, because of the kinematic constraints one
needs to consider only the cut separating the higgs vertex from the
photon vertices. All the appearing expressions have a very similar
structure to those that have been already considered in the
above-mentioned parts. Thanks to the presence of the cut, each
integral over the four-momentum flowing in the loop can be converted
into an integral over a sphere, which can be evaluated explicitly. We
stress the fact that, due to compactness of the region of integration,
the computation of the absorptive part does not involve any
regularization.
An important feature of electroweak theory in the $R_\xi$-gauges is
the fact that all interaction vertices have dimensions lower or equal
to four (because $\mathrm{dim}\,W^\mu=1$, in contrast to the value
$\mathrm{dim}\,W^\mu=2$ for the unitary gauge). In particular, a
straightforward power counting argument gives the upper bound
$\om\bigl(d^1_{\mu\nu}\bigr)\leq0$ for the singular order of the
off-shell distribution $d_{\mu\nu}^1$. Noting that $\tilde\rho=(k_1
k_2)/2M^2$ and $f(\tilde\rho)= O(\log\tilde\rho)$ we see that the
on-shell restriction of $d^1_{\mu\nu}(k_1,k_2)$ in
Eq.~\eqref{eq:absorptive_r_xi} grows logarithmically for big values of
$\tilde\rho$. For the off-shell~$d_{\mu\nu}^1$, this implies the
equality $\om\bigl(d^1_{\mu\nu} \bigr)=0$. This should be contrasted
with the bounds $6\geq\om\bigl(d_{\mu\nu} \bigr) \geq 2$ in the case
of the absorptive part computed in the unitary gauge.
The off-shell distribution $d_{\mu\nu}^1$ is again of the type
considered in Sec.~\ref{ssc:do-ut-des}. In particular, the method of
distribution splitting developed in subsection
\ref{ssc:hasta-ahi-podiamos-llegar} is applicable. For photons
on-shell with physical polarizations, the central solution reads
\begin{align}
&t^{1\,c}_{\mu\nu}(k_1,k_2)
= -\frac{1}{2^3\,(2\pi)^6\,M^2}\,\x
\label{eq:txi-on}
\\
&\biggl\{g_{\mu\nu} (k_1k_2) \biggl[ \Bigl(
-\frac{3}{\tilde\rho^2} + \frac{7}{\tilde\rho} -
\frac{\rho}{\tilde\rho^2}\Bigr) f(\tilde\rho) + \frac{3}{\tilde\rho} +
\frac{2\rho}{\tilde\rho} \biggr] - k_{1\nu} k_{2\mu} \biggl[ \Bigl(
-\frac{3}{\tilde\rho^2} + \frac{8}{\tilde\rho} -
\frac{2\rho}{\tilde\rho^2} \Bigr) f(\tilde\rho) + \frac{3}{\tilde\rho}
+ \frac{2\rho}{\tilde\rho} \biggr] \biggr\}.
\notag
\end{align}
According to the postulate `Divergence degree' (in Sect.
\ref{ssc:porca-miseria}), we have to demand for the off-shell
$t^1_{\mu\nu}$ that
\begin{equation*}
\om\bigl(t^1_{\mu\nu} \bigr) = \om\bigl(d^1_{\mu\nu}\bigr) = 0.
\end{equation*}
This implies that the pertaining normalization freedom consists of a
constant term which is a tensor with two indices. By the Lorentz
invariance such a term has to be proportional to the metric tensor.
Consequently, the general off-shell solution of the splitting problem
is of the form
\begin{equation}
t^{1}_{\mu\nu}(k_1,k_2) = t^{1\,c}_{\mu\nu}(k_1,k_2) + g_{\mu\nu} D,
\end{equation}
where $D$ is an arbitrary constant; note that this relation holds also
after restriction to on-shell photons with physical polarizations.
Observe that, in contrast to the unitary gauge, as long as the higgs
is off-shell, the distributions~\eqref{eq:absorptive_r_xi}
and~\eqref{eq:txi-on} are \textit{not} electromagnetically
gauge-invariant. This was to be expected and is related to the
presence of unphysical degrees of freedom in electroweak theory in the
$R_\xi$-gauges. However, entirely on-shell EGI can be satisfied:
setting $\tilde\rho := \rho$ in~\eqref{eq:txi-on}, we plainly get
\begin{equation}
t^1_{\mu\nu}(k_1,k_2)\bigr|_{\tilde\rho=\rho}
= - \frac{1}{2^3\,(2\pi)^6\,M^2}\, \bigl(
P_{\mu\nu}(k_1,k_2)\, F_1(\rho) + D\,g_{\mu\nu} \bigr),
\label{eq:txi-onon}
\end{equation}
and one sees that $D$ must be put equal to zero. This fixes completely
the normalization freedom in the construction of~$t_{\mu\nu}^1$ in the
Feynman gauge. At this level there is of course coincidence with the
result in \cite{HerreroMorales}, despite different game rules.
Recall that in the unitary gauge, for on-shell photons with physical
polarizations, the general normalization freedom fulfilling
electromagnetic gauge invariance and Lorentz covariance is given by
the last term in \eqref{eq:gratatio-capitis}, where $\om\equiv
\om(d_{\mu\nu})$. We stress that the constants~$C_k$ appearing in that
term cannot be fixed without imposing some further normalization
conditions. To address this problem, observe
that it is possible to adjust the coefficients $C_k$ of the polynomial
in the expression~\eqref{eq:gratatio-capitis} for $t_{\gi,\mu\nu}$ in
the unitary gauge in such a way that the following equality
\begin{equation}
t_{\gi,\mu\nu}(k_1,k_2)\bigr|_{\tilde\rho=\rho}
= t^1_{\mu\nu}(k_1,k_2)\bigr|_{\tilde\rho=\rho}.
\label{eq:gauindep}
\end{equation}
holds entirely on-shell, i.e. for $\tilde\rho=\rho$. In fact, we must
set $C_0 := 2$ and $C_k := 0$ for all $k \geq 1$
in~\eqref{eq:gratatio-capitis}, which fixes completely the
normalization freedom of $t_{\gi,\mu\nu}$. Eq.~\eqref{eq:gauindep}
expresses the independence of the physical amplitude of the diphoton
decay of the higgs of the choice of the gauge. We
regard~\eqref{eq:gauindep} as a normalization condition of
time-ordered products. We have shown that this condition can be
satisfied in the case at hand and determines uniquely the
indeterminate normalization polynomial of~$t_{\gi,\mu\nu}$ in the
expression~\eqref{eq:gratatio-capitis}.
In summary, our final result for the entirely on-shell EW
$h \to \ga\ga$ decay reads:
$$
t_{\mu\nu}(k_1,k_2)\bigr|_{\tilde\rho=\rho}
= - \frac{1}{2^3\,(2\pi)^6\,M^2}\, P_{\mu\nu}(k_1,k_2)\, F_1(\rho),
$$
in agreement with the majority of the literature.
\subsection{On settling the controversy}
\label{ssc:dondedijedigo}
Should one infer that by computing in the ``physical'' unitary gauge
there is no way to entirely settle the controversy that motivates
this work, by removing the remaining ambiguity in determining the
amplitude in question? Not without at least pondering credible
``heavy-higgs'' (or $M \to 0$) and ``light-higgs'' (or $M \to \infty$)
arguments to bolster the case of $F_1(\rho)$ versus $F_1(\rho) - 2$,
that have been made in the literature.
Now, for the present authors the question is not whether either class
of arguments is compelling enough. Instead, the question is whether
they can be made within the BEG prescriptions, and at the level of
rigour of this paper. The arguments in the first-named class involve
plays with field transformations, power counting rules and the
adiabatic limit that we find hard to countenance in the BEG formalism.
However, those of the second class are persuasive within our purview.
Note that $F(0)$, for both scalar and vector boson charged fields, as
well as for Dirac fermions, must coincide with (the first coefficient
of) the $\beta$-function series associated to \textit{electric charge
renormalization}.%
\footnote{$F_0(0)=-1/3$, which has been calculated in this paper,
means precisely this.}
It was a fortunate historical fact that a calculation of the effective
Lagrangian for charged Proca particles~\cite{RussianCharge1} was
already available when the first ``exact'' computation of the higgs to
digamma process that we are aware of was performed
\cite{RussianCharge2} -- thus making possible a dependable
``light-higgs'' argument. A computation of the renormalization of the
electric charge of massive vector bosons in the unitary gauge by means
of BEG technology is in principle feasible -- cf. in this
respect~\cite[Sect.~7]{BDF09} and \cite{Duetsch15} -- and expected to
yield the required value $F_1(0)=7$. That would complete the analysis
of this paper, without going beyond the unitary gauge framework.
\section{Conclusion}
\label{sec:die-Eule}
Contrary to custom, we begin this section by declaring what we have
\textit{not} done in the paper. Finite QFT \textit{\`a~la}
Bogoliubov--Epstein--Glaser is mathematically a rigorous method. So,
referring to what is found in the literature -- like that cited in the
Introduction -- we have not employed dimensional regularization,
deemed an ``artifact'' by some. Nor do we borrow Pauli--Villars', nor
cutoff regularizations, for that matter. We did not have to practice
``judicious routings of the external momenta''~\cite{GastmansWuWu2},
nor adopt the ``loop regularization method''~\cite{TresChinosAgudos},
or any of the techniques to handle divergent integrals, resulting from
the blind application of Feynman graph technology on momentum space.
We do not pore over divergent integrals, at all. Each and every one of
the integrals appearing in this paper produces an unambiguous result;
each amplitude is finite.
We expected the BEG procedure to yield a conceptually clear
understanding of the EW $h \to \ga\ga$ decay in the unitary gauge. We
have succeeded in this -- at a price. According to Epstein and Glaser,
the adiabatic limit is to be performed \textit{after} distribution
splitting. Such an off-shell procedure for the $h \to \ga\ga$ decay in
the unitary gauge demands computations more than one order of
magnitude greater than the ones performed in this paper -- compare the
computation of the QED vertex function in~\cite[Chap.~3.8]{Scharf14}
and in~\cite{DuetschKS93}.
We were not disposed to inflict this on ourselves, nor our surviving
readers. Thus we were forced to innovate on the method, generalizing
the splitting dispersion integral to production of massless particles,
and showing that in the present situation the adiabatic limit may be
performed before distribution splitting. Only, then one may have to
add to the result so obtained an a~priori indeterminate polynomial in
the external momenta, of a degree given by the singular order of the
amplitude off-shell. It is precisely the addition of this polynomial
that is missing in references \cite{GastmansWuWu1,GastmansWuWu2} and
\cite{ChristovaI}. We have resolved the ambiguity by recourse to
gauge-fixing independence of the entirely on-shell amplitude.
Alternatively, the ambiguity could be resolved within the unitary
gauge in the BEG scheme, by invoking the low-energy argument.%
\footnote{Variants of the ``light-higgs'' or ``low energy'' argument
besides \cite{RussianCharge2,RussianCharge3} and
\cite{MelnikovVainshtein} are found for instance in
\cite[Ch.~24.8]{GranLev}, in \cite{KSpira} and in \cite{HStoehr}.}
We have not attempted here a rigorous proof of this argument, nor
computed the relevant coefficient of the beta function, leaving the
task for a separate analysis in future work.
\subsection*{Acknowledgements}
We are grateful to E. Alvarez, L. Alvarez-Gaum\'e, M. Herrero, C. P.
Mart\'in, J. C. V\'arilly and T. T. Wu for comments, discussions and
helpful remarks. We particularly thank I.~T.~Todorov for keen help in
the beginning, and his continued and thought-provoking, if contrarian,
interest in this work. As well we thank an anonymous referee for
knowledgeable reporting, definitely contributing to improve the paper.
During the inception and writing of this article, PD received funding
from the National Science Center, Poland, under the grant
UMO-2017/25/N/ST2/01012. He also gratefully acknowledges the
hospitality of the University of Zaragoza. JMG-B received funding from
the European Union's Horizon~2020 research programme under the Marie
Sk{\l}odowska-Curie grant agreement RISE~690575; from Project
FPA2015--65745--P of MINECO/Feder; from CERN; from the COST actions
MP1405 and~CA18108. Hospitality of CERN, IFT-Madrid, ITP-G\"ottingen
and ZiF-Bielefeld is gratefully acknowledged.
|
\section{Introduction}
\label{sec:introduction}
\thispagestyle{FirstPage}
Supervised learning has been well-studied for tackling audio classification problems, such as acoustic scene classification~\cite{Barchiesi2015Acoustic} and environmental sound classification~\cite{Piczak2015Environmental, Piczak2015ESC}.
To obtain classifiers with satisfactory performance, existing supervised learning techniques require large amounts of annotated training data from target sound classes, which is labor-intensive and costly to acquire.
Moreover, with the increasing diversity of observed sound classes, it becomes even more challenging for humans to collect sufficient annotated training data for all possible sound classes.
Recent work~\cite{Salamon2017Deep, Koluguri2020Meta, Shi2020FewShot, Chou2019Learning} in the audio recognition literature that deal with the lack of adequate training data mainly apply~\emph{data augmentation}~\cite{Brian2015ASoftware},~\emph{meta learning}~\cite{Finn2017Model} and~\emph{few-shot learning}~\cite{Wang2020Generalizing} methods.
However, a certain amount of representative training data from target classes is still indispensable to make these methods work.
Furthermore, to classify instances from novel classes, a supervised learning classifier would require retraining for the novel classes, which can be time-consuming and requires exhaustive parameter tuning.
In this paper, we consider the extreme case of audio classification where target sound classes have no available training samples but only class side information (e.g., textual descriptions).
This problem is generally referred to as the~\emph{zero-shot learning}~\cite{Palatucci2009Zero}, which has been increasingly studied in the context of image classification.
In contrast to conventional supervised learning, zero-shot learning uses training data from only predefined classes (i.e., seen classes) to obtain classifiers that can be generalized to novel classes (i.e., unseen classes).
There is only limited work that has been done for zero-shot learning in audio classification.
Due to the lack of training data from unseen classes, class side information is used as a compensation for exploring the relationship between seen classes and unseen classes to make zero-shot learning possible.
Prior work~\cite{Islam2019SoundSemantics, Xie2019Zero, Xie2020ZeroShotAudio} tackled zero-shot learning by leveraging semantic side information of sound classes, such as textual labels, with a two-phase learning process.
First, intermediate-level representations were learned for audio instances and sound classes, respectively.
Audio instances were embedded into a low-dimensional acoustic space with feature learning techniques, such as~\emph{VGGish}~\cite{Hershey2017CNN}.
Sound classes were represented by word embeddings in a semantic space, which were extracted from their semantic side information with pre-trained language models, such as~\emph{Word2Vec}~\cite{Mikolov2013Efficient}.
Then, an acoustic-semantic projection was learned to associate acoustic embeddings with semantic embeddings.
\mbox{Islam et al.}~\cite{Islam2019SoundSemantics} employed a two-layer fully-connected neural network to model a nonlinear projection of acoustic embeddings onto semantic embeddings.
In our previous work~\cite{Xie2019Zero, Xie2020ZeroShotAudio}, a bilinear model was used to learn a bidirectional linear projection between acoustic embeddings and semantic embeddings.
Another prior work~\cite{Choi2019Zero} was conducted by integrating the acoustic embedding learning phase with the acoustic-semantic projection learning phase to optimize them holistically.
Thus, a nonlinear acoustic-semantic projection was inherently built into their model.
In this paper, we extend our previous work~\cite{Xie2019Zero, Xie2020ZeroShotAudio} by introducing matrix decomposition and nonlinear activation functions (e.g., tanh) to the bilinear model.
We develop factored linear and nonlinear acoustic-semantic projections for zero-shot learning in audio classification.
Experimental results show that the proposed projection methods are effective for improving classification performance of zero-shot learning in audio classification.
The remainder of this paper is organized as follows.
In Section~\ref{sec:zero-shot-learning}, we introduce the concept of zero-shot learning in audio classification.
Then, we present the proposed factored linear and nonlinear projections in Section~\ref{sec:factored-linear-and-nonlinear-projections}.
We describe the training algorithm in Section~\ref{sec:training-algorithm}, and discuss the experimental results in Section~\ref{sec:experiments}.
Finally, we conclude this paper in Section~\ref{sec:conclusion}.
\section{Zero-Shot Learning}
\label{sec:zero-shot-learning}
In this section, we introduce the concept of zero-shot learning via semantic side information in audio classification.
We denote the audio sample space by $X$, the set of seen sound classes by $Y$, and the set of unseen sound classes by $Z$.
Note that $Y$ and $Z$ are disjoint.
Let $\theta(x) \in \mathbb{R}^{d_{a}}$ be the acoustic embedding of an audio instance $x \in X$ in an acoustic space, and $\phi(y) \in \mathbb{R}^{d_{s}}$, $\phi(z) \in \mathbb{R}^{d_{s}}$ be the semantic embeddings of sound classes $y \in Y$ and $z \in Z$ in a semantic space, respectively.
We are given the training data $S_{tr}=\{(x_{n}, y_{n}) \in X \times Y | n=1,\dots,N\}$, where $x_{n}$ is an annotated audio sample belonging to a seen sound class $y_{n}$.
In zero-shot learning, our goal is to learn an audio classifier $f:X \rightarrow Z$, which can predict the correct sound class for an audio instance $x \in X$, given its acoustic embedding $\theta(x)$ and the semantic embeddings $\phi(z)$ of every candidate sound class $z \in Z$.
In prior work~\cite{Islam2019SoundSemantics, Xie2019Zero, Xie2020ZeroShotAudio, Choi2019Zero}, this is done via learning an acoustic-semantic projection $T:\mathbb{R}^{d_{a}} \rightarrow \mathbb{R}^{d_{s}}$ of acoustic embeddings onto semantic embeddings.
Given an audio instance $x \in X$ belonging to a sound class $z_{x} \in Z$, it is generally assumed that the projected acoustic embedding $T(\theta(x))$ in the semantic space should be closer to the semantic embedding $\phi(z_{x})$ of its correct sound class $z_{x}$ rather than those of other sound classes.
A similarity scoring function $F:\mathbb{R}^{d_{s}} \times \mathbb{R}^{d_{s}} \rightarrow \mathbb{R}$, as known as the~\emph{compatibility function}, is then defined to measure how similar/compatible a projected acoustic embedding and a semantic embedding are.
To measure the similarity of two embedding vectors, the popular choices of $F$ could be Euclidean distance~\cite{Islam2019SoundSemantics}, cosine similarity~\cite{Choi2019Zero}, dot product~\cite{Xie2019Zero, Xie2020ZeroShotAudio}, etc.
Therefore, the classifier $f:X \rightarrow Z$ is formulated as\footnote{Note that the $\argmin$ operation is used for other compatibility functions, such as Euclidean distance, etc.}
\begin{equation}
\label{eq:ZSL_classifier}
z_{x} = f(x) = \argmax_{z \in Z} F(T(\theta(x)),\phi(z)).
\end{equation}
During the training stage, the projection $T$ is trained with audio samples $(x_{n}, y_{n}) \in S_{tr}$ such that
\begin{equation}
\label{eq:training_classifier}
y_{n} = f(x_{n}) = \argmax_{y \in Y} F(T(\theta(x_{n})),\phi(y)).
\end{equation}
For prediction, an audio instance will be classified into the sound class, the semantic embedding of which is most compatible with its projected acoustic embedding.
\section{Factored Linear and Nonlinear Projections}
\label{sec:factored-linear-and-nonlinear-projections}
In this section, we introduce our approach for developing factored linear and nonlinear acoustic-semantic projections for zero-shot learning in audio classification.
First, we present the bilinear model used in our previous work~\cite{Xie2019Zero, Xie2020ZeroShotAudio}, on which we build factored linear and nonlinear acoustic-semantic projections.
Then, we describe a factored linear projection obtained by applying matrix decomposition to the bilinear model.
After that, we introduce the nonlinear projections derived from the factored linear projection by introducing nonlinear activation functions, such as tanh\@.
\subsection{Bilinear Model}\label{subsec:bilinear-model}
Inspired by prior work~\cite{Akata2016Label, Xian2016Latent} in computer vision, we employed a bilinear model in~\cite{Xie2019Zero, Xie2020ZeroShotAudio} to learn a bidirectional linear projection between acoustic embeddings and semantic embeddings.
The audio classifier $f:X \rightarrow Z$ was then written in a bilinear form as
\begin{equation}
\label{eq:bilinear_model}
f(x) = \argmax_{z \in Z} \theta(x)^{'} W \phi(z),
\end{equation}
where $W$ was the learned projection matrix.
Considering a projection of acoustic embeddings onto semantic embeddings, $T$ could be formulated as
\begin{equation}
\label{eq:cross_domain_proj}
T(\theta(x)) = W^{'} \theta(x).
\end{equation}
Thus, in~\eqref{eq:bilinear_model}, the dot product was inherently defined as the compatibility function $F$
\begin{equation}
\label{eq:dot_production}
F(T(\theta(x)),\phi(z)) = T(\theta(x))^{'} \phi(z).
\end{equation}
\subsection{Factored Linear Projection}\label{subsec:factored-linear-projection}
In the case where $d_{a}$ and $d_{s}$ are large, it would be valuable to decompose $W$ into a product of two low-rank matrices $U_{d_{a} \times r}$ and $V_{r \times d_{s}}$ to reduce the effective number of parameters in the bilinear model.
Thus, we consider the rank decomposition $W=UV$ in~\eqref{eq:cross_domain_proj} to build a factored linear projection
\begin{equation}
\label{eq:linear_projection}
T(\theta(x)) = (UV)^{'} \theta(x) = V^{'} U^{'} \theta(x).
\end{equation}
\subsection{Nonlinear Projection}\label{subsec:nonlinear-projection}
Based on the linear projection~\eqref{eq:linear_projection}, we consider introducing nonlinear activation functions (e.g., tanh) into it to model the potential nonlinear relationship between acoustic embeddings and semantic embeddings.
We apply a nonlinear activation function $t$ to $U^{'} \theta(x)$.
Therefore, a nonlinear projection is formulated as
\begin{equation}
\label{eq:nonlinear_projection1}
T(\theta(x)) = V^{'} t(U^{'} \theta(x)).
\end{equation}
We notice the fact that~\eqref{eq:nonlinear_projection1} also defines a two-layer fully-connected neural network\footnote{Note that the bias parameters of each layer are ignored from~\eqref{eq:nonlinear_projection1} and later formulas in this paper.} with input $\theta(x)$, layer weights $U^{'}$ and $V^{'}$.
To describe deeper fully-connected neural networks, we can simply add more activation functions and matrices between $U^{'}$ and $V^{'}$.
For example, a three-layer fully-connected neural network is formulated as
\begin{equation}
\label{eq:nonlinear_projection2}
T(\theta(x)) = V^{'} t(Q~t(U^{'} \theta(x))).
\end{equation}
where the matrix $Q$ denotes the weight parameters of the second fully-connected layer.
\section{Training Algorithm}
\label{sec:training-algorithm}
In this section, we introduce the algorithm for learning an acoustic-semantic projection $T$ with training data $S_{tr}$.
Given an audio sample $(x_{n},y_{n}) \in S_{tr}$, we consider the task of sorting sound classes $y \in Y$ in descending order according to their compatibility values $F(T(\theta(x_{n})),\phi(y))$.
Our objective is to optimize $T$ so that the correct class $y_{n}$ would be ranked at top of the sorted class list, i.e., having the maximal compatibility value for $x_{n}$.
Let $r_{y}$ be the position index of a sound class $y$ in the sorted class list.
We define $r_{y}=0$ when $y$ is sorted at the first position.
By applying a ranking error function~\cite{Usunier2009Ranking}, we transform position index $r$ into loss $\beta(r)$:
\begin{equation}
\label{eq:ranking_error}
\beta(r) = \sum_{i=1}^{r} \alpha_{i},
\end{equation}
with $\alpha_{1} \geq \alpha_{2} \geq \dots \geq 0$ and $\beta(0)=0$.
Specifically, $\alpha_{i}$ denotes a penalty to a class losing a position from $i-1$ to $i$.
In this paper, we follow previous work~\cite{Akata2016Label, Usunier2009Ranking} and choose $\alpha_{i}=1/i$.
To learn an acoustic-semantic projection $T$ with audio samples $(x_{n},y_{n}) \in S_{tr}$, we minimize the weighted approximate-rank pairwise objective~\cite{Weston2011WSABIE}
\begin{equation}
\label{eq:weighted_ranking}
\dfrac{1}{N} \sum_{n=1}^{N} \dfrac{\beta(r_{y_{n}})}{r_{y_{n}}} \sum_{y\in Y} \max \{0, l(x_{n},y_{n},y)\},
\end{equation}
with the convention $0/0=0$ when $y_{n}$ is top-ranked.
In this paper, we define the hinge loss $l(x_{n},y_{n},y)$ as
\begin{equation}
\label{eq:hinge_loss}
\begin{split}
l(x_{n},y_{n},y) = \Delta(y_{n},y) &+ F(T(\theta(x_{n})),\phi(y)) \\ &- F(T(\theta(x_{n})),\phi(y_{n})),
\end{split}
\end{equation}
where $\Delta(y_{n},y)=0$ if $y_{n}=y$ and 1 otherwise.
The objective~\eqref{eq:weighted_ranking} is convex and can be optimized through stochastic gradient descent.
To prevent over-fitting, we regularize~\eqref{eq:weighted_ranking} with L2 norms of parameter matrices in $T$.
%
\begin{table}[!t]
\centering
\begin{tabular}{c||c||c}
\hline
Class Fold & Sound Class & Audio Sample \\
\hline
Fold0 & 104 & 23007 \\
\hline
Fold1 & 104 & 22889 \\
\hline
Fold2 & 104 & 22762 \\
\hline
Fold3 & 104 & 22739 \\
\hline
Fold4 & 105 & 21377 \\
\hline
\end{tabular}
\caption{Class folds in the selected subset of AudioSet.}
\label{tab:AudioSet_subset}
\end{table}
\section{Experiments}
\label{sec:experiments}
In this section, we evaluate the proposed method with AudioSet~\cite{Gemmeke2017AudioSet} and report the effectiveness of factored linear and nonlinear projections.
\subsection{Dataset}\label{subsec:dataset}
AudioSet~\cite{Gemmeke2017AudioSet} is a large unbalanced audio dataset, which contains over two million weakly labeled audio clips covering 527 sound classes.
Most of these audio clips are multi-label.
In this work, we focus on zero-shot learning in single-label classification problems.
We follow the same experimental setup as in ~\cite{Xie2020ZeroShotAudio}.
An audio subset containing 112,774 single-label 10-second audio clips and 521 sound classes is selected from AudioSet.
We randomly split the selected subset into five disjoint class folds.
The number of sound classes and audio samples in each class fold is shown in Table~\ref{tab:AudioSet_subset}.
\subsection{Acoustic Embeddings}\label{subsec:acoustic-embeddings}
A pre-trained VGGish~\cite{Hershey2017CNN} was used to generate acoustic embeddings from audio clips in~\cite{Xie2019Zero}.
In zero-shot learning, it is generally assumed that unseen classes are unknown at training stage.
Since a pre-trained VGGish may have already embedded knowledge about unseen sound classes, using it to generate acoustic embedding can lead to a biased evaluation of zero-shot learning in audio classification.
Therefore, we train VGGish from scratch with audio data excluding unseen sound classes in this work.
Following~\cite{Xie2019Zero, Xie2020ZeroShotAudio}, an 10-second audio clip is first split into ten one-second audio segments without overlapping.
Then, a 128-dimensional embedding vector is generated for each audio segments with the trained VGGish.
To obtain the clip-level acoustic embedding for an audio clip, we take the average of the 128-dimensional embedding vectors extracted from its one-second audio segments.
\subsection{Class Semantic Embeddings}\label{subsec:class-semantic-embeddings}
In AudioSet~\cite{Gemmeke2017AudioSet}, a sound class is described by one or several textual labels (i.e., words and phrases) and an additional short description (i.e., sentences).
In this work, we consider only textual labels as class semantic side information.
We adopt Word2Vec~\cite{Mikolov2013Efficient} as a word embedding model for generating semantic embeddings from these textual labels.
For the sake of simplicity, we use a publicly available pre-trained Word2Vec\footnote{Word2Vec: https://code.google.com/archive/p/word2vec.}, which embeds roughly three million English words and phrases.
It outputs a 300-dimensional semantic word vector for a single word or a phrase.
To represent a sound class with semantic word vectors, we calculate the average of these word vectors extracted from its textual labels.
\subsection{Experimental Setup}\label{subsec:experimental-setup}
In the following experiments, we first train an VGGish for generating acoustic embeddings with class folds ``Fold0'' and ``Fold1''.
Then, we conduct zero-shot learning in audio classification with ``Fold2'' for training, ``Fold3'' for parameter validation and ``Fold4'' for test, respectively.
\textbf{VGGish Training}.
Audio samples of class folds ``Fold0'' and ``Fold1'' are first randomly split into training/validation partitions with a class-specific proportion of 75/25.
Then, we train an VGGish from scratch by feeding log mel spectrogram extracted from audio clips into it.
After training, the trained VGGish achieves a classification accuracy (TOP-1) of 27.4\% on the validation partition.
\textbf{Zero-Shot Learning}.
We conduct zero-shot learning in audio classification with the proposed factored linear and nonlinear projections, respectively.
We use the bilinear model as the baseline method.
For the factored linear projection~\eqref{eq:linear_projection}, we experiment with low-rank decomposition and full-rank decomposition of $W$ to investigate the effect of the rank $r$ and the L2 norm regularization.
For the nonlinear projection~\eqref{eq:nonlinear_projection1}, we experiment on three widely used activation functions, i.e., ReLU, sigmoid and tanh.
We implement~\eqref{eq:nonlinear_projection1} by two-layer fully-connected neural networks, which are denoted by FC2$_{relu}$, FC2$_{sigmoid}$ and FC2$_{tanh}$, respectively.
Similarly, we implement~\eqref{eq:nonlinear_projection2} by an three-layer fully-connected neural network with tanh, which is denoted by FC3$_{tanh}$.
To prevent randomness, each projection method is evaluated twenty times with random initialization.
The averages and standard deviations of their TOP-1 accuracies are reported in Table~\ref{tab:zsl_results}.
\subsection{Result and Analysis}\label{subsec:result-and-analysis}
As a baseline method, the bilinear model achieves an averaged TOP-1 of 5.7\% with a standard deviation of 1.1\%.
For factored linear projections with either low-rank decomposition or full-rank decomposition of $W$, we obtain similar results (roughly 6.4 $\pm$ 0.6\%).
We conclude that, with the L2 norm regularization, the rank $r$ has a limited influence on classification performance.
Here, we report the result from a factored linear projection with the full-rank decomposition (i.e., \mbox{$r$=128}) of $W$, which has an averaged TOP-1 of 6.3\% with a standard deviation of 0.8\%.
Unpaired t-test with $\alpha$=0.05 is used to measure the statistical significance among different methods.
The results of the factored linear projection are significantly different from those of the bilinear model (\mbox{$t(38)$=2.09}, \mbox{$p$=0.04}).
It shows that classification performance is improved by applying rank decomposition to $W$ with L2 norm regularization.
For nonlinear projections, we set \mbox{$r$=128}.
Classification performance is impaired with FC2$_{relu}$ (5.5 $\pm$ 0.9\%) while it is improved with FC2$_{sigmoid}$ (7.0 $\pm$ 0.5\%) and FC2$_{tanh}$ (7.2 $\pm$ 0.6\%).
Particularly, the averaged TOP-1 of FC2$_{tanh}$ is significantly better than those of the bilinear model (\mbox{$t(38)$=5.60}, \mbox{$p$=$4.50\mathrm{e}{-6}$}) and the factored linear projection (\mbox{$t(38)$=3.88}, \mbox{$p$=$4.59\mathrm{e}{-4}$}).
Compared with sigmoid and tanh, the ReLU function introduces non-linearity by simply dropping negative values from its inputs.
In a two-layer fully-connected neural network, this can lead to a poor acoustic-semantic projection and results in an impaired performance.
For FC2$_{sigmoid}$ and FC2$_{tanh}$, we think both of them capture non-linearity between acoustic embeddings and semantic embeddings, which is useful for improving classification performance.
However, compared with FC2$_{tanh}$, classification performance is impaired with FC3$_{tanh}$ (6.0 $\pm$ 0.6\%).
It seems that it would not be helpful for improving classification performance in zero-shot learning by simply introducing more nonlinear layers in a fully-connected neural network.
\begin{table}[!t]
\centering
\begin{tabular}{c|l||c}
\hline
\multicolumn{2}{c||}{\bfseries Acoustic-Semantic } & \bfseries TOP-1 (\%) \\
\multicolumn{2}{c||}{\bfseries Projection} & (avg $\pm$ std) \\
\hline
\multicolumn{2}{c||}{Bilinear (baseline)} & 5.7 $\pm$ 1.1 \\
\hline
\multicolumn{2}{c||}{Factored Linear} & 6.3 $\pm$ 0.8 \\
\hline
\multirow{4}{*}{Nonlinear} & FC2$_{relu}$ & 5.5 $\pm$ 0.9 \\
\cline{2-3}
& FC2$_{sigmoid}$ & 7.0 $\pm$ 0.5 \\
\cline{2-3}
& FC2$_{tanh}$ & \bfseries 7.2 $\pm$ 0.6 \\
\cline{2-3}
& FC3$_{tanh}$ & 6.0 $\pm$ 0.6 \\
\hline
\end{tabular}
\caption{Zero-Shot learning in audio classification with different acoustic-semantic projections.}
\label{tab:zsl_results}
\end{table}
\section{Conclusion}
\label{sec:conclusion}
In this paper, we present an approach for zero-shot learning in audio classification with factored linear and nonlinear acoustic-semantic projections.
We develop factored linear and nonlinear projections by applying rank decomposition and nonlinear activation functions to a bilinear model.
We evaluate our proposed approach with a large unbalanced audio dataset.
The experimental results show that both factored linear and nonlinear projections are effective for zero-shot learning in audio classification.
With a factored linear projection, it achieves an averaged TOP-1 accuracy of 6.3\%, which is better than the prior bilinear model (5.7\%).
By introducing nonlinear activation functions into it, classification performance can be further improved.
Classification performance achieves an averaged TOP-1 accuracy of 7.2\% by using a nonlinear projection with the tanh activation function.
\section{Acknowledgement}
\label{sec:acknowledgement}
The research leading to these results has received funding from the European Research Council under the European Unions H2020 Framework Programme through ERC Grant Agreement 637422 EVERYSOUND. OR was funded by Academy of Finland grant no. 314602.
\bibliographystyle{IEEEbib}
|
\section{Introduction}
\label{introduction}
Question Answering (QA) has attracted extensive interest in the fields of Computer Vision (CV) and Natural Language Processing, such as Visual Question Answering (VQA) \citep{yu2019deep, khademi2020multimodal} in CV and Machine Reading Comprehension (MRC) \citep{nie2019revealing, saxena2020improving} in NLP. Recently, a new task named Textbook Question Answering (TQA) \citep{kembhavi2017you} that aims to answer diagram/non-diagram questions given a large multi-modal context consisting of diagrams, text and few natural images was proposed. The task is developed from middle school science curricula and describes the real-life process of a student who learns new knowledge from books and assesses learning achievement. Compared with MRC and VQA, TQA is more complex and more realistic.
In real-life scenarios, the TQA robot should not only answer a question accurately, but also give \emph{students} the reason to choose the answer. Although existing neural-symbolic works \citep{yi2018neural, mao2018neuro} on the CLEVR dataset make significant progress on the explainability, they are not applicable to the TQA dataset. Here are two reasons for this. First, the curriculum is diverse in the TQA dataset, which makes the design of the domain-specific language (DSL) difficult.
Second, there is no any supervision information except the answer label.
An analysis \citep{kembhavi2017you} of the information scope required to answer questions in the TQA dataset shows that about $80\%$ of the questions require a single sentence or multiple sentences. Inspired by this, we may be able to provide the evidence spans\footnote{In this paper, spans are represented by the indexes of sentences instead of words.} about questions for students, which would give them explainability to some extent. For example, when the TQA robot gives students the correct answer \emph{B} to the question \emph{1}, it also provides an evidence span as the explanation in Figure \ref{fig.1}. We make an assumption owing to the lack of supervision information of evidence spans in the TQA dataset. \textbf{Assumption: if a question is answered accurately after fusing a generated span and other essential information, the span is a gold evidence to the question.} For example, the span marked in green in the middle part of Figure \ref{fig.1} is a gold evidence to the question \emph{1}.
\begin{figure}[!htbp]
\centering
\includegraphics[width=\linewidth]{images/tqa-example.pdf}
\caption{An example of the TQA task and a brief illustration of our work. The abundant essays and diagrams of a lesson are shown in the left part. The diagram/non-diagram questions of this lesson are shown in the right part.}
\label{fig.1}
\end{figure}
Based on the assumption, we devise a novel architecture towards span-level eXplanations of the Textbook Question Answering (XTQA) as illustrated in Figure \ref{fig.1}. First, XTQA uses our proposed coarse-to-fine grained algorithm to generate evidence spans of questions. In the coarse grained phase, top $M$ paragraphs relevant to questions are chosen from abundant essays by the TF-IDF score. In the fine grained phase, top $K$ evidence spans are generated from all candidate spans within the top $M$ paragraphs by computing the information gain of each span to questions. The larger information gain indicates the more uncertainty of questions reduced by spans. Second, XTQA fuses the information of generated evidence spans, questions, options and diagrams to provide answers. Finally, not only the answers but also the evidences to choose them are provided to students.
The contributions of our work are as follows. (1) We devise a novel architecture that can give students evidences as well as answers for TQA. (2) We propose a coarse-to-fine grained algorithm to generate span-level evidences by computing the information gain of each span to questions. (3) We conduct extensive experiments and ablation studies on the TQA dataset \citep{kembhavi2017you} to verify the effectiveness of the XTQA and our proposed assumption. Experimental results show that XTQA significantly improves the state-of-the-art performance compared with baselines. (4) We also explore how to effectively integrate the self-supervised learning method SimCLR \citep{chen2020simple} into the proposed architecture to further improve the accuracy of the TQA task.
\section{Related Work}
Since the TQA task is similar to MRC and VQA, we briefly review some of their research relevant to our work.
\textbf{MRC} requires a machine to answer questions accurately given a textual context \citep{lehnert1977process}. \citet{wang2018multi} proposed a multi-granularity hierarchical attention fusion network to answer questions for a given narrative paragraph. \citet{lee2018ranking} proposed a paragraph ranker to improve the answer recall with less noise. \citet{ding2019cognitive} proposed CogQA that builds a cognitive graph by an implicit extraction module and an explicit reasoning module to address the multi-hop question answering. Although the above studies achieved success, they did not estimate the relationship and combined design of retrieval and comprehension. \citet{nie2019revealing} proposed a hierarchical pipeline model that reveals the importance of semantic retrieval to give general guidelines on the system design for MRC.
Unlike the MRC task, the TQA task not only involves the textual question but also the visual question. In this paper, we first apply the self-supervised learning method SimCLR \citep{chen2020simple} to learn the diagram representation in the TQA dataset to the best of our knowledge.
\textbf{VQA} requires a machine to answer questions accurately given an image \citep{antol2015vqa}. \citet{yang2016stacked} proposed stacked attention networks that perform multi-step reasoning to infer answers. \citet{KimOLKHZ17} proposed a low-rank bi-linear pooling method using Hadamard product to address the complex computation. However, the above methods consider little on the effectiveness of the explicit regions for answering. To address this issue, \citet{anderson2018bottom} proposed a combined bottom-up and top-down attention mechanism that computes attentions at the level of salient image regions and objects. \citet{gao2019multi} proposed a multi-modality latent interaction module to model the summarization of language and visual information. However, these methods may not perform complex reasoning. To address this issue, \citet{yi2018neural} proposed a neural-symbolic visual question answering architecture that disentangles question and image understanding from reasoning. Based on this paper, \citet{mao2018neuro} proposed a neuro-symbolic reasoning module that executes generated programs on the latent scene representations to perform reasoning. These methods can execute complex logic reasoning based on the manually designed domain-specific language.
Unlike the VQA task, the TQA task contains multi-modal context except questions and diagrams. In this paper, we argue that the explainability of the TQA task should place \emph{students} as a key aspect to be considered. We propose a coarse-to-fine grained algorithm to provide the evidence spans of questions for students.
\section{Task Formulation}
\label{sec.3}
The TQA dataset contains two types of questions: diagram questions and non-diagram questions. Therefore, the TQA task can be classified into two categories: diagram question answering and non-diagram question answering. Owing to the difference only at this point between them, here we give the task formulation of the diagram question answering.
Given a dataset $\mathcal{S}$ consisting of $n$ triplets $(c_i, d_i, q_i, \mathcal{A}_i)$ with $c_i \in \mathcal{C}$ representing multi-modal context of a lesson, $d_i \in \mathcal{D}$ representing a diagram, $q_i \in \mathcal{Q}$ representing a question and $\mathcal{A}_i \subsetneqq \mathcal{A}$ representing the candidate answers of $q_i$, one must optimize the parameters $\theta$ of the function $f: \mathcal{C} \times \mathcal{D} \times \mathcal{Q} \rightarrow \mathbb{R}^{|\mathcal{A}|}$ to produce accurate predictions. $a_{i,j} \in \mathcal{A}_i$ denotes the $j$-th candidate answer of $q_i$.
A candidate evidence span $e_{i,k} \in \mathcal{E}_i \subsetneqq \mathcal{E}$ to $q_i$ is represented by its start $\mathrm{START}(k)$ and end $\mathrm{END}(k)$ indexes respectively following \citep{ma2020jointly}, where $\mathcal{E}_i$ represents the candidate evidence span set of $q_i$, $1 \leq k \leq N$, $1 \leq \mathrm{START}(k) \leq \mathrm{END}(k) \leq t$ and $ N = \frac{t(t+1)}{2}$ is the number of candidate evidence spans supposing $c_i$ containing $t$ sentences. We rewrite the function $f: \mathcal{E} \times \mathcal{D} \times \mathcal{Q} \rightarrow \mathbb{R}^{|\mathcal{A}|}$ since evidence spans and other essential information are fused to answer questions. To summarize, we optimize $\theta$ to obtain not only the correct answer $a_i$ but also the evidence span $e_i$ to $q_i$ based on the assumption in Section \ref{introduction}.
\section{Method}
The architecture of the XTQA is shown in Figure \ref{fig.2}. It consists of four modules: question/answer representation, evidence span generation, diagram representation and answer prediction.
\textbf{Question/Answer representation.} This module uses GRUs and learned attention mechanisms to obtain the word-level representations $q_i^{'}$/$a_{i,j}^{'}$ and global representations $q_i^{''}$/$a_{i,j}^{''}$ of the question/candidate answer $q_i$/$a_{i,j}$ respectively.
\textbf{Evidence span generation.} This module uses our proposed coarse-to-fine grained algorithm to obtain the representations $e_{i}^{'''}$ of the top $K$ evidence spans to $q_i$ and their indexes $[\mathrm{START}(k), \mathrm{END}(k)]$ from all candidate spans within top $M$ paragraphs $p_i$.
\textbf{Diagram representation.} This module uses CNNs to obtain the representation $d_i^{'}$ of the diagram $d_i$.
\textbf{Answer prediction.} This module gives students not only the predicted answer $\hat{a}_{i}$ but also the evidence span $e_i$ to choose it after multi-modal information fusing.
\begin{figure*}[!htbp]
\centering
\includegraphics[width=\linewidth]{images/xtqa.pdf}
\caption{The architecture towards span-level eXplanations of the Textbook Question Answering (XTQA). The left part shows a lesson and a question in the TQA dataset. The index of the evidence span to a question is obtained by our proposed coarse-to-fine grained algorithm in the middle module of the middle part.}
\label{fig.2}
\end{figure*}
\subsection{Question/Answer Representation}
We use uni-directional GRUs to obtain the $d_1$-dimensional word-level representations $q_i^{'} \in \mathbb{R}^{X \times d_1}$ and $a_{i,j}^{'} \in \mathbb{R}^{Y \times d_1}$ of $q_i$ and $a_{i,j}$ respectively as follows:
\begin{equation}
\label{eq.1}
\begin{split}
& q_i^{'} = \mathrm{GRU_s}(embeding(q_i)), \\
& a_{i,j}^{'} = \mathrm{GRU_s}(embedding(a_{i,j})),
\end{split}
\end{equation}
where $q_i \in \mathbb{R}^{X \times 1}$ denotes the $i$-th question, $a_{i,j} \in \mathrm{R}^{Y \times 1}$ denotes the $j$-th candidate answer of $q_i$, $X$ and $Y$ denotes the maximum length of $q_i$ and $a_{i,j}$ respectively, and $embedding(\cdot)$ is used to get word embeddings.
We use learned attention mechanisms to obtain the $d_1$-dimensional global information $q_i^{''} \in \mathbb{R}^{d_1 \times 1}$ and $a_{i,j}^{''} \in \mathbb{R}^{d_1 \times 1}$ of $q_i$ and $a_{i,j}$ respectively as follows:
\begin{equation}
\label{eq.2}
\begin{split}
& \alpha = \mathrm{softmax}(\mathrm{MLP_s}(q_i^{'})), \\
& q_i^{''} = \sum_{u=1}^{X} \alpha_u \circ q_{i,u}^{'},
\end{split}
\end{equation}
where $\alpha \in \mathbb{R}^{X \times 1}$ is the learned attention weight matrix by MLPs, $\circ$ denotes the element-wise product, and $q_{i,u}^{'}$ is the $u$-th word representations of $q_i$. $a_{i,j}^{''}$ is obtained by analogy.
\subsection{Evidence Span Generation}
Although the multi-modal context $c_i$ contains abundant essays, only a single sentence or multiple sentences would be required to answer $q_i$. Inspired by this, we propose a coarse-to-fine grained algorithm to generate evidence spans of $q_i$.
In the coarse phase, the TF-IDF method is used to narrow down the scope of textual context from a lesson to top $M$ paragraphs $p_i$ relevant to $q_i$. $p_i \in \mathbb{R}^{M \times L \times O}$ can be denoted as follows:
\begin{equation}
\label{eq.3}
p_i = \mathrm{TFIDF}(q_i, c_i),
\end{equation}
where $L$ is the maximum number of sentences in each paragraph, and $O$ is the maximum length of each sentence. The shared $\mathrm{GRU_s}$ in Equation \ref{eq.1} is used to obtain the $d_1$-dimensional word-level representation $p_i^{'} \in \mathbb{R}^{M \times L \times O \times d_1}$ of $p_i$. We also use the shared learned attention mechanism in Equation \ref{eq.2} to obtain the $d_1$-dimensional global representations $p_i^{''} \in \mathbb{R}^{M \times L \times d_1}$ of sentences within $p_i$.
\begin{algorithm}[!htbp]
\caption{Evidence span generation}
\label{algorithm1}
\LinesNumbered
\KwIn{question $q_i$, multi-modal context $c_i$.}
\KwOut{representation $e_i^{''}$ of evidence span $e_i$ and its index.}
Choose top $M$ paragraphs using Equation \ref{eq.3}; \\
Obtain the whole candidate evidence span representations according to their indexes using Equation \ref{eq.4}; \\
Compute the information gain of each span to $q_i$ using Equation \ref{eq.6}; \\
Choose top $K$ evidence span.
\end{algorithm}
In the fine-grained phase, top $K$ evidence spans are generated from all candidate spans within top $M$ paragraphs by computing the information gain of each span to questions. Concretely, the representation at the $\mathrm{START}(k)$ and $\mathrm{END}(k)$ index as described formally in Section \ref{sec.3} are concatenated to generate the candidate evidence span representation $e_{i,k}^{'} \in \mathbb{R}^{M \times N \times 2d_1}$ as follows:
\begin{equation}
\label{eq.4}
e_{i,k}^{'} = [p_{i, \mathrm{START}(k)}^{''}; p_{i, \mathrm{END}(k)}^{''}],
\end{equation}
where $N = \frac{L(L+1)}{2}$ is the number of candidate evidence spans within each paragraph. To obtain the final candidate evidence span representation $e_{i,k}^{''} \in \mathbb{R}^{M \times N \times d_1}$, we apply the average pooling $\mathrm{AP}$ with kernel size $2$ on $e_{i,k}^{'}$ as follows:
\begin{equation}
e_{i,k}^{''} = \mathrm{AP}(e_{i,k}^{'}).
\end{equation}
We compute the information gain $g(q_i, e_{i,k})$ of each candidate evidence span $e_{i,k}$ to the question $q_i$ for obtaining the top $K$ evidence span. $g(q_i, e_{i,k})$ can be denoted as follows:
\begin{equation}
\label{eq.6}
g(q_i, e_{i,k}) = H(q_i) - H(q_i|e_{i,k}),
\end{equation}
where $H(q_i)$ is the entropy of $q_i$, i.e., the uncertainty of $q_i$, and $H(q_i|e_{i,k})$ is the conditional entropy of $q_i$ given $e_{i,k}$, i.e., the uncertainty of $q_i$ given $e_{i,k}$. The larger information gain indicates the more uncertainty of $q_i$ reduced by $e_{i,k}$. $H(q_i)$ can be denoted as follows:
\begin{equation}
\begin{split}
& H(q_i) = \mathbb{E}[-\log(p(q_i))], \\
& p(q_i) = \sigma(\mathrm{MLP_h}(q_i^{''})),
\end{split}
\end{equation}
where $\mathbb{E}$ is the expected value operator, $p(q_i)$ denotes the probability of $q_i$ being answered accurately, and $\sigma$ is the sigmoid function. $H(q_i|e_{i,k})$ can be denoted as follows:
\begin{equation}
\begin{split}
& H(q_i|e_{i,k}) = \mathbb{E}[-\log(p(q_i, e_{i,k}))], \\
& p(q_i, e_{i,k}) = \sigma(\mathrm{MLP_h}(\mathrm{AP}([q_i^{''}, e_{i,k}^{''}]))),
\end{split}
\end{equation}
where $p(q_i, e_{i,k})$ is the probability of $q_i$ being answered accurately given $e_{i,k}$, $\mathrm{AP}$ denotes the average pooling with kernel size 2. A formal description about this algorithm is shown in Algorithm \ref{algorithm1}.
\begin{table*}[!htbp]
\centering
\resizebox{\textwidth}{!}{\begin{tabular}{lccccc}
\toprule \toprule
Model & Non-diagram T/F & Non-diagram MC & Non-diagram All & Diagram & All \\ \midrule
CMR \cite{ZhengGK20} & $51.14$ & $30.65$ & $38.72$ & $30.73$ & $34.53$ \\
MUTAN \citep{ben2017mutan} & $51.72$ & $31.18$ & $39.27$ & $30.29$ & $34.56$ \\
MFB \citep{yu2017multi} & $51.73$ & $30.65$ & $38.95$ & $30.76$ & $34.65$ \\
BAN \cite{kim2018bilinear} & $51.70$ & $31.11$ & $39.22$ & $30.65$ & $34.73$ \\
MCAN \citep{yu2019deep} & $51.72$ & $\textbf{32.55}$ & $40.10$ & $30.58$ & $35.10$ \\ \hline
XTQA & $\textbf{58.24}$& $30.33$ & $\textbf{41.32}$ & $\textbf{32.05}$ & $\textbf{36.46}$ \\ \bottomrule \bottomrule
\end{tabular}}
\caption{Experimental results ($\%$ accuracy) of different-type questions on the validation split. We show the accuracy of True/False (Non-diagram T/F) questions, multiple-choice questions (Non-diagram MC), the whole non-diagram questions (Non-diagram All = Non-diagram T/F $\cup$ Non-diagram MC), multiple-choice diagram questions (Diagram) and total questions (All = Non-diagram $\cup$ Diagram). \textbf{Note that all the baselines use the same BERT embeddings and the SimCLR model as XTQA for a fair comparison.}}
\label{tab.1}
\end{table*}
After generating the top $K$ evidence spans and their representations $e_i^{''} \in \mathbb{R}^{K \times d_1}$, the learned attention mechanism in Equation \ref{eq.2} is used to obtain the final evidence span representation $e_{i}^{'''} \in \mathbb{R}^{d_1 \times 1}$.
\subsection{Diagram Representation}
We use CNNs to obtain the $d_2$-dimensional representation $d_i^{'} \in \mathbb{R}^{d_2 \times 1}$ of $d_i$ as follows:
\begin{equation}
d_i^{'} = \mathrm{CNN_s}(d_i).
\end{equation}
The CNNs could be ResNet \citep{he2016deep}, SimCLR \citep{chen2020simple} and so on.
\subsection{Answer Prediction}
After the above modules, we obtain the word-level/global representations $q_i^{'}/q_i^{''}$ of $q_i$, the diagram representations $d_i^{'}$ of $d_i$, the global representations $a_{i,j}^{''}$ of $a_{i,j}$, the evidence span representations $e_i^{'''}$ of $e_{i}$ and the indexes $[\mathrm{START}(k), \mathrm{END}(k)]$ of spans.
The mentioned representations are fused to obtain the global fusion feature $g_{i,j} \in \mathbb{R}^{9d_1 \times 1}$ with $j$-th candidate answer as follows:
\begin{equation}
\begin{split}
& g_{i,j} = [q_i^{''}; d_i^{'}; a_{i,j}^{''}; e_i^{'''}; g_{i}^{\beta}; g_{i,j}^{\gamma}; g_{i}^{\mu}; g_{i,j}^{\eta};], \\
& g_{i}^{\beta} = \mathrm{BAN}(q_i^{'}, d_i^{'}), ~ g_{i,j}^{\gamma} = Wq_i^{''} \circ Wa_{i,j}^{''}, \\
& g_{i}^{\mu} = Wq_i^{''} \circ We_i^{'''}, ~ g_{i,j}^{\eta} = We_i^{'''} \circ Wa_{i,j}^{''}, \\
& g_{i,j}^{\psi} = Wq_i^{''} \circ Wa_{i,j}^{''} \circ g_{i}^{\beta},
\end{split}
\end{equation}
where $\mathrm{BAN}$ is the bi-linear attention networks \citep{kim2018bilinear}, $W \in \mathbb{R}^{d_1 \times d_1}$ is the learned weight matrix, $g_{i}^{\beta}, g_{i,j}^{\gamma}, g_{i}^{\mu}, g_{i,j}^{\eta} \in \mathbb{R}^{d_1 \times 1}$ denote the pairwise similarity, and $g_{i,j}^{\psi} \in \mathbb{R}^{d_1 \times 1}$ is the triple-wise similarity.
To obtain the estimation $\hat{a}_i$ of the correct answer, $g_{i,j}^\mathrm{T}$ is projected into a scalar score as follows:
\begin{equation}
\hat{a}_i = \mathop{\arg\max}_{j \in |\mathcal{A}_i|} \mathrm{softmax}(\mathrm{MLP_c}(g_{i,j}^\mathrm{T})).
\end{equation}
Eventually, not only the answer $\hat{a}_i$ but also the evidence span is provided for students by optimizing the cross-entropy loss function.
\section{Experiments}
\subsection{Datasets and Evaluations}
We use the accuracy to evaluate XTQA on the TQA dataset \citep{kembhavi2017you} that consists of $1,076$ lessons with $78,338$ sentences and $3,455$ diagrams (including very few natural images). The lessons are obtained from the Physical Science, Life Science and Earth Science textbooks of the middle school on-line curricula. The TQA dataset is split into a training set with $666$ lessons and $15,154$ questions, a validation dataset with $200$ lessons and $5,309$ questions, and a test set with $210$ lessons and $5,797$ questions. Among of the total $26,260$ questions, $12,567$ of them have an accompanying diagram. There are four candidate answers for each diagram question. The non-diagram questions can be classified into two categories: True/False (T/F) with two candidate answers and Multiple Choice (MC) with two to seven candidate answers.
\begin{table*}[!htbp]
\centering
\resizebox{\textwidth}{!}{\begin{tabular}{lccccc}
\toprule \toprule
Model & Non-diagram T/F & Non-diagram MC & Non-diagram All & Diagram & All \\ \midrule
BAN \cite{kim2018bilinear} & $48.080$ & $32.96$ & $38.44$ & $27.28$ & $32.11$ \\
MFB \citep{yu2017multi} & $48.081$ & $31.83$ & $37.72$ & $28.17$ & $32.30$ \\
MCAN \citep{yu2019deep} & $48.077$ & $33.15$ & $38.56$ & $27.56$ & $32.32$ \\
MUTAN \citep{ben2017mutan} & $48.075$ & $32.90$ & $38.40$ & $28.29$ & $32.67$ \\
CMR \cite{ZhengGK20} & $52.022$ & $33.15$ & $39.99$ & $29.54$ & $34.06$ \\
Challenge & $-$ & $-$ & \textbf{$42.08$} & $31.75$ & $36.22$ \\ \hline
XTQA & $\textbf{56.218}$& $\textbf{33.40}$ & $41.67$ & $\textbf{33.34}$ & $\textbf{36.95}$ \\ \bottomrule \bottomrule
\end{tabular}}
\caption{Experimental results ($\%$ accuracy) of different-type questions on the test split. We show the accuracy of True/False (Non-diagram T/F) questions, multiple-choice questions (Non-diagram MC), the whole non-diagram questions (Non-diagram All = Non-diagram T/F $\cup$ Non-diagram MC), multiple-choice diagram questions (Diagrams) and total questions (All = Non-diagram $\cup$ Diagram). \textbf{Note that all the baselines use the same BERT embeddings and the SimCLR model as XTQA for a fair comparison.}}
\label{tab.2}
\end{table*}
\subsection{Implementation Details}
In \emph{Question/Answer Representation}, we use BERT \citep{devlin2019bert} to obtain $768$-dimensional word embeddings and uni-directional one-layer GRUs with $d_1=1024$ hidden units to encode questions and candidate answers. The shared MLPs (FC($1024$)-Dropout($0.2$)-FC($1$)) are used to learn attention coefficients. In \emph{Evidence Span Generation}, the pylucene\footnote{https://lucene.apache.org/pylucene/} is used to conduct paragraphs indexing and searching. The maximum number of paragraphs $M$, the maximum number of sentences within each paragraph $L$, the maximum length of each sentence $O$ and the maximum number of evidence span $K$ are set to $1/1, 5/15, 20/15$ and $1/1$ for non-diagram/diagram question answering respectively. we set the maximum widths of candidate evidence spans to $2$. In \emph{Diagram Representation}, we resize the diagrams to $224$ owing to the different sizes of them in the TQA dataset. To obtain $d_2=2048$-dimensional diagram representations, we first train the SimCLR\footnote{We open this code on https://github.com/keep-smile-001/SimCLR-TQA-master} \citep{chen2020simple} on the diagrams within TQA dataset with default hyper-parameters, and then fine-tune the pre-trained model by the task-specific supervision (TQA). In \emph{Answer Prediction}, the MLPc (FC($2048$)-ReLU-Dropout($0.2$)-FC($1$)) is used to obtain the candidate answer scores.
XTQA is trained by the Adam optimizer with $\beta_1=0.9, \beta_2=0.98$. The base learning rate is $\min(2.5\tau e^{-4}, 1e^{-4})$, where $\tau$ is the current epoch. The rate is decayed by $0.1$ after $8$ epochs. XTQA converges at the end of the $10$-th epoch with the batch size $2$.
\subsection{Baselines}
We compare XTQA with several methods that focus on multi-modality fusion as follows\footnote{All the baselines are implemented in our open source code \emph{opentqa}.}:
\noindent \textbf{MFB} \citep{yu2017multi} is a multi-modal factorized bi-linear pooling approach. It aims at addressing the high dimensionality of the output features and the huge number of parameters caused by the bi-linear pooling based models \citep{tenenbaum1997}.
\noindent \textbf{MUTAN} \citep{ben2017mutan} is a multi-modal tensor-based decomposition approach with a low-rank matrix constraint. It also aims at addressing huge dimensionality issues.
\noindent \textbf{BAN} \citep{kim2018bilinear} is a bi-linear attention network that aims at learning effective interactions between images and questions using the proposed bi-linear attention mechanism.
\noindent \textbf{MCAN} \citep{yu2019deep} is a deep modular co-attention network. It aims at obtaining sufficient multi-modality interactions by modularly composing the self-attention of questions and images, as well as the question-guided-attention of images.
\noindent \textbf{CMR} \cite{ZhengGK20} is a cross-modality relevance module that can be used in an end-to-end framework. It learns the relevance representations between entities of input modalities and modeling the higher-order relevance between entity relations to perform language and vision reasoning.
\noindent \textbf{Challenge}\footnote{https://competitions.codalab.org/competitions/16931\#results} is a competition of the TQA task. We cite the best results on each type of questions of the test split. The results come from different teams.
\subsection{Results}
\begin{table*}[!htbp]
\centering
\resizebox{\textwidth}{!}{\begin{tabular}{@{}lcccccc@{}}
\toprule \toprule
Models & Non-diagram All & $\Delta$ & Diagram & $\Delta$ & All & $\Delta$ \\ \midrule
XTQA & $\textbf{41.32}$ & & $\textbf{32.05}$ & & $\textbf{36.46}$ & \\ \hline
w/o fine-tuning the SimCLR model & $\textbf{41.32}$ & $0$ & $30.36$ & $-1.69$ & $35.57$ & $-0.89$ \\
w/o the BERT embeddings & $41.20$ & $-0.12$ & $29.82$ & $-2.23$ & $35.24$ & $-1.22$ \\
w/o fine-grained evidence spans & $38.95$ & $-2.37$ & $30.36$ & $-1.69$ & $34.45$ & $-2.01$ \\ \bottomrule \bottomrule
\end{tabular}}
\caption{Ablation results ($\%$ accuracy) of the XTQA. Non-diagram All denotes the accuracy on the non-diagram questions. Diagram denotes the accuracy on the diagram questions. All denotes the accuracy on the total questions. $\Delta$ denotes the accuracy reduction without the specific module.}
\label{tab.3}
\end{table*}
The experimental results on the validation/test split are shown in Table \ref{tab.1}/Table \ref{tab.2} respectively.
In Table \ref{tab.1}, we can see that XTQA outperforms the best baseline MCAN by $1.36\%$ on the total questions of the validation split. For diagram questions, XTQA outperforms the best baseline MFB by $1.29\%$. For non-diagram questions, XTQA outperforms the best baseline MCAN by $1.22\%$. For T/F-type questions, XTQA outperforms the best baseline MFB by $6.51\%$. XTQA performs worst on the MC-type questions, which may be caused by the different data distributions between T/F (two candidate answers) and MC (four to seven candidate answers) questions. Note that all the baselines perform information fusion between top $1$ paragraph and questions for non-diagram questions considering the specificity of the TQA task.
In Table \ref{tab.2}, we can see that XTQA outperforms the best baseline Challenge by $0.73\%$ on the total questions of the test split. For diagram questions, XTQA outperforms the best baseline Challenge by $1.59\%$. For non-diagram questions, XTQA under-performs the best baseline Challenge by $0.41\%$. For TF-type questions, XTQA outperforms the best baseline CMR by $4.196\%$. All the methods perform poorly on this type of questions except XTQA and CMR, which may be caused by the different data distributions or the difficult questions. For MC-type questions, XTQA outperforms the best baseline MCAN and CMR by $0.25\%$. Note that XTQA achieves the best performance on the two splits, which shows the effectiveness of our method. In addition, the performance of baselines is different on the two splits, which may be caused by the following reasons. (1) The generation abilities of the baselines are different. (2) The data distribution is different between the validation split and test split. For example, the information of Lesson \emph{earth science and its branches} is mutually exclusive in the two splits.
\subsection{Ablation Studies}
We perform ablation studies as shown in Table \ref{tab.3} to verify the effectiveness of each module.
The accuracy on the diagram questions drops by $1.69\%$ without fine-tuning the SimCLR model, which shows that fine-tuning the pre-trained model helps models to learn effective diagram representations to improve the TQA performance.
The accuracy on the non-diagram/diagram questions drop by $0.12\%$ and $2.23\%$ respectively without the BERT embeddings, which shows the significant performance difference of the BERT embeddings on the different task. This phenomenon may be caused by unstable optimizing on the diagram question answering.
The accuracy on the non-diagram/diagram questions drop by $2.37\%$ and $1.69\%$ respectively without fine-grained evidence spans, which shows the effectiveness of our proposed coarse-to-fine grained evidence span generation algorithm.
In short, each component makes its contributions to the performance of the XTQA and our proposed coarse-to-fine grained evidence span generation algorithm plays the biggest role.
\subsection{Case Studies}
We conduct the case studies as shown in Figure \ref{fig.3} to present the strengths and weaknesses of our method intuitively.
\begin{figure*}[!htbp]
\centering
\includegraphics[width=\linewidth]{images/case-study.pdf}
\caption{Case studies to show the strengths and weaknesses of the XTQA. The left part shows the multi-modal context of the Lesson \emph{Seafloor Spreading}. The middle parts show the evidence generation processes by our proposed algorithm. The right part shows the questions of this lesson. These cases come from the validation split.}
\label{fig.3}
\end{figure*}
\noindent \textbf{Strengths.} XTQA can provide the explicit evidence span for students by the coarse-to-fine grained algorithm. In addition, it can generate evidence spans of different lengths according to the specific question. For example, XTQA provides evidence span of length $1$ of the diagram question, and provides evidence span of length $2$ of the non-diagram question for students respectively as shown in the middle part of Figure \ref{fig.3}.
\noindent \textbf{Weaknesses.} If the coarse-grained algorithm makes errors, it will cause the failure to find evidence spans in the fine-grained phase. For example, XTQA finds the wrong top $1$ paragraph for the non-diagram question as shown in the middle part of Figure \ref{fig.3}, which causes the failure to find the evidence span.
In short, XTQA gives students explainability to some extent, and optimizing the coarse-grained algorithm may be the further research direction.
\subsection{Discussions}
We make an assumption about the evidence generation in Section \ref{introduction} owing to the lack of supervision. To verify the effectiveness of this assumption, we take some samples that are answered correctly from the test split and manually check whether the generated evidence spans can provide certain explainability for students. The results are show in Figure \ref{fig.4}.
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.55\linewidth]{images/span-accuracy.pdf}
\caption{The accuracy of evidence spans on the non-diagram questions (NDQ), diagram questions (DQ), total questions (ALL).}
\label{fig.4}
\end{figure}
The accuracy on the non-diagram questions, diagram questions and total questions are $83.33\%$, $86.67\%$ and $85\%$ respectively. This shows the evidences for most of the questions answered correctly provide explainability for students and proves the validity of our assumption to some extent.
In addition, we find XTQA and other baselines have poor generalization abilities on the validation split and the test split, which may be caused by the following reasons. (1) Small datasets may easily lead to overfitting. (2) There may exist different data distributions between splits. For example, the information of Lesson \emph{earth science and its branches} is mutually exclusive in the data splits.
\section{Conclusion and Future Work}
In this paper, we propose a novel architecture towards span-level eXplanations of the Textbook Question Answering (XTQA). It can provide not only the answers but also the evidences to choose them for students. To generate evidence spans, we propose a coarse-to-fine grained algorithm. In the coarse phase, the algorithm uses the TF-IDF method to find top $M$ paragraphs relevant to questions within the multi-modal context. In the coarse-grained phase, it generates the top $K$ evidence spans by computing the information gain of each candidate evidence span to questions. Experimental results show that XTQA achieves the state-of-the-art performance on the validation split and the test split of the TQA dataset.
In the future, the following directions will be explored: (1) Error reduction of the coarse-grained algorithm may improve the accuracy of evidence span generation. We will explore how to devise an end-to-end architecture to optimize the process of the coarse-grained evidence span generation. (2) External knowledge helps to improve the performance of other tasks such as named entity recognition and visual question answering. An analysis about the TQA dataset \citep{kembhavi2017you} also shows that about $10\%$ of the questions need external knowledge to answer. We will explore how to integrate the external knowledge into XTQA to improve its performance.
|
\section{Introduction}
\label{sec:intro}
Visual tracking plays an active role in a wide range of applications, including driverless vehicles, robotics, surveillance, human-computer interaction, surveillance, motion analysis. Recent years have witnessed significant developments in visual tracking, where an enormous amount of research effort has gone into tasks such as short-term single-object tracking where the target appears in the first frame. However, many challenges remain, such as deformation, target rotation, scale variation and fast motion.
Correlation filters (CFs) have recently been introduced for visual-tracking and have been shown to achieve high speed as well as robust performance~\cite{bertinetto2016staple,bolme2010visual,henriques2012exploiting,boddeti2013correlation,danelljan2017discriminative,danelljan2015learning,galoogahi2017learning,danelljan2017eco}. For instance, David S. Bolme et al.~\cite{bolme2010visual} proposed one based on the minimum output sum of squared errors (MOSSE), which works by manipulating the maximum cross-correlation response between the model and the candidate patch. Henriques et al.~\cite{henriques2012exploiting} exploited the circulate structure and Fourier transformation in a kernel space (so-called the CSK tracker), offering excellent performance on a range of computer vision problems. A vector correlation filter (VCF) was devised by Boddeti et al.~\cite{boddeti2013correlation} to minimize localization errors while improving the tracking speed. The DSST tracker~\cite{danelljan2017discriminative} employs adaptive multi-scale correlations filters using HOG features in an effort to handle scale changes in target objects. To learn a model that is inherently robust to both color changes and deformations, Staple~\cite{bertinetto2016staple} combines two image patch representations that are sensitive to potential competing factors such as rotation and deformation. Danelljan et al.~\cite{danelljan2015learning} utilized a spatial-regularization component to penalize correlation filter coefficients as a function of their spatial location. The work in~\cite{galoogahi2017learning} presents a background-aware correlation filter (BACF) that can model how the background, as well as the foreground, of an object may vary over time. To drastically reduce the number of modeling parameters, ECO~\cite{danelljan2017eco} introduces a factorized convolution operator into traditional CF, where a compact generative model on the training sample distribution significantly reduces the memory and time complexity, while enhancing sample diversity.
Recognizing the success of deep convolutional neural networks (CNNs) on a wide range of visual-recognition tasks, a number of tracking methods based on deep features and correlation filters have been developed ~\cite{ma2015hierarchical,qi2016hedged}. Empirical studies using large object-tracking benchmark datasets show that such CNN-based trackers perform favorably against methods based on the use of hand-crafted features. However, extracting CNN features from each frame, and training or updating CF trackers over high dimensional deep features, is computationally expensive. Therefore for visual tracking, such an approach often leads to poor real-time performance.
\textbf{Contribution}: In this paper, we propose a robust correlation tracking method (RCT) via the exploitation of feature fusion and reliable response. A fused feature herein describes the gradient and color information conjunctively in a more natural way as compared to existing techniques. The novel fused features are then embedded into a correlation filter that is background-aware (in the sense that the filter is capable of learning from real, negative examples densely extracted from the background). In addition, an adaptive optimization strategy is introduced to remove the untrusted part of the response map that is caused by deformation or other challenging factors.
We evaluate the proposed tracker on the OTB50, OTB100 and Temple-Color128 datasets. The results demonstrate that our method obtains a very competitive accuracy level in comparison with the state-of-the art CF-based trackers, but does so with a real-time tracking speed of 17 FPS on a standard desk-top CPU.
\section{Related Work}
This section first provides an overview of feature representation in correlation tracking and then outlines the existing methods for addressing model drift that are most relevant to the present work.
\noindent\textbf{Feature representation in correlation tracking.} It is critical to employ an efficient feature-representation mechanism in order to improve performance during object-tracking. Gradient and color features are the most popular single types of feature. In particular, color features help capture rich color characteristics, while features reflecting the histogram of oriented gradient (HOG)~\cite{dalal2005histograms} are adept in capturing abundant gradient information. Based on these feature descriptions, a variety of techniques on target-tracking have been proposed. For instance, CSK~\cite{henriques2012exploiting} exploits the circulant structure while only utilizing the grey-level features over single channels. Upon this basis, the kernelized correlation filter (KCF)~\cite{henriques2015high} employs HOG features to improve filter performance. Danelljan et al.~\cite{danelljan2014adaptive} investigated color attributes to reinforce correlation-filter tracking, drawing the conclusion that using CN features can greatly improve the performance of correlation filters. Also, the powerful features including HOG and color-naming have been integrated together to further boost the overall tracking performance in~\cite{li2014scale}. Furthermore, applications of features extracted by CNNs~\cite{ma2015hierarchical} have been widely studied, preliminarily in performing the task of visual tracking, but are still limited due to expensive computational costs incurred. As such, whilst the aforementioned techniques have shown good potential, how to jointly utilize different features to increase the tracking performance remains an open issue.
\noindent\textbf{Robustness to model drift.} Model drifts lead to inaccurate model-based predictions. In addressing this problem, Kalal et al.~\cite{kalal2012tracking} proposed an approach that decomposed the ultimate task of tracking into subtasks of tracking, learning and detection (TLD), where tracking and detection reinforce each other. Also, the technique of Multiple-Instance Learning~\cite{babenko2011robust}(MILTrack) was introduced to train with bags of positive examples. Recently, discriminative correlation filters (DCF) which relies on a periodic assumption of the training and detection samples have been shown to be able to achieve tracking by employing a circular correlation. However, the circulant shifted samples in such trackers suffer from periodic repetitions on boundary positions, thereby leading to model drift and significantly degrading the tracking performance. Spatial regularization methods have since been suggested to alleviate the unwanted boundary effects. For example, using the Alternating Direction Method of Multipliers (ADMM), Galoogahi et al.~\cite{galoogahi2015correlation} resolved a constrained optimization problem for single-channel DCF. Somewhat differently, the SRDCF formulation~\cite{danelljan2015learning} allows correlation filters to be trained on a significantly larger set of negative training samples, without corrupting the positive samples, where a spatial regularization component is introduced to the training process to penalize the correlation filter coefficients in relation to their spatial location. Unlike previous DCF-based trackers in which negative examples are restricted to circular shifted patches, BACF~\cite{galoogahi2017learning} utilizes a correlation filter whose spatial size is much smaller than that of the training samples; real negative training examples, densely extracted from the background are utilized. These works have aimed to prevent model drift through modifying the training strategy rather than improving the underlying model-based predictions themselves. In the present work, we improve the predictions by obtaining and manipulating a more reliable response map, leading to an enhanced tracking result.
Compared with the existing methods, our proposed tracker has several merits: (1) while RCT may be viewed as an (improved) approximation to the work of~\cite{galoogahi2017learning} on multiple training samples, the filter works more efficiently owing to the use of a more reliable response map; and (2) with the introduction of fused features, the RCT tracker can learn more robust features than the previous work, thereby leading to superior tracking performance.
\begin{algorithm}
\caption{Tracking algorithm}\label{algorithm}
\KwData{Frames $I_f$, initial target location $p_1$ ( $f$ means the number of the current frame)}
\KwResult{Target location $p_f$}
\textbf{repeat}\;
Crop an image region from $I_f$ at the last location and extract its fused-feature vector $x_f$\;
Compute the optimum correlation filter (via Eq.2) and obtain the original response map\;
Construct the mask to yield a reliable response map\;
Detect the target location $p_f$ via the reliable response map\;
Estimate the scale of the target and update the tracking model (as summarized in Section 3.4)\;
\textbf{until} end of video sequence\;
\end{algorithm}
\section{Proposed Approach}
We aim to develop a robust tracking algorithm that is adaptive to significant changes without being prone to drifting. We first propose a fused feature mechanism which describes the gradient and color information in an integrated way. Then, a background-aware correlation filter based on the exploitation of fused features is designed to obtain a response map. Furthermore, the mask obtained according to the value of the response map will be multiplied with a given original image to form a more reliable response map, which help alleviate possible model drifts. The proposed method is presented in Algorithm \ref{algorithm}.
\subsection{Multi-channel Fused Features}
Features play an important role in computer vision; much of the impressive progress in object detection can be attributed to the improvement in the representation power of features. Gradient and color features are the most widely exploited in object detection and tracking. Indeed, previous work~\cite{khan2013discriminative} has verified that there exists a strong complementarity between gradient and color features.
How to best utilize different features jointly for visual tracking is still an open question. In the medical profession, Duplicity theory~\cite{hecht1935intermittent} concerns the comparison (in terms of both differences and similarities) and interaction between the cone and rod systems in the visual pathways, with the assumption that the cone system is active during daylight vision and the rod system functions in low light (night time). The correct (whether cone or rod) system is first selected according to the luminance of environment; then, the chosen system will work to obtain detailed information of the surroundings.
Inspired by the duplicity theory of vision, we construct a more natural feature representation. In our setting, instead of concatenating the color and gradient features directly, we first transform the original image into HSV (Hue, Saturation, Value) color space, which is based more upon how colors are organized and conceptualized in human vision. In such a color space, brightness and colorfulness are absolute measures, which usually describe the spectral distribution of light entering the eye. Benefiting from this, our fused feature performs robust to the illumination variation. Secondly, HOG gradient information is extracted from each channel of the HSV color space, separately. Finally, all the HOG features are concatenated to form our proposed fused feature, in the form of a matrix. Hence, for terminology, we think of the output feature as the combination of fused-(input)-features or, as a (singular) fused feature. Without losing generality and for conciseness, we term the resultant feature descriptor a fused feature. The process to extract the final fused feature is shown in Figure \ref{fig:feature}.
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{feature.eps}
\caption{Extraction of our fused features.}
\label{fig:feature}
\end{figure}
\subsection{CF Tracking through Fused Features}
In this section, we introduce our fused feature into background-aware correlation filter \cite{galoogahi2017learning} to construct a better correlation-tracking framework. We utilize a correlation filter with a spatial size which is smaller than the size of training examples to reduce the boundary effects. Denote $x_k$ as the fused feature vector of a cardinality ${x_k}\in {R^T}$ , respectively. We consider $y \in {R^T}$ as the desired correlation output corresponding to a given sample $x_k$ . A correlation filter $w$ with the dimensionality of $D$ (where $T > >D$) is then learned by solving the following minimization problem:
\begin{align}
E(w) = \sum\limits_{j = 1}^T {||{y_j} - \sum\limits_{k = 1}^K {{w_k}^{\top}\cdot{\rm{P}}{x_k}[\Delta {\tau _j}]} |{|^2} + \lambda \sum\limits_{k = 1}^K {||{w_k}||_2^2}}
\label{Eq:1}
\end{align}
where $\lambda$ is a regularization parameter, and ${\rm{P}}{x_k}[\Delta {\tau _j}]$ generates all circular shifts of size $D$ from the entire frame over all $j = \left[ {0,...,T - 1} \right]$ steps.
Note that the Eq.\ref{Eq:1} can be readily transformed into frequency domain (using discrete Fourier transform) in order to improve the computational efficiency. We introduce $\hat g = {[\hat g_1^{\rm T}, \cdot \cdot \cdot ,\hat g_K^{\rm T}]^{\rm T}}$ as an auxiliary variable. The trained filter in the frequency domain will be written as
\begin{align}
\begin{array}{l}
E(w,\hat g) = ||\hat y - \hat X\hat g||_2^2 + \lambda ||w||_2^2\\
s.t.\hat g = \sqrt T ({\rm{F}}{{\rm{P}}^{\top}} \otimes {{\rm{I}}_k})w
\end{array}
\label{Eq:2}
\end{align}
where $\hat X$ is denoted by $\hat X = {[diag{({\hat x_1})^{\top}},...,diag{({\hat x_k})^{\top}}]^{\top}}$, $I_k$ is the $K\times K$ identity matrix, and $\otimes$ denotes the Kronecker product. In particular, $\hat A $ represents the Discrete Fourier Transform (DFT) of a signal $A$,where $F$ is the orthonormal $T\times T$ matrix of complex basis vectors, mapping any $T$-dimensional vectorized signal to its Fourier domain. The transpose operator $\top$ on a complex vector or matrix gives the conjugate transpose.
By directly employing the Augmented Lagrangian Method (ALM)~\cite{galoogahi2017learning}, we can solve Eq.\ref{Eq:2} and obtain the required correlation filter ${\hat g^{(f - 1)}}$, where $f$ is the current frame number.
\subsection{Object Location by Reliable Response}
Representing the response value of every pixel, the response map $R$ in frame $f$ can be computed by applying the filter ${\hat g^{(f - 1)}}$ that has been updated in the previous frame. Due to the challenges typically faced in performing real-world tracking tasks, such as deformation and rotation, the similarity between the target and modeling template may be decreased, leading to great risk of model drift or locating mistakenly.
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{reliable_mp.PNG}
\caption{Object location by reliable response.}
\label{fig:reliable_response_map}
\end{figure}
How to remove a lot of potentially misleading redundant information (responses to similar objects) contained in the original response map then? As shown in Figure \ref{fig:reliable_response_map}, when noise exists, the position with the maximum value in the response map does not necessarily correspond to the real target. In this case, simply taking the position with the highest response as the target position is rather unreliable. Through a large number of experiments, empirically we find that the response peak of the real target often changes gradually, while the response peak of the disturbed object is usually very steep and looks very abrupt. Accordingly, in order to exclude the anomalies, we first try to identify the target proposals which are associated with a relatively high value in the response map. In order to achieve this goal, we set a threshold $Th$, which divides the response map into two parts. The pixels with a gray value greater than $Th$ belong to the target proposal set $A$ and the remaining are deemed to attribute to the background part $B$. The number of pixels contained in the two parts is represented with $NA$ and $NB$ respectively. We vary $Th$ from 0 to 255, each time, $NA$ and $NB$ are counted to calculate the ratio of the target proposals $P_f$ in the current frame, such that
\begin{align}
\begin{array}{l}
{P_f} = \frac{{{N_A}}}{{{N_A} + {N_B}}}
\end{array}
\label{Eq:3}
\end{align}
Repeat Eq.\ref{Eq:3} until the difference between the $P_f$ and a predefined value $P$ is less than the error range $G$,
\begin{align}
\begin{array}{l}
|{P_f} - P| < G
\end{array}
\label{Eq:4}
\end{align}
When Eq.\ref{Eq:4} is satisfied, the grey value of pixels in the set $A$ is reset to 255, while each of the rest pixels is set to 0. From this, a number of connected domains are obtained. Then, any connected domain whose area is less than a fixed threshold $AR$ is deleted to form the fine mask $K$. By multiplying $K$ with the original response map $R$, the reliable response map results. Finally the position with the maximum value in the reliable response map is treated and recognized as the target location.
\subsection{Model Updating and Scale Estimation}
To obtain a robust approximation, we update the model of the correlation filter $M$ at the $f-th$ frame using a moving average:
\begin{align}
\begin{array}{l}
M_{}^f = (1 - \eta ){M^{f - 1}} + \eta {M^f}
\end{array}
\label{Eq:5}
\end{align}
where $\eta$ is the learning rate.
In order to be adaptable to any change of the scale of a target, the filter is applied on multiple resolutions of the searching area where tracking takes place~\cite{danelljan2015learning}. The searching area has the same spatial size as that of the filter ${\hat g}$ . This returns $S$ correlation outputs with different scales, where $S$ is the number of scales. The scale with the maximum correlation output is used to update the object location and the subsequent scale.
\section{Experimental Results}
In order to present an objective evaluation regarding the performance of the proposed approach, we examine our RCT tracker on three standard datasets,
including OTB50~\cite{WuLimYang13}, OTB100~\cite{wu2015object}, and Temple-Color128 (TC128)~\cite{Liang2015Encoding}. Both the general capability and the special scenarios-handling ability are tested. We compare our tracker with a range of excellent state-of-the-art trackers, including: SRDCFdecon \cite{danelljan2016adaptive}, MCPF \cite{zhang2017multi}, ECO\_HC \cite{danelljan2017eco}, BACF \cite{galoogahi2017learning}, CF\_AT \cite{bibi2016target}, CACF \cite{mueller2017context}, BIT \cite{cai2016bit}, fDSST \cite{danelljan2017discriminative}, Staple \cite{bertinetto2016staple}. Different metrics may be used for evaluation depending on preferred perspectives, amongst which one-pass evaluation (OPE) is arguably the most commonly used evaluation method. OPE runs a tracker on each sequence once: it initializes a tracker using the ground truth object state in the first frame, and reports the average precision or success rate of all subsequent results. Having recognized this, OPE is also used herein to comparatively evaluate the present work.
\subsection{Implementation Details}
We evaluate our tracker on the challenging Visual Tracker Benchmark \cite{wu2015object}, by following rigorously the existing evaluation protocols. The experiments are performed in Matlab2014a on an Inteli5 3.2GHz CPU with 4G RAM. In all the experiments carried out, we use the same parameter values for all image sequences. We employ 31-channel HOG features using 4 x4 cells to obtain the fused feature. The regularization factor is empirically set to 0.001 and the number of scales is set to 5 with a scale-step of 1.01. A 2D Gaussian function with bandwidth of $\sqrt {wh/16}$ is used to define the correlation output for an object of size $[h,w]$. The learning rate ($\eta$ ) of the correlation filter is 0.013.The area threshold $AR$ is set to 20\%, and the error range $G$ is set to be within 10.
\subsection{Evaluation on OTB Dataset}
We implement the one-pass experiment on the OTB50 and OTB100 benchmark datasets. All these image sequences are annotated with 11 attributes which cover various challenging factors, including scale variation (SV), occlusion (OCC), illumination variation (IV), motion blur (MB), deformation (DEF), fast motion (FM), out-of plane rotation (OPR), background clutters (BC), out-of-view (OV), in-plane rotation (IPR) and low resolution (LR).
Let $r_T$ represent the tracking result of trackers and $r_G$ denote the given groudtruth, the success plots below show the ratios of successful frames whose overlap score ($OS$) is defined as $OS{\rm{ = }}\frac{{{r_T} \cap {r_G}}}{{{r_T} \cup {r_G}}}$. The trackers are ranked by precision values (with a threshold of 20) in the precision plots, and by the area-under-curve (AUC) depicted by the success plots. If $OS>0.5$, the tracking result is deemed correct.
\begin{figure}
\centering
\subfigure[\tiny{Success plots on OTB50}]{
\label{fig:subfig:a}
\includegraphics[width=0.235\linewidth]{S_50.png}}
\subfigure[\tiny{Precision plots on OTB50}]{
\label{fig:subfig:b}
\includegraphics[width=0.235\linewidth]{P_50.png}}
\subfigure[\tiny{Success plots on OTB100}]{
\label{fig:subfig:a}
\includegraphics[width=0.235\linewidth]{S_100.png}}
\subfigure[\tiny{Precision plots on OTB100}]{
\label{fig:subfig:b}
\includegraphics[width=0.235\linewidth]{P_100.png}}
\caption{Results of the proposed tracker and other compared trackers on OTB dataset.}
\label{fig:OTB}
\end{figure}
Figures \ref{fig:OTB} shows the success and precision plots on the OTB50 and OTB100 dataset, respectively. Only the top 10 trackers are displayed for clarity. Overall, our RCT is better than the other state-of-the-art trackers and achieves a significant gain in the OTB100 dataset. BIT is a tracker that extracts low-level biologically-inspired features while imitating an advanced learning mechanism to combine generative and discriminative models for target location. Our RCT improves on BIT by an average of 27.58\% in the AUC scores on OTB50, and by 19.22\% on OTB100. This testifies to the extraordinary performance of the fused features. Since ECO\_HC ranks first amongst all of the HOG-based trackers, it makes for a good representation of the existing trackers compared. Our RCT improves on ECO\_HC in the average AUC scores on both datasets. The BACF and SRDCFdecon trackers primarily focus on the boundary effects, with reported speeds of 30 FPS and 1 FPS respectively. Our RCT employs the ''background-aware'' mechanism from BACF, but performs favorably over BACF as well as SRDCFdecon.
\subsection{Evaluation on TC128 dataset}
TC128 is the first comprehensive color-tracking benchmark. The results of ten trackers on the 128 videos are summarized in Figure \ref{fig:TC128}, which shows the average rates of both the success plots and the precision plots. As can be seen from these results, our tracker always performs reliably and can achieve the optimal or at the least, very close to the best with respect to both metrics. Note that MCPF (which ranks highly) utilizes deep features obtained using the VGG-19 convolutional neural network. Extracting CNN-features from each frame and training or updating CF trackers over high dimensional deep features is computationally expensive (which results in poor real-time performance). Although the average AUC scores of our tracker is a little weaker compared with MCPF, it runs more than 34 times faster as compared to MCPF, which has a reported speed of 0.5 FPS.
\begin{figure}
\centering
\subfigure[Success plots]{
\label{fig:subfig:a}
\includegraphics[width=0.44\linewidth]{S_128.png}}
\subfigure[Precision plots]{
\label{fig:subfig:b}
\includegraphics[width=0.45\linewidth]{P_128.png}}
\caption{Results of proposed tracker and other compared trackers on TC128 dataset.}
\label{fig:TC128}
\end{figure}
\subsection{Further State-of-the-art Comparison}
Table \ref{tab:1} compares our method with the state-of-the-art CF based trackers on the OTB50, OTB100 and TC128 datasets, where our RCT achieves the highest accuracy across all HOG-based trackers and, including even MPCF, which utilizes deep features (and hence involves substantially more computation). More specifically, as shown in this table, RCT achieves the best AUC (76.99\%) on OTB100 followed by ECO\_HC (76.60\%) and SRDCFdecon (76.44\%). On the OTB50 dataset, RCT (75.96\%) outperforms ECO\_HC (74.44\%) and SRDCFdecon (73.90\%). Finally, our RCT (72.92\%) is ranked first on average, closely followed by ECO\_HC (72.86\%), which is the winner of the comparison on TC128.
\begin{table}
\large
\centering
\caption{Success rates (OS$>$0.50) of the proposed trackers versus other state-of-art trackers. The first, second and third best methods are shown in color (The red ranks first, green means second and blue means third).}
\begin{tabular}{ccccc}
\hline
\multicolumn{1}{r}{} & \multicolumn{1}{p{4em}}{OTB50} & \multicolumn{1}{p{4em}}{OTB100} & \multicolumn{1}{p{3em}}{TC128} & \multicolumn{1}{p{6em}}{Avg.succ.rate} \\
\hline
SRDCFdecon & \textcolor[rgb]{ 0, .69, .941}{73.90} & \textcolor[rgb]{ 0, .69, .941}{76.44} & 64.05 & \textcolor[rgb]{ 0, .69, .941}{71.46} \\
MCPF & 71.44 & 75.80 & \textcolor[rgb]{ 0, .69, .314}{66.69} & 71.31 \\
ECO\_HC & \textcolor[rgb]{ 0, .69, .314}{74.44} & \textcolor[rgb]{ 0, .69, .314}{76.60} & \textcolor[rgb]{ 1, 0, 0}{67.54} & \textcolor[rgb]{ 0, .69, .314}{72.86} \\
BACF & 75.91 & 76.54 & 61.64 & 73.36 \\
CF\_AT & 64.21 & 67.52 & 61.40 & 64.37 \\
CACF & 69.97 & 72.50 & 63.57 & 68.68 \\
BIT & 48.39 & 57.77 & 53.83 & 53.33 \\
fDSST & 63.95 & 66.87 & 52.10 & 60.97 \\
Staple & 67.18 & 69.64 & 62.36 & 68.39 \\
RCT & \textcolor[rgb]{ 1, 0, 0}{75.97} & \textcolor[rgb]{ 1, 0, 0}{76.99} & \textcolor[rgb]{ 0, .69, .941}{65.76} & \textcolor[rgb]{ 1, 0, 0}{72.91} \\
\hline
\end{tabular}%
\label{tab:1}%
\end{table}%
\section{Conclusion}
In this paper, we have proposed a robust correlation filter-based tracking method via the use of multi-channel fused features and reliable response maps. The correlation filter that utilizes multi-channel fused features leads to a significant improvement in tracking performance while dealing with challenging factors such as deformation and scale variation. We have also proposed a novel strategy to obtain a more reliable response map, thereby locating the target through it. This allows our tracker to reduce the probability of incorrect locating when target occlusion and rotation exist severely. Comparative experimental investigations have proven, both quantitatively and qualitatively that our approach outperforms the state-of-the-art tracking methods. Future work will involve investigating more powerful fused features that combine intensity and color information while keeping the speed sufficiently fast for real-time applications.
\bibliographystyle{unsrtnat}
|
\section{#1}\lb{#2}} \def\sscn#1#2{\subsection{#1}\lb{#2}}
\def\begin{flushleft}{\begin{flushleft}}
\def\end{flushleft}{\end{flushleft}}
\def\begin{flushright}{\begin{flushright}}
\def\end{flushright}{\end{flushright}}
\def\begin{center}{\begin{center}}
\def\end{center}{\end{center}}
\def\begin{equation}{\begin{equation}}
\def\end{equation}{\end{equation}}
\def\begin{subequations}{\begin{subequations}}
\def\end{subequations}{\end{subequations}}
\def\begin{eqnarray}{\begin{eqnarray}}
\def\end{eqnarray}{\end{eqnarray}}
\def\baa#1{\begin{array}{#1}}
\def\end{array}{\end{array}}
\def} \def\bw{\begin{widetext}{} \def} \def\bw{\begin{widetext}{\begin{widetext}}
\def} \def\ew{\end{widetext}{} \def} \def\ew{\end{widetext}{\end{widetext}}
\def\nonumber {\nonumber }
\def\lb#1{\label{#1}}
\def\begin{itemize}{\begin{itemize}}
\def\end{itemize}{\end{itemize}}
\def\begin{comment}} \def\eco{\end{comment}{\begin{comment}} \def\eco{\end{comment}}
\def\begin{cases}{\begin{cases}}
\def\end{cases}{\end{cases}}
\newcommand{\vecs}[1]{\mathbf{#1}}
\newcommand{\bullet}{\bullet}
\def\Sign#1{\, \text{sign}\left(#1\right) }
\def\diag#1{\, \text{diag}\left(#1\right) }
\def\Der#1#2{\frac{d #1}{d #2}}
\def\pDer#1#2{\frac{\partial #1}{\partial #2}}
\def\fDer#1#2{\,\frac{\delta #1}{\delta #2}}
\def\Lim#1#2{\, \lim\limits_{#1 \to #2}}
\def\Int#1#2{\, \int\limits_{#1}^{#2}}
\def\,\verb|!|{\,\verb|!|}
\def\JacobiSn#1#2{\, \text{sn}\left(#1,\,#2 \right) }
\def\JacobiCn#1#2{\, \text{cn}\left(#1,\,#2 \right) }
\def\Sech#1#2{\, \text{sech}^{#1}\left(#2 \right) }
\def\Exp#1#2{\, \text{exp}^{#1}\left[#2 \right] }
\def\Cosech#1#2{\, \text{cosech}^{#1}\left(#2 \right) }
\def\Cosh#1#2{\, \text{cosh}^{#1}\left(#2 \right) }
\def\Coth#1#2{\, \text{coth}^{#1}\left(#2 \right) }
\def\Cos#1#2{\, \text{cos}^{#1}\!\left(#2 \right) }
\def\Sinh#1#2{\, \text{sinh}^{#1}\left(#2 \right) }
\def\Sin#1#2{\, \text{sin}^{#1}#2}
\def\Arccosh#1{\, \text{arccosh}\, #1}
\def\Kummer#1#2#3{\,_1F_1\!\left(#1, #2; #3 \right)}
\def\Auxfun#1#2{\, \breve{Z}\left(#1, #2 \right)}
\def\Prlog#1{\,{\cal W}\!\left(#1 \right)}
\def{\cal U}_\Psi{{\cal U}}
\def\mathbb{U}{\mathbb{U}}
\def{\rm U}{{\rm U}}
\def{\rm H}{{\rm H}}
\def{\rm T}{{\rm T}}
\defm^{-1}{\eta}
\def{\cal S}{{\cal S}}
\defV{{\cal V}}
\def{\ell_0}{{\ell_0}}
\def{N}{{\cal N}}
\def\textbf{a}{\textbf{a}}
\def{M}{{\cal M}}
\def{\cal L}{{\cal L}}
\def{V}{{V}}
\def\epsilon{{E}}
\def\boldsymbol{\nabla}{\boldsymbol{\nabla}}
\def\ve#1{\boldsymbol{#1}}
\defd{d}
\def\drm\vol{\drmV}
\def{\bar d}{{\bar d}}
\def\delta{\delta}
\def\rho} \def\dnc{\bar{\dn}{\rho} \def\dnc{\bar{\rho} \def\dnc{\bar{\dn}}}
\def\bar{\ell}{\bar{\ell}}
\def\len#1{\ell_{#1}}
\def\epsilon{\epsilon}
\def\vartheta{\vartheta}
\def\sigma{\sigma}
\defq{q}
\def\vena^2}%_D}\def\lapl{\Delta}%_D{\boldsymbol{\nabla}^2
\defb_0{b_0}
\def{\cal U}_\Psi{{\cal U}_\Psi}
\defT_\Psi{T_\Psi}
\def\tilde{T}_\Psi - \tilde{T}_\Psi^{(0)}{\tilde{T}_\Psi - \tilde{T}_\Psi^{(0)}}
\def{T}_\Psi - {T}_\Psi^{(0)}{{T}_\Psi - {T}_\Psi^{(0)}}
\def{\rm U}{{\rm U}}
\def{\rm H}{{\rm H}}
\def{\rm T}{{\rm T}}
\def\Psi_{\text{vac}}{\Psi_{\text{vac}}}
\def\Psi_\circ{\Psi_{(0)}}
\def\Phi{\Phi}
\def\grpot} %\def\grpott{\grpot_{(\Psi)}{\Phi}
\def\grpot_{\text{smi}}{\Phi_{\text{smi}}}
\def\grpot_{\text{RN}}{\Phi_{\text{RN}}}
\def\grpot_{\text{N}}{\Phi_{\text{N}}}
\def\grpot_{\text{gal}}{\Phi_{\text{gal}}}
\def\grpot_{\text{mgl}}{\Phi_{\text{mgl}}}
\def\grpot_{\text{dS}}{\Phi_{\text{dS}}}
\defL_{\text{smi}}{L_{\text{smi}}}
\defL_{\text{RN}}{L_{\text{RN}}}
\defL_{\text{N}}{L_{\text{N}}}
\defK_{\text{gal}}{K_{\text{gal}}}
\defL_{\text{mgl}}{L_{\text{mgl}}}
\defL_{\text{dS}}{L_{\text{dS}}}
\defL_\chi{L_\chi}
\defR_{\text{dS}}{R_{\text{dS}}}
\def\rho_{\text{dS}}{\rho_{\text{dS}}}
\def\phi_{\text{mgl}}{\phi_{\text{mgl}}}
\def{\mathfrak M}{{\virgo}}
\def{\mathfrak M}{{\text{mgl}}}
\def{\mathfrak M}{{M}}
\def{\mathfrak M}{{\mathfrak M}}
\defr{r}
\def\ve{r}{\ve{r}}
\def\chi{\chi}
\def{\phi}{{\phi}}
\defR{R}
\def\rho_\grpott{\rho_\grpot} %\def\grpott{\grpot_{(\Psi)}}
\def\tilde{\rho}{\tilde{\rho}}
\def{\dnir}_\grpott{{\tilde{\rho}}_\grpot} %\def\grpott{\grpot_{(\Psi)}}
\def\adn#1{\widetilde{\Omega}_{\text{#1}}}
\def\sigma^2} %\Omega_{(2)}^2{\sigma^2}
\begin{document}
\preprint{\scriptsize \it
17th Russian Gravitational Conference - International Conference on Gravitation, Cosmology and Astrophysics (RUSGRAV-17)~}
\preprint{\footnotesize Universe \textbf{6}, 180 (2020)
\ \
[\href{https://doi.org/10.3390/universe6100180}{DOI: 10.3390/universe6100180}]
}
\title{ ~\\
An alternative to dark matter and dark energy:
Scale-dependent gravity\\ in superfluid vacuum theory}
\author{Konstantin G. Zloshchastiev}
\email{https://orcid.org/0000-0002-9960-2874}
\affiliation{Institute of Systems Science, Durban University of Technology, P.O. Box 1334, Durban 4000, South Africa
\begin{abstract}
We derive an effective gravitational potential,
induced by the quantum wavefunction of a physical vacuum of a self-gravitating configuration, while the vacuum itself is viewed as the superfluid described by the logarithmic quantum wave equation. We determine that gravity has a multiple-scale pattern, to such an extent that one can distinguish sub-Newtonian, Newtonian, galactic, extragalactic and cosmological terms. The last of these dominates at the largest length scale of the model, where superfluid vacuum induces an asymptotically Friedmann-Lema\^itre-Robertson-Walker-type spacetime, which provides an explanation for the accelerating expansion of the Universe.
The model describes different types of expansion mechanisms, which could explain the discrepancy between measurements of the Hubble constant using different methods.
On a galactic scale, our model explains the non-Keplerian behaviour of galactic rotation curves, and also why their profiles can vary depending on the galaxy. It also makes a number of predictions about the behaviour of gravity at larger galactic and extragalactic scales. We demonstrate how the behaviour of rotation curves varies with distance from a gravitating center, growing from an inner galactic scale towards a metagalactic scale: a squared orbital velocity's profile crosses over from Keplerian to flat, and then to non-flat. The asymptotic non-flat regime is thus expected to be seen in the outer regions of large spiral galaxies.
\end{abstract}
\date{received:
17 June 2020 [RG], 29 August 2020 [MDPI], 26 November 2020 [arXiv]}
\pacs{04.60.Bc, 95.35.+d, 95.36.+x, 03.75.Kk\\
~\\
Keywords: quantum gravity; cosmology; superfluid vacuum; emergent spacetime; dark matter; galactic rotation curve;
quantum Bose liquid; logarithmic fluid; logarithmic wave equation}
\maketitle
\scn{Introduction}{s:intro}
Astronomical observations over many length scales
support the existence of a number of novel phenomena, which are usually attributed
to dark matter (DM) and dark energy (DE).
Dark matter was introduced to explain a range of observed phenomena
at a galactic scale, such as flat rotation curves,
while
dark energy
is expected to account for cosmological-scale dynamics,
such as the accelerating expansion of the Universe.
For instance, the~$\Lambda$CDM model, which is currently the most popular approach
used
in cosmology and galaxy-scale astrophysics,
makes use of both DE and cold DM concepts~\cite{car01}.
In spite of being a generally successful framework purporting to explain the large-scale structure of the Universe,
it currently faces certain challenges ~\cite{bb17,ty18}.
There is also growing consensus that a convincing theory of DM- and DE-attributed phenomena cannot
be a stand-alone model; but should, instead, be a part of a fundamental theory involving all known interactions.
In turn, we contend that formulating this fundamental theory will be impossible without
a clear understanding of the dynamical structure of the physical vacuum,
which underlies all interactions that we know of.
Moreover, this theory must operate at a quantum level, which~necessitates us rethinking
of the concept of gravity using basic notions of quantum~mechanics.
One of the promising candidates for a theory of physical vacuum is superfluid vacuum theory (SVT),
a post-relativistic approach to high-energy physics and gravity.
Historically,
it evolved from Dirac's idea of
viewing the physical vacuum as a nontrivial quantum object,
whose phase and derived velocity are non-observable in a {quantum-mechanical sense
~\cite{dir51}.
The term `post-relativistic',
in this context,
means that SVT can generally be a non-relativistic theory;
which nevertheless contains relativity
as a special case, or~limit, with~respect to some dynamical value
such as momentum
(akin to general relativity being a superset of the Newton's theory of gravity).
Therefore, underlying three-dimensional space would not be physically observable
until an observer goes beyond the above-mentioned limit, as~will be discussed in more detail later in this~article.
The dynamics and structure of superfluid vacuum are being studied, using various approaches
which agree upon the main paradigm (physical vacuum being a background quantum liquid of a certain kind,
and elementary particles being excitations thereof),
but differ in their physical details, such as an underlying model of the
liquid~\cite{volbook,huabook,z11appb}.
It is important to work with a precise definition of superfluid,
to ensure that we avoid the most common misconceptions which otherwise might arise when one attempts to apply superfluid models to astrophysics and cosmology, some details can be found in Appendix \ref{s:sf}.
In fact, some superfluid-like models of dark matter based on classical perfect fluids, scalar field theories or scalar-tensor gravities,
turned out to be vulnerable to experimental verification~\cite{lmo19}.
Moreover, superfluids are often confused not only with perfect fluids,
but also
with
the concomitant phenomenon of
Bose-Einstein condensates (BEC),
which is
another kind of quantum matter occurring in low-temperature condensed matter
~\cite{psbook04}.
However,
even though BEC's do share certain features with superfluids,
this does not imply that they are
superfluidic in~general.
In particular, quantum excitations in laboratory superfluids that we know of
have dispersion relations of a distinctive shape
called the Landau ``roton'' spectrum.
Such a shape of the spectral curve is crucial, as~it ensures
the suppression of dissipative fluctuations at { a quantum level
~\cite{z12eb,z20ijmpa},
which~results in inviscid flow~\cite{kap38,am38}.
If plotted as an excitation energy versus momentum,
the curve starts from the origin, climbs up to a local maximum (called the {maxon} peak),
then slightly descends to a local nontrivial minimum (called the ``roton'' energy gap); then grows again, this time all the way up,
to the boundary of the theory's applicability range.
In fact it is not the roton energy gap alone, but~the energy barrier formed by the maxon peak and roton minimum in momentum space,
which~ensures the above-mentioned suppression of quantum fluctuations in quantum liquid
and, ultimately, causes its flow to become inviscid.
In other words, it is the global characteristics of the dispersion curve, not just the existence of a nontrivial
local minimum and related energy gap,
which is important for superfluidity to occur.
Obviously, these are non-trivial properties,
which cannot possibly occur in all quantum liquids and condensates.
Further details and aspects are discussed in Appendix \ref{s:sf}.
This Letter~ is organized as follows.
Theory of physical vacuum based on the logarithmic superfluid model
is outlined in \Sec~\ref{s:mod},
where we also demonstrate how four-dimensional spacetime can emerge from the
three-dimensional dynamics of quantum liquid.
In \Sec~\ref{s:gr}, we derive the gravitational potential,
induced by the logarithmic superfluid vacuum in a given state,
using certain simplifying assumptions.
Thereafter, in~\Sec~\ref{s:phy},
we give a brief physical interpretation of different parts of the derived
gravitational potential and estimate their characteristic length scales.
In \Sec~\ref{s:den},
profiles of induced matter density
are derived and discussed for the case of spherical symmetry.
Galactic scale phenomena are discussed
in \Sec~\ref{s:frc},
where
the phenomenon of galactic rotation curves is explained
without introducing any exotic matter ad~hoc.
In \Sec~\ref{s:uexp},
we discuss the various mechanisms of the accelerating expansion of the Universe,
as well as the
cosmological singularity,
``vacuum catastrophe''
and
cosmological coincidence
problems.
Conclusions are drawn in \Sec~\ref{s:con}.\\
\scn{Logarithmic Superfluid Vacuum}{s:mod}
Superfluid vacuum theory assumes that
the physical vacuum is described,
when disregarding
quantum fluctuations, by~
the
fluid
condensate wavefunction $\Psi (\ve{r},\, t)$,
which is a three-dimensional Euclidean scalar.
The state itself is described by a ray in the corresponding Hilbert space,
therefore
this wavefunction obeys a normalization condition
\begin{equation}\lb{e:norm}
\langle \Psi | \Psi \rangle
= \int_V \rho} \def\dnc{\bar{\dn} \, \drm\vol
= {M}
,\end{equation}
where
${M}$ and $V$
are the
total mass and volume of the fluid, respectively,
and
$\rho} \def\dnc{\bar{\dn} = |\Psi|^2$ is the fluid mass density.
The wavefunction's dynamics is governed by an equation of
a $U(1)$-symmetric Schr\"odinger~form:
\begin{equation}\lb{e:becgeneq}
\left[
- i \hbar \, \partial_t
- \frac{\hbar^2}{2 m} \boldsymbol{\nabla}^2
+
V_{\text{ext}} (\ve{r},\, t)
+
F(|\Psi|^2)
\right]
\Psi
= 0,
\end{equation}
where
$m$
is the constituent particles'
mass,
$V_\text{ext} (\ve{r},t)$ is an external or trapping potential
and $F (\rho} \def\dnc{\bar{\dn})$ is a duly chosen function, which effectively
takes into account many-body effects inside the fluid.
This~wave equation can be formally derived as a minimizing condition of an action functional
with the following
Lagrangian:
} \def\bw{\begin{widetext}
\begin{eqnarray}
{\cal L}
&=&
\frac{i \hbar}{2}(\Psi \partial_t\Psi^* - \Psi^*\partial_t\Psi)+
\frac{\hbar^2}{2 m}
|\boldsymbol{\nabla} \Psi|^2
+ V_\text{ext} (\ve{r},\, t)\,
|\Psi|^2
+
{V} (|\Psi|^2)
,
\lb{e:ftlan}
\end{eqnarray}
} \def\ew{\end{widetext}
where ${V} (\rho} \def\dnc{\bar{\dn})$ equals to a primitive of $F(\rho} \def\dnc{\bar{\dn})$ up to an additive constant:
$F (\rho} \def\dnc{\bar{\dn}) = {V}' (\rho} \def\dnc{\bar{\dn}) $;
throughout the paper the prime denotes a derivative with respect to the function's~argument.
In this picture,
massless excitations, such as photons, are analogous to
acoustic waves propagating with velocity
$
c_s \propto \sqrt{|p' (\rho} \def\dnc{\bar{\dn}) |}
$,
where fluid pressure $p =p (\rho} \def\dnc{\bar{\dn})$ is determined via the equation of state.
For~the system \eqref{e:becgeneq},
both the equation of state and speed of sound can be derived using
the fluid-Schr\"odinger analogy, which was established for a special case
in Ref.} \def\Refs{Refs.} \def\eq{Eq. } \def\eqs{Eqs. } %\def\Sec{Section } \def\Secs{Sections ~\cite{dmf03},
and generalized for an arbitrary $F (\rho)$ in works~\cite{z11appb,z19mat}.
In a leading-order approximation
with respect
to the Planck constant,
we obtain
\begin{equation}\lb{e:cappapp}
p
=
- \frac{1}{m}
\int\!\rho F'(\rho)\, d\rho
,
\ \
c_s^2
=
\frac{1}{m}
\rho} \def\dnc{\bar{\dn}
|F' (\rho} \def\dnc{\bar{\dn})|
,
\end{equation}
while higher-order corrections
would
induce Korteweg-type effects, thus significantly
complicating the subject matter~\cite{z19mat}.
Furthermore,
it is natural to require that
superfluid vacuum theory must
recover Einstein's theory of relativity at a certain limit.
One can show that at a limit of low momenta of quantum excitations,
often~called a ``phononic'' limit by analogy with laboratory quantum liquids,
Lorentz symmetry does emerge.
This can be easily shown by virtue of the fluid/gravity analogy~\cite{blv05},
which~was subsequently used to formulate the BEC-spacetime correspondence~\cite{z11appb};
it can also be demonstrated by using dispersion relations~\cite{z20ijmpa,z11pla},
which are generally become deformed in theories with non-exact Lorentz symmetry~\cite{sch17,csa18,ol19,plj20}.
This correspondence states that Lorentz symmetry is approximate,
while
four-dimensional spacetime is an induced phenomenon,
determined by the dynamics of quantum Bose liquid
moving in Euclidean three-dimensional space.
The latter is only observable by a certain kind of observer, a~F(ull)-observer.
Other observers,
R(elativistic)-observers, perceive this superfluid
as a non-removable background, which can be modeled as a four-dimensional pseudo-Riemannian manifold.
What is the difference between these types of observers?
F-observers can perform measurements using objects of
arbitrary momenta and ``see'' the fundamental
superfluid wavefunction's evolution in three-dimensional
Euclidean space according to \eq \eqref{e:becgeneq} or an analogue thereof.
On the other hand,
R-observers are restricted
to measuring only small-momentum small-amplitude excitations of the background superfluid.
This is somewhat analogous to listening to acoustic waves (phonons) in the conventional Bose-Einstein condensates,
but~being unaware of higher-energy particles such as photons or~neutrons.
According to BEC-spacetime correspondence,
a R-observer
``sees'' himself located inside four-dimensional curved spacetime with a pseudo-Riemannian metric.
The latter can be written in Cartesian coordinates as~\cite{z11appb}:
\begin{equation}\lb{e:metr1}
g_{\mu\nu}
\propto
\frac{\rho} \def\dnc{\bar{\dn}}{c_s}
\left(
\baa{ccc}
-\left[ c_s^2 - m^{-1}^2 (\boldsymbol{\nabla} S)^2 \right] & \vdots & - m^{-1} \boldsymbol{\nabla}~S \\
\cdots\cdots & \cdot & \cdots \\
- m^{-1} \boldsymbol{\nabla} S & \vdots & \textbf{I}
\end{array}
\right)
,
\end{equation}
where $m^{-1} = \hbar/m$,
$S = S (\ve{r},\, t) = - i \ln{\left(\Psi (\ve{r},\, t) /|\Psi (\ve{r},\, t)|\right)}$ is a phase of the condensate wavefunction written
in the Madelung representation,
$\Psi = \sqrt\rho} \def\dnc{\bar{\dn} \, \exp{(i S)}$,
and
$\textbf{I}$ is a three-dimensional unit matrix.
To~maintain the correct metric signature in \eq \eqref{e:metr1},
condition
$|c_s | > m^{-1} \, |\boldsymbol{\nabla} S|$ must be imposed, which indicates
that $c_s$ is the maximum attainable velocity of test particles
(i.e.,~small-amplitude excitations of the condensate),
moving along geodesics on this induced spacetime.
Therefore, $c_s$ is the velocity of those excitations of vacuum,
which describe massless particles in the low-momentum limit,
whereas massive test particles move along geodesics of a pseudo-Riemannian manifold with metric \eq \eqref{e:metr1}.
According to a R-observer, they are freely falling, independently of their properties including their rest~mass.
In this approach,
we
interpret Einstein field equations not as differential
equations for an unknown metric; but as a definition for an induced stress-energy tensor,
describing
some effective matter to which test particles couple.
Therefore,
this would be the gravitating matter observed by a R-observer.
We thus obtain
\begin{equation}\lb{e:setdef}
\widetilde{T}_{\mu\nu}
\equiv
\kappa^{-1}
\left[
R_{\mu\nu} (g) - \frac{1}{2} g_{\mu\nu} R (g)
\right]
,
\end{equation}
where $\kappa = {8 \pi G}/{c_{(0)}^2}$ is the Einstein's gravitational constant.
An example of usage of this procedure will be considered in \Sec~\ref{s:cfs}.
While \eq \eqref{e:setdef} is in fact an assumption,
it should hold not only under the validity of conventional general relativity, but~also in other Lorentz-symmetric theories of gravity which are
linear with respect to the Riemann tensor,
because the form of Einstein equations is quite universal (up to a conformal transform).
For other Lorentz-symmetric theories,
whose field equations cannot be transformed into this form, definition \eqref{e:setdef} can be adjusted accordingly.
Furthermore,
one can see from \eq \eqref{e:cappapp}, that $c_s$ contains
an unknown function $F (\rho} \def\dnc{\bar{\dn})$.
To~determine its form, let us recall
that one of the relativistic postulates
implies that velocity $c_s$
should not depend on density, at~least
in the classical limit.
More specifically, at~low momenta, this~velocity
should tend to the value $ c_{(0)} \approx c $,
where
$c = 2.9979 \times 10^{10}$ $\text{cm}\, \text{s}^{-1}$
is
called the {speed of light in vacuum},
for~historical reasons.
Recalling \eq \eqref{e:cappapp},
this requirement can be written
as a differential equation \citen{z11appb}:
\begin{equation}\lb{e:Feq}
\rho} \def\dnc{\bar{\dn}
|F' (\rho} \def\dnc{\bar{\dn})|
=
m c_s^2
\approx
\text{const} (\rho} \def\dnc{\bar{\dn})
, \end{equation}
where
$\text{const} (\rho} \def\dnc{\bar{\dn})$ denotes a function which does not depend on density.
The solution of
this differential equation
is a logarithmic function:
\begin{equation}\lb{e:oFln}
F (\rho} \def\dnc{\bar{\dn}) = - b \ln{(\rho} \def\dnc{\bar{\dn}/\dnc)}
, \end{equation}
where $b$ and $\dnc$ are generally real-valued functions of coordinates.
The wave Equation \eqref{e:becgeneq}
thus narrows down to
\begin{equation}\label{e:oF}
i \hbar \partial_t \Psi
=
\left[-\frac{\hbar^2}{2 m} \boldsymbol{\nabla}^2
+
V_\text{ext} (\ve{r},\, t)
-
b
\ln{(|\Psi|^{2}/\dnc)}
\right]\Psi
,
\end{equation}
where $b$ is the nonlinear coupling; $b = b (\ve{r},\, t)$ in general.
Correspondingly,
{Equation
~\eqref{e:cappapp} yields
\begin{equation}\lb{e:cappappln}
p
=
- (b/m) \rho} \def\dnc{\bar{\dn}
,
\ \
c_s
=
\sqrt{|b|/m}
,
\end{equation}
thus indicating that logarithmic Bose liquid behaves like barotropic perfect fluid;
but only when one
neglects quantum corrections, and~assumes classical averaging.
This reaffirms the statement made in the previous section about the place of perfect-fluid
models when it comes to gravitational phenomena.
The way gravity emerges in the superfluid vacuum picture is entirely
different from those models, as~will be demonstrated shortly,
after we have specified our working~model.
Some special cases of \eq \eqref{e:oF}, for~example when $b \to b_0 =$ const,
were extensively studied
in the past, although~not for reasons related to quantum liquids~\cite{ros68,bb76}.
There were also extensive mathematical studies of these equations, to~mention just some
very recent literature~\cite{ss18,af18,aff19,wz19,bcs19,ct19,wtc19,lzh19,zh20,mm20,aj20,ss20,tb20je,tb20mj}.
Interestingly, wave equations with logarithmic nonlinearity can be also introduced
into fundamental physics
independently of relativistic arguments~\cite{z11appb,z10gc,szm16}.
This nonlinearity readily occurs in the theory of open quantum systems, quantum entropy
and information~\cite{yas78,bra91}; as well as in the theory of general condensate-like
materials, for~which characteristic kinetic energies are significantly smaller than interparticle
potentials~\cite{z18zna}.
One example of such a material is helium II, the~superfluid phase of helium-4.
For the latter, the~logarithmic superfluid model is known to have been well verified by experimental data~\cite{z12eb,sz19}.
Among other things, the~logarithmic superfluid model does reproduce the
sought-after Landau-type spectrum of excitations, discussed in the previous section;
detailed derivations can be found in~\cite{z12eb}.
One of underlying reasons for such phenomenological success is that the ground-state wavefunction of
free (trapless) logarithmic liquid
is not a de Broglie plane wave, but~a spatial Gaussian modulated by a de Broglie plane wave.
This explains the liquid's inhomogenization followed by
the formation of fluid elements or parcels;
which indicates that
such models do describe fluids, rather than gaseous matter~\cite{az11,bo15,z17zna,z18epl,kdk19,z19ijmpb}.
To summarize, a~large number of arguments to date, both theoretical and experimental,
demonstrate the robustness of logarithmic models in the general theory of superfluidity.
In the next \Sec~we shall demonstrate the logarithmic superfluid model's capabilities
when assuming
superfluidity of the physical vacuum~itself.
In what follows,
we shall make use of a minimal inhomogeneous model for the logarithmic superfluid
which was proposed in Ref.} \def\Refs{Refs.} \def\eq{Eq. } \def\eqs{Eqs. } %\def\Sec{Section } \def\Secs{Sections ~\cite{z18zna},
based on statistical and thermodynamics arguments.
In the F-observer's picture,
its wave equation can be written as
\begin{equation}
i\hbar \partial_t \Psi
=
\left[
-\frac{\hbar^{2}}{2 m} \vena^2}%_D}\def\lapl{\Delta}%_D
+V_\text{ext} (\ve{r}, t)
-
\left(
b_0 - \frac{q}{r^2}
\right)\!
\ln{\!\left(\frac{|\Psi|^{2}}{\dnc} \right)}
\right]\!\Psi
,\label{e:vcm}
\end{equation}
where $r = |\ve{r}| = \sqrt{\ve{r} \cdot \ve{r}}$
is a radius-vector's absolute value,
and $b_0$ and $q$ are real-valued constants.
For definiteness, let us assume that $b_0 > 0$, because~one can always change the overall signs of the nonlinear term
$F (\rho} \def\dnc{\bar{\dn})$ and the corresponding field-theoretical potential ${V} (\rho} \def\dnc{\bar{\dn})$.
As always, this wave equation must be supplemented with a normalization condition
\eqref{e:norm}, boundary and initial conditions of a quantum-mechanical type;
which ensure the fluid interpretation of $\Psi$ \cite{ry99}.
One can show that
nonlinear coupling $b = b (\ve{r}) = b_0 - q/r^2$
is a linear function of the quantum temperature $T_\Psi$,
which is defined as
a thermodynamic conjugate
of quantum information entropy sometimes dubbed as the Everett-Hirschman information entropy.
The latter can be written as
$
S_\Psi =
-
\langle \Psi | \ln{(|\Psi|^2/\dnc)}|\Psi \rangle
=
-\int_V |\Psi|^2 \ln{(|\Psi|^2/\dnc)} \, \drm\vol
$,
where a factor $1/\dnc$ is introduced for the sake of correct dimensionality,
and can be absorbed into an additive constant due to the normalization condition \eqref{e:norm}.
Therefore, one can expect that the thermodynamical parameters
\begin{equation}\lb{e:bqtem}
b_0 = b_0 (T_\Psi), \ q = q (T_\Psi)
\end{equation}
are constant at a fixed temperature $T_\Psi$.
Thus, for~a trapless version of the model \eqref{e:vcm}
we have four parameters, but~only two of them, $m$ and $\dnc$,
are \textit{a priori} fixed, whereas the other two, $b_0$ and $q$,
can vary depending on
the~environment.
\scn{Induced Gravitational Potential}{s:gr}
Invoking model \eqref{e:vcm}, while neglecting quantum fluctuations, let us assume that physical vacuum is a collective
quantum state described
by wavefunction $\Psi = \Psi_{\text{vac}} (\ve{r},\, t)$,
which forms a self-gravitating configuration with a center at $\ve{r} = 0$.
Therefore, for~this state, the~solution of \eq \eqref{e:vcm} is equivalent to the solution
of the linear Schr\"odinger equation,
\begin{equation}
i\hbar \partial_t \Psi
=
\left[
-\frac{\hbar^{2}}{2 m} \vena^2}%_D}\def\lapl{\Delta}%_D
+V_\text{eff} (\ve{r},\, t)
\right]
\Psi
,\label{e:vcmef}
\end{equation}
for a particle of mass $m$
driven by an effective potential
\begin{eqnarray}
V_\text{eff}
(\ve{r}, t)
&=&
V_\text{ext} (\ve{r}, t)
-
b
\ln{\!\left(\frac{|\Psi_{\text{vac}} (\ve{r}, t)
|^{2}}{\dnc} \right)}
\nonumber \\
&=&
V_\text{ext} (\ve{r}, t)
-
\left(
b_0 - \frac{q}{r^2}
\right)\!
\ln{\!\left(\frac{|\Psi_{\text{vac}} (\ve{r}, t)
|^{2}}{\dnc} \right)}\!,~~~~~
\lb{e:vind}
\end{eqnarray}
when written in Cartesian coordinates~\cite{z18zna}.
If working in curvilinear coordinates, the~last formula must be supplemented with terms
which arise after separating out the angular variables in the wave~equation.
In the absence of quantum excitations and other interactions, it is natural to associate
this effective quantum-mechanical potential
with the only non-removable fundamental interaction that we know of: gravity.
This interpretation will be further justified in \Sec~\ref{s:phy}.
Therefore, in~Cartesian coordinates one can write the induced gravitational potential as
\begin{equation}\lb{e:igrav}
\Phi (\ve{r}, t)
=
- \frac{1}{m}
V_\text{eff} (\ve{r}, t)
=
\frac{1}{m}
\left(
b_0 - \frac{q}{r^2}
\right)
\ln{\!\left(\frac{|\Psi_{\text{vac}} (\ve{r}, t)|^{2}}{\dnc} \right)}
,
\end{equation}
where we assume that the background superfluid is trapless, i.e.,~we set $V_\text{ext} = 0$.
It should also be remembered that in curvilinear coordinates,
this formula must be modified according to the remark after \eq \eqref{e:vind}; but for now we
shall disregard any anisotropy
and~rotation.
It should be noticed that if one regards this potential as a multiplication operator
then its quantum-mechanical average would be related to
the~Everett-Hirschman information entropy discussed
in the previous section:
$
\langle
\Phi
\rangle
\sim T_\Psi\left\langle \Psi\right|\, \ln{(|\Psi|^{2})} \left|\Psi \right\rangle
\sim T_\Psi
S_\Psi
$.
This not only makes theories of entropic gravity
(which are essentially based on the ideas of Bekenstein, Hawking, Jacobson and others) a subset of the logarithmic superfluid vacuum
approach, but~also endows them with an underlying physical meaning and origin of the entropy~implied.
We can see that the induced potential maintains its form as long as the physical vacuum stays in the state
$|\Psi_{\text{vac}} \rangle$.
If the vacuum were to transition into a different state, then it would change its wavefunction;
hence the induced gravitational potential
would also change.
We expect that our vacuum is currently in a stable state,
which is close to a ground state
or at least to a metastable state, with~a sufficiently large lifetime.
It is thus natural to assume that the state
$|\Psi_{\text{vac}} \rangle$
is stationary and rotationally~invariant.
As we established earlier, the~wavefunction describing such a state
should be the solution
of a quantum wave equation containing logarithmic nonlinearity.
In the case of trivial spatial topology and infinite extent,
the amplitude
of such a solution is known to be the product of a Gaussian function,
which was mentioned in the previous section,
and a conventional quantum-mechanical part, which~is a product of an exponential function,
power function and a polynomial.
Thus we can write the amplitude's general form as:
\begin{equation}\lb{e:avex}
|\Psi_{\text{vac}} | =
\sqrt{\dnc}\,
\left(
\frac{r}{\bar{\ell}}
\right)^{\chi_0/2}
P (r)
\,
\exp{\!
\left(
-
\frac{a_2 }{2 \bar{\ell}^2} r^2
+
\frac{a_1}{2 \bar{\ell}} r
+
\frac{a_0}{2}
\right)
}
,
\end{equation}
where
$P (r)$ is a polynomial function,
$\chi_0$
and $a$'s are constants,
and $\bar{\ell} = (m/\dnc)^{1/3}$
is a classical characteristic length scale of the logarithmic nonlinearity
(alternatively, one can choose $\bar{\ell}$ being equal to the quantum characteristic length,
$\hbar/\sqrt{m b_0}$, which might be more useful for $\hbar$-expansion techniques).
If quantum liquid occupies an infinite spatial domain then the
normalization condition \eqref{e:norm}
requires
\begin{equation}
a_2 > 0
,
\end{equation}
which is also confirmed by analytical and numerical studies of differential equations with
logarithmic nonlinearity of
various types~\cite{bb76,ss18,af18,bcs19,lzh19,z18zna,ss20}.
Both the form of a function $P (r)$ and the values of $\chi_0$ and $a$'s
must be determined by a solution of an eigenvalue problem
for the
wave equation
under normalization
and boundary conditions.
At~this stage, those conditions are not yet precisely known;
even if they were, we do not yet know which quantum state our vacuum is currently in.
Therefore, these constants' values remain theoretically unknown at this stage,
yet can be determined empirically.
Furthermore, for~the sake of simplicity,
let us approximate
the power-polynomial term $(r/\bar{\ell})^{\chi_0/2} P (r)$,
by the single power function $(r/\bar{\ell})^{\chi/2}$, where the constant $\chi$
is
the best fitting parameter.
Therefore, we can approximately rewrite \eq \eqref{e:avex} as
\begin{equation}\lb{e:ava}
|\Psi_{\text{vac}} |^2 \approx
\dnc\,
\exp{\!
\left[
-
\frac{a_2 }{\bar{\ell}^2} r^2
+
\frac{a_1 }{\bar{\ell}} r
+ \chi
\ln{\!\left(
\frac{r}{\bar{\ell}}
\right)}
+ a_0
\right]
}
,
\end{equation}
which is more convenient for further analytical studies than the original expression \eqref{e:avex}.
From the empirical point of view, the~function \eqref{e:ava} can be considered as a
trial function, whose parameters can be fixed using experimental data
following the procedure we describe~below.
For the trial solution \eqref{e:ava},
the
normalization condition \eqref{e:norm} immediately imposes a constraint for one
of its parameters:
\begin{equation}
\text{exp}{(a_0)}
\approx
\frac{{M}
a_2^{(\chi +3)/2}
}{2 \pi m}
\left[
\Auxfun{\frac{3}{2}}{\frac{1}{2}}
+
\frac{a_1}{\sqrt{a_2}}
\Auxfun{2}{\frac{3}{2}}
\right]^{-1}
,
\end{equation}
where we introduced an auxiliary function
$
\Auxfun{a}{b}=
\Gamma\left(a + \chi/2 \right)
\Kummer{a+ \chi/2}{b}{a_1^2/4 a_2}
$,
where
$\Gamma (a)$ and $\Kummer{a}{b}{z}$ are the gamma function and Kummer confluent hypergeometric function,
respectively.
If values of $a$'s and $\chi$ are determined, e.g.,~empirically, then this formula
can be used to estimate the ratio ${M}/m$.
Furthermore,
by
substituting the trial solution \eqref{e:ava} into the definition \eqref{e:igrav},
we derive the induced gravitational potential as a sum of seven terms:
\begin{eqnarray}
\grpot} %\def\grpott{\grpot_{(\Psi)} (r)
&=&
\grpot_{\text{smi}}
(r)
+
\grpot_{\text{RN}}
(r)
+
\grpot_{\text{N}}
(r)
\nonumber \\&&
+
\grpot_{\text{gal}}
(r)
+
\grpot_{\text{mgl}}
(r)
+
\grpot_{\text{dS}}
(r)
+
\grpot} %\def\grpott{\grpot_{(\Psi)}_0
,\lb{e:grev}
\end{eqnarray}
where
\begin{eqnarray}
\grpot_{\text{smi}} (r)
&=&
-\frac{\chi \, q}{m}
\frac{\ln{(r/\bar{\ell})}}{r^2}
=
- \zeta_{\chi q} c_b^2
\frac{L_{\text{smi}}^2 \ln{(r/\bar{\ell})}}{r^2},
~~~\lb{e:grsmi}\\
\grpot_{\text{RN}} (r)
&=&
-\frac{a_0 \, q}{m}
\frac{1}{r^2}
=
- \zeta_{a_0 q} c_b^2
\frac{L_{\text{RN}}^2}{r^2}
\, , \lb{e:grmic}\\
\grpot_{\text{N}} (r)
&=&
-\frac{a_1 q}{m \bar{\ell}} \frac{1}{r} =
- \frac{G M}{r}
, \lb{e:grnew}\\
\grpot_{\text{gal}} (r)
&=&
\frac{\chi \, b_0}{m}
\ln{(r/\bar{\ell})}
=
c_b^2\,
\chi \ln{(r/\bar{\ell})}
, \lb{e:grgal}\\
\grpot_{\text{mgl}} (r)
&=&
\frac{a_1 b_0}{m \bar{\ell}}
r
=
\zeta_{a_1} c_b^2
\frac{r}{L_{\text{mgl}}}
, \lb{e:grmgl}\\
\grpot_{\text{dS}} (r)
&=&
-\frac{a_2 b_0}{m \bar{\ell}^2}
r^2
=
- c_b^2
\frac{r^2}{L_{\text{dS}}^2}
, \lb{e:grcos}
\end{eqnarray}
and
\begin{equation}
\grpot} %\def\grpott{\grpot_{(\Psi)}_0
=
\frac{1}{m}
\left(
a_0 b_0
+ \frac{a_2 q}{\bar{\ell}^2}
\right)
=
\frac{1}{m}
\left(
a_0 b_0
+ \frac{q}{L_{\text{dS}}^2}
\right)
\, \lb{e:gr0}
\end{equation}
is the additive constant.
Here, and~throughout the paper, we denote the sign functions
by $\zeta$'s:
$\zeta_\alpha = \Sign{\alpha}$,
and use the following notations:
\begin{eqnarray}
&&
c_b = \sqrt{\frac{b_0}{m}}, \
G M
= \frac{a_1 q}{m \bar{\ell}},
\
L_{\text{smi}} = \sqrt{\frac{|\chi\, q|}{b_0}}
,
\nonumber \\&&
L_{\text{RN}} = \sqrt{\frac{|a_0 q|}{b_0}}
,\
L_{\text{mgl}} = \frac{\bar{\ell}}{|a_1|}
= \frac{|q|}{m G M}
,
\lb{e:lengr}\\&&
L_{\text{dS}} = \frac{\bar{\ell}}{\sqrt{a_2}}
, \
L_\chi = \frac{\chi \bar{\ell}}{a_1}
= \frac{\chi q}{m G M}
,
\nonumber
\end{eqnarray}
where
$G$ is the Newton's gravitational constant as per~usual.
Furthermore,
Lorentz symmetry emerges
in the ``phononic'' low-momentum limit of the theory, as discussed in the previous section.
Therefore,
a R-observer
would perceive the gravity induced by potential \eqref{e:grev}
as curved four-dimensional spacetime,
which is a local perturbation (not necessarily small)
of the background flow metric,
such as the one derived in \Sec~5.3 of Ref.} \def\Refs{Refs.} \def\eq{Eq. } \def\eqs{Eqs. } %\def\Sec{Section } \def\Secs{Sections ~\cite{z11appb},
see \Sec~\ref{s:cfs} below.
In a rotationally invariant case,
the line element of this spacetime
can be written in the Newtonian gauge;
if $\grpot} %\def\grpott{\grpot_{(\Psi)} (r)/c_{(0)}^2 \ll 1$,
then it can be approximately
rewritten in the form
\begin{equation}\lb{e:metrst}
d s^2
\approx
- c_{(0)}^2\!\left[
1 +
\frac{2 \grpot} %\def\grpott{\grpot_{(\Psi)} (r)}{c_{(0)}^2}
\right] d t^2
+ \frac{d r^2}{1 + 2 \grpot} %\def\grpott{\grpot_{(\Psi)} (r)/c_{(0)}^2} +
R^2 (r) d \sigma^2} %\Omega_{(2)}^2
,
\end{equation}
where
$R (r) = r \left[1 + {\cal O} \left(\grpot} %\def\grpott{\grpot_{(\Psi)} (r)/c_{(0)}^2\right)\right] \approx r$,
$d\, \sigma^2} %\Omega_{(2)}^2 = d\theta^2 + \Sin{2}{\theta} \, d \varphi^2$
is the line element of a unit two-sphere,
and a leading-order approximation
with respect
to the Planck constant is implied,
as~usual.
The mapping \eqref{e:metrst} is valid for regions where
the induced metric maintains a signature `$- + + +$',
and its matrix is non-singular.
In other regions, such as close vicinities of spacetime singularities
or horizons,
the relativistic approximation is likely to fall
outside its applicability range,
thus it should be replaced with
the F-observer's description of~reality.
The main
simplifying assumptions and approximations underlying
the derivation of our gravitational potential
are summarized and enumerated in the Appendix \ref{s:aa}.
\scn{Physical Interpretation}{s:phy}
It should be noticed that if we did not have a logarithm
in the original model \eqref{e:vcm}, then~in \mbox{\eqs \eqref{e:vind}
and \eqref{e:igrav}},
then
we would not have arrived at the polynomial functions
in \mbox{\eqs~\eqref{e:grev}--\eqref{e:grcos}},
which are easily recognizable.
This reaffirms our expectations that the underlying model
can be successfully confirmed by experiment;
but first those functions must be endowed with precise physical~meaning.
In this \Sec,
we shall assign a physical interpretation to each term of the derived gravitational potential.
For the sake of brevity,
we shall be omitting an additive constant $\grpot} %\def\grpott{\grpot_{(\Psi)}_0$,
assuming that it is small compared to $c_{(0)}^2$.
\sscn{Potential $\grpot_{\text{N}}$ and gravitational mass generation}{s:pnew}
We begin with term \eqref{e:grnew}, which has the most obvious meaning.
In a non-relativistic picture, it~represents Newton's model of gravity.
According
to the BEC-spacetime correspondence manifested through
the mapping \eqref{e:metrst},
a R-observer can observe an effect of the potential $\grpot_{\text{N}}$ by
measuring probe particles moving along geodesics in the Schwarzschild spacetime:
\begin{equation}\lb{e:meschw}
d s^2_{(\text{N})}
\approx
- c_{(0)}^2
\left(
1 - \frac{r_H}{r}
\right) d t^2
+ \frac{d r^2}{1 - r_H / r} +
r^2 d \sigma^2} %\Omega_{(2)}^2
,
\end{equation}
where
$r_H = 2 G M/c_{(0)}^2$ is the Schwarzschild radius.
Therefore,
in absence of asymptotically non-vanishing terms, $M$ can be interpreted as the gravitational mass of
the~configuration.
This mass can be expressed in terms of superfluid parameters as
\begin{equation}\lb{e:schrad}
r_H =
\frac{2 a_1 q}{m c_{(0)}^2 \bar{\ell}}
,
\ \
\Sign{a_1 q} =
\left\{
\baa{rll}
1 & &\text{gravity},\\
-1 & &\text{anti-gravity},
\end{array}
\right.
\end{equation}
thus assigning physical meaning to a combination of parameters $a_1 q/m \bar{\ell}$.
In particular,
one can see
that
a sign of the product $a_1 q$
determines
whether the $\grpot_{\text{N}}$ interaction is attractive (gravity) or repulsive (anti-gravity).
For most systems that we know of, anti-gravitational effects have not yet been observed,
therefore one can assume that $M > 0$ or
\begin{equation}
a_1 q > 0
\end{equation}
from now~on.
Nevertheless,
it should be emphasized that the anti-gravity case is not
\textit{a priori}
forbidden in superfluid vacuum theory.
Indeed, the~spacetime singularity occurs at
$r = 0$ in a relativistic picture only, which poses certain issues for a R-observer,
especially in the case of anti-gravity when a singularity is not covered by an event horizon
(the existence of naked singularities is often doubted, on~grounds of the cosmic censorship hypothesis).
However, a~F-observer would see no singular behavior in either case, because~the wavefunction
$\Psi_{\text{vac}}$
remains
regular and normalizable at each point of space and at any given time -- as it should be in a quantum-mechanical theory.
This reaffirms the fact that spacetime singularities are an artifact of incomplete information accessible
to observers operating with
relativistic particles~\cite{z11appb}.
Thus, the~mapping from \eqs \eqref{e:grnew}--\eqref{e:meschw}
can be used to reformulate black hole phenomena in the language of continuum mechanics and
the theory of superfluidity;
which can resolve certain long-standing problems occurring in the relativistic theory of gravity.
For instance,
neglecting asymptotically non-flat terms for simplicity,
one can view \eqs \eqref{e:grnew}, \eqref{e:meschw} and \eqref{e:schrad}
as the gravitational mass generation mechanism:
such mass is not a fundamental notion,
but a composite quantum phenomenon induced by the background superfluid's dynamics
(through the elementary inertial mass $m$ and
critical density $\dnc$), its quantum temperature (through $q$),
and an
exponential part of the condensate's wavefunction (through $a_1$).
Such a mechanism can be thus considered as the quantum-mechanical version of the Mach principle~\cite{z11appb}.
For example,
if either $a_1$ or $q$ vanish, then the system would not possess any gravitational mass,
but it still would be gravitating in a non-Newtonian way, if~other potentials from \eq \eqref{e:grev} are non-zero.
\sscn{Potential $\grpot_{\text{RN}}$ and abelian charges}{s:pmic}
Equation \eqref{e:grmic} represents another potential which can be easily recognized.
According
to the mapping \eqref{e:metrst},
potential $\grpot_{\text{RN}}$ is observed by a R-observer as Reissner-Nordstr\"om spacetime,
when~taken together with the $\grpot_{\text{N}}$ potential:
\begin{eqnarray}
d s^2_{(\text{N}+\text{RN})}
&\approx&
- c_{(0)}^2
\left(
1 - \frac{r_H}{r} + \frac{r_Q^2}{r^2}
\right) d t^2
\nonumber \\&&
+ \frac{d r^2}{1 - r_H / r + r_Q^2/r^2}
+
r^2 d \sigma^2} %\Omega_{(2)}^2
,\lb{e:memic}
\end{eqnarray}
where $r_Q$
is a characteristic length scale:
\begin{equation}
r_Q
=
\sqrt 2 L_{\text{RN}}
\frac{ c_b }{c_{(0)}}
=
\sqrt{\frac{2 |a_0 q|}{m c_{(0)}^2}}
,
\end{equation}
provided \
$a_0 q < 0$.
The Reissner-Nordstr\"om metric is known to be associated with the gravitational field caused
by a charge related to an abelian group.
An example would be an electric charge $Q$,
which is related to
the abelian group $U(1)$ of electromagnetism.
This charge can be thus revealed through the formula
\begin{equation}
Q^2 = \frac{r_Q^2 c_{(0)}^4}{k_e G}
\approx
\frac{2 | a_0 q| c^2}{k_e G m}
,
\end{equation}
where $k_e$ is the Coulomb constant.
In other words,
potential $\grpot_{\text{RN}}$ describes, together with $\grpot_{\text{N}}$,
the~gravitational field created by an object of a charge $Q$ and gravitational mass $M$.
Thus, from~a F-observer's viewpoint,
an electrical charge is not
an elementary notion,
but a composite quantum phenomenon, induced by the background superfluid's dynamics (through the elementary inertial mass $m$), its quantum temperature (through $q$),
and
overall constant coefficient of the condensate's wavefunction (through $a_0$).
It should be noted also that $\grpot_{\text{RN}}$ is a short-range potential,
therefore, it becomes substantial only at those microscopical length scales,
of an order
$r_{QH} = r_Q^2/r_H = \bar{\ell}\,|a_0/a_1|$ or below.
Since $r_Q < r_H$ for most objects we know of,
we have $0 \leqslant r_{QH} < r_H$.
Thus, those scales would be causally inaccessible to a R-observer, but~a F-observer would have no problem accessing them, per~usual.
\sscn{Potential $\grpot_{\text{smi}}$ and strong gravity}{s:psmi}
As the distance from a gravitating center decreases,
it is term \eqref{e:grsmi} which eventually predominates.
According
to the mapping \eqref{e:metrst},
this potential $\grpot_{\text{smi}}$,
when taken together with the $\grpot_{\text{N}}$ and $\grpot_{\text{RN}}$ potentials,
induces spacetime with the line element:
} \def\bw{\begin{widetext}
\begin{eqnarray}
d s^2_{(\text{N}+\text{RN}+\text{smi})}
\!&\approx&\!
- c_{(0)}^2\!\left[
1 - \frac{r_H}{r} + \frac{r_Q^2}{r^2}
-
\zeta_{\chi q}
\frac{r_W^2}{r^2}
\ln{\left(\frac{r}{\bar{\ell}}\right)}
\right]\! d t^2
+ \frac{d r^2}{1 - r_H / r + r_Q^2/r^2
-
\zeta_{\chi q}
r_W^2
\ln{\left(r/\bar{\ell}\right)}/r^2
}
+
r^2 d \sigma^2} %\Omega_{(2)}^2,~~~~
\lb{e:mesmi}
\end{eqnarray}
} \def\ew{\end{widetext}
where $r_W$
is a characteristic length scale:
\begin{equation}
r_W
=
\sqrt 2
L_{\text{smi}}
\frac{ c_b }{c_{(0)}}
=
\sqrt{
\frac{2 |\chi\, q|}{m c_{(0)}^2}
}
.
\end{equation}
The potential $\grpot_{\text{smi}}$ has a distinctive property:
unlike other sub-Newtonian potentials in \mbox{\eq~\eqref{e:grev}},
it can switch between repulsive and attractive regimes,
depending on whether the distance is larger or smaller than $\bar{\ell}$.
The
magnitudes of $\grpot_{\text{smi}}$ and $\grpot_{\text{RN}}$ become comparable
at two values of $r$,
shown by the formula
\begin{equation}
r_{WQ}^{(\pm)} =
\bar{\ell}
\exp{(\pm r_Q^2/r_W^2)}
=
\bar{\ell}
\exp{(\pm |a_0/\chi|)}
,
\end{equation}
which indicates
that $|\grpot_{\text{smi}} |$ overtakes $|\grpot_{\text{RN}} |$
either at $r < r_{WQ}^{(-)}$ or at $r > r_{WQ}^{(+)}$.
If $|a_0/\chi|$ is large then $r_{WQ}^{(+)}$ is exponentially large and
$r_{WQ}^{(-)}$ is exponentially~small.
Magnitudes of $\grpot_{\text{smi}}$ and $\grpot_{\text{N}}$ become comparable
at a certain value of $r$,
which is
shown by the~formula
\begin{eqnarray}
r_{WM}
&=&
-
\ell_W
\Prlog{
-
\bar{\ell}/\ell_W
}
\leqslant \ell_W,
\end{eqnarray}
where
$
\ell_W
\equiv
{r_W^2}/{|r_H|}
=
|L_\chi|
$,
and by $\Prlog a$ we denoted the Lambert function.
The function $|\grpot_{\text{smi}} (r)|$ becomes larger than $|\grpot_{\text{N}} (r)|$
at $r < r_{WM} \leqslant |L_\chi| $.
Furthermore,
considering $\grpot_{\text{smi}}$ as a perturbation of $\grpot_{\text{N}}$,
one can deduce that
the effective gravitational coupling,
\begin{equation}\lb{e:Geff}
G_\text{eff} \approx
G
\left[
1 +
\zeta_{\chi q}
L_\chi
\frac{\ln{(r/\bar{\ell})}}{r}
\right]
,
\end{equation}
becomes larger as $r$ gets smaller.
Here
an approximation symbol reminds us that we are working with
wavefunction \eqref{e:ava}, and~assume a leading-order approximation
with respect
to the Planck constant.
One can see that
gravity naturally becomes stronger at shorter scales,
without introducing any additional effects or matter,
which suggests the way towards resolving the hierarchy problem.
Indeed, the~latter states the large discrepancy between magnitudes of forces of Standard Model
interactions and classical gravity,
but
in this case,
the gravitational force
grows stronger than inverse square as $r$ decreases, as~$\ln r/r^3$.
Moreover,
it is possible to have even stronger short-length behaviour in our approach:
if one goes beyond a minimal model \eqref{e:vcm} and generalizes
nonlinear coupling to a series: $b (\ve{r}) = b_0 - q/r^2 + q_1/r^3 + ...$,
then this will induce terms $\ln r/r^4$, $\ln r/r^5$, \textit{et cetera}.
On the other hand, as~$r$ grows, the~function
$G_\text{eff} $
converges to $G$,
especially if $|L_\chi |$ or $1/\bar{\ell}$ are sufficiently small;
which makes $G_\text{eff}$ approximately constant for a large range of values of $r$.
\sscn{Potential $\grpot_{\text{dS}}$ and de Sitter spacetime}{s:pcos}
Let us turn our attention to terms
which do not tend to zero at $r \to \infty$.
The physical meaning of one of these
terms, given by \eq \eqref{e:grcos},
becomes clear upon
using the mapping \eqref{e:metrst}:
\begin{equation}\lb{e:meds}
d s^2_{(\text{dS})}
\approx
- c_{(0)}^2
\left(
1 - \frac{r^2}{R_{\text{dS}}^2}
\right) d t^2
+ \frac{d r^2}{1 - r^2 / R_{\text{dS}}^2} +
r^2 d \sigma^2} %\Omega_{(2)}^2
,
\end{equation}
where
\begin{equation}
R_{\text{dS}} =
\frac{1}{\sqrt 2}
\frac{c_{(0)}}{c_b}
L_{\text{dS}}
=
\bar{\ell}
\sqrt{\frac{m c_{(0)}^2}{2 a_2 b_0}}
\end{equation}
is a radius of de Sitter~horizon.
This metric represents de Sitter spacetime (written
in static coordinates),
which belongs to a class of Friedmann-Lema\^itre-Robertson-Walker (FLRW)
spacetimes.
Indeed, by~applying a coordinate transformation
$\tau - t = \tau_\text{dS}
\ln{\!\left(1 - r^2/R_{\text{dS}}^2
\right)}
$,
$
\varrho = \alpha_0^{-1} r
\exp{\!\left(
- t/2 \tau_\text{dS}
\right)}
\left(1 - r^2/R_{\text{dS}}^2
\right)^{-1/2}
$,
one can rewrite \eq \eqref{e:meds} in isotropic coordinates
\begin{equation}\lb{e:meds2}
d s^2_{(\text{dS})}
\to
- c_{(0)}^2 d \tau^2
+
\alpha_0^2
\exp{\!\left(
\frac{\tau}{\tau_\text{dS}}
\right)
}
\left(
d \varrho^2 + \varrho^2 d \sigma^2} %\Omega_{(2)}^2
\right)
,
\end{equation}
where
$\tau_\text{dS} = R_{\text{dS}}/ 2 c_{(0)}$
and $\alpha_0$ is an integration~constant.
The physical implications of the term $\grpot_{\text{dS}}$
will be further examined in \Sec~\ref{s:uexp}.
\sscn{Potentials $\grpot_{\text{mgl}}$ and $\grpot_{\text{gal}}$ and gravity on astronomical scales}{s:pmgal}
The remaining asymptotically non-vanishing potentials are given by \eqs \eqref{e:grgal} and \eqref{e:grmgl}.
With respect to the dependence upon radial distance from the gravitating center,
they occupy an intermediate
place between de Sitter term \eqref{e:grcos} and Newtonian potential \eqref{e:grnew}.
It is thus natural to expect that
these terms are responsible for large scale dynamics --
from galaxies (a kiloparsec scale) to metagalactic objects, such as voids and superclusters
(a megaparsec scale).
According to the mapping \eqref{e:metrst},
the terms $\grpot_{\text{gal}}$ and $\grpot_{\text{mgl}}$ modify de Sitter metric \eqref{e:meds}:
causing the resulting spacetime to be asymptotically de Sitter only.
The line element, which corresponds to the terms $\grpot_{\text{dS}}$,
$\grpot_{\text{mgl}}$ and $\grpot_{\text{gal}}$ taken together,
can be written in static coordinates as
} \def\bw{\begin{widetext}
\begin{eqnarray}
d s^2_{(\text{gal+mgl+dS})}
&\approx&
- c_{(0)}^2
\left[
1
+ \beta_\chi^2
\ln{\!\left(
\frac{r}{\bar{\ell}}
\right)}
+ \frac{r}{R_{\mathfrak M}} - \frac{r^2}{R_{\text{dS}}^2}
\right] d t^2
+ \frac{d r^2}{
1
+
\beta_\chi^2 \ln{\left(r/\bar{\ell}\right)}+ r/R_{\mathfrak M} - r^2 / R_{\text{dS}}^2
} +
r^2 d \sigma^2} %\Omega_{(2)}^2
,
\lb{e:medsa}
\end{eqnarray}
} \def\ew{\end{widetext}
where
$
\beta_\chi
=
\sqrt{
{2 \chi \, b_0}/{m c_{(0)}^2}
}
=
\sqrt{
2 \chi
}
{c_b}/{c_{(0)}}
$,
and
\begin{equation}
R_{\mathfrak M}
=
\frac{m c_{(0)}^2 \bar{\ell}}{2 a_1 b_0}
=
\frac{\zeta_{a_1}}{2}\frac{c_{(0)}^2}{c_b^2}
L_{\text{mgl}}
\end{equation}
is a characteristic length scale constant,
which can be positive or negative depending on the sign of $a_1$.
Note that in \eq \eqref{e:medsa},
we included the $\grpot_{\text{dS}}$-induced (de Sitter) term
to remind us that at large $r$ we still have a spacetime of a
FLRW~type.
Interestingly, linear terms in metrics
occur also in an alternative theory of gravity, Weyl gravity~\mbox{\cite{mk89,man93}}.
This coincidence can be explained by conformal symmetry, which often emerges in logarithmic models
at the relativistic limit,
see for example \Sec~\ref{s:cfs}.
However, Weyl gravity does not produce a logarithmic term in metric,
while in our theory it is induced by $\grpot_{\text{gal}}$.
This term is responsible for the flat rotation curves phenomenon in galaxies,
which will be discussed, along with other astronomical implications of the terms
$\grpot_{\text{mgl}}$ and $\grpot_{\text{gal}}$, in~further detail in \Sec~\ref{s:frc}.
It is also useful to know that the metric induced by the term \eqref{e:grmgl} alone,
\begin{equation}
d s^2_{(\text{mgl})}
\approx
- c_{(0)}^2
\left(
1
+ \frac{r}{R_{\mathfrak M}}
\right) d t^2
+ \frac{d r^2}{
1
+ r/R_{\mathfrak M}
} +
r^2 d \sigma^2} %\Omega_{(2)}^2
,
\lb{e:memgl}
\end{equation}
transforms to the conformally FLRW-type metric
} \def\bw{\begin{widetext}
\begin{eqnarray}
d s^2_{(\text{mgl})}
&\to&
\frac{
(1+\varrho/4 R_{\mathfrak M})^2
}{
a_{\mathfrak M}^2 (\tau)\, (1-\varrho/4 R_{\mathfrak M})^2
}
\left[
- c_{(0)}^2 d \tau^2
+ \frac{
a_{\mathfrak M}^2 (\tau)
}{(1-\varrho^2/16 R_{\mathfrak M}^2)^2
}
\left(
d \varrho^2 + \varrho^2 d \sigma^2} %\Omega_{(2)}^2
\right)
\right]
,
\lb{e:memgl2}
\end{eqnarray}
} \def\ew{\end{widetext}
upon applying the coordinate transformation
$r = \varrho/(1-\varrho/4 R_{\mathfrak M})^2$,
$t = \int d \tau/a_{\mathfrak M} (\tau)$.
For R-observers,
this represents a surrounding
homogeneous and
isotropic spacetime with
scale factor $a_{\mathfrak M} (\tau)$
and negative-definite spatial scalar curvature,
$- 1/(2 R_{\mathfrak M})^2$.
Finally, one could mention that
potentials of type \eqref{e:grgal} were studied, albeit in the context
of a linear Schr\"odinger equation,
in \Refs~\cite{zs18a,zs18b}.
\scn{Density of Effective Gravitating Matter
}{s:den}
In this \Sec, let us derive spherically-symmetric density profiles
of effective gravitating matter $\rho_\grpott$, which
formally corresponds to the superfluid-vacuum induced potential \eqref{e:grev}.
Contrary to the directly observable value of orbital velocity derived in \Sec~\ref{s:frc},
such density
depends more substantially
on the choice of an observer.
It can be
defined
either via the Lorentz-covariant definition
~\eqref{e:setdef},
or via
the Poisson equation,
which is a non-relativistic version of \eq \eqref{e:setdef}.
Correspondingly,
one would obtain different results,
which will be discussed~below.
In accordance with the last paragraph of Appendix
\ref{s:aa},
this \Sec's computations cannot take into account any secondary induced matter,
such as the equilibrium configurations of mass-energy emerging as a result of interaction between scalar and tensor modes
of superfluid vacuum's excitations.
One can show that such equilibria do exist, manifesting themselves in a form of general relativistic nonsingular horizon-free stellar-like objects
or particle-like Q-balls,
therefore,
in reality such objects would definitely contribute to the density associated with dark matter.
In other words, here
we are restricting ourselves to background values of~density.
In this \Sec~only,
we shall be temporarily assuming
that
$
M = M (r)
$,
otherwise the corresponding contribution to the density profile would be identically zero,
and thus~non-indicative.
\sscn{Galilean symmetry
}{s:denn}
In this case,
one defines an effective matter density by virtue of the Poisson equation.
Taking the whole potential \eqref{e:grev} and assuming spherical symmetry,
we obtain
} \def\bw{\begin{widetext}
\begin{eqnarray}
\rho_\grpott
(r)
&\equiv&
\frac{1}{4 \pi G}
\left[
\grpot} %\def\grpott{\grpot_{(\Psi)}'' (r)
+
\frac{2}{r} \grpot} %\def\grpott{\grpot_{(\Psi)}' (r)
\right]
=
\rho_{\text{smi}} (r)
+
\rho_{\text{RN}} (r)
+
\rho_{\text{N}} (r)
+
\rho_{\text{gal}} (r)
+
\rho_{\text{mgl}} (r)
+
\rho_{\text{dS}} (r)
,
\lb{e:rhodm}
\end{eqnarray}
} \def\ew{\end{widetext}
where
\begin{eqnarray}
\rho_{\text{smi}}
(r)
&=&
\frac{3 \chi \, q}{4 \pi G m r^4}
\left[1 -
\frac{2}{3}
\ln{\left(
\frac{r}{\bar{\ell}}
\right)}
\right]
,\lb{e:rhosmi}\\
\rho_{\text{RN}}
(r)
&=&
-
\frac{a_0 q}{2 \pi G m r^4}
=
\frac{k_e Q^2}{4 \pi r^4}
,\lb{e:rhomic}\\
\rho_{\text{N}}
(r)
&=&
-
\frac{1}{4 \pi r}
M''(r)
,\lb{e:rhonew}\\
\rho_{\text{gal}}
(r)
&=&
\frac{\chi \, b_0}{4 \pi G m r^2}
=
\frac{v_\text{gal}^2}{4 \pi G r^2}
,\lb{e:rhogal}\\
\rho_{\text{mgl}}
(r)
&=&
\frac{a_1 \, b_0}{2 \pi G m \bar{\ell} r}
=
\frac{\chi \, b_0}{2 \pi G m L_\chi r}
,\lb{e:rhomgl}\\
\rho_{\text{dS}}
(r)
&=&
-
\frac{3 a_2 \, b_0}{2 \pi G m \bar{\ell}^2}
=
-
\frac{3 b_0}{2 \pi G m L_{\text{dS}}^2}
= \text{const}
,\lb{e:rhocos}
\end{eqnarray}
whereby
the
sum
of the last three densities
can be regarded as a density corresponding to the astronomical-scale ``dark matter'' and ``dark energy'',
which will be further justified in \mbox{\Secs \ref{s:frc} and \ref{s:uexp}}.
Furthermore,
these
formulae were derived on the assumption
that the gravitational coupling constant $G$ is the same for all length scales;
which is valid when any influence from the term \eqref{e:grsmi} can be disregarded.
However, this term might cause an additional effect, discussed in \Sec~\ref{s:psmi}:
it makes the gravitational coupling constant vary with distance.
If this does happen, then in \eqs \eqref{e:rhodm}--\eqref{e:rhocos}
one should replace $G$ with a running constant given by \eq \eqref{e:Geff}.
\sscn{Lorentz symmetry
}{s:denr}
In this case,
one defines effective matter density by virtue of Einstein field
equations with the induced stress-energy tensor defined by \eq \eqref{e:setdef}.
Using it together with \eq \eqref{e:grev} and metric~\eqref{e:metrst},
we obtain
} \def\bw{\begin{widetext}
\begin{eqnarray}
{\dnir}_\grpott
(r)
&\equiv&
-
\frac{1}{4 \pi G r}
\left[
\grpot} %\def\grpott{\grpot_{(\Psi)}' (r)
+
\frac{1}{r} \grpot} %\def\grpott{\grpot_{(\Psi)} (r)
\right]
=
\tilde{\rho}_{\text{smi}} (r)
+
\tilde{\rho}_{\text{RN}} (r)
+
\tilde{\rho}_{\text{N}} (r)
+
\tilde{\rho}_{\text{gal}} (r)
+
\tilde{\rho}_{\text{mgl}} (r)
+
\tilde{\rho}_{\text{dS}} (r)
+
\tilde{\rho}_{\grpot} %\def\grpott{\grpot_{(\Psi)}_0} (r)
,
\lb{e:rhodmr}
\end{eqnarray}
} \def\ew{\end{widetext}
where
\begin{eqnarray}
\tilde{\rho}_{\text{smi}}
(r)
&=&
\frac{\chi \, q}{4 \pi G m r^4}
\left[1 -
\ln{\left(
\frac{r}{\bar{\ell}}
\right)}
\right]
,\lb{e:rhosmir}\\
\tilde{\rho}_{\text{RN}}
(r)
&=&
-
\frac{a_0 q}{4 \pi G m r^4}
=
\frac{k_e Q^2}{8 \pi r^4}
,\lb{e:rhomicr}\\
\tilde{\rho}_{\text{N}}
(r)
&=&
\frac{1}{4 \pi r^2}
M'(r)
,\lb{e:rhonewr}\\
\tilde{\rho}_{\text{gal}}
(r)
&=&
-
\frac{\chi \, b_0}{4 \pi G m r^2}
\left[1 +
\ln{\left(
\frac{r}{\bar{\ell}}
\right)}
\right]
,\lb{e:rhogalr}\\
\tilde{\rho}_{\text{mgl}}
(r)
&=&
-
\frac{a_1 \, b_0}{2 \pi G m \bar{\ell} r}
=
-
\frac{\chi \, b_0}{2 \pi G m L_\chi r}
,\lb{e:rhomglr}\\
\tilde{\rho}_{\text{dS}}
(r)
&=&
\frac{3 a_2 \, b_0}{4 \pi G m \bar{\ell}^2}
=
\frac{3 b_0}{4 \pi G m L_{\text{dS}}^2}
= \text{const}
,\lb{e:rhocosr}\\
\tilde{\rho}_{\grpot} %\def\grpott{\grpot_{(\Psi)}_0}
(r)
&=&
-
\frac{\Phi_0}{4 \pi G r^2}
=
-
\frac{1}{4 \pi G m r^2}
\left(
a_0 b_0
+ \frac{q}{L_{\text{dS}}^2}
\right)\!,~~~
\lb{e:rhophior}
\end{eqnarray}
whereby
the
sum
of last four densities
can be regarded corresponding to the astronomical-scale ``dark matter'' and ``dark energy'',
which will be further justified in \Secs \ref{s:frc} and \ref{s:uexp}.
Density \eqref{e:rhophior}
is a somewhat surprising contribution, because~it corresponds to the constant term $\Phi_0$ in \eq \eqref{e:grev},
which is not supposed to affect trajectories; at least, in~classical mechanics.
Unless its presence can be confirmed by observations,
it must be regarded as a gauge term, or~as an artifact of approximations underlying \eq \eqref{e:metrst}.
Furthermore,
a running gravitational coupling constant can not be implemented in the relativistic case
as simply as in \Sec~\ref{s:denn}.
To preserve Lorentz invariance,
one has to associate this coupling with a four-dimensional scalar;
which automatically upgrades general relativity to a scalar-tensor gravity
with a non-minimally coupled scalar field.
This theory cannot be written by hand, but~it must be derived
in a way which is similar to that used in \Sec~\ref{s:cfs}.
Comparing results of Sections~\ref{s:denn} and \ref{s:denr},
one can conclude that the definition of effective matter density is somewhat ambiguous:
in particular, it drastically depends on symmetry assumptions.
Therefore, further experimental studies should help to empirically
establish which symmetry is more appropriate to use when dealing with
``dark''~phenomena.
\scn{Galactic Rotation Curves}{s:frc}
In this \Sec, we demonstrate how induced gravitational potential
can
explain
various phenomena, which are usually attributed
to dark matter.
Let us focus on the terms \eqref{e:grgal} and \eqref{e:grmgl}, which~were partially discussed in \Sec~\ref{s:pmgal}.
Because they become significant at a galactic scale and above (i.e.,~a kiloparsec to megaparsec scale),
it is natural to conform them to astronomical observations; such as those of rotation curves in~galaxies.
In a spherically symmetric case,
velocity curves of stars orbiting
with non-relativistic velocities
on a plane in a central gravitational potential $\grpot} %\def\grpott{\grpot_{(\Psi)} (r)$
can be estimated using a simple formula
$
v^2
= R a_c =
R \,
\grpot} %\def\grpott{\grpot_{(\Psi)}' (R)
$,
where $v$ is the orbital velocity,
$a_c$ is the centripetal acceleration,
and $R$ is the orbit's radius.
The~cylindrically symmetric case can be considered by analogy,
by assuming various
disk models~\cite{too63,cas83}.
Considering the terms \eqref{e:grgal} and \eqref{e:grmgl} in conjunction with the Newtonian term \eqref{e:grnew},
we thus obtain
\begin{eqnarray}
v (R)
&=&
\left\{
R \Der{}{R}
\left[
\grpot_{\text{N}} (R) + \grpot_{\text{gal}} (R) + \grpot_{\text{mgl}} (R)
\right]
\right\}^{1/2}
\nonumber \\&=&
\sqrt{
v_\text{N}^2
+ v_\text{gal}^2
+
\grpot_{\text{mgl}} (R)
}
, \lb{e:ov}
\end{eqnarray}
where
\begin{eqnarray}
v_\text{N}^2
&=&
\frac{G M}{R}
=
\frac{a_1 q}{m \bar{\ell}} \frac{1}{R}
=
- \grpot_{\text{N}} (R)
,\\
v_\text{gal}^2
&=&
\chi c_b^2 = \frac{\chi b_0}{m}
= \text{const}
, \lb{e:ovgal}
\end{eqnarray}
while the contribution from the term \eqref{e:grcos} is disregarded for now,
due the assumed smallness of the ``local'' cosmological constant
$1/R_{\text{dS}}^2$;
and
the contribution from the term \eqref{e:grsmi} is disregarded due the assumed smallness of
the corresponding characteristic length, according to discussion in \Sec~\ref{s:psmi}.
In the case of a galaxy,
the contribution from the Newtonian term $\grpot_{\text{N}}$ rapidly decreases as $R$ grows.
Correspondingly, the~main contribution would then come
from
the second term in a row, $\grpot_{\text{gal}}$, and~then from
the third term, $\grpot_{\text{mgl}}$.
From \eq \eqref{e:ovgal} one can see that the contribution from $\grpot_{\text{gal}}$ is constant,
which explains the average flatness of galactic rotation~curves.
Notice that the value of velocity $v_\text{gal}$
depends on one of the wavefunction parameters $\chi$
and one of quantum temperature parameters $b_0$.
Both are not \textit{a priori} fixed parameters of the model,
cf. \mbox{\eqs \eqref{e:bqtem} and \eqref{e:ava}},
but vary depending on the
environment and conditions:
background superfluid gets affected by the gravitational potential it induces,
because this potential acts upon the surrounding conventional matter, thus creating density inhomogeneity.
Therefore, $\chi$ and $b_0$ should generally be different for each galaxy;
and hence should $v_\text{gal}$ be.
Similar to the case of $v_\text{gal}$,
the
parameter $a_1$ hence a value $\grpot_{\text{mgl}} (R)$ at a fixed $R$
will also be dependent on the gravitating object they refer to.
This potential
should usually be negligible on the inner scale length of a galaxy,
but
as $R$ grows towards the extragalactic length scale, $R
\gtrsim
10 \, \text{kpc}$,
rotation curves should start to deviate from flat:
\begin{equation}\lb{e:grclin}
v (R)
\approx
v_\text{gal}
\sqrt{
1+
\frac{R}{L_\chi}
},
\end{equation}
which can be used for estimating the combination
of superfluid vacuum parameters
$\chi \bar{\ell}/a_1 $ empirically.
In cases where
the contribution from other terms of the induced potential cannot be neglected,
{Equation
~\eqref{e:ov} must be generalized to include
those~too.
Possible galactic-scale regions, where this non-flat asymptotics
should become visible,
depending on a value $L_\chi$,
are the outer regions of large spiral galaxies, such as M31 or M33~\cite{rvt80,cs00,ccf09,clw10,kcc17},
where $\grpot_{\text{mgl}} (r)$ can not only overtake $\grpot_{\text{N}} (r)$ but also become
comparable with $\grpot_{\text{gal}} (r)$.
\scn{Accelerating Expansion of the Universe}{s:uexp}
This phenomenon is usually explained by introducing exotic forms of relativistic matter,
such~as dark energy; usually modeled by various long-range scalar fields, which are
assumed not to affect the numerous
particle physics experiments on the Earth.
The superfluid vacuum approach offers a simple framework,
which can explain the Universe's expansion as an observer-dependent effect,
without~involving any matter other than the background superfluid~itself.
\sscn{Conformally flat spacetime and dilaton field}{s:cfs}
Following work~\cite{z11appb},
let us consider the most simple possible special case:
laminar flow of a logarithmic background condensate in a state
$| \Psi_\circ \rangle$,
described by \eq \eqref{e:oF} at $b = \text{const}$,
with a constant velocity $\ve u^{(0)}$,
if viewed as,
from the F-observer's perspective, an~embedding into underlying Euclidean space.
On the other hand, what does a R-observer see?
Due to a well-known separability property
of the logarithmic Schr\"odinger equation~\mbox{\cite{bb76,af18,bcs19,lzh19,z18zna}},
the phase of
its simplest ground-state solutions is
a linear function of a
radius-vector:
\begin{equation}\lb{e-phaselin}
i \ln{
\left(
\frac{\Psi_\circ (\ve{r}, t)}{|\Psi_\circ (\ve{r}, t)|}
\right)}
\propto
\ve u^{(0)} \cdot \ve{r} + f (t)
,
\end{equation}
where $\ve u^{(0)}$
is a constant 3-vector, and~$f(t)$ is an arbitrary function of time.
In this case, the~fluid-Schr\"odinger analogy
confirms
that the background condensate does flow with a constant
velocity
$
\ve u
= - i \eta \boldsymbol{\nabla} \ln{
\left(
\Psi_\circ
/ |\Psi_\circ|
\right)
}
\propto
\ve u^{(0)}
$.
Recalling \eq \eqref{e:metr1}, it means that
the background
geometry induced by such solutions is conformally flat:
\begin{equation}\lb{e-metlog}
d s^{2}_{(0)}
\propto
\frac{1}{\dnc}
|\Psi_\circ (\ve{r}, t)|^2
\left[
-
c_s^2
d t^2
+
\left(d \ve{r} - \ve u^{(0)} d t\right)^2
\right]
,
\end{equation}
where
$c_s$ is given by \eq \eqref{e:cappappln},
in a leading-order approximation
with respect
to the Planck~constant.
Spacetime of a type \eqref{e-metlog} lies within a large class
of manifolds with the
vanishing Weyl tensor -- a type $\mathbf{O}$ in the
Petrov classification.
This is the class that all FLRW spacetimes belong to, including those which describe the Universe
with accelerating expansion -- simply written in conformally-flat coordinates
instead of comoving ones.
Using definition \eqref{e:setdef},
we obtain
an induced stress-energy tensor
for our system:
\begin{eqnarray}\lb{e-setelog}
\kappa
\widetilde{T}_{\mu\nu}
&=&
\tilde D
\Bigl[
\nabla_\mu \nabla_\nu {\phi}
-
\nabla_\mu {\phi} \nabla_\nu {\phi}
\nonumber \\&&
-
g_{\mu\nu}
\left(
\nabla_\lambda \nabla^\lambda {\phi}
+
\tfrac{1}{2}
(\tilde D-1)
\nabla_\lambda {\phi} \nabla^\lambda {\phi}
\right)
\Bigr]\!,~~~
\end{eqnarray}
where $\tilde D = D -2 = 2$,
$\nabla$ is a covariant derivative
with respect to metric $g$,
and by ${\phi}$
we denote the induced scalar field:
\begin{equation}\lb{e-phipsi}
{\phi} =
\ln{
\left(
|\Psi_\circ (\ve{r}, t)|^2/\dnc
\right)
}
,
\end{equation}
up to an additive~constant.
This stress-energy tensor strongly resembles
the one occurring in
the theory of gravity with a scalar field.
One can verify that
it
can be indeed derived
from the following scalar-tensor
gravity action~functional
\begin{equation}\lb{e:actv}
\tilde{\cal S}
[g,\, \not\!{{\phi}}]
\propto
\int d^D x
\sqrt{- g}
\,
\text{e}^{\tilde D {\phi}}
\left[
R + \tilde D (\tilde D + 1)
(\nabla {\phi})^2
\right]
,
\end{equation}
where the notation ``$\not\!{{\phi}} $'' reminds us
that
the field ${\phi}$ is fixed by the solution
of the original quantum wave equation, cf. \eq \eqref{e-phipsi},
while the variation of action must be taken with respect to the metric only.
In other words,
both metric $g$ and dilaton ${\phi}$ are
induced
by the superfluid vacuum being in a state described by $\Psi_\circ (\ve{r}, t)$.
Thus, we have found yet another example of the differences between the F-observer's
and R-observer's pictures of reality.
While the former sees
a background quantum fluid flowing with a constant velocity
in three-dimensional Euclidean space,
the latter observes itself as being inside four-dimensional spacetime governed
by a Lorentz-covariant scalar-tensor~gravity.
An action functional \eqref{e:actv}
therefore explains why covariant models involving
scalars
provide a robust description of the large-scale
evolution of the Universe,
agreeing with current observational data;
yet no quanta of relativistic dilaton have thus far been~detected.
This correspondence
also reveals the
limitations of the relativistic description
itself:
if the superfluid vacuum goes into a different quantum state,
then
one gets a different expression for the induced metric,
scalar, stress-energy tensor and covariant
action.
In fact, for~more complicated superfluid flows, even the condition \eqref{e-phaselin},
leading to a conformal flatness,
can become relaxed to an asymptotic one.
Therefore, depending on the physical configuration (determined by external potential,
if any,
and boundary conditions), nonlinear coupling behaviour
and the quantum state the vacuum is in,
small fluctuations and probe particles
would obey different covariant actions.
Consequently,
a~R-observer would have to tweak its field-theoretical models by hand;
with
the unified picture being observable only
at the level
of a F-observer.
\sscn{Cosmological constant and local expansion mechanism}{s:cco}
In \Sec~\ref{s:cfs} we considered an example of the global superfluid flow
which would be seen by a relativistic observer
as the accelerating expansion of the observable Universe.
What about locally induced gravity \eqref{e:grev}, can it cause similar effects?
Let us consider once again the induced potentials of
Sections~\ref{s:pcos} and \ref{s:pmgal}.
The de Sitter term from \Sec~\ref{s:pcos} predominates if we consider
the physical vacuum alone, without~any generated matter therein.
This is perhaps only valid for the early Universe,
such as the one which existed during the
inflationary epoch.
In the current epoch, our spacetime can only be de Sitter asymptotically or
even
approximately;
therefore the potential $\grpot_{\text{dS}}$ must be negligible, unless~one considers
very large length scales.
For example, if~a length scale $R_{\text{dS}}$,
which generally depends on a massive object defining the frame of reference
of the function \eqref{e:avex},
is comparable to a size of the observable Universe
($\sim$ 10 Gpc),
then $R_{\text{dS}}$ can be related to the cosmological constant $\Lambda$
as
\begin{equation}\lb{e:rhlam}
R_{\text{dS}}^
\text{(cos)}
= \sqrt{3/\Lambda}, \ \
\Lambda \sim
10^{-56} \ \text{cm}^{-2}
,
\end{equation}
therefore,
the term \eqref{e:grcos} becomes substantial only at a gigaparsec scale.
This relation yields an empirical constraint for a
combination of characteristic parameters
of superfluid vacuum, average quantum temperature and $\Psi_{\text{vac}}$:
\begin{equation}\lb{e:ablam}
\frac{
a_2^
\text{(cos)}
b_0^
\text{(cos)}
}{m \bar{\ell}^2}
\approx
\frac{1}{6} \Lambda c^2
\sim
10^{-36} \ \text{s}^{-2}
,
\end{equation}
see also
the discussion around
\eq \eqref{e:bqtem}.
All this essentially means that
$\Lambda$
is not a fundamental constant of Nature,
but a combination of various parameters of superfluid vacuum, including quantum temperature
and Gaussian width of the condensate function $\Psi_{\text{vac}}$.
Therefore, its smallness, sometimes referred as ``vacuum catastrophe'' \cite{car01},
can easily be explained by the fact that:
either average quantum temperature across
the Universe, or~wavefunction's width, or~both,
are sufficiently small compared to $m c^2$ and $\bar{\ell}$, respectively;
thus resulting in the overall smallness of the ratio on the left-hand side of \eq \eqref{e:ablam}.
The line element, which results from taking $\grpot_{\text{dS}}$,
$\grpot_{\text{mgl}}$, $\grpot_{\text{gal}}$, $\grpot} %\def\grpott{\grpot_{(\Psi)}_0$ and $\grpot_{\text{N}}$ together,
can be written in static coordinates as
} \def\bw{\begin{widetext}
\begin{eqnarray}
d s^2_{(\text{cos+N})}
\!&\approx&\!
- c_{(0)}^2
\left[
1
+ \delta_0
- \frac{r_H}{r}
+ \beta_\chi^2
\ln{\!\left(
\frac{r}{\bar{\ell}}
\right)}
+ \frac{r}{R_{\mathfrak M}} - \frac{r^2}{R_{\text{dS}}^2}
\right] d t^2
\nonumber \\&&
+ \frac{d r^2}{
1
+
\delta_0
- r_H/r
+
\beta_\chi^2 \ln{\left(r/\bar{\ell}\right)}+ r/R_{\mathfrak M} - r^2 / R_{\text{dS}}^2
} +
r^2 d \sigma^2} %\Omega_{(2)}^2
,
\lb{e:medsa2}
\end{eqnarray}
} \def\ew{\end{widetext}
where $\delta_0 = 2 \grpot} %\def\grpott{\grpot_{(\Psi)}_0/c_{(0)}^2$.
In this metric,
the
Schwarzschild term
ensures that spacetime singularity at $r = 0$ is ``dressed'' by the black hole horizon,
while at large $r$ we still have spacetime of a
FLRW type.
While from~the viewpoint of a F-observer, no spacetime singularity would
pose a problem, because~quantum wavefunction remains regular and normalizable
at any non-negative value of $r$.
This reaffirms our earlier statement that cosmological singularity is an artifact
of the low-momentum approximation of superfluid vacuum~\cite{z11appb,z20rss}.
According to \eq \eqref{e:memgl2},
the linear potential term \eqref{e:grmgl}, when taken alone,
induces
the universal acceleration
$c_{(0)}^2/2 R_{\mathfrak M}$,
occurring due to the spatial curvature,
when seen by a R-observer in its own local static coordinate system.
This contributes
to the Hubble expansion induced by the quadratic term \eqref{e:grcos}.
Thus,
in the relativistic picture,
the non-small fluctuation of superfluid vacuum produces an effect at the center of a gravitating
configuration;
and therefore contributes to the explicit
rotational motions of the stars inside this configuration,
which can be seen as a consequence of curved~spacetime.
\sscn{Expansion mechanisms and cosmological coincidences}{s:coi}
Comparing \Secs \ref{s:cfs} and
\ref{s:cco},
one can see that they describe different expansion mechanisms.
The mechanism of \Sec~\ref{s:cfs} occurs due to the global flow of background superfluid,
assumed to be laminar, which is ``seen'' by a R-observer as a FLRW-type spacetime;
the resolution of various cosmological problems
related to this mechanism
was discussed in \Sec~5 of Ref.~\cite{z11appb}.
On the other hand,
in \Sec~\ref{s:cco}, expansion is explained as a cumulative effect from terms
in metric \eqref{e:medsa2},
which~do not vanish at spatial infinity,
induced by a ``local'' wavefunction associated with a gravitating configuration or body.
This wavefunction can be regarded as a fluctuation (not necessarily small) of the
global wavefunction from \Sec~\ref{s:cfs}.
The interplay between these mechanisms depends on the length scales of the quadratic and linear terms,
$R_{\text{dS}}$ and $R_{\mathfrak M}$.
Unless
a cumulative expansion effect from
asymptotically non-vanishing terms taken together
is,
by some extraordinary coincidence, exactly equal
to the expansion due to the global flow mechanism,
its rate must be different
from that of the global flow-induced~expansion.
The occurrence of an additional expansion mechanism,
at the scale of a supercluster, such as our Virgo or Laniakea,
could explain the remarkable discrepancy between measurements of the Hubble constant using different methods from
those
based on the whole Universe expansion, such as cosmic microwave background radiation (CMB).
Among non-CMB methods one could mention
Cepheid calibration,
time-delay cosmography,
and
geometric distance measurements to megamaser-hosting galaxies~\cite{fmg01,cfs19,pbr20}.
From a theoretical point of view,
a scenario with different expansion rates seems slightly more plausible, because~it does not require
an explanation why accelerations from different mechanisms should be exactly equal to each other
(this coincidence should not be confused with the conventional cosmological coincidence
which we will discuss next).
Furthermore,
our approach offers a simple explanation of the cosmological coincidence problem itself~\cite{car01},
in both the simplified and quantitative versions~thereof.
The simplified formulation of the cosmological coincidence states that if dark matter and dark energy were different kinds of matter, then during the Universe's evolution they
should have evolved independently of each other;
therefore their distributions would be uncorrelated by now - which does not seem to be the case.
Superfluid vacuum theory trivially resolves this paradox: because ``dark matter'' and ``dark energy'' are
actually induced phenomena
and manifestations of the same object, superfluid vacuum,
they cannot be independent from each~other.
The quantitative formulation of the coincidence problem is
an explanation requirement for why the ratio of DM- and DE-associated densities is of order one,
despite the reasons given in the simplified formulation.
In our approach, we can regard all terms of the induced
potential, which do not vanish at spatial infinity,
as being associated with ``dark'' effects;
but inside this group we cannot unambiguously separate DM-attributed effects from DE ones.
For example,
potential \eqref{e:grmgl} is intermediate
between de Sitter and logarithmic, thus it affects both galactic rotation curves and Hubble expansion,
cf. \mbox{\Secs \ref{s:frc} and \ref{s:cco}}.
If, for~simplicity,
we consider only the local expansion mechanism of \Sec~\ref{s:cco}
and omit the contribution from $\grpot} %\def\grpott{\grpot_{(\Psi)}_0$,
then the cosmological coincidence can be reformulated
as the following~condition:
\begin{equation}\lb{e:coi}
\frac{
\adn{gal}
+
\adn{mgl}
}{
\adn{mgl}
+
\adn{dS}
}
=
{\cal O} ( 1 )
,
\end{equation}
where
$\adn{}
$'s
are average values corresponding to densities from \Sec~\ref{s:denr}.
The numerator of this ratio represents average density
of effective ``dark matter'',
while the denominator represents effective ``dark energy'' density; or at least
the predominating proportions thereof.
In general, this condition
simply imposes yet another constraint for the parameters
of the theory.
It is trivially satisfied if the
involved parts of the gravitational potential, hence the associated
densities,
are of the same order of magnitude,
if averaged on a large scale.
Moreover, even if
$\adn{gal}$
or
$\adn{dS}$ are much smaller
than the remaining involved densities,
but the value
$\adn{mgl}$
is substantial,
then
the
relation~\eqref{e:coi} still holds,
due to the presence of
$\adn{mgl}$
in both parts of the ratio.
\scn{Conclusions}{s:con}
Working within the framework of the post-relativistic theory of physical vacuum,
based on the logarithmic superfluid model,
we derived induced gravitational potential, corresponding to
a generic quantum wavefunction of the vacuum.
This mechanism is radically different from the one used in models of
relativistic classical fluids and fields,
which are based on modifying the stress-energy tensor in Einstein field~equations.
The form of such a wavefunction is motivated by ground-state
solutions of quantum wave equations of a logarithmic type.
Such equations find
fruitful applications in the theory of strongly-interacting quantum fluids,
and have been successfully applied to laboratory
superfluids~\mbox{\cite{z12eb,z19ijmpb,sz19}}.
We note that,
in principle, one is not precluded from adding other types of nonlinearity,
such as polynomial ones,
into the condensate wave equations,
but the role of logarithmic nonlinearity is
crucial
Thus, we used a logarithmic superfluid model with variable nonlinear coupling,
because it accounts for an effect of the environment in a more realistic way than
the logarithmic model with a constant coupling.
As a result,
for the trapless version of our model,
we have four parameters, but~only two of them
are \textit{a priori} fixed, whereas the other two
can vary, depending on the
quantum thermodynamic properties of
the environment under consideration.
Additionally, a~number of parameters come from the wavefunction solution itself.
Those are not independent parameters of the theory, but~functions thereof.
Because we do not yet know the exact form of the superfluid wavefunction,
see remarks at the end of \Sec~\ref{s:gr},
we leave those parameters to be empirically estimated, or~bound, at~the stage
of current~knowledge.
It turns out that gravitational interaction has a multiple-scale structure
in our theory: induced potential is dominated by different terms at each length scale;
such that
one can distinguish
sub-Newtonian, Newtonian (inverse-law), galactic (logarithmic-law), metagalactic (linear-law), and~cosmological (square-law) parts.
A relativistic observer,
who operates with low-momentum small-amplitude fluctuations of superfluid
vacuum, observes this induced potential by
measuring the
trajectories of probe particles
moving along
geodesics in induced four-dimensional pseudo-Riemannian spacetime.
The metric of the latter is determined by virtue of the BEC-spacetime correspondence
and fluid-Schr\"odinger analogy,
applied~jointly.
The sub-Newtonian part of the induced gravitational potential is
defined as one which grows faster than the inverse law, as~distance tends towards zero.
It
can be naturally divided into the following two parts.
One part has an inverse square law behaviour,
and thus can be associated with the gravitational field caused by a $U(1)$ gauge charge,
such as an electric charge.
On a relativistic level, it is described by Reissner-Nordstr\"om spacetime.
The other part has `inverse square times logarithm' law behaviour,
which might become substantial at both ultra-short and macroscopic distances,
depending on the values of the corresponding parameters.
If it ``survives'' at macroscopic distances, then it upgrades Newton's
gravitational constant to a function of length, such that gravity has both strong and weak~regimes.
With the potential or spacetime metric in hand,
one can, in~principle, assign effective fictitious matter density to our potential,
which corresponds to ``dark matter'' and ``dark energy''.
This can be done in two ways:
either by Einstein field equations in a relativistic case, or~the Poisson equation in a non-relativistic one.
It should be noted that the resulting density in each of the cases can be modified, depending on whether the gravitational constant is considered to be running or not.
This will require more verification from future experimental and theoretical studies.
Furthermore,
on a galactic scale and above, the~potential
is dominated by non-Newtonian terms,
which do not vanish at spatial infinity.
This explains the non-Keplerian behaviour of rotation curves in galaxies,
which is often attributed to dark matter.
Our model, not only explains the average flatness of
galactic rotation curves,
but also makes a number of new predictions.
One of them is
the approximately linear law behavior of gravitational
potential on a metagalactic scale, which is an
intermediate scale between galactic distances and the size of the observable universe.
This should partially affect galactic rotation curves too:
as the distance from the gravitating center grows further towards the metagalactic length scale, a~squared velocity's profile
asymptotically changes from being
flat towards linear, cf. \eq \eqref{e:grclin}.
On the other hand,
at the largest length scale, the~induced potential displays square law behaviour.
If the quadratic term is negative-definite, then
the corresponding metric describes (asymptotically) de Sitter space,
merely written in static coordinates.
Taken together with the contribution from the linear potential term,
this explains the accelerating expansion of the corresponding spacetime region,
which~is usually associated with dark~energy.
Such expansion could supplement the ``global'' one,
caused by
laminar flow of background logarithmic superfluid absent any other matter,
which
induces
a FLRW-type
spacetime.
The occurrence of more than one type of expansion mechanism,
could be responsible for the discrepancy between measurements of the Hubble constant using different~methods.
The relevant problems, such as smallness of cosmological constants and cosmological coincidence,
were also~discussed.
To conclude,
we used the BEC-spacetime correspondence and fluid-Schr\"odinger analogy
to argue that the description of reality and fundamental symmetry crucially depend on the choice of an observer.
We demonstrated that
both dark matter and dark energy are related phenomena, and~different manifestations of the same object,
superfluid vacuum, which acts by
inducing both gravitational potential and~spacetime.\\
\noindent
\textbf{Abbreviations}\\
The following abbreviations are used in this manuscript:\\
\noindent
\begin{tabular}{@{}ll}
BEC & Bose-Einstein condensate\\
CDM & Cold dark matter\\
CMB & Cosmic microwave background\\
DE & Dark energy\\
DM & Dark matter\\
dS & de Sitter\\
FLRW & Friedmann-Lema\^itre-Robertson-Walker\\
F-observer & Full observer\\
RN & Reissner-Nordstr\"om\\
R-observer & Relativistic observer\\
SVT & Superfluid vacuum theory
\end{tabular}
\begin{acknowledgments}This research is supported by Department of Higher Education and Training of South Africa and in part by National Research Foundation of South Africa.
Proofreading of the manuscript by P. Stannard is greatly appreciated.
\end{acknowledgments}
|
\section{Introduction}
The precise shaping of the time structure of laser pulses has been the basis for time-resolved spectroscopy and ultrafast science \cite{Zewail1645,krausz2009attosecond}. Engineering the local polarization state and/or the spatial phase of the wavefront offers further opportunities for applications. Such local engineering of optical fields can be brought about by various means: Appropriately designed photonic elements such as waveguides \cite{born2013principles, saleh2019fundamentals} host eigenmodes with desirable polarization distribution. Also, specifically decorated plasmonic structures allow tuning the spatial distribution of the spin (polarization) and orbital (phase) structure of the electromagnetic field \cite{tsesses2018optical, Spektor1187, PhysRevX.9.021031, dai2019ultrafast, davis2020ultrafast, D0NR00618A, omatsu2017focus, Kerber_2018, ayuso2019synthetic}. In this way, optical skyrmions and plasmonic waves carrying orbital angular momentum were realized. While plasmonic fields overcome the limitation on the spatial resolution of diffraction-limited propagating (laser) wave in free space, the latter offers a large flexibility in tuning the frequency, intensity, and carrier-envelope phase at very low power losses. Therefore, much efforts were devoted to the spatial structuring of the polarization and wavefront phase of freely propagating waves. For instance, laser pulses with azimuthal and radial polarization \cite{born2013principles, saleh2019fundamentals} were realized in a wide frequency range \cite{Mitchell:s,zhan2009cylindrical,Erd_lyi_2008,kozawa2005generation,hernandez2017extreme}. \\
The topic in general is attracting much research recently due to the great potential for fundamental and applied science \cite{Rubinsztein_Dunlop_2016}. For instance, orbital (OAM) and spin (SAM) angular momentum carrying pulses can generate unidirectional charge currents \cite{Quinteiro:09, watzel2016optical, sederberg2020Vectorized} which is interesting for opto (spin)electronic applications \cite{PhysRevB.90.115401,Watzel:12, solyanik2019spin,PhysRevB.93.045205,KOC2015599,PhysRevB.100.115308,doi:10.10631.5027667, ji2020photocurrent}. For molecules \cite{PhysRevLett.89.143601,PhysRevA.71.055401,PhysRevLett.96.243001}, structured laser pulses are expected to yield new information, particularly on chiral and helical molecular aggregates \cite{PhysRevA.99.023837,Wozniak:19,ayuso2019synthetic}. \\
For atoms, the electron wave function is extremely localized with respect to variation in the spin or orbital parts of optical fields. Thus, at first glance, it seems that the local structure of the laser field is marginal when considering the response of a random distribution of non-interacting atoms in the laser spot \cite{kaneyasu2017limitations}. On the other hand, the local spatial variation of the SAM and phase of the laser-field are not diffraction-limited and may change on the sub-wavelength scale.
However, already in the perturbative regime, one can identify an optimal position of the atom within the laser spot where the phase structure of the laser is important, and at the same time the transition probability is sizable \cite{PhysRevA.86.063812}. For a trapped cold atom, theory and experiment revealed much details on which types of bound-bound transitions are triggered by OAM-carrying light (for instance, \cite{schmiegelow2016transfer, afanasev2018experimental,PhysRevA.97.023422,Duan_2019,peshkov2016absorption}).
Interaction with OAM carrying pulses with resulting in photoemission has also been the subject of various theoretical studies \cite{picon2010photoionization, seipt2016two, baghdasaryan2019dichroism,watzel2016discerning}. \\
Summarizing the status of knowledge on electrons in structured propagating fields, one may say that the direct ionization by an OAM carrying pulse is relatively well understood. Continuum-continuum (CC) transitions involving OAM exchange are less unstudied, however. The investigation of the two-photon transition matrix element corresponding to a conventional XUV field and an IR vortex illustrated the impact of the (transferred) OAM on the CC phase and the associated time delay \cite{giri2020signatures}. The recent experimental and theoretical work \cite{de2020photoelectric} (cf. in particular the supplemental materials of \cite{de2020photoelectric}) clearly highlights the importance of using the correct structured-light-matter interaction, including the longitudinal field component and also the role of the position of the atom in the laser spot. These two aspects (among others) are inherent features of the interaction of matter with structured light and will be discussed at length within the framework of our developed theoretical model.\\
Interaction with SAM structured fields (vector beams) with atoms is much less studied. How a spin-orbitally coupled electronic system react to vector beams was addressed in Refs. \cite{watzel2019magnetoelectric, watzel2020nanostructures}. High harmonic generation (HHG) upon a strongly nonlinear driving of atoms with vector beams was reported in Refs.\, \cite{hernandez2017extreme,watzel2020multipolar,watzel2020topological}.
The interaction of atoms with optical skyrmions was recently formulated in \cite{watzel2020topological}, and nonlinear electron dynamics was simulated.\\
In this study we will deal with intense propagating (laser) fields having
a spatial variation in the orbital or spin or in both (such as in skyrmions) \cite{allen1992orbital, andrews2012angular,bliokh2015transverse,Barnett_2016,ALLEN1999291,Bouchard_2014}. One key goal is to derive a unified quasi-analytical description of non-linear electron dynamics in such structured laser beams.
Moreover, the derived expressions allow for the incorporation of a laser pulse with arbitrarily (but reasonably) SAM or/and OAM structured pulse. Utilizing the strong-field approximation \cite{keldysh1965ionization, faisal1973multiple, reiss1980effect, lewenstein1994theory}, the derived electron state in the presence of structured light fields are used for the calculations of laser-induced electron emission in dependence on the optical OAM of the laser-assisting fields. In addition, it is demonstrated how intense and tightly focused SAM-structured vector beams \cite{zhan2009cylindrical} can be employed for linear momentum texturing of electronic wave packet.
Furthermore, we demonstrate that (photo)electron dynamics can sample the spatio-temporal structure of intense propagating optical skyrmions where the optical OAM and SAM are intertwined \cite{tsesses2018optical,Spektor1187,dai2019ultrafast,watzel2020topological}.
\section{Mathematical description of structured propagating laser fields}
\label{sec:vortex_math}
\subsection{General considerations}
In the vicinity of the optical axis, Bessel \cite{volke2002orbital} and Laguerre-Gaussian modes \cite{allen1992orbital} exhibit locally similar functional dependencies \cite{watzel2020multipolar}. Cylindrical coordinates $\pmb{r}=\left\{ \rho,\varphi,z \right\}$ allow a convenient description of several types of structured beams, including beams carrying OAM, radially and azimuthally polarized vector beams as well as propagating optical skyrmions. Generally, for these beams the key ingredient is the vortex vector field $\pmb{A}_{\rm OV}^{m^{(a)},\sigma_{\rm L}}(\pmb{r},t)$, whose mathematical expression is given explicitly below. $m^{(a)}$ is the topological charge with the superscript $a=\pm$ signaling the vortex chirality reflecting the direction of the embodied OAM. $\sigma_{\rm L}=\pm1$ indicates the polarization state (direction of the SAM). For the behavior of the non-paraxial vector potential employed below,
the direction of the beam's OAM relative to SAM is important. We distinguish between the \emph{parallel} case, i.e., ${\rm sgn}(\sigma_{\rm L})=a$ and the \emph{antiparallel} case where ${\rm sgn}(\sigma_{\rm L})=-a$. Of a particular interest is the region near the optical axis $\rho\approx 0$ (on the scale of the beam waist $w_{\rm L}$) \cite{quinteiro2015formulation}, as discussed for instance in \cite{PhysRevA.71.055401} for the case of OAM carrying laser beam.
\subsubsection{Parallel SAM and OAM, ${\rm sgn}(\sigma_{\rm L})=a$}
The vector potential in the parallel case can be written as
\begin{equation}
\begin{split}
\pmb{A}^{m^+,+1}_{\rm OV}(\pmb{r},t)=&A_0F_m(\rho)e^{im\varphi} e^{i(q_zz-\omega_{\rm L} t)}\hat{e}_{\sigma_{\rm L}=+1} \\
& + {\rm c.c.}
\end{split}
\label{eq:parallel1}
\end{equation}
and
\begin{equation}
\begin{split}
\pmb{A}^{m^-,-1}_{\rm OV}(\pmb{r},t)=&(-1)^mA_0F_m(\rho)e^{-im\varphi} \\
&\times e^{i(q_zz-\omega_{\rm L} t)}\hat{e}_{\sigma_{\rm L}=-1} + {\rm c.c.} .
\end{split}
\label{eq:parallel2}
\end{equation}
A striking feature is the absence of a longitudinal component so that the approximate (i.e., for $\rho\approx 0$ ) vector potential is fully transverse. The laser field propagates effectively along the $z$ axis with the wave vector $q_z$, the amplitude is set by $A_0$ which determines the laser intensity and $\hat{e}_{\sigma_{\rm L}}=(\hat{e}_\rho + i\sigma_{\rm L}\hat{e}_\varphi)e^{i\sigma_{\rm L}\varphi}$ is the circular polarization vector. The dispersion relation is $$q_\perp^2+q_z^2=q_{\rm L}^2=\omega_{\rm L}^2/c^2,$$
and the radial distribution reads
$$F_m(\rho)=(q_r\rho)^m.$$
The transverse wave vector $q_\perp $ is related to the beam waist $w_{\rm L}$ as $q_\perp \simeq 1/w_{\rm L}$. The parallel class vector potentials are fully transverse (within the adopted approximation) so that OAM and SAM are separable \cite{bliokh2015transverse}. The carried total angular momentum is $\hbar(m+\sigma_{\rm L})$.
\subsubsection{Anti-parallel SAM and OAM, ${\rm sgn}(\sigma_{\rm L})=-a$}
For the antiparallel case one finds
\begin{equation}
\begin{split}
\pmb{A}^{m^+,-1}_{\rm OV}(\pmb{r},t)=&\left(F_m(\rho) \hat{e}_{\sigma_{\rm L}=-1} +i2m\frac{q_\perp}{q_z}F_{m-1}(\rho)e^{-i\varphi}\hat{e}_{z}\right) \\
& \times A_0e^{im\varphi} e^{i(q_zz-\omega_{\rm L} t)} + {\rm c.c.}
\end{split}
\label{eq:antiparallel1}
\end{equation}
and
\begin{equation}
\begin{split}
\pmb{A}^{m^-,+1}_{\rm OV}(\pmb{r},t)=&\left(F_m(\rho)\hat{e}_{\sigma_{\rm L}=+1} + i2m\frac{q_\perp}{q_z}F_{m-1}(\rho) e^{+i\varphi}\hat{e}_{z}\right) \\ & \times (-1)^mA_0e^{-im\varphi} e^{i(q_zz-\omega_{\rm L} t)} + {\rm c.c.}.
\end{split}
\label{eq:antiparallel2}
\end{equation}
These equations evidence the presence of a longitudinal component, { whose strength (relative to the transverse component) is determined by the focusing condition.}\\
For the following discussion, it is important to note that the longitudinal component scales as $r^{m-1}$ in the antiparallel case. Consequently, for $m=1$ the on-axis field does not vanish along the propagation direction. The accuracy of the chosen approximation is demonstrated in Fig.\,\ref{fig1} (first row), where the cartesian components of the vector potential function for both classes are presented. Up to distances of $25/q$, an optical vortex is well-described by the approximation given in Eqs.\,\eqref{eq:parallel1} and \eqref{eq:antiparallel1}. Note, the presence of a longitudinal component does not invalidate $\pmb{\nabla}\cdot\pmb{A}^{m^{(a)},\sigma_{\rm L}}_{\rm OV}(\pmb{r},t)=0$ for all vector potentials in Eqs.\eqref{eq:parallel1}-\eqref{eq:antiparallel2}. Due to the non-vanishing longitudinal component, OAM and SAM are not separable \cite{bliokh2015transverse}.
\subsection{Optical vortices, polarization structured beams and propagating optical skyrmions}
{Propagating optical fields with desired polarization and spatial phase textures can be constructed as a linear combination of the vector functions $\pmb{A}^{m^{(a)},\sigma_{\rm L}}_{\rm OV}(\pmb{r},t)$:}
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{Fig1}\\
\caption{Upper row: Spatial distributions of the vector potential components corresponding to the parallel and antiparallel OAM and SAM. The vortex topological charge is $+1$ while the laser focus is set by $\alpha=\arctan(q_\perp/q_z)$ and is chosen to be $1^\circ$. Lower row: the gradients in both cases of the vector potentials along the $x$-axis.}
\label{fig1}
\end{figure}
\begin{itemize}
\item A \emph{linearly polarized} beam carrying the OAM value $m^{(a)}\hbar$ can be viewed as an optical vortex with topological charge $m^{(a)}$ and is expressible as the superposition $\pmb{A}^{m^{(a)},+1}_{\rm OV}(\pmb{r},t)\pm\pmb{A}^{m^{(a)},-1}_{\rm OV}(\pmb{r},t)$. The vector field does not transfer a net SAM during the interaction.
\item An \emph{azimuthally polarized vector beam} \cite{zhan2009cylindrical} (AVB) can be written as the coherent sum $\pmb{A}^{+1,-1}_{\rm OV}(\pmb{r},t) + \pmb{A}^{-1,+1}_{\rm OV}(\pmb{r},t)$. The explicit expression of the vector potential is given by \cite{watzel2019magnetoelectric}
\begin{equation}\label{eq:avb}
\pmb{A}_{\rm AVB}(\pmb{r},t) = A_0q_\perp\rho\sin(q_zz-\omega_{\rm IR} t)\hat{e}_\varphi.
\end{equation}
Note the absence of the longitudinal component.
\item A \emph{radially polarized vector beam} (RVB) is expressible as $\pmb{A}^{+1,-1}_{\rm OV}(\pmb{r},t) - \pmb{A}^{-1,+1}_{\rm OV}(\pmb{r},t)$ \cite{watzel2020multipolar}. For a tightly focused beam \cite{dorn2003sharper}, it is sometimes useful to Taylor-expand the spatial radial distributions $F_{m=0}$ to second or higher orders. For instance, $F_0(\rho)\approx 1 - (q_\perp\rho_0)^2/4$ and $F_1(\rho)=q_\perp\rho_0 - (q_\perp\rho_0)^3/8$. The corresponding vector potential reads then
\begin{equation}
\begin{split}
\pmb{A}_{\rm RVB}(\pmb{r},t) =& A_0\left[\vphantom{\frac{q}{q}} \left(q_\perp\rho-(q_\perp\rho_0)^3/8\right) \cos(q_zz-\omega_{\rm IR} t) \hat{e}_\rho \right. \\
&\left. - 2\frac{q_\perp}{q_z}\left(1-(q_\perp\rho_0)^2/4\right)\sin(q_zz-\omega_{\rm IR} t) \hat{e}_z \right],
\end{split}
\end{equation}
where a longitudinal field is present at the optical axis, and still $\pmb{\nabla}\cdot\pmb{A}_{\rm RVB}(\pmb{r},t)=0$ is sustained. As for AVB, the radially polarized vector beam does not exhibit a net OAM nor SAM. Yet, the well-defined spatial structuring of SAM does affect electron dynamics in a unique way. For instance, the AVB can act on electrons as a gauge-invariant vector potential, leading to a transient Aharonov-Bohm-type, non-dissipative current (meaning, AVB triggers a time-dependent orbital magnetic moment even if the net SAM of the field vanishes) \cite{watzel2019magnetoelectric}.
\item An \emph{optical propagating skyrmion} \cite{watzel2020topological} reveals a rich phase and position-dependent polarization landscape. A convenient mathematical representation is $\alpha\pmb{A}^{m_1^+,+1}_{\rm OV}(\pmb{r},t)+\beta\pmb{A}^{m_2^{-},-1}_{\rm OV}(\pmb{r},t)$ ($\alpha,\beta$ are real numbers) resulting in the vector potential
\begin{equation} \label{eq:optskr}
\begin{split}
&\pmb{A}_{\rm OS}^{m_1^+,m_2^-}(\pmb{r},t) = A_0e^{i(q_zz-\omega_{\rm IR}t-(m_2+1)\varphi)} \\
&\quad\times\left[\left( \vphantom{e^{i(m_1+2)}} \alpha e^{i(m_1+m_2+2)\varphi}(q_\perp\rho)^{m_1} + \beta(-q_\perp\rho)^{m_2} \right)\hat{e}_\rho \right. \\
&\quad+ i\left.\left( \vphantom{e^{i(m_1+2)}} \alpha e^{i(m_1+m_2+2)\varphi}(q_\perp\rho)^{m_1} - \beta(-q_\perp\rho)^{m_2} \right)\hat{e}_\varphi\right]\\
&\quad + c.c..
\end{split}
\end{equation}
The vector potential is not transverse and the carried OAM and SAM \cite{bliokh2015transverse} are intertwined in a way that may be characterized by a topological quantity in analogy to the skyrmion number of a magnetic skyrmion (for details on magentic skyrmions we refer to Ref. \cite{Fert2017} and references therein). In Ref.\cite{watzel2020topological} we discussed a possible definition of optical skyrmionic topological index but we should note a key difference to magnetic skyrmions. In \cite{jia2019twisting} for example, we discussed how by increasing the radius of a magnetic disc (corresponding to changing the beam waist in the laser beam) the magnetic ordering transforms from a vortex to a skyrmionic state eventually reaching a uniform magnetic ordering; an OAM carrying wave \cite{Jia2019} may also occur. This behavior can be described within one unified, mathematically consistent picture. In contrast to optics, in magnetism we are dealing with static vector field stabilized by internal interactions. For linear materials, the time-average of propagating (or plasmonic) electromagnetic fields vanishes, regardless of their spatial or spin structure. The relevance of geometry or topology of optical fields is manifested by the type of the processes they trigger when interacting with matter \cite{ayuso2019synthetic,hernandez2017extreme,Watzeltoroid,watzel2019magnetoelectric,watzel2020multipolar,watzel2020topological}, as illustrated below.
\end{itemize}
\section{Light-matter interaction}
The interaction of an electron with an arbitrarily structured laser field propagating in free space with the wave vector $k$
and described by the vector potential $\pmb {A}_L$
follows from a minimal coupling scheme. The Langrangian density is cast as
${\cal L}={\cal L}_{mech}+{\cal L}_{field}+ \pmb {j}\cdot\pmb {A}_L -\rho_c \Phi_L$, where ${\cal L}_{mech}$ and ${\cal L}_{field}$ are the mechanical and field parts, and $\pmb{j}$ and $\rho_c$ are the current and charge densities, respectively. $\Phi_L$ is the scalar potential. Thus, the interaction of matter with the field delivers
in general two contributions to the Hamiltonian. The current-current coupling term yields
(atomic units (a.u.) are used, unless stated otherwise)
\begin{equation}
\hat{H}_{\rm CC-int}=\frac{1}{2}\left[ -i\pmb{\nabla}\cdot\pmb{A}_L(\pmb{r},t) -2i \pmb{A}_L(\pmb{r},t)\cdot\pmb{\nabla} + \pmb{A}_L^2(\pmb{r},t)\right]
\end{equation}
Even in free space the charge density couples to the laser scalar potential $\Phi_L$, where
$\partial_t \Phi_L(\pmb{r},t)=- c^2 \pmb{\nabla}\cdot\pmb{A}_L$.
It is possible to transform from this Lorenz gauge to an instantaneous (Coulomb) gauge \cite{peshkov2017photoexcitation, de2020photoelectric} (which is adopted henceforth) by introducing the vector potential $\pmb{A}=\pmb{\nabla}(\pmb{\nabla}\cdot \pmb{A}_L)/k^2+ \pmb{A}_L$.
In this gauge $\Phi(\pmb{r},t)$ does not appear in the light-matter interaction, however the longitudinal component of the vector potential $\pmb{A}$ can be decisive \cite{watzel2020multipolar,watzel2020topological} affecting $\rho_c$ via non-dipolar transitions. Thus, denoting with $\hat{\pmb{p}}_i$ the momentum operator of the $i^{th}$ electron, and with $\pmb{A}(\pmb{r},t)=\sum_j \pmb{A}_{\rm j}(\pmb{r},t)$ the total (sum) vector potential of all present fields $\pmb{A}_{\rm j}$, we may write
in general for the light-matter interaction
\begin{equation}
\hat{H}_{\rm int}(t)= \sum_i\pmb{A}(\pmb{r}_i,t)\cdot\hat{\pmb{p}}_i + \frac{1}{2}\pmb{A}^2(\pmb{r}_i,t).
\end{equation}
The expression is formally similar to the case of uniform fields but physically leads to the same effects as in Lorenz gauge such as the possible excitation of volume charge-density modes \cite{watzel2019magnetoelectric}.\\
Figure\,\ref{fig1} shows the local gradient of the vector potential (for parallel and antiparallel classes) in the vicinity of the optical axis. In contrast to the transversal components, the corresponding gradient remains finite when approaching $\rho=0$. Consequently, the dynamics around the optical axis in the parallel case is mainly driven by this gradient. Note, $\pmb{\nabla}\pmb{A}_{\rm OV}^{m^+,+1}$ points into the $\varphi$-direction which is associated with the intrinsic
phase structure of the vortex field and is proportional to $m/\rho$. In the antiparallel case, the near-axis dynamics is dominated by the longitudinal component (also on the level of the gradient).
\subsection{Volkov-type states in arbitrarily structured laser field}
\label{sec:SFA}
Let us consider the simplest example of a unbound electron subject to a strong structured laser field.
In the case of spatially homogeneous vector potential such a state is described by a Volkov wave \cite{wolkow1935klasse}.
Analytical (Volkov-like) solutions for the unbounded electron motion in generally structured fields were not reported sofar. Below, we derive under certain conditions "structured-light Volkov wave" (SL-VW).
From the discussion so far and considering Fig.\,\ref{fig1}c-d, we conclude that reasonable approximations should capture the action of the vector potential gradient $\pmb{\nabla}\pmb{A}^{m^{(a)},\sigma_{\rm L}}_{\rm OV}$.\\
An atom at the position $\pmb{r}_0=(x_0,y_0,0)^T$ in the laser focal plane experiences
a vector potential at $\pmb{r}_0$ that varies smoothly in space. Suppressing for clarity sub and superscripts of $\pmb{A}$ and Taylor expanding around $\pmb{r}_0$ yields for the $j$-th component to a first order
${A}_j(\pmb{r},t)={A}_j(\pmb{r}_0,t) + \sum_i {r}_i {M}_{ij}$ or equivalently (${r}_i$ is $i^{th}$ component of $\pmb{r}$)
\begin{equation}
\pmb{A}(\pmb{r},t)=\pmb{A}(\pmb{r}_0,t) + \pmb{r}\cdot\doubleunderline{M}(t).
\label{eq:Aloc}
\end{equation}
The matrix elements of $\doubleunderline{M}(t)$ are $M_{ij}= \left. \partial_{r_i} {A}_j(\pmb{r},t) \right|_{\pmb{r}=\pmb{r}_0}$. The treatment of the first order term enables the inclusion of non-dipolar effects \cite{walser2000high, chirilua2002nondipole}. In the presence of $\pmb{A}(\pmb{r},t)$, the Hamiltonian of an electron bound by the potential
$V(\pmb r)$ reads then
\begin{equation}
\hat{H}(t)=\frac{1}{2}\left[\hat{\pmb{p}} + \pmb{A}(\pmb{r}_0,t) + \pmb{r}\cdot\doubleunderline{M}(t) \right]^2 + V(r).
\label{eq:Ht}
\end{equation}
With $\pmb{\mathcal{E}}=-\partial_t\pmb{A}(\pmb{r}_0,t)$ being the electric field and using the gauge transformation $|\Psi_L\rangle = e^{i\pmb{r}\cdot\pmb{A}(\pmb{r}_0,t)}|\Psi\rangle$ one obtains
\begin{equation}
\hat{H}_L(t)=\frac{1}{2}\left[\hat{\pmb{p}}+ \pmb{r}\cdot\doubleunderline{M}(t) \right]^2 + \pmb{r}\cdot\pmb{\mathcal{E}}(t) + V(r),
\end{equation}
Noting that $[\hat{\pmb{p}},\pmb{r}\cdot\doubleunderline{M}(t)]_-=0$,
and neglecting higher order terms in the local variation of $\pmb{A}$ (i.e., $[\pmb{r}\cdot\doubleunderline{M}(t)]^2\approx 0$) we write
\begin{eqnarray}\label{eq:volkov13}
\hat{H}_L(t)&=& \hat{H}_{\rm Volkov}(t) + V(r), \nonumber\\
\hat{H}_{\rm Volkov}(t)&=&\frac{1}{2}\hat{\pmb{p}}^2 + \pmb{r}\cdot\doubleunderline{M}(t)\cdot\hat{\pmb{p}} + \pmb{r}\cdot\pmb{\mathcal{E}}(t).
\end{eqnarray}
There is an opportunity for a nonperturbative analytical treatment if considering $\hat{H}_{\rm Volkov}(t)$ and $V$ to act separately, which is the basis of the strong field approximation \cite{faisal1973multiple,keldysh1965ionization,reiss1980effect} (strong means that the field terms dominates $V$ when considering the unbound electron dynamics). Such an approximation is worthwhile doing, for a series of important phenomena and experiments can be
described reasonably well within this strong field approximation \cite{Amini_2019}. For us the key issue here is to find the function $|\Psi_{\pmb{p}}^{\rm (V)}(t)\rangle$ obeying
\begin{equation}\label{eq:volkovH}
i\partial_t|\Psi_{\pmb{p}}^{\rm (V)}(t)\rangle = \hat{H}_{\rm Volkov}|\Psi_{\pmb{p}}^{\rm (V)}(t)\rangle.
\end{equation}
To derive the expression for this state which we termed above SL-VW, one may proceed at first as for the conventional Volkov state by writing the ansatz \cite{wolkow1935klasse, pisanty2018high}
\begin{equation}
|\Psi_{\pmb{p}}^{\rm (V)}(t)\rangle = e^{-\frac{i}{2}\int^t\pi^2(\pmb{p},\tau){\rm d}\tau}|\pmb{\pi}(\pmb{p},t)\rangle.
\label{eq:PsiVolkov}
\end{equation}
The states $|\pmb{\pi}(\pmb{p},t)\rangle$ are plane waves propagating with the kinematic momenta $\pmb{\pi}(\pmb{p},t)$, meaning $\hat{\pmb{p}}|\pmb{\pi}(\pmb{p},t)\rangle= \pmb{\pi}(\pmb{p},t)|\pmb{\pi}(\pmb{p},t)\rangle$.
Thus, Eq. (\ref{eq:volkovH}) amounts to integrating
\begin{equation}
\frac{\partial\pmb{\pi}(\pmb{p},t)}{\partial t} + \pmb{\mathcal{E}}(t) = - \doubleunderline{M}(t)\cdot\pmb{\pi}(\pmb{p},t).
\end{equation}
Recalling that $\int_t\doubleunderline{M}(t){\rm d}t\sim(q/\omega)\pmb{A}(\pmb{r}_0,t)=(1/c)\pmb{A}(\pmb{r}_0,t)$,
we write $\pmb{\pi}(\pmb{p},t)=\pmb{p}+\pmb{A}(\pmb{r}_0,t) + \delta\pmb{\pi}(\pmb{p},t)$ and seek a solution to first order in (1/c) which yields
\begin{equation}
\pmb{\pi}(\pmb{p},t)=\pmb{p}+\pmb{A}(\pmb{r}_0,t) - \int^t{\rm d}\tau \doubleunderline{M}(\tau)\cdot(\pmb{p}+\pmb{A}(\pmb{r}_0,\tau)).
\label{eq:kin_mom}
\end{equation}
As detailed below, even in regions where $\pmb{A}(\pmb{r}_0,t)$ is very small $\doubleunderline{M}$ may be large enough such that the last term in eq.(\ref{eq:kin_mom}) may even dominate the behaviour of the Volkov phases ($A_j$ and gradient of ${A}_j$ are independent). Such a case is encountered when an atom resides in the vicinity of the optical vortex core in the parallel class, described by $\pmb{A}_{\rm OV}^{m^+,+1}$ or $\pmb{A}_{\rm OV}^{m^-,-1}$.\\
The key quantity of SL-VW is its
phase $S_V(\pmb{p},t,\pmb{r}_0)=\frac{1}{2}\int^t {\rm d}\tau\,\pmb{\pi}(\pmb{p},\tau,t')^2$, or explicitly \\
\begin{widetext}
\begin{equation}
S_V(\pmb{p},t,\pmb{r}_0)
=E_pt + \frac{1}{2}\int^t{\rm d}\tau\pmb{A}^2(\pmb{r}_0,\tau) +\pmb{p}\cdot\int^t{\rm d}\tau\left[ \pmb{A}(\pmb{r}_0,\tau)-\int^\tau{\rm d}t'' \doubleunderline{M}(t'')\cdot(\pmb{p}+\pmb{A}(\pmb{r}_0,t''))\right]
\label{eq:SV}
\end{equation}
\end{widetext}
where $E_p=p^2/2$. The second term is related to the action of the local ponderomotive potential
(terms containing higher powers of $\pmb{A}(\pmb{r}_0,t)$ are neglected). In principle, having Eq.(\ref{eq:SV}) one may in retrospect insert the determined SL-VW into Eq.(\ref{eq:volkovH}), and assures the consistency of the approximations. The explicit form of the SL-VW depends on the type of the vector potentials and is discussed below for some typical cases.
\subsection{Electrons in a strong OAM carrying laser field}
Let us consider the phase of SL-VW for the case where an optical vortex of the parallel class acts on an electron that has been released from an atom residing at $\pmb{r}_0=(\rho_0\cos\varphi_0,\rho_0\sin\varphi_0,0)^T$. It reads
\begin{widetext}
\begin{equation}
S_V^{(m^+,+1)}(\pmb{p},t,\pmb{r}_0)= \frac{1}{2}\left(p^2 + A_0^2(q_\perp\rho_0)^{2m}\right)t + \alpha_m\sin\vartheta_{\pmb{p}}\left[\frac{m}{\rho_0} \frac{pq_z}{\omega q_\perp}\sin\vartheta_{\pmb{p}}\cos(2\varphi_{\pmb{p}}-\omega t) - \left(1 + \frac{q_zp\cos\vartheta_{\pmb{p}}}{\omega}\right) \sin(\varphi_{\pmb{p}}-\omega t) \right]
\label{eq:OV1}
\end{equation}
\end{widetext}
We expressed $\pmb{p}$ with its amplitude $p$ and the spherical angles $\vartheta_{\pmb{p}},\varphi_{\pmb{p}}$. In Eq.(\ref{eq:OV1}) $\alpha_m=A_0p(q_\perp\rho_0)^m/\omega$ characterizes the displacement of the electron at the position $\pmb{r}_0$ in the structured laser field.
Note that $\doubleunderline{M}(t)\cdot\pmb{A}(\pmb{r}_0,t) \propto qA_0^2 $ and was therefore neglected.
In the antiparallel case which occurs for instance for a topological charge $m^-$ and $\sigma_{\rm L}=+1$ (opposite chiralities of SAM and OAM), the most relevant contributions to the SL-VW phase are
\begin{widetext}
\begin{equation}
\begin{split}
S_V^{(m^-,+1)}(\pmb{p},t,\pmb{r}_0)=&\frac{1}{2}\left(p^2 + A_0^2(q_\perp\rho_0)^{2m}\right)t + \alpha_m\left[ (-1)^{m+1} \frac{m}{\rho_0} \left(\frac{2}{q_z}\cos\vartheta_{\pmb{p}} +
\frac{p}{2\omega}(1+3\cos^2\vartheta_{\pmb{p}})\right)\cos\omega t \right. \\
&\left. - \left(1 + \frac{q_zp\cos\vartheta_{\pmb{p}}}{\omega}\right)\sin\vartheta_{\pmb{p}} \sin(\varphi_{\pmb{p}}-\omega t) \right].
\end{split}
\label{eq:OV2}
\end{equation}
\end{widetext}
The influence of the orbital angular momentum $m$ of the laser fields enters the SL-VW for both cases as terms which scale as $m/\rho_0$. For very large (compared with $q_z^{-1}$) axial distances $\rho_0=\sqrt{x_0^2+y_0^2}$, differences between Eq.\,\eqref{eq:OV1} and Eq.\,\eqref{eq:OV2} vanish, and the SL-VW converges to the conventional Volkov wave for spatially uniform circularly polarized light. This is to be expected, as $m$ is related to the optical axis. Hence, for an atom at large $\rho_0$, the phase of the vector potential is basically constant. This observation can be exploited for spatially resolved photoemission on the scale below the optical diffraction limit: Photoelectrons that show dependence on $m$ must have started from regions around the optical axis, or in general from regions where the spatial phase of the vector potential varies significantly on the scale of the atomic wave functions \cite{watzel2016discerning}. This argument may also serve for using the photoelectrons to map the structure of the optical fields, as demonstrated below.
The independence of the SL-VW given by (\ref{eq:OV1},\ref{eq:OV2}) on the atom-position polar angle $\varphi_0=\arctan(-y_0/x_0)$ reflects the symmetry of the considered system (the vector potential and the atom).
\subsection{Electrons driven by polarization textured vector beams}
For an unbound electron in a vector beam with azimuthal polarization we find for the SL-VW phase the form
\begin{widetext}
\begin{equation}
\begin{split}
S_V^{\rm AVB}(\pmb{p},t,\pmb{r}_0)=&\frac{1}{2}\left(p^2t + (A_0q_\perp\rho_0)^2 \right)t + \alpha_1\left( 1 + \frac{pq_z}{\omega}\cos\vartheta_{\pmb{p}}\right) \sin\vartheta_{\pmb{p}}\sin(\varphi_{\pmb{p}}-\varphi_0)\cos(\omega t).
\end{split}
\label{eq:AVB}
\end{equation}
\end{widetext}
For a radially polarized vector beam the expression is markedly different encompassing the influence of the longitudinal component which becomes more important for tighter focusing. We infer the expression
\begin{widetext}
\begin{equation}
\begin{split}
S_V^{\rm RVB}(\pmb{p},t,\pmb{r}_0)=\frac{1}{2}\left[p^2 + A_0^2\left(q_\perp^2\rho_0^2 + \frac{q_\perp^2}{q_z^2}\left(4-q_\perp^2\rho_0^2\right)\right)\right]t
+ \alpha_1&\left[ \vphantom{\frac{2}{q_z\rho_0}} \left(1-\frac{q_\perp^2\rho_0^2}{8}+\frac{pq_z}{\omega}\cos\vartheta_{\pmb{p}}\right) \sin\vartheta_{\pmb{p}}\cos(\varphi_{\pmb{p}}-\varphi_0)\sin(\omega t) \right. \\
&\left.
- \frac{2}{q_z\rho_0}\left(1-(q_\perp\rho_0)^2/4\right) \cos\vartheta_{\pmb{p}}\cos(\omega t)\right].
\end{split}
\label{eq:RVB}
\end{equation}
\end{widetext}
\subsection{Electron quantum dynamics in intense optical skyrmionic fields}
In the case of an optical propagating skyrmion, some properties of the modified Volkov phase can be inferred from the two vortices with different topological charges $m_1^+$ and $m_2^-$ that form the skyrmion but the cylindrical symmetry cannot be exploited as in previous cases and the expression is thus more involved:
\begin{widetext}
\begin{equation}
\begin{split}
S_V^{\rm OS}(\pmb{p},t,\pmb{r}_0)=&\frac{1}{2}\left[p^2 + \left(\alpha^2(q_\perp\rho_0)^{2m_1} + \beta^2(q_\perp\rho_0)^{2m_2} \right)\, A_0^2\right]t\\
& + \frac{A_0p\sin\vartheta_{\pmb{p}}}{\omega_{\rm L}}\left[ \vphantom{\frac{A_0p\sin{\vartheta}}{\omega_{\rm L}}} \beta(q_\perp\rho_0)^{m_2}
\sin[m_2(\pi+\varphi_0)+\varphi_{\pmb{p}} + \omega_{\rm L}t] - \alpha(q_\perp\rho_0)^{m_1}
\sin[m_1\varphi_0 + \varphi_{\pmb{p}} - \omega_{\rm L}t] \right] \\
& + \frac{A_0p^2\sin^2\vartheta_{\pmb{p}}}{\omega_{\rm L}^2}\left[ \vphantom{\frac{A_0p\sin{\vartheta}}{\omega_{\rm L}}} \alpha \frac{m_1}{\rho_0}(q_\perp\rho_0)^{m_1}
\cos[2\varphi_{\pmb{p}} - \omega_{\rm L}t - (m_1-1)\varphi_0] \right. \\
&\quad\quad\quad\quad\quad\quad\quad
\left.+ \beta \frac{m_2}{\rho_0}(q_\perp\rho_0)^{m_2}
\cos[2\varphi_{\pmb{p}} + \omega_{\rm L}t - m_2(\pi+\varphi_0) -\varphi_0] \right].
\end{split}
\label{eq:OSphase}
\end{equation}
\end{widetext}
The meaning of the various terms entering $S_V^{\rm OS}(\pmb{p},t,\pmb{r}_0)$ follows from the discussions of Eq.(\ref{eq:OV2}), as the skyrmionic field receives contributions from two vortices with winding numbers $m_{1,2}$. Whether terms associated with $m_1$ or $m_2$ are locally more important depends on the ratio
$(\alpha(q_\perp\rho_0)^{m_1})/(\beta(q_\perp\rho_0)^{m_2})$.
\section{Applications}
Having derived the electronic wave function in the presence of a structured intense laser field, we utilize it for the description of selected physical processes, namely a) photoionisation assisted by structured intense laser fields, b) for steering and momentum texturing of electronic wave packets with radially polarized vector beam, and c) for spatio-temporal mapping of skyrmionics optical fields. For concreteness we use in all calculations below a He atom as a typical target. The potential $V$ confining the electrons to the atom is modelled within the effective single-particle approach, discussed and mathematically detailed in Ref. \cite{tong2005empirical}.
\subsection{Photoionization of atoms assisted by intense optical vortices}
Let us consider the liberation of a valence shell electron upon the absorption of one (X)UV photon with energy $\hbar\omega_{\rm X}$. In addition, an intense structured laser field $L$ with frequency $\hbar\omega_{\rm L}$ is present. This laser $L$ affects strongly the photoelectron wavepacket dynamics in a way that can be quantified as follows:
The photoionization amplitude in the presence of the two laser fields reads \cite{faisal1973multiple, milovsevic2019atom}
\begin{equation}
\mathcal{A}_{\pmb{p}}(\pmb{r}_0)=-i\int_{-\infty}^{\infty}{\rm d}t'\langle\Psi^{-}_{\pmb{p},\rm SV}(\pmb{r}_0,t')|\hat{H}_{\rm X}(t')|\Psi_i(t')\rangle.
\label{eq:amplitude0}
\end{equation}
Here, $\hat{H}_{\rm X}(t')$ is the interaction Hamiltonian of the valence shell electron with the (X)UV field and $|\Psi_i(t)\rangle$ is the initial state. The time-dependent final state is
\begin{equation}
|\Psi^{(-)}_{\pmb{p},\rm SV}(\pmb{r}_0,t)\rangle=e^{i[\pmb{A}(\pmb{r}_0,t)-\pmb{K}(\pmb{p},t)]\cdot\pmb{r}} |\Psi^{(-)}_{\pmb{p}}\rangle e^{-iS_V(\pmb{p},t,\pmb{r}_0)}.
\label{eq:Psi_ASV}
\end{equation}
where $\pmb{K}(\pmb{p},t)=\int^t{\rm d}\tau\,\doubleunderline{M}(\tau)\cdot(\pmb{p}+\pmb{A}(\pmb{r}_0,\tau)$ and $|\Psi^{(-)}_{\pmb{p}}\rangle$ satisfies the time-independent Schr\"{o}dinger equation for the atomic Hamiltonian $\hat{H}_{\rm at}=\hat{\pmb{p}}^2/2 + V(r)$ for the kinetic energies $E_p>0$. The Volkov phase $S_V(\pmb{p},t,\pmb{r}_0)$ is given in Eq.\,\eqref{eq:SV}. In contrast to Eq.\,\eqref{eq:PsiVolkov}, we use the full scattering states in the atomic potential $V(r)$ instead of plane waves.
The justification of using the final states $|\Psi^{(-)}_{\pmb{p},\rm SV}(t)\rangle$ follows the formal steps when deriving the Coulomb-Volkov ansatz, given in Refs.\,[\onlinecite{faisal2016strong,milovsevic2019atom}].
Similarly, it can be shown that Eq.\,\eqref{eq:amplitude0} is the zeroth order amplitude corresponding to the integral equation describing the time evolution operator
\begin{equation}
\hat{U}(t,t')=\hat{U}_{\rm SV}(t,t') - i\int_{t'}^t{\rm d}\tau\,\hat{U}_{\rm SV}(t,\tau)\hat{V}_{\rm SV}(\tau)\hat{U}(\tau,t')
\label{eq:TimeEvolution}
\end{equation}
which involves the full Hamiltonian $\hat{H}(t)$ [cf.\,Eq.\,\eqref{eq:Ht}]. We note that $|\Psi^{(-)}_{\pmb{p},\rm SV}(t\rightarrow\infty)\rangle=|\Psi^{(-)}_{\pmb{p}}\rangle$, where $t\rightarrow\infty$ is the time when measurement is conducted (at the photoelectron detector) while all laser fields are off.\\
While the formulation applies to all types of structured fields, we select the case when
the assisting laser field L is an infrared intense (IR) optical vortex carrying OAM and it propagates collinearly with the homogeneous (on the scale of the atoms), weaker XUV field, a situation which has been experimentally realized recently \cite{de2020photoelectric, mazza2016angular}. Our focus here is on the theoretical aspects. Details of the experiments and comparison with theory are discussed at length in Ref.\cite{de2020photoelectric} where the target was a thermal gas cell of helium atoms \cite{meyer2010two}.\\
During detection, via the photoelectron energy selection we may zoom to those events where one photon from the XUV laser and one IR photon are involved. Interestingly, we may even achieve a time ordering on which photon is absorbed first by choosing a tightly focused XUV and less focused IR laser beam, in which case our photoelectrons first absorb the XUV photon in the region where the IR laser has a very low (or vanishing) intensity and then experience the IR laser on their way out to the detector (cf. Fig.\,\ref{fig:fields}a-b ). This scenario implies also a spatial resolution on the position of the involved atom on the scale of the laser spot of the XUV laser. In a way, our setting resembles the case of STED-microscopy (STED=stimulated emission depletion) \cite{Westphal246}. In fact, if we would investigate few XUV photon processes (that we can select via the photoelectron energy), we would increase the spatial dependence to around the intensity peak of the XUV laser.\\
Generally, if we are interested in effects related to the spatial structure of the laser,
the photoelectron should be slow enough such that the first kinetic energy term in Eq.(\ref{eq:volkov13}) does not completely overwhelm the field terms.
Interestingly, the slow photoelectrons that take notice of the local phase structure of the IR laser, absorb the IR photon in the vicinity of the optical axis and not at the IR field maximum (where the local field resembles for the atom a Gaussian field).
Hence, our approach in deriving Eqs.\,\eqref{eq:parallel1}-\eqref{eq:antiparallel2} is indeed useful.\\
For concreteness, we choose the XUV photon energy to be $\hbar\omega_{\rm X}=30$\,eV. The durations of both laser fields are assumed very long (short pulses can be associated with the streaking regime \cite{kazansky2010angle}), in which case on both sides of the main-photonline, additional lines are well-developed and are separated by $\hbar\omega_{\rm L}=1.55$\,eV.\\
The spatial phase of the IR pulse is reflected in the difference of the photoelectron yields corresponding to the IR pulses with $m^+$ or $m^-$, meaning an OAM-induced dichroism. The conventional circular dichroism and how it relates to target's orientation and/or alignment is well-established in the literature (for example, in Refs.\cite{kazansky2011circular, mazza2014determining} and references therein), similar arguments apply to circularly polarized optical vortices \cite{baghdasaryan2019dichroism, seipt2016two}. The circular dichroism in photoexcitation by using optical vortices was presented in Ref.\,\cite{afanasev2017circular}, where transitions involving higher multipolarity revealed a strong difference. \\
We consider an XUV field with a fixed helicity of $\sigma_{\rm X}=+1$ so that the ejected photoelectron is orbitally oriented.
The XUV laser field spatial distribution is assumed to be Gaussian $f(\rho_0)=e^{-(\rho_0/(2w_{\rm X}))^2}$, where $w_{\rm X}$ is the effective width of the focal spot.
\begin{figure}[t!]
\centering
\includegraphics[width=0.97\columnwidth]{Fig2}
\caption{XUV photoionization assisted with an optical vortices laser pulse. a) Spatial distributions of both laser fields in the focal plane where (b) a gas of atoms is located. The extent of the interaction region is set by the profile of the X-ray field, which ionizes an atom directly upon absorption of one photon. c) Acquired OAM by the photoelectron for circularly polarized optical vortices with winding numbers $m_{\rm OAM}=\pm1$ (solid lines) and $m_{\rm OAM}=\pm3$ (dashed lines). The detected energy belongs to the first sideband.}
\label{fig:fields}
\end{figure}
In the rotating wave approximation \cite{landau2013quantum}, the X-ray interaction Hamiltonian can be expressed as $\hat{H}_{\rm X}(t)=H_{\rm X}e^{-i\omega_{\rm X}t}$ with $H_{\rm X}\propto rY_{1,1}(\Omega_{\pmb{r}})$, which can be inserted into Eq.\,\eqref{eq:amplitude0}. \\
To trace the transfer of the optical OAM to the electrons, we investigate the angular momentum acquired by the photoelectron
\begin{equation}
\langle L_z\rangle = \frac{\langle\psi_{\rm SB}|\hat{L}_z|\psi_{\rm SB}\rangle}{\langle\psi_{\rm SB}|\psi_{\rm SB}\rangle},
\end{equation}
where $\hat{L}_z=-i\partial_\varphi$ and the wave function belonging to a specific side band (SB) is found by the projection
\begin{equation}
\psi_{\rm SB}(\pmb{r},t)=\int_{\rm SB}{\rm d}\pmb{p}\,\mathcal{A}_{\pmb{p}} \Psi_{\pmb{p}}^{(-)}(\pmb{r})e^{-iE_pt}.
\end{equation}
Here, the integration is performed (numerically) around the $n$th sideband's energy $E_n=\hbar(\omega_{\rm X} + n\omega_{\rm L})+E_i$, i.e. $E_p\in[E_n-\epsilon,E_n+\epsilon]$ where $\epsilon$ is determined by the energy width of sideband. \\
Fig.\,\ref{fig:fields}c) shows the acquired OAM of the photoelectrons for final energies in the first sideband (i.e., one IR vortex photon is absorbed) depending on the IR laser field's winding number and on the atom's distance $\rho_0$ to the optical axis. Let us inspect the case $m=1$: At $\rho_0\rightarrow0$, the OAM transfer converges against $(m^{(a)}+\sigma_{\rm L}+\sigma_{\rm X})\hbar$, meaning that in the parallel case ($m^+,\sigma_{\rm L}=+1$) the vortex field boosts the angular momentum of the photoelectron. This can be explained by the modified selection rules \cite{picon2010photoionization}, i.e., $|1s^2\rangle\xrightarrow{\rm X}|Y_{11}\rangle\xrightarrow{\rm L}|Y_{33}\rangle$ by absorption of one photon from each the X and the L laser fields. For the antiparallel case ($m^-,\sigma_{\rm L}=+1$), $|1s^2\rangle\xrightarrow{\rm X}|Y_{11}\rangle\xrightarrow{\rm L}\alpha|Y_{31}\rangle+\beta|Y_{11}\rangle$ indicating that no total angular momentum is transferred to the photoelectron.\\
Increasing the axial distance $\rho_0$, we verify that $\langle L_z\rangle$ converges to $(\sigma_{\rm L}+\sigma_{\rm X})\hbar=2\hbar$, i.e. the L laser field s locally homogeneous and circularly polarized (vortex' helicity $\sigma_{\rm L}$ is spatially independent).\\
We note the different "decay rates" of the OAM transfer, which can be traced back to the spatial components of the vector potential: In the \emph{antiparallel} case, the decay is slower due to the emergence of the longitudinal component. However, OAM transfer in the \emph{parallel} case
is mediated by the first derivative of $\pmb{A}_{\rm OV}^{m^+,\sigma_{\rm L}=+1}$, whose effective distance is limited. In Fig.\,\ref{fig:fields}c), we demonstrate that the decay rates can be slowed down by increasing the topological charge as presented by the dashed curves. We recall however that the behavior of the curves for $m>1$ and $\rho_0\rightarrow0$ is not correctly described by our theory in the full range, because the theory accounts up to the first derivative in Eq.\,\eqref{eq:Aloc}. For $m>1$, higher derivatives of order $m$ are necessary for a correct description (the transverse component scales like $\rho^m$). In those cases however, the ionization probabilities at the origin are negligibly small, which justifies our restriction to the first order in the series expansion of $\pmb{A}_{\rm OV}$.
\subsubsection{OAM-dependent dichrosim}
The theory presented so far proves the OAM transfer in vortex laser-assisted photoionization. On the other hand, in a photoionisation experiment typically differential cross sections (DCS) $\propto\mathcal{W}(\pmb{p})$ are measured.
\begin{figure}[t!]
\centering
\includegraphics[width=0.97\columnwidth]{Fig3}
\caption{Dichroism in the photoionization probability corresponding to circularly polarized optical vortices. a) Dichroism dependence on the width $w_{\rm X}$ of the ionizing X-ray field spot for $m=1$. b) the same as a) for $m=3$. c) Angular-resolved, averaged (over X-ray spot) DCS and dichroism for $w=1\,\mu$m and $m=1$. d) the same as c) for $m=3$. }
\label{fig3}
\end{figure}
The question is then whether the laser-matter OAM transfer may show up in angular or energy-resolved ionization probability. To quantify the answer, two different measurements with fixed SAM state $\sigma_{\rm L}$ and pulse parameters ($A_0, w_{\rm L}$) are mandatory. On this basis, we define the orbital dichroism $\mathcal{D}^{\sigma_{\rm L}}_{m}$ as the normalized difference in the (measured) photoelectron yields using two OAM-carrying lasers that have oppositely directed orbital chiralities, while the polarization state is fixed. More precisely, we define:
\begin{equation}
\mathcal{D}^{\sigma_{\rm L}}_{m}=\frac{ \mathcal{W}_{m^+}^{\sigma_{\rm L}}(\pmb{p}) - \mathcal{W}_{m^-}^{\sigma_{\rm L}}(\pmb{p}) }{ \mathcal{W}_{m^+}^{\sigma_{\rm L}}(\pmb{p}) + \mathcal{W}_{m^-}^{\sigma_{\rm L}}(\pmb{p}) }.
\end{equation}
Here, the ionization probabilities are statistically averaged over a macroscopic distribution of atoms (gas sample)
\begin{equation}
\mathcal{W}^{\sigma_{\rm L}}_{m}(\pmb{p})=2\pi\int_0^\infty{\rm d}\rho_0\,\rho_0\left|\mathcal{A}_{\pmb{p}}^{m,\sigma_{\rm L}}(\rho_0)\right|^2.
\end{equation}
In Fig.\,\ref{fig3}, we present the results for the dichroism in the laser-assisted photoionization by circularly polarized optical vortices ($\sigma_{\rm L}=+1$). We concentrate here on the first sideband in the electron spectra, which corresponds to the absorption of one photon from the vortex laser L yielding a continuum-continuum transition.
Panel a) shows the dichroism depending on the size of the interaction region (determined by the width $w_{\rm X}$ of the ionizing laser (X) field for different (asymptotic) directions $\vartheta_{\pmb{p}}$ of the photoelectrons in the case of $m=1$. The limited (spatial) range of the dichroism is ubiquitous: similar to the OAM transfer mediated by the modified Volkov phases, we observe a fast decay of the dichroism in all directions by increasing the effective interaction region. The dichroism represents the different actions of the transversal and longitudinal field components: the ionization probability $\mathcal{W}_{m^+}^{\sigma_{\rm L}=+1}(\pmb{p})$ belongs to the vortex in the parallel class (cf.\,Eq.\,\eqref{eq:parallel1}) while
$\mathcal{W}_{m^-}^{\sigma_{\rm L}=+1}(\pmb{p})$ is associated with the antiparallel class, where the longitudinal component dominates the dynamics near the optical axis. Therefore, it is not surprising that we find a \emph{positive} dichroism in the transversal plane, which we can trace back to the transversal field component of the parallel class vector potential $\pmb{A}_{\rm OV}^{m^+,\sigma_{\rm L}=+1}$. Detecting, however, the photoelectron more in the direction of the light propagation axis, i.e., in the vicinity of $\vartheta_{\pmb{p}}=0$, results in a \emph{negative} dichroism, which we attribute to the action of the longitudinal component present in the antiparallel vector potential $\pmb{A}_{\rm OV}^{m^-,\sigma_{\rm L}=+1}$. Hence, similar trends as in the OAM transfer can be inferred: the angular dependence of the dichroism can be related to a smooth transition between the short-ranged effect of the transversal component (around the transverse plane) and the long-ranged effect of the longitudinal component. \\
Fig.\,\ref{fig3}c) shows the angular resolved DCS and the corresponding dichroism for an interaction region $w=1\,\mu$m. Both probabilities peak in the transverse plane, which is usual for laser-assisted photoionization.The dichroism is positive around $\vartheta_{\pmb{p}}=\pi/2$ changing sign rapidly when the photoelectrons emerge near the optical axis.\\
Figures.\,\ref{fig3}b-d) present the same results for a higher winding number, i.e., $m=3$. Increasing the vortex' carried orbital momentum increases the dichroism and the range, which is particularly apparent when comparing the blue curves, belonging to $\vartheta_{\pmb{p}}=\pi/2$, between Figs\,\ref{fig3}a) and \ref{fig3}b). Furthermore, the domain where $\mathcal{D}_{m=3}^{\sigma_{\rm L}=+1}>0$ is increased. This is in line with our observation of the acquired angular momentum, as highlighted in Fig.\,\ref{fig:fields}a) and by the inspection of Eqs.\,\eqref{eq:OV1}-\eqref{eq:OV2}. The terms representing the impact of the OAM are proportional to $m/\rho_0$ so that increasing the topological charge enhances the effect.\\
The results so far underline that the OAM transfer can be linked to the different behaviors of the corresponding photoionization probabilities, particularly when compared to the photoelectron yield in the transverse plane.
Increasing the photoelectron's angular momentum by the absorption of a vortex photon with a suitable OAM direction (i.e., parallel to the photon's helicity) results in an enlarged photoionization probability in the transverse plane. The absorption of a vortex photon carrying \emph{antiparallel} OAM (relative the helicity) decreases the cross-section in transverse direction giving rise to a dichroism.
\subsubsection{Coherence and thermal average}
\label{sub:coh}
As schematically depicted in Fig.\ref{fig:fields}b), the atoms are stochastically distributed in the laser spot (red ring in Fig.\ref{fig:fields}b) and have an extension way below the optical wavelength. The spatial phase of the OAM-carrying laser is related to the spatial angular coherence of the laser wavefront. Any spatial fluctuations of the laser phase blurs the value of the carried OAM. So how comes that a phase of a classical field defined on such a length scale can be imprinted on thermally distributed, extremely localized electronic quantum state \cite{de2020photoelectric}. In principle, we may pose the same question regarding the sensitivity of photoelectrons to circular polarization of a homogeneous field, for such a polarization is nothing but the coherent oscillation (in time) of two independent (but equal in strength) transverse components which are phase shifted by $\pm\pi/2$. This phase shift is everywhere the same. Each of the independent and thermally distributed atoms reacts hence locally in the same way to this phase shift and therefore the thermal average does not affect the circular dichroism. This same argument applies to the spatial phase of the laser. What is constant here (for OAM carrying fields for instance) is the angular gradient of the vector potential (which is proportional to $m$). The additional caveat however is that the radial distribution is not homogeneous (in contrast to the case of a circular polarization) and $m$ is defined with respect to the optical axis. On the other hand, right on the optical axis the intensity is very low or vanishing so that the light-matter interaction is very weak. Substantially away from the optical axis the atoms are insensitive to $m$ \cite{PhysRevA.86.063812}. Thus, contributions to the orbital dichroism in $m$ stem from a narrow ring around the optical axis (whose radius for weak fields is discussed in \cite{PhysRevA.86.063812}) where the OAM-transfer is independent of the atom's angular position and hence unaffected by thermal averaging. In fact, a denser gas cell is more favorable for an experimental observation (note the absorption-emission time is instantaneous on the scale of the thermal atomic motion).
\begin{figure}[t!]
\centering
\includegraphics[width=0.95\columnwidth]{Fig_repr_RVB}
\caption{Schematic representation of laser-assisted photoionization setup with a tightly focussed RVB. A focused $z$-linearly polarized XUV field ionizes a valence shell electron into the external IR RVB. Depending on the region from which the electronb wave packet is launched, the photoelectron experiences the distribution of the longitudinal and the transversal field components of the vector beam. The double-headed arrows represent the local polarization directions of the RVB fields.}
\label{Fig_repr_RVB}
\end{figure}
\subsection{Steering and momentum-texturing of electronic wave packet via radially polarized vector beams}
\begin{figure}[t!]
\centering
\includegraphics[width=0.95\columnwidth]{Fig_RVB}
\caption{Laser-modified photoelectron wavepacket motion in a tightly focussed RVB. a) The dependence of the longitudinal and transversal field components on the axial distance $\rho_0$. b-c) Photoionization spectra in the dependence on the final momenta $p_x$ and $p_z$ for two different launching positions within the laser spot. $p_0$ belongs to the sole absorptions of an XUV photon with $\hbar\omega_{\rm X}=30$\,eV (zeroth side band). Higher side bands belong to absorption/emission of several IR photons. d) Photoionization probability for a fixed asymptotic direction within the first side band (SB I) as a function of the extent of the laser spot (given by the waist $w_{\rm L}=1/q_\perp$).}
\label{figRVB}
\end{figure}
By focusing a radially polarized vector beam we can realize a transition from a dominating longitudinal component in the vicinity of the optical axis to a transversal component for outer radii \cite{watzel2020multipolar}. This transition occurs on the sub-wavelength scale. An example of the individual field components is shown in Fig.\,\ref{figRVB}a).\\
Depending on the atom's position within the laser spot, the liberated photoelectron is exposed to an external laser field with different spatial components and varying (local) amplitude, as presented schematically by Fig.\,\ref{Fig_repr_RVB}. For a demonstration, we picked up two positions, which are represented by 1 and 2 in Fig.\,\ref{figRVB}. By absorbing IR photons of the assisting vector beam, the photoelectron wave packet is accelerated, which is visible in the (measured) momentum distribution. Moreover, the direction of the acceleration is crucially dependent on which field component is locally dominating.\\
In the following example, the ionizing XUV field is $z$-linearly polarized with $\hbar\omega_{\rm X}=30$\,eV, while the focussing of the assisting RVB is determined by $\tan\alpha=q_\perp/q_z$ ($\alpha=28^\circ$). For this condition, the IR laser spot size's diameter in the focal plane is around the wavelength $\lambda_{\rm L}$. A rather high amplitude of $A_0=0.15$\,a.u. together with an infinitely long pulse lengths enable the emergence of higher side bands \cite{kazansky2010angle}, as highlighted by Figs.\,\ref{figRVB}b-c. A photoelectron liberated from an atom located in the vicinity of the optical axis is exposed to the strong longitudinal component of the external IR vector beam, and - upon absorption of IR photons, it is accelerated into the propagation direction. Moreover, boosting the photoelectron wave packet's kinetic energy results in a pronounced concentration of ionization probability around $\vartheta_{\pmb{p}}=0$. As mentioned before, the relative strengths between the longitudinal and the transversal field components of the tightly focused RVB vary very quickly on a sub-wavelength scale. Hence, for an atom located further away from the optical axis, the measured photoelectron's characteristics changes strongly. If the electron wave packet is ejected into the $z$-direction by the XUV field, the transversal component forces the photoelectron into a transverse trajectory, as illustrated in Fig.\,\ref{figRVB}c, where higher side bands corresponding to the exchange of several IR photons are visible. The spectrum reveals that the center of the respective angular-dependent ionization probability wanders more and more towards the vicinity of $\vartheta_{\pmb{p}}=\pi/2$. \\
Therefore, the direction of the photoelectron wave packet's acceleration can be manipulated by the vector beam components via focusing: Keeping all the laser parameters unchanged, the photoionization probability exhibits a strong dependence on the laser spot size determined by $w_{\rm L}=1/q_\perp$. This connection is demonstrated in Fig.\,\ref{figRVB}d), where the ionization probability for the first side band (SB I) is presented for two fixed asymptotic directions and positions of the atoms. Interestingly, for a tightly focused RVB, $\mathcal{W}(p_1,\vartheta_{\pmb{p}}=0)$ (belonging to 1) related to the longitudinal component is larger than the one (position 2) related to the transverse component. Broadening the beam waist changes the situation: While both probabilities decrease with increasing $w_{\rm L}$, the individual decay rates are different. As a consequence and as a general rule, for larger extents of the focal spots, interaction with the assisting vector beam is much more probable via the transversal field component. For weakly focused RVBs, with a waist in the range of 10 microns, the acceleration due to the longitudinal component is practically not present.
\subsubsection{Coherence and role of atom spatial distribution}
In Sec.\,\ref{sub:coh}, we argued why the OAM transfer to a stochastic atom distribution is not washed away by configurational averaging. For AVB and RVB the situation has some subtleties. For a strongly focused RVB, the longitudinal component is dominant and its action resembles the case of a linear polarized field (along the propagation direction). Therefore, the response of the atom distribution in the focus is linearly related to the response of a single atom (we suppressed so far the well-established propagation and phase-matching issues related to a finite length (along $z$ direction) of the sample).
For de-focused vector beams, the intensity distribution is similar to vortex beams. Let us for concreteness consider AVB, as given by Eq.~\ref{eq:avb}. Alternatively, we can also write for this beam $\pmb{A}_{\rm AVB}(\pmb{r},t) = i/(2\sigma_{\rm L})\; A_0q_\perp\rho( \hat{e}_{-\sigma_{\rm L}} e^{i\sigma_{\rm L}\varphi} - \hat{e}_{\sigma_{\rm L}} e^{-i\sigma_{\rm L}\varphi})\sin(q_zz-\omega_{\rm L}t)$. The coherence in the time oscillations of the transverse field components are reflected by the value of $\hat{e}_{\sigma_{\rm L}}$. In AVB (or RVB) $\hat{e}_{\sigma_{\rm L}}$ is everywhere the same except for a fixed spatial rotation angle as $\varphi$ evolves (signified by $e^{\pm i\sigma_{\rm L}\varphi}$ in $\pmb{A}_{\rm AVB}(\pmb{r},t) $). Therefore, the response of a strongly inhomogeneous atomic distribution will be different from that of a statistically distributed one. Indeed, this fact is reflected in the non-trivial dependence of the SL-VW phases in Eqs.\,\eqref{eq:AVB} and \eqref{eq:RVB} on the atom angular position $\varphi_0$ from which the electron is launched. Clearly, one can reverse the argument and retrieves from the photoelectron distributions information on the spatial structure of the atom distribution in the vector beam laser spot, on a scale below the optical wavelength. Similar arguments also apply to optical skyrmions.
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{Fig4}
\caption{Photoelectrons traversing an optical skyrmion. a) Radial component of the vector potential in the focal plane. b) Setup: An XUV radial vector beam with variable focusing liberates an photoelectron at its intensity maximum with high probability, where the local spatio-temporal IR optical skyrmion is present. (1) and (2) present two different XUV focusing. c) Streaking spectrum for a photoelectron detected in the asymptotic direction $\vartheta_{\pmb{p}}=\pi/2$ and $\varphi_{\pmb{p}}=0$. The center of energy (COE) represents the maximum of the photoionization probability depending on the delay time $\Delta t$.}
\label{fig4}
\end{figure}
\subsection{Reconstruction of an optical propagating skyrmionic field via attosecond streaking}
Let us consider as an example the optical skyrmion $\pmb{A}_{\rm OS}^{m_1=3,m_2=-1}(\pmb{r},t)$ ($\alpha=7$, $\beta=1$) with a moderate focusing of $w_L=7\,\mu m$. The corresponding radial component of the vector potential is shown in Fig.\,\ref{fig4}a), revealing a strong azimuthal variation. Note that the optical skyrmion is shown only in the range where our approximation of the spatial distribution function $F_m(\rho)$ is valid. Our goal is to sample the local structure of the skyrmionic field via a traversing electronic wave packet. Generally, one may use the attosecond streaking technique \cite{goulielmakis2004direct, krausz2009attosecond, pazourek2015attosecond} which has been established as a key element of attosecond spectroscopy. If the XUV laser field is a short pulse, its large bandwidth allows for several quantum paths, involving absorption and emission of several IR photons, to a final energy state.
\begin{figure}[b!]
\centering
\includegraphics[width=0.95\columnwidth]{Fig4a}
\caption{Schematic representation of the light field - detector geometry. Photoelectrons are primarily emitted in the focal plane due to the radially polarized XUV field and are streaked by the local IR optical vortex which is characterized by a varying polarization and phase landscape. The Photoelectrons are measured in directions $\{\vartheta_{\pmb{p}}=\pi/2,\varphi_{\pmb{p}}\}$.}
\label{fig4a}
\end{figure}
Interference between those quantum paths results in a modulation of the final momentum, which depends on the temporal delay $\Delta t$ between both pulses (in our case, $\Delta t$ refers to the temporal difference between the maxima of both pulses) \cite{dahlstrom2013theory}. Classically, the detected momentum of the photoelectron (in a specific asymptotic direction $\Omega_{\pmb{p}}=\left\{\vartheta_{\pmb{p}},\varphi_{\pmb{p}}\right\}$) follows $p(\Delta t)\approx\sqrt{2(\omega_{\rm X}+E_i)}-\tilde{A}_{\rm L}(\Delta t)$, where $\tilde{A}_{\rm L}$ is the projection of the vector potential in the direction of the (measured) asymptotic momentum. For low amplitudes of the external laser (L) field, the above-mentioned classical relationship is an excellent approximation.\\
The temporal relation between the photoelectron's final momentum and the vector field's amplitude at the moment of ionization allows for the imaging of the optical skyrmion using the photoemission streaking measurements.
If the photoemission spectra are recorded in every asymptotic direction $\varphi_{\pmb{p}}$ in the focal plane ($\vartheta_{\pmb{p}}=\pi/2$), a phase shift should be visible since the streaking field (the optical skyrmion) has an internal phase structure along the azimuthal direction. In other words, for fixed delay times $\Delta t$, the final asymptotic momentum in the presence of both laser fields becomes explicitly directionally dependent, reflecting the unique phase structure of the optical skyrmion, as shown in Fig.\,\ref{fig4}a). Moreover, the phase variation of the structured field depends on the axial distance.
\begin{figure}[t!]
\centering
\includegraphics[width=0.95\columnwidth]{Fig5}
\caption{Reconstructed information retrieved from the photoemission spectra. a-b) Directional dependent COEs for two different focusing setups of the ionizing XUV RVB revealing the temporal information about the local field of the IR optical skyrmion. c-d) Reconstructed phase variation for a fixed delay time $\Delta t$. The curves follow from an interpolation of the data points with periodic boundary conditions.}
\label{fig5}
\end{figure}
To deal with this feature, one may use a radially polarized vector beam \cite{hernandez2017extreme}, as shown schematically in Fig.\,\ref{fig4}b). By adjusting the focus of the donut-shaped intensity distribution, we can select atoms within the laser spot that are photoionized in the radial direction in the transverse plane (with respect to the optical axis). Only atoms near the intensity maximum of the XUV field have a sizable ionization probability via one-photon processes. Once liberated, the photoelectrons are affected (streaked) by the local field of the IR optical skyrmion. Measuring now the photoelectron in the asymptotic direction $\left\{\vartheta_{\pmb{p}}=\pi/2, \varphi_{\pmb{p}}\right\}$ exploits the radial field component $A^{+3,-1}_{\rm OS}\cdot\hat{e}_\rho$ of the optical skyrmion in the photoemission spectrum. A schematic representation is given in Fig.\,\ref{fig4a}.\\
Mathematically, we gain access to the streaking spectrum by numerical integration of Eq.\,\eqref{eq:amplitude0} and by calculating the modified Volkov phases for $A^{+3,-1}_{\rm OS}(\pmb{r},t)$, given in Eq.\,\eqref{eq:OSphase}. In addition, we introduce the delay time $\Delta t$ in the Hamiltonian describing the interaction between the XUV-photon and an atom located at the axial distance $\rho_0$ within the laser spot:
\begin{equation}
\hat{H}_{\rm X}(\rho_0,t-\Delta t)=\rho\mathcal{E}_{\rm RVB}(\rho_0)f(t-\Delta t)\cos[\omega_{\rm X}(t-\Delta t)].
\end{equation}
Here, $\mathcal{E}_{\rm RVB}(\rho_0)=E_0(\rho_0/w_{\rm X})\exp(-\rho_0^2/w_{\rm X}^2)$ is the radial distribution function of the vector beam, where $w_{\rm X}$ determines the focusing. The action of the longitudinal component can be neglected since the XUV beam is weakly focused [cf.\,Fig.\,\ref{fig4}b)]. The temporal envelope of the short pulse is given by $f(t)=\cos(\omega_{\rm X}t/(2n))^2$ for $t\in[-n\pi/\omega_{\rm X},n\pi/\omega_{\rm X}]$ (and zero otherwise). In our simulations, we chose $\hbar\omega_{\rm X}=60$\,eV and a number of $n=7$ optical cycles, which means we are in the streaking regime \cite{kazansky2010angle}. For the setups shown in Fig.\,\ref{fig4}b), $w_{\rm X}=2.4\,\mu$m (case 1) and $w_{\rm X}=5.3\,\mu$m (case 2) were chosen. The fixed peak intensity (at the maximal field amplitude) of $2\times10^{14}$\,W/cm$^2$ ensures that the single photon processes are initiated by the radially polarized XUV field.\\
In Figs.\,\ref{fig5}a-b), we present the center of energy (COE), extracted from the corresponding streaking spectra, for different asymptotic directions $\varphi_{\pmb{p}}$ and two different focusing setups of the ionizing XUV pulse. Due to the spatial extent of the XUV field, we can "scan" the IR field characteristics in the radial direction, whereas the azimuthal dependencies can be retrieved from rotating the photoelectron detector [cf.\,Fig.\,\ref{fig4a}]. By varying $\varphi_{\pmb{p}}$, the COEs reveals modulation both in the amplitude and on the time axis. This is, in particular, evident in the second focusing setup presented in Fig.\,\ref{fig5}b). A further important aspect is the $2\pi$-periodicity revealing the imprinting of the optical field's phase onto the photoelectron distributions. By fixing the time delay $\Delta t$, we can reconstruct the phase information of the optical skyrmion, as shown in Figs.\,\ref{fig5}c-d). As anticipated, the phase structure of a skyrmion is very involved due to the interplay of the two contributing vortices with different orbital angular momenta. By varying the focusing of the XUV field (i.e., adjusting the waist parameter $w_{\rm X}$), the full radial field component of the optical skyrmion can be scanned. The other spatial components $A^{+3,-1}_{\rm OS}\cdot\hat{e}_\varphi$ and $A^{+3,-1}_{\rm OS}\cdot\hat{e}_z$ can retrieved from Maxwell's equations. Usually, plasmonic optical skyrmions are mapped via PEEM (photoemission electron microscopy) \cite{tsesses2018optical,Spektor1187,dai2019ultrafast,PhysRevX.9.021031}. The complementary method proposed here can also map the spatio-temporal structure of freely propagating optical skyrmions.
\section{conclusions}
The goal has been to derive quasi-analytical expressions for the electron motion in an intense laser field that exhibits a non-trivial distribution in space of the spin and/or of the wavefront. Besides, we aim at exploiting the derived expression for describing physical processes such as laser-assisted ionization of atoms, particle acceleration, or spatio-temporal mapping of topological photonic fields. The well-known Volkov wave fully describes the motion of an unbound electron in an unstructured laser field. In contrast, the structured laser field Volkov state (SL-VW) we were able to obtain only under certain approximations that can be in principle improved systematically, but such improvements were not treated here. Fortunately, for a number of physical processes of interest, the derived approximate SL-VW is valid and useful. Generally, the SL-VW is fundamentally different from the conventional Volkov state since it receives contributions from the spin-angular momentum spatial distribution as well as from the space structure of the vector potential. Notably, also the field scalar potential affects SL-VW, and its influence can be encapsulated in a longitudinal component of the vector potential. As a demonstration, we considered photoionization assisted by propagating optical vortices, meaning a field with a well-defined spatial phase structure (but no spin-angular momentum structure). Orbital angular momentum of the laser field is related to this phase and is found to be transferable to the photoelectrons, even for atoms that are stochastically distributed in the laser spot. The assisting laser pulse may carry no orbital angular momentum but its spin angular momentum can be structured, as for radially polarized laser pulse. In this case we found that this beam can be employed to manipulate the momentum distribution of the electronic wavepacket. Thereby, the longitudinal component of the laser's vector potential is decisive. A further application in photonics concerns topology. An electronic wavepacket traversing a topologically no-trivial optical field such as optical skyrmions samples spatio-temporal information on the skyrmionic field which enables a mapping of the phase and the spin-angular momentum texturing in the skyrmion, even if it is not in the form of a localized plasmonic field. Rich important phenomena results from applying a combination of time delayed or frequency shifted structured fields. For example, applying a RVB and a time delayed linearly polarized pulse to a torus generates a field-free toroidal moment as (excited) electronic eigenstates \cite{Watzeltoroid}. In addition to application in electronics and magnetism, future applications include understanding high harmonic emission in arbitrarily structured fields as well as field-assisted particle trapping and stabilization.
\section{Acknowledgements}
This work was supported by the DFG through SFB TRR 227 and WA 4352/2-1.
|
\section{Introduction}
Consider a (rigid) body, $\mathscr B$, translating with constant nonzero velocity, $v_\infty$, in a viscous (Navier-Stokes) liquid, $\mathscr L$, that occupies the whole space outside $\mathscr B$. Without loss of generality, we assume that $v_\infty$ is directed along the positive $x_1$-axis, namely,
$v_\infty=\lambda\e_1$ with $\lambda>0$. We also assume that $\mathscr L$ is subject to a body force and a distribution of boundary velocity, both being time-periodic of period $\mathcal T$.
Then, the time-periodic dynamics of the liquid around the body
are governed by the
following set of equations
\begin{equation}\label{sys:NavierStokesTP_ExteriorDomain}
\begin{pdeq}
\partial_tu-\Deltau-\lambda\partial_1u+u\cdot\nablau+\nabla\mathfrak{p}&=f
&&\text{in }{\mathbb T}\times\Omega, \\
\Divu&=0
&&\text{in }{\mathbb T}\times\Omega, \\
u&=u_\ast
&&\text{on }{\mathbb T}\times\partial\Omega, \\
\lim_{\snorm{x}\to\infty}u(t,x)&=0
&&\text{for } t\in{\mathbb T},
\end{pdeq}
\end{equation}
where $\Omega:=\mathbb{R}^3\setminus\mathscr B$ is the domain occupied by the liquid.
Moreover, $u\colon{\mathbb T}\times\Omega\to\mathbb{R}^3$ and $\mathfrak{p}\colon{\mathbb T}\times\Omega\to\mathbb{R}$ are velocity and pressure fields
of the liquid,
$f\colon{\mathbb T}\times\Omega\to\mathbb{R}^3$ is the external body force,
and $u_\ast\colon{\mathbb T}\times\partial\Omega\to\mathbb{R}^3$
the velocity field at the boundary.
The time-axis is given by the torus group
${\mathbb T}:=\mathbb{R}/\calt\mathbb{Z}$, which
ensures that all functions appearing in \eqref{sys:NavierStokesTP_ExteriorDomain}
are time-periodic with a prescribed period $\calt>0$.
Note that for
a body at rest, that is, for $\lambda=0$, the mathematical and physical characteristics of the flow
are very different from those for $\lambda\neq0$. For this issue, we refer the reader to the recent papers \cite{galdi:zeroaverage,galdi:transarbitrary}.
\par
Existence, uniqueness and spatial asymptotic behavior of solutions to \eqref{sys:NavierStokesTP_ExteriorDomain} have been the object of several recent researches \cite{GaldiKyed_TPflowViscLiquidpBody,GaldiKyed_TPSolNS3D_AsymptoticProfile,Eiter_SpatiallyAsymptoticStructureTPNS_2020}. In particular, under suitable assumptions on the data, these results provide sharp pointwise {\em algebraic} decays for the velocity field and its first spatial derivatives; see \cite{Eiter_SpatiallyAsymptoticStructureTPNS_2020} and Theorem \ref{thm:VelocityDecay} below. However, as suggested by physical grounds, the vorticity field $\omega:=\curl u$ is expected to decay at an {\em exponential} rate, at least outside the ``wake region" behind $\mathscr B$. It is just to this question that the present paper is devoted.
\par
More precisely, we shall study the asymptotic behavior of
the vorticity field $\curlu(x,t)$ for $\snorm{x}\to\infty$, uniformly in time.
In these regards, we recall that in the case of a steady-state flow, that is, when $\np{v,p}$ is a time-independent
solution to \eqref{sys:NavierStokesTP_ExteriorDomain},
a famous result of \textsc{Clark} \cite{Clark_VorticityAtInfinityNavierStokes_1970}
and \textsc{Babenko} and \textsc{Vasil'ev} \cite{BabenkoVasilev_AsymptoticBehaviorSteadyFlow_1973}
shows that for $\snorm{x}$ sufficiently large one has
\begin{equation}\label{est:Curlu_SteadyStateCase}
\snorm{\curlv(x)}\leq C \snorm{x}^{-3/2}\e^{-\alpha \wakefct{x}}
\end{equation}
for some constants $C,\,\alpha>0$,
where
\[
\wakefct{x}:=\snorm{x}+x_1.
\]
In particular, this reflects the anisotropic behavior of the fluid flow and
translates, in mathematical terms, the presence of a ``wake region" behind $\mathscr B$.
Estimate \eqref{est:Curlu_SteadyStateCase}
implies that the vorticity, $\curlv$, decays exponentially fast on rays
$\setc{x\in\mathbb{R}^3}{x_1=\theta\snorm{x}}$ for $\theta\in(-1,1]$,
while inside parabolic regions $\setc{x\in\mathbb{R}^3}{\wakefct{x}\leq \beta}$, $\beta>0$,
estimate \eqref{est:Curlu_SteadyStateCase} merely yields an algebraic decay rate.
Since time-independent solutions are trivially also time-periodic,
one would expect a similar behavior in the time-periodic case.
As a matter of fact, we show that this is indeed true and that
the vorticity field associated to a time-periodic flow is subject to an analogus estimate.
Actually, as proved in \cite{Eiter_SpatiallyAsymptoticStructureTPNS_2020},
if we split $u$ into
its time average $v$ and a purely periodic part $w:=u-v$,
then the decay rates of $v$ and $\nablav$ are much slower than those of $w$ and $\nablaw$.
Thus, also in the problem at hand, it seems reasonable to derive separate pointwise estimates for the two parts
$\curlv$ and $\curlw$ of the vorticity $\curlu$.
In doing so, we are indeed able to show that
the time-independent part $v$ satisfies \eqref{est:Curlu_SteadyStateCase}
whereas the other part obeys the estimate
\begin{equation}\label{est:Curlu_PurelyPeriodic}
\snorm{\curlw(t,x)}\leq C \snorm{x}^{-9/2}\e^{-\alpha \wakefct{x}}
\end{equation}
for all sufficiently large $\snorm{x}$,
and therefore decays faster.
It is worth emphasizing that we establish this result for {\em any} weak solution to \eqref{sys:NavierStokesTP_ExteriorDomain} (see Definition \ref{def:WeakSolution_NStp}), whose purely periodic part {\em only} satisfies the Serrin-like condition \eqref{el:VelocityAdditionalIntegrability}, provided the data are sufficiently smooth with $f$ of bounded spatial support; see Theorem \ref{thm:VorticityDecay}.
A main tool in our approach is the introduction of a time-periodic fundamental solution
associated to the vorticity field $\curlu$.
The concept of time-periodic fundamental solutions in the field of fluid dynamics is new
and was recently introduced by \textsc{Kyed} \cite{Kyed_FundsolTPStokes2016}
and \textsc{Galdi} and \textsc{Kyed} \cite{GaldiKyed_TPSolNS3D_AsymptoticProfile}
in the case of a three-dimensional Navier--Stokes flow,
and further extended by \textsc{Eiter} and \textsc{Kyed} \cite{EiterKyed_etpfslns}
to the general $n$-dimensional case.
The fundamental solution $\varGamma^\lambda$ introduced there consists of the fundamental solution $\fundsolvel_0$
to the steady-state problem and a second so-called purely periodic part $\fundsolvel_\perp$.
Analogously, we define the time-periodic vorticity fundamental solution $\phi^\lambda$
as the sum
of the corresponding steady-state fundamental solution $\fundsolvort_0$
and a purely periodic part $\fundsolvort_\perp$.
After introducing these time-periodic fundamental solutions,
our procedure parallels that of \cite{DeuringGaldi_ExpDecayVorticity_2016},
where \textsc{Deuring} and \textsc{Galdi} studied the
vorticity field associated to the steady-state flow past a rotating body.
Note that this problem is directly related to
the one investigated here
since a time-independent solution in the frame attached to the rotating body
corresponds to a time-periodic solution in the inertial frame.
By means of the above time-periodic fundamental solutions we deduce representation formulas
for $u$ and $\curlu$,
which enable us to express
$u$ as a fixed point of a nonlinear map $F_S$ of convolution type; see eq. \eqref{eq:uFixedPoint_VortDecay}.
We then
establish the existence of a fixed point $z=F_S(z)$ of this map
in a class of functions such that $\curlz$ decays in the expected way; see Corollary \ref{cor:ExistenceFixedPoint} .
Successively, we show that this fixed point is, in fact, unique in the {\em larger} class of functions
that merely satisfy the pointwise estimates of $u$ and $\nablau$ established in
\cite{Eiter_SpatiallyAsymptoticStructureTPNS_2020}; see Theorem \ref{thm:VorticityDecay_WholeSpace}.
Since $u$ is a fixed point of $F_S$ by construction,
we thus conclude $u=z$ and that $u=v+w$ satisfies
\eqref{est:Curlu_SteadyStateCase} and \eqref{est:Curlu_PurelyPeriodic}.
Observe that, in order to employ the contraction mapping principle,
the existence of the fixed point $z$ is established
in a class of functions that satisfy a slightly weaker estimate
than that given in \eqref{est:Curlu_PurelyPeriodic}.
However, by another application of the representation formulas
via the vorticity fundamental solution, we finally
obtain the asserted decay rates \eqref{est:Curlu_SteadyStateCase} and \eqref{est:Curlu_PurelyPeriodic}.
The result just described is proved in the case where $\Omega$ is the whole space $\mathbb{R}^3$. However, we show that it can be readily
transferred to the case of an exterior domain
by a classical cut-off argument, provided $u_*$ and $f$ are sufficiently smooth, with $u_*$ having zero total net flux at $\partial\Omega$. We leave it as an open question whether this condition can indeed be removed.
Finally, we observe that some of the intermediate results are contained in the first author's PhD thesis \cite{Eiter_Diss}. However, they were derived
under the stringent assumption that both external force $f$ and solution $u$ are of class $\CR \infty$. In contrast, here we merely require summability assumptions on
$f$ and $u$ (see \eqref{cond:forcing} and \eqref{el:VelocityAdditionalIntegrability}) which
represents a rather significant improvement
The paper is structured as follows.
After introducing the basic notation in Section \ref{sec:Notation},
we present our main result on the decay of the vorticity field in Section \ref{sec:MainResults}.
In Section \ref{sec:TPFundSol} we recall the notion of a time-periodic fundamental solution to the Navier--Stokes equations
and introduce the concept of a time-periodic vorticity fundamental solution.
Section \ref{sec:Regularity} is dedicated to the study of regularity
of weak solutions to the time-periodic Navier--Stokes problem.
The introduced fundamental solutions
are employed in Section \ref{sec:FixedPointProblem}
in order to conclude a suitable fixed-point equation.
After the derivation of appropriate estimates for the terms in this equation in Section \ref{sec:Estimates},
we finish the proof of the main result in Section \ref{sec:ConclusionProof}.
\section{Notation}
\label{sec:Notation}
Points in ${\mathbb T}\times\Omega$ for $\Omega\subset\mathbb{R}^3$ are usually denoted by $(t,x)$
and consist of a time variable $t\in{\mathbb T}$
and a spatial variable $x\in\Omega$.
For a sufficiently regular function $u\colon{\mathbb T}\times\Omega\to\mathbb{R}^3$
we write $\partial_ju:=\partial_{x_j}u$,
and we set $\Deltau:=\partial_j\partial_ju$
and $\Divu:=\partial_ju_j$.
Here we employ Einstein's summation convention,
which we do frequently in the following.
By $\delta_{jk}$ and $\varepsilon_{jk\ell}$ we denote
the Kronecker delta and the Levi-Civita symbol, respectively.
For $R>0$ and $x\in\mathbb{R}^3$ we set $\mathrm{B}_R(x):=\setc{y\in\mathbb{R}^3}{\snorm{x-y}<R}$ and
$\mathrm{B}^R(x):=\setc{y\in\mathbb{R}^3}{\snorm{x-y}>R}$,
and in the case $x=0$ we write $\mathrm{B}_R:=\mathrm{B}_R(0)$ and $\mathrm{B}^R:=\mathrm{B}^R(0)$.
Moreover, for $R>r>0$ we set $\mathrm{B}_{r,R}:=\mathrm{B}_R\cap\mathrm{B}^r$.
For vectors $a,b\in\mathbb{R}^3$ their vector product $a\wedge b$ and their tensor product $a\otimes b$
are given by $(a\wedge b)_{j}=\varepsilon_{jk\ell}a_k b_\ell$ and
$(a\otimes b)_{jk}=a_jb_k$, respectively.
Moreover, we call a subset $\Omega\subset\mathbb{R}^3$ an exterior domain,
if it is the complement of a non-empty compact subset of $\mathbb{R}^3$.
Without loss of generality, we always assume that $0$ is contained
in the interior of $\mathbb{R}^3\setminus\Omega$.
In order to include the time periodicity in the formulation
of the Navier--Stokes equations \eqref{sys:NavierStokesTP_ExteriorDomain},
we formulated the system
on ${\mathbb T}\times\Omega$.
In the case $\Omega=\mathbb{R}^3$, which plays a prominent role in our approach,
the time-space domain is given by the locally compact Abelian group $G:={\mathbb T}\times\mathbb{R}^3$.
The dual group of $G$ can be identified with $\widehat{G}=\mathbb{Z}\times\mathbb{R}^3$,
the elements of which we denote by $(k,\xi)\in\mathbb{Z}\times\mathbb{R}^3$.
We equip the group ${\mathbb T}$ with the normalized Haar measure given by
\[
\forall f\in\CR{}({\mathbb T}):\qquad
\int_{\mathbb T} f(t)\,{\mathrm d}t=\frac{1}{\calt}\int_0^\calt f(t)\,{\mathrm d}t,
\]
the group $\mathbb{Z}$ with the counting measure,
and $G$ and $\widehat{G}$ with the corresponding product measures.
The Fourier transform $\mathscr{F}_G$ on $G$
and its inverse $\mathscr{F}^{-1}_G$ are formally given by
\[
\begin{aligned}
\mathscr{F}_G\nb{f}(k,\xi)
&:=\int_{\mathbb T}\int_{\mathbb{R}^3} f(t,x)\e^{-i\frac{2\pi}{\per} kt-ix\cdot\xi}\,{\mathrm d}x{\mathrm d}t,\\
\mathscr{F}^{-1}_G\nb{f}(t,x)
&:=\sum_{k\in\mathbb{Z}}\int_{\mathbb{R}^3} f(k,\xi)\e^{i\frac{2\pi}{\per} kt+ix\cdot\xi}\,{\mathrm d}\xi.\\
\end{aligned}
\]
This defines an isomorphism $\mathscr{F}_G\colon\mathscr{S}(G)\to\mathscr{S}(\widehat{G})$ with inverse $\mathscr{F}^{-1}_G$,
provided that the Lebesgue measure ${\mathrm d}\xi$ is normalized appropriately.
Here $\mathscr{S}(G)$ is the so-called Schwartz--Bruhat space, which is a generalization
of the classical Schwartz space in the Euclidean setting;
see \cite{Bruhat61,EiterKyed_tplinNS_PiFbook}.
By duality, this yields an isomorphism $\mathscr{F}_G\colon\mathscr{S^\prime}(G)\to\mathscr{S^\prime}(\widehat{G})$
between the corresponding dual spaces $\mathscr{S^\prime}(G)$ and $\mathscr{S^\prime}(\widehat{G})$,
the spaces of tempered distributions.
For an open set $\Omega\subset\mathbb{R}^3$ or $\Omega\subset{\mathbb T}\times\mathbb{R}^3$ and $q\in[1,\infty]$, $m\in\mathbb{N}$,
we denote the classical Lebesgue and Sobolev spaces by $\LR{q}(\Omega)$
and $\WSR{m}{q}(\Omega)$, respectively.
Moreover, $\LRloc{1}(\Omega)$ is the set of all locally integrable functions,
and $\WSRloc{1}{1}(\Omega)$ is the subset of $\LRloc{1}(\Omega)$ with
locally integrable weak derivatives.
For an open subset $\Omega\subset\mathbb{R}^3$,
homogeneous Sobolev spaces are denoted by
\[
\DSR{m}{q}(\Omega)
:=\setcl{u\in\LRloc{1}(\Omega)}{\nabla^m u\in\LR{q}(\Omega)},
\]
where $\nabla^mu$ denotes the collection of all $m$-th
weak derivatives of $u$.
We further set
\[
\CR{\infty}_{0,\sigma}(\Omega):=\setc{\varphi\in\CR \infty_0(\Omega)^3}{\Div\varphi=0},
\]
where $\CR \infty_0(\Omega)$ is the class of
compactly supported smooth functions on $\Omega$.
For $q\in[1,\infty]$ and a (semi-)normed vector space $X$,
$\LR{q}({\mathbb T};X)$ denotes the corresponding Bochner--Lebesgue space on ${\mathbb T}$,
and
\[
\WSR{1,2}{q}({\mathbb T}\times\Omega)
:=
\setcl{u\in\LR{q}({\mathbb T};\WSR{2}{q}(\Omega))}
{\partial_tu\in\LR{q}({\mathbb T}\times\Omega)}.
\]
We further define the projections
\[
\calp f (x):= \int_{\mathbb T} f(t,x)\,{\mathrm d}t,
\qquad \calp_\bot f:= f-\calp f,
\]
which decompose $f\in\LRloc{1}\np{{\mathbb T}\times\Omega}$
into a time-independent \emph{steady-state} part $\calp f$
and a remainder \emph{purely periodic} part $\calp_\bot f$.
One readily sees that $\calp$ and $\calp_\bot$ are bounded operators on $\LR{q}({\mathbb T}\times\Omega)$ for all $q\in[1,\infty]$
and that
\[
\calp f = \mathscr{F}^{-1}_G \bb{\delta_\mathbb{Z}(k) \mathscr{F}_G\nb{f}},
\qquad
\calp_\bot f = \mathscr{F}^{-1}_G \bb{\np{1-\delta_\mathbb{Z}(k)} \mathscr{F}_G\nb{f}},
\]
where $\delta_\mathbb{Z}$ is the delta distribution on $\mathbb{Z}$.
The letter $C$ always denotes a generic positive constant,
the value of which may change from line to line.
When we want to specify the dependence of the constant $C$ on quantities $a, b, \ldots$,
we write $C(a, b, \ldots)$.
\section{Main result}
\label{sec:MainResults}
As emphasized earlier on, our focus
is the pointwise estimates of the vorticity field $\curl u$
associated to a solution $\np{u,\mathfrak{p}}$ of
\eqref{sys:NavierStokesTP_ExteriorDomain}.
More precisely, we study the vorticity field of weak solutions
to \eqref{sys:NavierStokesTP_ExteriorDomain}
defined as follows.
\begin{defn}\label{def:WeakSolution_NStp}
Let $f\in\LRloc{1}({\mathbb T}\times\Omega)^3$.
A function $u\in\LRloc{1}({\mathbb T}\times\Omega)^3$
is called \emph{weak solution} to \eqref{sys:NavierStokesTP_ExteriorDomain}
if
\begin{enumerate}
\item[i.]
$\nablau\in\LR{2}({\mathbb T}\times\Omega)^{3\times3}$,
$u\in\LR{2}({\mathbb T};\LR{6}(\Omega))^3$,
$\Divu=0$ in ${\mathbb T}\times\Omega$,
$u=u_\ast$ on ${\mathbb T}\times\partial\Omega$,
\label{item:WeakSolution_NStp_L2}
\item[ii.]
$\calp_\botu\in\LR{\infty}({\mathbb T};\LR{2}(\Omega))^3$,
\label{item:WeakSolution_NStp_LInfty}
\item[iii.]
the identity
\[
\int_{{\mathbb T}\times\Omega}\bb{-u\cdot\partial_t\varphi
+\nablau:\nabla\varphi
-\lambda\partial_1u\cdot\varphi
+\np{u\cdot\nablau}\cdot\varphi}\,{\mathrm d}(t,x)
=\int_{{\mathbb T}\times\Omega} f\cdot\varphi\,{\mathrm d}(t,x)
\]
holds for all test functions $\varphi\in\CR{\infty}_{0,\sigma}({\mathbb T}\times\Omega)$.
\label{item:WeakSolution_NStp_WeakFormulation}
\end{enumerate}
\end{defn}
Let us explain the choice of the functional class for weak solutions.
When $\Omega=\mathbb{R}^3$, condition
i.~is equivalent to
$u\in\LR{2}({\mathbb T};\DSRNsigma{1}{2}(\mathbb{R}^3))$,
where $\DSRNsigma{1}{2}(\mathbb{R}^3)$ is the closure
of $\CR{\infty}_{0,\sigma}(\mathbb{R}^3)$ with respect to the homogeneous norm $\norm{\nabla\cdot}_2$.
In this case, the class of solutions from Definition \ref{def:WeakSolution_NStp}
is the same as considered in \cite{GaldiKyed_TPSolNS3D_AsymptoticProfile}
and \cite{Eiter_SpatiallyAsymptoticStructureTPNS_2020},
where the asymptotic behavior of the velocity field $u$ and its gradient $\nablau$
was investigated.
Moreover, for any $f\in\LR{2}({\mathbb T};\DSRN{-1}{2}(\mathbb{R}^3))^3$
the existence of a weak solution in the above sense
was shown by \textsc{Kyed} \cite{Kyed_habil}.
Therefore, this class of solutions is a natural candidate
for further investigation of the associated vorticity field $\curlu$.
The goal of the present article is to prove the following result.
\begin{thm}\label{thm:VorticityDecay}
Let $\Omega\subset\mathbb{R}^3$ be an exterior domain
with boundary of class $\CR{2}$,
and let $\lambda> 0$.
Let $f$ and $u_\ast$ be such that
\begin{align}
\forall q\in(1,\infty): \ f\in\LR{q}({\mathbb T}\times\Omega)^3,
&&
\supp f \text{ bounded}
\label{cond:forcing},
\\
u_\ast\in\CR{}\np{{\mathbb T};\CR{2}(\partial\Omega)}^3
\cap\CR{1}\np{{\mathbb T};(\partial\Omega)}^3,
&&
\int_{\partial\Omega}u_\ast\cdot\mathrm{n}\,{\mathrm d}S=0,
\label{cond:BoundaryData}
\end{align}
where $\mathrm{n}$ denotes the unit outer normal at $\partial\Omega$.
Let $u$ be a
weak time-periodic solution to \eqref{sys:NavierStokesTP_ExteriorDomain}
in the sense of Definition \ref{def:WeakSolution_NStp}, which satisfies
\begin{equation}
\label{el:VelocityAdditionalIntegrability}
\exists r\in(5,\infty): \quad \calp_\botu\in\LR{r}({\mathbb T}\times\Omega)^3.
\end{equation}
Then there exist constants $C_1>0$ and
$\alpha=\alpha(\lambda,\calt)>0$ such that
\begin{align}
\snorm{\curl\calpu(x)}
&\leq C_1\snorm{x}^{-3/2}\e^{-\alpha\wakefct{x}},
\label{est:VorticityDecay_ss}\\
\snorm{\curl\calp_\botu(t,x)}
&\leq C_1\snorm{x}^{-9/2}\e^{-\alpha\wakefct{x}}
\label{est:VorticityDecay_pp}
\end{align}
for all $t\in{\mathbb T}$ and $x\in\Omega$.
\end{thm}
\begin{rem}The constant $C_1$ depends on $\Omega,\lambda$ and on norms of the solution $u$ which, in turn, can be estimated in terms of the body force $f$. So, ultimately, $C_1$ depends on $\Omega,\lambda$ and $f$.
If not specified otherwise, this may always be the case for all other constants $C$, $C_i$ that we will introduce throughout the paper.
\end{rem}
\begin{rem} In our proof, we need the zero-flux condition \eqref{cond:BoundaryData}$_4$ on the boundary velocity $u_*$, which, instead, is not needed in the particular case of steady-state solutions \cite{Clark_VorticityAtInfinityNavierStokes_1970,BabenkoVasilev_AsymptoticBehaviorSteadyFlow_1973}. Though it is probable that our result continues to hold if the flux is only ``sufficiently small," it is not clear whether the same conclusion may be drawn for flux of arbitrary magnitude.
\end{rem}
\begin{rem}Condition \eqref{el:VelocityAdditionalIntegrability}
is merely a technical assumption.
As pointed out in \cite{GaldiKyed_TPSolNS3D_AsymptoticProfile}
for the case $\Omega=\mathbb{R}^3$,
it leads to additional local regularity of the solution
but does not improve its spatial decay properties.
\end{rem}
\begin{rem} If $f$ is time-independent, then $u\equiv{\mathcal P} u$, and our result reduces to that of \textsc{Clark} \cite{Clark_VorticityAtInfinityNavierStokes_1970}
and \textsc{Babenko} and \textsc{Vasil'ev} \cite{BabenkoVasilev_AsymptoticBehaviorSteadyFlow_1973}. Actually --as it becomes clear from our proof-- in such a case, we do not need the assumptions \eqref{cond:BoundaryData}, and \eqref{el:VelocityAdditionalIntegrability}.
\end{rem}
\section{Time-periodic fundamental solutions}
\label{sec:TPFundSol}
In this section, we consider the so-called Oseen linearization of \eqref{sys:NavierStokesTP_ExteriorDomain} in the
whole space
given by
\begin{equation}\label{sys:NSlintp_TPFS}
\begin{pdeq}
\partial_tu-\Deltau - \lambda\partial_1u +\nabla\mathfrak{p} &= f
&& \text{in }{\mathbb T}\times\mathbb{R}^3, \\
\Divu &=0 && \text{in }{\mathbb T}\times\mathbb{R}^3
\end{pdeq}
\end{equation}
for $\lambda>0$.
In \cite{GaldiKyed_TPSolNS3D_AsymptoticProfile, EiterKyed_etpfslns},
a velocity fundamental solution $\varGamma^\lambda$ to the time-periodic problem
\eqref{sys:NSlintp_TPFS} was introduced such that
\[
u=\varGamma^\lambda\ast f
\]
with convolution taken with respect to the
locally compact abelian group $G={\mathbb T}\times\mathbb{R}^3$.
It is given by
\begin{equation}
\varGamma^\lambda := \fundsolvel_0\otimes 1_{{\mathbb T}} + \fundsolvel_\perp,
\label{eq:tpfundsol_decompVel}
\end{equation}
where
\begin{align}
&\fundsolvel_0\colon\mathbb{R}^3\setminus\set{0}\to\mathbb{R}^{3\times3}, \quad
\fundsolvel_{0,j\ell}(x)
:=\frac{1}{4\pi\lambda}\bb{\delta_{j\ell}\Delta-\partial_j\partial_\ell}
\int_0^{\wakefct{\lambda x} /2} \frac{1-\e^{-\tau}}{\tau}\,{\mathrm d}\tau,
\label{eq:OseenFundSolss_3d}\\
&\fundsolvel_\perp := \mathscr{F}^{-1}_G\Bb{
\frac{1-\delta_\mathbb{Z}(k)}{\snorm{\xi}^2 + i(\frac{2\pi}{\per} k - \lambda \xi_1)}\,
\Bp{I - \frac{\xi\otimes\xi}{\snorm{\xi}^2}}},
\label{eq:tpfundsol_deffundsolcompl}
\end{align}
the symbol $1_{{\mathbb T}}$ denotes the constant $1$ distribution,
and
$\wakefct{x}= \snorm{x}+x_1$ as above.
In particular, the fundamental solution $\varGamma^\lambda$ decomposes into
a \emph{steady-state} part $\fundsolvel_0$ and a \emph{purely periodic} part $\fundsolvel_\perp$.
The steady-state part $\fundsolvel_0$ is the fundamental solution to the steady-state Oseen problem
\begin{align}\label{sys:NSlinss_TPFS}
\begin{pdeq}
-\Deltav - \lambda\partial_1v +\nablap &= f && \text{in }\mathbb{R}^3, \\
\Divv &=0 && \text{in }\mathbb{R}^3;
\end{pdeq}
\end{align}
see \cite[Section VII.3]{GaldiBookNew}.
This function shows strongly anisotropic behavior, which is reflected in the pointwise estimates
\begin{equation}
\forall \alpha\in\mathbb{N}_0^3\ \forall \epsilon>0\ \exists C>0\ \forall \snorm{x}\geq \epsilon:\quad
\snorm{{\mathrm D}_x^\alpha \fundsolvel_0(x)} \leq C\bb{\snorm{x}\np{1+\wakefct{\lambda x}}}^{-1-\frac{\snorm{\alpha}}{2}};
\label{est:fundsolss_Decay}
\end{equation}
see \cite[Lemma 3.2]{Farwig_habil}.
For the purely periodic part $\fundsolvel_\perp$ one can show the estimates
\begin{equation}
\forall \alpha \in \mathbb{N}_0^3 \
\forall r\in [1,\infty)\ \forall\epsilon>0\ \exists C>0\ \forall \snorm{x}\geq \epsilon:\
\norm{{\mathrm D}_x^\alpha \fundsolvel_\perp(\cdot,x)}_{\LR{r}({\mathbb T})} \leq C\snorm{x}^{-3-\snorm{\alpha}};
\label{est:tpfundsol_ComplPointwiseEst}
\end{equation}
see \cite{EiterKyed_etpfslns}.
Observe that estimate \eqref{est:tpfundsol_ComplPointwiseEst} does not have an anisotropic character
and that the purely periodic part $\fundsolvel_\perp$ decays faster than
the steady-state part $\fundsolvel_0$.
In order to derive estimates of the solution $u$ from those of the fundamental solution
$\varGamma^\lambda$, one thus has to study convolutions of functions
that satisfy pointwise estimates similar to those in
\eqref{est:fundsolss_Decay} and \eqref{est:tpfundsol_ComplPointwiseEst}.
Convolutions of the first type were examined by
\textsc{Farwig} \cite{Farwig_habil,FarwigOseenAnisotropicallyWeightedSob}
in dimension $n=3$,
and later by
\textsc{Kra\v cmar}, \textsc{Novotn\'y} and \textsc{Pokorn\'y}
\cite{KracmarNovotnyPokorny2001} in the general $n$-dimensional case.
We collect some of their results in the following theorem,
which gives estimates of convolutions with $\fundsolvel_0$ and $\nabla\fundsolvel_0$.
\begin{thm}\label{thm:ConvFundsolss}
Let $A\in[2,\infty)$ and $B\in[0,\infty)$, and let $g\in\LR{\infty}(\mathbb{R}^3)$
such that $\snorm{g(x)}\leq M\np{1+\snorm{x}}^{-A}\np{1+\wakefct{x}}^{-B}$.
Then there exists a constant
$ C= C(A,B,\lambda)>0$
with the following properties:
\begin{enumerate}
\item
If $A+\min\set{1,B}>3$, then
\begin{equation}
\snorml{\snorm{\fundsolvel_0}\ast g (x)}
\leq C M \bb{\np{1+\snorm{x}}\bp{1+\wakefct{\lambda x}}}^{-1}.
\end{equation}
\item
If $A+\min\set{1,B}>3$ and $A+B\geq7/2$, then
\begin{equation}
\snorml{\snorm{\nabla\fundsolvel_0}\ast g (x)}
\leq C M \bb{\np{1+\snorm{x}}\bp{1+\wakefct{\lambda x}}}^{-3/2}.
\end{equation}
\end{enumerate}
\end{thm}
\begin{proof}
These are special cases of \cite[Theorems 3.1 and 3.2]{KracmarNovotnyPokorny2001}.
\end{proof}
An analogous result for convolutions with $\fundsolvel_\perp$ and $\nabla\fundsolvel_\perp$
was derived in \cite{Eiter_SpatiallyAsymptoticStructureTPNS_2020}.
\begin{thm}\label{thm:ConvFundsolpp}
Let $A\in\mathbb{R}$ and $g\in\LR{\infty}({\mathbb T}\times\mathbb{R}^3)$
such that $\snorm{g(t,x)}\leq M\np{1+\snorm{x}}^{-A}$.
Then for any $\varepsilon>0$ there exists a constant
$C= C(A,\lambda,\calt,\varepsilon)>0$
with the following properties:
\begin{enumerate}
\item
If $A>3$, then
\begin{equation}
\forall \snorm{x}\geq \epsilon: \qquad
\snorml{\snorm{\fundsolvel_\perp}\ast_{G} g (t,x)}
\leq C M
\np{1+\snorm{x}}^{-3}.
\label{est:ConvFundsolpp}
\end{equation}
\item
If $A>4$, then
\begin{equation}
\forall \snorm{x}\geq \epsilon: \qquad
\snorml{\snorm{\nabla\fundsolvel_\perp}\ast_{G} g (t,x)}
\leq C M
\np{1+\snorm{x}}^{-4}.
\label{est:ConvFundsolpp_Grad}
\end{equation}
\end{enumerate}
\end{thm}
\begin{proof}
We refer to \cite[Theorem 3.3]{Eiter_SpatiallyAsymptoticStructureTPNS_2020}.
\end{proof}
Next we derive a fundamental solution for the vorticity field $\curlu$.
For $u=\varGamma^\lambda\ast f$ a direct computation yields
\[
\np{\curlu}_m
= \varepsilon_{mhj}\partial_h\fundsolvel_{0,j\ell}\ast\calp f _\ell
+ \varepsilon_{mhj}\partial_h\fundsolvel_{\perp,j\ell}\ast f_\ell
=\varepsilon_{mh\ell} \partial_h\fundsolvort_0\ast\calp f _\ell
+\varepsilon_{mh\ell}\partial_h\fundsolvort_\perp\ast f_\ell
\]
with
\begin{align}
\fundsolvort_0(x)
&:= \frac{1}{4\pi\snorm{x}}\e^{-\wakefct{\lambda x}/2},
\label{eq:fundsolvortss}
\\
\fundsolvort_\perp
&:= \mathscr{F}^{-1}_G\Bb{\frac{1-\delta_\mathbb{Z}(k)}{\snorm{\xi}^2-i\lambda \xi_1+i\frac{2\pi}{\per} k}}.
\label{eq:fundsolvortpp_multiplier}
\end{align}
In conclusion, we obtain
\begin{equation}\label{eq:vorticityOseen}
\curl u(t,x) = \int_{G} \nabla\phi^\lambda(t-s,x-y) \wedge f(s,y) \,{\mathrm d}(s,y),
\end{equation}
where
\begin{equation}\label{eq:fundsolvort_decomp}
\phi^\lambda:=\fundsolvort_0\otimes1_{{\mathbb T}}+\fundsolvort_\perp.
\end{equation}
We have thus found an integral formula for the vorticity $\curlu$.
We call $\phi^\lambda$ the \emph{vorticity fundamental solution}.
As for the velocity fundamental solution $\varGamma^\lambda$,
the vorticity fundamental solution $\phi^\lambda$
decomposes into a steady-state and a purely periodic part,
which can be analyzed separately.
A direct computation leads to the the following estimate of $\nabla\fundsolvort_0$.
\begin{thm}\label{thm:FundsolVortss_est}
There exists $ C= C\np{\lambda}>0$
such that for all $x\in\mathbb{R}^3\setminus\set{0}$ it holds
\begin{equation}
\snorm{\nabla\fundsolvort_0(x)}
\leq C\bp{\snorm{x}^{-2}+\snorm{x}^{-3/2}\wakefct{\lambda x}^{1/2}}
\e^{-\wakefct{\lambda x}/2}.
\label{est:FundsolVortss_grad}
\end{equation}
\end{thm}
\begin{proof}
The estimate follows directly by taking derivatives in \eqref{eq:fundsolvortss}
and using the identity
$\snorm{\nabla\nb{\wakefct{\lambda x}}}^2=2\lambda^2\wakefct{x}/\snorm{x}$.
\end{proof}
The remainder of this section is dedicated to the derivation of
an analogous estimate of $\nabla\fundsolvort_\perp$.
More precisely, we show the following result.
\begin{thm}\label{thm:FundsolVortpp_est}
There exist constants $C= C(\lambda,\calt,q,\gamma)>0$
and $C_3= C_3(\lambda,\calt)>0$
such that
for all $\gamma\in(0,1)$, $q\in[1,\frac{1}{1-\gamma})$ and
$x\in\mathbb{R}^3\setminus\set{0}$
it holds
\begin{align}
\norm{\fundsolvort_\perp(\cdot,x)}_{\LR{q}({\mathbb T})}
&\leq C\snorm{x}^{-(1+2\gamma)} \e^{- C_3\snorm{x}},
\label{est:FundsolVortpp_fct}\\
\norm{\nabla\fundsolvort_\perp(\cdot,x)}_{\LR{q}({\mathbb T})}
&\leq C\snorm{x}^{-(2+2\gamma)} \e^{- C_3\snorm{x}}.
\label{est:FundsolVortpp_grad}
\end{align}
\end{thm}
For the proof of Theorem \ref{thm:FundsolVortpp_est}
we represent $\fundsolvort_\perp$ in a different way.
From
$\mathscr{F}^{-1}_{G}=\mathscr{F}^{-1}_{{\mathbb T}}\otimes\mathscr{F}^{-1}_{\mathbb{R}^3}$ we conclude the identity
\begin{equation}
\fundsolvort_\perp(t,x)
= \mathscr{F}^{-1}_{\mathbb T} \bb{ k\mapsto\bp{1-\delta_\mathbb{Z}(k)}\varGamma_{\mathrm{H}}^{\frac{2\pi}{\per} k,\lambda}(x)}(t),
\label{eq:fundsolvortpp_series}
\end{equation}
where
\[
\varGamma_{\mathrm{H}}^{\eta,\lambda}:=\mathscr{F}^{-1}_{\mathbb{R}^3}\Bb{\frac{1}{\snorm{\xi}^2-i\lambda \xi_1+i\eta}}
\]
is the fundamental solution to the equation
\begin{equation}
\label{eq:HelmholtzWithDrift}
i\eta\,v-\Deltav -\lambda \partial_1v =f \qquad \text{in }\mathbb{R}^3.
\end{equation}
This function is explicitly given by
\begin{equation}\label{eq:fundsolaux3D}
\varGamma_{\mathrm{H}}^{\eta,\lambda}\colon\mathbb{R}^3\setminus\set{0}\to\mathbb{C}, \qquad
\varGamma_{\mathrm{H}}^{\eta,\lambda}(x)=\frac{1}{4\pi\snorm{x}}\e^{i\sqrt{-\mu}\snorm{x}-\frac{\lambda}{2}x_1}
\end{equation}
for $\eta\neq0$ and $\mu := \mu(\eta,\lambda):= (\lambda/2)^2+i\eta\in\mathbb{C}\setminus\mathbb{R}$;
see \cite[Lemma 3.3]{EiterKyed_etpfslns}.
Here $\sqrt{z}$ is the square root of $z$ with nonnegative imaginary part.
We first derive pointwise estimates of $\varGamma_{\mathrm{H}}^{\eta,\lambda}$.
\begin{lem}\label{lem:fundsolaux_pointwiseEst3D}
Let $\eta_0>0$. Then there exists
$C_4= C_4(\lambda,\eta_0)>0$ such that
\begin{align}
\snorm{\varGamma_{\mathrm{H}}^{\eta,\lambda}(x)}
&\leq C\snorm{x}^{-1}\e^{-C_4\snorm{\eta}^\frac{1}{2}\snorm{x}},
\label{est:fundsolaux_pointwiseEst3D_fct}\\
\snorm{\nabla\varGamma_{\mathrm{H}}^{\eta,\lambda}(x)}
&\leq C\bp{\snorm{x}^{-2}+\snorm{\eta}^\frac{1}{2}\snorm{x}^{-1}}
\e^{-C_4\snorm{\eta}^\frac{1}{2}\snorm{x}},
\label{est:fundsolaux_pointwiseEst3D_grad}
\end{align}
for all $\eta\in\mathbb{R}$ with $\snorm{\eta}>\eta_0$ and $x\in\mathbb{R}^3\setminus\set{0}$.
\end{lem}
\begin{proof}
As in \cite[Lemma 3.2]{EiterKyed_etpfslns},
we show the existence of a constant
$C_4= C_4(\lambda,\eta_0)>0$ such that
\[
\impart( \sqrt{-\mu})-\frac{\snorm{\lambda}}{2} \geq C_4 \snorm{\eta}^\frac{1}{2}
\]
for $\snorm{\eta}\geq\eta_0$ and $\mu = (\lambda/2)^2+i\eta$.
We thus have
\[
\snormL{\e^{i\sqrt{-\mu}\snorm{x}-\frac{\lambda}{2}x_1}}
\leq \e^{-\impart\np{\sqrt{-\mu}}\snorm{x}+\frac{\snorm{\lambda}}{2}\snorm{x}}
\leq \e^{-C_4\snorm{\eta}^\frac{1}{2}\snorm{x}}.
\]
This directly implies \eqref{est:fundsolaux_pointwiseEst3D_fct}.
Computing derivatives and employing this estimate again, we further deduce
\[
\snorml{\nabla\varGamma_{\mathrm{H}}^{\eta,\lambda}(x)}
\leq C\bp{
\snorm{x}^{-2}+\snorm{x}^{-1}\np{\snorm{\sqrt{-\mu}}+\snorm{\lambda}}
}\e^{-C_4\snorm{\eta}^\frac{1}{2}\snorm{x}},
\]
which implies \eqref{est:fundsolaux_pointwiseEst3D_grad}
by using $\snorm{\lambda}\leq2\snorm{\sqrt{-\mu}}\leq C\snorm{\eta}^\frac{1}{2}$
for $\snorm{\eta}\geq\eta_0$.
\end{proof}
Now let
$\chi\in\CR \infty(\mathbb{R})$, $0\leq \chi \leq 1$, with
$\chi(\eta)=0$ for $\snorm{\eta}\leq \frac{1}{2}$ and $\chi(\eta)=1$ for $\snorm{\eta}\geq 1$.
For $\alpha\in\mathbb{N}_0^3$ with $\snorm{\alpha}\leq1$, $\gamma\in(0,1)$
and $x\in\mathbb{R}^3\setminus\set{0}$
define the function
\begin{align}\label{eq:FundsolVortpp_est_multiplier}
m_{\alpha,x}\colon\mathbb{R}\to\mathbb{R}, \qquad
m_{\alpha,x}(\eta):=\chi(\eta)\snorm{\eta}^\gamma
{\mathrm D}^\alpha\varGamma_{\mathrm{H}}^{\frac{2\pi}{\per}\eta,\lambda}(x).
\end{align}
We show that $m_{\alpha,x}$ is an $\LR{q}(\mathbb{R})$ multiplier
and give an estimate of the multiplier norm
by means of the Marcinkiewicz Multiplier Theorem.
\begin{lem}\label{lem:FundsolVortpp_est_multiplier}
Let $\alpha\in\mathbb{N}_0^3$ with $\snorm{\alpha}\leq1$, $\gamma\in(0,1)$
and $x\in\mathbb{R}^3\setminus\set{0}$.
Then $m_{\alpha,x}$ is an $\LR{q}(\mathbb{R})$ multiplier for any $q\in(1,\infty)$,
and there exist constants
$ C
=C\np{\lambda,\calt,q,\alpha,\gamma}>0$
and $ C_5= C_5\np{\lambda,\calt}>0$
such that
\[
\norml{\mathscr{F}^{-1}_\mathbb{R}\bb{m_{\alpha,x}\mathscr{F}_\mathbb{R}\nb{f}}}_{\LR{q}(\mathbb{R})}
\leq C\snorm{x}^{-1-\snorm{\alpha}-2\gamma}
\e^{- C_5\snorm{x}}\norm{f}_{\LR{q}(\mathbb{R})}.
\]
\end{lem}
\begin{proof}
At first, let $\alpha=0$.
From \eqref{est:fundsolaux_pointwiseEst3D_fct} we conclude
\[
\snorm{m_{0,x}(\eta)}\leq C \chi(\eta)\snorm{\eta}^\gamma
\snorm{x}^{-1}\e^{- C_4\snorm{\frac{2\pi}{\per}\eta}^\frac{1}{2}\snorm{x}}
\leq C \snorm{x}^{-1-2\gamma}\e^{- C_4\snorm{\frac{2\pi}{\per}\eta}^\frac{1}{2}\snorm{x}/2}
\]
for $\snorm{\eta}\geq\frac{1}{2}$.
Moreover, differentiating $\varGamma_{\mathrm{H}}^{\frac{2\pi}{\per}\eta,\lambda}$ with respect to $\eta$, we obtain
\[
\snorml{\partial_\eta\varGamma_{\mathrm{H}}^{\frac{2\pi}{\per}\eta,\lambda}(x)}
\leq C \snorm{\partial_\eta\sqrt{-\mu}}\,\snorm{x}\,\snorml{\varGamma_{\mathrm{H}}^{\frac{2\pi}{\per}\eta,\lambda}(x)}
\leq C \snorm{\eta}^{-\frac{1}{2}}\snorm{x}\,\snorml{\varGamma_{\mathrm{H}}^{\frac{2\pi}{\per}\eta,\lambda}(x)},
\]
so that \eqref{est:fundsolaux_pointwiseEst3D_fct} yields
\begin{align*}
\snorm{\eta\partial_\etam_{0,x}(\eta)}
&\leq\snorml{\chi'(\eta)\snorm{\eta}^{\gamma+1}\varGamma_{\mathrm{H}}^{\frac{2\pi}{\per}\eta,\lambda}(x)}
+\snorml{\chi(\eta)\gamma\snorm{\eta}^\gamma\varGamma_{\mathrm{H}}^{\frac{2\pi}{\per}\eta,\lambda}(x)}
+\snorml{\chi(\eta)\snorm{\eta}^{\gamma+1}\partial_\eta\varGamma_{\mathrm{H}}^{\frac{2\pi}{\per}\eta,\lambda}(x)}\\
&\quad\leq C \bp{\snorm{\eta}^\gamma\snorm{x}^{-1}+\snorm{\eta}^{\gamma+\frac{1}{2}}}
\e^{- C_4\snorm{\frac{2\pi}{\per}\eta}^\frac{1}{2}\snorm{x}}
\leq C \snorm{x}^{-1-2\gamma}\e^{- C_4\snorm{\frac{2\pi}{\per}\eta}^\frac{1}{2}\snorm{x}/2}
\end{align*}
for $\snorm{\eta}\geq\frac{1}{2}$.
Collecting these estimates and utilizing $m_{0,x}(\eta)=0$ for
$\snorm{\eta}\leq \frac{1}{2}$, we have
\begin{equation}\label{est:fundsolvortpp_fct_multiplierEst}
\snorm{m_{0,x}(\eta)}+\snorm{\eta\partial_\etam_{0,x}(\eta)}
\leq C \snorm{x}^{-1-2\gamma}\e^{- C_5\snorm{x}}
\end{equation}
with $ C_5=\sqrt{\pi/\calt} C_4/2$ for all $\eta\in\mathbb{R}$.
Next consider the case $\alpha=\e_j$ for some $j\in\set{1,2,3}$.
Then \eqref{est:fundsolaux_pointwiseEst3D_grad} leads to
\[
\snorm{m_{\alpha,x}(\eta)}
\leq C \chi(\eta)\snorm{\eta}^\gamma
\bp{\snorm{x}^{-2}+\snorm{\eta}^\frac{1}{2}\snorm{x}^{-1}}
\e^{- C_4\snorm{\frac{2\pi}{\per}\eta}^\frac{1}{2}\snorm{x}}
\leq C \snorm{x}^{-2-2\gamma}\e^{- C_4\snorm{\frac{2\pi}{\per}\eta}^\frac{1}{2}\snorm{x}/2}
\]
for $\snorm{\eta}\geq\frac{1}{2}$.
Moreover, a straightforward calculation yields
\[
\snorml{\partial_\eta\partial_j\varGamma_{\mathrm{H}}^{\frac{2\pi}{\per}\eta,\lambda}(x)}
\leq C \bp{\snorm{\mu}^{-\frac{1}{2}}+\snorm{x}}
\snorml{\varGamma_{\mathrm{H}}^{\frac{2\pi}{\per}\eta,\lambda}(x)}
\leq C \bp{\snorm{\eta}^{-\frac{1}{2}}+\snorm{x}}
\snorml{\varGamma_{\mathrm{H}}^{\frac{2\pi}{\per}\eta,\lambda}(x)},
\]
so that we can employ Lemma \ref{lem:fundsolaux_pointwiseEst3D} to estimate
\begin{align*}
\snorm{\eta\partial_\eta\partial_jm_{\alpha,x}(\eta)}
&\leq\snorml{\chi'(\eta)\snorm{\eta}^{\gamma+1}\partial_j\varGamma_{\mathrm{H}}^{\frac{2\pi}{\per}\eta,\lambda}(x)}
+\snorml{\chi(\eta)\gamma\snorm{\eta}^\gamma\partial_j\varGamma_{\mathrm{H}}^{\frac{2\pi}{\per}\eta,\lambda}(x)}\\
&\qquad\qquad+\snorml{\chi(\eta)\snorm{\eta}^{\gamma+1}
\partial_\eta\partial_j\varGamma_{\mathrm{H}}^{\frac{2\pi}{\per}\eta,\lambda}(x)}\\
&\leq C \bp{\snorm{\eta}^\gamma\snorm{x}^{-2}+\snorm{\eta}^{\gamma+\frac{1}{2}}\snorm{x}^{-1}
+\snorm{\eta}^{\gamma+1}}
\e^{- C_4\snorm{\frac{2\pi}{\per}\eta}^\frac{1}{2}\snorm{x}}
\\
&\leq C \snorm{x}^{-2-2\gamma}\e^{- C_4\snorm{\frac{2\pi}{\per}\eta}^\frac{1}{2}\snorm{x}/2}
\end{align*}
for $\snorm{\eta}\geq\frac{1}{2}$.
Collecting these estimates and utilizing $m_{\alpha,x}(\eta)=0$ for
$\snorm{\eta}\leq \frac{1}{2}$, we have
\begin{align}\label{est:fundsolvortpp_grad_multiplierEst}
\snorm{m_{\alpha,x}(\eta)}+\snorm{\eta\partial_\etam_{\alpha,x}(\eta)}
\leq C \snorm{x}^{-2-2\gamma}\e^{- C_5\snorm{x}}
\end{align}
with $ C_5=\sqrt{\pi/\calt} C_4/2$ as above.
By the Marcinkiewicz Multiplier Theorem (see \cite[Corollary 5.2.5]{Grafakos1}),
the assertion is now a direct consequence of
\eqref{est:fundsolvortpp_fct_multiplierEst} and \eqref{est:fundsolvortpp_grad_multiplierEst}.
\end{proof}
Using this result, we establish the pointwise estimates of $\fundsolvort_\perp$ asserted in Theorem \ref{thm:FundsolVortpp_est}
by means of the so-called transference principle for Fourier multipliers.
\begin{proof}[Proof of Theorem \ref{thm:FundsolVortpp_est}]
It suffices to consider $q\in(1,\infty)$.
Due to \eqref{eq:fundsolvortpp_series}, we have
\begin{equation}\label{eq:FundsolVortpp_TorusMult}
{\mathrm D}_x^\alpha\fundsolvort_\perp(\cdot,x)= \mathscr{F}^{-1}_{\mathbb T}\bb{M_{\alpha,x}\mathscr{F}_{\mathbb T}\nb{\varphi_\gamma}}
\end{equation}
with
\[
M_{\alpha,x}(k):= \bp{1-\delta_\mathbb{Z}(k)} \snorm{k}^\gamma {\mathrm D}_x^\alpha\varGamma_{\mathrm{H}}^{\frac{2\pi}{\per} k,\lambda} (x),
\qquad
\varphi_\gamma:=\mathscr{F}^{-1}_{\mathbb T}\bb{k\mapsto\bp{1-\delta_\mathbb{Z}(k)}\snorm{k}^{-\gamma}}.
\]
First, note that $M_{\alpha,x}=\restriction{m_{\alpha,x}}{\mathbb{Z}}$.
Since $m_{\alpha,x}$ is a continuous $\LR{q}(\mathbb{R})$ multiplier
by Lemma \ref{lem:FundsolVortpp_est_multiplier},
the transference principle
(see \cite[Theorem B.2.1]{EdwardsGaudryBook} or \cite[Theorem 2.15]{EiterKyed_tplinNS_PiFbook})
implies that $M_{\alpha,x}$
is an $\LR{q}({\mathbb T})$ multiplier for any $q\in(1,\infty)$
and that
\[
\norm{\mathscr{F}^{-1}_{\mathbb T}\bb{M_{\alpha,x}\mathscr{F}_{\mathbb T}\nb{f}}}_{\LR{q}({\mathbb T})}
\leq C \snorm{x}^{-1-\snorm{\alpha}-2\gamma}\e^{- C_3\snorm{x}}\norm{f}_{\LR{q}({\mathbb T})}.
\]
Moreover, we have
$\varphi_\gamma\in\LR{q}({\mathbb T})$
provided $q<1/(1-\gamma)$, which is a direct consequence of
\cite[Example 3.1.19]{Grafakos1} for example.
Finally, the assertion follows from \eqref{eq:FundsolVortpp_TorusMult}.
\end{proof}
\section{Regularity results}\label{sec:Regularity}
Here we collect some results concerning the
regularity of weak solutions to \eqref{sys:NavierStokesTP_ExteriorDomain}
and its linearization,
which is given by
\begin{equation}\label{sys:NSlintp_extdomain}
\begin{pdeq}
\partial_tu-\Deltau - \lambda\partial_1u +\nabla\mathfrak{p} &= f
&& \text{in }{\mathbb T}\times\Omega,
\\
\Divu &=0 && \text{in }{\mathbb T}\times\Omega,
\\
u & =u_\ast
&&\text{on }{\mathbb T}\times\partial\Omega.
\end{pdeq}
\end{equation}
First of all, we derive the following regularity theorem for solutions to
\eqref{sys:NSlintp_extdomain}
in the case $\calp f=0$.
\begin{lem}\label{lem:NSlintp_Regularity}
Let $\Omega\subset\mathbb{R}^3$ be an exterior domain of class $\CR{2}$,
let $u_\ast$ be as in \eqref{cond:BoundaryData},
and let $f\in\LR{q}({\mathbb T}\times\Omega)$ for some $q\in(1,\infty)$.
Assume $\calp f=0$ and that $u$ is a weak
solution to \eqref{sys:NSlintp_extdomain},
that is, $u=u_\ast$ on ${\mathbb T}\times\partial\Omega$,
$\Divu=0$ and
\begin{equation}\label{eq:NSlintp_weak}
\int_{{\mathbb T}\times\Omega}\bb{-u\cdot\partial_t\varphi
+\nablau:\nabla\varphi
-\lambda\partial_1u\cdot\varphi
}\,{\mathrm d}(t,x)
=\int_{{\mathbb T}\times\Omega} f\cdot\varphi\,{\mathrm d}(t,x)
\end{equation}
for all $\varphi\in\CR{\infty}_{0,\sigma}({\mathbb T}\times\Omega)$.
Assume that $u\in\LR{\infty}({\mathbb T};\LR{2}(\Omega))^3$ and
$\nablau\in\LR{2}({\mathbb T}\times\Omega)^{3\times3}$.
Then
$u\in\WSR{1,2}{q}({\mathbb T}\times\Omega)^3$,
and there exists $\mathfrak{p}\in\LR{q}({\mathbb T};\DSR{1}{q}(\Omega))$
such that $\np{u,\mathfrak{p}}$ is a strong solution to
\eqref{sys:NSlintp_extdomain}.
\end{lem}
\begin{proof}
First of all, using classical arguments
(see \cite[Section III.3]{GaldiBookNew} for example)
one can show the existence of a function
$U\in\WSR{1,2}{q}({\mathbb T}\times\Omega)^3$
such that $U=u_\ast$ on ${\mathbb T}\times\partial\Omega$ and
$\DivU=0$.
Moreover, since $\calp f=0$,
by \cite[Theorem 5.1]{GaldiKyed_TPflowViscLiquidpBody}
there exist $z\in\WSR{1,2}{q}({\mathbb T}\times\Omega)^3$ and
$\mathfrak{p}\in\LR{q}({\mathbb T};\DSR{1}{q}(\Omega))$
such that
\begin{equation}
\begin{pdeq}
\partial_tz-\Deltaz - \lambda\partial_1z +\nabla\mathfrak{p}
&= f - \partial_tU-\DeltaU - \lambda\partial_1U
&& \text{in }{\mathbb T}\times\Omega,
\\
\Divz &=0 && \text{in }{\mathbb T}\times\Omega,
\\
z & =0
&&\text{on }{\mathbb T}\times\partial\Omega.
\end{pdeq}
\end{equation}
Then $\np{\tilde{u},\mathfrak{p}}:=\np{z+U,\mathfrak{p}}$
solves \eqref{sys:NSlintp_extdomain},
and for the completion of the proof it remains to show $u=\tilde{u}$.
For this purpose, we employ a duality argument.
Let $\psi\in\CR \infty_0({\mathbb T}\times\Omega)^3$.
By \cite[Theorem 5.1]{GaldiKyed_TPflowViscLiquidpBody}
there exist functions
$w\in\WSR{1,2}{2}({\mathbb T}\times\Omega)^3
\cap\WSR{1,2}{q'}({\mathbb T}\times\Omega)^3$
and
$\mathfrak{q}\in\LR{2}({\mathbb T};\DSR{1}{2}(\Omega))
\cap\LR{q'}({\mathbb T};\DSR{1}{q'}(\Omega))$,
where $q'=q/(q-1)$,
which satisfy
\begin{equation}\label{sys:NSlintp_extdomain_hombdry}
\begin{pdeq}
\partial_tw-\Deltaw + \lambda\partial_1w +\nabla\mathfrak{q} &= \calp_\bot\psi
&& \text{in }{\mathbb T}\times\Omega,
\\
\Divw &=0 && \text{in }{\mathbb T}\times\Omega,
\\
w & =0
&&\text{on }{\mathbb T}\times\partial\Omega.
\end{pdeq}
\end{equation}
By a standard density argument one shows that we can let $\varphi=w$ in
the weak formulation \eqref{eq:NSlintp_weak}.
Then, by an integration by parts, we get
\[
\begin{aligned}
\int_{{\mathbb T}\times\Omega}\np{u-\tilde{u}}\cdot\calp_\bot\psi\,{\mathrm d}(t,x)
&=\int_{{\mathbb T}\times\Omega}\np{u-\tilde{u}}
\cdot\bp{\partial_tw-\Deltaw + \lambda\partial_1w +\nabla\mathfrak{q}}\,{\mathrm d}(t,x)
\\
&=\int_{{\mathbb T}\times\Omega}\bb{u\cdot\partial_tw+\nablau:\nablaw
- \lambda\partial_1u\cdotw}\,{\mathrm d}(t,x)
\\
&\qquad\qquad-\int_{{\mathbb T}\times\Omega} \bp{\partial_t\tilde{u}-\Delta\tilde{u} - \lambda\partial_1\tilde{u} +\nabla\mathfrak{p}}\cdotw\,{\mathrm d}(t,x)
\\
&=\int_{{\mathbb T}\times\Omega} f\cdotw-\int_{{\mathbb T}\times\Omega} f\cdotw
=0.
\end{aligned}
\]
Since $\calpu=\calp\tilde{u}=0$, we thus conclude
\[
\int_{{\mathbb T}\times\Omega}\np{u-\tilde{u}}\cdot\psi\,{\mathrm d}(t,x)
=\int_{{\mathbb T}\times\Omega}\np{u-\tilde{u}}\cdot\calp\psi\,{\mathrm d}(t,x)
+\int_{{\mathbb T}\times\Omega}\np{u-\tilde{u}}\cdot\calp_\bot\psi\,{\mathrm d}(t,x)
=0
\]
for arbitrary $\psi\in\CR \infty_0\np{{\mathbb T}\times\Omega}^3$,
which implies $u=\tilde{u}$ and completes the proof.
\end{proof}
Based on this result for the linearized problem \eqref{sys:NSlintp_extdomain},
we can now show that the additional integrability condition
assumed in \eqref{el:VelocityAdditionalIntegrability}
leads to higher regularity of the weak solution.
\begin{lem}\label{lem:NavierStokesTP_Regularity}
In the hypotheses of Theorem \ref{thm:VorticityDecay}
we have $u=v+w$, with
$v=\calpu$, $w=\calp_\botu$ such that
\begin{align}
&\forall q\in(1,\infty):\, v\in\DSR{2}{q}(\Omega),\ \
\forall r\in(\frac{4}{3},\infty]:\, v\in\DSR{1}{r}(\Omega), \ \
\forall s\in(2,\infty]:\, v\in\LR{s}(\Omega),
\label{el:NavierStokesTP_Regularity_ss}
\\
&\forall q\in(1,\infty): \ w\in\WSR{1,2}{q}({\mathbb T}\times\Omega).
\label{el:NavierStokesTP_Regularity_pp}
\end{align}
Moreover, there exists a pressure field $\mathfrak{p}$ with
\begin{equation}\label{el:NavierStokesTP_Regularity_pressure}
\forall q\in(1,\infty): \ \mathfrak{p}\in\LR{q}({\mathbb T};\DSR{1}{q}(\Omega))
\end{equation}
such that \eqref{sys:NavierStokesTP_ExteriorDomain}
is satisfied in the strong sense.
\end{lem}
\begin{proof}
At first, observe that $v$ is a weak solution to
the steady-state Navier--Stokes problem
\begin{equation}\label{sys:NavierStokesTP_ExteriorDomain_sspart}
\begin{pdeq}
-\Deltav-\lambda\partial_1v+v\cdot\nablav+\nablap
&=\calp f-\calp\nb{w\cdot\nablaw}
&&\text{in }\Omega, \\
\Divv&=0
&&\text{in }\Omega, \\
v&=\calpu_\ast
&&\text{on }\partial\Omega.
\end{pdeq}
\end{equation}
H\"older's inequality and Definition \ref{def:WeakSolution_NStp} yield
$w\cdot\nablaw
\in\LR{1}({\mathbb T};\LR{3/2}(\Omega))\cap\LR{2}({\mathbb T};\LR{1}(\Omega))$.
Therefore, we have
$\calp f-\calp(w\cdot\nablaw)\in\LR{1}(\Omega)\cap\LR{3/2}(\Omega)$,
and \cite[Lemma X.6.1]{GaldiBookNew} implies
\begin{equation}\label{el:regularity_ss1}
\forall q\in(1,\frac{3}{2}]:\, v\in\DSR{2}{q}(\Omega),\ \
\forall r\in(\frac{4}{3},3]:\, v\in\DSR{1}{r}(\Omega), \ \
\forall s\in(2,\infty):\, v\in\LR{s}(\Omega)
\end{equation}
and the existence of $p\in\DSR{1}{q}(\Omega)$ for all $q\in(1,\frac{3}{2}]$
such that \eqref{sys:NavierStokesTP_ExteriorDomain_sspart} is satisfied
in the strong sense.
Moreover,
$w$ is a weak solution to
\begin{equation}\label{sys:NavierStokesTP_ExteriorDomain_pppart}
\begin{pdeq}
\partial_tw-\Deltaw - \lambda\partial_1w +\nabla\mathfrak{q}
&= \calp_\bot f -v\cdot\nablaw-w\cdot\nablav
-\calp_\bot\np{w\cdot\nablaw}
&& \text{in }{\mathbb T}\times\Omega,
\\
\Divw &=0 && \text{in }{\mathbb T}\times\Omega,
\\
w & =\calp_\botu_\ast
&&\text{on }{\mathbb T}\times\partial\Omega.
\end{pdeq}
\end{equation}
By a standard interpolation argument, the assumptions from
Definition \ref{def:WeakSolution_NStp} imply
$w\in\LR{10/3}({\mathbb T}\times\Omega)$.
Since $v\in\LR{10/3}(\Omega)$ by \eqref{el:regularity_ss1}
and $\nablau\in\LR{2}({\mathbb T}\times\Omega)$ by assumption,
this implies
\begin{equation}\label{el:rhs_pp}
\calp_\bot f -v\cdot\nablaw-w\cdot\nablav
-\calp_\bot\np{w\cdot\nablaw} \in\LR{s}({\mathbb T}\times\Omega)
\end{equation}
for $s=5/4$.
Now Lemma \ref{lem:NSlintp_Regularity} shows
that there exists $\mathfrak{q}$
such that
\begin{equation}\label{el:regularity_pp_5/4}
w\in\WSR{1,2}{5/4}({\mathbb T}\times\Omega), \
\mathfrak{q}\in\LR{5/4}({\mathbb T};\DSR{1}{5/4}(\Omega)),
\end{equation}
and \eqref{sys:NavierStokesTP_ExteriorDomain_pppart}
holds in a strong sense.
Starting from \eqref{el:regularity_pp_5/4},
we now employ a boot-strap argument to conclude the proof.
If $w\in\WSR{1,2}{q}({\mathbb T}\times\Omega)$
for some $q\in(1,15/8)$,
then the embedding theorem from
\cite[Theorem 4.1]{GaldiKyed_TPflowViscLiquidpBody}
implies $\nablaw\in\LR{5q/(5-q)}({\mathbb T}\times\Omega)$.
In virtue of \eqref{el:VelocityAdditionalIntegrability}
and \eqref{el:regularity_ss1},
this implies $w\cdot\nablaw,\
v\cdot\nablaw\in\LR{s}({\mathbb T}\times\Omega)$
for $\frac{1}{s}=\frac{1}{q}+\frac{1}{r}-\frac{1}{5}$.
Moreover, \cite[Theorem 4.1]{GaldiKyed_TPflowViscLiquidpBody}
yields $w\in\LR{5q/(5-q)}({\mathbb T};\LR{15q/(15-8q)}(\Omega))$,
so that
$w\cdot\nablav\in\LR{s}$ by \eqref{el:regularity_ss1}.
In total, we thus obtain
\eqref{el:rhs_pp} for $\frac{1}{s}=\frac{1}{q}+\frac{1}{r}-\frac{1}{5}$,
and Lemma \ref{lem:NSlintp_Regularity} leads to the implication
\begin{equation}\label{impl:bootstrap_1}
\exists\, q\in\bp{1,\frac{15}{8}}:\, w\in\WSR{1,2}{q}({\mathbb T}\times\Omega)
\implies
\forall\, \frac{1}{s}\in\bb{\frac{1}{q}+\frac{1}{r}-\frac{1}{5},\frac{1}{q}}:\,
w\in\WSR{1,2}{s}({\mathbb T}\times\Omega).
\end{equation}
If $w\in\WSR{1,2}{q}({\mathbb T}\times\Omega)$
for some $q\in[5/3,5/2)$,
then \cite[Theorem 4.1]{GaldiKyed_TPflowViscLiquidpBody}
yields $w\in\LR{5q/(5-2q)}({\mathbb T}\times\Omega)$ and
$\nablaw\in\LR{5q/(5-q)}({\mathbb T}\times\Omega)$,
which implies $w\cdot\nablaw\in\LR{s_1}({\mathbb T}\times\Omega)$
for all $\frac{1}{s_1}\in[\frac{2}{q}-\frac{3}{5},\frac{2}{q}]$.
Hence we have $\calp(w\cdot\nablaw)\in\LR{s_1}(\Omega)$,
and another application of \cite[Lemma X.6.1]{GaldiBookNew} in view of
\eqref{el:regularity_ss1} yields $\nablav\in\LR{t}(\Omega)$ for all $t\in[4/3,15/4]$ and $v\in\LR{t}(\Omega)$ for all $t\in(2,\infty]$.
We thus conclude
$w\cdot\nablav\in\LR{s_2}({\mathbb T}\times\Omega)$ for
$\frac{1}{s_2}\in[\frac{1}{q}-\frac{2}{15},\frac{1}{q}+\frac{3}{4}]$
and $v\cdot\nablaw\in\LR{s_3}({\mathbb T}\times\Omega)$
for $\frac{1}{s_3}\in[\frac{1}{q}-\frac{1}{5},\frac{1}{q}+\frac{1}{2}]$.
In particular, we obtain \eqref{el:rhs_pp}
for $\frac{1}{s}=\frac{2}{q}-\frac{3}{5}$ if $q\leq15/7$,
and $\frac{1}{s}=\frac{1}{q}-\frac{2}{15}$ if $q\geq15/7$.
By Lemma \ref{lem:NSlintp_Regularity}, this implies
\begin{align}
\exists\, q\in\big(\frac{5}{3},\frac{15}{7}\big]:\, w\in\WSR{1,2}{q}({\mathbb T}\times\Omega)
\implies
&\forall\, \frac{1}{s}\in\bb{\frac{2}{q}-\frac{3}{5},\frac{1}{q}}:\,
w\in\WSR{1,2}{s}({\mathbb T}\times\Omega),
\label{impl:bootstrap_2}
\\
\exists\, q\in\big[\frac{15}{7},\frac{5}{2}\big):\, w\in\WSR{1,2}{q}({\mathbb T}\times\Omega)
\implies
&\forall\, \frac{1}{s}\in\bb{\frac{1}{q}-\frac{2}{15},\frac{1}{q}}:\,
w\in\WSR{1,2}{s}({\mathbb T}\times\Omega).
\label{impl:bootstrap_3}
\end{align}
Arguing in a similar fashion, one shows the further implications
\begin{align}
\exists\, q\in\big[\frac{5}{2},5):\, w\in\WSR{1,2}{q}({\mathbb T}\times\Omega)
\implies
&\forall\, \frac{1}{s}\in\big(\frac{1}{q}-\frac{1}{5},\frac{1}{q}\big]:\,
w\in\WSR{1,2}{s}({\mathbb T}\times\Omega),
\label{impl:bootstrap_4}
\\
\exists\, q\in[5,\infty):\, w\in\WSR{1,2}{q}({\mathbb T}\times\Omega)
\implies
&\forall\, s\in[q,\infty):\,
w\in\WSR{1,2}{s}({\mathbb T}\times\Omega).
\label{impl:bootstrap_5}
\end{align}
Using now \eqref{el:regularity_pp_5/4} as starting point,
we can iteratively employ \eqref{impl:bootstrap_1}--\eqref{impl:bootstrap_5}
to obtain $w\in\WSR{1,2}{s}({\mathbb T}\times\Omega)$
for all $s\in[5/4,\infty)$.
Firstly, this
yields $w\cdot\nablaw\in\LR{\infty}({\mathbb T}\times\Omega)$,
so that
$\calp f-\calp\np{w\cdot\nablaw}\in\LR{q}(\Omega)$
for all $q\in[1,\infty)$.
Now \eqref{el:NavierStokesTP_Regularity_ss}
is a direct consequence of \cite[Theorem X.6.4]{GaldiBookNew}.
Secondly, this shows that
\eqref{el:rhs_pp} holds for all $s\in[1,\infty)$,
whence Lemma \ref{lem:NSlintp_Regularity}
implies \eqref{el:NavierStokesTP_Regularity_pp}.
Finally, the claimed regularity \eqref{el:NavierStokesTP_Regularity_pressure}
of $\mathfrak{p}=p+\mathfrak{q}$ is
a direct consequence.
This completes the proof.
\end{proof}
\section{The fixed-point problem}
\label{sec:FixedPointProblem}
In this section we derive a suitable fixed-point equation
satisfied by weak solutions in the whole space,
and we introduce the necessary functional framework.
More precisely, the main focus of the subsequent analysis
lies on the study of problem \eqref{sys:NavierStokesTP_ExteriorDomain}
when $\Omega=\mathbb{R}^3$, namely,
\begin{equation}\label{sys:NavierStokesTP}
\begin{pdeq}
\partial_tu-\Deltau-\lambda\partial_1u+u\cdot\nablau+\nabla\mathfrak{p}&=f
&&\text{in }{\mathbb T}\times\mathbb{R}^3, \\
\Divu&=0
&&\text{in }{\mathbb T}\times\mathbb{R}^3, \\
\lim_{\snorm{x}\to\infty}u(t,x)&=0
&&\text{for } t\in{\mathbb T}\,.
\end{pdeq}
\end{equation}
The case of an exterior domain will be treated
at the end of the last section.
\par
We begin to observe that asymptotic properties of weak solutions to \eqref{sys:NavierStokesTP}
were studied in
\cite{GaldiKyed_TPSolNS3D_AsymptoticProfile}
and \cite{Eiter_SpatiallyAsymptoticStructureTPNS_2020},
where the following
decay estimates of $u$ and $\nablau$ were derived.
\begin{thm}\label{thm:VelocityDecay}
Let $\lambda>0$ and $f\in\LR{q}({\mathbb T}\times\mathbb{R}^3)^3$ for all $q\in(1,\infty)$
and let $\supp f$ be compact.
Let $u$ be a weak solution to \eqref{sys:NavierStokesTP} that satisfies \eqref{el:VelocityAdditionalIntegrability}
for $\Omega=\mathbb{R}^3$.
Then there is $ C_2>0$
such that for all $\np{t,x}\in{\mathbb T}\times\mathbb{R}^3$
the function $u$ satisfies
\begin{align}
\snorm{\calpu(x)}
&\leq C_2\bb{\bp{1+\snorm{x}}\bp{1+\wakefct{\lambda x}}}^{-1},
\label{est:VelocityDecay_fct_ss}
\\
\snorm{\nabla\calpu(x)}
&\leq C_2\bb{\bp{1+\snorm{x}}\bp{1+\wakefct{\lambda x}}}^{-\frac{3}{2}},
\label{est:VelocityDecay_grad_ss}
\\
\snorm{\calp_\botu(t,x)}
&\leq C_2\bp{1+\snorm{x}}^{-3},
\label{est:VelocityDecay_fct_pp}
\\
\snorm{\nabla\calp_\botu(t,x)}
&\leq C_2\bp{1+\snorm{x}}^{-4}.
\label{est:VelocityDecay_grad_pp}
\end{align}
\end{thm}
\begin{proof}
Under the assumption $f\in\CR \infty_0({\mathbb T}\times\mathbb{R}^3)^3$,
this result was shown in \cite[Theorem 4.5]{Eiter_SpatiallyAsymptoticStructureTPNS_2020}
based on estimates of the velocity field $u$ derived in \cite{GaldiKyed_TPSolNS3D_AsymptoticProfile}.
However, a careful study of the proofs shows
that these results continue to be valid under the stated weaker assumption on $f$.
\end{proof}
To derive a suitable fixed-point equation,
we exploit the following representation formulas
that result from the time-periodic fundamental solutions introduced in the previous section.
\begin{prop}\label{prop:RepresentationVorticity}
Let $u$ be a weak solution as in Theorem \ref{thm:VelocityDecay}.
Then
\begin{equation}\label{eq:RepresentationVelocity_CurluWedgeu}
D_x^\alphau=D_x^\alpha\varGamma^\lambda\ast\nb{f-\curlu\wedgeu}
\end{equation}
for all $\alpha\in\mathbb{N}_0^3$ with $\snorm{\alpha}\leq1$.
In particular, the steady-state part
$v:=\calpu$ and the purely periodic part $w:=\calp_\botu$ satisfy
\begin{align}
D_x^\alphav&=D_x^\alpha\fundsolvel_0\ast\bb{\calp f - \curlv\wedgev
-\calp\np{\curlw\wedgew}},
\label{eq:RepresentationVelocity_CurluWedgeu_ss}
\\
D_x^\alphaw&=D_x^\alpha\fundsolvel_\perp\ast
\bb{\calp_\bot f -\curlv\wedgew -\curlw\wedgev+\calp_\bot\np{\curlw\wedgew}}.
\label{eq:RepresentationVelocity_CurluWedgeu_pp}
\end{align}
Moreover, we have
\begin{equation}\label{eq:RepresentationVorticity_CurluWedgeu}
\curl u(t,x)
=\int_{G} \nabla\phi^\lambda(t-s,x-y) \wedge \bb{f-\curlu\wedgeu}(s,y)\,{\mathrm d}(s,y),
\end{equation}
and
\begin{equation}
\curlv(x)
=\int_{\mathbb{R}^3} \nabla\fundsolvort_0(x-y) \wedge \bb{\calp f - \curlv\wedgev
-\calp\np{\curlw\wedgew}}(y)\,{\mathrm d}y,
\label{eq:RepresentationVorticity_CurluWedgeu_ss}
\end{equation}
as well as
\begin{equation}
\begin{aligned}
\curlw(t,x)
&=\int_{G} \nabla\fundsolvort_\perp(t-s,x-y) \wedge \bb{\calp_\bot f -\curlv\wedgew \\
&\qquad\qquad\qquad-\curlw\wedgev-\calp_\bot\np{\curlw\wedgew}}(s,y)\,{\mathrm d}(s,y).
\end{aligned}
\label{eq:RepresentationVorticity_CurluWedgeu_pp}
\end{equation}
\end{prop}
\begin{proof}
Since $u\cdot\nablau=\frac{1}{2}\nabla\bp{\snorm{u}^2}+\curlu\wedgeu$ and
$\varGamma^\lambda\ast\nabla\bp{\snorm{u}^2}=\Div\bp{\varGamma^\lambda\ast\snorm{u}^2}=0$,
the equations \eqref{eq:RepresentationVelocity_CurluWedgeu},
\eqref{eq:RepresentationVelocity_CurluWedgeu_ss} and \eqref{eq:RepresentationVelocity_CurluWedgeu_pp}
are direct consequences of \cite[Proposition 4.8]{Eiter_SpatiallyAsymptoticStructureTPNS_2020}.
The remaining identities follow by
applying the $\curl$ operator to both sides of these formulas and repeating the computations from Section \ref{sec:TPFundSol}.
\end{proof}
\begin{rem}\label{rem:VortDecay_pp_NotAsInLinear}
In view of Proposition \ref{prop:RepresentationVorticity}
and the pointwise estimates of $\phi^\lambda$ from
Theorem \ref{thm:FundsolVortss_est} and Theorem \ref{thm:FundsolVortpp_est},
we can explain, at this point, the origin of the pointwise estimates
stated in Theorem \ref{thm:VorticityDecay}.
Comparing \eqref{est:VorticityDecay_ss} and \eqref{est:FundsolVortss_grad},
we see that the asserted decay rates of the steady-state parts $\curlv$
and $\nabla\fundsolvort_0$ coincide,
which is the optimal result one can expect to derive
from equation \eqref{eq:RepresentationVorticity_CurluWedgeu_ss}
for general $f\in\CR \infty_0(G)^3$.
In contrast, the asserted decay rates of the purely periodic parts
$\curlw$ and $\fundsolvel_\perp$ given in
\eqref{est:VorticityDecay_pp} and \eqref{est:FundsolVortpp_grad}, respectively,
do not coincide.
The reason is due to the presence of the term $\curlv\wedgew$
in \eqref{eq:RepresentationVorticity_CurluWedgeu_pp}.
By assuming the---to some extent---optimal decay rate \eqref{est:VorticityDecay_ss} for $\curlv$,
the pointwise estimate of $w$ from \eqref{est:VelocityDecay_fct_pp} implies
\[
\snorm{\curlv\wedgew}(t,x)\leq C \snorm{x}^{-9/2}\e^{- \alpha \wakefct{\lambda x}}.
\]
In the end, this term dominates the decay of the right-hand side
of \eqref{eq:RepresentationVorticity_CurluWedgeu_pp}
and thus the pointwise estimates of $\curlw$. As a result,
the decay of $\curlw$ is slower than that of $\nabla\fundsolvort_\perp$ but, however, still faster than the decay rate of the steady-state vorticity field $\curlv$.
\end{rem}
Proposition \ref{prop:RepresentationVorticity} yields fixed-point equations for $u$
and $\curlu$ and the respective steady-state and purely periodic parts,
which we now decompose in an appropriate way.
Let $\chi\in\CR \infty_0(\mathbb{R};[0,1])$
with $\chi(s)=1$ for $\snorm{s}\leq 5/4$
and $\chi(s)=0$ for $\snorm{s}\geq 7/4$.
For $S>0$ define
$\chi_S\in\CR \infty_0(\mathbb{R}^3;[0,1])$ by $\chi_{S}(x):=\chi(S^{-1}\snorm{x})$,
and fix $S_0>0$ such that $\supp f\subset {\mathbb T}\times\mathrm{B}_{S_0}$.
For $S\in[2S_0,\infty)$
we express \eqref{eq:RepresentationVelocity_CurluWedgeu}
as the sum of two terms, namely
\[
u
=\varGamma^\lambda\ast\bb{-\np{1-\chi_S}\curlu\wedgeu}
+\varGamma^\lambda\ast\bb{f-\chi_S\curlu\wedgeu}.
\]
Due to $\supp\np{1-\chi_S}\subset\mathrm{B}^S$,
this yields
\begin{align}\label{eq:uFixedPoint_VortDecay}
\restriction{u}{{\mathbb T}\times\mathrm{B}^S}={\mathcal F}_S(\restriction{u}{{\mathbb T}\times\mathrm{B}^S})+{\mathcal H}_S,
\end{align}
where
\[
\begin{aligned}
{\mathcal F}_S(z)
&:=\restriction{\bp{\varGamma^\lambda\ast\bb{-\np{1-\chi_S}\curlz\wedgez}}}{{\mathbb T}\times\mathrm{B}^S}, \\
{\mathcal H}_S
&:=\restriction{\bp{\varGamma^\lambda\ast\bb{f-\chi_S\curlu\wedgeu}}}{{\mathbb T}\times\mathrm{B}^S}.
\end{aligned}
\]
We set $\cala(z):=-\curlz\wedgez$ and
\begin{align}
\cala_0(z)
:=\calp\cala(z)
&=-\curlz_0\wedgez_0-\calp\np{\curlz_\perp\wedgez_\perp},
\label{eq:nonlinCurl_ss}\\
\cala_\perp(z)
:=\calp_\bot\cala(z)
&=-\curlz_0\wedgez_\perp-\curlz_\perp\wedgez_0
-\calp_\bot\np{\curlz_\perp\wedgez_\perp},
\label{eq:nonlinCurl_pp}
\end{align}
with $z_0:=\calpz$ and $z_\perp:=\calp_\botz$.
For $(t,x)\in{\mathbb T}\times\mathrm{B}^S$ from Proposition \ref{prop:RepresentationVorticity} we then obtain
\begin{align}
D_x^\alpha\calp{\mathcal F}_S(z)(x)
&=D_x^\alpha\fundsolvel_0\ast\bb{\np{1-\chi_S}
\cala_0(z)}(x),
\label{eq:Representation_Fss}\\
D_x^\alpha\calp_\bot{\mathcal F}_S(z)(t,x)
&=D_x^\alpha\fundsolvel_\perp\ast\bb{\np{1-\chi_S}\cala_\perp(z)}(t,x),
\label{eq:Representation_Fpp}\\
\curl\calp{\mathcal F}_S(z)(x)
&=\int_{\mathbb{R}^3} \nabla\fundsolvort_0(x-y) \wedge
\bb{\np{1-\chi_S}\cala_0(z)}(y)\,{\mathrm d}y,
\label{eq:Representation_CurlFss}\\
\curl\calp_\bot{\mathcal F}_S(z)(t,x)
&=\int_{{\mathbb T}\times\mathbb{R}^3} \nabla\fundsolvort_\perp(t-s,x-y)
\wedge\bb{\np{1-\chi_S}\cala_\perp(z)}(s,y)\,{\mathrm d}(s,y),
\label{eq:Representation_CurlFpp}
\end{align}
and
\begin{align}
D_x^\alpha\calp{\mathcal H}_S(x)
&=D_x^\alpha\fundsolvel_0\ast\bb{\calp f+\chi_S\cala_0(u)}(x),
\label{eq:Representation_Hss}\\
D_x^\alpha\calp_\bot{\mathcal H}_S(t,x)
&=D_x^\alpha\fundsolvel_\perp\ast\bb{\calp_\bot f+\chi_S\cala_\perp(u)}(t,x),
\label{eq:Representation_Hpp}\\
\curl\calp{\mathcal H}_S(x)
&=\int_{\mathbb{R}^3} \nabla\fundsolvort_0(x-y) \wedge
\bb{\calp f+\chi_S\cala_0(u)}(y)\,{\mathrm d}y,
\label{eq:Representation_CurlHss}\\
\curl\calp_\bot{\mathcal H}_S(t,x)
&=\int_{{\mathbb T}\times\mathbb{R}^3} \nabla\fundsolvort_\perp(t-s,x-y)
\wedge\bb{\calp_\bot f+\chi_S\cala_\perp(u)}(s,y)\,{\mathrm d}(s,y).
\label{eq:Representation_CurlHpp}
\end{align}
In the next step we introduce the functional framework
for the analysis of the fixed-point equation \eqref{eq:uFixedPoint_VortDecay}.
Let $\varepsilon\in\np{0,\frac{1}{4}}$ and fix a radius $S>S_0$.
We define the following (semi-)norms,
which take into account different decay rates of the steady-state and the purely periodic parts:
\begin{align*}
\velnorm{z}
&:=
\esssup_{x\in\mathrm{B}^{S}}\Bb{\snorm{x}\np{1+\wakefct{x}}\snorm{\calpz(x)}
+\bb{\snorm{x}\np{1+\wakefct{x}}}^{3/2}\snorm{\nabla\calpz(x)}}
\\
&\qquad\qquad\qquad\qquad
+\esssup_{\np{t,x}\in{\mathbb T}\times\mathrm{B}^{S}}\Bb{\snorm{x}^3\snorm{\calp_\botz(t,x)}
+\snorm{x}^4\snorm{\nabla\calp_\botz(t,x)}},
\\
\vortnorm{z}
&:=\esssup_{x\in\mathrm{B}^{S}}\snorm{x}^{3/2}
\e^{\frac{\wakefct{K x}}{1+S}}\snorm{\curl\calpz(x)}
\\
&\qquad\qquad\qquad\qquad
+\esssup_{\np{t,x}\in{\mathbb T}\times\mathrm{B}^{S}}\snorm{x}^{9/2-\varepsilon}
\e^{\frac{\wakefct{K x}}{1+ S}}\snorm{\curl\calp_\botz(t,x)},
\end{align*}
where $K:=\frac{1}{4}\min\set{\lambda, C_3}$
with $ C_3$ from Theorem \ref{thm:FundsolVortpp_est}.
The function spaces associated to these (semi-)norms are given by
\begin{align*}
\calm_S&:=\setcl{z\in\WSRloc{1}{1}({\mathbb T}\times\mathrm{B}^S)}{\velnorm{z}<\infty},
\\
\caln_S^\varepsilon&:=\setcl{z\in\calm_S}{\vortnorm{z}<\infty},
\end{align*}
which are Banach spaces with respect to the norms
\[
\norm{z}_{\calm_S}:=\velnorm{z},
\qquad
\norm{z}_{\caln_S^\varepsilon}:=\velnorm{z}+\vortnorm{z},
\]
respectively.
\begin{rem}
Let us explain the terms appearing in these definitions.
The definition of $\velnorm{z}$ is chosen to capture the asymptotic behavior of $u$ and $\nablau$
described in Theorem \ref{thm:VelocityDecay}.
A justification for the denominator $1+S$
in the exponential term in the definition of $\vortnorm{z}$
is given by
Lemma \ref{lem:Hs_EstExpFct} below.
The choice of the constant $K$ ensures the validity of the inequalities
\begin{align}\label{est:Kconst}
\e^{2\wakefct{K x}}
\leq \e^{\wakefct{\lambda x}/2}, \qquad\qquad
\e^{2\wakefct{K x}}
\leq \e^{ C_3\snorm{x}},
\end{align}
so that the exponential term can be related with the exponential terms
in the decay rates of $\nabla\fundsolvort_0$ and $\nabla\fundsolvort_\perp$
from Theorem \ref{thm:FundsolVortss_est} and Theorem \ref{thm:FundsolVortpp_est}, respectively.
Moreover, the second term in the definition of $\vortnorm{z}$
contains the factor $\snorm{x}^{9/2-\varepsilon}$ instead of $\snorm{x}^{9/2}$,
which one would expect, in view of the asserted estimate \eqref{est:VorticityDecay_pp}.
Later on we shall see that this discrepancy is necessary to ensure that ${\mathcal F}_S$
is a contraction in the underlying function space.
\end{rem}
\section{Estimates}
\label{sec:Estimates}
In this section, we collect estimates of ${\mathcal H}_S$ and ${\mathcal F}_S(z)$
with respect to the (semi-)norms introduced above,
which ensure that $z\mapsto{\mathcal F}_S(z)+{\mathcal H}_S$
is a contractive self-mapping when we choose $S$ sufficiently large.
We begin with the following elementary lemma,
which explains the term $1+S$
in the definition of $\vortnorm{z}$.
\begin{lem}\label{lem:Hs_EstExpFct}
Let $a,\,S>0$.
If $x,y\in\mathbb{R}^3$ with $\snorm{y}\leq2S$, then
\begin{align}
\e^{-\wakefct{a\np{x-y}}}
&\leq\e^{4a}\e^{-\frac{\wakefct{a x}}{1+S}},
\label{est:Hs_vortest_ExpWakefct}
\\
\e^{-a\snorm{x-y}}
&\leq\e^{2a}\e^{-\frac{a\snorm{x}}{1+S}}.
\label{est:Hs_vortest_ExpAbs}
\end{align}
\end{lem}
\begin{proof}
For $\snorm{y}\leq2S$ we have
$\wakefct{a y}/\np{1+S}\leq 2a\snorm{y}/\np{1+S}\leq4a$.
Together with $\wakefct{a\np{x-y}}\geq\wakefct{ax}-\wakefct{ay}$, this implies
\[
\e^{-\wakefct{a\np{x-y}}}\leq\e^{-\frac{\wakefct{a\np{x-y}}}{1+S}}
\leq\e^{-\frac{\wakefct{a x}}{1+S}}\e^{\frac{\wakefct{a y}}{1+S}}
\leq\e^{-\frac{\wakefct{a x}}{1+S}}\e^{4a}.
\]
Similarly, we have
$\snorm{y}/(1+S)\leq 2$,
which implies
\[
\e^{-a\snorm{x-y}}\leq\e^{-\frac{a\snorm{x-y}}{1+S}}
\leq\e^{-\frac{a\snorm{x}}{1+S}}\e^{\frac{a\snorm{y}}{1+S}}
\leq\e^{-\frac{a\snorm{x}}{1+S}}\e^{2a}.
\]
This completes the proof.
\end{proof}
We further employ the following lemma in order to estimate convolutions
of functions with anisotropic decay behavior.
\begin{lem}\label{lem:ConvEst_WakeFct_ThreeHalfs}
Let $A\in(2,\infty)$, $B\in[0,\infty)$ with $A+\min\set{1,B}>3$.
Then there exists $C=C\np{A,B}>0$ such that
for all $x\in\mathbb{R}^3$ it holds
\[
\int_{\mathbb{R}^3}\bb{\np{1+\snorm{x-y}}\np{1+\wakefct{x-y}}}^{-3/2}
\np{1+\snorm{y}}^{-A}\np{1+\wakefct{y}}^{-B}\,{\mathrm d}y
\leq C\np{1+\snorm{x}}^{-3/2}.
\]
\end{lem}
\begin{proof}
See \cite[Theorem 5]{DeuringGaldi_ExpDecayVorticity_2016}.
\end{proof}
The next lemma
treats convolutions of functions that are homogeneous in space.
\begin{lem}\label{lem:ConvolutionExp}
Let $A\in(0,3)$, $B\in(0,\infty)$, $\alpha\in(0,\infty)$.
Then there exists a constant $C=C\np{A,B,\alpha}>0$
such that for all $x\in\mathbb{R}^3$ it holds
\[
\int_{\mathbb{R}^3}\snorm{x-y}^{-A}
\e^{-\alpha\snorm{x-y}}\np{1+\snorm{y}}^{-B}\,{\mathrm d}y
\leq C\np{1+\snorm{x}}^{-B}.
\]
\end{lem}
\begin{proof}
For $x=0$ the integral is finite,
so that it remains to consider $x\neq0$.
We split the integral into two parts
\begin{align*}
I_1
&:=\int_{\mathrm{B}_{\snorm{x}/2}(x)}\snorm{x-y}^{-A}
\e^{-\alpha\snorm{x-y}}\np{1+\snorm{y}}^{-B}\,{\mathrm d}y,\\
I_2
&:=\int_{\mathrm{B}^{\snorm{x}/2}(x)}\snorm{x-y}^{-A}
\e^{-\alpha\snorm{x-y}}\np{1+\snorm{y}}^{-B}\,{\mathrm d}y,
\end{align*}
which we estimate separately.
On the one hand,
since $\snorm{x-y}\leq\snorm{x}/2$ implies $\snorm{y}\geq\snorm{x}-\snorm{x-y}\geq\snorm{x}/2$,
we have
\[
I_1
\leq C\np{1+\snorm{x}}^{-B}\int_{\mathbb{R}^3}\snorm{x-y}^{-A}\e^{-\alpha\snorm{x-y}}\,{\mathrm d}y
\leq C\np{1+\snorm{x}}^{-B},
\]
where the integral is finite due to $A<3$.
On the other hand, we obtain
\[
I_2
\leq C\e^{-\alpha\snorm{x}/4}
\int_{\mathbb{R}^3}\e^{-\alpha\snorm{x-y}/2}\,{\mathrm d}y
\leq C\e^{-\alpha\snorm{x}/4}
\leq C\np{1+\snorm{x}}^{-B}.
\]
This completes the proof.
\end{proof}
Since our assumptions do not provide pointwise information
on the body force $f$,
we estimate the convolutions of the fundamental
solutions with $f$ in a different way,
which leads to the following lemma.
\begin{lem}\label{lem:Conv_fLp}
There exists a constant $C>0$ such that
for $\alpha\in\mathbb{N}_0$, $\snorm{\alpha}\leq1$,
we have
\begin{align}
\snorml{D_x^\alpha\fundsolvel_0\ast{\calp f}(x)}
&\leq C\bb{\snorm{x}\bp{1+\wakefct{\lambda x}}}^{-1-\frac{\snorm{\alpha}}{2}},
\label{est:Conv_fLp_velss}
\\
\snorml{D_x^\alpha\fundsolvel_\perp\ast{\calp_\bot f}(t,x)}
&\leq C\snorm{x}^{-3-\snorm{\alpha}},
\label{est:Conv_fLp_velpp}
\\
\snorml{\int_{\mathbb{R}^3} \nabla\fundsolvort_0(x-y) \wedge {\calp f}(y)\,{\mathrm d}y}
&\leq C\snorm{x}^{-3/2}\e^{-\frac{\wakefct{\lambda x}}{4\np{1+S_0}}},
\label{est:Conv_fLp_vortss}
\\
\snorml{\int_{{\mathbb T}\times\mathbb{R}^3} \nabla\fundsolvort_\perp(t-s,x-y)
\wedge{\calp_\bot f}(s,y)\,{\mathrm d}(s,y)}
&\leq C\snorm{x}^{-9/2}\e^{-\frac{C_3\snorm{x}}{2\np{1+S_0}}}.
\label{est:Conv_fLp_vortpp}
\end{align}
for all $t\in{\mathbb T}$ and $\snorm{x}\geq 2S_0$.
\end{lem}
\begin{proof}
For $\snorm{x}\geq 2S_0\geq 2\snorm{y}$
we have
\[
(1+2\lambdaS_0)\np{1+\lambda\wakefct{x-y}}
\geq 1+\lambda\wakefct{x}+2\lambdaS_0-\lambda\wakefct{y}
\geq 1+\lambda\wakefct{x}
\]
and $\snorm{x-y}\geq\snorm{x}-\snorm{y}\geq\snorm{x}/2\geq S_0$.
Therefore, \eqref{est:fundsolss_Decay} and
$\supp f \subset {\mathbb T}\times\mathrm{B}_{S_0}$
imply
\[
\begin{aligned}
\snorml{D_x^\alpha\fundsolvel_0\ast{\calp f}(x)}
&\leq C\int_{\mathrm{B}_{S_0}}
\bb{\snorm{x-y}\bp{1+\lambda\wakefct{x-y}}}^{-1-\frac{\snorm{\alpha}}{2}}
\snorm{\calp f(y)}\,{\mathrm d}y
\\
&\leq C
\bb{\snorm{x}\bp{1+\lambda\wakefct{x}}}^{-1-\frac{\snorm{\alpha}}{2}}
\int_{\mathrm{B}_{S_0}} \snorm{\calp f(y)}\,{\mathrm d}y,
\end{aligned}
\]
which yields \eqref{est:Conv_fLp_velss}.
Using H\"older's inequality in time
and \eqref{est:tpfundsol_ComplPointwiseEst}, for any $q\in(1,\infty)$
we obtain in a similar way
\[
\begin{aligned}
\snorml{D_x^\alpha\fundsolvel_\perp\ast{\calp_\bot f}(t,x)}
&\leq \int_{\mathrm{B}_{S_0}}
\Bp{\int_{\mathbb T} \snorm{D_x^\alpha\fundsolvel_\perp(s,x-y)}^{\frac{q}{q-1}}\,{\mathrm d}s}^{\frac{q-1}{q}}
\Bp{\int_{\mathbb T}\snorm{\calp_\bot f(s,y)}^q\,{\mathrm d}s}^{1/q}\,{\mathrm d}y
\\
&\leq C\int_{\mathrm{B}_{S_0}}
\snorm{x-y}^{-3-\snorm{\alpha}}
\Bp{\int_{\mathbb T}\snorm{\calp_\bot f(s,y)}^q\,{\mathrm d}s}^{1/q}\,{\mathrm d}y
\\
&\leq C
\snorm{x}^{-3-\snorm{\alpha}}
\Bp{\int_{\mathbb T}\int_{\mathrm{B}_{S_0}}\snorm{\calp_\bot f(s,y)}^q\,{\mathrm d}s{\mathrm d}y}^{1/q},
\end{aligned}
\]
which shows \eqref{est:Conv_fLp_velpp}.
In virtue of the estimates \eqref{est:FundsolVortss_grad}
and \eqref{est:FundsolVortpp_grad} and
Lemma \ref{lem:Hs_EstExpFct},
for $\snorm{x}\geq2S_0\geq2\snorm{y}$ we further derive
\[
\begin{aligned}
\snorm{\nabla\fundsolvort_0(x-y)}
&\leq C\snorm{x-y}^{-3/2}
\e^{-\wakefct{\lambda\np{x-y}}/4}
\leq C \snorm{x}^{-3/2}\e^{-\frac{\wakefct{\lambda x}}{4\np{1+S_0}}},
\\
\norm{\nabla\fundsolvort_\perp(\cdot,x-y)}_{\LR{q}({\mathbb T})}
&\leq C\snorm{x-y}^{-9/2}\e^{- C_3\snorm{x-y}/2}
\leq C \snorm{x}^{-9/2}\e^{-\frac{-C_3\snorm{x}}{2\np{1+S_0}}}.
\end{aligned}
\]
From these estimates
we conclude \eqref{est:Conv_fLp_vortss} and \eqref{est:Conv_fLp_vortpp}
with the same argument as above.
\end{proof}
After these preparations,
we show in the next two lemmas that the norm of ${\mathcal H}_S$
in both $\calm_S$ and $\caln_S^\varepsilon$ is bounded by a constant
independent of $S\geq 2S_0$.
\begin{lem}\label{lem:Hs_velest}
There exists a constant $ C_6>0$ such that for all $S\in[2S_0,\infty)$
we have
\[
\velnorm{{\mathcal H}_S}\leq C_6.
\]
\end{lem}
\begin{proof}
From the decay estimates of $u$ and $\nablau$ from
Theorem \ref{thm:VelocityDecay}
we conclude
\begin{align}
\snorml{\chi_S(x) \cala_0\np{u}(x)}
&\leq C \bb{\np{1+\snorm{x}}\np{1+\wakefct{x}}}^{-5/2},
\label{est:Hs_rhsss}\\
\snorml{\chi_S(x) \cala_\perp\np{u}(t,x)}
&\leq C \np{1+\snorm{x}}^{-9/2}.
\label{est:Hs_rhspp}
\end{align}
By Theorem \ref{thm:ConvFundsolss} and Theorem \ref{thm:ConvFundsolpp},
these estimates and the formulas \eqref{eq:Representation_Hss}
and \eqref{eq:Representation_Hpp}
together with Lemma \ref{lem:Conv_fLp}
imply
\[
\begin{aligned}
\snorm{\calp{\mathcal H}_S(x)}
&\leq C \bb{\np{1+\snorm{x}}\np{1+\wakefct{x}}}^{-1},
\\
\snorm{\nabla\calp{\mathcal H}_S(x)}
&\leq C \bb{\np{1+\snorm{x}}\np{1+\wakefct{x}}}^{-3/2},
\\
\snorm{\calp_\bot{\mathcal H}_S(t,x)}
&\leq C \np{1+\snorm{x}}^{-3},
\\
\snorm{\nabla\calp_\bot{\mathcal H}_S(t,x)}
&\leq C \np{1+\snorm{x}}^{-4}
\end{aligned}
\]
for all $t\in{\mathbb T}$ and $\snorm{x}\geqS_0$.
Collecting these, we arrive at the claimed estimate.
\end{proof}
\begin{lem}\label{lem:Hs_vortest}
There exists a constant $ C_7>0$ such that for all $S\in[2S_0,\infty)$ we have
\[
\vortnorm{{\mathcal H}_S}\leq C_7.
\]
\end{lem}
\begin{proof}
At first, let $x\in\mathbb{R}^3$ with $\snorm{x}\geq2S$.
For $\snorm{y}\leq7S/4$ we have
\[
\snorm{x-y}\geq\snorm{x}-\snorm{y}\geq\snorm{x}-7S/4
\geq\snorm{x}-7\snorm{x}/8
=\snorm{x}/8
\geqS/4\geqS_0/2.
\]
From \eqref{est:FundsolVortss_grad} and Lemma \ref{lem:Hs_EstExpFct}, we then conclude
\begin{align*}
\snorm{\nabla\fundsolvort_0(x-y)}
&\leq C \bp{\snorm{x-y}^{-2}+\snorm{x-y}^{-3/2}\wakefct{\lambda \np{x-y}}^{1/2}}
\e^{-\wakefct{\lambda \np{x-y}}/2}\\
&\leq C \bp{1+\snorm{x-y}^{-3/2}\bp{1+\wakefct{\lambda\np{x-y}}}^{-3/2}}
\e^{-\wakefct{\lambda \np{x-y}}/4}\\
&\leq C \bb{\np{1+\snorm{x-y}}\bp{1+\wakefct{\lambda\np{x-y}}}}^{-3/2}
\e^{-\frac{\wakefct{\lambda x}}{4\np{1+S}}}.
\end{align*}
In virtue of
\eqref{eq:Representation_CurlHss}, \eqref{est:Conv_fLp_vortss}
and \eqref{est:Hs_rhsss}
we thus obtain
\begin{align*}
&\snorm{\curl\calp{\mathcal H}_S(x)}
\leq C\int_{\mathrm{B}_{7S/4}} \snorml{\nabla\fundsolvort_0(x-y)} \,
\snorml{\calp f+\chi_S\cala_0(u)}(y)\,{\mathrm d}y\\
&\leq C\snorm{x}^{-3/2}\e^{-\frac{\wakefct{\lambda x}}{4\np{1+S_0}}} \\
&\qquad + C \e^{-\frac{\wakefct{\lambda x}}{4\np{1+S}}}
\int_{\mathbb{R}^3}\bb{\bp{1+\snorm{x-y}}\bp{1+\wakefct{x-y}}}^{-3/2}
\bb{\np{1+\snorm{y}}\np{1+\wakefct{y}}}^{-5/2}\,{\mathrm d}y
\end{align*}
for $\snorm{x}\geq2S\geq4S_0$.
By estimating the remaining integral with the help of Lemma \ref{lem:ConvEst_WakeFct_ThreeHalfs}
and employing \eqref{est:Kconst}, we deduce
\begin{equation}\label{est:Hs_vortest_ss}
\snorm{\curl\calp{\mathcal H}_S(x)}
\leq C \e^{-\frac{\wakefct{K x}}{1+S}}\snorm{x}^{-3/2}
\end{equation}
for $\snorm{x}\geq2S$.
If $S\leq\snorm{x}\leq2S$,
then Lemma \ref{lem:Hs_velest} yields
\[
\snorm{\curl\calp{\mathcal H}_S(x)}
\leq C \snorm{\nabla\calp{\mathcal H}_S(x)}
\leq C \bb{\np{1+\snorm{x}}\np{1+\wakefct{x}}}^{-3/2}
\leq C \snorm{x}^{-3/2}.
\]
Since $\snorm{x}\leq2S$ implies
$
\wakefct{K x}/\np{1+S}
\leq 2\snorm{K x}/\np{1+S}
\leq 4\KconstS/\np{1+S}
\leq 4K,
$
we have
$1\leq\e^{4K}\e^{-\wakefct{K x}/\np{1+S}}$,
so that \eqref{est:Hs_vortest_ss} also holds for $S\leq\snorm{x}\leq2S$.
Now let us turn to $\curl\calp_\bot{\mathcal H}_S$.
From \eqref{est:FundsolVortpp_grad}
and \eqref{est:Hs_vortest_ExpAbs}, for $\snorm{y}\leq2S$
we conclude
\[
\int_{\mathbb T}\snorm{\nabla\fundsolvort_\perp(t-s,x-y)}\,{\mathrm d}s
\leq C \snorm{x-y}^{-5/2} \e^{-\frac{ C_3\snorm{x-y}}{2}}
\e^{-\frac{ C_3\snorm{x}}{2\np{1+S}}},
\]
so that \eqref{eq:Representation_CurlHpp}, \eqref{est:Conv_fLp_vortpp}
and \eqref{est:Hs_rhspp} lead to
\begin{align*}
&\snorm{\curl\calp_\bot{\mathcal H}_S(t,x)}
\leq C \int_{\mathrm{B}_{7S/4}}\int_{\mathbb T}\snorml{\nabla\fundsolvort_\perp(t-s,x-y)}
\,\snorml{\calp_\bot f+\chi_S\cala_\perp\np{u}}(s,y) \,{\mathrm d}s{\mathrm d}y\\
&\quad\qquad
\leq C\snorm{x}^{-9/2}\e^{-\frac{C_3\snorm{x}}{2\np{1+S_0}}}
+C \e^{-\frac{ C_3\snorm{x}}{2\np{1+S}}}
\int_{\mathbb{R}^3}\snorm{x-y}^{-5/2}\e^{- C_3\snorm{x-y}/2}
\np{1+\snorm{y}}^{-9/2}\,{\mathrm d}y.
\end{align*}
The remaining integral can be estimated with Lemma \ref{lem:ConvolutionExp}.
Further using \eqref{est:Kconst}, we end up with
\[
\snorm{\curl\calp_\bot{\mathcal H}_S(t,x)}
\leq C \e^{-\frac{ C_3\snorm{x}}{2\np{1+S}}}
\snorm{x}^{-9/2}
\leq C \e^{-\frac{\wakefct{K x}}{1+S}}
\snorm{x}^{-9/2+\varepsilon}
\]
for $\snorm{x}\geqS\geq2S_0$ and $t\in{\mathbb T}$.
A combination of this estimate with \eqref{est:Hs_vortest_ss} finishes the proof.
\end{proof}
In the next two lemmas we provide appropriate estimates of ${\mathcal F}_S(z)$.
Observe that, in contrast to ${\mathcal H}_S$, this term depends on the (unknown) function
$z$.
In order to eventually obtain a contraction for large $S$,
we factor out the term $S^{-\varepsilon}$ in the estimates.
\begin{lem}\label{lem:Fs_velest}
There exists a constant $ C_8>0$ such that for all $S\in[2S_0,\infty)$
and all $z_1,z_2\in\calm_S$ we have
\begin{align}
\velnorm{{\mathcal F}_S(z_1)}
&\leq \CFsvelnormS^{-\varepsilon}\velnorm{z_1}^2,
\label{est:Fs_quadratic}
\\
\velnorm{{\mathcal F}_S(z_1)-{\mathcal F}_S(z_2)}
&\leq \CFsvelnormS^{-\varepsilon}\bp{\velnorm{z_1}+\velnorm{z_2}}
\velnorm{z_1-z_2}.
\label{est:Fs_contraction}
\end{align}
\end{lem}
\begin{proof}
For $z\in\calm_S$
we immediately deduce
\begin{align*}
\snorml{\np{1-\chi_S(x)}\cala_0(z)(x)}
&\leq C \velnorm{z}^2\np{1-\chi_S(x)}\bb{\np{1+\snorm{x}}\np{1+\wakefct{x}}}^{-5/2} \\
&\leq C S^{-\varepsilon}\velnorm{z}^2\np{1+\snorm{x}}^{-5/2+\varepsilon}\np{1+\wakefct{x}}^{-5/2},
\\
\snorml{\np{1-\chi_S(x)} \cala_\perp\np{z}(t,x)}
&\leq C \velnorm{z}^2\np{1-\chi_S(x)}\np{1+\snorm{x}}^{-9/2}\\
&\leq C S^{-\varepsilon}\velnorm{z}^2\np{1+\snorm{x}}^{-9/2+\varepsilon}
\end{align*}
for $\snorm{x}\geqS$.
By Theorem \ref{thm:ConvFundsolss} and Theorem \ref{thm:ConvFundsolpp},
from these estimates and the formulas \eqref{eq:Representation_Fss}
and \eqref{eq:Representation_Fpp}
we conclude
\[
\begin{aligned}
\snorm{\calp{\mathcal F}_S(z)(x)}
&\leq C S^{-\varepsilon}\velnorm{z}^2\bb{\np{1+\snorm{x}}\np{1+\wakefct{x}}}^{-1},
\\
\snorm{\nabla\calp{\mathcal F}_S(z)(x)}
&\leq C S^{-\varepsilon}\velnorm{z}^2\bb{\np{1+\snorm{x}}\np{1+\wakefct{x}}}^{-3/2},
\\
\snorm{\calp_\bot{\mathcal F}_S(z)(t,x)}
&\leq C S^{-\varepsilon}\velnorm{z}^2\np{1+\snorm{x}}^{-3},
\\
\snorm{\nabla\calp_\bot{\mathcal F}_S(z)(t,x)}
&\leq C S^{-\varepsilon}\velnorm{z}^2\np{1+\snorm{x}}^{-4}.
\end{aligned}
\]
Collecting these estimates, we obtain \eqref{est:Fs_quadratic}.
The inequality \eqref{est:Fs_contraction} is proved in the same fashion.
\end{proof}
\begin{lem}\label{lem:Fs_vortest}
There exists a constant $ C_9>0$ such that for all $S\in[2S_0,\infty)$
and all $z_1,z_2\in\caln_S^\varepsilon$ we have
\begin{align}
\vortnorm{{\mathcal F}_S(z_1)}&\leq \CFsvortnormS^{-\varepsilon}\velnorm{z_1}\vortnorm{z_1},
\label{est:Fs_curlquadratic}\\
\vortnorm{{\mathcal F}_S(z_1)-{\mathcal F}_S(z_2)}
&\leq \CFsvortnormS^{-\varepsilon}
\bp{\norm{z_1}_{\caln_S^\varepsilon}+\norm{z_2}_{\caln_S^\varepsilon}}
\norm{z_1-z_2}_{\caln_S^\varepsilon}.
\label{est:Fs_curlcontraction}
\end{align}
\end{lem}
\begin{proof}
For $z\in\caln_S^\varepsilon$ we have
\begin{align}
&\begin{aligned}
\snorml{\np{1-\chi_S(x)}\cala_0(z)(x)}
&\leq C \velnorm{z}\vortnorm{z}\np{1-\chi_S(x)}
\snorm{x}^{-5/2}\np{1+\wakefct{x}}^{-1}\e^{-\frac{\wakefct{K x}}{1+S}} \\
&\leq C S^{-\varepsilon}\velnorm{z}\vortnorm{z}
\snorm{x}^{-5/2+\varepsilon}\np{1+\wakefct{x}}^{-1}\e^{-\frac{\wakefct{K x}}{1+S}},
\end{aligned}
\label{est:FsCurl_rhsss}
\\
&\begin{aligned}
\snorml{\np{1-\chi_S(x)} \cala_\perp\np{z}(t,x)}
&\leq C \velnorm{z}\vortnorm{z}\np{1-\chi_S(x)}\snorm{x}^{-9/2}
\e^{-\frac{\wakefct{K x}}{1+S}}\\
&\leq C S^{-\varepsilon}\velnorm{z}\vortnorm{z}\snorm{x}^{-9/2+\varepsilon}
\e^{-\frac{\wakefct{K x}}{1+S}}
\end{aligned}
\label{est:FsCurl_rhspp}
\end{align}
for $\snorm{x}\geqS$.
Exploiting the representation formula \eqref{eq:Representation_CurlFss},
we can employ \eqref{est:FundsolVortss_grad} and \eqref{est:FsCurl_rhsss} to estimate
\begin{align*}
\snorm{\curl\calp{\mathcal F}_S(z)(x)}
&\leq C\int_{\mathbb{R}^3}\snorml{\nabla\fundsolvort_0(x-y)}\,\snorml{
\np{1-\chi_S(y)}\cala_0(z)(y)}\,{\mathrm d}y\\
&\leq C S^{-\varepsilon}\velnorm{z}\vortnorm{z}\bp{I_1+I_2},
\end{align*}
where
\[
\begin{aligned}
I_1
&:=\int_{\mathrm{B}^S\cap\mathrm{B}_{S_0}(x)}
\snorm{x-y}^{-2}\e^{-\frac{\wakefct{\lambda\np{x-y}}}{4}}
\snorm{y}^{-5/2+\varepsilon}\np{1+\wakefct{y}}^{-1}\e^{-\frac{\wakefct{K y}}{1+S}}\,{\mathrm d}y,
\\
I_2
&:=\int_{\mathrm{B}^S\cap\mathrm{B}^{S_0}(x)}
\bb{\snorm{x-y}\wakefct{\lambda \np{x-y}}}^{-3/2}\e^{-\frac{\wakefct{\lambda\np{x-y}}}{4}} \snorm{y}^{-5/2+\varepsilon}\np{1+\wakefct{y}}^{-1}
\e^{-\frac{\wakefct{K y}}{1+S}}\,{\mathrm d}y.
\end{aligned}
\]
To give estimates of these integrals, we first note that by
$\wakefct{\lambda(x-y)}\geq\wakefct{\lambda x}-\wakefct{\lambda y}$
and \eqref{est:Kconst}, we have
\begin{align}\label{est:Fs_vortest_ExpConv}
\e^{-\frac{\wakefct{\lambda\np{x-y}}}{4}}\e^{-\frac{\wakefct{K y}}{1+S}}
\leq\e^{-\frac{\wakefct{\lambda x}}{4\np{1+S}}}
\e^{\frac{\wakefct{\lambda y}}{4\np{1+S}}}
\e^{-\frac{\wakefct{K y}}{1+S}}
\leq\e^{-\frac{\wakefct{K x}}{1+S}}
\end{align}
for all $x,\,y\in\mathbb{R}^3$.
On the one hand, exploiting this estimate and
that $\snorm{x-y}\leqS_0\leq\snorm{x}/2$ implies
$\snorm{y}\geq\snorm{x}-\snorm{x-y}
\geq\snorm{x}-S_0
\geq\snorm{x}/2$,
we conclude
\[
I_1\leq C \e^{-\frac{\wakefct{K x}}{1+S}}\snorm{x}^{-5/2+\varepsilon}
\int_{\mathrm{B}_{S_0}(x)}\snorm{x-y}^{-2}\,{\mathrm d}y
\leq C \e^{-\frac{\wakefct{K x}}{1+S}}\snorm{x}^{-3/2}
\]
for $\snorm{x}\geqS\geq2S_0$.
On the other hand, due to \eqref{est:Fs_vortest_ExpConv}
and the fact that $\snorm{y}\geqS\geq2S_0$ implies
$\snorm{y}\geq C \np{1+\snorm{y}}$, we obtain
\begin{align*}
I_2
&\leq C
\e^{-\frac{\wakefct{K x}}{1+S}}
\int_{\mathbb{R}^3}\bb{\np{1+\snorm{x-y}}\wakefct{x-y}}^{-3/2}
\np{1+\snorm{y}}^{-5/2+\varepsilon}\np{1+\wakefct{y}}^{-1}\,{\mathrm d}y\\
&\leq C \e^{-\frac{\wakefct{K x}}{1+S}}\snorm{x}^{-3/2}
\end{align*}
by Lemma \ref{lem:ConvEst_WakeFct_ThreeHalfs}.
From the estimates of $I_1$ and $I_2$
we deduce
\[
\snorm{\curl\calp{\mathcal F}_S(z)(x)}
\leq C S^{-\varepsilon}\velnorm{z}\vortnorm{z}
\e^{-\frac{\wakefct{K x}}{1+S}}
\snorm{x}^{-3/2}.
\]
Now let us turn to the purely periodic part $\calp_\bot{\mathcal F}_S(z)$.
From \eqref{est:FundsolVortpp_grad} (with $q=1$ and $\gamma=1/4$) we conclude
\[
\int_{\mathbb T}\snorm{\nabla\fundsolvort_\perp(t-s,x-y)}\,{\mathrm d}s
\leq C \snorm{x-y}^{-5/2} \e^{- C_3\snorm{x-y}}.
\]
With formula \eqref{eq:Representation_CurlFpp} and estimate \eqref{est:FsCurl_rhspp}
we thus obtain
\begin{align*}
\snorm{\curl\calp_\bot{\mathcal F}_S(z)(t,x)}
&\leq C\int_{\mathbb T}\int_{\mathbb{R}^3}\snorml{\nabla\fundsolvort_\perp(t-s,x-y)}\,\snorml{
\np{1-\chi_S(y)}\cala_\perp(z)(s,y)}\,{\mathrm d}y{\mathrm d}s\\
&\leq C S^{-\varepsilon}\velnorm{z}\vortnorm{z}
\int_{\mathrm{B}^S}\snorm{x-y}^{-5/2} \e^{- C_3\snorm{x-y}}
\snorm{y}^{-9/2+\varepsilon}\e^{-\frac{\wakefct{K y}}{1+S}}\,{\mathrm d}y.
\end{align*}
By \eqref{est:Kconst} we have
\[
\e^{-\frac{ C_3\snorm{x-y}}{2}}\e^{-\frac{\wakefct{K y}}{1+S}}
\leq\e^{-\wakefct{K\np{x-y}}}\e^{-\frac{\wakefct{K y}}{1+S}}
\leq\e^{-\frac{\wakefct{K\np{x-y}}}{1+S}}\e^{-\frac{\wakefct{K y}}{1+S}}
\leq\e^{-\frac{\wakefct{K x}}{1+S}}.
\]
This yields
\[
\begin{aligned}
&\snorm{\curl\calp_\bot{\mathcal F}_S(z)(t,x)}\\
&\qquad\leq C S^{-\varepsilon}\velnorm{z}\vortnorm{z}
\e^{-\frac{\wakefct{K x}}{1+S}}
\int_{\mathbb{R}^3}\snorm{x-y}^{-5/2} \e^{-\frac{ C_3\snorm{x-y}}{2}}
\np{1+\snorm{y}}^{-9/2+\varepsilon}\,{\mathrm d}y.
\end{aligned}
\]
Employing Lemma \ref{lem:ConvolutionExp} to estimate the remaining integral,
we end up with
\[
\snorm{\curl\calp_\bot{\mathcal F}_S(z)(t,x)}
\leq C S^{-\varepsilon}\velnorm{z}\vortnorm{z}
\e^{-\frac{\wakefct{K x}}{1+S}}\snorm{x}^{-9/2+\varepsilon}.
\]
In total, we have thus shown \eqref{est:Fs_curlquadratic}.
Estimate \eqref{est:Fs_curlcontraction} is derived in the same way.
\end{proof}
\section{Conclusion of the proof}
\label{sec:ConclusionProof}
After the preparatory results from the previous section,
we now prove the existence of a function $z\in\caln_S^\varepsilon$
satisfying the fixed-point equation
\[
z={\mathcal F}_S(z)+{\mathcal H}_S
\]
provided $S\geq2S_0$ is chosen sufficiently large.
Afterwards, we show uniqueness of this fixed point in the function class $\calm_S$.
Since $\restriction{u}{{\mathbb T}\times\mathrm{B}^S}$ is another solution to
this fixed-point equation and belongs to $\calm_S$ by Theorem \ref{thm:VelocityDecay},
we then conclude that $z$ coincides with $\restriction{u}{{\mathbb T}\times\mathrm{B}^S}$.
This yields the decay rate of the vorticity field asserted in Theorem \ref{thm:VorticityDecay}
up to a factor $\snorm{x}^{-\varepsilon}$ for the purely periodic part.
Returning to the representation formula \eqref{eq:RepresentationVorticity_CurluWedgeu_pp},
we finally omit this factor and complete the proof of Theorem \ref{thm:VorticityDecay}.
To begin with, for $S\in[2S_0,\infty)$ consider the closed subset
\[
{\mathcal B}_S:=\setcl{z\in\caln_S^\varepsilon}
{\norm{z}_{\caln_S^\varepsilon}\leq C_6+ C_7+1}
\]
of the Banach space $\caln_S^\varepsilon$.
Choose $S_1\in[2S_0,\infty)$ so large that
for all $S\in[S_1,\infty)$ we have
\begin{align*}
\np{ C_8+ C_9}
\np{ C_6+ C_7+1}^2S^{-\varepsilon}
&\leq 1,
\\
\np{ C_8+ C_9}
\np{ C_6+ C_7+1}S^{-\varepsilon}
&\leq\frac{1}{4}.
\end{align*}
Thus, we obtain the existence of a fixed point of $z\mapsto{\mathcal F}_S(z)+{\mathcal H}_S$.
\begin{cor}\label{cor:ExistenceFixedPoint}
For any $S\in[S_1,\infty)$ there is a function $z_S\in{\mathcal B}_S$ with
$z_S={\mathcal F}_S(z_S)+{\mathcal H}_S$.
\end{cor}
\begin{proof}
By the Lemma \ref{lem:Hs_velest}, Lemma \ref{lem:Hs_vortest},
Lemma \ref{lem:Fs_velest} and Lemma \ref{lem:Fs_vortest}
and the choice of $S_1$, the mapping
\[
F_S\colon{\mathcal B}_S\to{\mathcal B}_S,\qquad
F_S(z):={\mathcal F}_S(z)+{\mathcal H}_S
\]
is a well-defined contractive self-mapping for any $S\geqS_1$.
The contraction mapping principle thus implies
the existence of the asserted fixed point $z_S\in{\mathcal B}_S$
of $F_S$.
\end{proof}
Next we show that $z_S$ coincides with
$\restriction{u}{{\mathbb T}\times\mathrm{B}^S}$
for $S$ sufficiently large.
This yields pointwise estimates of $u$.
\begin{lem}\label{lem:SlowerDecay}
There exists $S_2\in[S_1,\infty)$ such that for all $S\in[S_2,\infty)$ we have
\begin{align*}
\snorm{\curl\calpu(x)}
&\leq\np{ C_6+ C_7+1}
\snorm{x}^{-3/2}\e^{-\frac{\wakefct{K x}}{1+S}}, \\
\snorm{\curl\calp_\botu(t,x)}
&\leq\np{ C_6+ C_7+1}
\snorm{x}^{-9/2+\varepsilon}\e^{-\frac{\wakefct{K x}}{1+S}}
\end{align*}
for all $t\in{\mathbb T}$ and $x\in\mathrm{B}^{S}$.
\end{lem}
\begin{proof}
For $S\geq2S_0$ we set $U_S:=\restriction{u}{{\mathbb T}\times\mathrm{B}^S}$.
By Theorem \ref{thm:VelocityDecay} we know $U_S\in\calm_S$
with $\velnorm{U}\leq C_2$,
and by \eqref{eq:uFixedPoint_VortDecay} we have
$U_S={\mathcal F}_S(U_S)+{\mathcal H}_S$ for any $S\geq 2S_0$.
Now let $S\geqS_1$ and let $z_S\in{\mathcal B}_S$ be the function from
Corollary \ref{cor:ExistenceFixedPoint}.
Then Lemma \ref{lem:Fs_velest} implies
\[
\begin{aligned}
\velnorm{z_S-U_S}
=\velnorm{{\mathcal F}_S(z_S)-{\mathcal F}_S(U_S)}
&\leq \CFsvelnormS^{-\varepsilon}
\bp{\velnorm{z_S}+\velnorm{U_S}}
\velnorm{z_S-U_S}\\
&\leq \CFsvelnormS^{-\varepsilon}
\bp{ C_6+ C_7+1+ C_2}
\velnorm{z_S-U_S}.
\end{aligned}
\]
Choosing $S_2\in[S_1,\infty)$ such that for all $S\in[S_2,\infty)$ we have
\[
C_8
\bp{ C_6+ C_7+1+ C_2}S^{-\varepsilon}
\leq\frac{1}{2},
\]
we conclude $\velnorm{z_S-U_S}\leq\velnorm{z_S-U_S}/2$
and hence $\velnorm{z_S-U_S}=0$ for all $S\in[S_2,\infty)$.
This implies $z_S=U_S=\restriction{u}{{\mathbb T}\times\mathrm{B}^S}$.
In particular, we have
$\vortnorm{\restriction{u}{{\mathbb T}\times\mathrm{B}^S}}=\vortnorm{z_S}
\leq C_6+ C_7+1$
for all $S\in[S_2,\infty)$.
This completes the proof.
\end{proof}
Another application of the convolution formula \eqref{eq:RepresentationVorticity_CurluWedgeu_pp}
enables us to omit the term $\varepsilon$ in the estimate of $\curl\calp_\botu$,
which yields the estimates from Theorem \ref{thm:VorticityDecay} in the case $\Omega=\mathbb{R}^3$.
\begin{thm}\label{thm:VorticityDecay_WholeSpace}
Let $\lambda> 0$ and let $f\in\LR{q}({\mathbb T}\times\mathbb{R}^3)^3$ for all $q\in(1,\infty)$
have bounded support.
Let $u$ be a weak time-periodic solution to \eqref{sys:NavierStokesTP}
in the sense of Definition \ref{def:WeakSolution_NStp}, which satisfies
\eqref{el:VelocityAdditionalIntegrability}.
Then there exist constants $C_1>0$ and
$\alpha=\alpha(\lambda,\calt)>0$ such that
the estimates
\eqref{est:VorticityDecay_ss} and
\eqref{est:VorticityDecay_pp}
hold for all $(t,x)\in{\mathbb T}\times\mathbb{R}^3$.
\end{thm}
\begin{proof}
We decompose $u=v+w$ into steady-state part $v:=\calpu$
and purely periodic part $w:=\calp_\botu$.
Since $\curlu$ is bounded
by Theorem \ref{thm:VelocityDecay},
Lemma \ref{lem:SlowerDecay} implies
\begin{equation}\label{est:VorticityDecay_BadEst}
\begin{aligned}
\snorm{\curlv(x)}
&\leq C
\np{1+\snorm{x}}^{-3/2}\e^{-\alpha\wakefct{x}}, \\
\snorm{\curlw(t,x)}
&\leq C
\np{1+\snorm{x}}^{-9/2+\varepsilon}\e^{-\alpha\wakefct{x}}
\end{aligned}
\end{equation}
for all $(t,x)\in{\mathbb T}\times\mathbb{R}^3$, where $\alpha=\np{1+S_2}^{-1}K$.
In particular, this implies \eqref{est:VorticityDecay_ss},
and for \eqref{est:VorticityDecay_pp}
it remains to remove $\varepsilon$ in the second inequality.
Due to Theorem \ref{thm:VelocityDecay},
the estimates \eqref{est:VorticityDecay_BadEst} further yield
\[
\snorml{
\curlv(y)\wedgew(s,y)
+\curlw(s,y)\wedgev(y)
+\calp_\bot\nb{\curlw\wedgew}(s,y)
}
\leq C \np{1+\snorm{y}}^{-9/2}\e^{-\alpha\wakefct{\lambda y}}
\]
for all $\np{t,x}\in{\mathbb T}\times\mathbb{R}^3$.
Moreover, by Theorem \ref{thm:FundsolVortpp_est} we have
\[
\int_{\mathbb T}\snorm{\nabla\fundsolvort_\perp(t-s,x-y)}\,{\mathrm d}s
\leq C \snorm{x-y}^{-5/2} \e^{- C_3\snorm{x-y}}.
\]
Using these estimates and \eqref{est:Conv_fLp_vortpp}
in the representation formula
\eqref{eq:RepresentationVorticity_CurluWedgeu_pp}, we conclude
\[
\snorm{\curlw(t,x)}
\leq C\snorm{x}^{-9/2}\e^{-\frac{C_3\snorm{x}}{2\np{1+S_0}}}
+C \int_{\mathbb{R}^3}\snorm{x-y}^{-5/2} \e^{- C_3\snorm{x-y}}
\np{1+\snorm{y}}^{-9/2}\e^{-\alpha\wakefct{\lambda y}}\,{\mathrm d}y.
\]
Due to $2\wakefct{K x}\leq C_3\snorm{x}$, we have
\[
\frac{1}{2} C_3\snorm{x-y}+\alpha\wakefct{\lambda y}
\geq\wakefct{K\np{x-y}}+\frac{\wakefct{K y}}{1+S_2}
\geq\frac{\wakefct{K x}}{1+S_2}
=\alpha\wakefct{x},
\]
and we obtain
\[
\begin{aligned}
\snorm{\curlw(t,x)}
&\leq
C\np{1+\snorm{x}}^{-9/2}\e^{-\alpha\wakefct{x}}\\
&\qquad+C \e^{-\alpha\wakefct{x}}\int_{\mathbb{R}^3}\snorm{x-y}^{-5/2}
\e^{- C_3\snorm{x-y}/2}\np{1+\snorm{y}}^{-9/2}\,{\mathrm d}y,
\end{aligned}
\]
where we used \eqref{est:Kconst}.
We estimate the remaining integral with Lemma \ref{lem:ConvolutionExp},
which leads to
\eqref{est:VorticityDecay_pp}
and completes the proof.
\end{proof}
Finally, we employ a classical cut-off argument
to extend the result to an exterior domain and to
finish the proof of Theorem \ref{thm:VorticityDecay}.
\begin{proof}[Proof of Theorem \ref{thm:VorticityDecay}]
First of all, Lemma \ref{lem:NavierStokesTP_Regularity}
implies the existence of a pressure field $\mathfrak{p}$
such that $\np{u,\mathfrak{p}}$ is a strong solution
to \eqref{sys:NavierStokesTP_ExteriorDomain}
satisfying
\eqref{el:NavierStokesTP_Regularity_ss}--\eqref{el:NavierStokesTP_Regularity_pressure}.
Fix radii $R>r>0$ such that $\partial\Omega\subset\mathrm{B}_r$,
and let $\chi\in\CR \infty(\mathbb{R}^3)$ be a cut-off function such that
$\chi(x)=0$ for $\snorm{x}\leq r$ and
$\chi(x)=1$ for $\snorm{x}\geq R$.
By the divergence theorem, $\Divu=0$ and \eqref{cond:BoundaryData},
we have
\[
\begin{aligned}
\int_{\mathrm{B}_{r,R}}u\cdot\nabla\chi\,{\mathrm d}x
&=\int_{\mathrm{B}_R} \Div\bp{u\np{\chi-1}}\,{\mathrm d}x
=-\int_{\partial\Omega}u\cdot\mathrm{n}\,{\mathrm d}S
=0.
\end{aligned}
\]
Therefore, there exists a function $V$ with
$V\in\WSR{1,2}{q}({\mathbb T}\times\mathbb{R}^3)^3$ for all $q\in(1,\infty)$
and $\suppV\subset{\mathbb T}\times\mathrm{B}_{r,R}$
such that
$\DivV=u\cdot\nabla\chi$;
see \cite[Section III.3]{GaldiBookNew} for example.
We define $U:=\cutoffu-V$ and
$\mathfrak{P}:=\chi\mathfrak{p}$.
Then $U\in\LR{r}({\mathbb T}\times\mathbb{R}^3)$ for some $r\in(5,\infty)$
and $U$ is a weak solution to
\begin{equation}\label{sys:NavierStokesTP_CutOff}
\begin{pdeq}
\partial_tU-\DeltaU-\lambda\partial_1U+U\cdot\nablaU+\nabla\mathfrak{P}&=F
&&\text{in }{\mathbb T}\times\mathbb{R}^3, \\
\DivU&=0
&&\text{in }{\mathbb T}\times\mathbb{R}^3, \\
\lim_{\snorm{x}\to\infty}U(t,x)&=0
&&\text{for } t\in{\mathbb T},
\end{pdeq}
\end{equation}
in the sense of Definition \ref{def:WeakSolution_NStp},
where $F\in\LR{q}({\mathbb T}\times\mathbb{R}^3)^3$ for all $q\in(1,\infty)$,
and $\supp F$ is compact.
Now the assertion follows from Theorem \ref{thm:VorticityDecay_WholeSpace}
and the identity $U(t,x)=u(t,x)$ for $\snorm{x}>R$.
\end{proof}
|
\section{Introduction}
\noindent
First identified in 1873,\cite{vdW} the van der Waals (vdW) interactions are forces that today attract more interest than ever. They are present everywhere, but their manifestations still pose challenging questions which are relevant for such varied systems as soft matter, surfaces, and DNA, and in phenomena as different as supramolecular binding, surface reactions, and the dynamic properties of water. vdW interactions play a central role even in the characterization of layered materials, such as transition metal dichalcogenides (TMDs).
These represent a large class of materials with mild stiffness, which are not as soft as tissues and not as strong as metals.\cite{LM}
They are two-dimensional systems which have unique properties and
are central to current research in solid-state science.
In TMDs, a transition metal atom (M) layer is sandwiched between two chalcogen
atom (X) layers and it is commonly assumed that the MX$_2$ slabs are stacked
by vdW interactions, whereas the intralayer M-X interactions are instead
covalent. The weak, interlayer vdW interactions play a key role in the
formation, intercalation, exfoliation and layer-by-layer building of TMD
materials, as well as being decisive for their characteristic properties.\\
\noindent
Like all non-relativistic electronic effects, the vdW interactions are present in the (unknown) exact Density Functional Theory (DFT) exchange-correlation (XC) functional, but they are not properly described by standard approximate DFT schemes, such as local density approximation (LDA)\cite{LDA} or semilocal generalized gradient approximation (GGA).\cite{PBE}
Hence the need to introduce truly non local XC functionals, able to accurately account for vdW corrections, arises.
Despite a considerable amount of work focused on the development of accurate vdW-corrected DFT methods, a precise characterization of the layer-layer distance and of the interlayer binding energy of vdW layered materials, still represents a challenging issue.\cite{Kohn,Pernal,Stohr,Ruiz,Su}
This is due to the nonlocal and long-range nature of the vdW interaction, as well as to the coexistence of the weak vdW bonding and much stronger intralayer chemical bonding. Fully accounting for the vdW interaction is achievable by high-level methods such as quantum Monte Carlo (QMC),\cite{Foulkes} coupled-cluster singles and doubles with perturbative triples (CCSD(T)),\cite{Raghavachari} and exploiting the adiabatic-connection fluctuation-dissipation theorem within the random-phase approximation (RPA),\cite{Eshuis} which are, however, only feasible for limited-size systems because of their high computational cost.\\
The physical quantity lying at the DFT core, is the electron density (ED),
which fully determines the total energy of any many-electron system
in its ground state.
The ED distribution probably represents the most information-rich observable available since it has become possible to determine EDs from analysis of structure factors obtained from accurate X-ray diffraction data. During the past decade, the accuracy of experimental X-ray diffraction data has increased dramatically owing to the use of high-energy synchrotron sources, which significantly limit systematic errors in the data and, thanks to these modern techniques, nowadays it is possible to evaluate both structural and electronic properties of a wide class of layered materials in an extremely accurately way.
In this respect, Medvedev \textit{et al.}\cite{Medvedev} have pointed out that modern DFT functionals are constructed on the basis of empirical fitting on energetic and geometrical benchmarks, neglecting the ED as a parameterization parameter.
If, from one side, this leads to improved predictions of energetic and structural features, from the other, the exact reproduction of the ED has worsened.
In a recent study, Kasai \textit{et al.}\cite{Kasai} studied the archetypal vdW material, TiS$_2$, to compare the ED distribution derived from X-ray diffraction data with the theoretical one obtained by several DFT functionals. They obtained a good agreement between theory and experiment for the description of the intralayer, covalent Ti-S interaction, but, at the same time, significant discrepancies were found for the interlayer vdW interactions, particularly considering the ED distribution in the region between S atoms belonging to adjacent layers. In fact, while
the properties at the bond critical point (BCP) were quite similar, noticeable differences were instead observed away from the BCP: considering the theoretical DFT density, the S atoms behave as if they are practically not interacting with the neighboring layer, showing charge concentration and accumulation in the region expected for a lone pair in an \textit{sp3}-hybridized S atom with three bonds to nearby Ti atoms and no other bonds; in the experiment, however, an appreciable charge accumulation and concentration was observed in the interlayer region, which was interpreted as a sign of a stronger and more directional interaction between S atoms.
Thus, this study suggested the inability of current DFT functionals to accurately describe the interlayer ED for vdW layered materials, thus supporting the conclusions of Medvedev \textit{et al.}
In order to shed light on this issue, in the first part of this study, we apply the LDA, the semilocal PBE\cite{PBE} and the vdW-corrected rVV10\cite{rVV10} DFT functionals, with standard Projector-Augmented Wave (PAW) pseudopotentials,\cite{PAW} to explore the behavior of the interlayer ED in different regimes, obtained by also modifying the parameters of the rVV10 functional, so that to artificially tuning the intensity of the vdW interactions. The basic result is that, in order to produce an ED
distribution appreciably closer to the experimental profile, we have to
select the rVV10 parameters in such a way to unphysically increase the
strength of the vdW interactions, thus leading to an exaggeratedly high
interlayer binding energy.
In the second part a standard rVV10 approach is used, but a novel
pseudopotential is adopted, which explicitly accounts for the 3$d$ orbitals of the S atom. While these orbitals are empty in the isolated, neutral S atom, they
can play a role\cite{dstate} in molecules and condensed-matter systems where S atoms interact with other atoms and form bonds. Interestingly,
using this novel, more flexible S pseudopotential, we obtain not only an ED
distribution closer to the experimental profile, but also a better estimate
of the interlayer binding energy. We finally show that this improvement in
the theoretical DFT description is not limited to TiS$_2$ but also applies
to other, similar layered systems involving S atoms, such as
TaS$_2$ HfS$_2$, and MoS$_2$.
\section{Theoretical background and Computational details}
\noindent
In DFT the properties of a many-electron system can be determined by introducing suitable functionals of the electron density $n$(\textbf{r}); the most interesting many-body physical effects are described by the so-called exchange-correlation (XC) functional, related to the XC potential: $V_{xc}[n(\textbf{r})] = V_{x}[n(\textbf{r})] + V_{c}[n(\textbf{r})]$.
In principle $V_{xc}$ describes all the many-body effects, including the
vdW interactions, however, in practice, a number of approximation must be introduced, since, unfortunately, the exact form of the XC functional is not known. Nowadays several kinds of approaches to approximately include vdW interaction in DFT calculations are available. Some of these are based on semiempirical, atom-based pair potentials, able to give approximate vdW corrections;\cite{grimme1,grimme2,xdm} instead others are built introducing a suitable non-local XC density functional, such as the vdW-DF family by Dion \textit{et al.}\cite{Dion} and the scheme proposed by Vydrov and Van Voorhis,\cite{VV10} with the revised, more efficient version rVV10.\cite{rVV10}
rVV10 certainly represents one of the best vdW-corrected functionals nowadays available; it takes into account the entire range of vdW interactions at a reasonable computational cost using only the ED $n$(\textbf{r}) as input.
In particular, the rVV10 non-local correlation energy has the following expression:
\begin{equation}
E_c^{NL}\,=\,
\int\int\,n(\textbf{r})\Phi(\textbf{r}\,,\textbf{r'}) n(\textbf{r'})\label{1}
\end{equation}
where $\Phi$ is the non-local correlation kernel, which is:
\begin{equation}
\Phi = -\frac{3e^4}{2m^2} \frac{ 1 }
{(q(\textbf{r}) R^2 + 1)(q(\textbf{r'}) R^2 + 1)
[(q(\textbf{r})\,+\,q(\textbf{r'})) R^2 + 2]} \label{2}
\end{equation}
In the above expression $R = | \textbf{r} - \textbf{r'} |$ and $q(\textbf{r})$ is a function of the ED and its gradient:
\begin{equation}
q(\textbf{r}) =
\frac{\omega_0 (n(\textbf{r}), \nabla n(\textbf{r}))}
{k(n(\textbf{r}))}
\end{equation}
where
\begin{equation}
\omega_0 =
\sqrt{ C \frac{\hbar^2}{m^2}
\left|\frac{ \nabla n(\textbf{r})} {n(\textbf{r})}\right|^4 +
\frac{1}{3} \frac{4\pi n(\textbf{r}) e^2}{m}}
\label{C}
\end{equation}
and
\begin{equation}
k(n (\textbf{r})) = 3 \pi b \left(\frac{n(\textbf{r})}{9 \pi} \right)^{1/6} \label{B}\;.
\end{equation}
The total XC functional is then obtained by adding the non-local correlation energy to the refitted Perdew-Wang exchange functional\cite{rPW86} and the Perdew-Wang LDA correlation\cite{LDAc} functional, as proposed in the seminal work by Vydrov and Van Voorhis:\cite{VV10}
\begin{equation}
E_{xc}^{NL}\,=\,E_{x}^{rPW86} + E_{c}^{LDA} + E_{c}^{NL}\;.
\end{equation}
The rPW86 exchange functional was chosen mainly because it is nearly vdW-free,\cite{rPW86} in such a way to avoid double-counting effects. In eq. \eqref{C} the $C$ parameter represents the so-called local band gap and is tuned to give accurate asymptotic vdW $C_6$ coefficients, thus regulating the behavior of the long-range component of vdW interactions. The fitting procedure, aimed to minimize the average error for a benchmark set of 54 $C_6$ coefficients, was originally carried out by Vydrov and Van Voorhis,\cite{VV10} leading to an optimal value of $C$ = 0.0093.
Another essential aspect of the VV10 (and its successor rVV10) formalism is the presence of a second adjustable parameter, $b$ (see eq. \eqref{B}), which controls the short range behavior of the non-local correlation energy.
This means that when E$_c^{NL}$ is added to the other energetic terms, the $b$ parameter is adjusted to merge the interaction energy contributions at short and intermediate ranges.
With an empirical fitting procedure on the S22 binding-energy data set,\cite{BE} Vydrov and Van Voorhis proposed a value of $b$ = 5.9.\cite{VV10} After the implementation of the efficient Roman-Perez, Soler\cite{RPS} interpolation scheme in the reciprocal space, this value was revised by Sabatini and coworkers to $b$ = 6.3.\cite{rVV10}\\
We have chosen rVV10 as the basic DFT functional to investigate the effect of vdW interactions in the ED distribution of TiS$_2$ (although extensive tests were
also performed using other vdW-corrected functionals), also because of the
possibility to separately modify the intensity of both short- and long-range vdW interactions in a transparent way by simply tuning the two adjustable parameters $b$ and $C$ mentioned above.
First the short-range component is analyzed. This is done by considering two different regimes: the former in which the value of the $b$ parameter is increased to $b$=10.0, thus damping the intensity of short range vdW interactions, and the latter in which $b$ is reduced to $b$=1.0, that instead results in a substantial increase of the intensity of the short-range vdW interaction.
While tuning the parameters of the rVV10 functional can improve the
description of some properties of the TiS$_2$ system, clearly this could
lead to a reduction of the transferability of the modified functional
to other systems.
Our ab-initio calculations have been performed using the version 6.5 of the Quantum ESPRESSO (QE) DFT package.\cite{QE1,QE2}
We first used for Ti and S atoms standard PAW\cite{PAW} pseudopotentials taken
from the QE database.\cite{pseudo}
Subsequently, for further testing, we have generated, for the S atom,
a novel pseudopotential, using the \textit{atomic} QE program.
We report DFT results obtained using the LDA, the semilocal GGA PBE
functional, and the vdW-corrected functional rVV10,
both in its standard form and considering the
variants (parameter changes and combination with a novel S pseudopotential)
described above.
Self-consistent calculations and relaxation processes, leading to the ground-state equilibrium geometry of bulk TiS$_2$, were carried out within the \textit{PWscf} QE program. The calculations adopted plane-wave and density cutoffs of 100 and 1000 Ry, respectively, to get highly-converged results, a cold-smearing parameter of 0.01 Ry, and a 12$\times$12$\times$4 k-point mesh for the sampling of the Brillouin Zone (BZ).
The ED analysis is carried out using the \textit{PostProc} QE program, employing the B-Spline method\cite{spline} for the $n(x, y, z)$ interpolation; in particular, for the ED profile computed along the S-S, S-center, and Ti-S lines, a real-space grid of 600$\times$600$\times$800 (x, y, z) points, generated inside a parallelepiped containing the axis of interest, is used.
To get a reliable estimate of the $\nabla^{2}n$(\textbf{r}) function, in the case of the S-S line, the distance between the parallelepiped axis and the edges of its sides is 0.5 \si{\angstrom}, for the S-center case 0.25 \si{\angstrom}, and finally 0.35 \si{\angstrom} for the Ti-S line.
The ED is then processed by using the \textit{Critic2} program.\cite{Critic2_1,Critic2_2}
More specifically, we compute the ED charge deformation,
$\Delta n$(\textbf{r}), by considering a spatial mesh of 72$\times$72$\times$108 points inside the unit cell (containing one Ti atom and two S atoms)
and taking the difference between the ED $n$(\textbf{r}) and the
sum of the EDs of the isolated Ti and S atoms.
$\Delta n$(\textbf{r}) profiles and Laplacian $\nabla^{2}n$(\textbf{r}) maps, in the interlayer and intralayer areas, are then evaluated on the plane of interest using \textit{Critic2}. The calculation of the atomic properties of a given S atom is carried out by making an integration of the so-called ''atomic basin'' of the S atom, represented by the space region whose surface has zero-density gradient flux.
This is done again with the \textit{Critic2} program, that allows to compute spherical multipolar moments defined as
\begin{equation}
Q_{lm} = \int_A n(\textbf{r})\,r^l\,Y_{lm}
\label{m1}
\end{equation}
where the integration domain $A$ is the atomic basin volume. In the above equation, $n(\textbf{r})$ is the charge ED and $Y_{lm}$ are the spherical harmonics, in which the $m$ index runs over the interval [-$l$\,;\,$l$]. With this definition, the magnitude of the dipole moment attributed to the S atom is obtained as the square root of the sum of the squared $l$=1 terms:
\begin{equation}
d(q_A) = \sqrt{\sum_{m=-1,0,1} Q_{1m}^2}\,\;.
\end{equation}
\section{Results and discussions}
\noindent
Starting from the experimental lattice parameters, $a$ = 3.398 \si{\angstrom} and $c$ = 5.665 \si{\angstrom}, as reported by Kasai \textit{et al.},\cite{Kasai} we relax the crystal structure in the intralayer atomic plane until the force acting on the system is smaller than $10^{-5}$ Ry/a.u. This relaxation process, carried out with both PBE and rVV10 functionals, leads to very similar results,
as expected,\cite{Kasai} since vdW interactions do not substantially influence the geometry of the intralayer plane.
By further optimizing the interlayer geometry, the standard rVV10 functional
predicts that
the distance between two S atoms belonging to adjacent layers is 3.443 \si{\angstrom} while the intralayer S-Ti distance is 2.421 \si{\angstrom}, in excellent agreement with the reference DFT values (at the SCAN+rVV10 level\cite{SCAN+rVV10}) of 3.446 \si{\angstrom} and 2.420 \si{\angstrom}, respectively, reported in ref. \citenum{Kasai}.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=\linewidth]{TiS2_tetrahedron.png}
\end{center}
\caption{\scriptsize a) \& b): bulk structure of TiS$_2$. The unit cell is repeated twice along the $x$, $y$, and $z$ directions. c) On the left hand side the TiS$_2$ octahedron is represented: the bigger S atom in the middle is the reference S atom, while the grey balls denote Ti atoms. The lower tetrahedron sides indicate the S-S lines, connecting the reference atom to the S atoms belonging to the adjacent layer. On the right hand side, the S-center line (connecting the reference S atom to the center of the lower tetrahedron) is represented by dots. The green spheres denote the positions of the (3, -3) critical points (CPs) along the S-S lines, the red spheres are referred to the (3, -1) CPs and, finally, the violet sphere shown inside the triangular section (right side of the c) panel) represents the position of the (3, -3) CP belonging to the S-center line.}
\end{figure}
As already mentioned, Kasai \textit{et al.}\cite{Kasai} pointed out that
DFT simulations are characterized by a
an interlayer S-S ED deformation significantly lower than the experimental one, implying that the X-ray diffraction data predict a stronger S-S interlayer interaction than DFT.
Moreover, the experimental ED is characterized by 3 maxima in the direction of the 3 neighboring, symmetry-related S atoms, while the theoretical one exhibits only one maximum directed towards the center of the S-atom tetrahedron (see Figure 1), suggesting that the
theoretical DFT description of the interlayer interactions indeed
qualitatively differs from the experimental one.
In order to investigate the reasons of this discrepancy, we analyze the results of our DFT calculations and focus on the two directions of main interest, namely the S-center and the S-S line. Using the Bader nomenclature, for each line we characterize the (3, -3) critical point (CP) evaluating the position, \textbf{r}$_{max}$, where the charge density reaches its maximum value, the density at this point,
$n$(\textbf{r}$_{max}$), and the Laplacian, $\nabla^{2}n$(\textbf{r}$_{max}$).
Furthermore, we do the same for the (3, -1) CP, computing \textbf{r}$_{min}$,
$n$(\textbf{r}$_{min}$), and $\nabla^{2}n$(\textbf{r}$_{min}$).
As a preliminary analysis, we have computed the ED and the Laplacian along the Ti-S line, identifying the main features of the polar covalent bond in the intralayer plane.
The values that characterize the ED are reported in Table S1 of
\textit{Supporting information}.
Two non-equivalent (3, -3) CPs are found along the Ti-S line (see top left panel in Figure S1 of \textit{Supporting information}).
The (3, -1) CP is instead located at 1.12 \si{\angstrom} from the Ti atom.
As can be seen, looking at Table S1 of \textit{Supporting information}, very similar values are predicted by
all the DFT functionals applied and most of these are also in good agreement with the experimental data.\\
A similar analysis can be also applied to characterize the interlayer bonding:
the properties of the (3, -3) CP for both the S-S and S-center lines, and for the (3, -1) CP along the the S-S line are summarized in Table 1 and 2, respectively.
In the interesting case of the S-S line, the PBE, rVV10, and rVV10($b$=10.0) functionals predict a maximum value of the ED at a distance \textbf{r}$_{max}$ $\sim$ 0.61 \si{\angstrom} from the reference S atom, with a relative error of $\sim$ 13 \% with respect to the experimental result,\cite{Kasai} while the error is slightly larger with rVV10($b$=1.0) and LDA functionals.
\begin{table}[h!]
\caption{\scriptsize Properties at the (3, -3) CP along the S-S and S-center lines, evaluated using different DFT functionals.}
\begin{center}
\begin{tabular}{l| l| c| c| c}
&line&\textbf{r}$_{max}$ (\si{\angstrom}) & $n$(\textbf{r}$_{max}$) (e/\si{\angstrom}$^3$) &
$\nabla^{2}n$(\textbf{r}$_{max}$) (e/\si{\angstrom}$^5$) \\
\hline
\multirow{2}{*}{rVV10}&S - S&0.614&1.102&-9.163\\
&S - center&0.614&1.132&-10.052\\
\hline
\multirow{2}{*}{rVV10($b$=1.0)}&S - S& 0.607& 1.050&-8.663\\
&S - center&0.605&1.081&-9.479\\
\hline
\multirow{2}{*}{rVV10($b$=10.0)}&S - S&0.614&1.105&-9.222\\
&S - center&0.614&1.132&-9.985\\
\hline
\multirow{2}{*}{PBE}&S - S&0.614&1.095&-9.093\\
&S - center&0.614&1.129&-9.934\\
\hline
\multirow{2}{*}{LDA}&S - S&0.579&1.141&-10.016\\
&S - center&0.579&1.172&-11.419\\
\hline
\multirow{2}{*}{rVV10$opt$}&S - S&0.609&1.074&-8.950\\
&S - center&0.607&1.102&-9.766\\
\hline
Experiment\cite{Kasai}&S - S&0.707&1.150&-9.460\\
Theory\cite{Kasai}(SCAN+rVV10)&S - center&0.711&1.080&-8.910\\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[!h]
\begin{center}
\includegraphics[width=\linewidth]{charge_lap_PBE.png}
\end{center}
\caption{\scriptsize a) ED computed along the S-S and
S-center line with the PBE functional. b) Corresponding Laplacian profiles. The small oscillations of the Laplacian profile are due to the finite-size of the real space mesh considered for the ED computation.
In this case $\nabla^{2}n$(\textbf{r}$_{max}$) and $\nabla^{2}n$(\textbf{r}$_{min}$) are estimated by a parabolic interpolation $f(r \sim c) = c_2 (r-c)^2 + k$, where
the expansion center is $c$ = \textbf{r}$_{max}$ or $c$ = \textbf{r}$_{min}$.
c) ED deformation $\Delta$ $n$(\textbf{r}) computed using the PBE functional and d) the rVV10($b$=1.0) functional. Curves obtained by rVV10 and rVV10($b$=10.0) functionals are very similar to the PBE ones and therefore are not reported.}
\end{figure}
\newpage
The $n$(\textbf{r$_{max}$}) values are all very similar and in reasonable agreement with the experimental data, with the LDA functional that shows the best agreement with the experimental value. The rVV10($b$=1.0) result, for which the maximum discrepancy occurs, has a deviation of only $\sim$ 9 \%.
The Laplacians $\nabla^{2}n$(\textbf{r}$_{max}$) are also in acceptable agreement with the experimental values. Again, the result of the rVV10($b$=1.0) functional exhibits the maximum deviation from the experiment, with a discrepancy of $\sim$ 8 \%.
This means that, at a distance r$_{max}$ from the reference S atom, along the S-S line, with rVV10($b$=1.0), the ED deformation is slightly smaller than for the other functionals.\\
For the S-center line a direct comparison with experimental data is not available, so we take as reference the theoretical SCAN+rVV10 result reported by Kasai \textit{et al.}\cite{Kasai}
The S-center line is characterized by the same \textbf{r}$_{max}$ value found for the S-S line, again with a small deviation of the value obtained with rVV10($b$=1.0) and LDA. Even in this case, the results for the ED and the Laplacians are in line with the theoretical reference data.\cite{Kasai} Interestingly, for each functional employed, a slightly larger
$n$(\textbf{r$_{max}$}) value than that relative to the S-S line is observed, with a stronger ED deformation described by the associated Laplacian, indicating that the charge deformation along the S-center line is stronger than for the S-S line.
The maximum deviation from the reference values, for both ED and Laplacians, occurs again for the LDA functional with an error of $\sim$ 8 \% and $\sim$ 28\%, respectively.\\
\begin{table}[h!]
\caption{\scriptsize Properties at the (3, -1) CP along the S - S line, evaluated using different DFT functionals. The theoretical (SCAN+rVV10) result reported, is the result achieved after a multipolar refinement of the pure DFT density (see ref. \citenum{Kasai} for details).}
\begin{center}
\begin{tabular}{l| c| c| c}
&\textbf{r}$_{min}$ (\si{\angstrom}) & $n$(\textbf{r}$_{min}$) (e/\si{\angstrom}$^3$) &
$\nabla^{2}n$(\textbf{r}$_{min}$) (e/\si{\angstrom}$^5$) \\
\hline
rVV10&1.718&0.068&0.486\\
rVV10($b$=1.0)&1.718&0.079&0.461\\
rVV10($b$=10.0)&1.718&0.067&0.481\\
PBE&1.718&0.067&0.460\\
LDA&1.718&0.070&0.468\\
rVV10$opt$&1.718&0.074&0.464\\
\hline
Theory\cite{Kasai}(SCAN+rVV10)& &0.058&0.727\\
Experiment\cite{Kasai}& &0.086&0.691\\
\hline
\end{tabular}
\end{center}
\end{table}
\newpage
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{diff_fin.PNG}
\end{center}
\caption{\scriptsize ED deformation map in the interlayer plane using four different functionals: a) PBE, b) rVV10, c) rVV10($b$=10.0), and d) rVV10($b$=1.0). The Ti atom are represented by grey spheres, while the S atoms are the yellow spheres. In the present maps, the smallest isoline value reported is 0.001 e/\si{\angstrom}$^3$, with red and blue contours that represent charge accumulation and depletion, respectively.}
\end{figure}
The S-S bond shows a symmetric ED distribution along the line connecting the two S atoms (see Figure 2a)), with \textbf{r}$_{min}$ exactly halfway between the two S atoms.
Our computed $n$(\textbf{r}$_{min}$) values, with the PBE, rVV10, and rVV10($b$=10.0) functionals, as well as LDA, are in very good agreement with the DFT densities presented in ref. \citenum{Kasai}, exhibiting a significant underestimate (by about 22 \%) of the ED at the interlayer BCP, with respect to the experimental value.
Interestingly, using the rVV10($b$=1.0) functional, the discrepancy with the
experiment is instead considerably reduced to about 8 \%.
Note that the reported, theoretical SCAN+rVV10 value, \cite{Kasai} is slightly below (and therefore slightly worse than) all our results.
For the (3, -1) CP, the Laplacian sign is always positive, as expected, indicating that the interlayer BCP located at \textbf{r}$_{min}$ is a point where a charge depletion occurs.
In Figure 2a) and 2b), an example of ED profiles and Laplacians, at the PBE level, along the S-S and S-center lines are shown, while in the bottom part, a comparison between the PBE ED deformation $\Delta n$(\textbf{r}) (Figure 2c)) and that obtained by rVV10($b$=1.0) (Figure 2d)) is made along the same lines, highlighting that rVV10($b$=1.0) predicts a stronger charge depletion in the vicinity of the reference S atom than PBE and the other functionals. A consequent charge accumulation in the interlayer region appears, as confirmed by the higher ED $n$(\textbf{r}$_{min}$)$_{\mbox{S}\,-\,\mbox{S}}$ at the (3, -1) CP accompanied by a smaller Laplacian value $\nabla^2 n$(\textbf{r}$_{min}$)$_{\mbox{S}\,-\,\mbox{S}}$ (see Table 2) that indicates a slightly less charge depletion than for the other functionals considered.
The ED deformation map in the interlayer plane, presented in Figure 3, is very similar with the PBE, rVV10, and rVV10($b$=10.0) functionals (as well as the LDA case which it is not reported), while the same is not true for rVV10($b$=1.0), which predicts a more pronounced ED deformation in the region between the two S atoms belonging to adjacent layers.
The Laplacian map of the interlayer plane is presented in Figure S2 of the \textit{Supporting information}: in this case, no significant differences for the functional schemes adopted up to now are found, as well as for LDA and the other schemes presented below.
\begin{table}[h!]
\caption{\scriptsize Cohesive energy (CE) and interlayer binding energy (ILBE), see text for the definitions, obtained using the experimental lattice parameter $c$ (first and second column) and the
corresponding quantities (denoted by the $^{*}$ symbol, in the fourth and
fifth column) relative to the $c$ value (third column) optimized for each adopted functional. The reported reference CE and $c$ values are experimental data, while, for the ILBE, we provide theoretical RPA\cite{bj} and MBD estimates.
The quantity V, in the sixth column, represents the total charge belonging to the atomic basin $A$, surrounding the S atom. The charge q$_{A}$, contained in $A$, and the atomic dipole moment of the S atom, d(q$_{A}$), are reported in column seven and eight respectively.}
\begin{center}
\begin{adjustbox}{max width=\textwidth}
\begin{tabular}{l| c| c| c| c| c| c| c| c}
&CE (\si{\electronvolt})&ILBE (m\si{\electronvolt}/\si{\angstrom}$^2$)
&$c$ (\si{\angstrom})&CE$^{*}$ (\si{\electronvolt})&ILBE$^{*}$ (m\si{\electronvolt}/\si{\angstrom}$^2$)&V (\si{\angstrom}$^3$)&q$_{A}$ (e)&d(q$_{A}$) (e\si{\angstrom})\\
\hline
rVV10&-15.602&-27.828&5.728&-15.603&-27.888&23.775&-0.860&0.380\\
rVV10($b$=1.0)&-27.928&-192.073&4.934&-42.140&-1613.663&23.781&-0.861&0.383\\
rVV10($b$=10.0)&-15.035&-16.609&5.878&-15.045&-17.698&23.574&-0.796&0.324\\
PBE&-15.150&-3.013&6.595&-15.189&-0.861&23.804&-0.871&0.404\\
LDA&-18.633&-17.275&5.470&-18.633&-18.357&23.648&0.829&0.383\\
rVV10$opt$&-25.889&-151.580&5.123&-25.310&-175.506&23.615&-0.809&0.373\\
\hline
Experiment&-14.805\cite{ce}& &5.665\cite{Kasai}&-14.805& &23.470\cite{Kasai}&-0.820\cite{Kasai}&0.030\cite{Kasai}\\
Theory (RPA)& &-18.900\cite{bj}& & &-18.900&23.630\cite{Kasai}&-0.800\cite{Kasai}&0.340\cite{Kasai}\\
Theory (MBD)& &-19.031& & &-19.031& & & \\
\hline
\end{tabular}
\end{adjustbox}
\end{center}
\end{table}
\newpage
A summary of the basic energetic quantities of TiS$_2$ and the atomic
properties of the reference S atom is reported in Table 3.
We optimized the interlayer lattice constant $c$ for each functional adopted (third column of Table 3), obtaining results close to the experimental reference value using the rVV10 and rVV10($b$=10.0) functionals, with an error of only $\sim$ 3 \%. Instead, the quality of the PBE, rVV10($b$=1.0), and LDA estimates is worse, with PBE that substantially overestimates, and LDA and rVV10($b$=10.0) which instead underestimate. This behavior does not come as a surprise: in fact LDA and the semilocal PBE functional are not able to properly describe the vdW interactions, while the rVV10($b$=1.0) functional was deliberately built to artificially increase the strength of short-range vdW interactions, thus leading to a larger interlayer bonding energy and to a shorter layer-layer distance.
The atomic cohesion energy (CE) is defined as E(TiS$_2$) - E(Ti) - 2E(S), where
E(TiS$_2$) is the total energy of the TiS$_2$ systems, while E(Ti) and E(S)
are the total energies of the isolated T and S atoms. This quantity mainly
reflects the intensity of the strong, intralayer Ti-S covalent bonds.
Considering both the experimental configuration (first column) and the one in which the lattice constant $c$ is optimized (fifth column), quite similar results are achieved using
PBE, rVV10, and rVV10($b$=10.0), with an average discrepancy with the
experimental reference value of $\sim$5 \%.
However, the CE value obtained with the rVV10($b$=1.0) functional is largely
overestimated implying that artificially increasing the strength of the
short-range vdW interaction leads to a severe overbinding even of the
Ti-S covalent bond. Therefore, the improvement in the ED distribution using
rVV10($b$=1.0) comes at the expense of an unacceptable deterioration in the
description of the energetic properties.
This is also confirmed by the analysis of the interlayer binding energy (ILBE), defined as E(TiS$_2$) - E(TiS$_2$)$_{\mbox{c/a=5}}$, representing the interaction energy between two
adjacent layers of the material. Here, E(TiS$_2$)$_{\mbox{c/a=5}}$ indicates
the total energy of the system when the $c$ parameter is set to
$5a \sim$17 \si{\angstrom}, which corresponds to a distance so large that
the interlayer interaction is negligible.
In order to have a more sound comparison with reference values,
in addition to the RPA value reported in the literature,\cite{bj}
we have also computed the ILBE using the Many-Body Dispersion (MBD)
approach\cite{mbd} which allows for an effective RPA description of
long-ranged vdW interactions, beyond conventional pairwise approximations.
Within MBD one maps the atomic response functions into a set of
quantum harmonic oscillators, mutually coupled at the dipole-dipole level.
Due to the inclusion of many-body effects, MBD can
reach chemical accuracy for a broad variety of systems (from small molecules to large nanostructure and periodic materials)\cite{mbd}.
In particular, many-body effects were shown to be crucial in two-dimensional vdW
materials~\cite{prb}, as they can
qualitatively alter the power law scaling of dispersion interactions.
Remarkably (see Table 3) the MBD estimate is very close to the
RPA value reported in ref.\citenum{bj}.
As expected, PBE dramatically underestimates the ILBE (the discrepancy is
even worse considering the optimal configuration predicted by PBE), a
behavior that can be clearly ascribed to the inadequate inclusion of
vdW effects by PBE. As can be seen in Table 3, the standard rVV10
functional significantly overestimates the ILBE, while rVV10($b$=10.0) gives
a result reasonably close to the reference values. Interestingly, the LDA estimation is very good, showing the best agreement with the reference values.
However, one must point out that this good performance is accidental:
the well-known LDA overbinding due to the overestimate of
the long-range part of the exchange contribution, somehow mimics the
missing vdW interactions, thus leading to reasonable estimates of the
binding energies in a wide range of systems.\cite{LDAprob}
Instead, similarly to what observed for the CE, rVV10($b$=1.0) predicts a very large, unphysical ILBE.
Looking at Table 3 our computed atomic basin volume V turns out to be
in good agreement
with the theoretical and experimental reference values, and the same is true
for the total charge q$_{A}$ contained in the atomic basin, with all the
adopted functionals that predict quite similar values (which the partial
exception
of rVV10($b$=10.0) which gives a significantly less negative accumulated
charge). The negative value of q$_{A}$ is consistent with the higher
electronegativity of the S atom, which therefore behaves as the
electron-acceptor in the polar covalent Ti-S bond.
The values of the dipole moment of the S atom, d(q$_{A}$),
computed through the spherical multipolar
expansion (see section II), are, for all the considered functionals, in
line with the theoretical reference data but much larger than the
experimental estimate, thus reproducing the discrepancy between theory and
experiment discussed above. In more detail, PBE exhibits the
largest discrepancy, thus confirming that this functional is not adequate,
while the rVV10($b$=10.0) functional gives a slightly less pronounced dipole
moment, suggesting a description a little closer to the experiment,
although even the rVV10($b$=10.0) value of d(q$_{A}$) is still quite far from the experimental one.
So far, our results for the ED distribution, and for the energetic and atomic
properties of TiS$_2$, are in line with the conclusions of
Kasai \textit{et al.}\cite{Kasai} concerning the theoretical DFT description.
In particular, our analysis of CPs (see Table 1), confirms that the ED predicted at DFT level along the S-S line is smaller than the one along the S-center line, at variance with the experimental evidence.
Small, quantitative differences between our calculations and the theoretical
SCAN+rVV10 results reported by Kasai \textit{et al.}\cite{Kasai} can be easily
explained by the different DFT functional adopted and by other
technical differences: for example, in ref. \citenum{Kasai} a finer mesh
of 22$\times$22$\times$13 k-points was used for the sampling of the BZ,
this choice being motivated by the need of describing core electron
oscillations to carry out a multipole refinement procedure.\cite{Kasai}
Coming to a more detailed analysis, we observe (see Figure S2 and Figure S3 of \textit{Supporting information}) that the maps of $\Delta n$(\textbf{r}) and $\nabla^2 n$(\textbf{r}) obtained by
rVV10($b$=10.0) are very similar to those given by the standard rVV10
functional and also by PBE, thus suggesting that the interlayer ED
distribution is weakly affected by an augmented damping applied to
the short-range vdW interactions; this is also quantitatively confirmed
by the the values of $n$(\textbf{r}) and $\nabla^2 n$(\textbf{r}$_{max}$)
evaluated at the (3, -3) CP (see Table 1 ). A small difference can be
noticed only at the (3, -1) CP, where the rVV10($b$=10.0) value of
$\nabla^2 n$(\textbf{r}$_{max}$) almost coincides with that obtained by
the standard rVV10 result but weakly differs from the PBE one.
The situation is different considering rVV10($b$=1.0), namely
a functional where the short-range vdW interaction is artificially
increased; in fact, although the ED and the Laplacian at the CPs defined
above are similar to the same quantity evaluated by the other functionals,
more pronounced differences can be observed away from the BCP, reflecting
an accumulation of the ED in the interlayer region. As a consequence,
rVV10($b$=1.0) predicts an ILBE value which appears to be dramatically
overestimated.
Moreover, with rVV10($b$=1.0) one can observe (Figure 3d)) a considerable ED
deformation between the reference S atom and the other S atom belonging to
the same layer, thus indicating a strong intralayer S-S interaction,
a feature not observed with the other functionals and absent in
the experimental description.
To overcome (at least partially) these evident shortcomings of rVV10($b$=1.0),
we have also tested another DFT functional, rVV10($b$=1.1,$C$=0.00207), named as rVV10$opt$, which is characterized by a new $C$ parameter, thus modifying the value of the
effective C$_6$ coefficient responsible for the long-range behavior of
the vdW interactions, and also by a slightly changed $b$ parameter.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{charge_deformation_opt2.png}
\end{center}
\caption{ \scriptsize a) ED along the S-S and S-center line, obtained using the
rVV10$opt$ functional. b) Associated ED deformation
$\Delta n$(\textbf{r}). c) ED deformation map in the Ti-S intralayer plane.
d) ED deformation map in the interlayer plane with a contour of $\pm$ 0.01 e/\si{\angstrom}$^3$. e) ED deformation map in the interlayer plane with a smaller contour of $\pm$ 0.001 e/\si{\angstrom}$^3$.}
\end{figure}
The ED deformation map obtained with this new functional is shown in Figure 4.
As can be seen, by artificially manipulating the long-range component of vdW interactions, the unphysical ED accumulation in the intralayer S-S region
disappears and, more interestingly, a small tendency to form a localized
S-S interlayer bond can be observed, although the intensity of
the interlayer ED deformation is again underestimated, being 7 times smaller than the
experimental one.\cite{Kasai}
rVV10$opt$ preserves the nature of the S-Ti intralayer
bond and the basic properties of the ED at the two non equivalent (3, -3)
CPs and at the (3, -1) CP; in fact the positions of these points are
unchanged and the ED and Laplacian values, reported in Table 1 are in line with the other functionals employed.
Similar conclusions hold considering the S-S and S-center lines.
Finally, one must point out that, even with this new functional, the
ED accumulation along the S-S line is predicted to be smaller than along
the S-center line, differently from the experimental findings.
Therefore, considering that (see Table 3 ) rVV10$opt$ still
largely overestimate the ILBE, although it slightly reduces the dramatic
overbinding of rVV10($b$=1.0), one can conclude that it is not possible
to achieve, at the same time, a description of energetic and
ED deformation features, in reasonable
agreement with the experimental results, by simply tuning the values of the $b$ and $C$ parameters of the rVV10 functional, which determine
the strength of short/medium- and long-range vdW effects.
As an alternative strategy to reduce the discrepancy between the theoretical
DFT description of TiS$_2$ and the experimental evidence, we have also
investigated the effect of introducing a modified pseudopotential
for S. The S atom is characterized by the [Ne]3s$^2$3p$4$ ground-state
configuration, so that in neutral S atom $d$ orbitals are empty, and,
in simple molecules, they are nearly unpopulated. Nonetheless, it
was shown\cite{dstate} that they can play a significant role: for instance,
even in the small S$_2$ molecule an appreciable improvement in the
binding-energy estimate was observed by adopting for S a pseudopotential
built by taking into account $d$ orbitals, which are expected\cite{dstate} to
give S enhanced polarizability and enable a great variety of bonding
configurations through hybridization.
In fact, although some orbitals (for instance, in this case the $d$ ones)
are not bound states of the neutral atom, it is necessary\cite{dstate}
to choose a partially ionized atomic reference configuration, with a
charge slightly different from the neutral atom, to get
pseudopotentials that are better tailored for applications to
solid-state and extended systems, since the effective configuration
of these systems differs from that of the neutral atom.
Following this approach,
we have generated a novel S ultra-soft pseudopotential,
based on the configuration [Ne]3s$^2$3p$^{3.0}$3d$^{1.0}$, using the
Troullier-Martins pseudization method\cite{tm} with non-linear core corrections.\cite{nlcc} Coupling it with the rVV10 functional, we have developed the rVV10$d$ scheme.
First we have applied the rVV10$d$ functional to the simple case of the S$_2$ molecule, which is in itself an interesting system, being an ubiquitous intermediate in the combustion, atmosphere, and interstellar space.\cite{s2}
With rVV10$d$, while the equilibrium distance is unchanged with respect to
rVV10, we obtain a substantial improvement of the molecular
binding-energy and a slight improvement of the vibrational frequency
(see table S3 of \textit{Supporting information}),
thus confirming the importance of including the $d$ orbitals in such a system.
In particular, considering the binding energy, the more than 20 \% discrepancy with the experimental reference value of the rVV10 original functional, reduces to only about 3 \% with rVV10$d$. We have therefore recomputed the energetic and electronic properties of TiS$_2$ using the new rVV10$d$ functional.
\begin{table}[h!]
\caption{\scriptsize Properties at the (3, -3) CPs along both the S-S and S-center line and at the (3, -1) CP on the S-S line, computed using rVV10$d$ and compared with the results achieved with rVV10.}
\begin{center}
\begin{tabular}{ l| l| c| c| c }
(3, -3) CPs & & & &\\
&line&\textbf{r}$_{max}$ (\si{\angstrom}) & $n$(\textbf{r}$_{max}$) (e/\si{\angstrom}$^3$)&
$\nabla^{2}n$(\textbf{r}$_{max}$) (e/\si{\angstrom}$^5$) \\
\hline
\hline
\multirow{2}{*}{rVV10}&S - S&0.614&1.102&-9.163\\
&S - center&0.614&1.132&-10.052\\
\hline
\multirow{2}{*}{rVV10$d$}&S - S&0.643&1.094&-9.253\\
&S - center&0.643&1.126&-9.990\\
\hline
Experiment\cite{Kasai}&S - S&0.707&1.150&-9.460\\
Theory\cite{Kasai}(SCAN+rVV10)&S - center&0.711&1.080&-8.910\\
\hline
\hline
(3, -1) CP & & & &\\
&line&\textbf{r}$_{min}$ (\si{\angstrom}) & $n$(\textbf{r}$_{min}$) (e/\si{\angstrom}$^3$)&
$\nabla^{2}n$(\textbf{r}$_{min}$) (e/\si{\angstrom}$^5$) \\
\hline
\hline
rVV10&S - S&1.718&0.068&0.486\\
rVV10$d$&S - S&1.718&0.068&0.482\\
\hline
Experiment\cite{Kasai}&S - S& &0.086&0.691\\
Theory\cite{Kasai}(SCAN+rVV10)&S - S& &0.058&0.727\\
\hline
\end{tabular}
\end{center}
\end{table}
As can be seen in Table 4, at the (3, -3) and (3, -1) CPs, with rVV10$d$,
the basic ED properties are very similar to those obtained by the standard
rVV10 functional. However, the effect of replacing rVV10 with rVV10$d$ is
evident looking at Figure 5: one can appreciate a significant ED accumulation
along the S-S line, which is not present using the other DFT functionals
considered, including that (SCAN+rVV10) adopted by
Kasai \textit{et al.}\cite{Kasai}
Although the discrepancy with the experimental description is only partially
resolved (even with rVV10$d$ the ED accumulation along the S-S line is
predicted to be slightly smaller than along the S-center line), a clear
improvement is evident, which we attribute to the increased flexibility of the new S pseudopotentials that is able to better
reproduce the variety of bonding structures and hybridization processes
involving S atoms.
\begin{table}[h!]
\caption{\scriptsize Energetic and structural properties of TiS$_2$, together with the atomic basin properties of the S atom, obtained with the rVV10$d$ and compared with rVV10. The notation used is the same as in Table 3.}
\begin{center}
\begin{adjustbox}{max width=\textwidth}
\begin{tabular}{l| c| c| c| c| c| c| c| c}
&CE (\si{\electronvolt})&ILBE (m\si{\electronvolt}/\si{\angstrom}$^2$)
&$c$ (\si{\angstrom})&CE$^{*}$ (\si{\electronvolt})&ILBE$^{*}$ (m\si{\electronvolt}/\si{\angstrom}$^2$)&V (\si{\angstrom}$^3$)&q$_{A}$ (e)&d(q$_{A}$) (e\si{\angstrom})\\
\hline
rVV10&-15.602&-27.828&5.728&-15.603&-27.888&23.775&-0.860&0.380\\
rVV10$d$&-15.640&-20.888&5.809&-15.646&-20.771&23.761&-0.860&0.390\\
\hline
Experiment&-14.805\cite{ce}& &5.665\cite{Kasai}&-14.805& &23.470\cite{Kasai}&-0.820\cite{Kasai}&0.030\cite{Kasai}\\
Theory (RPA)& &-18.900\cite{bj}& & &-18.900 &23.630\cite{Kasai}&-0.800\cite{Kasai}&0.340\cite{Kasai}\\
Theory (MBD)& &-19.031& & & -19.031& & & \\
\hline
\end{tabular}
\end{adjustbox}
\end{center}
\end{table}
Remarkably, with rVV10$d$, an even more substantial improvement in energetic properties
takes place. In fact, while the CE remains very close to that obtained by
rVV10, suggesting that $d$ orbitals of S do not play a significant role in the
formation of strong, intralayer covalent bonds, the change in the
ILBE is instead considerable: assuming the experimental structure, our
computed ILBE differs by only $\sim$ 2 m\si{\electronvolt}/\si{\angstrom}$^2$ from the theoretical references, i.e. high quality, RPA,\cite{bj} and MBD estimates, which represents a relative error of about 10 \%; such an error is comparable to that obtained by the much more expensive meta-GGA approaches and is much smaller than that
(about 47 \%) found with the original rVV10 functional which clearly overbinds.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\linewidth]{diff_d_inf.PNG}
\end{center}
\caption{ \scriptsize Comparison between the ED deformation obtained by the standard rVV10 functional (panel a) and rVV10$d$ (panel b). The smallest isoline value reported is 0.001 e/\si{\angstrom}$^3$, with red and blue contours that represent charge accumulation and depletion, respectively.}
\end{figure}
As can be seen in Table 5, the good energetic description of rVV10$d$
is preserved even considering a TiS$_2$ configuration where the experimental
lattice constant $c$ is replaced by the $c$ value optimized with rVV10$d$ (whose prediction is only slightly worse than the rVV10 one), thus increasing the confidence in the reliability of this functional.
\begin{table}[h!]
\caption{\scriptsize Optimized lattice constant $c$, ILBE and ILBE* computed for TiS$_2$ and three other TMDs, employing both rVV10 and rVV10$d$ functionals.
The last two lines report the mean error(ME) and the mean absolute
relative error (MARE).}
\begin{center}
\begin{adjustbox}{max width=\textwidth}
\begin{tabular}{l| c| c| c| c| c| c}
&\multicolumn{3}{c}{rVV10$d$}&\multicolumn{3}{c}{rVV10}\\
&ILBE (m\si{\electronvolt}/\si{\angstrom}$^2$)&$c$ (\si{\angstrom})& ILBE* (m\si{\electronvolt}/\si{\angstrom}$^2$)&ILBE (m\si{\electronvolt}/\si{\angstrom}$^2$)&$c$ (\si{\angstrom})& ILBE* (m\si{\electronvolt}/\si{\angstrom}$^2$)\\
\hline
TiS$_2$&-20.771&5.809&-20.888&-27.828&5.728&-27.888\\
ref.\citenum{bj}&-18.900&5.665&-18.900&-18.900&5.665&-18.900\\
\hline
MoS$_2$&-20.296&12.810&-22.147&-29.242&12.530&-29.683\\
ref.\citenum{bj}&-20.530&12.300& -20.530& -20.530& 12.300&-20.530\\
\hline
TaS$_2$&-18.924&6.152&-19.344&-29.367&6.009&-29.369\\
ref.\citenum{bj}&-17.680&5.900& -17.680&-17.680&5.900& -17.680\\
\hline
HfS$_2$&-18.391&5.938&-18.530&-23.452&5.858&-23.603\\
ref.\citenum{bj}&-16.130&5.840& -16.130&-16.130&5.840& -16.130\\
\hline
\hline
ME&-1.285&0.251&-1.917&-9.162&0.105&-9.362\\
\hline
MARE (\%)&7.5 &3.2 &10.7 &50.3 &1.3 &51.1\\
\hline
\end{tabular}
\end{adjustbox}
\end{center}
\end{table}
As far as the dipole moment is concerned, with rVV10$d$ this quantity
is not improved with respect to the experimental estimate (see
Table 5), which can be rationalized as follows:
as a consequence of applying the modified rVV10d functional,
the electronic charge is more widespread around the S atoms,
which leads to a better agreement with experimental ED findings
and to an improved estimate of the ILBE. However this increased
charge delocalization does not occur exclusively in the interlayer
region, but also, for instance, in the intralayer region, particularly
along lines connecting neighboring S atoms. Therefore, the net
delocalization effect (see Figure 5) is approximately isotropic so that
the S dipole moment is not appreciably changed from
that estimated by the original rVV10 functional.
We have also performed additional calculations on TiS$2$ by
adopting vdW-corrected DFT functionals different from rVV10.
Interestingly, while the inclusion of the new pseudopotential involving
$d$ orbitals leads to improvements similar to those observed with
rVV10 in the description of the ED profiles
(with a small charge accumulation in the interlayer region)
and in the ILBE estimate, using vdW-corrected DFT functionals
based on non-local functionals, such as VDW-DF-cx,\cite{VDW-DF-cx}
the same is not the case when vdW-corrected DFT functionals
are adopted, which are more empirical in character (and not directly
dependent on the detailed electron density distribution), such as the
DFT-D3 approach\cite{DFT-D3}, which indicates that the new pseudopotential
is indeed better suited to describe the subtle charge deformations
occurring in this system.
Finally, to further assess the validity of this novel approach, we have applied the rVV10$d$ functional to other three different TMDs, whose chalcogen is always the S atom: MoS$_2$, TaS$_2$, and HfS$_2$. The results for the equilibrium lattice parameter $c$, as well as the ILBE and ILBE* (see Table 6 ), are very promising. The major flexibility of the new pseudopotential, embedded in rVV10$d$ scheme, dramatically improves the energetic description of S-based TMDs, maintaining an appreciable quality of the estimate of the longitudinal lattice parameter $c$: in fact, although the error relative to $c$
is $\sim$ 2 \% bigger than with rVV10, the mean absolute relative
error (MARE) of the ILBE is reduced by about \textit{five} times; this means that
the pronounced overbinding tendency of rVV10 is almost totally eliminated.
\section{Conclusions}
\noindent
We have presented the results of a first principles, DFT investigation,
of the energetic and electronic properties of TiS$_2$, that is a system
where a puzzling discrepancy, between the distribution of the electron
density obtained by X-ray diffraction data and that computed by
state-of-the DFT schemes, has been recently reported, mainly concerning
the interlayer region between S atoms belonging to adjacent layers.
We basically confirm this observation, using the LDA, the (semilocal GGA)
PBE, the (vdW-corrected) rVV10 DFT functionals, and also a modified rVV10
functional (rVV10($b$=10.0)) where the $b$ parameter has been increased so
that short/medium-range vdW interactions are artificially damped.
If instead the $b$ parameter is considerably reduced to $b$=1.0
(rVV10($b$=1.0)) significant changes occur: a stronger ED deformation in
the interlayer region is accompanied by a dramatic (and physically
unrealistic) overestimate of the CE and ILBE terms. Substantially
decreasing (by a 4.5 factor) the other ($C$) rVV10 parameter
and further adjusting the $b$ parameter (rVV10($b$=1.1,$C$=0.00207)
functional) only slightly reduces the dramatic
overbinding of rVV10($b$=1.0) scheme.
Interestingly, a more substantial improvement, in the direction of
making the theoretical description in closer agreement with the
experiment, is obtained by adopting for S a modified, more flexible,
pseudopotential, involving $d$ orbitals (rVV10$d$ functional).
Although the rVV10$d$ estimate of the atomic S dipole moment
still remains substantially larger (see Table 5) than
the experimental one, a consistent improvement in the description of
the ED deformation in the interlayer region and of the ILBE is achieved.
Probably, the residual discrepancy with the experiment could be only
eliminated by an higher-level theoretical, first-principle approach,
such as a Quantum Monte Carlo scheme, in principle able to include the
whole set of correlation effects. Such an approach would be however
extremely expensive computationally because very small ED deformations
could only be accurately reproduced by very long simulations to get a
sufficiently small statistical error associated to differential densities.
In any case, although all fine details in the ED deformation map cannot be
captured by rVV10$d$, nonetheless we suggest that this DFT functional
indeed represents a reasonable compromise between accuracy and efficiency
for improving the theoretical description of TiS$_2$. Additional
investigations indicate that similar improvements, particularly in the
evaluation of energetic terms, can be obtained by DFT calculations
based on the rVV10$d$ functional, applied also to other TMD materials
containing S atoms and characterized by the same structure of TiS$_2$,
such as TaS$_2$, HfS$_2$, and MoS$_2$.
\section{Supporting information}
\section{Intralayer bonding}
The analysis of the Ti-S covalent bond, which characterize the intralayer interactions, is presented here. The electron density (ED) (panel a) ) and Laplacian (panel b) ) profiles computed along the Ti-S line, as well as the ED deformation (panel c) ) and the laplacian map (panel d) ) of the Ti-S intralayer plane, are reported. Since the results achieved with all the functional schemes employed are very close to each other, we show only the PBE results as an example (Figure S1). In Table S1, the characterization of the two (3, -3) CPs and the (3, -1) CP is presented.
As can be seen, very similar values are predicted by
all the DFT functionals applied and these are also in good agreement with
experimental data, with the exception of the LDA estimate of the Laplacian of (3, -3) CP \#1, that exhibits a larger error and the Laplacian values at the (3, -1) CP, which are also substantially overestimated.
\begin{figure}[!]
\begin{center}
\includegraphics[width=14cm, height = 9.3cm]{intra_fin.png}
\end{center}
\caption{\scriptsize Ti-S intralayer bond. a) ED along the Ti-S line; local maxima denote the positions of the (3,-3) CPs, while the local minimum that of the (3,-1) one; b) Laplacian of the ED; c) ED deformation density map in the Ti-S plane; d) Laplacian map in the Ti-S plane. Here results obtained at the PBE level are reported as an
example.}
\end{figure}
\newpage
\begin{table}[!]
\caption{\scriptsize Properties at the (3, -3) CPS and (3, -1) CP identified along the Ti-S line.}
\begin{center}
\begin{adjustbox}{max width=\textwidth}
\begin{tabular}{ l| c| c| c }
(3, -3) CP \#1 &\textbf{r}$_{max}$ (\si{\angstrom}) & $n$(\textbf{r}$_{max}$) (e/\si{\angstrom}$^3$)& $\nabla^{2}n$(\textbf{r}$_{max}$) (e/\si{\angstrom}$^5$)\\
\hline
rVV10&0.353&9.490&-253.367\\
rVV10($b$=1.0)&0.350&9.487&-251.214\\
rVV10($b$=10.0)&0.353&9.495&-252.613\\
PBE&0.352&9.497&-257.543\\
LDA&0.340&9.217&-221.444\\
rVV10$opt$&0.349& 9.477&-255.304\\
\hline
Experiment\cite{Kasai}&0.407&9.350&-253.000\\
Theory\cite{Kasai} (SCAN+rVV10) &0.407&9.310&-250.000\\
\hline
\hline
(3, -3) CP \#2 &\textbf{r}$_{max}$ (\si{\angstrom}) & $n$(\textbf{r}$_{max}$) (e/\si{\angstrom}$^3$)& $\nabla^{2}n$(\textbf{r}$_{max}$) (e/\si{\angstrom}$^5$)\\
\hline
rVV10&1.777&1.171&-10.315\\
rVV10($b$=1.0)&1.782&1.175&-9.978\\
rVV10($b$=10.0)&1.780&1.175&-10.789\\
PBE&1.781&1.172&-10.430\\
LDA&1.819&1.216&-10.091\\
rVV10$opt$&1.778&1.157&-10.165\\
\hline
Experiment\cite{Kasai}&1.714&1.260&-11.300\\
Theory\cite{Kasai}(SCAN+rVV10)&1.707&1.200&-9.870\\
\hline
\hline
(3, -1) CP &\textbf{r}$_{min}$ (\si{\angstrom}) & $n$(\textbf{r}$_{min}$) (e/\si{\angstrom}$^3$)& $\nabla^{2}n$(\textbf{r}$_{min}$) (e/\si{\angstrom}$^5$)\\
\hline
rVV10&1.122&0.453&1.455\\
rVV10($b$=1.0)&1.126&0.448&1.211\\
rVV10($b$=10.0)&1.122&0.450&1.484\\
PBE&1.120&0.458&1.231\\
LDA&1.125&0.461&1.002\\
rVV10$opt$&1.124&0.453&1.245\\
\hline
Experiment\cite{Kasai}& &0.429&3.791\\
Theory\cite{Kasai} (SCAN+rVV10)& &0.421&3.956\\
\hline
\end{tabular}
\end{adjustbox}
\end{center}
\end{table}
\newpage
\section{ED interlayer Laplacians}
\noindent
Here we present the Laplacian maps of the interlayer plane of interest, computed for all the functional employed (the LDA map is very similar to the ones shown).
\begin{figure}[!]
\begin{center}
\includegraphics[width=\linewidth]{lap_supp.PNG}
\end{center}
\caption{\scriptsize Laplacian map in the interlayer plane using different functionals: a) PBE, b) rVV10, c) rVV10($b$=1.0), d) rVV10($b$=10.0), e) rVV10$opt$, and f) rVV10$d$. The Ti atom are represented by grey spheres, while the S atoms are the yellow spheres. The results are reported by using Bader contour lines of [1, 2, 4, 8]$\times$ 10 $^{[-3, \,-2, \,-1, \,0, \,1]}$ e/\si{\angstrom}$^5$. In these maps red (blue) contours are associated to negative (positive) laplacian values, which represents charge accumulation (depletion).}
\end{figure}
\newpage
\section{S$_2$ molecule characterization}
\noindent
The calculations for the S$_2$ molecule, were carried out using the same
approach adopted for the simulations of the TiS$_2$ system (see
section 2 in the main text); the only differences were the chosen
supercell (a cube with sides of 13.2 \si{\angstrom}) and the sampling
of the BZ limited to the $\Gamma$ point.
Calculations were spin-polarized and the
S$_2$ vibrational frequency was computed using the \textit{PHonon} program of QE.
\begin{table}[!]
\caption{\scriptsize Equilibrium distance ($d$), binding energy, and vibrational frequency of the S$_2$ molecule, computed with the rVV10 and rVV10$d$ functionals.}
\begin{center}
\begin{tabular}{ l| c| c| c}
& $d$ (\si{\angstrom}) & binding energy (\si{\electronvolt})& vibrational frequency (cm$^{-1}$)\\
\hline
rVV10&1.91&5.73&689\\
rVV10$d$&1.91&4.57&692\\
\hline
Experiment&1.89\cite{dstate}&4.74\cite{dstate}& 714 $\pm$ 12\cite{s2}\\
\hline
\end{tabular}
\end{center}
\end{table}
\section{Acknowledgements}
We acknowledge funding from Fondazione Cariparo, Progetti di Eccellenza 2017,
relative to the project: ''Engineering van der Waals
Interactions: Innovative paradigm for the control of Nanoscale
Phenomena''.
\newpage
|
\section*{Introduction}
An Ulrich sheaf on a closed subscheme $X$ of ${\mathds {P}}^N$
of dimension $n$ and degree $d$ is a non-zero coherent sheaf ${\mathcal F}$ on $X$ satisfying ${\mathrm H}^*(X,{\mathcal F}(-j))=0$ for $1 \le j \le n$.
In particular, the cohomology table $\{ h^i (X, {\mathcal F}(j)) \}$ of ${\mathcal F}$ is a multiple of the cohomology table of ${\mathds {P}}^n$. It turns out that the reduced Hilbert polynomial ${\mathrm p}_{\mathcal F}(t)=\chi({\mathcal F}(t))/\operatorname{rk}({\mathcal F})$ of an Ulrich
sheaf ${\mathcal F}$ must be:
\[
{\mathrm u}(t):=\frac{d}{n!}\prod_{i=1}^n(t+i).
\]
Ulrich sheaves first appeared in commutative algebra in the 1980s,
namely, in the form of maximally generated maximal Cohen-Macaulay modules
\cite{Ulr84}. Pioneering work of Eisenbud and Schreyer
\cite{ESW03} popularized them in algebraic geometry in
view of their many connections and applications. Eisenbud
and Schreyer asked whether every projective scheme supports an Ulrich
sheaf. That this should be the case is now called a conjecture of
Eisenbud-Schreyer (see also \cite{ES11}). They also proposed another
question about what is the smallest possible rank of an Ulrich sheaf on
$X$. This is called the \emph{Ulrich complexity} $\mathrm{uc}(X)$ of $X$
(cf. \cite{BES17}).
Both the Ulrich existence problem and the Ulrich complexity problem
have been elucidated only for a few cases.
We focus here on the case when $X$ is a hypersurface in ${\mathds {P}}^{n+1}$
over an algebraically closed field
${\mathds k}$.
Using the generalized Clifford algebra, Backelin and Herzog proved in
\cite{BH89} that any hypersurface $X$ has an Ulrich sheaf.
However, their
construction yields an Ulrich sheaf of rank
$d^{\tau(X)-1}$, where $\tau(X)$ is the Chow rank of
$X$ (i.e. the smallest length of an expression of the defining
equation of $X$ as sums of products of $d$ linear forms),
often much bigger than $\mathrm{uc}(X)$.
Looking in more detail at the Ulrich complexity problem for smooth
hypersurfaces of degree $d$ in ${\mathds {P}}^{n+1}$, the situation is well-understood for
arbitrary $n$ only for $d=2$.
Indeed, in this case the only indecomposable Ulrich bundles on $X$ are spinor
bundles, which have rank $2^{\lfloor (n-1)/2 \rfloor}$ \cite{BEH87}. On the
other hand, for $d \ge 3$, the Ulrich
complexity problem is wide open except for a very few
small-dimensional cases.
For instance, smooth cubic curves and surfaces
always $X$ satisfy $\mathrm{uc}(X)=1$, while for smooth cubic threefolds
$X$ we have $\mathrm{uc}(X)=2$,
(cf. \cite{Bea00, Bea02, LMS15}).
When $X$ is a smooth cubic fourfold, which is the main object of this
paper, there can be several possibilities. In any case, $X$ does not
have an Ulrich bundle of rank $1$, but some $X$ can have an Ulrich
bundle ${\mathcal F}$ of rank $2$. Since ${\mathcal F}$ is globally generated, one can
consider the locus defined by a general global section of ${\mathcal F}$, which
is a del Pezzo surface of degree $5$. The moduli space of cubic
fourfolds containing a del Pezzo surface of degree $5$ forms a divisor
${\mathcal C}_{14} \subset {\mathcal C}$, so a general cubic fourfold $X$ has
$\mathrm{uc}(X) \ge 3$.
A few more cubic fourfolds which have an Ulrich bundle of rank $3$ or $4$
have been reported very recently by Troung and Yen \cite{TY20}. However, all
these cases are special cubic fourfolds which contain a surface
not homologous to a complete intersection. Indeed, it turns out that
the Ulrich complexity of a very general cubic fourfold is divisible by $3$
and at least $6$, see \cite{KS20}. On the other hand, it is known that a
general cubic fourfold has a rank $9$ Ulrich bundle (cf. \cite{IM14,
Man19, KS20}). Therefore, the Ulrich complexity of a (very) general
cubic fourfold is either $6$ or $9$.
The goal of this paper is to prove the following result. This implies
that a general cubic fourfold $X$ satisfies $\mathrm{uc}(X)=6$.
\begin{thm*}
Any smooth cubic fourfold admits an Ulrich bundle
of rank $6$.
\end{thm*}
To achieve this, we use a deformation theoretic
argument. Let us sketch briefly the strategy of our proof.
As a warm-up it, let us review a construction of a rank-$2$
Ulrich bundle on a smooth cubic threefold.
First, starting from a line $L$ contained in the threefold, one
constructs an ACM bundle of rank $2$ having $(c_1, c_2) = (0, L)$.
Such a bundle is unstable
since it has a unique global section which vanishes along $L$.
By choosing a line $L'$ disjoint from $L$, we may take an elementary
modification of it so that we have a simple and semistable sheaf ${\mathcal E}$
of $(c_1, c_2) = (0,2L)$. The sheaf ${\mathcal E}$ is not Ulrich, but one can show that its general deformation becomes Ulrich. A similar argument
is used to prove the existence of rank $2$ Ulrich bundles on K3
surfaces \cite{Fae19} and prime Fano threefolds \cite{BF11}.
For fourfolds, twisted cubics play a central role in the
construction, rather than lines.
Note that twisted cubics in $X$ form a $10$-dimensional family. For each twisted
cubic $C \subset X$, its linear span $V = \langle C \rangle $
defines a linear section $Y \subset X$ which is a cubic surface. When
$Y$ is smooth, the rank-$3$ sheaf $\mathcal G = \ker[3 \mathcal {\mathcal O}_X
\to {\mathcal O}_Y(C)]$ is stable.
The family of such stable sheaves of rank $3$ forms an $8$-dimensional
moduli space, which is indeed a very well-studied
smooth hyperk\"ahler manifold \cite{LLSS17, LLMS18}.
We will call them Lehn-Lehn-Sorger-van Straten sheaves and the
Lehn-Lehn-Sorger-van Straten eightfold.
To construct an Ulrich bundle of rank 6,
we first consider two Lehn-Lehn-Sorger-van Straten
sheaves of rank 3 associated with two
points of this manifold. We choose the points so that the associated
pair of twisted cubics
spans the same smooth surface section of $X$ and intersects at
$4$ points. Then, we define a simple sheaf ${\mathcal E}$ as extension of such
sheaves and show that this enjoys a large part of the
cohomology vanishing required to be Ulrich.
In particular, this will show that ${\mathcal E}$ lies
in the Kuznetsov category $\operatorname{Ku}(X)$ of $X$, \cite{Kuz04}.
Finally, we obtain an Ulrich sheaf by taking a generic deformation
${\mathcal F}$ of ${\mathcal E}$
in the moduli space of simple sheaves over $X$ and using that the cohomology
vanishing of ${\mathcal E}$ propagates to ${\mathcal F}$ by semicontinuity.
This step relies on
deformation-obstruction theory of the sheaf as developed in \cite{KM09,BLMNPS19} and makes substantial use
of the fact that $\operatorname{Ku}(X)$ is a K3 category.
The structure of this paper is as follows. In Section
\ref{section:background}, we recall basic notions and set up some
background. In Section \ref{section:twisted cubics}, we introduce an ACM
bundle of rank $6$ which arises as a (higher) syzygy sheaf of a
twisted cubic and review some material on Lehn-Lehn-Sorger-van Straten sheaves as syzygy sheaves. Then
we take an elementary modification to define a strictly semistable
sheaf ${\mathcal E}$ of rank $6$ whose reduced Hilbert polynomial is
${\mathrm u}(t)$. In Section \ref{section:smoothing}, we show that a general
deformation of ${\mathcal E}$ is Ulrich. We first prove this claim for cubic
fourfolds which do not contain surfaces of small degrees other than
linear sections. Then we extend it for every smooth cubic fourfold.
\medskip
\begin{ack}
We wish to thank F\'ed\'eration Bourgogne Franche-Comt\'e Math\'ematiques FR CNRS 2011 for supporting the visit of Y.K. in Dijon.
We would like to thank Frank-Olaf Schreyer and Paolo Stellari for valuable advice and helpful discussion.
\end{ack}
\section{Background} \label{section:background}
Let us collect here some basic material. We work over an algebraically
closed field ${\mathds k}$ of characteristic other than $2$.
\subsection{Background definitions and notation}
Consider a smooth connected $n$-dimensional projective subvariety $X
\subseteq {\mathds {P}}^N$ and denote by $H_X$ the hyperplane divisor on
$X$ and ${\mathcal O}_X(1) = {\mathcal O}_X(H_X)$. Given a coherent sheaf ${\mathcal F}$ on $X$
and $t \in {\mathds Z}$, write ${\mathcal F}(t)$ for ${\mathcal F} \otimes {\mathcal O}_X(t H_X)$.
Let ${\mathcal F}$ be a torsion-free sheaf on $X$.
The reduced Hilbert polynomial of ${\mathcal F}$ is defined as
\[
{\mathrm p}_{\mathcal F}(t) := \frac{1}{\operatorname{rk}({\mathcal F})}\chi({\mathcal F}(t)) \in {\mathds Q}[t].
\]
Let ${\mathcal F}, {\mathcal G}$ be torsion-free sheaves on $X$. We say that ${\mathrm p}_{\mathcal F} < {\mathrm p}_{\mathcal G}$ if ${\mathrm p}_{\mathcal F}(t) < {\mathrm p}_{\mathcal G}(t)$ for $t \gg 0$.
The slope of ${\mathcal F}$
is defined as:
\[
\mu({\mathcal F})=\frac{c_1({\mathcal F})\cdot H_X^{n-1}}{\operatorname{rk}({\mathcal F})}.
\]
A torsion-free sheaf ${\mathcal F}$ on $X$ is \emph{stable}
(respectively, \emph{semistable, $\mu$-stable, $\mu$-semistable})
if, for any subsheaf $0 \ne {\mathcal F}' \subsetneq {\mathcal F}$, we have:
\[
{\mathrm p}_{{\mathcal F}'} < {\mathrm p}_{{\mathcal F}}, \qquad \mbox{(respectively, ${\mathrm p}_{{\mathcal F}'} \le
{\mathrm p}_{{\mathcal F}}$, $\mu({\mathcal F}') < \mu({\mathcal F})$, $\mu({\mathcal F}') \le \mu({\mathcal F})$}.
\]
A polystable sheaf is a direct sum of stable sheaves having the same
reduced Hilbert polynomial.
\subsection{ACM and Ulrich sheaves}
We are mostly interested in coherent sheaves on $X$ which admit nice
minimal free resolutions over ${\mathds {P}}^N$, namely ACM and Ulrich
sheaves. Equivalently, such properties are characterized by cohomology
vanishing conditions as follows:
\begin{defn}
Let $X \subseteq {\mathds {P}}^N$ be as above, and let ${\mathcal F}$ be a coherent sheaf
on $X$. Then ${\mathcal F}$ is:
\begin{enumerate}[label=\roman*)]
\item \emph{ACM} if it is locally Cohen-Macaulay and ${\mathrm H}^i(X, {\mathcal F}(j))=0$ for $0<i<n$ and $j \in {\mathds Z}$.
\item \emph{Ulrich} if ${\mathrm H}^i(X, {\mathcal F}(-j))=0$ for $i \in {\mathds Z}$ and $1 \le j \le n$.
\end{enumerate}
\end{defn}
We refer to \cite[Proposition 2.1]{ESW03} for several equivalent
definitions for Ulrich sheaves. In particular, every Ulrich sheaf is
ACM. If $X$ is smooth, then a coherent sheaf is locally Cohen-Macaulay if and only if is locally free. Moreover, for Ulrich sheaves,
(semi-)stability is equivalent to $\mu$-(semi-)stability, see \cite{CH12}.
Let us review some previous works on the existence of Ulrich bundles
on a smooth cubic fourfold $X$, possibly of small rank.
In terms of Hilbert polynomial, an Ulrich bundle ${\mathcal U}$ satisfies:
\[
{\mathrm p}_{\mathcal U}(t)={\mathrm u}(t) := \frac 18 (t+4)(t+3)(t+2)(t+1).
\]
Note that $X$ carries an Ulrich line bundle if and only if it is linearly determinantal, which is impossible since a determinantal hypersurface
is singular along a locus of codimension $3$. $X$ carries a rank $2$
Ulrich bundle if and only if it is linearly Pfaffian. Equivalently,
such an $X$ contains a quintic del Pezzo surface \cite{Bea00}. Note that
a Pfaffian cubic fourfold also carries a rank $5$ Ulrich
bundle \cite{Man19}. For rank $3$ and $4$, Truong and Yen provided computer-aided construction of a rank $3$ Ulrich bundle on a general element in the moduli of special cubic fourfolds $\mathcal C_{18}$ of discriminant
$18$, and of a rank $4$ Ulrich bundle on a general element in
$\mathcal C_{8}$ \cite{TY20}.
All the above cases were made over special cubic fourfolds,
i.e., they contain a surface which is not homologous to a
complete intersection. Such cubic fourfolds form a countable union of
irreducible divisors in the moduli space of smooth cubic fourfolds
$\mathcal C$. We refer to \cite{Has00} for the convention and more
details. On a very general cubic fourfold $X$ (so that any surface
contained in $X$ is homologous to a complete intersection), it is easy
to find the following necessary condition on Chern classes of a
coherent sheaf to be Ulrich:
\begin{prop}[{\cite[Proposition 2.5]{KS20}}]\label{Proposition:KS20}
Let ${\mathcal E}$ be an Ulrich bundle of rank $r$ on a very general cubic fourfold $X \subset \mathbb{P}^5$. Let $c_i := c_i ({\mathcal E}(-1))$. Then $r$ is divisible by $3$, $r \ge 6$, and
\[
c_1 = 0, \qquad c_2 = \frac{1}{3}r H^2, \qquad c_3 = 0, \qquad c_4 = \frac{1}{6}r(r-9).
\]
\end{prop}
The existence of rank $9$ Ulrich bundles on a general cubic fourfold
$X$ is known according to \cite{IM14,
Man19, KS20}. Therefore, the Ulrich complexity of a very general
cubic fourfold is either $6$ or $9$.
It is thus natural to ask the question: \textit{Does a smooth cubic fourfold carry an Ulrich bundle of rank $6$}? The goal of this paper is to give a
positive answer to this question.
In particular, the Ulrich complexity $\mathrm{uc}(X)$ of a (very)
general cubic fourfold $X$ is $6$.
\subsection{Reflexive sheaves}
Let ${\mathcal E}$ be a torsion-free sheaf on a smooth connected projective $n$-dimensional
variety $X$. The following lemma is standard.
\begin{lem} \label{reflexive}
For each $k \in \{0,\ldots,n-2\}$ there is $p_k \in {\mathds Q}[t]$, with
$\deg(p_k) \le k$ such that:
\[
{\mathrm h}^{k+1}({\mathcal E}(-t))=p_k(t), \qquad {\mathrm h}^0({\mathcal E}(-t))=0\qquad \mbox{for $t \gg 0$},
\]
Assume ${\mathcal E}$ is reflexive. Then $\forall k \in \{0,\ldots,n-3\}$ there is $q_k \in {\mathds Q}[t]$, with
$\deg(q_k) \le k$ such that:
\[
{\mathrm h}^{k+2}({\mathcal E}(-t))=q_k(t), \qquad {\mathrm h}^0({\mathcal E}(-t))={\mathrm h}^1({\mathcal E}(-t))=0 \qquad \mbox{for $t \gg 0$},
\]
Moreover, ${\mathcal E}$ is locally free if and only if $p_k=0$ for all $k
\in \{0,\ldots,n-2\}$, equivalently if $q_k=0$ for all $k \in \{0,\ldots,n-3\}$.
\end{lem}
\begin{proof}
Given positive integers $p,q$ with $p+q \le n$, Serre duality and the
local-global spectral sequence give, for all $t \in {\mathds Z}$:
\begin{equation}
\label{SS}
{\mathrm H}^{n-p-q}({\mathcal E}(-t))^\vee \simeq \operatorname{Ext}^{p+q}_X({\mathcal E},\omega_X(t))
\Leftarrow {\mathrm H}^p(\mathcal{E}xt_X^q({\mathcal E},\omega_X) \otimes {\mathcal O}_X(t)) = E_2^{p,q}.
\end{equation}
For $t \gg 0$ and $p > 0$ we have ${\mathrm H}^p(\mathcal{E}xt_X^q({\mathcal E},\omega_X)
\otimes {\mathcal O}_X(t))=0$ by Serre vanishing. Then:
\[
{\mathrm h}^{n-q}({\mathcal E}(-t)) \simeq {\mathrm h}^0(\mathcal{E}xt_X^q({\mathcal E},\omega_X) \otimes
{\mathcal O}_X(t)), \qquad \mbox{for $t \gg 0$}.
\]
Hence ${\mathrm h}^{n-q}({\mathcal E}(-t))$ is a rational polynomial function of
$t$ for $t \gg 0$.
By \cite[Proposition 1.1.10]{HL10}, since ${\mathcal E}$ is torsion-free, for $q \ge 1$ we have:
\[
\operatorname{codim}(\mathcal{E}xt_X^q({\mathcal E},\omega_X)) \ge q+1,
\]
while when ${\mathcal E}$ is reflexive, for $q \ge 1$:
\[
\operatorname{codim}(\mathcal{E}xt_X^q({\mathcal E},\omega_X)) \ge q+2.
\]
Thus, for $t \gg 0$, the degree of the polynomial function
${\mathrm h}^{n-q}({\mathcal E}(-t))$ is at most $n-q-1$, actually of $n-q-2$ if
${\mathcal E}$ is reflexive.
Finally, ${\mathcal E}$ is locally free if and only if
$\mathcal{E}xt_X^q({\mathcal E},\omega_X)=0$ for all $q>1$. Since this happens if and
only if ${\mathrm h}^0(\mathcal{E}xt_X^q({\mathcal E},\omega_X) \otimes {\mathcal O}_X(t))$ for $t \gg
0$, the last statement follows.
\end{proof}
\subsection{Minimal resolutions and syzygies}
We recall some notions from commutative algebra.
Let $R = {\mathds k} [x_0,\cdots, x_N]$ be a polynomial ring over a field ${\mathds k}$ with the
standard grading, and let $R_X = R/I_X$ be the homogeneous coordinate
ring of $X$ where $I_X$ is the ideal of $X$. Let $\Gamma$ be a
finitely generated graded $R_X$-module.
The minimal free resolution of $\Gamma$ over $R_X$ is constructed by
choosing minimal generators of $\Gamma$ of degrees
$(a_{0,0},\ldots,a_{0,r_0})$ so that there is a surjection
\[
F_0 = \bigoplus_{j=0}^{r_0} R_X (-a_{0,j}) \twoheadrightarrow \Gamma.
\]
Taking a minimal set of generators of degrees
$(a_{1,0},\ldots,a_{1,r_1})$ of its kernel we get a minimal
presentation of $\Gamma$ of the form $F_1 = \bigoplus_{j=0}^{r_1}
R_X(-a_{1,j}) \to F_0$. Repeating this process, we have a free resolution of $\Gamma$:
\[
F_{\bullet} (\Gamma) : \cdots \to F_i \stackrel{d_i} \longrightarrow
F_{i-1} \stackrel{d_{i-1}} \longrightarrow \cdots \longrightarrow F_1
\stackrel{d_1} \longrightarrow F_0 \to \Gamma \to 0, \qquad
\mbox{with:} \qquad F_i = \bigoplus_{j=0}^{r_i} R_X (-a_{i,j}).
\]
Note that the resolution obtained this way is minimal, i.e. $d_i
\otimes_{R_X} {\mathds k} = 0$ for every $i$, and is unique up to homotopy,
see \cite[Corollary 1.4]{Eis80}. In general, it has infinitely many terms.
We define the \textit{minimal resolution} of a coherent sheaf ${\mathcal F}$ on
$X$ as the sheafification of the minimal graded free resolution of its
module of global sections $\Gamma_{*}({\mathcal F}) = \bigoplus_{j \in {\mathds Z}} \Gamma(X, {\mathcal F} (j))$, provided that this is finitely
generated. In this case, for $i \in {\mathds N}$, we call \textit{i-th syzygy}
of ${\mathcal F}$ the sheafification of $\operatorname{Im}(d_i)$ and we denote this by
$\Sigma^X_i({\mathcal F})$. Of course for positive $j$ we have $\Sigma^X_{i+j}({\mathcal F}) \simeq \Sigma^X_j\Sigma^X_i({\mathcal F})$.
\subsection{Matrix factorizations and ACM/Ulrich sheaves}
We recall the notion of matrix factorization which is introduced by Eisenbud \cite{Eis80} to study free resolutions over hypersurfaces.
\begin{defn}
Let $X \subseteq {\mathds {P}}^N$ be a hypersurface defined by a homogeneous polynomial $f$ of degree $d$, and let ${\mathcal F}$ and ${\mathcal G}$ are free graded ${\mathcal O}_{{\mathds {P}}^N}$-modules. A pair of morphisms $\varphi : {\mathcal F} \to {\mathcal G}$ and $\psi : {\mathcal G} (-d) \to {\mathcal F}$ is called a \emph{matrix factorization} of $f$ (of $X$) if
\[
\varphi \circ \psi = f \cdot id_{{\mathcal G}(-d)}, \qquad \psi (d) \circ \varphi = f \cdot id_{{\mathcal F}}.
\]
\end{defn}
Matrix factorizations have a powerful application to ACM/Ulrich bundles as follows:
\begin{prop}[{\cite[Corollary 6.3]{Eis80}}]
The association
\[
(\varphi, \psi) \mapsto M_{(\varphi, \psi)} := \operatorname{coker} \varphi
\]
induces a bijection between the set of equivalence classes of reduced matrix factorizations of $f$ and the set of isomorphism classes of indecomposable ACM sheaves. In particular, when $(\varphi, \psi)$ is completely linear, that is, $\varphi : {\mathcal O}_{{\mathds {P}}^N}(-1)^{\oplus t} \to {\mathcal O}_{{\mathds {P}}^N}^{\oplus t}$ for some $t \in {\mathds Z}$ then the corresponding sheaf is Ulrich.
\end{prop}
\subsection{Twisted cubics and Lehn-Lehn-Sorger-van Straten eightfold}\ Let us briefly recall how can we construct a rank 2 Ulrich bundle on a cubic threefold $X$ via deformation theory. If there is such an Ulrich bundle ${\mathcal F}$, then ${\mathcal F}(-1)$ must have the Chern classes $(c_1, c_2) = (0, 2)$ by Riemann-Roch. Note that $X$ has an ACM bundle ${\mathcal F}_1$ of rank $2$ with $(c_1, c_2) = (0,1)$ which fits into the following short exact sequence
\[
0 \to {\mathcal O}_X \to {\mathcal F}_1 \to {\mathcal I}_{\ell} \to 0
\]
where $\ell \subset X$ is a line. We see that ${\mathcal F}_1$ is unstable due to its unique global section. We can take an elementary modification with respect to ${\mathcal O}_{\ell^{\prime}}$ where $\ell^{\prime} \subset X$ is a line disjoint to $\ell$. The resulting sheaf ${\mathcal F}_2 := \ker \left[ {\mathcal F}_1 \to {\mathcal O}_{\ell^{\prime}} \right]$ is simple, strictly semistable, and non-reflexive. One can check that its general deformation is stable and locally free, and becomes Ulrich after twisting by $O_X(1)$. One major difference between the case of cubic threefolds is that not lines but twisted cubics play a significant role both in finding an ACM bundle (of same $c_1$ as Ulrich) and taking an elementary modification.
Let $X \subset {\mathds {P}}^5$ be a smooth cubic fourfold which does not contain a plane, and let $M_3 (X)$ be the irreducible component of the Hilbert scheme of $X$ containing the twisted cubics. Then $M_3(X)$ is a smooth irreducible projective variety of dimension $10$ \cite[Theorem A]{LLSS17}. Let $C$ be a twisted cubic contained in $X$, and $V \simeq {\mathds {P}}^3$ be its linear span. According to \cite{LLSS17}, the natural morphism $C \mapsto V \in Gr(4,6)$ factors through a smooth projective eightfold ${\mathcal Z}^{\prime}$ so that $M_3(X) \to {\mathcal Z}^{\prime}$ is a ${\mathds {P}}^2$-fibration. In ${\mathcal Z}^{\prime}$, there is an effective divisor coming from non-CM twisted cubics on $X$ which induces a further contraction ${\mathcal Z}^{\prime} \to {\mathcal Z}$ so that ${\mathcal Z}$ is a smooth hyperk\"ahler eightfold which contains $X$ as a Lagrangian submanifold, and the map ${\mathcal Z}^{\prime} \to {\mathcal Z}$ is the blow-up along $X$ \cite[Theorem B]{LLSS17}. The variety ${\mathcal Z}$ is called the Lehn-Lehn-Sorger-van Straten eightfold.
We are interested in a moduli description of
${\mathcal Z}^{\prime}$. Let $Y := V \cap X$ be a cubic surface containing
$C$. The sheaf ${\mathcal I}_{C/Y} (2)$ is indeed an Ulrich line bundle on $Y$, and hence it fits into the following short exact sequence
\[
0 \to {\mathcal G}_C \to 3 {\mathcal O}_X \to {\mathcal I}_{C/Y}(2) \to 0.
\]
Lahoz, Lehn, Macr\`i and Stellari showed that the sheaf ${\mathcal G}_C$ is
stable, and the moduli space of Gieseker stable sheaves with the same Chern character is isomorphic to ${\mathcal Z}^{\prime}$ \cite{LLMS18}. Since we are only interested in general CM twisted cubics and corresponding Lehn-Lehn-Sorger-van Straten sheaves, we may regard that a general point of the Lehn-Lehn-Sorger-van Straten eightfold ${\mathcal Z}$ corresponds to a rank 3 sheaf ${\mathcal G}_C$, where $C$ is a CM twisted cubic on $X$, even when $X$ potentially contains a plane.
\section{Syzygies of twisted cubics} \label{section:twisted cubics}
Let $X \subset {\mathds {P}}^5$ be a smooth cubic fourfold.
\subsection{Twisted cubics and $6$-bundles}
\label{Subsection:construction of S}
Here we show that taking the fourth syzygy of the structure sheaf of
a twisted cubic $C$ is a vector
bundle of rank $6$ which admits a trivial subbundle of rank
$3$. Factoring out this quotient gives back the second syzygy of
$C$, with a degree shift. We will use this filtration later on.
\begin{prop} \label{syzygy of twisted}
Let $C \subset X$ be a twisted cubic, $V$ its linear span and set $Y=X\cap V$. Put:
\[
{\mathcal S}=\Sigma^X_4({\mathcal O}_C(5)), \qquad
{\mathcal G}_C=\Sigma^X_1({\mathcal I}_{C/Y}(2)).
\]
Then ${\mathcal S}$ is an ACM sheaf of rank $6$ on $X$ with:
\[
{\mathrm p}_{\mathcal S}(t)=\frac 1 8 (t+2)^2(t+1)^2, \qquad {\mathrm H}^*({\mathcal S}(-1))={\mathrm H}^*({\mathcal S}(-2))=0.
\]
Moreover, ${\mathrm h}^0(X,{\mathcal S})=3$ and there is an exact sequence:
\begin{equation}
\label{hatSO}
0 \to 3{\mathcal O}_X \to {\mathcal S} \to {\mathcal G}_C \to 0.
\end{equation}
\end{prop}
To keep notation lighter, we remove the subscript $C$ from ${\mathcal G}_C$ so we just write ${\mathcal G}$, as soon as no confusion occurs, i.e. until \S \ref{second conic}.
\begin{proof}
For the sake of this proof, for any integer $i$ we omit writing
${\mathcal O}_C$ from expressions of the form $\Sigma_i^X({\mathcal O}_C)$ and
$\Sigma_i^Y({\mathcal O}_C)$, so that for instance:
\begin{equation}
\label{SigmaI}
\Sigma_1^X \simeq {\mathcal I}_{C/X}, \qquad
\Sigma_1^Y \simeq {\mathcal I}_{C/Y}.
\end{equation}
The curve $C$ is Cohen-Macaulay of degree $3$ and arithmetic genus
$0$, its linear span $V$ is a ${\mathds {P}}^3$, and the linear section $Y$ is a cubic surface equipped with the Ulrich line bundle ${\mathcal I}_{C/Y}(2)$.
Hence, we have a linear resolution on $V$:
\[
0 \to 3{\mathcal O}_V(-3) \xrightarrow{M} 3{\mathcal O}_V(-2) \to {\mathcal I}_{C/Y} \to 0,
\]
where $M$ is a matrix of linear forms whose determinant is an
equation of $Y$ in $V$. Put $G_1=3{\mathcal O}_Y(-2)$ and
$G_2=3{\mathcal O}_Y(-3)$.
Thanks to \cite[Theorem 6.1]{Eis80},
taking the adjugate matrix $M'$ of $M$ forms a matrix factorization $(M, M')$
of $Y$ which provides the following $2$-periodic resolution on $Y$
(we still denote by $M, M'$ the reduction of $M$ and $M'$ modulo $Y$):
\begin{equation}
\label{infiniteY}
\cdots \xrightarrow{M'} G_2(-3) \xrightarrow{M} G_1(-3) \xrightarrow{M'} G_2 \xrightarrow{M} G_1 \to {\mathcal I}_{C/Y} \to 0.
\end{equation}
This gives, for all $i \in {\mathds N}$:
\begin{equation}
\label{periodic}
\Sigma^Y_{2i+1} \simeq {\mathcal I}_{C/Y}(-3i).
\end{equation}
Next, set $K_0 = {\mathcal O}_X$, $K_1=2{\mathcal O}_X(-1)$, $K_2 = {\mathcal O}_X(-2)$ and write the obvious Kozsul resolution:
\begin{equation}
\label{koszulY}
0 \to K_2 \to K_1 \to K_0 \to {\mathcal O}_Y \to 0.
\end{equation}
We look now at the exact sequence:
\begin{equation}
\label{ideals}
0 \to {\mathcal I}_{Y/X} \to {\mathcal I}_{C/X} \to {\mathcal I}_{C/Y} \to 0.
\end{equation}
Set $F_1=3{\mathcal O}_X (-2)$, $F_2=3{\mathcal O}_X(-3)$.
We proceed now in two directions. On one hand, the composition
$F_1 \to G_1 \to \Sigma_1^Y$ lifts to $F_1 \to \Sigma_1^X$ to give
a diagram (we
omit zeroes all around for brevity):
\[
\xymatrix@-2ex{
K_2 \ar[d] \ar[r] & K_1 \ar[d] \ar[r] & {\mathcal I}_{Y/X} \ar[d]\\
\Sigma_2^X \ar[d] \ar[r] & F_1 \oplus K_1 \ar[d] \ar[r] & \Sigma_1^X \ar[d] \\
\Sigma_1^X\Sigma_1^Y\ar[r] &F_1 \ar[r] & \Sigma_1^Y
}
\]
Looking at the above diagram and using that $\Gamma_*(K_2)$ is free, we get:
\begin{equation}
\label{greater3}
0 \to K_2 \to \Sigma_2^X \to \Sigma_1^X\Sigma_1^Y \to 0, \qquad
\Sigma_{i+1}^X \simeq \Sigma_i^X\Sigma_1^Y, \qquad \forall i \ge 2.
\end{equation}
Next, \eqref{ideals},
\eqref{SigmaI} and \eqref{infiniteY}
induce a diagram
\[
\xymatrix@-2ex{
F_1 \otimes {\mathcal I}_{Y/X} \ar@{=}[d]\ar[r] & \Sigma_1^X \Sigma_1^Y
\ar[r] \ar[d] &
\Sigma_2^Y\ar[d] \\
F_1 \otimes {\mathcal I}_{Y/X} \ar[r] & F_1 \ar[r] \ar[d] & G_1 \ar[d]\\
& \Sigma_1^Y \ar@{=}[r] & \Sigma_1^Y
}
\]
This in turn gives the exact sequence
\begin{equation}
\label{hatS}
0 \to F_1 \otimes {\mathcal I}_{Y/X} \to \Sigma_1^X\Sigma_1^Y \to \Sigma_2^Y \to 0. \end{equation}
Lifting $F_2 \to \Sigma_2^Y$ to $F_2 \to \Sigma_1^X\Sigma_1^Y$,
we get the exact diagram:
\[
\xymatrix@-2ex{
F_1 \otimes K_2 \ar[r] \ar[d] &F_1 \otimes K_1 \ar[d] \ar[r] & F_1 \otimes {\mathcal I}_{Y/X} \ar[d] \\
\Sigma_2^X\Sigma_1^Y \ar[d] \ar[r] & F_1 \otimes K_1 \oplus F_2
\ar[r] \ar[d] & \Sigma_1^X\Sigma_1^Y \ar[d] \\
\Sigma_1^X\Sigma_2^Y \ar[r] & F_2 \ar[r]& \Sigma_2^Y.
}
\]
Using the diagram and the fact that $\Gamma_*(F_1 \otimes K_2)$ is
free we get:
\begin{equation}
\label{greater2}
0 \to F_1 \otimes K_2 \to \Sigma_2^X\Sigma_1^Y \to \Sigma_1^X
\Sigma_2^Y \to 0, \qquad \Sigma_{i+1}^X\Sigma_1^Y \simeq
\Sigma_i^X\Sigma_2^Y, \forall i \ge 2.
\end{equation}
Repeating once more this procedure and using the periodicity of \eqref{infiniteY} we get:
\[
0 \to F_2 \otimes {\mathcal I}_{Y/X} \to \Sigma_1^X \Sigma_2^Y \to \Sigma_3^Y \to 0.
\]
Then, using \eqref{periodic} and lifting $F_1(-3) \to {\mathcal I}_{C/Y}(-3) \simeq \Sigma_3^Y$ to $F_1(-3) \to \Sigma_1^X \Sigma_2^Y$, we have the exact sequence:
\[
0 \to F_2 \otimes K_2 \to \Sigma_2^X \Sigma_2^Y \to \Sigma_1^X
\Sigma_3^Y \to 0.
\]
Summing up, \eqref{greater3} and \eqref{greater2} give $\Sigma_4^X \simeq
\Sigma_3^X\Sigma_1^Y \simeq \Sigma_2^X\Sigma_2^Y$, so that the above
sequence tensored with ${\mathcal O}_X(5)$ becomes:
\[
0 \to 3{\mathcal O}_X \to {\mathcal S} \to \Sigma_1^X({\mathcal I}_{C/Y}(2)) \to 0,
\]
which is the sequence appearing in the statement. The fact that
${\mathrm h}^0(X,{\mathcal S})=3$ is clear from the sequence.
Since $X$ is smooth and $C
\subset X$ is arithmetically Cohen-Macaulay of
codimension $3$, the syzygy sheaf $\Sigma_4^X$ is ACM and hence locally free.
Looking at the above resolution we compute the following invariants of
${\mathcal S}$:
\[
\operatorname{rk}({\mathcal S})=6, \qquad c_1({\mathcal S}) = 0, \qquad c_2({\mathcal S})=H^2, \qquad
{\mathrm p}_{\mathcal S}(t)=\frac 1 8 (t+1)^2(t+2)^2.
\]
It remains to prove ${\mathrm H}^*({\mathcal S}(-1))={\mathrm H}^*({\mathcal S}(-2))=0$. By
\eqref{hatSO}, it suffices to show ${\mathrm H}^*({\mathcal G}(-1))={\mathrm H}^*({\mathcal G}(-2))=0$. By definition we have
\begin{equation}
\label{SI}
0 \to {\mathcal G} \to 3{\mathcal O}_X \to {\mathcal I}_{C/Y}(2) \to 0,
\end{equation}
and ${\mathcal I}_{C/Y}(2)$ is Ulrich on $Y$ so
${\mathrm H}^*({\mathcal I}_{C/Y}(1))={\mathrm H}^*({\mathcal I}_{C/Y})=0$. We conclude that ${\mathrm H}^*({\mathcal G}(-1))={\mathrm H}^*({\mathcal G}(-2))=0$.
\end{proof}
Along the way we found the following minimal free resolution of
${\mathcal O}_C$ over $X$:
\[
\begin{array}{ccccccccc}
& 3 {\mathcal O}_X(-5) & & 9{\mathcal O}_X(-4) & & {\mathcal O}_X(-2) & & 2{\mathcal O}_X(-1) & \\
\cdots \to & \oplus & \stackrel{d_4} \to & \oplus &\stackrel{d_3} \to & \oplus & \stackrel{d_2} \to & \oplus & \stackrel{d_1} \to {\mathcal O}_X \to {\mathcal O}_C \to 0.\\
& 9{\mathcal O}_X(-6) & & 3{\mathcal O}_X(-5) & & 9 {\mathcal O}_X(-3)& & 3{\mathcal O}_X(-2) &
\end{array}
\]
This is an instance of Shamash's resolution. It becomes periodic after three steps.
We record that ${\mathcal S}$ fits into:
\[
\xymatrixrowsep{10pt}
\xymatrixcolsep{10pt}
\xymatrix{
\cdots \ar[r] & 9 {\mathcal O}_X(-2) \oplus 3 {\mathcal O}_X(-3) \ar^-{d_5}[rr] \ar[dr] && 3 {\mathcal O}_X \oplus 9 {\mathcal O}_X(-1)
\ar[dr] \ar^-{d_4}[rr]&& 9 {\mathcal O}_X(1) \oplus 3 {\mathcal O}_X \ar[r] & \cdots \\
& & \Sigma_5^X({\mathcal O}_C(5)) \ar[ur] & & {\mathcal S} \ar[ur] & & &
}
\]
Let us fix the notation:
\[
{\mathcal R} = \Sigma_1^Y({\mathcal I}_{C/Y}(2)) \simeq \operatorname{Im}(M), \qquad \mbox{with}
\qquad M : 3{\mathcal O}_Y(-1) \to 3{\mathcal O}_Y.
\]
The following lemma is essentially \cite[Proposition 2.5]{LLMS18}, we
reproduce it here for self-containedness. In fact, given a Cohen-Macaulay
twisted cubic $C \subset X$, the sheaf ${\mathcal G}={\mathcal G}_C$
represents uniquely a point of the Lehn-Lehn-Sorger-van Straten
eightfold ${\mathcal Z}$ associated with the cubic fourfold $X$.
\begin{lem} \label{hat S is stable}
Assume that $Y$ is integral. Then the sheaf ${\mathcal G}$ is stable with:
\[
{\mathrm p}_{{\mathcal G}}(t)={\mathrm u}(t-1)=\frac 18 (t+3)(t+2)(t+1)t, \qquad {\mathrm H}^*(X,{\mathcal G}(-t))=0, \qquad \mbox{for $t=0,1,2$}.
\]
Finally, we have $\mathcal{E}xt_X^i({\mathcal G},{\mathcal O}_X)=0$ except for $i=0,1$,
in which case:
\[{\mathcal G}^\vee \simeq 3{\mathcal O}_X,
\qquad \mathcal{E}xt_X^1({\mathcal G},{\mathcal O}_X) \simeq \mathcal{H}om_Y({\mathcal I}_{C/Y},{\mathcal O}_Y) = {\mathcal O}_Y(C).
\]
\end{lem}
\begin{proof}
The Hilbert polynomial of ${\mathcal G}$ is computed directly from the
previous proposition. Next, we use the sheaf ${\mathcal R}$ which satisfies
${\mathcal R} \simeq \Sigma_2^Y({\mathcal O}_{C}(2))$.
Recall from the proof of the previous proposition the sequence
\eqref{hatS} that we rewrite as:
\begin{equation}
\label{IR}
0 \to 3 {\mathcal I}_{Y/X} \to {\mathcal G} \to {\mathcal R} \to 0.
\end{equation}
By definition of ${\mathcal G}
= \Sigma_1^X({\mathcal I}_{C/Y}(2))$, the map $3{\mathcal O}_X \to {\mathcal I}_{C/Y}(2)$ in \eqref{SI}
induces an isomorphism on global sections, hence
${\mathrm H}^*(X,{\mathcal G})=0$. The vanishing ${\mathrm H}^*(X,{\mathcal G}(-1))={\mathrm H}^*(X,{\mathcal G}(-2))=0$ was proved in the previous proposition.
Next, we show first that ${\mathcal G}$ is simple.
Applying $\operatorname{Hom}_X(-,{\mathcal G})$ to \eqref{SI}, we get:
\[
\operatorname{End}_X({\mathcal G}) \simeq \operatorname{Ext}_X^1({\mathcal I}_{C/Y}(2),{\mathcal G}).
\]
We note that ${\mathcal I}_{C/Y}$ is simple,
$\operatorname{Hom}_X({\mathcal I}_{C/Y}(2),{\mathcal O}_X)=0$ as ${\mathcal I}_{C/Y}$ is torsion
and:
\[
\operatorname{Ext}_X^1({\mathcal I}_{C/Y}(2),{\mathcal O}_X) \simeq {\mathrm H}^3({\mathcal I}_{C/Y}(-1))^\vee=0
\]
since $\dim(Y)=2$. Hence applying $\operatorname{Hom}_X({\mathcal I}_{C/Y}(2),-)$ to
\eqref{SI}, we observe that ${\mathcal G}$ is simple:
\[
\operatorname{End}_X({\mathcal G}) \simeq \operatorname{Ext}_X^1({\mathcal I}_{C/Y}(2),{\mathcal G}) \simeq \operatorname{End}_X({\mathcal I}_{C/Y}) \simeq {\mathds k}.
\]
Suppose that ${\mathcal G}$ is not stable. Consider a saturated destabilizing subsheaf ${\mathcal K}$ of ${\mathcal G}$ so $\operatorname{rk}({\mathcal K})
\in \{1,2\}$ and ${\mathrm p}_{\mathcal K} \ge {\mathrm p}_{{\mathcal G}}$ so that ${\mathcal Q} = {\mathcal G} / {\mathcal K}$ is torsion-free with $\operatorname{rk}({\mathcal Q})=3-\operatorname{rk}({\mathcal K})$. Since
${\mathcal K} \subset {\mathcal G} \subset 3{\mathcal O}_X$, we have $\mu({\mathcal K}) \le
0$. From ${\mathrm p}_{\mathcal K} \ge {\mathrm p}_{{\mathcal G}}$ we deduce that
$c_1({\mathcal K})=c_1({\mathcal Q})=0$.
We look at the two possibilities for $\operatorname{rk}({\mathcal K})$.
If $\operatorname{rk}({\mathcal K})=1$, then ${\mathcal K}$ is torsion-free with $c_1({\mathcal K})=0$
so there is a closed subscheme $Z \subset X$ of codimension at
least $2$ such that ${\mathcal K} \simeq
{\mathcal I}_{Z/X}$. If $Z=\emptyset$ then ${\mathcal K} \simeq {\mathcal O}_X$, which is
impossible as ${\mathrm H}^0(X,{\mathcal G})=0$. Now for $Z \ne \emptyset$
consider the inclusion ${\mathcal I}_{Z/X} \subset {\mathcal G}
\subset 3{\mathcal O}_X$. Taking reflexive hulls, we see that this factors
through a single copy of ${\mathcal O}_X$ in $3{\mathcal O}_X$. Looking at
\eqref{SI}, we get that the quotient
${\mathcal O}_Z={\mathcal O}_X/{\mathcal I}_{Z/X}$ inherits a non-zero map to
${\mathcal I}_{C/Y}(2)$. The image of this map is ${\mathcal O}_Y$ itself
because ${\mathcal I}_{C/Y}(2)$ is torsion-free of rank $1$ over $Y$ as $Y$ is integral.
Note that ${\mathrm p}_{{\mathcal I}_{Z/X}} = {\mathrm p}_{{\mathcal G}}$ precisely when
$Z$ is a linear subspace ${\mathds {P}}^2$ contained in $X$, and that
${\mathrm p}_{{\mathcal I}_{Z/X}} < {\mathrm p}_{{\mathcal G}}$ if $\deg(Z) \ge 2$ and
$\dim(Z)=2$.
Hence, the image of ${\mathcal O}_Z \to {\mathcal I}_{C/Y}(2)$ cannot be the whole ${\mathcal O}_Y$ as
then $Y \subseteq Z$, so we have $\dim(Z) = 2$ and $\deg(Z) \ge 3$, while we
are assuming ${\mathrm p}_{{\mathcal I}_{Z/X}} \ge {\mathrm p}_{{\mathcal G}}$.
Therefore, the possibility $\operatorname{rk}({\mathcal K})=1$ is ruled out.
\smallskip
We may now assume $\operatorname{rk}({\mathcal K})=2$. Arguing as in the previous case,
we deduce that there is a closed
subscheme $Z \subset X$ of codimension at least $2$ such that ${\mathcal Q} \simeq {\mathcal I}_{Z/X}$.
Using \eqref{IR} and noting that $3{\mathcal I}_{Y/X}$ cannot be contained in ${\mathcal K}$
for $\operatorname{rk}({\mathcal K})=2$, we get a non-zero map $3{\mathcal I}_{Y/X} \to
{\mathcal I}_{Z/X}$ by composing $3 {\mathcal I}_{Y/X} \hookrightarrow {\mathcal G}$ with ${\mathcal G}
\twoheadrightarrow {\mathcal I}_{Z/X}$. The image of this map is of the form ${\mathcal I}_{Z'/X}
\subset {\mathcal I}_{Z/X}$ for some closed subscheme $Z' \supseteq Z$ of $X$. Since $3{\mathcal I}_{Y/X}$ is polystable and
$3{\mathcal I}_{Y/X} \twoheadrightarrow {\mathcal I}_{Z'/X}$, we have ${\mathcal I}_{Z'/X} \simeq {\mathcal I}_{Y/X}$
so $Z'=Y$. In particular, we have $Z \subseteq Y$.
Again, we use that ${\mathrm p}_{{\mathcal I}_{Z/X}} > {\mathrm p}_{{\mathcal G}}$ as soon as
$\dim(Z) \le 1$, so the
assumption that ${\mathcal K}$ destabilizes ${\mathcal G}$ forces
$\dim(Z) \ge 2$. Hence, $Z$ is a surface contained in $Y$ so that $Z=Y$
since $Y$ is integral. Then ${\mathcal I}_{Y/X}$ is a direct summand of ${\mathcal G}$ which therefore splits as ${\mathcal G} = {\mathcal K} \oplus {\mathcal I}_{Y/X}$. But
this contradicts the fact that ${\mathcal G}$ is simple.
We conclude that ${\mathcal G}$ must be stable.
Finally, we apply $\mathcal{H}om_X(-,\omega_X)$ to \eqref{SI} and use
Grothendieck duality to compute $\mathcal{E}xt^i_X({\mathcal G},{\mathcal O}_X)$ using
that ${\mathcal I}_{C/Y}$ is reflexive on $Y$ to get:
\[\mathcal{E}xt^1_X({\mathcal G},\omega_X) \simeq \mathcal{E}xt^2_X({\mathcal I}_{C/Y} (2),\omega_X) \simeq
\mathcal{H}om_Y({\mathcal I}_{C/Y} (2),\omega_Y).
\]
Since $\omega_X \simeq {\mathcal O}_X(-3)$ and $\omega_Y \simeq {\mathcal O}_Y(-1)$,
the conclusion follows.
\end{proof}
The next lemma analyzes the restriction of ${\mathcal S}$ onto $Y$.
\begin{lem} \label{SonY}
There is a surjection $\xi : {\mathcal S}|_Y \to {\mathcal R}$ whose kernel fits into:
\begin{equation}
\label{kerxi}
0 \to {\mathcal R}(1) \to \ker(\xi) \to 2 {\mathcal I}_{C/Y}(1) \to 0.
\end{equation}
\end{lem}
\begin{proof}
First of all, restricting the Koszul resolution \eqref{koszulY} to
$Y$ we find:
\[
\mathcal{T}or_1^X({\mathcal I}_{C/Y},{\mathcal O}_Y) \simeq 2 {\mathcal I}_{C/Y}(-1), \qquad
\mathcal{T}or_2^X({\mathcal I}_{C/Y},{\mathcal O}_Y) \simeq {\mathcal I}_{C/Y}(-2).
\]
Therefore, restricting \eqref{SI} to $Y$ we get:
\[
0 \to 2 {\mathcal I}_{C/Y}(1) \to {\mathcal G}|_Y \to 3{\mathcal O}_Y \to {\mathcal I}_{C/Y}(2)
\to 0,
\]
and hence:
\begin{equation}
\label{hatS|Y}
0 \to 2 {\mathcal I}_{C/Y}(1) \to {\mathcal G}|_Y \to {\mathcal R} \to 0.
\end{equation}
We also get:
\[
\mathcal{T}or_1^X({\mathcal G},{\mathcal O}_Y) \simeq \mathcal{T}or_2^X({\mathcal I}_{C/Y}(2),{\mathcal O}_Y) \simeq {\mathcal I}_{C/Y}.
\]
Next, we restrict \eqref{hatSO} to $Y$ to obtain:
\[
0 \to {\mathcal I}_{C/Y} \to 3{\mathcal O}_Y \to {\mathcal S}|_Y \to {\mathcal G}|_Y \to 0.
\]
Looking at \eqref{periodic}, we see that the image of the middle
map is ${\mathcal R}(1)$, so we obtain:
\begin{equation}
\label{sS|Y}
0 \to {\mathcal R}(1) \to {\mathcal S}|_Y \to {\mathcal G}|_Y \to 0.
\end{equation}
Composing ${\mathcal S}|_Y \to {\mathcal G}|_Y$ with the surjection appearing in
\eqref{hatS|Y} we get the surjection $\xi$. Using \eqref{hatS|Y}
and \eqref{sS|Y} we get the desired filtration for $\ker(\xi)$.
\end{proof}
\subsection{Elementary modification along a cubic surface}
\label{second conic}
In \S \ref{Subsection:construction of S} we constructed an ACM bundle
${\mathcal S}$ of rank $6$. Recall that ${\mathrm h}^0(X,{\mathcal S})=3$, and these three global
sections of ${\mathcal S}$ make it unstable. Hence, it is natural to consider
an elementary modification of ${\mathcal S}$ by a sheaf ${\mathcal A}$ such that
$H^0({\mathcal S}) \stackrel{\sim} \to H^0({\mathcal A})$. Moreover, Proposition
\ref{Proposition:KS20} suggests a good candidate for ${\mathcal A}$ to get
closer to an Ulrich bundle on $X$. Indeed, we should have:
\[
\chi_{\mathcal A}(t)=6{\mathrm p}_{\mathcal S}(t)-6{\mathrm u}(t-1)= \frac 3 2(t+2)(t+1).
\]
A natural choice for ${\mathcal A}$ would thus be an Ulrich line bundle on $Y$.
In terms of Chern classes (as a coherent sheaf on $X$), we should have:
\[
c_1({\mathcal A})=0, \qquad c_2({\mathcal A})=-H_X^2.
\]
Since an Ulrich line bundle on a cubic surface comes from a twisted cubic,
we need to choose another twisted cubic $D$ in $Y$,
construct a surjection ${\mathcal S} \to {\mathcal O}_Y(D)$ so that the induced map on $H^0$ is an isomorphism, and take the kernel to perform an elementary modification.
To do this, from now on in this section, we assume
that $Y$ is the blow-up of ${\mathds {P}}^2$ at the six points $p_1,\ldots,p_6$
in general position and that the blow-down map $\pi : Y \to {\mathds {P}}^2$ is
associated with the linear system $|{\mathcal O}_Y(C)|$. Write $L$ for the
class of a line
in ${\mathds {P}}^2$ and denote by $E_1,\ldots,E_6$ the exceptional
divisors of $\pi$, so that $C=\pi^*L$ and $H_Y=3C-E_1-\cdots - E_6$.
Note that ${\mathcal R}(1) \simeq \pi^*(\Omega_{{\mathds {P}}^2}(2))$.
\begin{lem}
Let $Z = \{p_1,p_2,p_3\}$. Then we have:
\[
0 \to {\mathcal O}_{{\mathds {P}}^2}(-2) \to \Omega_{{\mathds {P}}^2}(1) \to {\mathcal I}_{Z/{\mathds {P}}^2}(1) \to 0.
\]
\end{lem}
\begin{proof}
By assumption $Z$ is contained in no line, hence
by the Cayley-Bacharach property (see for instance \cite[Theorem
5.1.1]{HL10}) there is a vector bundle ${\mathcal F}$ of rank $2$
fitting into:
\[
0 \to {\mathcal O}_{{\mathds {P}}^2}(-2) \to {\mathcal F} \to {\mathcal I}_{Z/{\mathds {P}}^2}(1)
\to 0.
\]
Note that $c_1({\mathcal F})=-L$ and $c_2({\mathcal F})=L^2$. By the above sequence
${\mathrm H}^0({\mathcal F})=0$ so ${\mathcal F}$ is stable. But the only stable bundle on
${\mathds {P}}^2$ with $c_1({\mathcal F})=-L$ and $c_2({\mathcal F})=L^2$ is $\Omega_{{\mathds {P}}^2}(1)$.
\end{proof}
Set $D=2 C-E_1-E_2-E_3$. This is a class of a twisted
cubic in $Y$ with:
\[
D \cdot C=2.
\]
\begin{lem}
There is a surjection $\eta : {\mathcal R}(1) \to {\mathcal O}_Y(D)$ such that the
induced map on global sections
${\mathrm H}^0({\mathcal R}(1)) \to {\mathrm H}^0({\mathcal O}_Y(D))$ is an isomorphism.
\end{lem}
\begin{proof}
Recall the exact sequence:
\[
0 \to {\mathcal O}_Y(-C) \to 3{\mathcal O}_Y \to {\mathcal R}(1) \to 0,
\]
so that ${\mathcal R}(1) \simeq \pi^*(\Omega_{{\mathds {P}}^2}(2))$. There is an exact
sequence:
\begin{equation}
\label{pi*IZ}
0 \to \bigoplus_{i=1}^3{\mathcal O}_{E_i}(-1) \to \pi^* ({\mathcal I}_{Z/{\mathds {P}}^2}(2))
\to {\mathcal O}_Y(D) \to 0.
\end{equation}
By the previous lemma, we have
\begin{equation}
\label{Omega2}
0 \to {\mathcal O}_{{\mathds {P}}^2}(-1) \to \Omega_{{\mathds {P}}^2}(2) \to {\mathcal I}_{Z/{\mathds {P}}^2}(2)
\to 0
\end{equation}
and thus via $\pi^*$ an exact sequence:
\[
0 \to {\mathcal O}_Y(-C) \to {\mathcal R}(1) \to \pi^*({\mathcal I}_{Z/{\mathds {P}}^2}(2)) \to 0.
\]
Composing ${\mathcal R}(1) \to \pi^*({\mathcal I}_{Z/{\mathds {P}}^2}(2))$ with the surjection
appearing in \eqref{pi*IZ}, we get the following:
\[
0 \to {\mathcal O}_Y(-C+E_1+E_2+E_3) \to {\mathcal R}(1) \to {\mathcal O}_Y(D) \to 0.
\]
The map on global sections ${\mathrm H}^0({\mathcal R}(1)) \to {\mathrm H}^0({\mathcal O}_Y(D))$ is
induced by the map ${\mathrm H}^0(\Omega_{{\mathds {P}}^2}(2)) \to
{\mathrm H}^0({\mathcal I}_{Z/{\mathds {P}}^2}(2))$ arising from \eqref{Omega2} and as such it
is an isomorphism since ${\mathrm H}^*({\mathcal O}_{{\mathds {P}}^2}(-1))=0$.
\end{proof}
Given the class of a twisted cubic $C$ in $Y$, we observe that
$C^{\mathrm t} = 2H_Y-C$ is also the class of a twisted cubic. We denote:
\[
C^{\mathrm t} = 2H_Y-C.
\]
This notation is justified by the fact that ${\mathcal I}_{C^{\mathrm t}/Y}$ is
presented by the transpose matrix $M^{\mathrm t}$ of $M$. We have:
\[
C^{\mathrm t} \cdot D = C \cdot D^{\mathrm t} = 4.
\]
\begin{lem}
There is a surjection $\zeta : {\mathcal S} \to {\mathcal O}_Y(D)$ inducing an
isomorphism:
\[
{\mathrm H}^0(X,{\mathcal S}) \to {\mathrm H}^0({\mathcal O}_Y(D)).
\]
\end{lem}
\begin{proof}
According to the previous lemma, we have $\eta : {\mathcal R}(1) \to {\mathcal O}_Y(D)$
inducing an isomorphism on global sections.
We would like
to use Lemma \ref{SonY} to lift $\eta$ to a surjection ${\mathcal S}|_Y \to {\mathcal O}_Y(D)$ and compose this
lift with the restriction ${\mathcal S} \to
{\mathcal S}|_Y$ preserving the isomorphism on global sections.
So in the notation of Lemma \ref{SonY} we first lift $\eta$ to
$\ker(\xi)$. To do this, we apply $\operatorname{Hom}_Y(-,{\mathcal O}_Y(D))$ to
\eqref{kerxi} and get:
\[
\cdots \to \operatorname{Hom}_Y(\ker(\xi),{\mathcal O}_Y(D)) \to \operatorname{Hom}_Y({\mathcal R}(1),{\mathcal O}_Y(D)) \to
2{\mathrm H}^1({\mathcal O}_Y(C+D-H_Y)) \to \cdots
\]
Now, $C+D-H_Y=E_4+E_5+E_6$ so ${\mathrm H}^1({\mathcal O}_Y(C+D-H_Y))=0$. Therefore $\eta$
lifts to $\hat \eta : \ker(\xi) \to {\mathcal O}_Y(D)$. Note that by
\eqref{kerxi} the map ${\mathcal R}(1) \to \ker(\xi)$ induces an isomorphism on
global sections, so $\hat \eta$ gives an isomorphism
${\mathrm H}^0(\ker(\xi)) \simeq {\mathrm H}^0({\mathcal O}_Y(D))$.
Next, write:
\[
0 \to \ker(\xi) \to {\mathcal S}|_Y \to {\mathcal R} \to 0,
\]
and apply $\operatorname{Hom}_Y(-,{\mathcal O}_Y(D))$. We get an exact sequence:
\[
\cdots \to \operatorname{Hom}_Y({\mathcal S}|_Y,{\mathcal O}_Y(D)) \to \operatorname{Hom}_Y(\ker(\xi),{\mathcal O}_Y(D))
\to \operatorname{Ext}^1_Y({\mathcal R},{\mathcal O}_Y(D)) \to \cdots
\]
So $\hat \eta$ lifts to ${\mathcal S}|_Y \to {\mathcal O}_Y(D)$ if we prove
$\operatorname{Ext}^1_Y({\mathcal R},{\mathcal O}_Y(D))=0$. To do it, write again the defining sequence
of ${\mathcal R}$ as:
\[
0 \to {\mathcal R} \to 3{\mathcal O}_Y \to {\mathcal O}_Y(C^{\mathrm t}) \to 0.
\]
Applying $\operatorname{Hom}_Y(-,{\mathcal O}_Y(D))$ to this sequence we get:
\[
\cdots \to 3{\mathrm H}^1({\mathcal O}_Y(D)) \to \operatorname{Ext}^1_Y({\mathcal R},{\mathcal O}_Y(D)) \to
{\mathrm H}^2({\mathcal O}_Y(C+D-2H_Y)) \cdots
\]
Now, ${\mathcal O}_Y(D)$ is Ulrich so ${\mathrm H}^1({\mathcal O}_Y(D))=0$, and $H_Y-C-D=-E_4-E_5-E_6$:
\[
{\mathrm h}^2({\mathcal O}_Y(C+D-2H_Y))={\mathrm h}^0({\mathcal O}_Y(H_Y-C-D))=0.
\]
This provides a lift $\tilde \eta : {\mathcal S}|_Y \to {\mathcal O}_Y(D)$ of $\hat \eta$ and again
$\ker(\xi) \hookrightarrow {\mathcal S}|_Y$ induces an isomorphism on global sections,
hence so does $\tilde \eta$.
Finally we define $\zeta : {\mathcal S} \to {\mathcal O}_Y(D)$ as composition of the
restriction ${\mathcal S} \to {\mathcal S}|_Y$ and $\tilde \eta$. Since
$H^*({\mathcal S}(-1))=H^*({\mathcal S}(-2))=0$, tensoring the Koszul resolution
\eqref{koszulY} by ${\mathcal S}$ we see that ${\mathcal S} \to {\mathcal S}|_Y$ induces an
isomorphism on global sections. Therefore, so does $\zeta$ and the
lemma is proved.
\end{proof}
Consider $D^{\mathrm t} = 2H_Y-D$ and ${\mathcal G}_{D^{\mathrm t}} = \ker(3{\mathcal O}_X \to {\mathcal O}_Y(D))$.
Let ${\mathcal E} = \ker(\zeta)$, so we have:
\begin{equation}
\label{ESD}
0 \to {\mathcal E} \to {\mathcal S} \to {\mathcal O}_Y(D) \to 0.
\end{equation}
\begin{lem} The sheaf ${\mathcal E}$ is simple and
has a Jordan-H\"older filtration :
\begin{equation}
\label{sE}
0 \to {\mathcal G}_{D^{\mathrm t}} \to {\mathcal E} \to {\mathcal G}_C \to 0.
\end{equation}
Also, we have:
\[
{\mathcal E}^\vee \simeq {\mathcal S}^\vee, \qquad {\mathrm p}_{{\mathcal E}}(t) = {\mathrm u}(t-1), \qquad {\mathrm H}^*({\mathcal E}(-t))=0,
\qquad \mbox{for $t=0,1,2$}.
\]
\end{lem}
\begin{proof}
The sheaves ${\mathcal G}_{D^{\mathrm t}}$ and ${\mathcal G}_C$ are stable by
Lemma \ref{hat S is stable} and the reduced Hilbert polynomial of
both of them is ${\mathrm u}(t-1)$. Hence ${\mathcal E}$ is
semistable and has the same reduced Hilbert polynomial as soon
as it fits in \eqref{sE}. Also, ${\mathcal E}$ is simple if
this sequence is non-split.
Moreover, by Lemma \ref{hat S is stable}, we get ${\mathrm H}^*({\mathcal E}(-t))=0$,
for $t=0,1,2$ as well by \eqref{sE}.
Summing up, it suffices to prove that ${\mathcal E}$ fits in \eqref{sE}
and that this sequence is non-split. To do it, use the previous
lemma to show that the evaluation of global sections gives an exact
commutative diagram:
\[
\xymatrix@-2ex{
& 0 \ar[d] & 0\ar[d] & & \\
0 \ar[r] & {\mathcal G}_{D^{\mathrm t}} \ar[r] \ar[d] & {\mathcal E} \ar[r] \ar[d]
& {\mathcal G}_C \ar[r] \ar@{=}[d] & 0 \\
0 \ar[r] & 3{\mathcal O}_X \ar[r] \ar[d] & {\mathcal S} \ar[r] \ar[d] & {\mathcal G}_C \ar[r] & 0 \\
& {\mathcal O}_Y(D) \ar@{=}[r] \ar[d] & {\mathcal O}_Y(D) \ar[d] & & \\
& 0 & 0 & &
}
\]
We thus have \eqref{sE}. By contradiction, assume that it
splits as ${\mathcal E} \simeq {\mathcal G}_C \oplus {\mathcal G}_{D^{\mathrm t}}$.
Note that, since ${\mathcal S}$ is locally free, \eqref{ESD} gives:
\[
{\mathcal E}^\vee \simeq {\mathcal S}^\vee, \qquad \mathcal{E}xt^1_X({\mathcal E},{\mathcal O}_X) \simeq
\mathcal{E}xt_X^2 ({\mathcal O}_Y(D), {\mathcal O}_X) \simeq
{\mathcal O}_Y(D^{\mathrm t}).
\]
On the other hand, if ${\mathcal E} \simeq {\mathcal G}_C \oplus {\mathcal G}_{D^{\mathrm t}}$
then by Lemma \ref{hat S is stable} we would have ${\mathcal E}^\vee \simeq
6{\mathcal O}_X$ and $\mathcal{E}xt_X^1({\mathcal E},{\mathcal O}_X) \simeq {\mathcal O}_Y(C) \oplus {\mathcal O}_Y(D^{\mathrm t})$, which is not the case.
\end{proof}
\section{Smoothing the modified sheaves} \label{section:smoothing}
In the previous section we constructed a simple and semistable sheaf
${\mathcal E}$ with ${\mathrm p}_{{\mathcal E}}(t) = u(t-1)$. In particular, the sheaf ${\mathcal E}(1)$
has the same reduced Hilbert polynomial as an Ulrich bundle
${\mathcal U}$. However, ${\mathcal E}(1)$ itself cannot be Ulrich: for instance it is not locally free since $\mathcal{E}xt_X^1({\mathcal E},
{\mathcal O}_X) \simeq {\mathcal O}_Y(D^{\mathrm t})$.
The goal of this section is to show that ${\mathcal E}(1)$ admits a flat deformation to an Ulrich bundle.
\subsection{The Kuznetsov category}
The bounded derived category ${\mathrm D}(X)$ of coherent sheaves on $X$ has
the semiorthogonal decomposition:
\[
\langle \operatorname{Ku}(X),{\mathcal O}_X , {\mathcal O}_X(1), {\mathcal O}_X(2)\rangle,
\]
where $\operatorname{Ku}(X)$ is a K3 category. Indeed, $\operatorname{Ku}(X)$ equips with the K3-type Serre duality
\[
\operatorname{Ext}^i ({\mathcal F}, {\mathcal G}) \simeq \operatorname{Ext}^{2-i} ({\mathcal G}, {\mathcal F})^{\vee}
\]
for any ${\mathcal F}, {\mathcal G} \in \operatorname{Ku}(X)$ \cite{Kuz04}. We have:
\[
{\mathrm H}^*(X,{\mathcal E})={\mathrm H}^*(X,{\mathcal E}(-1))={\mathrm H}^*(X,{\mathcal E}(-2))=0,
\]
and therefore:
\[
{\mathcal E} \in \operatorname{Ku}(X).
\]
Lemma \ref{hat S is stable} says that for a Cohen-Macaulay twisted
cubic $C \subset X$ spanning an irreducible cubic surface we have
that ${\mathcal G}_C$ is stable and:
\[
{\mathcal G}_C \in \operatorname{Ku}(X).
\]
We also know that ${\mathcal G}_C$ represents a point of the
Lehn-Lehn-Sorger-van Straten irreducible symplectic eightfold ${\mathcal Z}$ and
that ${\mathcal Z}$ contains a Zariski-open dense subset ${\mathcal Z}^\circ$ whose
points are in bijection with the sheaves of the form ${\mathcal G}_C$ \cite{LLSS17,LLMS18}.
\begin{lem}
We have ${\mathrm h}^3({\mathcal E}(-3)) = {\mathrm h}^4({\mathcal E}(-3)) = 3$, $\operatorname{ext}^i_X ({\mathcal E},
{\mathcal E})=0$ for $i \ge 3$ and:
\[
\operatorname{ext}^1_X ({\mathcal E}, {\mathcal E})=26, \qquad
\operatorname{ext}^2_X ({\mathcal E}, {\mathcal E})=1.
\]
\end{lem}
\begin{proof}
First note that ${\mathrm h}^3 ({\mathcal E}(-3)) = {\mathrm h}^2({\mathcal O}_Y(D-3H_Y))=3$
since ${\mathcal O}_Y(D)$ is an Ulrich line bundle on a cubic surface
$Y$. We also have ${\mathrm h}^4({\mathcal E}(-3))= {\mathrm h}^4 ({\mathcal S}(-3))=3$
since $\chi({\mathcal S}(-3)) = 6 {\mathrm p}_{\mathcal S} (-3) = 3$ and ${\mathrm h}^i({\mathcal S}(-3))=0$ for $i<4$.
Recall that ${\mathcal E}$ is a simple sheaf and that ${\mathcal E}$ lies in
$\operatorname{Ku}(X)$. Since $\operatorname{Ku}(X)$ is a K3 category, we have
$\operatorname{ext}^2_X ({\mathcal E}, {\mathcal E})=\operatorname{Hom}_X({\mathcal E},{\mathcal E})=1$ and $\operatorname{ext}^i_X ({\mathcal E}, {\mathcal E})=0$
for $i \ge 3$.
The equality $\operatorname{ext}^1_X ({\mathcal E}, {\mathcal E})=26$ now follows from Riemann-Roch.
\end{proof}
\subsection{Deforming to Ulrich bundles}
We assume in this subsection that $X$ does not contain an integral surface of degree up to $3$ other than linear sections. In other words, $X$
does not contain a plane (equivalently, a quadric surface) nor a smooth cubic scroll, nor a cone over a rational normal cubic curve.
The content of our main result is that there is a smooth
connected quasi-projective variety $T^\circ$ of
dimension $26$ and a sheaf ${\mathcal F}$ on $T^\circ \times X$, flat over $T^\circ$,
together with a distinguished point $s_0 \in T^\circ$ such that ${\mathcal F}_{s_0}
\simeq {\mathcal E}$ and such that ${\mathcal F}_s(1)$ is an Ulrich bundle on $X$, for all
$s$ in $T^\circ \setminus \{s_0\}$ -- here we write
${\mathcal F}_s = {\mathcal F}|_{\{s\} \times X}$ for all $s \in T^\circ$.
Stated in short form this gives the next result.
\begin{thm} \label{theorem:no_plane}
If $Y$ is smooth, then the sheaf ${\mathcal E}(1)$ deforms to an Ulrich bundle on $X$.
\end{thm}
\begin{proof} We divide the proof into several steps.
\begin{step} \label{step:cohomology}
Compute negative cohomology of ${\mathcal E}$, i.e. ${\mathrm h}^k({\mathcal E}(-t))$ for
$t \gg 0$ and $k \in \{0,1,2,3\}$.
\end{step}
Let $C \subset Y \subset X$ be a twisted cubic with $Y$ smooth.
The sheaf ${\mathcal G}_C$ is stable and lies in $\operatorname{Ku}(X)$ by Lemma
\ref{hat S is stable}.
We note that ${\mathrm h}^1({\mathcal O}_Y(D+t H_Y))=0$ for $t \in {\mathds Z}$, while
${\mathrm h}^2({\mathcal O}_Y(D-t H_Y))$ for $t \le 1$ while:
\[
{\mathrm h}^2({\mathcal O}_Y(D-t H_Y))=\frac 32 (t-1)(t-2), \qquad \mbox{for $t
\ge 2$}.
\]
Also, ${\mathrm H}^0({\mathcal E})={\mathrm H}^1({\mathcal E})=0$ since the surjection \eqref{ESD} induces an
isomorphism on global sections. This also implies that, since
${\mathrm H}^0({\mathcal O}_Y(D)) \otimes
{\mathrm H}^0({\mathcal O}_X(t))$ generates ${\mathrm H}^0({\mathcal O}_Y(D+t H_Y))$ for all $t \ge 0$,
the map ${\mathrm H}^0({\mathcal S}(t)) \to {\mathrm H}^0({\mathcal O}_Y(D+t H_Y))$ induced by
\eqref{ESD} is surjective. Since
${\mathrm H}^1({\mathcal S}(t))=0$ for all $t \in {\mathds Z}$, we obtain ${\mathrm H}^1({\mathcal E}(t))=0$ for $t
\in {\mathds Z}$.
By \eqref{ESD} we have:
\[\left\{
\begin{array}{ll}
{\mathrm h}^0({\mathcal E}(-t))=0, & \mbox{$t \ge 0$}, \\
{\mathrm h}^1({\mathcal E}(t))=0, & \mbox{$t \in {\mathds Z}$},\\
{\mathrm h}^2({\mathcal E}(t))=0, & \mbox{$t \in {\mathds Z}$},\\
{\mathrm h}^3({\mathcal E}(-t))=\frac 32 (t-1)(t-2), & \mbox{$t \ge 2$}.
\end{array}\right.
\]
\begin{step} \label{step:deformation}
Argue that ${\mathcal E}$ is unobstructed.
\end{step}
This follows from the argument of \cite[\S 31]{BLMNPS19}, which
applies to the sheaf ${\mathcal E}$ as it is simple and lies in
$\operatorname{Ku}(X)$.
To sketch this, recall that the framework is based on a combination
of Mukai's unobstructedness
theorem \cite{mukai:symplectic} and Buchweitz-Flenner's approach
to the deformation theory of ${\mathcal E}$, see \cite{BF00,BF03}.
To achieve this step, we use the proof of \cite[Theorem
31.1]{BLMNPS19} which goes as follows. Let $\mathrm{At}({\mathcal E}) \in
\operatorname{Ext}^1_X({\mathcal E},{\mathcal E} \otimes \Omega_X)$ be the Atiyah class of ${\mathcal E}$.
\begin{itemize}
\item Via a standard use of the infinitesimal lifting
criterion, one reduces to show that ${\mathcal E}$ has a
formally smooth deformation space.
\item We show that the deformation space of ${\mathcal E}$ is formally
smooth by observing that ${\mathcal E}$ extends over any square-zero
thickening of $X$, conditionally to
the vanishing of the product of the Atiyah class $\mathrm{At}({\mathcal E})$ and the
Kodaira-Spencer class $\kappa$ of the thickening, see
\cite{huybrechts-thomas:deformation} -- note that this holds
in arbitrary characteristic.
\item We use \cite{KM09} in order to show $\kappa \cdot \mathrm{At}({\mathcal E})
= 0$. Indeed, in view of \cite[Theorem 4.3]{KM09}, this takes place if
the trace $\mathrm{Tr}(\kappa \cdot \mathrm{At}^2({\mathcal E}))$ vanishes as element of
${\mathrm H}^3(\Omega_X)$.
\item We use that $\mathrm{Tr}(\kappa \cdot \mathrm{At}({\mathcal E})^2)=2\kappa \cdot
\mathrm{ch}({\mathcal E})$, (cf. the proof of \cite[Theorem 31.1]{BLMNPS19}) and
note that this vanishes as the Chern character
$\mathrm{ch}({\mathcal E})$ remains algebraic under any deformation of $X$. It holds
as all components of $\mathrm{ch}({\mathcal E})$ are multiples of powers of the hyperplane
class, while $\kappa \cdot \mathrm{ch}({\mathcal E})$ is the obstruction to
algebraicity of $\mathrm{ch}({\mathcal E})$ along the thickening of $X$ -- again,
cf. the proof of \cite[Theorem 31.1]{BLMNPS19}.
Note that the assumption that that ${\mathds k}$ has characteristic other
than $2$ is needed to use the formula $\mathrm{Tr}(\kappa \cdot
\mathrm{At}({\mathcal E})^2)=2 \mathrm{Tr}(\kappa \cdot \exp(\mathrm{At}({\mathcal E})))=2\kappa \cdot
\mathrm{ch}({\mathcal E})$.
\end{itemize}
According to the above deformation argument, there is a smooth quasi-projective scheme $T$ representing an open piece of the
moduli space of simple sheaves over $X$ containing ${\mathcal E}$. In other
words,
there is a point $s_0 \in T$, together with a coherent sheaf
${\mathcal F}$ over $T \times X$, such that ${\mathcal F}_{s_0} \simeq {\mathcal E}$, and the
Zariski tangent space of $T$ at $s_0$ is identified with
$\operatorname{Ext}^1_X({\mathcal E},{\mathcal E})$.
Note that all the sheaves ${\mathcal F}_s$ are simple.
By the openness of semistability and torsion-freeness, there is a
connected open dense subset $T_0 \subset T$, with $s_0
\in T_0$, such that ${\mathcal F}_s$ is simple, semistable and torsion-free
for all $s \in T_0$.
\begin{step} \label{step:hull}
Compute the cohomology of the reflexive hull ${\mathcal F}_s^{\vee \vee}$ of
${\mathcal F}_s$ and of ${\mathcal F}_s^{\vee \vee}/{\mathcal F}_s$.
\end{step}
For $s \in T_0$, let us consider the reflexive hull
${\mathcal F}_s^{\vee \vee}$ and the torsion sheaf ${\mathcal Q}_s = {\mathcal F}_s^{\vee \vee}/{\mathcal F}_s$. Let us write the reflexive hull sequence:
\begin{equation} \label{Fs**}
0 \to {\mathcal F}_s \to {\mathcal F}_s^{\vee \vee} \to {\mathcal Q}_s \to 0.
\end{equation}
By the upper-semicontinuity of cohomology, there is a Zariski-open dense
subset $s_0 \in T_1$
of $T_0$ such that for all $s \in T_1$ we have:
\[
{\mathrm H}^*({\mathcal F}_s) = {\mathrm H}^*({\mathcal F}_s(-1)) = {\mathrm H}^*({\mathcal F}_s(-2)) = 0.
\]
In particular, for $s \in T_1$ and $t \ge 0$:
\[
{\mathrm h}^k({\mathcal F}_s(-t))=0, \qquad
\mbox{for $k \le 2$}.
\]
By Lemma \ref{reflexive}, since ${\mathcal F}_s$ is torsion-free there is a
polynomial $q_2 \in {\mathds Q}[t]$, with $\deg(q_2) \le 2$, such that
${\mathrm h}^3({\mathcal F}_s(-t))=q_2(t)$ for $t \gg 0$. By semicontinuity,
there is a Zariski-open dense subset $T_2$ of $T_1$, with $s_0 \in T_2$, such that for all
$s \in T_2$ we have $q_2(t) \le \frac 32 (t-1)(t-2)$.
Since $\operatorname{codim}({\mathcal Q}_s) \ge 2$, we get ${\mathrm H}^k({\mathcal Q}_s(t))=0$ for each $k
\ge 3$ and $t \in {\mathds Z}$. Using Lemma \ref{reflexive}, we get the vanishing
${\mathrm H}^0({\mathcal F}_s^{\vee \vee}(-t))={\mathrm H}^1({\mathcal F}_s^{\vee \vee}(-t))=0$ for $t
\gg 0$, and the existence of polynomials $q_0,q_1 \in {\mathds Q}[t]$ with $\deg(q_k) \le k$
such that ${\mathrm h}^{k+2}({\mathcal F}_s^{\vee \vee}(-t))=q_k(t)$ for $t \gg 0$.
By \eqref{Fs**}, for $t \gg 0$ and $k \ne 2$ we have:
\[\left\{
\begin{array}{l}
{\mathrm h}^2({\mathcal Q}_s(-t))=q_2(t)-q_1(t)+q_0, \\
{\mathrm h}^k({\mathcal Q}_s(-t))= 0.
\end{array}\right.
\]
Next, we use again the local-global spectral sequence
\[
\operatorname{Ext}_X^{p+q}({\mathcal Q}_s,{\mathcal O}_X(t-3)) \Leftarrow {\mathrm H}^p(\mathcal{E}xt_X^q({\mathcal Q}_s,{\mathcal O}_X(t-3))) = E_2^{p,q}.
\]
Via Serre vanishing for $t \gg 0$ and Serre duality this gives:
\begin{align}
\nonumber &{\mathrm h}^2({\mathcal Q}_s(-t)) = {\mathrm h}^0(\mathcal{E}xt_X^2({\mathcal Q}_s,{\mathcal O}_X(t-3))), \\
\label{extQ} &\mathcal{E}xt_X^k({\mathcal Q}_s,{\mathcal O}_X)=0, \quad \mbox{for $k \ne 2$}.
\end{align}
Assume ${\mathcal Q}_s \ne 0$. By the above discussion, ${\mathcal Q}_s$ is a non-zero reflexive sheaf supported
on a codimension 2 subvariety $Y_s$ of $X$, in which case
$h^2 ({\mathcal Q}_s (-t))$ must agree with a polynomial function of degree
$2$ of
$t$ for $t \gg 0$.
Hence the
sheaf $\hat {\mathcal Q}_s = \mathcal{E}xt_X^2({\mathcal Q}_s,{\mathcal O}_X(-3))$ supported on $Y_s$ satisfies:
\begin{equation}
\label{boundchi}
\chi(\hat {\mathcal Q}_s(t)) = h^2({\mathcal Q}_s(-t)) = q_2(t)-q_1(t)+q_0 \le \frac 32 (t-1)(t-2),
\quad \deg(\chi(\hat {\mathcal Q}_s(t)))=2.
\end{equation}
Note that \eqref{extQ} and \cite[Proposition 1.1.10]{HL10} imply
$\mathcal{E}xt_X^k({\mathcal F}_s^{\vee \vee},{\mathcal O}_X) = 0$ for $k \ge 3$ and therefore,
via \eqref{Fs**},
also $\mathcal{E}xt_X^k({\mathcal F}_s,{\mathcal O}_X)=0$ for $k \ge 3$.
We prove along the way that:
\begin{equation}
\label{local Ext2}
{\mathrm H}^1(\mathcal{E}xt_X^2({\mathcal F}_s(t),{\mathcal O}_X))= 0, \qquad \mbox{for $t \gg 0$.}
\end{equation}
Indeed, dualizing
\eqref{Fs**} and using \eqref{extQ} we get an epimorphism, for $t
\in {\mathds Z}$:
\[
\mathcal{E}xt_X^2({\mathcal F}_s^{\vee \vee}(t),{\mathcal O}_X) \twoheadrightarrow \mathcal{E}xt_X^2({\mathcal F}_s(t),{\mathcal O}_X).
\]
By \cite[Proposition 1.1.10]{HL10}, if the sheaf
$\mathcal{E}xt_X^2({\mathcal F}_s^{\vee \vee},{\mathcal O}_X)$ is non-zero then it is
supported on a zero-dimensional subscheme of $X$, hence the same
happens to $\mathcal{E}xt_X^2({\mathcal F}_s,{\mathcal O}_X(-t))$ by the above epimorphism.
Therefore ${\mathrm H}^1(\mathcal{E}xt_X^2({\mathcal F}_s,{\mathcal O}_X(-t)))=0$ for $t \gg 0$.
\begin{step} \label{step:extension}
Show that, if ${\mathcal F}_s$ is not reflexive, then it is an extension of
sheaves coming from ${\mathcal Z}^\circ$.
\end{step}
We have proved that, if ${\mathcal F}_s$ is not reflexive, the support $Y_s$
of $\hat {\mathcal Q}_s$ is a surface of degree at most
$3$. But then, since $X$ contains no integral surface of degree up to $3$
other than complete intersections, the reduced structure of each
primary component of $Y_s$ must be a
cubic surface contained in $X$, and hence $Y_s$ itself must be a cubic surface.
So the open subset $T_2
\subset T_1$ provides a family of cubic surfaces ${\mathcal Y} \to T_2$ whose
fibre over $T_2$ is the cubic surface $Y_s$, where $Y_{s_0}=Y$
is smooth.
Since smoothness is an open
condition, there is a Zariski-open dense subset $T_3$ of $T_2$, with
$s_0 \in T_3$, such
that $Y_s$ is smooth for all $s \in T_3$.
It follows again by \eqref{boundchi} that ${\mathcal Q}_s$ is reflexive of rank
$1$ over $Y_s$, i.e. ${\mathcal Q}_s$ is a line bundle on $Y_s$ since $Y_s$ is smooth.
Hence, we have a family of sheaves $\{{\mathcal Q}_s \mid s \in T_3 \}$ where
${\mathcal Q}_s$ is a line bundle over $Y_s$ and ${\mathcal Q}_{s_0}\simeq {\mathcal O}_Y(D)$. Since the Picard
group of $Y_s$ is discrete, this family must be locally constant. In
other words, for each $s \in T_3$ there is a divisor class $D_s$ on $Y_s$
corresponding to a twisted
cubic contained in $Y_s$ such that ${\mathcal Q}_s \simeq {\mathcal O}_{Y_s}(D_s)$ and $D_{s_0} \equiv D$.
Since ${\mathrm H}^0({\mathcal F}_s)={\mathrm H}^1({\mathcal F}_s)=0$, the evaluation map of global
sections $3{\mathcal O}_X \to {\mathcal O}_{Y_s}(D_s)$ lifts to a non-zero map
$\beta_s : 3 {\mathcal O}_X \to {\mathcal F}_s^{\vee \vee}$. The snake lemma yields an exact
sequence:
\[
0 \to \ker(\beta_s) \to {\mathcal G}_{D_s^{\mathrm t}} \to {\mathcal F}_s \to \operatorname{coker}(\beta_s) \to 0.
\]
Since the sheaves ${\mathcal F}_s$ and ${\mathcal G}_{D_s^{\mathrm t}}$ share the same reduced
Hilbert polynomial, with ${\mathcal F}_s$ semistable and ${\mathcal G}_{D_s^{\mathrm t}}$
stable, we must have $\ker(\beta_s)=0$. By semistability of ${\mathcal F}_s$,
we note that ${\mathcal D}_s=\operatorname{coker}(\beta_s)$ is
torsion-free, since otherwise the reduced Hilbert polynomial of the
torsion-free part of ${\mathcal D}_s$ would be strictly smaller than
${\mathrm p}_{{\mathcal D}_s} = {\mathrm p}_{{\mathcal F}_s}$.
Therefore, we have a flat family of sheaves over $T_3$ whose fibre
over $s$ is ${\mathcal D}_s$, with ${\mathcal D}_{s_0} \simeq {\mathcal G}_C$. Hence, for
all $s \in T_3$, the sheaf ${\mathcal D}_{s}$ corresponds to a point of the
open part ${\mathcal Z}^\circ$ of the
Lehn-Lehn-Sorger-van Straten eightfold, i.e. ${\mathcal D}_{s_0} \simeq {\mathcal G}_{C_s}$ for some twisted cubic $C_s \subset X$.
We take a further Zariski-open dense subset $T_4$ of $T_3$ such that
$C_s$ is Cohen-Macaulay and spans a smooth cubic surface, for all $s
\in T_4$.
Summing up, in a Zariski-open neighbourhood $T_4$ of $s_0$, dense in
$T$, the
hypothesis ${\mathcal Q}_s \ne 0$ for $s \in T_4$ implies the existence of
twisted cubics
$D_s$, $C_s$ in $X$ such that ${\mathcal F}_s$ fits into:
\[
0 \to {\mathcal G}_{D_s^{\mathrm t}} \to {\mathcal F}_s \to {\mathcal G}_{C_s} \to 0,
\]
where the twisted cubic $C_s$ is Cohen-Macaulay, so that ${\mathcal G}_{C_s}$
lies in $\operatorname{Ku}(X)$. Therefore the sheaves ${\mathcal G}_{C_s}$ and
${\mathcal G}_{D_s^{\mathrm t}}$ correspond uniquely to well-defined points of ${\mathcal Z}^\circ$.
\begin{step} \label{step:ulrich}
Conclude that ${\mathcal F}_s(1)$ is Ulrich for generic $s \in T$.
\end{step}
We compute the dimension of the family ${\mathcal W}$ of
sheaves ${\mathcal F}_s$ fitting into extensions as in the previous display.
Indeed, ${\mathcal W}$ is
equipped with a regular map ${\mathcal W} \to {\mathcal Z}^\circ \times {\mathcal Z}^\circ$
defined by ${\mathcal F}_s \mapsto ({\mathcal G}_{D_s^{\mathrm t}},{\mathcal G}_{C_s})$,
whose fibre is ${\mathds {P}}(\operatorname{Ext}^1_X({\mathcal G}_{C_s}, {\mathcal G}_{D_s^{\mathrm t}}))$. Since
$D^{\mathrm t}=D_{s_0}^{\mathrm t}$ and $C=C_{s_0}$ are contained in $Y$ and satisfy
$C \cdot D^{\mathrm t}=4$, $C \cdot C^{\mathrm t}=5$, we have $C \not \equiv D^{\mathrm t}$ so ${\mathcal G}_{D_{s_0}^{\mathrm t}} \not \simeq {\mathcal G}_{C_{s_0}}$.
Therefore ${\mathcal G}_{D_s^{\mathrm t}} \not \simeq {\mathcal G}_{C_s}$ for all
$s$ in a Zariski-open dense subset $T_5 \subset T_4$, with $s_0 \in T_5$.
Since ${\mathcal G}_{D_s^{\mathrm t}}$ and ${\mathcal G}_{C_s}$ lie in $\operatorname{Ku}(X)$ and
represent stable non-isomorphic sheaves, we have:
\[
\operatorname{Hom}_X({\mathcal G}_{C_s}, {\mathcal G}_{D_s^{\mathrm t}})=0, \qquad
\operatorname{ext}^2_X({\mathcal G}_{C_s}, {\mathcal G}_{D_s^{\mathrm t}}) \simeq
\operatorname{Hom}_X({\mathcal G}_{D_s^{\mathrm t}}, {\mathcal G}_{C_s})=0.
\]
Also, $\operatorname{ext}^k_X({\mathcal G}_{C_s}, {\mathcal G}_{D_s^{\mathrm t}})=0$ for $k \ge 3$ hence :
\[
\operatorname{ext}_X^1({\mathcal G}_{C_s}, {\mathcal G}_{D_s^{\mathrm t}})=-\chi({\mathcal G}_{C_s}, {\mathcal G}_{D_s^{\mathrm t}}) = 6.
\]
Therefore the fibre of ${\mathcal W} \to {\mathcal Z}^\circ \times {\mathcal Z}^\circ$ is $5$-dimensional and:
\[
\dim({\mathcal W})=2 \cdot \dim({\mathcal Z}^\circ) + \operatorname{ext}^1_X ({\mathcal G}_{C_s}, {\mathcal G}_{D_s^{\mathrm t}}) -1 = 21.
\]
So there is a Zariski-open dense
subset $T_6 \subset T_5$ with $s_0 \in T_6$, such that ${\mathcal Q}_s=0$ for
all $s \in T_6 \setminus \{s_0\}$. Hence ${\mathcal F}_s$ is reflexive for
all $s \in T_6 \setminus \{s_0\}$.
Then ${\mathcal F}_s^\vee$ is also semistable and shares
the same reduced Hilbert polynomial as ${\mathcal F}_s$, hence we have:
\[
{\mathrm h}^4({\mathcal F}_s(-3))=\operatorname{ext}^4_X({\mathcal O}_X(3),{\mathcal F}_s) = {\mathrm h}^0({\mathcal F}_s^\vee)=0.
\]
Since ${\mathrm h}^k({\mathcal F}_s(-3))=0$ for $k \le 2$, by Riemann-Roch we obtain
${\mathrm h}^3({\mathcal F}_s(-3))=0$, i.e. ${\mathrm H}^*({\mathcal F}_s(-3))=0$. We have now proved
that ${\mathcal F}_s(1)$ is Ulrich for $s \in T_6 \setminus \{s_0\}$.
Put $T^\circ=T_6$. We have proved that, for all $s \in T^\circ
\setminus \{s_0\}$, the sheaf ${\mathcal F}_s(1)$ is an Ulrich bundle of rank $6$.
\end{proof}
\subsection{Fourfolds containing planes or cubic scrolls}
We turn now our attention to the case of smooth cubic fourfolds $X$
containing a surface of degree up to three, other than linear sections.
The goal is to prove our main theorem from the introduction, in other
words, we would like to extend Theorem \ref{theorem:no_plane} to these fourfolds.
Note that Steps \ref{step:cohomology}, \ref{step:deformation} and \ref{step:hull} of the proof of Theorem
\ref{theorem:no_plane} are still valid.
Also, the argument of Step \ref{step:ulrich} holds once
Step \ref{step:extension} is established.
Summing up, it remains to work out Step
\ref{step:extension}. Recalling the base scheme $T_2$ introduced in Step
\ref{step:hull}, we are done as soon as we prove the
following result.
\begin{prop} \label{a linear section}
There is a Zariski-open neighborhood of $s_0$ in $T_2^\circ$ such that,
for all $s \in T_2^\circ$, the sheaf ${\mathcal Q}_s={\mathcal F}_s^{\vee \vee}/{\mathcal F}_s$
is either zero or it is supported on a linear section surface of $X$.
\end{prop}
\begin{proof}
We proved in Step \ref{step:hull} that, for $s \in
T_2$, the sheaf ${\mathcal Q}_s$ is zero or it is
a locally Cohen-Macaulay sheaf supported on a projective surface $Y_s\subset
X$ with $\deg(Y_s) \le 3$. Assuming ${\mathcal Q}_s \ne 0$, we would like to
show that $Y_s$ does not contain any surface $Z$ other than linear
sections of $X$.
Passing to the purely two-dimensional part of the reduced structure of
each primary component of $Z$, we may assume without loss of
generality that $Z$ is integral, still of degree at most $3$ and not
a linear section: we must then seek a contradiction.
The surface $Z$ is either a plane,
or a quadric surface, or a smooth cubic scroll, or a cone over a
rational normal cubic. The Hilbert polynomial of ${\mathcal O}_Z$ is thus either
$r_1=(t+1)(t+2)/2$, $r_2=(t+1)^2$, or $r_3=(t+1)(3t+2)/2$, and $Z$ is locally a complete intersection in any case.
\smallskip
We denote by ${\mathcal H}$ union of primary components of $\mathrm{Hilb}_r(X)$
containing integral subschemes $Z\subset X$ having Hilbert polynomial
$r$, with $r \in \{r_1, r_2,r_3\}$.
Note that $\mathrm{Hilb}_{r_1}(X)$ is a finite reduced
scheme consisting of planes contained in $X$. For $r=r_2$ or $r=r_3$, a priori a surface in
$\mathrm{Hilb}_r(X)$ might be badly singular. However, we have the
following claim.
\begin{claim} \label{itsP2}
Each surface of $\mathrm{Hilb}_{r_2}(X)$ is a reduced quadric. For $r=r_3$, all surfaces of ${\mathcal H}$ are purely 2-dimensional Cohen-Macaulay. For $r=r_2$ or $r=r_3$, each component of ${\mathcal H}$ is a projective plane.
\end{claim}
\begin{proof}
Take a surface $Z=Z_h$ in $\mathrm{Hilb}_{r_2}(X)$. If $Z$ is reduced,
then $Z$ is a quadric. Otherwise, the reduced structure of a
component of $Z$ must be a plane $L \subset X$. By computing
the Hilbert polynomial of ${\mathcal I}_{L/Z}$, we see that this sheaf must be
supported on a plane $L' \subset Z$ and have rank one over
$L'$.
Hence its ${\mathcal O}_{L'}$-torsion-free part is of the form
${\mathcal I}_{B/L'}(b)$ for a subscheme $B \subset L'$ and some $b \in {\mathds Z}$.
Note that ${\mathcal I}_{L/Z} \simeq {\mathcal I}_{L/X}/{\mathcal I}_{Z/X}$, so the surjection
$3{\mathcal O}_X(-1) \to {\mathcal I}_{L/X}$ induces an epimorphism $3{\mathcal O}_{L'}(-1) \to
{\mathcal I}_{B/L'}(b)$, whence $b \in \{-1,0\}$.
Computing Hilbert polynomials and arguing that the leading term of the Hilbert polynomial of the possible
$L'$-torsion part of ${\mathcal I}_{L/Z}$ must be non-negative, we see that
actually $b=-1$. This in turn implies $B=\emptyset$, i.e. ${\mathcal I}_{L/Z}
\simeq {\mathcal O}_{L'}(-1)$.
This says that $Z$ is a quadric surface. A direct computation shows
that $Z$ must be reduced, for a cubic fourfold containing a
non-reduced quadric surface is singular at least along a subscheme
of length 4.
All surfaces of a component of $\mathrm{Hilb}_{r_2}(X)$ are residual to the
same plane in $X$ so each component of $\mathrm{Hilb}_{r_2}(X)$ is the
projective plane of linear sections of $X$ containing a given plane.
\medskip
Assume now $r=r_3$ and let $Z=Z_h \subset X$ be an integral surface,
so that $Z$ is a smooth cubic scroll or a cone over a rational
normal cubic. We work roughly like in Proposition \ref{syzygy of
twisted}. The linear span $V$ of $Z$ is a ${\mathds {P}}^4$ that cuts $X$ along a cubic threefold $W$ and ${\mathcal I}_{Z/W}(2)$ is an Ulrich sheaf of rank $1$ over $W$ so we have a presentation:
\begin{equation}
\label{againM}
0 \to 3{\mathcal O}_{V}(-1) \xrightarrow{M} 3{\mathcal O}_{V} \to {\mathcal I}_{Z/W}(2) \to 0.
\end{equation}
Note that the threefold $W$ can have only finitely many singular
points as if $W$ had a $1$-dimensional family of singular points
then $X$ would singular along the intersection of this family
and a quadric in $V$.
The idea is to prove that, on one hand, denoting by ${\mathcal N}_{Z/X}$ the normal sheaf of $Z$ in $X$, we
have ${\mathrm h}^0({\mathcal N}_{Z/X})=2$. On the other hand, inspired by \cite[\S 4.1.2]{Has00}, we describe an explicit
projective plane parametrizing elements of ${\mathcal H}$ by proving that
each global section of ${\mathcal I}_{Z/W}(2)$ gives an element of ${\mathcal H}$ and
that all elements obtained this way are Cohen-Macaulay and indeed
contained in $W$.
Let us first accomplish the second task. By \eqref{againM}, the
projectivization $P={\mathds {P}}({\mathcal I}_{Z/W}(2))$ is embedded into $V \times
{\mathds {P}}^2 = {\mathds {P}}(3{\mathcal O}_V)$ and the subscheme $P$ is
cut in $V \times {\mathds {P}}^2$ by $3$ linear equations defined by the
columns of $M$. Write $\pi$ and $\sigma$ for the projections to
$V$ and ${\mathds {P}}^2$ from $V \times {\mathds {P}}^2$ and by ${\mathfrak {h}}$, ${\mathfrak {l}}$ the
pull-back to $V \times {\mathds {P}}^2$ of the hyperplane divisors of $V$ and ${\mathds {P}}^2$ via $\pi$ and $\sigma$. Use the same letters upon
restriction to $P$. From the Koszul resolution we obtain:
\[
0 \to {\mathcal O}_{{\mathds {P}}^2 \times V}(-2{\mathfrak {h}}-3{\mathfrak {l}}) \to 3{\mathcal O}_{{\mathds {P}}^2 \times
V}(-{\mathfrak {h}}-2{\mathfrak {l}}) \to 3{\mathcal O}_{{\mathds {P}}^2 \times V}(-{\mathfrak {l}}) \to
{\mathcal O}_{{\mathds {P}}^2 \times V}({\mathfrak {h}}) \to {\mathcal O}_P({\mathfrak {h}}) \to 0.
\]
Taking $\sigma_*$, we get that the sheaf
${\mathcal V}=\sigma_*({\mathcal O}_P({\mathfrak {h}}))$ fits into:
\[
0 \to 3{\mathcal O}_{{\mathds {P}}^2}(-1) \xrightarrow{N} 5{\mathcal O}_{{\mathds {P}}^2} \to {\mathcal V} \to 0.
\]
Observe that ${\mathrm H}^0({\mathcal O}_{{\mathds {P}}^2}(1))$ is naturally identified with
${\mathrm H}^0({\mathcal I}_{Z/W}(2))$. The choice of a line $\ell \subset
{\mathds {P}}^2$ corresponds uniquely to surjection $\ell :
{\mathrm H}^0({\mathcal O}_{{\mathds {P}}^2}(1)) \twoheadrightarrow 2{\mathds k}$ and thus
to an epimorphism $3{\mathcal O}_V \to
2{\mathcal O}_{V}$. Composing with $M$, the line $\ell$ gives
uniquely a matrix $M_\ell : 3{\mathcal O}_{V}(-1) \to 2{\mathcal O}_{V}$.
We have ${\mathds {P}}({\mathcal V}) \simeq P$. Note that the map $\pi : P \to W$ is birational since ${\mathcal I}_{Z/W}(2)$ has rank $1$ over $W$ and ${\mathcal O}_W$ has the same Hilbert polynomial as ${\mathcal O}_P({\mathfrak {h}})$. Therefore, $P$ is
irreducible and thus ${\mathcal V}$ is torsion-free.
In particular, for any line $\ell \subset {\mathds {P}}^2$, the restriction
$N|_\ell$ is injective and yields by restriction of $\pi$:
\[
\pi_\ell : {\mathds {P}}({\mathcal V}|_\ell) \to Z_\ell \subset W,
\]
where $Z_\ell=\operatorname{Im}(\pi_\ell)$ is a surface in $W$.
The scheme ${\mathds {P}}({\mathcal V}|_\ell)$ is equipped with two divisor
classes inherited from $P$, which we still denote by ${\mathfrak {l}}$ and ${\mathfrak {h}}$.
The surface $Z_\ell$ is the image of ${\mathds {P}}({\mathcal V}|_\ell)$ by
the linear system $|{\mathcal O}_{{\mathds {P}}({\mathcal V}|_\ell)}({\mathfrak {h}})|$.
Now ${\mathcal V}|_\ell \simeq \operatorname{coker}(N|_\ell)$ is
of the form ${\mathcal O}_\ell(a_1) \oplus {\mathcal O}_\ell(a_2) \oplus {\mathcal B}$, where $0
\le a_1 \le a_2 \le 3$, and ${\mathcal B}$ is a torsion sheaf on $\ell$ of
length $b$, with $a_1+a_2+b=3$.
According to \cite[Chapter 19]{burgisser-clausen-shokrollahi}, in a
suitable basis of ${\mathrm H}^0({\mathcal V})$ and ${\mathrm H}^1({\mathcal V}(-1))$ and choosing coordinates $(y_0:y_1)$ on
$\ell$, we may write a normal form of $N_\ell$. Removing the cases forbidden by the smoothness of $X$, the possibilities are:
\begin{itemize}
\item $(a_1,a_2,b)=(1,2,0)$: $Z_\ell$ is a smooth cubic scroll and:
\[
N_\ell^{\mathrm t} =
\begin{pmatrix}
y_0 & y_1 & 0 & 0 & 0 \\
0 & 0 & y_0 & y_1 & 0 \\
0 & 0 & 0 & y_0 & y_1
\end{pmatrix}.
\]
\item $(a_1,a_2,b)=(0,3,0)$: $Z_\ell$ is a cone over a rational normal
cubic curve and:
\[
N_\ell^{\mathrm t} =
\begin{pmatrix}
y_0 & y_1 & 0 & 0 & 0 \\
0 & y_0 & y_1 & 0 & 0 \\
0 & 0 & y_0 & y_1 & 0
\end{pmatrix}.
\]
\item $(a_1,a_2,b)=(1,1,1)$: $Z_\ell$ is the union of a ${\mathds {P}}^2$ and a
smooth quadric meeting along a line.
\[
N_\ell^{\mathrm t} =
\begin{pmatrix}
y_0 & y_1 & 0 & 0 & 0 \\
0 & 0 & y_0 & y_1 & 0 \\
0 & 0 & 0 & 0 & y_0
\end{pmatrix}.
\]
\item $(a_1,a_2,b)=(0,2,1)$: $Z_\ell$ is the cone over the union of a
smooth conic and a line meeting at a single point, spanning a
${\mathds {P}}^3 \subset V$ and having apex at a point outside $V$.
\[
N_\ell^{\mathrm t} =
\begin{pmatrix}
y_0 & y_1 & 0 & 0 & 0 \\
0 & y_0 & y_1 & 0 & 0 \\
0 & 0 & 0 & y_0 & 0
\end{pmatrix}.
\]
\item $(a_1,a_2,b)=(0,1,2)$: $Z_\ell$ is the cone over the union of a
line and reducible conic, meeting at a length-two subscheme of the
line, spanning a
${\mathds {P}}^3 \subset V$. The apex of the cone is a point outside $V$.
\[
N_\ell^{\mathrm t} =
\begin{pmatrix}
y_0 & y_1 & 0 & 0 & 0 \\
0 & 0 & y_0 & 0 & 0 \\
0 & 0 & 0 & y_1 & 0
\end{pmatrix}.
\]
\item $(a_1,a_2,b)=(0,0,3)$: $Z_\ell$ is a cone over a non-colinear
subscheme of length $3$ in ${\mathds {P}}^2 \subset V$, having a skew ${\mathds {P}}^1
\subset V$ as apex.
\[
N_\ell^{\mathrm t} =
\begin{pmatrix}
y_0 & 0 & 0 & 0 & 0 \\
0 & y_1 & 0 & 0 & 0 \\
0 & 0 & y_0+y_1 & 0 & 0
\end{pmatrix}.
\]
\end{itemize}
In all these cases the resulting subscheme $Z_\ell$ lies in ${\mathcal H}$
and has projective dimension $2$ with a Hilbert-Burch resolution
given $M_\ell^{\mathrm t}$.
Then the dual plane parametrizing lines $\ell
\subset {\mathds {P}}^2$ describes an explicit projective plane of
arithmetically Cohen-Macaulay surfaces in ${\mathcal H}$.
\medskip
Finally we have to show that ${\mathrm h}^0({\mathcal N}_{Z/X})=2$. We have an exact
sequence:
\[
0 \to {\mathcal O}_X(-1) \to {\mathcal I}_{Z/X} \to {\mathcal I}_{Z/W} \to 0.
\]
Applying $\mathcal{H}om_X(-,{\mathcal O}_Z)$ we get:
\[
0 \to {\mathcal N}_{Z/W} \to {\mathcal N}_{Z/X} \to {\mathcal O}_Z(1) \xrightarrow{\delta} \mathcal{E}xt_X^1({\mathcal I}_{Z/W},{\mathcal O}_Z) \to \mathcal{E}xt_X^1({\mathcal I}_{Z/X},{\mathcal O}_Z) \to 0.
\]
Since the surfaces $Z$ under consideration are locally complete
intersection in $X$, we get that $\mathcal{E}xt_X^1({\mathcal I}_{Z/X},{\mathcal O}_Z) \simeq
\mathcal{E}xt_X^2({\mathcal O}_Z,{\mathcal O}_Z)$ is the determinant of the normal bundle
${\mathcal N}_{Z/X}$ and is thus identified with the line bundle ${\mathcal N}_{Z/W}(1)$.
On the other hand, using \eqref{againM} we see that the sheaf
$\mathcal{E}xt_X^1({\mathcal I}_{Z/W},{\mathcal O}_Z)$ fails to be locally free of rank $1$ at
the subscheme $\Upsilon \subset W$ defined by the $2$-minors of $M$. Since
$\Upsilon$ is contained in (though sometimes not equal to) the singular locus of $W$, we have
$\dim(\Upsilon) = 0$ so the resolution of $\Upsilon$ is
obtained by the Gulliksen-Negard complex:
\[
0 \to {\mathcal O}_V(-6) \to 9{\mathcal O}_V(-4) \to 16{\mathcal O}_V(-3) \to 9{\mathcal O}_V(-2) \to {\mathcal I}_{\Upsilon/V} \to 0.
\]
Thus $\Upsilon$ has length $6$ and ${\mathrm H}^0({\mathcal I}_{\Upsilon/V}(1))=0$, which
in turn implies ${\mathrm H}^0({\mathcal I}_{\Upsilon/Z}(1))=0$.
Therefore, $\ker(\delta)\subset {\mathcal I}_{\Upsilon/Z}(1)$ gives
${\mathrm H}^0(\ker(\delta))=0$. In turn we get ${\mathrm H}^0({\mathcal N}_{Z/X}) \simeq
{\mathrm H}^0({\mathcal N}_{Z/W})$ so it only remains to show
${\mathrm h}^0({\mathcal N}_{Z/W})=2$. To get this, since $Z$ and $W$ are locally complete
intersection, we may use adjunction to the effect that ${\mathcal N}_{Z/W}
\simeq \mathcal{H}om_W( \omega_W,\omega_Z)/{\mathcal O}_W$. Therefore, using
$\omega_W \simeq {\mathcal O}_W(-2)$ and restricting
\eqref{againM} to $W$ we get:
\[
0 \to {\mathcal I}_{Z/W}(-1) \to 3{\mathcal O}_W(-1) \to 2{\mathcal O}_W \to {\mathcal N}_{Z/W} \to 0.
\]
Taking cohomology we obtain ${\mathrm h}^0({\mathcal N}_{Z/W})=2$ as desired.
\end{proof}
Write ${\mathcal Z} \subset X \times {\mathcal H}$ for the tautological surface. For each
point $h \in {\mathcal H}$, we denote by $Z_h={\mathcal Z} \cap X \times \{h\}$ the
corresponding surface. Consider ${\mathcal X} = X \times T_2 \times {\mathcal H}$ and write
$\pi_{1,2}$, $\pi_{1,3}$ and $\pi_{2,3}$ for the projections from
${\mathcal X}$ onto $X
\times T_2$, $X \times {\mathcal H}$ and $T_2 \times {\mathcal H}$.
We have the following claim.
\begin{claim} \label{component}
For any $(s,h) \in T_2 \times {\mathcal H}$, the surfaces
$Z_h$ and $Y_s$ share a component if and only if:
\[
{\mathrm H}^2(\mathcal{E}xt_X^1({\mathcal F}_s(t),{\mathcal O}_{Z_h})) \ne 0, \qquad \mbox{for $t \gg 0$}.
\]
\end{claim}
\begin{proof}
Take $(s,h) \in T_2 \times {\mathcal H}$ and set $Z=Z_h$.
Since ${\mathcal F}_s$ is torsion-free
and ${\mathcal F}_s^{\vee \vee}$ is reflexive, we have, for any coherent sheaf
${\mathcal B}$ on $X$:
\begin{equation}
\label{van1}
\mathcal{E}xt^q_X({\mathcal F}_s^{\vee \vee},{\mathcal B})= \mathcal{E}xt^{q+1}_X({\mathcal F}_s,{\mathcal B})=0
\quad \mbox{for $q \ge 3$},
\end{equation}
and, for $q \in \{1,2\}$:
\begin{equation}
\label{van2}
\dim(\mathcal{E}xt^{q}_X({\mathcal F}_s^{\vee \vee},{\mathcal B})) \le 2-q, \qquad
\dim(\mathcal{E}xt^q_X({\mathcal F}_s,{\mathcal B})) \le 3-q.
\end{equation}
Indeed, this follows from \cite[Proposition 1.1.10]{HL10} if ${\mathcal B}$ is
locally free. Then, \eqref{van1} and \eqref{van2} hold for an
arbitrary coherent sheaf ${\mathcal B}$ as we see by applying
$\mathcal{H}om_X({\mathcal F}_s,-)$ and $\mathcal{H}om_X({\mathcal F}_s^{\vee \vee},-)$ to a finite
locally resolution of ${\mathcal B}$ and using that \eqref{van1} and \eqref{van2}
hold for the terms of the resolution.
Applying $\mathcal{H}om_X(-,{\mathcal O}_{Z})$ to \eqref{Fs**} we get, for
$q \ge 1$:
\[
\cdots \to \mathcal{E}xt^q_X({\mathcal F}_s^{\vee \vee},{\mathcal O}_{Z}) \to
\mathcal{E}xt^q_X({\mathcal F}_s,{\mathcal O}_{Z}) \to
\mathcal{E}xt^{q+1}_X({\mathcal Q}_s,{\mathcal O}_{Z}) \to
\mathcal{E}xt^{q+1}_X({\mathcal F}_s^{\vee \vee},{\mathcal O}_{Z}) \to \cdots
\]
We deduce that $\mathcal{E}xt^q_X({\mathcal F}_s,{\mathcal O}_Z)=0$ for $q \ge 3$ and
\[
\dim(\mathcal{E}xt^1_X({\mathcal F}_s,{\mathcal O}_{Z})) = 2 ~ \Longleftrightarrow ~
\dim(\mathcal{E}xt^2_X({\mathcal Q}_s,{\mathcal O}_{Z})) = 2.
\]
Therefore
\begin{equation}
\label{3}
{\mathrm H}^p(\mathcal{E}xt^1_X({\mathcal F}_s(t),{\mathcal O}_{Z})) = 0 ~ \mbox{ for $p \ge 3$ and
all $t \in {\mathds Z}$},
\end{equation}
and
\begin{equation}
\label{dim2}
{\mathrm H}^2(\mathcal{E}xt^1_X({\mathcal F}_s(t),{\mathcal O}_{Z})) \ne 0 ~ \mbox{ for $t \gg 0$}
~ \Longleftrightarrow ~
\dim(\mathcal{E}xt^2_X({\mathcal Q}_s,{\mathcal O}_{Z})) = 2.
\end{equation}
By Claim \ref{itsP2}, we may assume that $Z$ is a locally
Cohen-Macaulay in $X$. Let $M$ be a matrix of size $p \times (p+1)$ whose $p$-minors
define $Z$ locally in $X$, then the sheaf $\mathcal{E}xt_X^2({\mathcal Q}_s,{\mathcal O}_{Z})$ is
locally presented as cokernel of the rightmost map in:
\begin{equation}
\label{localpresentation}
0 \to p\hat {\mathcal Q}_s \xrightarrow{\mathcal{H}om(\hat {\mathcal Q}_s,M)} (p+1) \hat {\mathcal Q}_s
\xrightarrow{\mathcal{H}om(\hat {\mathcal Q}_s,\wedge^p M)} \hat {\mathcal Q}_s
\end{equation}
Now the $p$-minors of $M$ vanish on an irreducible component of $Y_s$
if and only if such component also lies in $Z$, in which case
\eqref{localpresentation} shows that the support of $\mathcal{E}xt_X^2({\mathcal Q}_s,{\mathcal O}_{Z})$ is
the whole component.
Conversely, if $Y_s$ and $Z$ share no irreducible component so that
the $p$-minors do not vanish identically on any component of $Y_s$, then again by
\eqref{localpresentation} the sheaf $\mathcal{E}xt_X^2({\mathcal Q}_s,{\mathcal O}_{Z})$ is
supported along a closed subset of $Z$ having
dimension at most $1$.
This shows that $\dim(\mathcal{E}xt_X^2({\mathcal Q}_s,{\mathcal O}_{Z}))=2$ if and only if
${\mathcal Q}_s$ and $Z$ share a common component. Together with \eqref{dim2},
this finishes the proof.
\end{proof}
\begin{claim} \label{relative ext}
For $t \in {\mathds Z}$, put
${\mathcal B} = \mathcal{E}xt^1_{\mathcal X}(\pi_{12}^*({\mathcal F}(-t)),\pi_{13}^*({\mathcal O}_{\mathcal Z}))$
and ${\mathcal P}={\mathrm R}^2\pi_{23 *}({\mathcal B})$.
For $(s,h) \in T_2 \times {\mathcal H}$, we have ${\mathcal P}_{(s,h)} \ne 0$ for
$t \gg 0$ if and
only if $Z_h$ and $Y_s$ have a common component.
\end{claim}
\begin{proof}
Since ${\mathcal F}$ and ${\mathcal O}_{{\mathcal Z}}$ are flat over $T_2 \times {\mathcal H}$, we have an identification ${\mathcal B}_{(s,h)} \simeq \mathcal{E}xt^1_X({\mathcal F}_s(-t),{\mathcal O}_{Z_h})$
for all $t \in {\mathds Z}$ and $(s,h) \in T_2 \times {\mathcal H}$.
By the vanishing results of the previous
paragraph and using the flattening stratification for ${\mathcal B}$ over $T_2 \times {\mathcal H}$ and working over each stratum, we
get ${\mathrm R}^{p} \pi_{23 *} ({\mathcal B})=0$ for all $p \ge 3$ and $t \in
{\mathds Z}$ so via base-change we obtain, for all $(s,h) \in T_2 \times
{\mathcal H}$, we have
\[
{\mathcal P}_{(s,h)} \simeq {\mathrm R}^2\pi_{23 *}({\mathcal B})_{(s,h)} \simeq {\mathrm H}^2
\left(\mathcal{E}xt^1_X({\mathcal F}_s(t),{\mathcal O}_{Z_h})\right)
\]
for all $t \in {\mathds Z}$. The conclusion follows from Claim \ref{component}.
\end{proof}
We now finish the proof of the proposition. Indeed, by Claim
\ref{itsP2} for the
special point $s_0 \in T_2$, the surface $Y=Y_{s_0}$ shares no
component with any surface $Z_h$ for $h \in {\mathcal H}$. Indeed, if $Z_h$
contains $Y$ then comparing Hilbert polynomials we get that $Z_h$
must contain a line as its further (possibly embedded) component, which
is forbidden since $Z_h$ would not be Cohen-Macaulay.
Now, by Claim
\ref{relative ext} we have ${\mathcal P}_{(s_0,h)}=0$ for all $h \in {\mathcal H}$.
In other words, the support of ${\mathcal P}$ is disjoint from $\{s_0\} \times
{\mathcal H}$. Since ${\mathcal H}$ is projective, the image of the support of ${\mathcal P}$ in $T_2$
is thus a closed subset of $T_2$, disjoint from $s_0$.
Therefore there exists an open neighborhood $T_2^\circ$ of $s_0$
disjoint from this subset.
Thus the support of
${\mathcal P}$ does not intersect $T_2^\circ \times {\mathcal H}$. This implies that,
for all $s \in T_2^\circ$, if ${\mathcal Q}_s$ is not zero then its support is
a surface $Y_s$ having degree at most $3$
and containing no surface of ${\mathcal H}$ as a component, in other words
$Y_s$ must be a linear section of $X$. This completes the proof of
Proposition \ref{a linear section} and consequently of the main
theorem.
\end{proof}
|
\section{Introduction}
\subsection{Background and Motivation}
It has been a common concept throughout the history of mathematics to translate problems in one area of research to another area of research to solve them there.
An especially fruitful example of this translating approach has been the study of {\em toric varieties}. These are special varieties whose algebro-geometric properties are completely determined by the combinatorial properties of certain polyhedral objects. This phenomenon has been used by many mathematicians to study more general varieties via flat degenerations to toric varieties. Notable results in this regard have been archived by Gonciulea and Lakshmibai \cite{GL}, Kogan and Miller \cite{KoM}, Caldero \cite{C}, Alexeev and Brion \cite{AB} as well as Feigin, Fourier and Littelmann \cite{FFL3}.
However, there are also constructions that do not originate from representation theory. For example, Okounkov \cite{O,O2}, Lazarsfeld and Musta\cb{t}\u a \cite{LM} as well as Kaveh and Khovanskii \cite{KK} defined and analyzed convex bodies for arbitrary projective varieties\,---\,thereby developing the theory of {\em Newton-Okounkov bodies}.
It has been shown that most of the representation theoretic toric degenerations of flag varieties can be realized via Newton-Okounkov bodies\,---\,for example by Kaveh \cite{Ka}, Kiritchenko \cite{Ki} and Fujita and Naito \cite{FN}. A unified approach has been developed by Fang, Fourier and Littelmann who presented a construction of these polytopes from representation theory via {\em birational sequences} and connected them to Newton-Okounkov Theory in \cite{FaFL}.
Lately, Anderson \cite{A} showed that Newton-Okounkov bodies yield toric degenerations under reasonable technical assumptions, thereby providing a general reason for the existence of the diverse classes of toric degenerations in representation theory mentioned in the beginning. Even more, based on an observation by Alexeev and Brion \cite{AB} he realized that these toric degenerations respect the choice of a given {\em embedding} or {\em polarization} of the original variety. We will formalize this definition to develop the concept of embedded toric degenerations.
The goal of this paper is to study the following combinatorial property in the context of toric degenerations. A rational convex polytope is called {\em reflexive} if the polytope itself and its polar dual are lattice polytopes (i.e. they have integral vertices). There are two main reasons why these polytopes are important in our context.
Batyrev proved in \cite{B} that reflexive polytopes are in one-to-one correspondence with anticanonically polarized (normal) Gorenstein Fano toric varieties. So essentially we are working towards an answer to the following question.
\begin{question}\label{question:general}
Which polarized Gorenstein Fano varieties admit a flat projective degeneration to an anticanonically polarized Gorenstein Fano toric variety?
\end{question}
It should be noted that we will only be interested in {\em normal} Gorenstein Fano varieties, so we will leave out the additional adjective.
This poses an interesting problem in itself but there is also another viewpoint towards this framework. In the same work \cite{B}, Batyrev explicitly constructed mirrors to Calabi-Yau hypersurfaces in toric varieties via reflexive polytopes. It deems a reasonable hope that his construction can be generalized to more arbitrary varieties if one were able to associate meaningful reflexive polytopes to those varieties. Of course, (embedded) toric degenerations would prove such a meaningful connection.
\subsection{Results}
The main achievement of this paper is the following partial answer to \cref{question:general}.
\begin{theorem}\label{thm:main}
The limit of a polarized Gorenstein Fano variety under a toric degeneration is $\QQ$-polarized by its anticanonical divisor if and only if the polarization on the original variety is given by its anticanonical line bundle.
\end{theorem}
This result essentially says the following. If one were able to construct a toric degeneration of an anticanonically polarized Gorenstein Fano variety, one immediately gets that the toric limit is an anticanonically polarized $\QQ$-Gorenstein Fano variety.
This fact however is not new, it can be found for example in \cite[Theorem 3.8]{AB}. The new insight is that one can never reach anticanonically polarized limit varieties using any other polarization on the original variety! This greatly restricts the possible answers to \cref{question:general}.
However, in general our limit variety would only be $\QQ$-Gorenstein Fano. By a standard fact of toric geometry, it will be Gorenstein Fano if and only if the associated polytope is a lattice polytope. One can summarize this as follows.
\begin{theorem}\label{thm:reflexive}
Assume that the polarized Gorenstein Fano variety $(X, \L)$ admits a toric degeneration to the polarized toric variety $(X_\P, D_\P)$ associated to a rational convex polytope $\P$. Then the polytope $\P$ is reflexive if and only if it is a lattice polytope and the line bundle $\L$ is the anticanonical line bundle over $X$.
\end{theorem}
So the question that remains to be answered is the following.
\begin{question}\label{question:simplified}
Does every anticanonically polarized Gorenstein Fano variety admit a flat projective degeneration to a (necessarily anticanonically polarized) Gorenstein Fano toric variety (whose associated polytope would necessarily be a lattice polytope)?
\end{question}
This question remains quite difficult but at least we can answer it in the case of partial flag varieties using results from \cite{S3}.
\begin{theorem}\label{thm:flag}
Let $G$ be a complex classical group or $\GG$. Then for any parabolic $P\subseteq G$ the partial flag variety $(G/P, \omega_{G/P}^{-1})$ admits a flat projective degeneration to an anticanonically polarized Gorenstein Fano toric variety.
\end{theorem}
The proof of \cref{thm:main} is not straightforward and requires methods from different branches of mathematics like toric geometry, polyhedral geometry and algebraic geometry. Its schematics are depicted in \cref{fig:proof}.
Apart from the results by Batyrev and Nill, we will use methods from Ehrhart theory\,---\,especially a beautiful result by Hibi \cite{Hi}\,---\,to translate the desired properties of the polarized toric variety into combinatorial properties of the Ehrhart polynomial of its associated polytope. As a by product we will prove the following generalization of a result that is well known for lattice polytopes.
\begin{theorem}\label{thm:toricehrhart}
Let $\P\subseteq\RR^d$ be a full-dimensional rational convex polytope. Let $X_\P$ denote the associated normal projective toric variety and $D_\P$ the associated torus invariant $\QQ$-Weil divisor. Then
\[\#(n\P\cap\ZZ^d) = \chi(X_\P, \O_{X_\P}(\lfloor nD_\P\rfloor))\]
and
\[\#(\interior n\P \cap \ZZ^d) = \chi(X, \O_{X_\P}(\lceil nD_\P\rceil + K_{X_\P}))\]
for all $n \in \NN$.
\end{theorem}
Using Invariance of Euler Characteristic in flat projective families (see \cite[Theorem 24.7.1]{V}) we will be able to deduce the following statement.
\begin{theorem}\label{cor:twopolynomials}
Under the assumptions of \cref{thm:main}, the Hilbert polynomial associated to $X$ and $\L$ coincides with the Ehrhart quasi-polynomial associated to $\P$.
\end{theorem}
It has been known for quite some time that Serre Duality on toric varieties implies Ehrhart-Macdonald Reciprocity (see for example \cite[11.12.4]{D}). However, this theorem allows us to conclude that for a broad class of varieties the reverse implication also holds true. Hence it is indeed possible to prove one of the most famous results of algebraic geometry via lattice point counting!
The final step of our proof uses Serre Vanishing, Kodaira Vanishing for Rational Singularities and two results by Elkik \cite{E} on rational singularities in flat families. The main ingredient is the following observation.
\begin{theorem}\label{thm:vanishing}
Let $X$ be a Gorenstein Fano variety of dimension $d$ that has rational singularities and let $\L$ be an ample line bundle. Then the line bundle $\L$ is isomorphic to the anticanonical line bundle $\omega_X^{-1}$ if and only if the Hilbert polynomial $P_\L(n) := \chi(X, \L^n)$ of $\L$ fulfills
\[ P_\L(n) = (-1)^d P_\L(-n-1)\]
for all $n\in\NN$.
\end{theorem}
Our statement bears resemblance to a result by Kaveh and Villella in \cite{KV}, who were able to classify {\em anticanonical objects} in families of polyhedra associated to flag varieties purely via combinatorial conditions. However, their result needs stronger assumptions like Minkowski property of the occurring polytopes, which we do not need.
During our proof we stumbled upon a delicate detail. First of all, the limit divisors in our setting share an interesting property.
\begin{theorem}\label{thm:dpweil}
Under the assumptions of \cref{thm:main}, the divisor $D_\P$ is Weil.
\end{theorem}
Additionally, we noticed in \cref{thm:twopolynomials} that the Ehrhart quasi-polynomial of the polytope $\P$ must be a polynomial. Both of these properties can be thought of as integrality properties and we propose that they are in fact connected. This leads us to the following opinion.
\begin{conj}\label{conj:quasilattice}
Let $\P$ be a full-dimensional rational convex polytope with associated toric variety $(X_\P, D_\P)$. Then the Ehrhart quasi-polynomial of $\P$ is a polynomial if and only if $D_\P$ is Weil.
\end{conj}
Unfortunately, we were not able to state a proof of this claim.
\subsection{Structure}
We will start our work by recalling the needed results from Ehrhart Theory and a proof of \cref{thm:toricehrhart} in \cref{sec:ehrhart}. The introduction and study of dual-integral polytopes will be done in \cref{sec:dual}. In \cref{sec:degen} we will formalize the concept of embedded toric degenerations and prove \cref{thm:dpweil,cor:twopolynomials} as well as reprove the known implication of \cref{thm:main}. \cref{sec:vanishing} is dedicated to the proof of \cref{thm:vanishing} which will allow us to prove \cref{thm:main,thm:reflexive} in \cref{sec:main}. We will finish this paper with a brief recap of the flag variety case from \cite{S2,S3}, proving \cref{thm:flag}. Especially, we will argue why the proof of \cref{thm:main} in the flag variety case\,---\,which has been presented in \cite{S2}\,---\,is easier than in the general case.
\subsection{Acknowledgments}
This paper is a rewritten excerpt from my PhD thesis \cite{S1} under the supervision of Peter Littelmann. I am very grateful to Peter Littelmann, Michel Brion and Xin Fang for many helpful discussions, for introducing me to the various mathematical concepts involved in this paper and for their invaluable, continued support. Additionally, I would like to thank my PhD referees Ghislain Fourier and Kiumars Kaveh for their kind and helpful comments.
\section{Preliminaries}
Let us first recall some basic facts on toric varieties. By a result of Sumihiro in \cite{Su}, every normal toric variety can be realized as the toric variety associated to a pointed rational polyhedral fan. Additionally, by \cite[Theorem 6.2.1]{CLS} a normal toric variety is projective if and only if the corresponding fan is the normal fan of a (full-dimensional) rational convex polytope. To any full-dimensional rational convex polytope $\P$ we associate a normal projective toric variety by $X_\P := X_{\Sigma_\P}$ via the normal fan $\Sigma_\P$ of $\P$.
Additionally, polytopes have a strong connection with torus invariant divisors. Let $X_\Sigma$ be a normal toric variety. Then there exist torus invariant prime divisors $D_\rho$ on $X_\Sigma$ for every ray $\rho\in\Sigma(1)$. These divisors generate the divisor class group of $X_\Sigma$ (see \cite[First Proposition of Section 3.4]{F}). So we can choose a torus invariant representative in every divisor class.
The canonical divisor has a particularly nice representative. By \cite[First Proposition of Section 4.3]{F} we have $K_{X_\Sigma} \sim -\sum_{\rho\in\Sigma(1)}D_\rho$.
Furthermore, divisors on toric varieties correspond to polytopes in the following way.
We will always denote by $u_\rho$ the \textit{primitive ray generator} of the ray $\rho\in\Sigma(1)$, i.e. the unique lattice vector in $\rho$ whose coordinates are coprime.
Let $X_\Sigma$ be a normal toric variety and let $\P$ be a polytope such that $\Sigma$ is a refinement of the normal fan $\Sigma_\P$ of $\P$. We can always write $\P$ as
\[ \P = \{ x\in\RR^d \mid \langle x, u_\rho \rangle \leq b_\rho \text{ for all } \rho\in\Sigma(1)\},\]
where $b_\rho = \max_{x\in\P}\langle x,u_\rho\rangle\in\QQ$. Then we associate to $\P$ the torus invariant $\QQ$-Weil divisor $D_\P := \sum_{\rho\in\Sigma(1)} b_\rho D_\rho$.
Vice versa, to any torus invariant $\QQ$-Weil divisor $D = \sum_{\rho\in\Sigma(1)}a_\rho D_\rho$ we associate the (possibly empty) rational convex polytope
\[ \P_D := \{ x\in\RR^d \mid \langle x, u_\rho \rangle \leq a_\rho \text{ for all } \rho\in\Sigma(1)\}.\]
The following observations are immediate.
\begin{prop}\label{prop:toricdivisors}
Let $X_\Sigma$ be the normal toric variety corresponding to the pointed rational polyhedral fan $\Sigma\subseteq\RR^d$. Let $D$ be a torus invariant $\QQ$-Weil divisor on $X_\Sigma$ and let $\P$ be a rational convex polytope such that $\Sigma$ is a refinement of the normal fan $\Sigma_\P$.
Then the following properties hold.
\begin{enumerate}
\item $\P_{D_\P} = \P$.
\item $D_{\P_D} = D$ if $\Sigma_{\P_D} = \Sigma$.
\item $\P_{kD} = k\P_D$ for any $k \in \RR_{\geq0}$.
\item If $D \sim E$ for some torus invariant $\QQ$-Weil divisor $E$, we have $\P_E = \P_D + s$ for some $s \in\ZZ^d$.
\end{enumerate}
\end{prop}
One divisor on projective normal varieties behave particularly well (see \cite[Propositions 4.2.10 and 6.1.10]{CLS}).
\begin{prop}\label{prop:dpample}
Let $X_\P$ be the normal projective toric variety associated to the rational convex polytope $\P$. Then the divisor $D_\P$ is $\QQ$-Cartier and ample.
\end{prop}
To go from $\QQ$-Weil divisors to Weil divisors, we will need a rounding operation.
\begin{defi}
Let $X_\Sigma$ be a normal toric variety and $D = \sum_{\rho\in\Sigma(1)}$ a $\QQ$-Weil divisor on $X_\Sigma$. The \textbf{round-down} of $D$ is defined as $\lfloor D \rfloor := \sum_{\rho\in\Sigma(1)} \lfloor a_\rho \rfloor D_\rho$. The \textbf{round-up} $\lceil D \rceil$ is defined analogously.
\end{defi}
For easier notation later-on we will define a similar operation for polytopes.
\begin{defi}
Let $\P$ be a rational convex polytope with normal fan $\Sigma_\P$. Let $u_\rho$ denote the primitive ray generators of $\Sigma_\P$. Then there exist unique rational coefficients $b_\rho\in\QQ$ such that
\[\P = \{ x\in\RR^d \mid \langle x, u_\rho \rangle \leq b_\rho \text{ for all } \rho\in\Sigma_\P(1)\}. \]
The \textbf{round-down} of $\P$ is defined as
\[ \P = \{ x\in\RR^d \mid \langle x, u_\rho \rangle \leq \lfloor b_\rho \rfloor \text{ for all } \rho\in\Sigma_\P(1)\}. \]
The \textbf{round-up} of $\P$ is defined analogously.
\end{defi}
The following connection is clear.
\begin{prop}\label{prop:toricdivisors2}
Let $X_\Sigma$ be a normal toric variety and $D$ a torus invariant $\QQ$-Weil divisor on $X_\Sigma$. Then $\lfloor \P_D\rfloor = \P_{\lfloor D \rfloor}$ and $\lceil \P_D\rceil = \P_{\lceil D \rceil}$.
\end{prop}
The connection between divisors and polytopes gives us possibilities to answer questions about cohomology. As an example we will state \cite[Proposition 4.3.3]{CLS}.
\begin{prop}\label{prop:toricehrhartweil}
Let $X_\Sigma$ be the normal toric variety and $D$ a torus invariant Weil divisor on $X_\Sigma$. Then
\[ h^0(X_\Sigma, \O_{X_{\Sigma}}(D)) = \#(\P_D\cap\ZZ^d). \]
\end{prop}
This standard fact will motivate the next section.
\section{Ehrhart Theory}\label{sec:ehrhart}
Given a subset $S\subseteq\RR^d$ and a positive integer $n$, one might be interested in the number of lattice points in the dilation $nS$\,---\,i.e. the cardinality of $nS\cap\ZZ^d$. It turns out that there lies a beautiful theory behind this simple question if one starts with a convex rational polytope\,---\,called {\em Ehrhart Theory}. A well-written introduction into this theory is given in \cite{BR}. The main object of this theory is the following.
\begin{nota}\label{defi:ehrhart}
Let $S \subseteq\RR^d$ be an arbitrary subset. Then for every integer $n\in\NN$ we will denote the number of lattice points in the $n$-th dilation of $S$ by $L_S(n)$, i.e.
\[L_S(n) := \#(nS\cap\ZZ^d).\]
\end{nota}
To formulate the birth result in this theory\,---\,called {\em Ehrhart-Macdonald Reciprocity}\,---\,, we need the following definition.
\begin{defi}
A \textbf{quasi-polynomial} over $\RR$ is a function $f\colon\RR\to\RR$ that can be written as
\[ f(x) = a_d(x) x^d + a_{d-1} (x)x^{d-1} + \ldots + a_1(x)x + a_0(x) \]
for some periodic functions $a_0, \ldots, a_d$ with integral period and $a_d \not\equiv 0$. We call $d$ the \textbf{degree} of $f$.
\end{defi}
\begin{rem}
An equivalent definition would be that a function ${f\colon\RR\to\RR}$ is called a quasi-polynomial if there exists an integer $T$ and polynomials \sloppy{$f_1, \ldots, f_T \in\RR[x]$} such that
\[ f(n) = f_{i}(n) \hspace{10pt}\text{ if } i\equiv n \mod T \]
for every integer $n\in\NN$.
\end{rem}
We can now state a beautiful result, found for example in \cite[Theorem 4.1]{BR}.
\begin{theorem}[Ehrhart-Macdonald Reciprocity]\label{thm:emr}
Let $\P$ be a rational convex polytope. Then there exists a quasi-polynomial $l_\P$ of degree $\dim\P$\,---\,called the \textbf{Ehrhart quasi-polynomial}\,---\,such that
\[l_p(n) = L_\P(n) \text{ for all } n \in\NN.\]
Any such quasi-polynomial $l_\P$ fulfills
\[ l_\P(-n) = (-1)^{\dim\P}L_{\interior \P}(n) \text{ for all } n\in\NN. \]
Furthermore, if $\P$ is a lattice polytope, $l_\P$ can be chosen to be a rational polynomial.
\end{theorem}
\begin{rem}
Since the period of a polynomial is not quite unique and quasi-polynomials are not uniquely determined by their values on integers, for a given polytope $\P$ we have many different quasi-polynomials $l_\P$ fulfilling Ehrhart-Macdo\-nald Reciprocity. Thankfully, we are generally only interested in evaluating those polynomials on integers\,---\,and these values are unique. So we will just call any quasi-polynomial {\em the} Ehrhart quasi-polynomial of $\P$ if it fulfills Ehrhart-Macdo\-nald Reciprocity and among all those quasi-polynomials there exists none of strictly smaller period. To simplify notation we will denote this quasi-polynomial by $L_\P$ too.
\end{rem}
\begin{ex}
Let $\P = [0, 1/2] \subset \RR$. Then $L_\P(n) = \lfloor\frac{n}{2}\rfloor$, so one could choose for example
\[ l_\P(x) = \frac{1}{2}x + 1-\frac{1}{2}\sin^2\left(\frac{\pi x}{2}\right) \text{ or }l_\P(x) = \frac{1}{2}x + \frac{3+\cos\pi x}{4}.\]
\end{ex}
We have seen that every lattice polytope will have an Ehrhart polynomial but there are also non-lattice convex polytopes whose Ehrhart quasi-polynomial is a polynomial. So the following category of polytopes should be quite interesting.
\begin{defi}
A \textbf{quasi-lattice polytope} is a rational convex polytope whose Ehrhart quasi-polynomial is a polynomial.
\end{defi}
Examples of quasi-lattice polytopes can be found for example in \cite[Section 5]{S2}.
We will conclude this overview by proving the following new observation.
\begin{theorem*}[\cref{thm:toricehrhart}]
Let $\P\subseteq\RR^d$ be a full-dimensional rational convex polytope. Let $X_\P$ denote the associated normal projective toric variety and $D_\P$ the associated torus invariant $\QQ$-Weil divisor. Then
\[\#(n\P\cap\ZZ^d) = \chi(X_\P, \O_{X_\P}(\lfloor nD_\P\rfloor))\]
and
\[\#(\interior n\P \cap \ZZ^d) = \chi(X, \O_{X_\P}(\lceil nD_\P\rceil + K_{X_\P}))\]
for all $n \in \NN$.
\end{theorem*}
\begin{rem}
\cref{thm:toricehrhart} would in theory also provide a method to compute the Ehrhart quasi-polynomial of an arbitrary rational convex polytope via cohomology groups\,---\,although in practice this might be quite challenging.
\end{rem}
Before we state the proof of this result, we show the following useful identities.
\begin{prop}\label{prop:round2}
Let $\P\subseteq\RR^d$ be a full-dimensional rational convex polytope and let $X_\P$ denote the associated normal projective toric variety. Then
\begin{enumerate}
\item $L_\P(n) = L_{\lfloor n\P\rfloor}(1)$,
\item $L_{\interior \P}(n) = L_{\interior n\P}(1) = L_{\interior \lceil n\P\rceil}(1) = L_{\P_{\lceil nD_\P \rceil + K_{X_\P}}}(1)$
\end{enumerate}
for all $n\in\NN$.
\end{prop}
\begin{proof}
Let $n\in\NN$. For every ray $\rho$ in the normal fan $\Sigma_\P$ let us denote its primitive ray generator by $u_\rho$. We know that $n\P$ can be written as
\[ n\P = \{ x \in \RR^d \mid \langle x, u_\rho \rangle \leq nb_\rho \text{ for all } \rho \in\Sigma_\P(1) \} \]
for some rational numbers $b_\rho\in\QQ$, $\rho\in\Sigma_\P(1)$.
For the first property, notice that for any lattice point $x\in\ZZ^d$ and any ray $\rho\in\Sigma_\P(1)$, the number $\langle u_\rho, x\rangle$ will be an integer, hence
\[ \langle x,u_\rho\rangle \leq nb_\rho \hspace{10pt}\Leftrightarrow\hspace{10pt} \langle x,u_\rho\rangle \leq \lfloor nb_\rho\rfloor,\]
which proves the claim
The first equality of the second part is clear since
\begin{align*}
x \in n\interior\P &\hspace{10pt}\Leftrightarrow\hspace{10pt} \left\langle\frac{1}{n}\cdot x, u_\rho \right\rangle < b_\rho \text{ for all } \rho \in\Sigma_\P(1) \\
&\hspace{10pt}\Leftrightarrow\hspace{10pt} \langle x, u_\rho \rangle < nb_\rho\text{ for all } \rho \in\Sigma_\P(1)\\
&\hspace{10pt}\Leftrightarrow\hspace{10pt} x \in \interior n\P
\end{align*}
for every $x\in\RR^d$.
The second equality follows for the same reason as the first property (on the round-down), since for any $x \in\ZZ^d$ we have
\[ \langle x,u_\rho\rangle < nb_\rho \hspace{10pt}\Leftrightarrow\hspace{10pt} \langle x,u_\rho\rangle < \lceil nb_\rho\rceil\]
for any ray $\rho\in\Sigma_\P(1)$ because the scalar products are integers.
For the third equality, recall that the canonical divisor is linearly equivalent to $-\sum_{\rho\in\Sigma_\P(1)}D_\rho$. So the two polytopes in the equation can be calculated as
\[ \lceil n\P\rceil = \{ x\in\RR \mid \langle x, u_\rho \rangle \leq \lceil nb_\rho \rceil \text{ for all } \rho \in\Sigma_\P(1)\} \]
and
\[ \P_{\lceil nD_\P\rceil+K_{X_\P}} = \{ x\in\RR \mid \langle x, u_\rho \rangle \leq \lceil nb_\rho \rceil - 1 \text{ for all } \rho \in\Sigma_\P(1)\}. \]
So for any $x\in\ZZ^d$ we get
\[ \langle x,u_\rho\rangle < \lceil nb_\rho\rceil \hspace{10pt}\Leftrightarrow\hspace{10pt} \langle x,u_\rho\rangle \leq \lceil nb_\rho\rceil-1\]
because the scalar products are integers again.
\end{proof}
We are now able to prove the important formulae.
\begin{proof}[Proof of \cref{thm:toricehrhart}]
Let us start with the first equality. The divisor $nD_\P$ will be $\QQ$-Cartier and ample, since $D_\P$ is $\QQ$-Cartier and ample (see \cref{prop:dpample}). So by Demazure Vanishing (see \cite[Theorem 9.3.5]{CLS}) we know that all higher cohomology groups of the round-down divisor $\lfloor nD_\P \rfloor$ must vanish. We can thus apply \cref{prop:toricehrhartweil} to the torus invariant Weil divisor $\lfloor nD_\P \rfloor$ to get
\[ \chi(X_\P, \O_{X_\P}(\lfloor nD_\P \rfloor)) = h^0(X_\P, \O_{X_\P}(\lfloor nD_\P \rfloor)) = L_{\P_{\lfloor nD_\P \rfloor}} (1). \]
But because of \cref{prop:toricdivisors2} this polytope is equal to
\[ \P_{\lfloor nD_\P \rfloor} = \lfloor \P_{nD_\P}\rfloor = \lfloor n\P_{D_\P}\rfloor = \lfloor n\P\rfloor, \]
which proves the first claim using the observation
\[ L_{\lfloor n\P\rfloor}(1) = L_\P(n)\]
from \cref{prop:round2}.
Let us now consider the second case. By the toric version of the Kawamata-Viehweg Vanishing Theorem (see \cite[Corollary 2.5]{M} for a proof without Serre Duality) we know that all higher cohomology groups of the torus invariant Weil divisor $\lceil nD_\P\rceil + K_{X_\P}$ vanish. Using \cref{prop:toricehrhartweil} again we get
\[ \chi(X_\P, \O_{X_\P}(\lceil nD_\P\rceil + K_{X_\P})) = h^0(X_\P, \O_{X_\P}(\lceil nD_\P\rceil + K_{X_\P})) = L_{\P_{\lceil nD_\P\rceil + K_{X_\P}}} (1). \]
By our previous observation in \cref{prop:round2} we know that this number can be rewritten as $L_{\interior \P}(n)$ which had to be proven.
\end{proof}
\section{Dual-Fano and Dual-Integral Polytopes}\label{sec:dual}
Before introducing some new classes of polytopes, let us introduce a shorthand notation.
\begin{nota}
Let $v\in\RR^d$ and $b\in\RR$. Then we define the halfspaces
\[ H^+_{v,b} := \{ x \in \RR^d\mid \langle x,v\rangle \geq b \} \text{ and } H^-_{v,b} := \{ x\in\RR^d\mid\langle x, v\rangle \leq b\} \]
as well as the affine hyperplane
\[ H_{v,b} := H^+_{v,b}\cap H^-_{v,b}. \]
If $b=0$ we will just write $H^+_v := H^+_{v,0}$, $H^-_v := H^-_{v,0}$ and $H_v := H_{v,0}$.
\end{nota}
We will now recall the standard duality results on polytopes.
\begin{defi}
Let $S\subseteq\RR^d$ be a set. The \textbf{(polar) dual} $S^*$ of the set $S$ is defined as
\[ \P^* := \{ x \in \RR^d \mid \langle x,s \rangle \leq 1 \text{ for all } s \in S \}. \]
\end{defi}
We will mostly omit the word {\em polar} when speaking about dual polytopes.
The following is a collection of standard facts that can be found for example in \cite[Theorem 2.11]{Z}).
\begin{theorem}\label{thm:polardual}
Let $\P = \bigcap_{i=1}^r H^-_{\alpha_i,b_i}\subseteq\RR^d$ be a convex polytope with vertices $v_1, \ldots, v_s$.
\begin{enumerate}
\item $\P \subseteq (\P^*)^*$ and equality holds if and only if $0 \in \P$.
\item If $0 \in \interior\P$ (implying that $\P$ is full-dimensional), then the $b_i$'s can be chosen non-zero. In this case, the polar dual $\P^*$ is a convex polytope and it can be calculated as \[\P^* = \conv(b_1^{-1}\alpha_1, \ldots, b_r^{-1}\alpha_r) = \bigcap_{i=1}^s H^-_{v_i,1}.\]
\item If $0 \in \interior \P$ then $\P$ is rational if and only if $\P^*$ is rational.
\item For every $\lambda \in \RR_{>0}$ we have $(\lambda\P)^* = \lambda^{-1}\P^*$.
\item If $\P^*$ is a convex polytope, there exists an inclusion-reversing bijection between the faces of $\P$ and the faces of $\P^*$.
\end{enumerate}
\end{theorem}
\begin{defi}
A convex polytope $\P$ is called \textbf{reflexive} if both $\P$ itself and its polar dual $\P^*$ are lattice polytopes.
\end{defi}
Since this notion is too rigid for our applications we want to look at some weaker properties.
\begin{defi}\label{defi:dualfano}
\begin{enumerate}
\item A convex polytope is called a \textbf{quasi-reflexive} polytope if there exists a lattice point $p$ such that the translated polytope $\P-p$ is reflexive.
\item A convex polytope is called a \textbf{Fano} polytope if its vertices are primitive lattice vectors.
\item A convex polytope $\P$ is called a \textbf{dual-Fano} polytope there exists a lattice point $p$ such that the dual of the translated polytope $\P-p$ is a Fano polytope.
\item A convex polytope $\P$ is called a \textbf{dual-integral} polytope if there exists a lattice point $p$ such that the dual of the translated polytope $\P-p$ is a lattice polytope.
\end{enumerate}
\end{defi}
\begin{rem}
While the notion of Fano polytopes is quite standard (see for example \cite{N}), the other notions are not. But the naming should be quite self-explanatory and it will be a useful shorthand.
It should be noted that these definitions deviate from the definitions in \cite{S1}. In fact, we not only changed the name of dual-integral polytopes from {\em weakly dual-Fano} polytopes, but we also incorporated the lattice point translations into the definition and added the new definition of quasi-reflexive polytopes.
\end{rem}
\begin{rem}\label{rem:dualfano}
Notice that every quasi-reflexive, dual-Fano or dual-integral polytope must contain a lattice point in its interior that we can chose as the translation vector. This is due to the fact that after translation by a lattice vector the translated dual of the translated polytope must be bounded, so it must contain the origin in its interior. Additionally, by \cref{thm:polardual} each of these polytopes must be rational since they are the lattice point translations of the dual of a lattice (hence rational) polytope.
\end{rem}
Before we get into the details, let us look at an example.
\begin{ex}\label{ex:dualfano}
In \cref{fig:dualfano} we see the sketch of three different polytopes in $\RR^2$ and their dual polytopes. We will notice in this example that the three classes of polytopes introduced before are distinct.
The first polytope is given by the inequalities
\[\{(x,y)\in\RR^2\mid x\geq -1, y\geq -1, 2x+3y\leq 1\} \]
and we clearly see from the sketch that it is reflexive, dual-Fano and dual-integral.
The second polytope is given by the inequalities
\[\{(x,y)\in\RR^2\mid x\geq -1, y\geq -1, x+3y\leq 1\} \]
and we clearly see from the sketch that it is dual-Fano and dual-integral. However, since the upper left vertex is not integral, this polytope is not reflexive.
Finally, the third polytope is given by the inequalities
\[\{(x,y)\in\RR^2\mid x\geq -1, y\geq -1, 3x+3y\leq 1\}. \]
From the sketch we see that it is dual-integral. But it is not a lattice polytope, so it cannot be reflexive. Additionally the upper right vertex of its dual polytope is the point $(3,3)$ which is an integral multiple of the lattice point $(1,1)$. Hence the dual polytope will not be Fano.
\end{ex}
\begin{figure}[htb]\centering
\caption{Sketch of the three different polytopes in \cref{ex:dualfano} and their dual polytopes. The first one is reflexive, dual-Fano and dual-integral; the second one is dual-Fano and dual-integral but not reflexive; the third one is only dual-integral.}\label{fig:dualfano}
\begin{tabular}{c|c|c|c}
&reflexive&dual-Fano&dual-integral\\
\begin{tikzpicture}
\node (1) at (0,1) { };
\node (2) at (0,-1) { };
\node (3) at (0,0) {$\P$};
\end{tikzpicture}&
\begin{tikzpicture}[scale=.5]
\draw[step=1,help lines,black!20] (-1.95,-1.95) grid (2.95,1.95);
\draw[->] (-2,0) -- (3,0);
\draw[->] (0,-2) -- (0,2);
\foreach \Point in {(-1,-1), (-1,0), (-1,1), (0,-1), (0,0), (0,1) , (1,-1), (1,0), (1,1), (2,-1), (2,0), (2,1)}
\draw[fill=black] \Point circle (0.075);
\coordinate (1) at (-1,1);
\coordinate (2) at (-1,-1);
\coordinate (3) at (2,-1);
\draw[thick] (1) -- (2) -- (3) -- node[xshift=10pt,yshift=10pt] {\tiny $2x+3y\leq1$} cycle;
\end{tikzpicture}
&
\begin{tikzpicture}[scale=.5]
\draw[step=1,help lines,black!20] (-1.95,-1.95) grid (4.95,1.95);
\draw[->] (-2,0) -- (5,0);
\draw[->] (0,-2) -- (0,2);
\foreach \Point in {(-1,-1), (-1,0), (-1,1), (0,-1), (0,0), (0,1) , (1,-1), (1,0), (1,1), (2,-1), (2,0), (2,1), (3,-1), (3,0), (3,1), (4,-1), (4,0), (4,1)}
\draw[fill=black] \Point circle (0.075);
\coordinate (1) at (-1,2/3);
\coordinate (2) at (-1,-1);
\coordinate (3) at (4,-1);
\draw[thick] (1) -- (2) -- (3) -- node[xshift=10pt,yshift=10pt] {\tiny $x+3y\leq1$} cycle;
\end{tikzpicture}
&
\begin{tikzpicture}[scale=.5]
\draw[step=1,help lines,black!20] (-1.95,-1.95) grid (1.95,1.95);
\draw[->] (-2,0) -- (2,0);
\draw[->] (0,-2) -- (0,2);
\foreach \Point in {(-1,-1), (-1,0), (-1,1), (0,-1), (0,0), (0,1) , (1,-1), (1,0), (1,1)}
\draw[fill=black] \Point circle (0.075);
\coordinate (1) at (-1,4/3);
\coordinate (2) at (-1,-1);
\coordinate (3) at (4/3,-1);
\draw[thick] (1) -- (2) -- (3) -- node[xshift=15pt,yshift=10pt] {\tiny $3x+3y\leq1$} cycle;
\end{tikzpicture}\\
&$\downarrow$&$\downarrow$&$\downarrow$\\
\begin{tikzpicture}
\node (1) at (0,1) { };
\node (2) at (0,-1) { };
\node (3) at (0,0) {$\P^*$};
\end{tikzpicture}
&
\begin{tikzpicture}[scale=.5]
\draw[step=1,help lines,black!20] (-1.95,-1.95) grid (2.95,3.95);
\draw[->] (-2,0) -- (3,0);
\draw[->] (0,-2) -- (0,4);
\foreach \Point in {(-1,-1), (-1,0), (-1,1), (-1,2), (-1,3), (0,-1), (0,0), (0,1), (0,2), (0,3), (1,-1), (1,0), (1,1), (1,2), (1,3), (2,-1), (2,0), (2,1), (2,2), (2,3)}
\draw[fill=black] \Point circle (0.075);
\coordinate (1) at (-1,0);
\coordinate (2) at (0,-1);
\coordinate (3) at (2,3);
\draw[thick] (1) -- (2) -- (3) -- cycle;
\end{tikzpicture}
&
\begin{tikzpicture}[scale=.5]
\draw[step=1,help lines,black!20] (-1.95,-1.95) grid (1.95,3.95);
\draw[->] (-2,0) -- (2,0);
\draw[->] (0,-2) -- (0,4);
\foreach \Point in {(-1,-1), (-1,0), (-1,1), (-1,2), (-1,3), (0,-1), (0,0), (0,1), (0,2), (0,3), (1,-1), (1,0), (1,1), (1,2), (1,3)}
\draw[fill=black] \Point circle (0.075);
\coordinate (1) at (-1,0);
\coordinate (2) at (0,-1);
\coordinate (3) at (1,3);
\draw[thick] (1) -- (2) -- (3) -- cycle;
\end{tikzpicture}
&
\begin{tikzpicture}[scale=.5]
\draw[step=1,help lines,black!20] (-1.95,-1.95) grid (3.95,3.95);
\draw[->] (-2,0) -- (4,0);
\draw[->] (0,-2) -- (0,4);
\foreach \Point in {(-1,-1), (-1,0), (-1,1), (-1,2), (-1,3), (0,-1), (0,0), (0,1), (0,2), (0,3), (1,-1), (1,0), (1,1), (1,2), (1,3), (2,-1), (2,0), (2,1), (2,2), (2,3), (3,-1), (3,0), (3,1), (3,2), (3,3)}
\draw[fill=black] \Point circle (0.075);
\coordinate (1) at (-1,0);
\coordinate (2) at (0,-1);
\coordinate (3) at (3,3);
\draw[thick] (1) -- (2) -- (3) -- cycle;
\end{tikzpicture}
\end{tabular}
\end{figure}
Judging from the example we might guess that there is an inclusion relation between the three classes of polytopes and we could even guess how a standard form of these different polytopes would look like. We will formalize this in the next theorem and its corollary.
\begin{theorem}\label{thm:dualfano}
\begin{enumerate}
\item Every quasi-reflexive polytope is dual-Fano.
\item Every dual-Fano polytope is dual-integral.
\item A polytope is quasi-reflexive if and only if it is a dual-integral lattice polytope.
\end{enumerate}
\end{theorem}
\begin{proof}
Let $\P\subseteq\RR^d$ be a convex polytope. We can assume $\P$ to be rational and of full-dimension since all properties occurring in the statements imply rationality and full-dimensionality. Let $\Sigma_\P$ denote its normal fan with primitive ray generators $u_\rho$, $\rho\in\Sigma(1)$, i.e. $u_\rho$ is the unique lattice vector in $\rho$ whose coordinates are coprime.
For the first claim, assume that $\P$ is a quasi-reflexive polytope. By \cref{rem:dualfano} we know that there exists a lattice point $p$ in the interior of $\P$. Let $\Q := \P-p$. Then $\Sigma_\Q = \Sigma_\P$. We can find rational numbers $b_\rho$ such that
\[ \Q = \{ x \in \RR^d \mid \langle x, u_\rho \rangle \leq b_\rho \text{ for all } \rho\in\Sigma_\P(1)\}. \]
Because $\P$ and hence $\Q$ is a lattice polytope, the hyperplanes $H_{u_\rho, b_\rho}$ must contain lattice points, which requires all $b_\rho$ to be integers. Since $0\in\interior\Q$, all $b_\rho$ must be strictly positive. By \cref{thm:polardual} we know that vertices of the dual polytope are given by
\[ \vert\Q* = \left\{\frac{u_\rho}{b_\rho}\,\,\middle|\,\,\rho\in\Sigma_\P(1)\right\}.\]
Since $\Q^*$ must be a lattice polytope, we know that all $b_\rho$ must be equal to $1$ (because the $u_\rho$ are primitive). So we get
\[ \vert\Q^* = \left\{u_\rho\,\,\middle|\,\,\rho\in\Sigma_\P(1)\right\},\]
which means that $\Q^*$ is a Fano polytope, i.e. $\P$ is a dual-Fano polytope.
The second claim is obvious.
Notice that claims (i) and (ii) already prove one direction of claim (iii). But the other direction follows immediately from the definitions.
\end{proof}
Interestingly, from the proof of \cref{thm:dualfano} we get the following descriptions.
\begin{cor}\label{cor:dualfano}
Let $\P\subseteq\RR^d$ be a full-dimensional rational convex polytope and let $u_\rho$, $\rho\in\Sigma_\P(1)$, denote the primitive ray generators of the normal fan $\Sigma_\P$ of $\P$.
\begin{enumerate}
\item The polytope $\P$ is dual-integral if and only if there exist strictly positive integers $k_\rho$, $\rho\in\Sigma_\P(1)$, such that
\[ \P = p + \{ x\in\RR^d \mid \langle x, u_\rho \rangle \leq k_\rho^{-1} \text{ for all } \rho\in\Sigma_\P(1)\}\]
for some $p\in\interior\P\cap\ZZ^d$.
\item The polytope $\P$ is dual-Fano if and only if
\[ \P = p + \{ x\in\RR^d \mid \langle x, u_\rho \rangle \leq 1 \text{ for all } \rho\in\Sigma_\P(1)\}\]
for some $p\in\interior\P\cap\ZZ^d$.
\item The polytope $\P$ is quasi-reflexive if and only if it is a lattice polytope and
\[ \P = p + \{ x\in\RR^d \mid \langle x, u_\rho \rangle \leq 1 \text{ for all } \rho\in\Sigma_\P(1)\}\]
for some $p\in\interior\P\cap\ZZ^d$.
\end{enumerate}
\end{cor}
In light of toric geometry we repeat the following observation. It can be found in in Nill's doctoral thesis \cite[Proposition 1.4]{N} and follows from a result by Batyrev \cite[Proposition 2.2.23]{B}. Recall that the anticanonical divisor on a normal toric variety $X_\Sigma$ is given by $-K_{X_\Sigma} = \sum_{\rho\in\Sigma(1)}D_\rho$, hence its associated polytope $\P_{-K_{X_\Sigma}}$ can be written as
\[ \P_{-K_{X_\P}} = \{ x\in\RR^d \mid \langle x, u_\rho \rangle \leq 1 \text{ for all } \rho\in\Sigma_\P(1)\}.\]
\begin{theorem}[Batyrev, Nill]\label{thm:toricfano}
The map $(X_\P, -K_{X_\P}) \to \P_{-K_{X_\P}}$ induces a one-to-one correspondence between isomorphism classes of anticanonically polarized $\QQ$-Gorenstein Fano toric varieties and dual-Fano polytopes up to lattice translations. The same statement holds true for Gorenstein Fano toric varieties and reflexive polytopes.
\end{theorem}
We will conclude this overview with a beautiful result by Hibi \cite{Hi}.
\begin{theorem}[Hibi]\label{thm:hibi}
A full-dimensional rational convex polytope $\P\subseteq\RR^d$ is dual-integral if and only if
\[ \#(n\P\cap\ZZ^d) = \#(\interior(n+1)\P\cap\ZZ^d) \]
for all $n \in \NN$.
\end{theorem}
\begin{rem}\label{rem:hibi}
This formulation of Hibi's result is not his original formulation but this version shows clearer which beautiful magic is actual happening in the background. First of all, Hibi did not use our notion of dual-integral polytopes. Secondly, for computational purposes it is helpful to notice that by Ehrhart-Macdonald Reciprocity (see \cref{thm:emr}) the condition of Hibi's Theorem is equivalent to the property that
\[ L_\P(n) = (-1)^dL_\P(-n-1) \]
for all $n \in \NN$\,---\,a condition that can be verified by studying the Ehrhart quasi-polynomial alone. This was his original statement.
\end{rem}
The following consequence is immediate.
\begin{cor}\label{cor:hibi}
Every dual-integral polytope contains precisely one lattice point in its interior.
\end{cor}
From this observation we see that the lattice vectors in \cref{defi:dualfano} are unique. Additionally, the following notation is verified.
\begin{nota}
Let $\P$ be a dual-integral polytope. The unique interior lattice point in $\P$ will be denoted by $p_\P$.
\end{nota}
We will conclude this section with an interesting little observation that does not generalize to higher dimensions.
\begin{theorem}
Every two-dimensional quasi-lattice polytope is dual-integral if and only if it contains exactly one interior lattice point.
\end{theorem}
\begin{proof}
The first direction is clear. So let $\P\subseteq\RR^2$ be a full-dimensional quasi-lattice polytope with $L_\P(-1) = \#(\interior \P\cap\ZZ^2) = 1$. By \cref{thm:emr} we can find rational coefficients $a$, $b$ and $c$ such that the Ehrhart polynomial of $\P$ is given by
\[ L_\P(x) = ax^2+bx+c. \]
Since $L_\P(0) = 1$ we have $c=1$. Additionally, $L_\P(-1) = 1$ implies that $a = b$. So we see that
\[ L_\P(x) = ax(x+1) + 1 \]
for some strictly positive rational coefficient $a$. Hence
\[ L_\P(-n-1) = a(-n-1)(-n-1+1) + 1 = an(n+1) + 1 = L_\P(n), \]
so Hibi's \cref{thm:hibi} implies the claim.
\end{proof}
\section{Embedded Toric Degenerations}\label{sec:degen}
The following definition could be seen as the standard definition of toric degenerations.
\begin{defi}[First try]
Let $X$ be a normal complex variety. We say that the variety $X$ admits a \textbf{toric degeneration} to the normal toric variety $X_\Sigma$ if there exists a complex variety $\X$ and a flat morphism $\X \to \AA^1$ such that all fibers $\X_t$, $t\in\AA^1$, are isomorphic to $X$ and the special fiber $\X_0$ is isomorphic to $X_\Sigma$.
\end{defi}
However, this definition would be to weak for our applications. So we need to strengthen it as follows. Additionally, we are only interested in the case of projective varieties.
\begin{defi}\label{defi:toricdegen}
Let $X$ be a normal projective complex variety of dimension $d$ and let $D$ be an ample Cartier Weil divisor on $X$. Let $X_\P$ be the normal projective toric variety associated to the rational convex polytope $\P\subseteq\RR^d$. Let $D_\P$ denote the ample $\QQ$-Cartier $\QQ$-Weil divisor on $X_\P$ associated to $\P$ (see \cref{prop:dpample} for these properties of $D_\P$). We say that the pair $(X,D)$ admits a \textbf{toric degeneration} to the pair $(X_\P, D_\P)$ if there exists a complex variety $\X$ and a morphism
\[\pi \colon \X \to \AA^1\]
such that
\begin{enumerate}
\item $\pi$ is projective and flat,
\item $\X_t \simeq X$ for all $t \neq 0$ and $\X_0 \simeq X_\P$, and
\item for every $n \in \NN$ there exists a divisorial sheaf $\F^{(n)}$ on $\X$ such that $\F^{(n)}|_{\X_t} \simeq \O_X(nD)$ and $\F^{(n)}|_{\X_0} \simeq \O_{X_\P}(\lfloor n D_\P \rfloor )$.
\end{enumerate}
The variety $X_\P$ is called the \textbf{limit} of $X$ under the degeneration $\pi$.
\end{defi}
\begin{rem}
Whenever we say that {\em a variety admits a toric degeneration}, we mean this in the sense of \cref{defi:toricdegen} and not in the standard sense! We acknowledge that our terminology is not standard. Yet, we will see that basically all known examples of toric degenerations in the usual sense are also toric degenerations in our sense. The main reason behind this change is basically that we want to keep track of the embedding of our varieties. So one could think of our degenerations as {\em embedded toric degenerations}.
\end{rem}
\begin{rem}
We will use the analogue definition of toric degenerations for pairs $(X, \L)$ where $\L$ is an ample line bundle on the normal projective variety $X$.
\end{rem}
\begin{rem}\label{rem:birational}
This definition should not come completely unexpected. It has been proven by Anderson that toric degenerations in our sense arise naturally in the setting of Newton-Okounkov bodies (see \cite[Theorem 1, Corollary 2 and Lemma 3]{A}).
In the case of flag varieties there is an even broader formalism. Fang, Fourier and Littelmann constructed toric degenerations via so called {\em birational sequences} in \cite{FaFL}. They proved in \cite[Theorem 6]{FaFL} that their toric degenerations satisfy the properties of \cref{defi:toricdegen}. Additionally, they were able to show that every birational sequence defines a valuation. Hence every toric degeneration via birational sequences can be seen as a toric degeneration associated to a Newton-Okounkov body. It is noticeable that basically all polytopes in representation theory can be constructed via birational sequences and hence are Newton-Okounkov bodies. We will make use of this fact in \cref{sec:flag}.
\end{rem}
\begin{defi}
Let $\X\to\AA^1$ be a toric degeneration. The variety $\X$ is called the \textbf{degeneration space}. The fiber $\X_0$ is called the \textbf{special fiber} and the fiber $\X_1$\,---\,being isomorphic to any $\X_t$, $t\neq0,$\,---\,is called the \textbf{general fiber}.
\end{defi}
We will collect some useful facts of the degeneration space $\X$. It should be noted that the following three propositions also hold under the usual definition of toric degenerations.
\begin{prop}\label{prop:degenerationspace}
Let $\X \to \AA^1$ be a toric degeneration in the sense of \cref{defi:toricdegen}. Then $\X$ is quasi-projective, normal and has rational singularities.
\end{prop}
\begin{proof}
$\X$ is clearly quasi-projective since it is the domain of a projective morphism. Normality follows directly from Serre's criterion of normality since every fiber $\X_t$, $t\in\RR$, is normal (see \cite[Théorème 5.8.6]{EGAIV2}). Rationality of singularities follows from the fact that normal toric varieties have rational singularities (see \cite[Theorem 11.4.2]{CLS}). So by a result of Elkik (see \cite[Théorème 4]{E}) there exists an open neighborhood $0\in U\subseteq \AA^1$ such that all fibers $\X_t$, $t\in U$, have rational singularities. But all fibers are isomorphic (except for the special fiber), hence all fibers have rational singularities. This implies that $\X$ has rational singularities by another result of Elkik (see \cite[Théorème 5]{E}).
\end{proof}
\begin{prop}\label{prop:x0}
Let $\X \to \AA^1$ be a toric degeneration in the sense of \cref{defi:toricdegen}. Then each fiber $\X_t$ is a principal prime divisor on $\X$.
\end{prop}
\begin{proof}
By \cite[Chapter III, Corollary 9.6]{Ha} we know that $\X_t$ is a subvariety of codimension 1. The irreducibility is clear since $\X_t \simeq X$ or $\X_t \simeq X_\Sigma$. Notice that $\pi\in\O_\X(\X) \subseteq \CC(\X)$, so $\pi$ is a rational function on $\X$ and $\X_t = \div(\pi-t)$ is principal.
\end{proof}
\begin{prop}\label{prop:clx}
Let $\X \to \AA^1$ be a toric degeneration in the sense of \cref{defi:toricdegen} and let $X$ be isomorphic to the general fiber. Then for any $t\neq0$ the restriction map $\D\mapsto \D|_{\X_t}$ induces an isomorphism of divisor class groups
\[\Cl(\X) \simeq \Cl(X).\]
\end{prop}
\begin{proof}
Consider the variety $X\times\AA^1$ and the natural projection $\tau\colon X\times\AA^1\to \AA^1$. Let $Z := \tau^{-1}(0)$.
Then the open set $U := \X\setminus\X_0$ is isomorphic to the open set $V := (X\times\AA^1 )\setminus Z$.
By \cref{prop:x0} we know that the special fiber $\X_0$ is a principal prime divisor on $\X$. By the same arguments $Z$ is a principal prime divisor on $X\times\AA^1$. So repeated application of \cite[Propositions 6.5 and 6.6]{Ha} gives a sequence of isomorphisms
\begin{align*} \Cl(\X) \to \Cl(\X)/\ZZ\lbrack\X_0\rbrack \to &\Cl(U)\\ &\downarrow\\ &\Cl(V) \to \Cl (X\times\AA^1)/\ZZ\lbrack Z\rbrack \to \Cl(X\times\AA^1) \to \Cl(X), \end{align*}
where each step is either a restriction, identity or induced by the isomorphism between $X$ and the general fiber.
\end{proof}
Although we allowed our limit divisor $D_\P$ to be a $\QQ$-Weil divisor, we will now prove that this divisor will in fact be integral if the Weil divisor on our original variety was Cartier. This in return restricts the polytopes that could appear in our setting. We will try to explain these restrictions at the end of this section.
\begin{theorem*}[\cref{thm:dpweil}]
Let $X$ be a normal projective complex variety of dimension $d$ and let $D$ be an ample Cartier divisor on $X$. Let $X_\P$ be the normal projective toric variety associated to the rational convex polytope $\P\subseteq\RR^d$. Suppose the pair $(X, D)$ admits a toric degeneration to the toric pair $(X_\P, D_\P)$. Then $D_\P$ is a Weil divisor.
\end{theorem*}
\begin{proof}
Notice that $\F^{(n)}$ is divisorial for every $n\in\NN$, so there exist divisors $\D_n$ on the degeneration space $\X$, $n\in\NN$, such that $\F^{(n)}\simeq\O_\X(\D_n)$. Fix $t\neq0$ and $n\in\NN$. Since $D$ is Cartier, we know that
\[\F^{(n)}|_{\X_t} \simeq \O_X(nD) \simeq \O_X(D)^{\otimes n} \simeq (\F^{(1)})^n|_{\X_t}.\]
Translated to divisors this means
\[ \D_n|_{\X_t} \sim n\D_1|_{\X_t}.\]
By \cref{prop:clx} this is only possible if $\D_n \sim n\D_1$.
So there exists a Weil divisor $\D := \D_1$ on $\X$ such that $\F^{(n)} \simeq \O_\X(n\D)$ for every $n\in\NN$. Let $E$ denote the divisor on $X_\P$ that is isomorphic to $\D|_{\X_0}$ via the isomorphism $\X_0\simeq X_\P$.
Since $\F^{(n)}|_{\X_0}\simeq\O_{X_\P}(\lfloor nD_\P \rfloor)$ for every $n\in\NN$, we have
\[ \lfloor nD_\P \rfloor \sim nE \]
for every $n\in\NN$. Chose $l\in\NN$ such that $lD_\P$ is Weil. Then
\[ lD_\P \sim lE\]
i.e. there exist rational functions $f,g\in\CC(X_\P)^\times$ such that
\[ \lfloor D_\P \rfloor = E+\div(f) \hspace{10pt}\text{ and }\hspace{10pt} lD_\P = lE + \div(g). \]
We want to show that $g = \lambda f^l$ for some $\lambda\in\CC$. Notice that
\[ D_\P = E + l^{-1}\div(g), \]
hence \[E+\div(f) = \lfloor D_\P\rfloor = \lfloor E+l^{-1}\div(g) \rfloor = E+\lfloor l^{-1}\div(g)\rfloor.\]
So in conclusion we know that $\div(f) = \lfloor l^{-1}\div(g)\rfloor$. Since every coefficient of $\lfloor l^{-1}\div(g)\rfloor $ is smaller than or equal to the corresponding coefficient of $l^{-1}\div(g)$, we see that $l^{-1}\div(g)-\div(f)$ is an effective divisor. Equivalently, the difference $\div(g) - l\div(f) = \div(f^{-l}g)$ is an effective divisor.
But this implies that $f^{-l}g$ is a rational function on $X_\P$ whose valuation $v_{D_i}(f^{-l}g)$ on every prime divisor $D_i$ is non-negative. This is only possible if $f^{-l}g$ is regular. But since $X_\P$ is projective, every regular function is constant, so there exist $\lambda \in\CC^\times$ such that $f^{-l}g=\lambda$\,---\,or equivalently $g = \lambda f^l$.
In conclusion we see that $\div(g)=\div(\lambda f^l)=\div(f^l)$ and thus
\[D_\P = E+l^{-1}\div(g) = E+l^{-1}\div(f^l) = E+\div(f)\]
is a Weil divisor on $X_\P$.
\end{proof}
Let us now prove an important consequence of \cref{thm:toricehrhart}.
\begin{theorem}\label{thm:twopolynomials}
Let $X$ be a normal projective complex Gorenstein variety of dimension $d$ and let $\L$ be an ample line bundle over $X$. Let $\P \subseteq\RR^d$ be a full-dimensional rational convex polytope. Suppose that the pair $(X, \L)$ admits a toric degeneration to the toric variety $(X_\P, D_\P)$. Then
\[ \chi(X, \L^n) = \#(n\P\cap \ZZ^d) \hspace{10pt}\text{ and }\hspace{10pt} \chi(X, \L^n\otimes\omega_X) = \#(\interior n\P \cap \ZZ^d)\]
for all $n \in \NN$.
\end{theorem}
\begin{proof}
By \cref{thm:dpweil} we know that $D_\P$ is a Weil divisor on $X_\P$.
By our assumption there exist divisorial sheaves $\F^{(n)}$ for every $n\in\NN$ such that
\[ \F^{(n)}|_{\X_t} \simeq \L^n \text{ for all } t\neq0 \text{ and } \F^{(n)}|_{\X_0} \simeq \O_{X_\P}(nD_\P).\]
By \cref{thm:toricehrhart} we have
\[ \#(n\P\cap\ZZ^d) = \chi(X_\P, \O_{X_\P}(nD_\P)).\]
Since Euler characteristic is constant in flat projective families (see for example \cite[Theorem 24.7.1]{V}), we conclude
\[ \#(n\P\cap\ZZ^d) = \chi(X_\P, \O_{X_\P}( nD_\P)) = \chi(X, \L^n)\]
for every $n\in\NN$.
For the second set of equations let $\D$ denote the Weil divisor from the proof of \cref{thm:dpweil}, i.e. $\F^{(n)} = \O_\X(n\D)$ for all $n\in\NN$. Consider the divisorial sheaves $\O_\X(n\D+K_\X)$ for each $n\in\NN$. Recall that each fiber $\X_t$ is a principle prime divisor on $\X$ by \cref{prop:x0}. So the Adjunction Formula (see for example \cite[30.4.8]{V}) yields
\[ \O_\X(n\D+K_\X)|_{\X_t}\simeq \O_{\X_t}((n\D+K_\X+\X_t)|_{\X_t}) \simeq \begin{cases}
\O_{X_\P}(nD_\P+K_{X_\P})\text{ if } t=0,\\
\L^n\otimes\omega_X \text{ else}
\end{cases} \]
since $K_X$ is Cartier. The claim follows from the same reasons as before.
\end{proof}
There exists a different formulation of this theorem.
\begin{theorem*}[\cref{cor:twopolynomials}]
Let $X$ be a normal projective complex Gorenstein variety of dimension $d$ and let $\L$ be an ample line bundle over $X$. Let $\P \subseteq\RR^d$ be a full-dimensional rational convex polytope. Suppose that the pair $(X, \L)$ admits a toric degeneration to the toric variety $(X_\P, D_\P)$. Then the Ehrhart quasi-polynomial $L_\P$ of $\P$ coincides with the Hilbert polynomial $P_\L(n) := \chi(X, \L^n)$ of $X$ and $\L$, i.e.
\[ L_\P(n) = P_\L(n) \]
for all $n \in \ZZ$.
\end{theorem*}
\begin{proof}
The claim is true for positive $n$ by \cref{thm:twopolynomials}. Since $L_\P$ is a quasi-polynomial that coincides with the polynomial $P_\L$ on all positive integers, it must be a polynomial itself. Since both polynomials $L_\P$ and $P_\L$ coincide on all positive integers, they must coincide on all integers.
\end{proof}
We want to save one important observation in this proof for later purposes. It does not even need the Gorenstein hypothesis.
\begin{cor}\label{cor:degenerationpolynomial}
Let $\P$ be a rational convex polytope. Then the associated pair $(X_\P, D_\P)$ can only be the limit of a normal projective variety and an ample line bundle under a toric degeneration if the Ehrhart quasi-polynomial of $\P$ is a polynomial.
\end{cor}
We will return to this observation at the end of this section. Another consequence is the following.
\begin{rem}\label{rem:emrserre}
It is already known (see for example \cite[11.12.4]{D}) that Serre Duality for toric varieties implies Ehrhart-Macdonald Reciprocity for lattice polytopes (notice that Cartier divisors on toric varieties correspond to lattice polytopes by \cite[Theorem 4.2.8]{CLS}). Our formulation in \cref{cor:twopolynomials} generalizes this in two regards.
On one side, we could prove Ehrhart-Macdonald Reciprocity for arbitrary rational convex polytopes by using toric Serre Duality (see for example \cite[Exercise 9.3.5]{CLS} for a rational version) since \cref{thm:toricehrhart} implies
\begin{align*}
L_\P(-n) &= \chi(X_\P, \O_{X_\P}(\lfloor -nD_\P \rfloor)) = \chi(X_\P, \O_{X_\P}(-\lceil nD_\P \rceil)) \\&= (-1)^d\chi(X_\P, \O_{X_\P}(\lceil nD_\P \rceil+K_{X_\P})) = (-1)^dL_{\interior\P}(n).
\end{align*}
Here we additionally used that the Ehrhart quasi-polynomial of $\P$ and the Hilbert quasi-polynomial of $D_\P$ must coincide on all integers since they coincide on all positive integers.
But most importantly, the reverse argument is also possible! Whenever we have a normal projective Gorenstein variety $(X, \L)$ that admits a toric degeneration to any (normal projective) toric pair $(X_\P, D_\P)$, we get
\[ \chi(X, \L^{-n}) = L_\P(-n) = (-1)^d L_{\interior\P}(n) = (-1)^d \chi(X, \L^n\otimes\omega_X).\]
This remarkable fact shows that Serre Duality can be proved by counting lattice points for a broad class of varieties. This fact is depicted in \cref{fig:emrserre}.
\end{rem}
\begin{figure}[htb]\centering
\caption{Let $X$ be a normal projective complex Gorenstein variety, $\L$ an ample line bundle over $X$ and $\P$ a rational convex polytope such that $(X,\L)$ admits a toric degeneration to $(X_\P, D_\P)$. Then the sketched equalities hold for every $n\in\NN$.}\label{fig:emrserre}
\begin{tikzpicture}
\matrix (m) [matrix of math nodes,row sep=3em,column sep=7em,minimum width=2em]
{
\chi(X, \L^{-n}) & (-1)^d\chi(X, \L^n\otimes\omega_X) \\
L_\P(-n) & (-1)^dL_{\interior \P}(n) \\};
\draw[double distance = 3pt] (m-1-1) -- node[above] {\scriptsize Serre} node[below] {\scriptsize Duality} (m-1-2);
\draw[double distance = 3pt] (m-1-2) -- (m-2-2);
\draw[double distance = 3pt] (m-2-1) -- node[above] {\scriptsize Ehrhart-Macdonald} node[below] {\scriptsize Reciprocity} (m-2-2);
\draw[double distance = 3pt] (m-1-1) -- (m-2-1);
\end{tikzpicture}
\end{figure}
\section{A Unique Combinatorial Feature of the Anticanonical Bundle}\label{sec:vanishing}
This section is dedicated to proving the following central observation.
\begin{theorem*}[\cref{thm:vanishing}]
Let $X$ be a Gorenstein Fano variety of dimension $d$ that has rational singularities and $\L$ an ample line bundle. Then the line bundle $\L$ is isomorphic to the anticanonical line bundle $\omega_X^{-1}$ if and only if the Hilbert polynomial of $\L$ fulfills
\[ P_\L(n) = (-1)^d P_\L(-n-1)\]
for all $n\in\NN$.
\end{theorem*}
\begin{proof}
Notice first that Serre Duality implies that
\[ P_\L(-n-1) = \chi(X, \L^{-n-1}) = (-1)^d \chi(X, \L^{n+1}\otimes\omega_X).\]
So the first implication is obvious, since
\[ P_{\omega_X^{-1}}(-n-1) = (-1)^d \chi(X, \omega_X^{-n-1}\otimes \omega_X) =(-1)^d\chi(X, \omega_X^{-n}) = (-1)^d P_{\omega_X^{-1}}(n)\]
for every $n\in\NN$.
For the other implication let us notice that Kodaira Vanishing for Rational Singularities (see \cite[Theorem 2.70]{KM}, which can be proved as a consequence of the Grauert-Riemenschneider Vanishing Theorem in \cite[Satz 2.3]{GR}) implies that
\[\chi(X, \L^{n+1}\otimes \omega_X) = h^0(X, \L^{n+1}\otimes \omega_X)\]
for every $n\in\NN$. So for the special case $n = 0$ the assumption on the Hilbert polynomial implies that
\begin{align*}h^0(X, \L\otimes\omega_X) &= \chi(X, \L\otimes\omega_X) = (-1)^d P_\L(-1) \\ &=P_\L(0) = \chi(X, \O_X) = h^0(X, \O_X) = 1.\end{align*}
Chose a non-zero section $s \in H^0(X, \L\otimes\omega_X)$ and consider the natural morphism $f\colon \O_X \to \L\otimes\omega_X$ given by
\[f(U) \colon \O_X(U) \to (\L\otimes\omega_X)(U), \hspace{10pt} \xi \mapsto \xi \cdot s|_U,\]
on every open set $U \subseteq X$.
Since $\L\otimes\omega_X$ is a line-bundle, it is torsion-free. So the maps $f(U)$ and hence the morphism $f$ are injective. Thus we get a short exact sequence
\[ 0 \to \O_X \stackrel{f}{\to} \L\otimes\omega_X \to \C \to 0 \]
for some coherent sheaf $\mathcal{C}$. We want to show that $\C\simeq 0$.
Recall that Euler characteristic is additive on short exact sequences (see for example \cite[Lemma 2.5.2]{EGAIII1}). Hence tensoring with $\L^n$ and taking Euler characteristic yields
\[\chi(X, \L^n\otimes\C) = \chi(X, \L^{n+1}\otimes\omega_X) - \chi(X, \L^n) = (-1)^dP_\L(-n-1) - P_\L(n) = 0 \]
for every $n \in \NN$.
By the Serre Vanishing Theorem (see for example \cite[Proposition 2.2.2]{EGAIII1}) we can chose $n$ big enough such that the sheaf $\mathcal{C}\otimes\L^n$ is globally generated and its higher cohomologies vanish. Hence
\[ h^0(X, \mathcal{C}\otimes\L^n) = \chi(X, \mathcal{C}\otimes\L^n) = 0. \]
Since $\mathcal{C}\otimes\L^n$ is globally generated, this is only possible if $\mathcal{C}\otimes\L^n \simeq 0$, i.e. $\mathcal{C} \simeq 0$.
In conclusion we see that $\L\otimes\omega_X \simeq \O_X$ holds\,---\,or equivalently $\L \simeq \omega_X^{-1}$, which proves the claim.
\end{proof}
\section{Anticanonically Polarized Toric Degenerations}\label{sec:main}
We are now able to prove our main result. A sketch of the proof is presented in \cref{fig:proof}.
\begin{proof}[Proof of \cref{thm:main}]
Let $(X, \L)$ be a polarized Gorenstein Fano variety and suppose that it admits a toric degeneration to the pair $(X_\P, D_\P)$, where $\P\subseteq\RR^d$ denotes a full-dimensional rational convex polytope, $X_\P$ the associated toric variety and $D_\P$ the associated torus invariant ample $\QQ$-Cartier Weil divisor on $X_\P$ (see \cref{prop:dpample,thm:dpweil}).
Let us first assume that $\L\simeq\omega_X^{-1}$. Notice that the isomorphism from \cref{prop:clx} yields $\F^{(1)} \simeq \O_\X(-K_\X)$ and recall that $\X_0$ is a principal prime divisor on $\X$ by \cref{prop:x0}. So the Adjunction Formula (see for example \cite[Proposition 30.4.8]{V}) shows that
\[ K_{\X_0} = (K_\X+\X_0)|_{\X_0} \sim K_\X|_{\X_0}\]
which implies
\[ \O_{X_\P}(D_\P) \simeq \F^{(1)}|_{\X_0} \simeq \O_{\X_0} (-K_{\X_0}) \]
and thus $D_\P\sim-K_{X_\P}$.
Additionally, since $-K_{X_\P} \sim D_\P$ is $\QQ$-Cartier and ample, the toric limit variety $X_\P$ is $\QQ$-Gorenstein Fano.
For the reverse implication let us assume that $D_\P\sim -K_{X_\P}$ and hence $X_\P$ is $\QQ$-Gorenstein Fano. By \cref{prop:toricdivisors} we know that
\[ \P = \P_{D_\P} = \P_{-K_{X_\P}} + s\]
for some lattice vector $s\in\ZZ^d$.
Since the toric variety $X_\P$ is $\QQ$-Gorenstein Fano, \cref{thm:toricfano} (following Nill \cite[Proposition 1.4]{N} and Batyrev \cite[Proposition 2.2.23]{B}) shows that the dual of the polytope $\P_{-K_{X_\P}}$ is Fano. Hence $\P$ is dual-integral (it would even be dual-Fano), so Hibi's \cref{thm:hibi} implies that the Ehrhart polynomial $L_\P$ of $\P$ fulfills
\[ L_\P(n) = (-1)^dL_{\P}(-n-1)\]
for all $n\in\NN$.
We have seen in \cref{cor:twopolynomials} that this Ehrhart quasi-polynomial and the Hilbert polynomial $P_\L$ of $X$ and $\L$ coincide, hence
\[ P_\L(n) = (-1)^dP_\L(-n-1) \]
for every $n\in\NN$. Since $X_\P$ has rational singularities (see for example \cite[Theorem 11.4.2]{CLS}), Elkik's result \cite[Théorème 4]{E} proves that $X$ has rational singularities as well. So we can apply \cref{thm:vanishing}, which implies that $\L$ is isomorphic to $\omega_X^{-1}$. This concludes the proof.
\end{proof}
\begin{figure}[htb]
\centering
\caption{Sketch of the proof of \cref{thm:main}. Notation as in the formulation of its proof. The equations are supposed to hold for every $n\in\NN$.}\label{fig:proof}
\begin{tikzpicture}[>=stealth,>=implies]
\matrix (m) [matrix of nodes,align=center,row sep=3em,column sep=4em,minimum width=2em]
{
$\L\simeq\omega_X^{-1}$ & $D_\P \sim -K_{X_\P}$ \\
& $\P$ is dual-Fano \\
& $\P$ is dual-integral\\
&$\#(n\P\cap\ZZ^d) = \#(\interior(n+1)\P \cap \ZZ^d)$\\
$P_\L(n) = (-1)^d P_\L(-n-1)$ &$L_\P(n) = (-1)^d L_\P(-n-1)$\\
$\chi(X, \L^n) = (-1)^d\chi(X, \L^{-n-1})$ & {\scriptsize\begin{tabular}{c}$\chi(X_\P, \O_{X_\P}(\lfloor nD_\P\rfloor)) = $\\$(-1)^d \chi(X_\P, \O_{X_\P}(\lfloor -(n+1)D_\P\rfloor))$\end{tabular}}\\};
\draw[->,double distance = 3pt] (m-1-1) -- node[above] {\scriptsize\begin{tabular}{c}Alexeev--Brion\end{tabular}} node[below] {\scriptsize\begin{tabular}{c}\cite[Theorem 3.8]{AB}\end{tabular}}(m-1-2);
\draw[<->,double distance = 3pt] (m-1-1) -- node[left,align=right] {\scriptsize \begin{tabular}{r}Serre Duality\\\\ Elkik's Theorem\\\\Kodaira Vanishing for \\Rational Singularities\\\\Serre Vanishing\end{tabular}} node[right] {\scriptsize \begin{tabular}{l}\\\\\\\cref{thm:vanishing}\\\\\\\\\end{tabular}}(m-5-1);
\draw[<->,double distance = 3pt] (m-5-1) -- (m-6-1);
\draw[<->,double distance = 3pt] (m-6-1) -- node[above,align=center]{\scriptsize\begin{tabular}{c}Invariance of \\Euler Characteristic\end{tabular}} node[below] {\scriptsize\begin{tabular}{c}\cref{thm:twopolynomials}\end{tabular}} ([yshift=1.25pt]m-6-2.west);
\draw[<->,double distance = 3pt] (m-5-1) -- node[below] {\scriptsize\begin{tabular}{c}\cref{cor:twopolynomials}\end{tabular}} (m-5-2);
\draw[<->,double distance = 3pt] (m-5-2) -- node[left] {\scriptsize\begin{tabular}{r}Ehrhart-Macdonald\\Reciprocity\end{tabular}} node[right] {\scriptsize\begin{tabular}{l}\cref{thm:emr}\end{tabular}} (m-4-2);
\draw[<->,double distance = 3pt] (m-6-2) -- node[right] {\scriptsize\begin{tabular}{l}\cref{thm:toricehrhart}\end{tabular}} (m-5-2);
\draw[<->,double distance = 3pt] (m-4-2) -- node[left] {\scriptsize\begin{tabular}{r}Hibi\end{tabular}} node[right] {\scriptsize\begin{tabular}{l}\cref{thm:hibi}\end{tabular}}(m-3-2);
\draw[<-,double distance = 3pt] (m-3-2) -- node[right] {\scriptsize\begin{tabular}{l}\cref{thm:dualfano}\end{tabular}} (m-2-2);
\draw[<->,double distance = 3pt] (m-2-2) -- node[left] {\scriptsize\begin{tabular}{r}Batyrev, Nill\end{tabular}} node[right] {\scriptsize\begin{tabular}{l}\cref{cor:dualfano}\end{tabular}}(m-1-2);
\end{tikzpicture}
\end{figure}
The following consequence follows directly. It might prove useful in the hunt for mirror symmetry since it yields a necessary condition for reflexive polytopes to appear in the setting of toric degenerations.
\begin{theorem*}[\cref{thm:reflexive}]
Let $X$ be a Gorenstein Fano variety and $\L$ an ample line bundle over $X$. Suppose that $(X, \L)$ admits a toric degeneration to the pair $(X_\P, D_\P)$ associated to the rational convex polytope $\P$. Then $\P$ is quasi-reflexive if and only if $\P$ is a lattice polytope and $\L$ is isomorphic to the anticanonical bundle over $X$.
\end{theorem*}
There is another, more delicate consequence. To find it, let us review the equivalences and implications of our proof schematically in \cref{fig:proof}. Essentially, we have proved that every dual-integral polytope appearing in the context of toric degenerations of polarized Gorenstein Fano varieties must already be a dual-Fano polytope. So the polytope must {\em know} that it is the limit of a polarized variety under a toric degeneration. We already noticed in \cref{thm:dpweil} that not every polytope can appear as the polytope associated to the toric limit divisor since that divisor must be Weil.
Philosophically speaking, this information should be contained in the combinatorics of the polytope. In fact, there is one on-the-fly result we obtained but never used in later arguments. In \cref{cor:degenerationpolynomial} we showed that every polytope whose associated toric pair is the limit of a polarized Gorenstein Fano variety must be a quasi-lattice polytope, i.e. its Ehrhart quasi-polynomial must be a polynomial.
By this reason\,---\,and from examples\,---\,we reach the following conjecture.
\begin{conj}\label{conj:dualfano}
A convex polytope is dual-Fano if and only if it is a dual-integral quasi-lattice polytope.
\end{conj}
In other words this conjecture can be formulated as follows (by \cref{cor:dualfano}).
\begin{conj}
Let $\P$ be a full-dimensional rational convex polytope and suppose for every ray $\rho$ in the normal fan $\Sigma_\P$ of $\P$ there exist primitive lattice vectors $u_\rho$ and a strictly positive integer $k_\rho > 0$ such that
\[ \P = \{ x \in \RR^d \mid \langle x, u_\rho \rangle \leq k_\rho^{-1} \text{ for all } \rho\in\Sigma_\P(1)\}.\]
Then the Ehrhart quasi-polynomial $L_\P$ of $\P$ is a polynomial if and only if all integers $k_\rho$, $\rho\in\Sigma_\P(1)$, are equal to $1$.
\end{conj}
This conjecture would immediately follow from the following one.
\begin{conj*}[\cref{conj:quasilattice}]
The divisor associated to a full-dimensional rational convex polytope is a Weil divisor on the toric variety associated to the polytope if and only if the polytope is a quasi-lattice polytope.
\end{conj*}
Again, we can give a more combinatorial formulation of this conjecture.
\begin{conj}
Let $\P$ be a full-dimensional rational convex polytope such that $0\in\interior\P$. Then the Ehrhart quasi-polynomial $L_\P$ of $\P$ is a polynomial if and only if there exist strictly positive integers $m_\rho > 0$ for every ray $\rho\in\Sigma_\P(1)$ such that
\[ \P = \{ x \in \RR^d \mid \langle x, u_\rho \rangle \leq m_\rho \text{ for all } \rho\in\Sigma_\P(1)\}.\]
Here $u_\rho$ denotes the primitive ray generator of the ray $\rho\in\Sigma_\P(1)$.
\end{conj}
\section{Recap: The Flag Variety Case}\label{sec:flag}
We will now briefly recall the special case of flag varieties. In the Main Theorem of \cite{S1} we essentially proved the following.
\begin{theorem}\label{thm:nobody}
Let $G$ be a simple complex algebraic group, $P\subseteq G$ a parabolic and $\lambda$ a $P$-regular dominant integral weight. Let $\Delta(\lambda)$ be a Newton-Okounkov body associated to a full-rank valuation on $\bigoplus_{n\in\NN}H^0(G/P, \L_{n\lambda})$ whose valuation semigroup is finitely generated and saturated. Then the following are equivalent.
\begin{enumerate}
\item The weight $\lambda$ is the weight of the anticanonical line bundle over $G/P$.
\item The Newton-Okounkov body $\Delta(\lambda)$ contains exactly one interior lattice point.
\item The Newton-Okounkov body $\Delta(\lambda)$ is dual-integral.
\end{enumerate}
\end{theorem}
However, we really would not have needed to restrict ourselves to Newton-Okounkov bodies. By \cref{rem:birational} these are just special cases of polytopes coming from toric degenerations of $(G/P, \L_\lambda)$. Hence the equivalence of claims (i) and (iii) in \cref{thm:nobody} is just a consequence of \cref{thm:main}. By \cref{cor:hibi} claim (iii) implies claim (ii). But the truly remarkable fact about flag varieties is that (ii) already implies (i) and (iii).
The reason for this is that we have very precise control over the Ehrhart polynomial $L_{\Delta(\lambda)}$ because it is just given by Borel-Weil-Bott and Weyl's Dimension Formula. As stated in \cite[Lemma 3.1]{S1}, one can compute
\[ L_{\Delta(\lambda)}(n) = \prod_{\beta\in\Phi_P^+}\frac{\langle n\lambda+\rho, \beta^\vee\rangle}{\langle \rho, \beta^\vee\rangle} \]
where $\Phi_P^+$ denotes some subset of positive roots given by the choice of parabolic. Via some basic root combinatorics it is possible to verify that this polynomial fulfills Hibi's criterion (see \cref{rem:hibi}) if and only if it evaluates to $1$ at $n=-1$.
So the question remains whether a given partial flag variety admits a flat projective degeneration to a Gorenstein Fano toric variety (see \cref{question:simplified}). The answer is the following theorem that is essentially \cite[Theorem 6]{S3}.
\begin{theorem*}[\cref{thm:flag}]
Let $G$ be a complex classical group or $\GG$. Then for any parabolic $P\subseteq G$ the partial flag variety $(G/P, \omega_{G/P}^{-1})$ admits a flat projective degeneration to an anticanonically polarized Gorenstein Fano toric variety.
\end{theorem*}
\begin{proof}
By \cref{thm:main} any toric limit of $(G/P, \omega_{G/P}^{-1})$ will automatically be $\QQ$-Gorenstein Fano and anticanonically polarized. So we just have to show that there exists a toric limit $(X_\P, D_\P)$ with $\P$ being a lattice (hence reflexive) polytope because this means that $D_\P \sim -K_{X_\P}$ is Cartier.
If $G$ is a complex classical group, we have shown in \cite[Theorem 5]{S3} that for every polarized partial flag variety $(G/P, \L_\lambda)$ there exists a string polytope that will be a lattice polytope if $\L_\lambda\simeq\omega_{G/P}^{-1}$. From \cite[Theorem 3.2]{AB} (or equivalently \cite[Theorem 1]{Ka} and \cite[Theorem 1, Corollary 2 and Lemma 3]{A}) it is known that $(G/P, \L_{\omega_{G/P}^{-1}})$ admits a toric degeneration (in our sense) to the toric variety associated to said string polytope, which must therefore be Gorenstein Fano.
If $G$ is of type $\GG$ we can chose the polytope defined by Gornitskii in \cite{G}. One can easily calculate that it is always a lattice polytope. Additionally, by \cite[Section 8.2]{FaFL} and \cite[Theorem 6]{FaFL} it is clear that each polarized partial flag variety admits a toric degeneration to the polarized toric variety associated to the Gornitskii polytope, which must therefore be Gorenstein Fano.
\end{proof}
It would be very interesting to see whether \cref{thm:flag} also holds true for the exceptional Lie groups of type $\EE$, $\EEE$, $\EEEE$ and $\FF$. Additionally, it would be nice to see a class of polytopes that can be used for every type.
|
\section{DAISY-CHAIN IMPLEMENTATION IN FPGA}
\label{sec:FPGA}
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.45\textwidth]{Figures/BlockDaisyChain.eps}
\end{center}
\vspace{-5mm}
\caption{Block diagram of new Network Processor.}
\label{fig:blockdaisychain}
\end{figure}
We improved the \textit{Network Processor} and developed the daisy-chain function.
Another SiTCP and SFP I/F were added in the network processor.
Data Path controllers for the daisy-chain function were also implemented around them.
Fig.~\ref{fig:blockdaisychain} shows a block diagram of the new \textit{Network Processor}. This block consists of \textit{Path Controller}, \textit{Data Carrier}, and two SFP I/Fs (SFP I/F0 and SFP I/F1), along with the two SiTCPs (SiTCP0 and SiTCP1).
The \textit{Path Controller} manages the Ethernet frame path.
The \textit{Data Carrier} controls the event data path.
SiTCP0 exchanges Ethernet frames with the previous ROESTI. SiTCP1 exchanges Ethernet frames with the next ROESTI or the DAQ PC.
\subsection{Path Controller}
\label{sec:pathcont}
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.45\textwidth]{Figures/BlockPathCont.eps}
\end{center}
\vspace{-5mm}
\caption{Block diagram of Path Controller.}
\label{fig:blockpathcont}
\end{figure}
The \textit{Path Controller} has the following two functions.
The first one is a receive function. Each ROESTI receives an Ethernet frame from the neighboring ROESTI. When the \textit{path controller} receives the Ethernet frame, the \textit{Path Controller} verifies the destination MAC address of the Ethernet frame and decides whether to send the Ethernet frame to own SiTCP or the other neighboring ROESTI. The second one is the send function. The \textit{Path Controller} receives the Ethernet frame from own SiTCP or the neighboring ROESTI.
The \textit{Path Controller} arbitrates between two Ethernet frames and sends the Ethernet frame to the neighboring ROESTI or the DAQ PC. The Ethernet frames are synchronized with 125~MHz clock.
The \textit{Path Controller} consists of two \textit{Selector}s (\textit{Selector}0 and \textit{Selector}1) and two \textit{Arbiter}s (\textit{Arbiter}0 and \textit{Arbiter}1) as shown in Fig.~\ref{fig:blockpathcont}.
Each \textit{Selector} receives the Ethernet frame and verifies the destination MAC address thereof. If the destination MAC address matches the MAC address of the SiTCP, the \textit{Selector} sends the data to the SiTCP. If these MAC addresses are different, the \textit{Selector} sends the data to an \textit{Arbiter}. Alternatively, if the destination MAC address matches the broadcast address (for Ethernet, FF:FF:FF:FF:FF:FF), the \textit{Selector} sends the data to both the SiTCP and the \textit{Arbiter}.
The \textit{Arbiter}0 can receive the Ethernet frame only from the \textit{Selector}1 or the SiTCP0. On the other hand, The \textit{Arbiter}1 can receive the Ethernet frame only from the \textit{Selector}0 or the SiTCP1.
When each \textit{Arbiter} receives the Ethernet frame, the \textit{Path Controller} starts transferring the Ethernet frame to the SFP I/F. If the \textit{Arbiter} receives other data during this process, the two Ethernet frames collide. In such cases, the \textit{Arbiter} discards the latter Ethernet frame.
For example, if the \textit{Arbiter}1 receives an Ethernet frame from the SiTCP1 when the \textit{Arbiter}1 does not send any data, the \textit{Arbiter}1 starts to transfer the Ethernet frame to SFP I/F1. During the transfer of the Ethernet frame, if the \textit{Arbiter}1 receives the other Ethernet frame from the \textit{Selector}0, the \textit{Arbiter}1 discards the Ethernet frame from the \textit{Selector}0.
\subsection{Data Carrier}
\label{sec:datacarrier}
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.45\textwidth]{Figures/BlockDataCarrier.eps}
\end{center}
\vspace{-5mm}
\caption{Block diagram of Data Carrier.}
\label{fig:blockdatacarieer}
\end{figure}
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.33\textwidth]{Figures/statemachine.eps}
\end{center}
\vspace{-5mm}
\caption{State transition diagram for the TCP Arbiter of the Data Carrier.}
\label{fig:State}
\end{figure}
In the \textit{Data Carrier}, the event data path is controlled. The \textit{Data Carrier} receives the event data from the \text{Data I/F}. The \textit{Data Carrier} also receives the event data for the previous ROESTI from SiTCP0. In order to avoid Ethernet frames collision in the Path Controller, Ethernet frames of TCP of previous ROESTI are sent to the Data Carrier. The \textit{Data Carrier} arbitrates between two event data and sends the event data to SiTCP1. The event data is synchronized with 133~MHz clock.
Fig.~\ref{fig:blockdatacarieer} shows the \textit{Data Carrier} block. The \textit{Data Carrier} consists of ring buffer, FIFO, and \textit{TCP Arbiter}.
The ring buffer stores the event data from the \text{Data I/F}. The FIFO stores the event data from SiTCP0. The ring buffer size is 64~bit $\times$ 4096~bit. The FIFO size is 8~bit $\times$ 65536~bit.
The \textit{TCP Arbiter} extracts the event data from the ring buffer or the FIFO and sends the event data to SiTCP1.
The \textit{TCP Arbiter} has three states: SUSPENSION, MYROESTI, and NEIGHBOR as shown in Fig.~\ref{fig:State}. When the \textit{TCP Arbiter} receives the reset signal, the \textit{TCP Arbiter} moves to the SUSPENSION state. If the ring buffer has the event data for one or more events and the FIFO does not have any event data, the state changes to MYROESTI (Transition to MYROESTI). In the opposite case, the state changes to NEIGHBOR (Transition to NEIGHBOR). When both the ring buffer and the FIFO have the event data for one or more events, the state is determined by the event number of the event data in the ring buffer and the FIFO. If the event number of the event data in the ring buffer is less than that in the FIFO, the \textit{TCP Arbiter} state changes to MYROESTI (Transition to MYROESTI); otherwise, it changes to NEIGHBOR (Transition to NEIGHBOR). When the state changes to MYROESTI, the \textit{TCP Arbiter} extracts one event data from the ring buffer and sends it to SiTCP1. In contrast, when the state changes to NEIGHBOR, the \textit{TCP Arbiter} extracts one event data from the FIFO and sends it to SiTCP1. Upon transmitting the event data, the state returns to SUSPENSION (One Event Finish).
\subsection{Slow Control over UDP/IP}
\label{sec:slowcontrol}
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.44\textwidth]{Figures/egUDP2.eps}
\end{center}
\vspace{-5mm}
\caption{Example of slow control over UDP/IP. (a) Ethernet frame passes through the ROESTI1 because of MAC address mismatch. (b) Ethernet frame also passes through the ROESTI2 because of MAC address mismatch. (c) Ethernet frame is received in the ROESTI3 because of MAC address match. The slow control signal is sent to the \textit{Module Control} of the ROESTI3. }
\label{fig:egUDP}
\end{figure}
As mentioned in Section~\ref{sec:ROESTI}, the DAQ PC must send the slow control signal to SiTCP in the slow control process. After processing the slow control, SiTCP sends a slow control signal to the DAQ PC. For daisy-chain communication, the same process takes place between the DAQ PC and the target ROESTI.
Fig.~\ref{fig:egUDP} shows an example of the slow control process over UDP/IP. The DAQ PC and three ROESTIs (ROESTI1, ROESTI2, ROESTI3) are connected with the daisy-chain. The ROESTI3 and the DAQ PC communicate for the slow control over UDP/IP. The DAQ PC sends the Ethernet frame such that the destination MAC address matches the MAC address of the SiTCP1 in the ROESTI3.
The SiTCP1 of each ROESTI has a different MAC address. Because the MAC address of the SiTCP1 of the ROESIT1 does not match the destination MAC address of the Ethernet frame, the Ethernet frame simply passes through the \textit{Path Controller} of the ROESTI1 and is sent to the ROESTI2 (Fig.~\ref{fig:egUDP} (a)). For the same reason, the Ethernet frame passes through the ROESTI2 and is sent to the ROESTI3 (Fig.~\ref{fig:egUDP} (b)).
Subsequently, the \textit{Path Controller} of the ROESTI3 identifies the Ethernet frame and sends it to the SiTCP1 of the ROESTI3 (Fig.~\ref{fig:egUDP} (c)). The Ethernet frame includes the slow control signal. The slow control signal is sent to the \textit{Module Control} of the ROESTI3.
After that, the ROESTI3 sends the Ethernet frame to the DAQ PC. The destination MAC address corresponds with the MAC address of the DAQ PC. The Ethernet frame passes through the ROESTI1 and the ROESTI2 again before reaching its destination.
The slow control process does not change even if the number of ROESTIs increases.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.5\textwidth]{Figures/datacollision.eps}
\end{center}
\vspace{-5mm}
\caption{Data collision in the Arbiter1 of the Path Controller. }
\label{fig:data collision}
\end{figure}
In the \textit{Arbiter} of the \textit{Path Controller}, if the timing of two Ethernet frame matches as shown in Fig.~\ref{fig:data collision}, they collide with each other. Either of the Ethernet frames disappears as mentioned in Section~\ref{sec:pathcont}. However, this is not a major problem in the case of the slow control over UDP/IP because the frequency of slow control is small enough compared to the bandwidth of Gigabit Ethernet and slow control usually process when ROESTI does not transfer event data with TCP.
If the slow control process when ROESTI transfers event data with TCP, It is possible to disappear Ethernet frame of slow control. However, the DAQ PC finds this anomaly and the DAQ PC resends the Ethernet frame for slow control to the target ROESTI until slow control succeed. This is not a problem because the slow control process does not need instantaneousness.
\subsection{Data Transfer over TCP/IP}
\label{sec:datatransfer}
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.5\textwidth]{Figures/egTCP1.eps}
\end{center}
\vspace{-5mm}
\caption{Example 1 of transferring event data over TCP/IP. (a) Two ROESTIs receives the trigger signal and the event data is generated. (b) Two Ethernet frames collide in the Path Controller of the Network Processor. }
\label{fig:egTCP1}
\end{figure}
Unlike the slow control process over UDP/IP, the ROESTIs and DAQ PC cannot communicate directly in the event data transfers over TCP.
Fig.~\ref{fig:egTCP1} shows an example of the event data transfer process over TCP/IP. The DAQ PC and two ROESTIs (ROESTI1, ROESTI2) are connected with the daisy-chain. The ROESTI2 attempts to send the Ethernet frame to the DAQ PC directly.
When each ROESTI receives the trigger signal, the event data is generated on each ROESTI (Fig.~\ref{fig:egTCP1}~(a)). The event data is sent to SiTCP1 of \textit{Network Processor} and encapsulated in the Ethernet frame. Each Ethernet frame is sent to the DAQ PC.
The destination MAC address of the Ethernet frame from ROESTI2 corresponds with the MAC address of the DAQ PC.
The Ethernet frame from the ROESTI2 attempts to pass through in the \textit{Path Controller} of the \textit{Network Processor} (Fig.~\ref{fig:egTCP1}~(b)).
If the Ethernet frame is sent to the \textit{Path Controller} from SiTCP1 at the same time, two Ethernet frames collide in the \textit{Path Controller} as shown in Fig.~\ref{fig:data collision}.
Either of the Ethernet frames disappears as mentioned in Section~\ref{sec:pathcont}.
In such cases, the SiTCP1 of the ROESTI1 or the ROESTI2 detects the loss of the Ethernet frame. The SiTCP1 subsequently resends the same Ethernet frame. This process is called TCP re-transmission~\cite{ref:TCP1}.
The TCP re-transmission might slow down the data transfer speed and cause unstable communication.
Therefore, this communication method cannot be used.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.5\textwidth]{Figures/nodatacollision2.eps}
\end{center}
\vspace{-5mm}
\caption{No data collision in the Arbiter1 of the Path Controller.}
\label{fig:nodatacollision}
\end{figure}
To avoid the TCP re-transmission, TCP connection is established between SiTCP1 of one ROESTI, which is the closest to the DAQ PC, and the DAQ PC as shown in Fig.~\ref{fig:nodatacollision}. This ROESTI sends the event data to the DAQ PC.
TCP connection is also established between SiTCP1 of another ROESTI and SiTCP0 of next ROESTI. These ROESTIs send the event data to the next ROESTI.
In such cases, the ROESTI receives the Ethernet frame, which the destination is own ROESTI, from the previous ROESTI. Because of MAC address match, the \textit{selector0} of the \textit{Path Controller} sends the Ethernet frame to SiCTP0 and the \textit{Arbiter} in the \textit{Path Controller} receives only one Ethernet frame as shown in Fig.~\ref{fig:nodatacollision}. Therefore, data collision does not occur in the \textit{Path Controller}.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.74\textwidth]{Figures/egTCP2.eps}
\caption{Example 2 of transferring event data over TCP/IP. (a) Two ROESTIs receive the trigger signal and the event data is generated in the Data I/F. (b) The Ethernet frame is sent to the ROESTI1 from the ROESTI2. (c) The event data in the ring buffer is sent to the DAQ PC. (d) The event data in the FIFO is sent to the DAQ PC. }
\label{fig:egTCP2}
\end{center}
\end{figure*}
Fig.~\ref{fig:egTCP2} shows an example of the event data transfer process over TCP/IP. In this example, the DAQ PC and two ROESTIs are connected with the daisy-chain.
When each ROESTI receives the trigger signal, the event data is generated on each ROESTI (Fig.~\ref{fig:egTCP2}~(a)). The event data is sent to SiTCP1 of the \textit{Network Processor} and encapsulated in the Ethernet frame. Each Ethernet frame is sent to the DAQ PC or the ROESTI1. Because the ROESTI2 sends the Ethernet frame to the next ROESTI, the destination of the Ethernet frame from the ROESTI2 is the ROESTI1.
Because of MAC address match, the Ethernet frame is sent to the \textit{SiTCP0} in the \textit{Network Processor} of the ROESTI1 (Fig.~\ref{fig:egTCP2}~(b)). The Ethernet frame is unencapsulated and the event data is stored in the FIFO of the \textit{Data Carrier}. If the ROESTI1 receives the next trigger signal, the event data is stored in the ring buffer of the \textit{Data Carrier}. Because \textit{TCP Arbiter} handles two event data, two event data never collide. The \textit{Data Carrier} arbitrates between two event data and sends the event data to the SiTCP1 as mentioned in Section~\ref{sec:datacarrier}. In the case of the example, the \textit{Data Carrier} extracts the event data of the ring buffer. The SiTCP1 makes new Ethernet frame and the event data is sent to the DAQ PC (Fig.~\ref{fig:egTCP2}~(c)). Subsequently, the \textit{Data Carrier} extracts the event data of the FIFO. The SiTCP1 makes new Ethernet frame and the event data is sent to the DAQ PC (Fig.~\ref{fig:egTCP2}~(d)).
Similar to this example, ROESTI receives the event data from the previous ROESTI and sends the event data to either the next ROESTI or the DAQ PC. This process is repeated in all the ROESTIs until all event data reach the DAQ PC.
The FIFO of the \textit{Data Carrier} acts as the TCP receive buffer. When the receive buffer is full, TCP temporarily suspends the data transmission process so that the buffer does not overflow. Therefore, event data are not lost.
In the \textit{Data Carrier}, the event data which contains smaller event number has priority. Therefore, the older event data has priority. By this process, all the ROESTIs have the same priority. It never happens that only a particular ROESTI sends event data over TCP/IP.
\section{ROESTI}
\label{sec:ROESTI}
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.47\textwidth]{Figures/ROESTI.eps}
\end{center}
\vspace{-5mm}
\caption{Photograph of ROESTI prototype.}
\label{fig:ROESTI}
\end{figure}
The ROESTI requires the gain of approximately 1\,V/pC and the signal-to-noise ratio (S/N) of larger than 5 at the minimum charge from the straw tube tracker. To achieve a momentum resolution of better than 200 keV/c on the tracker, a timing resolution of better than 1 ns is required.
Event pileup capability is also needed due to the high hit rate.
Straw tube tracker has 2400 channels and ROESTI needs to read all of these channel signals.
Based on these requirements, we have developed a ROESTI prototype, which consists of 16-channel signal input connector, ASD~\cite{ref:ASD}, DRS4~\cite{ref:DRS4}, ADC (AD9637, Analog Devices), FPGA (XC7A200T-2FBG676C, Xilinx Inc.), Trigger connector, and Small Form-Factor Pluggable (SFP) connector. Fig.~\ref{fig:ROESTI} shows a photograph of the ROESTI prototype. The ASD amplifies and shapes the detector signal. The DRS4 and the ADC digitize analog signals with high-speed and high-accuracy. We evaluated the performance of this prototype and confirmed that it satisfies all the requirements described above~\cite{ref:ROESTI1}.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.45\textwidth]{Figures/FPGA2.eps}
\end{center}
\vspace{-5mm}
\caption{Block diagram of FPGA.}
\label{fig:FPGA}
\end{figure}
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.35\textwidth]{Figures/network_before.eps}
\end{center}
\vspace{-5mm}
\caption{Block diagram of Network Processor.}
\label{fig:beforeNetwork}
\end{figure}
Fig.~\ref{fig:FPGA} shows a block diagram of the FPGA. The FPGA consists of \textit{Trigger I/F}, \textit{Data I/F}, \textit{Module Control}, and \textit{Network Processor}. After the FPGA receives the trigger signal from the trigger connector, the trigger signal is sent to the \textit{Trigger I/F}. In the \textit{Trigger I/F}, the trigger signal is handled and the event number is generated. Every time the trigger signal is received, the event number is incremented by one. The trigger signal and the event number are sent to \textit{Data I/F}.
After receiving the trigger signal, \textit{Data I/F} starts to receive digitized data from the ADC and to generate event data. The event data includes both the digitized data and the event number. The event data is sent to the \textit{Network Processor}.
Fig.~\ref{fig:beforeNetwork} shows a block diagram of the \textit{Network Processor}. The \textit{Network Processor} consists of ring buffer, SiTCP~\cite{ref:SiTCP}, and SFP I/F. After being temporarily stored in the ring buffer, the event data is sent to the SiTCP. The SiTCP is a hardware-based TCP processor for the Gigabit Ethernet and is designed for small devices limited by hardware size, such as an FPGA in front-end devices.
The SiTCP enables a user circuit to process receiving and transmitting over TCP/IP.
Using the SiTCP, the event data are encapsulated in the TCP packet.
The SiTCP provides a mechanism for slow control over UDP/IP~\cite{ref:UDP}.
In the slow control process with the SiTCP, at first, the DAQ PC must send a UDP packet for requests to the SiTCP.
When the SiTCP in the \textit{Network Processor} receives the UDP packet, the SiTCP extracts the slow control signal from the UDP packet and sends the signal to the \textit{Module control}.
The slow control signal includes address and data which are needed for the process of the slow control.
The \textit{Module control} receives the slow control signal and processes it. After completing the slow control process, the
\textit{Module control} sends the slow control signal to SiTCP as an acknowledge signal.
The SiTCP encapsulates the slow control signal in the UDP packet and sends the UDP packet to the DAQ PC.
In order to communicate by TCP/IP or UDP/IP, the SiTCP encapsulates the TCP packet or the UDP packet in an Ethernet frame~\cite{ref:ethernet} which includes the destination MAC address. By exchanging the Ethernet frame with other network devices, the SiTCP can communicate over Ethernet.
The SFP I/F provides physical layer process such as Physical Coding Sublayer (PCS) and Physical Medium Attachment (PMA). This module connects the Ethernet frame to the SFP connector and created using Xilinx 1G/2.5G BASE-X PCS/PMA Core~\cite{ref:IP}.
\section{Introduction}
\label{sec:introduction}
\IEEEPARstart{T}{he}
COherent Muon to Electron Transition (COMET) experiment at J-PARC aims to search for the neutrinoless transition of a muon to an electron ($\mu$--$e$ conversion) in a muonic atom. Since charged lepton flavor is violated in this process, a branching ratio is highly suppressed to $\mathcal{O}(10^{-54})$ in the Standard Model (SM). However, theoretical models beyond the SM predict that the branching ratio of this process is to be $\mathcal{O}(10^{-15})$~\cite{ref:COMET}. Therefore, the discovery of $\mu$--$e$ conversion should be a clear evidence of new physics.
In order to suppress the background and to achieve the goal sensitivity, we adopt a straw tube tracker for the electron detector~\cite{ref:straw}. Since the detector is composed of an extremely light material which is operational in a vacuum, an excellent momentum resolution of better than 200 keV/c is achieved. We have developed the readout electronics board called ROESTI (Read-Out Electronics for Straw Tube Instrument) which reads out the signal from the detector precisely~\cite{ref:ROESTI1}. In order to prevent the degradation of the detector signal, the ROESTI needs to be located near the detector. We plan to install the ROESTI in the gas manifold of the detector. Data of the detector signal is transferred to the data acquisition (DAQ) PC with Ethernet~\cite{ref:ethernet}. This is because Ethernet provides advantages such as high-speed, high reliability, and high maintainability.
Although various kinds of Ethernet network topologies are created by using network switches, commercial network switches cannot be used in the gas manifold due to the space limitations and the radiation hardness. The heat of the ROESTI is cooled by the gas flow in the gas manifold.
Therefore, the number of the feedthroughs required for data communication lines should be the same as the one of the ROESTIs as shown in Fig.~\ref{fig:daisy}~(a).
However, we cannot adopt this network topology, because the number of vacuum feedthroughs needs to be reduced due to space constraints and cost limitations.
To avoid this problem, we developed a daisy-chain function of Gigabit Ethernet for the FPGA on the ROESTI as shown in Fig.~\ref{fig:daisy}~(b).
The required function for the communication with the daisy-chain is to transfer the data with TCP/IP~\cite{ref:TCP1} over Ethernet which provides high-speed and reliable communication. Additionally, the slow control function is required for setting and reading the parameters of the FPGA and IC chips on the ROESTI.
The performance target in terms of data transfer throughput on the daisy-chain is to achieve close to the maximum rate of TCP/IP over Gigabit Ethernet and to reduce data loss as much as possible.
To enable the other ROESTIs to transfer data, the daisy-chain needs to transfer data bidirectionally as a fail-safe.
\begin{figure}[t]
\centerline{\includegraphics[width=3.0in]{Figures/datatransfer1.eps}}
\centerline{\includegraphics[width=3.0in]{Figures/datatransfer2.eps}}
\caption{(a) Normal network topology (b) Network topology with the daisy-chain. The number of vacuum feedthroughs is decreased drastically.}
\label{fig:daisy}
\end{figure}
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.45\textwidth]{Figures/SideDefine.eps}
\end{center}
\vspace{-5mm}
\caption{Next ROESTI and previous ROESTI. The ROESTI1 is the next ROESTI from the ROESTI2. The ROESTI3 is the previous ROESTI from the ROESTI2.}
\label{fig:sidedefine}
\end{figure}
Each ROESTI has two neighboring ROESTIs as shown in Fig.~\ref{fig:sidedefine}.
We define the neighboring ROESTIs of near and far sides of the DAQ PC as "next ROESTI" and "previous ROESTI", respectively.
\section*{Acknowledgment}
\input{ref.tex}
\end{document}
\section{Measurement and Result}
\label{sec:Measurement}
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.3\textwidth]{Figures/setup.eps}
\end{center}
\vspace{-5mm}
\caption{Measurement setup.}
\label{fig:setup}
\end{figure}
The performance of TCP data transfer was evaluated by using two to six ROESTIs. Fig.~\ref{fig:setup} shows the experimental setup. Each ROESTI and the DAQ PC were connected with the daisy-chain. The function generator generated the trigger signal periodically. The trigger signal was distributed to all the ROESTIs using the trigger distributor. After receiving the trigger signal, each ROESTI sends the one event data to the DAQ PC with the daisy-chain over TCP. The DAQ PC received the data from each ROESTI and calculated the transfer speed and throughput of the entire data. The data size of event was set to 37112 bytes, which was corresponding to the maximum event data size. The DAQ PC was a Sen-SV9R-LCi7EX-TMZ-VEditor4K, Intel Core i7-5960X (3.0-3.5 GHz, 8 Cores, 16 Threads, 20 MB Smart Cache, TDP140W) running Scientific Linux release 6.10.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.4\textwidth]{Figures/trigger_vs_speed2.eps}
\end{center}
\vspace{-5mm}
\caption{TCP Data transfer speed as a function of trigger frequency. The difference in color indicates the difference in the number of ROESTIs connected by the daisy-chain.}
\label{fig:freq_vs_speed}
\end{figure}
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.4\textwidth]{Figures/result_final2.eps}
\end{center}
\vspace{-5mm}
\caption{Throughput of TCP data transfer as a function of elapsed time.}
\label{fig:time_vs_speed}
\end{figure}
Fig.~\ref{fig:freq_vs_speed} shows the TCP data transfer speed as a function of trigger frequency with two to six ROESTIs connected by the daisy-chain. When the transfer speed was not saturated, each board sent all the event data without loss after receiving the trigger signal. Hence the data transfer speed was proportional to the trigger frequency. In any number of boards, the maximum data transfer speed was 950~Mbps, which corresponds to the theoretical limit speed of TCP over Gigabit Ethernet calculated with overheads, such as protocol headers; this implies that the throughput corresponds to the theoretical limit of TCP over Gigabit Ethernet. In the case of six boards, when the trigger frequency was 533~Hz, the transfer speed reached 950~Mbps ($= 533$~Hz~$\times$~$37112$~bytes/event~$\times$~$6$~ROESTIs) of the theoretical limit. When the trigger frequency was greater than 533~Hz, all the ROESTIs received approximately 533 trigger signals per second and other trigger signals were ignored.
Fig.~\ref{fig:time_vs_speed} shows the throughput of the TCP data transfer as a function of the elapsed time. In this measurement, six ROESTIs and the DAQ PC were connected with the daisy-chain and the trigger frequency was set at 570~Hz. To evaluate the stability, we measured the throughput of the TCP data transfer every time 1000 event data were sent to the DAQ PC. Throughput of the TCP data transfer constantly corresponded with the theoretical limit speed of TCP over Gigabit Ethernet. We also measured data loss rate when total data rate was less than the theoretical limit speed. In this measurement, we used six ROESTIs and total data transfer speed was set at 910~Mbps. All ROESTI constantly sent 100\% of the data for 1 hour. We confirmed that our Gigabit Ethernet daisy-chain function satisfied the target performance.
In the realistic COMET situation,
five or more ROESTIs are planned to be connected with one daisy-chain line in the COMET experiment. All the ROESTIs receive the same trigger signal. After receiving the trigger signals, each ROESTI generates one event data. Average trigger rate was estimated to be 1~kHz by the simulation study. Event data size is variable and the average data size is estimated to be 10300~bytes. Total average data rate in one daisy-chain with five boards is estimated to be 410~Mbps ($= 10300$~bytes~$\times$~$1$~kHz~$\times$~$5$~ROESTIs).
This data rate is small enough to be compared to the bandwidth of Gigabit Ethernet. Even if we connect more ROESTIs, network communication with daisy-chain is possible.
Therefore, we are going to adopt this function in the COMET experiment.
\section{Summary}
\label{sec:Summary}
We are promoting the preparation for the COMET experiment at J-PARC, which aims to search for the neutrinoless transition of a muon to an electron. We plan to adopt the straw tube tracker as a $\mu$--$e$ conversion detector to achieve an excellent momentum resolution of better than 200~keV/c. We have developed the readout electronics board called ROESTI for the detector. To prevent the degradation of the detector signal, we plan to install the ROESTI in the gas manifold of the detector. The number of vacuum feedthroughs needs to be reduced due to space constraints and cost limitations. To decrease the number of vacuum feedthroughs drastically, we developed a new network communication scheme with the daisy-chain over Gigabit Ethernet.
The required function for communication with the daisy-chain is to transfer data with TCP/IP over Ethernet. Additionally, a slow control function is required for setting and reading parameters of the FPGA and IC chips on each ROESTI. We implemented two SiTCPs and new circuits for the daisy-chain on the FPGA of ROESTI.
The FPGA on ROESTIs verifies the MAC address of the Ethernet frames and controls the path of the ones. This circuit enables the DAQ PC to communicate with the target ROESTI directly for the slow control over UDP/IP.
To avoid data transfer slowdown and unstable communication by TCP re-transmission, a ROESTI sends the event data to the next ROESTI.
This process is repeated until all the event data are finally sent to the DAQ PC.
We measured the throughput, the stability, and the data loss rate of the Gigabit Ethernet daisy-chain function and obtained reasonable results in any number of ROESTIs.
We are going to adopt this function in the COMET experiment.
|
\section{Introduction}
The galactic cosmic rays are believed mainly come from supernova remnants. After injected from the sources, cosmic rays propagate in the interstellar space.
When cosmic rays enter the heliosphere, the interaction with the solar wind and the embedded magnetic field results in the intensity and the spectral shape of low energy cosmic rays are different from the local interstellar spectrum (LIS) \cite{1971RvGSP...9...27J,2013LRSP...10....3P}.
These effects on cosmic rays are called solar modulation.
The solar modulation effect limits our understanding for cosmic rays outside the heliosphere.
Therefore, the study of solar modulation is important for studying the injection and propagation parameters of cosmic rays, dark matter indirect measurement and the diffusion theory in the galaxy and heliosphere \cite{2017ApJ...840..115B,2017arXiv171203178T,2018ApJ...858...61B,2018ApJ...854...94B,2018ApJ...863..119Z,1973ApJ...182..585J,2002A&A...393..703T,2003AdSpR..32..549B,2004GeoRL..3110805B,2004JGRA..109.1109P,2009ASSL..362.....S,2010JGRA..115.3103P,2018ApJ...856...94Z,2018ApJ...854..137S}.
The recent experimental results from Voyager 1, PAMELA, and AMS-02 have achieved great breakthroughs which are useful to understand the solar modulation effect.
The Voyager 1 flew outside the heliosphere on August 2012 and directly measured the LIS in the range from a few to hundreds MeV/nucleon \cite{2013Sci...341..150S,2013Sci...341.1489G,2016ApJ...831...18C}.
The monthly PAMELA measurements of proton spectra \cite{2013ApJ...765...91A,2018ApJ...854L...2M} shed light on some details of the solar modulation \cite{2014SoPh..289..391P,2015ApJ...815..119V,2016ApJ...829....8C,2016AdSpR..57.1965R,2017ApJ...849L..32T,2017ApJ...846...56Q,2019PhRvD.100f3006W}.
Recently, the AMS-02 collaboration has published the continuous proton and helium energy spectrum with rigidity above $1\,\mathrm{GV}$ for proton and $1.9\,\mathrm{GV}$ for helium between May 2011 and May 2017 \cite{Aguilar:2018wmi}.
Some important results have been obtained, such as the confirmation of the velocity dependence of cosmic ray diffusion and finding the increase of the slope of perpendicular mean free path during solar maximum for low rigidity particles \cite{2018PhRvL.121y1104T,2019ApJ...871..253C}.
In \cite{2019PhRvD.100f3006W} (hereafter Paper I), we build a modulation model
to well reproduce the long time PAMELA proton measurements between July 2006 to
February 2014.
This modulation model includes mainly physical processes affecting the
propagation of cosmic rays in the heliosphere: diffusion, convection, drift, and
energy loss.
Meanwhile, some main factors affecting solar modulation are taken from
observations, such as the magnitude of the heliospheric magnetic field, the
solar wind speed, and the tilt angle of the heliospheric current sheet (HCS).
We adopt the same model in this work.
We also deliberate to keep the model as simple as possible by only including the
minimal degree of freedom.
The 2D modulation code solarprop
\footnote{\url{http://www.th.physik.uni-bonn.de/nilles/people/kappl/}}\cite{
2016CoPhC.207..386K} is used to solve the cosmic ray propagation equation and
obtain the modulated spectra.
In Paper I, we find that
the modulation processes are different between negative and polarity reversal
period.
Therefore, we analyze the modulation effect with AMS-02 proton data separately
in three periods related to the magnetic field polarity.
After successfully reproducing the proton observations, the modulation
parameters for proton are applied to calculate the modulated helium spectrum.
The paper is organized as follows. In Section ~\ref{sec:model}, we briefly
introduce the modulation model and present the LIS for proton and helium.
In Section ~\ref{sec:proton}, we compute the modulated proton spectrum and
compare them to observations.
In Section ~\ref{sec:helium}, we check the assumption that proton and helium
have the same mean free path by computing the helium spectrum with the
modulation parameters for proton.
Finally, we give a summary in Section ~\ref{sec:conclusion}.
\section{A 2D solar modulation model}\label{sec:model}
There are four major modulation mechanisms for cosmic rays in the heliosphere:
diffusion on irregularities of the heliospheric magnetic field, convection by
the outward solar wind, particles drift in the non-uniform magnetic field and
adiabatic energy loss.
Several review articles discuss the modulation process in great detail
\cite{1971RvGSP...9...27J,2013LRSP...10....3P}.
The propagation process can be described by the Parker equation
\cite{1965P&SS...13....9P}:
\begin{equation}
\frac{\partial f}{\partial t} = -(\vec{V}_{sw} +\vec{V}_{drift})\cdot \nabla f +
\nabla \cdot[K^s\cdot \nabla f] + \frac{\nabla \cdot \vec{V}_{sw}}{3}
\frac{\partial{f}}{{\partial\ln p}} ,
\end{equation}
where $f(\vec{r},p,t)$ is the omni-directional distribution function, $\vec{r}$
is the position in the heliocentric spherical coordinate system, $p$ is the
particle momentum, $\vec{V}_{sw}$ is the solar wind speed, $\vec{V}_{drift}$ is
the drift speed, $K^s$ is the symmetric part of diffusion tensor. The
differential intensity related with the distribution function is given by $I =
p^2f$.
It is customary to assume that the diffusion coefficient can be separated into
spatial and rigidity components \cite{2013LRSP...10....3P}.
The generally assumption about the rigidity part is that all particle species
have a universal function of rigidity
\cite{1971Ap&SS..11..288G,1971JGR....76..221F,1975JGR....80.1701F,
2013SSRv..176..299M,2018AdSpR..62.2859B,2017ICRC...35...24V}.
In Paper I, we adopt a linear rigidity dependence of the diffusion coefficient
and can reproduce the PAMELA monthly proton measurements between 2006 to 2012.
But during the polarity reversal period which is assumed between November 2012
and March 2014 in Paper I based on Ref. \cite{2015ApJ...798..114S}, the time-dependent rigidity dependence is necessary
to reproduce the observations.
In the present work, the parallel diffusion coefficient is adopted as the
following form taking into account the finding in Paper I:
\begin{equation}
k_{\parallel} = \frac{1}{3} k \beta (\frac{R}{1\,\mathrm{GV}})^{\delta}
\frac{B_E}{B} \end{equation}
where $k = 3.6 \times 10^{22}k_0~\cm^2/\mathrm{s}$ is a scale factor to model the time
dependence of the diffusion coefficient, $\beta$ is the particle speed in the
unit of the speed of light,
$\delta$ determines the rigidity dependence of the diffusion coefficient (by
default $\delta =1$),
$B_E$ is the heliospheric magnetic field strength near the Earth,
$B=B_0/{r^2}\sqrt{1+\tan{\psi}}$ is the heliospheric magnetic field strength at
the particle position and $\psi$ is the angle between magnetic field direction
and its radial direction \cite{1958ApJ...128..664P}.
The standard Parker magnetic field model \cite{1958ApJ...128..664P} is used in
this work.
We take the perpendicular diffusion coefficient to be $k_{\perp} = 0.02
k_{\parallel}$ according to the test particle simulation
\cite{1999ApJ...520..204G}.
The diffusion coefficient is also often marked as $k_{\parallel/\perp}=\frac{1}{3}v \lambda_{\parallel/\perp}$,
where $v$ is particle speed and $\lambda_{\parallel}$ ($\lambda_{\perp}$) is called the parallel (perpendicular) mean free path.
The gradient and curvature drift speed is written as $V_{gc} = q \frac{\beta
R}{3} \, \nabla \times \frac{\vec{B}}{B^2}$ \cite{1977ApJ...213..861J}.
We describe the heliospherical current sheet (HCS) drift following Ref.
\cite{1989ApJ...339..501B}, where a thick, symmetric transition region
determined by the tilt angle is used to simulate a wavy neutral sheet.
The HCS drift speed $V_{ns}^{w}$ is given by
\begin{equation}
\begin{aligned}
\vec{V}_{ns}^{w} = & \begin{cases}
qA \frac{v \theta_{\triangle}
\cos(\alpha)}{6\sin(\alpha+\theta_{\triangle})} \vec{e}_{r} ,&
\pi/2-\alpha-\theta_{\triangle} < \theta <\pi/2 + \alpha +\theta_{\triangle} \\
0, & else
\end{cases}
\end{aligned}
\end{equation}
where $q$ is the charge sign and $v$ is the particle speed, $\theta_{\triangle}
\approx \frac{2RV_{sw}}{B_0\Omega \cos{\alpha}}$.
When the polar solar magnetic field directs outward in the north (southern)
hemisphere and inward in the southern (north) hemisphere, it is said that the
Sun is in a positive (negative) polarity cycle marked as $A>0$ ($A<0$).
$qA$ determines the drift direction.
Taking into account the possible suppression of the drift effect, a scale factor
$k_d$ (by default $k_d=1$) is introduced and the drift velocity is described as
$V_{drift} = k_d(V_{gc}+V_{ns}^{w})$
\cite{2003AdSpR..32..645F,2015ApJ...810..141P}.
The solar wind speed and the magnitude of magnetic field are taken from the
website {\url{omniweb.gsfc.nasa.gov}}.
The tilt angle of HCS is obtained from the website \url{wso.stanford.edu} with
the ``new'' model
These quantities are averaged over several months which corresponds to the time
of solar wind propagation from the Sun to the modulation boundary at $100
\,\mathrm{AU}$.
More elaborate description to our model is given in Paper I and references
therein.
The discussion about the modulation resulted from the merged interaction regions
\cite{1993ApJ...403..760P,1993AdSpR..13..239P,2019ApJ...878....6L} is not
included in the recent work.
As an initial input condition in the modulation model, the LIS are constrained
by the current experimental measurements.
Voyager 1 has directly measured the LIS in the range from a few to hundreds of
MeV/nucleon.
The monthly precise AMS-02 data provide important ingredients to reconstruct the
LIS.
The LIS for proton and helium are constructed by the cubic spline interpolation
method following the works \cite{2016A&A...591A..94G,2018ApJ...863..119Z,2019PhRvD.100f3006W}.
This method avoids the bias comes from the cosmic ray injection and propagation model which is still in debate.
We determine the proton and helium LIS by matching the low energy LIS to the
Voyager 1 measurements and fitting the calculated spectrum to the AMS-02 data
observed during Bartels rotation 2429, 2432, 2435 and 2438
\cite{2016ApJ...831...18C,Aguilar:2018wmi}. These time periods are all within
the negative polarity and the data can be well explained with one free
parameter $k_0$.
The GNU Scientific Library
(GSL)\footnote{\url{https://www.gnu.org/software/gsl/}} is used to perform the
least-squares fitting.
The energy knots and the corresponding intensities in the cubic spline
interpolation method are shown in Table \ref{tab:plis} and \ref{tab:helis}.
The difference between the LIS obtained by this work and that derived in Paper I
is very small (see Appendix \ref{sec:liss}).
Once the LIS have been derived, the effects of solar modulation are calculated
directly from the model.
\begin{table}[!htbp]
\caption{The parameterization of proton LIS with the cubic spline
interpolation method. $E_k$ is kinetic energy and $I$ is intensity.} \label{tab:plis}
\centering
\begin{tabular}{lcccccccl}\hline \hline
$\mathrm{log(E_k/\GeV)}$ &-2.42 &-1.41 &-0.50 &0
&0.50 &1.00 &1.50 &2.00 \\
$\mathrm{log(I/(\GeV\mathrm{m}^2\mathrm{sr}\,\mathrm{s}))}$ &4.2905 &4.4688 &4.0176 &3.4548
&2.5849 &1.4158 &0.0597 &-1.3497 \\
\hline
\end{tabular}
\end{table}
\begin{table}[!htbp]
\caption{The parameterization of helium LIS with the cubic spline
interpolation method. $E_k$ is kinetic energy and $I$ is intensity.} \label{tab:helis}
\centering
\begin{tabular}{lccccccl}\hline \hline
$\mathrm{log(E_k/\GeV)}$ &-2.27 &-1.28 &-0.30
&0.56 &1.22 &1.78 &2.29 \\
$\mathrm{log(I/(\GeV\mathrm{m}^2\mathrm{s}\,\mathrm{sr}))}$ &2.3812 &2.7324 &2.6735
&1.7705 &0.4619 &-0.8980 &-2.2892 \\ \hline
\end{tabular}
\end{table}
\section{Solar modulation for proton}\label{sec:proton}
The AMS-02 data are taken during different solar activity levels and different
magnetic field conditions: the negative polarity cycle, the undefined polarity
period around the solar maximum and finally the positive polarity cycle.
In Paper I, we show that the diffusion and the drift are both different during
the negative and the polarity reversal periods.
Thus, we investigate the modulation effect for proton separately in three
periods related to the magnetic field polarity.
\subsection{Modulation of CR proton with the assumption of the negative
polarity} \label{sec:negative}
During every solar maximum the polarity of the solar magnetic field and
subsequently the heliospherical magnetic field reverses direction.
After the polarity reversal took place around 2000 \cite{2003ApJ...598L..63G},
the polarity is negative until the recent reversal.
During the negative polarity cycle ($A<0$), positively charged particles drift into the
inner heliosphere along the HCS and out over the poles.
Due to the asymmetric solar activity, the reversal of the solar magnetic field
polarity is not simultaneous in two hemispheres.
The summary of some estimates of the solar polar field reversal times for the northern
and southern solar hemisphere is presented in Table \ref{tab:time}.
\begin{table}[!htbp]
\caption{Estimates of the time of solar polar magnetic field polarity
reversals in the northern and southern hemisphere.} \label{tab:time}
\centering
\begin{tabular}{ccc}\hline \hline
North &South &Ref. \\
2012/06 &-
&\cite{2014SoPh..289.3381K} \\
2012/11 &2014/03
&\cite{2015ApJ...798..114S} \\
2013/07 &2015/01
&\cite{2015GeAe..55..969T} \\
2012/05-2014/04 &2013/06-2015/03
&\cite{2016KPCB...32...78P} \\
2012/10-2015/09 &2014/06
&\cite{2016ApJ...823L..15G} \\
2012/06-2014/11 &2013/10
&\cite{2018AA...618A.148J} \\
\hline
\end{tabular}
\end{table}
One can see from Table \ref{tab:time} that the estimated reversal time can be
very different by means of different methods and data.
The polarity reversal period is not well determined.
However, we also see that the polarity reversal process had not happened before
the early of 2012. Thus, it is safe to set the heliospheric magnetic field polarity as negative in this
time period.
We compare the computed spectrum with the AMS-02 measurements with rigidity
below $40\,\mathrm{GV}$.
For the particle with higher rigidity, the modulation effect is negligible.
Following the scenario in Paper I which reproduces 6 years PAMELA proton
spectrum between 2006 to 2012, we fixed both $k_d$ and $\delta$ as 1 and
only adjust $k_0$ to fit the observation.
The resulting time profile of the reduced-$\chi^2$ ($\chi^2/(d.o.f.)$) is shown
in Figure \ref{fg:am_chi2}.
The reduced-$\chi^2$ is around or less than 1 until August 2012.
There is no significant need to introduce more free parameter for this period.
However, after the August 2012, the reduced-$\chi^2$ suddenly increases to an
unacceptable level with reduced-$\chi^2>2$. The model with only one free
parameter $k_0$ fails to correctly describe the modulation process after this
time node.
\begin{figure}[hbt!]
\centering
\includegraphics[width=1\textwidth]{ams_AM_chi2new.pdf}
\caption{The time profile of reduced-$\chi^2$ for the fit to the monthly
AMS-02 proton data between May 2011 and November 2012 under the assumption
$A<0$. The scale factor of diffusion coefficient $k_0$ is taken as the free
parameter.}
\label{fg:am_chi2}
\end{figure}
\subsection{Modulation of CR proton with the assumption of the positive
polarity} \label{sec:positive}
As showed in Table \ref{tab:time}, the estimated latest time of the completion
of the reversal is September 2015.
After the reversal, the solar magnetic field polarity becomes positive ($A>0$).
Positively charged particles drift into the inner heliosphere over the poles and
out of it along the HCS. They have less difficulty in reaching the Earth and
less modulation than that during the negative polarity period.
When the solar activity indicated by the sunspot number decreases to moderate
level, we expect the turbulence magnetic properties and the rigidity
dependence of the mean free path recover to the similar behavior with
$\delta=1$ as in the negative polarity cycle.
We attempted to adopt the full drift effect ($k_d=1$)
, but the required scale factor of diffusion coefficient $k_0$
is much smaller than 0.7 prior to October 2016 (see Appendix \ref{sec:full}). Under the simple framework of the
force-field approximation, the modulation potential is inversely proportional to
the diffusion coefficient characterized by $k_0$.
The ratio of solar modulation potential (also the diffusion coefficient) reconstructed from the neutron monitor count rate \cite{2017AdSpR..60..833G,2019JGRA..124.2367K} is within 2.1 during May 2011 to May 2017. If the full drift effect is
adopted, the required diffusion coefficient is too small and far from this relation.
In addition, this scenario results in larger $\chi^2$
than the suppressed drift case (see Appendix \ref{sec:full}). Thus, we set
$k_d$ as a free parameter in this time period to reduce the drift effect.
In Figure \ref{fg:ap_chi2} we show that the time profile of reduced-$\chi^2$ and
the scale factor for the drift speed $k_d$.
The reduced-$\chi^2$ values are less than or close to 1 during November 2015 to
May 2017. The scale factor of drift velocity $k_d$ is nearly 0 around the
second half of 2015, and it is about 0.8 in March 2017.
It indicates that the drift effect is suppressed during this period.
The increase of $k_d$ form 0 to 0.8 indicates the gradual recovery of the
drift effect and implies that the drift effect may fully recover around the
middle of 2017.
There are several mechanisms that may cause the suppression of
the drift effect. The large-scale fluctuations in the heliospheric magnetic
field, such as the interaction regions and the merged interaction regions, fill
the heliosphere so that drifts may only occur on a less scale during moderate
to high solar activity period \cite{1993AdSpR..13..239P}.
In addition, numerical simulation shows that the presence of scattering can
also suppress the drift effect. For an intermediate degree of scattering, the
drift velocity is typically suppressed by a larger degree; when the scattering
is very strong, there is no large-scale drift motions
\cite{2007ApJ...670.1149M}.
Note that there is some degeneracy between $k_0$ and $k_d$ as shown in
Appendix \ref{sec:deg}. As the drift effect is opposite for particle with
opposite charge, a simultaneous fit to the proton and future antiproton
spectrum is crucial to reduce the uncertainty and get a better understanding for
the drift effect.
\begin{figure}[hbt!]
\centering
\includegraphics[width=1\textwidth]{ams_AP_chi2new.pdf}
\caption{The time profile of reduced-$\chi^2$ and the drift scale factor $k_d$ from the fits to the
AMS-02 proton data during August 2015 to May 2017 under the assumption of $A>0$.
The scale factor for diffusion coefficient and drift velocity, $k_0$ and $k_d$
, are free parameters in this period.}
\label{fg:ap_chi2}
\end{figure}
\subsection{Modulation for proton between August 2012 and October
2015}\label{sec:ps}
From Section \ref{sec:negative}, we show that the default model with the linear
rigidity dependence on the mean free path and the full drift effect ($\delta=1$,
$k_d=1$) fails to describe the modulation process since August 2012.
Additionally, the model with linear rigidity dependence on the mean free path
and the suppressed drift effect ($\delta=1$, $k_d \in [0,1]$) used in Section
\ref{sec:positive} can not reproduce the observations before October 2015.
In Paper I, after we tested various configurations for the diffusion
coefficient and drift effect to reproduce the PAMELA proton observations during
November 2012 to February 2014
, we concluded that the combination of the time-dependent power-law rigidity
dependence on the mean free path and the zero drift configuration give the best
fit to the data.
So in this case, the free parameters are $k_0$ and $\delta$.
This scenario is adopted to reproduce the AMS-02 observations between August
2012 and October 2015 which coincides with some estimated polarity reversal periods.
The time profile of the slope of the mean free path $\delta$ and the
reduced-$\chi^2$ are shown in Figure \ref{fg:ps_chi2}.
We find that $\delta$ roughly keeps increasing until it reaches the maximum
value of 1.28 in October 2013 and then decreases to 1.07 in October 2015.
The variation of rigidity dependence should be noticed in all cosmic ray
species, such as helium.
Almost all the reduced-$\chi^2s$ are smaller than 1.
Obviously, the combination of the two parameters $k_0$ and $\delta$ is adequate
to reproduce the observations.
Although introducing $k_d$ as the third free parameter may further improve the
fit,
but the parameter space will not be constrained well because of the
degeneracy between diffusion and drift parameters.
The future monthly antiproton data is needed to reduce the degeneracy.
The drift velocity is assumed to be 0 in this subsection, which may not be
realistic in the whole period since it may have a transition process.
Because of the degeneracy between the scale factor ($k_0$), the slope ($\delta$)
of the diffusion coefficient and the drift speed ($k_d$),
these transition processes have to be studied separately in greater detail.
\begin{figure}[hbt!]
\centering
\includegraphics[width=1\textwidth]{ams_ps_chi2.pdf}
\caption{The time profile of reduced-$\chi^2$ and the slope of the mean free
path $\delta$ for the fits to the AMS-02 proton data between August 2012 and
October 2015. The scale factor for the diffusion coefficient $k_0$ and slope for
the diffusion coefficient $\delta$ are free parameters in this period.}
\label{fg:ps_chi2}
\end{figure}
\section{Solar modulation for helium}\label{sec:helium}
It is an important assumption that the mean free path is the same for all
species of nuclei.
The precise AMS-02 measurements provide a good opportunity to check this widely
adopted assumption. The main parameter $k_0$ for proton is shown in the bottom
panel of Figure \ref{fg:sum}. The time profiles of $k_d$ and $\delta$ are
shown in Figure \ref{fg:ap_chi2} and Figure \ref{fg:ps_chi2}, respectively.
We take the modulation parameters ($k_0$, $\delta$, $k_d$) obtained in the
previous section as inputs to directly compute the modulated spectrum for
helium.
The reduced-$\chi^2$ for helium and proton are summarized in the upper panel of
Figure \ref{fg:sum}. We find that the same modulation parameters can
well reproduce the proton and the helium observations simultaneously.
Meanwhile, we show the ratios of the computed intensities to the measured
intensities as functions of rigidity and time in Figure \ref{fg:ratio}.
It can be seen that most of the fits agree with the data within $\pm5\%$.
\begin{figure}[hbt!]
\centering
\includegraphics[width=1\textwidth]{results.pdf}
\caption{The upper panel shows the time profile of reduced-$\chi^2$ for
proton (red dots) and helium (blue dots). The bottom panel shows the best fit
$k_0$ for proton. Note that in the upper panel, the modulation parameters for
helium are taken from proton.}
\label{fg:sum}
\end{figure}
\begin{figure}[hbt!]
\centering
\begin{tabular}{@{}cc@{}}
\includegraphics[width=0.65\textwidth]{endp171vld.pdf}
\includegraphics[width=0.65\textwidth]{vldhe171.pdf}
\end{tabular}
\caption{The left (right) panel show the ratio of the computed proton
(helium) intensities to measured values. The same modulation parameters are
applied for proton and helium.}
\label{fg:ratio}
\end{figure}
Recalling the treatment to the modulation for boron and carbon in Paper I,
in which the same mean free paths are able to reproduce the ACE boron and
carbon observations, different nuclei could have
the same mean free path in the heliosphere.
The upcoming monthly helium data of PAMELA between July 2006 and January 2016
\cite{2019EPJWC.20901004M} and future time-dependent nuclei data of AMS-02 will
give it a further test.
\section{Conclusion}\label{sec:conclusion}
The precise measurements of monthly cosmic ray proton and helium spectra by
AMS-02 between May 2011 to May 2017 provide an important chance to improve our
understanding for the solar modulation.
Compared to the PAMELA data up to February 2014, the AMS-02 data cover the whole
solar magnetic field polarity reversal period around the solar maximum and part
of the positive polarity cycle.
Meanwhile, the precise measurements of monthly helium spectrum provide a chance
to check the important assumption that the cosmic ray proton and helium
have the same diffusive mean free path in the
heliosphere.
A two-dimensional model is used to describe the propagation of proton and helium
in the heliosphere.
The model includes all major modulation processes and the variation of the
heliosphere environment, such as the solar wind speed, the magnetic field
strength and the tilt angle of HCS.
We consider a simplest reasonable scenario to reproduce the observations.
With no more than two free parameters, the computed spectrum are able to match
the AMS-02 proton and helium observations.
We find that the rigidity dependence of the mean free path is varied with time.
The linear rigidity dependence is adequate to reproduce the observations before
August 2012 or after October 2015.
Within the possible polarity reversal period between August 2012 and October 2015,
the time varying power-law dependence of the mean free path is essential to fit the data.
We also find that the zero drift
effect can well reproduce the observations during the polarity reversal
period and the suppressed drift effect clearly keeps recovering after this
period.
Finally, with the help of the precise monthly helium measurements, we confirm
that the mean free path is the same for proton and helium.
The future monthly data from AMS-02 and PAMELA for other nuclei will provide further
checking for the assumption that all nuclei have a universal mean free path,
and the monthly antiproton data would provide invaluable help to understand the
role of the drift effect in different solar activity periods.
\section*{Acknowledgement}
This work is supported by the National Key R\&D Program of China (No.
2016YFA0400200),
the National Natural Science Foundation of China (Nos. U1738209 and 11851303).
|
\section{Introduction}
Ontology Reuse (OR) is a critical aspect for the evolution of the Semantic Web since its origins \cite{gruber1993translation,neches1991enabling,studer1998knowledge,katsumi2016ontology}. OR aims to foster semantic interoperability between data sources and facilitate data integration tasks. As a consequence, it is naturally intertwined with other ontology engineering areas, such as ontology selection (OS), ontology integration (OI), and ontology access (OA), which in turn affect OR strategies.
To date, several approaches to OR exist, due to different motivations or different implementations. For instance, institutions and community consortia typically encourage \emph{direct} reuse of existing standard or popular ontologies. Research communities and communities of practice also suggest \emph{indirect} approaches, i.e. by designing ontologies tailored to a use case and aligning them to standard or popular ones when appropriate, or \emph{hybrid} approaches, i.e. when mixing direct and indirect approaches.
The decision on the OR approach to be adopted is usually done by ontology engineers when bootstrapping a project. However, despite a vast literature on OR provides definitions and requirements, decision methods may be biased and shallow. Hence, solutions are often taken on a case-by-case basis, hampering the definition of shareable good practices.
On the one hand, reuse approaches are primarily accompanied or motivated by design methods. Those are ultimately based on the conceptualisation needs of an ontology project. Methods borrowed from software engineering, philosophy and cognitive science have emerged. Examples are competency questions \cite{Gruninger1994}, foundational ontologies \cite{gangemi2002sweetening}, task-oriented quality assessment \cite{DBLP:conf/esws/GangemiCCL06}, design patterns \cite{Gangemi05,Gangemi2009,DBLP:books/ios/HGJKP2016}, etc. Such methods aim at making explicit the general cognitive requirements that should guide an ontology project.
On the other hand, social argumentations often affect designers' decisions. International institutions, community consortia and standardisation bodies tend to support their own practices and shared ontologies; domain experts may perceive the reuse of known ontologies as simpler and more authoritative; ontologists tend to prioritise their expertise, etc.
In this chapter we argue that to date only a few solutions exist (e.g. \cite{hyland2014best,Presutti2016role}) for making pragmatic decisions about which OR approach should be adopted in the development of a new ontology, and a more comprehensive effort is needed.
The aim of this work is twofold: (i) to provide a comprehensive account of existing solutions, addressing benefits and limits, and (ii) to support developers’ decision-making process by discussing two representative real-world scenarios in the light of general considerations addressed in (i).
The chapter is organized as follows. In Section \ref{approaches} we describe motivations underlying OR approaches and strategies of OI. In Section \ref{challenges} we present the state of the art, benefits, and limits of current practices in ontology selection OS, ontology access OA, ontology integration OI, and OR implementation. In Section \ref{casestudies} we present two case studies adopting different OR approaches. In Section \ref{discussion} we discuss choices made in such contexts in the light of their portability. We conclude with a summary of dimensions to be taken into account when evaluating OR approaches.
\section{Approaches to ontology reuse} \label{approaches}
When bootstrapping a new project, developers search and select candidate ontologies to be reused, and decide strategies for OI. In this section we introduce common motivations guiding OS, policies for OR, and methods for OI.
\paragraph{\textbf{A. Motivations guiding ontology selection}.}
Some guidelines for Linked Open Data and vocabularies publication \cite{hyland2014best,berrueta2008best} support developers in the selection of valid and documented ontologies, by promoting either \emph{top-down agreement} as implemented by standards, or \emph{currency}, by reusing popular ontologies. The designer of an ontology decides that its requirements are fully (or mostly) identical to those that have inspired a standard or a popular ontology.
Research communities \cite{Gruninger1994,gangemi2002sweetening,DBLP:books/daglib/0028799,DBLP:books/ios/HGJKP2016,Presutti2016role} and some communities of practice as with some W3 Consortium recommendations (\url{https://www.w3.org/2001/sw/BestPractices/OEP/}) foster \emph{explicit cognitive analysis} in OR, notably the indirect reuse of ontology patterns rather than ontology terms.
\textbf{OR by standardisation.} OR by standardisation refers to the practice of reusing ontologies issued by authoritative organisations, like ISO, W3 Consortium, and professional or community consortia. Examples include ISO standard ontologies such as CIDOC-CRM \cite{cidoc2017}, W3C endorsed ontologies such as PROV-O \cite{lebo2013prov} or Time Ontology (\url{https://www.w3.org/TR/owl-time/}), and Community standard ontologies such as FRBRoo \cite{riva2017frbroo}. Standard ontologies provide reference models for representing cross-domain information, e.g. events, temporal aspects, provenance, and often provide domain-oriented models meant to allow stakeholders to perform lossless data transformation into RDF.
Standard ontologies are usually recommended for direct reuse to members of a community, or can be intended as reference documents, envisioning specialisations and extensions.
\textbf{OR by popularity.} OR by popularity refers to the practice of reusing ontologies that are popular, typically because they are reused in (i) many third-party ontologies \cite{hyland2014best} by import, specialisation or extension, and in (ii) existing datasets by instantiation of ontology terms \cite{sabou2006ontology}. Popular ontologies might be designed by standardisation bodies, or might include cognitive analysis, but that is probably not the main reason they are reused. For example, FOAF (\url{http://xmlns.com/foaf/spec/}) is the result of an early exemplification (1999) of a vocabulary for the Semantic Web, which did not emerge (did not intend to do so) from any standardisation or cognitive/formal analysis. Nonetheless, it has become a reference for semantic social network data. Another example is Good Relations \cite{hepp2008goodrelations}, which has been designed to address an important interoperability problem in eCommerce. It is not a standardisation result, but is based on formal requirements from industry and cognitive analysis. Other popular ontologies include DBpedia (\url{http://dbpedia.org/ontology/}), the result of collective cleaning of Wikipedia user-based template specifications, and schema.org (\url{https://schema.org/}), an output of an industrial standardisation body fostering search engine optimisation via semantics.
\textbf{OR by cognitive analysis.} OR by cognitive analysis refers to the definition of requirements of an ontology project as primary source of decisions. that may lead to novel constructs, or the usage of existing design components, such as foundational ontologies \cite{gangemi2002sweetening} or ontology patterns \cite{DBLP:books/ios/HGJKP2016}. It is associated with measures to justify the axioms contained in an ontology and supports decisions about the reuse or the novel design of axioms. In particular, whether to reuse something or not is supposed to be dependent from the requirements of an ontology project, both functional and non-functional (e.g. usability, multilingualism).
As aforementioned, cognitive analysis borrows methods from software engineering, philosophy and cognitive science, the most popular including {\em competency questions} (CQs) \cite{Gruninger1994}, formal ontological analysis using patterns already encoded in foundational ontologies, e.g. DOLCE \cite{gangemi2002sweetening}, task-oriented quality assessment \cite{DBLP:conf/esws/GangemiCCL06}, etc. A community has emerged to integrate and support cognitive analysis by means of Ontology Design Patterns (ODP, \url{http://www.ontologydesignpatterns.org}) \cite{Gangemi05,Gangemi2009,DBLP:books/ios/HGJKP2016}. According to~\cite{Gangemi2009}, an ODP is a modelling solution to solve a recurrent ontology design problem. ODPs show certain characteristics, i.e. they are: computational, small, autonomous, hierarchical, cognitively relevant, linguistically relevant, and best practices. Furthermore, an ODP is intuitive and compact, and catches ``core'' notions of a domain. Notions are gathered in sets of CQs, which are the key tool for designing and enabling reuse of ODPs.
\paragraph{\textbf{B. Policies and implementation strategies}.}
Policies for implementing OR take into account motivations and requirements of the project, developers' design choices, and constraints given by the state of the art technologies. We distinguish policies for OR as follows: direct reuse, indirect reuse (as in \cite{Presutti2016role}), and hybrid reuse.
\emph{Direct reuse} may be performed in two ways, namely: (i) the import of ontologies into a new ontology (i.e. by means of the axiom \texttt{owl:imports}); (ii) the inclusion of selected ontology terms in a new ontology (possibly referencing reused ontologies by means of e.g. \texttt{rdfs:isDefinedBy} axioms). In the former case, the semantics of reused ontologies is included into the new one. In the latter, the semantics of reused terms is delegated to external ontologies.
\emph{Indirect reuse} implies that terms from external ontologies are reused as templates, or just aligned, in the new ontology. Terms and their semantics (axioms) are natively described in the new ontology, and aligned to reused terms by means of e.g. \texttt{rdfs:subClassOf} or \texttt{owl:equivalentProperty} axioms.
\emph{Hybrid reuse} is a design choice where ontology terms are selected either for direct reuse or for being indirectly reused as templates, according to characteristics of reused ontologies and requirements of the project.
Actual methods for integrating ontologies depend on the extent of reusable knowledge and the extent of required changes.
When requirements to be addressed by the new ontology are largely satisfied by existing ontologies, the latter can be reused \emph{as-is}. However, some manipulation of existing ontologies in the new setting is likely to be required \cite{Simperl2009}. Possible changes - e.g. extension, adaptation, specialization, change of naming conventions - may be needed due to the heterogeneity or incompleteness of reused ontologies, or the lack of sufficient abstraction or specificity. Moreover, ontology harmonization tasks or orthogonality of terms may be required.
Two methods are applicable, namely: composition (also referred to as integration), and merging (also known as fusion) \cite{pinto2004ontologieshow}.
\emph{Modular composition} is the practice of reusing ontology modules, i.e. sets of terms and axioms that address a specific subset of requirements, into the new ontology. Reused ontologies usually cover different subjects, and are likely to be orthogonal, thus there is little overlap of concepts and semantics. A particular case of composition is when an ontology is built by integrating Ontology Design Patterns (ODPs). In this case, the fragment tackling specific requirements is clearly and formally defined in a dedicated ontology, and has been explicitly designed for reuse.
\emph{Merging} refers to a process where individual concepts, axioms, and statements from reused ontologies are merged together into a new model. In this case, there are several overlapping areas, since the ontologies are on the same subject, and harmonization tasks are required.
While modular composition allows to clearly identify regions or modules reused from source ontologies, merging terms does not allow to immediately understand which source contributes to address a representational issue (e.g. a competency question), as individual concepts may loose or change their semantics in the context of the new ontology.
\section{Ontology reuse: benefits, gaps, and challenges}\label{challenges}
In this section we survey the state of the art of current OR practices and we address benefits and challenges affecting OR decisions. We loosely adopt FAIR principles \cite{wilkinson2016fair} as a framework to characterise aspects that affect available ontologies and OR tasks, namely: (i) selection (relevant to findability of ontologies), (ii) access and preservation (relevant to accessibility), (iii) integration (relevant to interoperability) and (iv) implementation (relevant to reusability).
\subsection{Ontology selection}
\textbf{State of the art.} The number of existing ontologies increased significantly over the last years, and tools to support OS have become a fundamental aid. Existing tools include: (a) catalogues of general purpose ontologies, such as Linked Open Vocabularies (LOV) \cite{vandenbussche2017linked}, vocab.org (\url{http://purl.org/vocab/}), ontologi.es (\url{http://ontologi.es/}), and Joinup platform (\url{https://joinup.ec.europa.eu}); (b) catalogues of domain ontologies, e.g. BioPortal (biomedicine) \cite{whetzel2011bioportal}, SOCoP+OOR (geography, \url{https://ontohub.org/socop}), and SWEET (Earth and environment) \cite{raskin2005knowledge}; (c) catalogues of ODPs \cite{presutti2008content}; (d) semantic web search engines, e.g. Watson \cite{d2007watson}, Falcons \cite{cheng2008falcons}, prefix.cc (\url{http://prefix.cc/}), Swoogle \cite{finin2005swoogle}, and Schemapedia (\url{http://schemapedia.com/}).
Services support users in browsing, filtering and searching over ontologies metadata and terms \cite{dAquin12}. Swoogle and LOV also show ranking scores of ontology terms, based on their popularity across LOD datasets. Moreover, LOV offers users' reviews and insights on the relatedness of ontologies, e.g. imports, extensions or specialisations, to support ontology alignment.
\textbf{Benefits.} Popularity-based metrics are common when searching for ontologies and can avoid time-consuming selection activities when looking for general purpose ontologies. Moreover, reusing popular ontologies increases the chances that data will be reused by other applications~\cite{Heath2011}. Such a practice is very common in the Linked Data community. It is essentially driven by the assumption that the semantics of an ontology is primarily based on the intuitiveness of their names and their popularity in the Web of Data.
Reusing well-known ontologies fosters semantic interoperability and homogenization at Web scale. In fact, re-using well established ontological solutions allows the construction of a shared understanding over the Web knowledge with a decentralised approach.
Such a scenario facilitates rational agents to leverage the Web knowledge and consequently interact at the knowledge level as envisioned by~\cite{Newell1982}. Indeed, building rational agents able to automatically share, reuse, exchange, and reason on the knowledge represented at Web scale is the El Dorado of the Semantic Web Community~\cite{Berners-Lee2001}.
\textbf{Gaps and challenges.} To date, OS is a highly subjective task, often manually performed by experienced ontology engineers, who select ontologies that intuitively fit for the purpose.
Secondly, only popularity-based metrics are available, which may foster biased behaviours in OR practices, such as boosting OR by popularity and OR by standardisation only. A correct OS must rely on clear ontological requirements gathered by analysing the modelling problem and the domain. In literature~\cite{Gruninger1994} ontological requirements are commonly identified by competency questions. The latter are the key tools that enable many ontology design methodologies based on ontology reuse. A notable example is eXtreme Design (XD)~\cite{Blomqvist2016}.
Moreover, internationalisation \cite{Gracia12} and licensing for reuse \cite{Poblet16} should be taken into account, as well as more sophisticated statistics on the co-occurrence of ontologies in third-party ontologies and datasets. The latter would show how ontologies are combined (i.e. by composition or merging), shedding light on orthogonality and harmonisation between ontologies. Detailed information on the extent ontologies are actually reused in other ontologies (e.g. whether single classes, properties or modules) along with information on the application domain, would help to classify reusing ontologies in families of ontologies, assuming that ontologies in a certain domain share commonalities \cite{Ochs17}.
Lastly, clustering datasets where ontologies co-occur would allow to classify ontologies on the basis of their application domain. While statistics on the LOD cloud are available \cite{Ermilov13}, and instance-level relations in data can be discovered \cite{Asprino19}, more sophisticated systems for ontology usage tracking are not available, and consequently supporting ontology selection is still a manual, error-prone activity.
\subsection{Ontology access and preservation}
\textbf{State of the art.} Third parties that reuse ontologies create a dependency on the original vocabulary that includes the semantics of reused terms. Hence, access, maintenance, and long-term preservation of ontologies are critical aspects for the development of new ontologies. Moreover, access methods must address versioning, since ontologies may evolve over time.
Well-known services to ensure persistence of HTTP URIs and correct content negotiation - e.g. purl.org (\url{https://archive.org/services/purl/}), w3id.org (\url{https://w3id.org/}) - and methods to access ontology contents - e.g. dump an OWL file, query a SPARQL endpoint, send requests to APIs - are available. Popular vocabularies such as schema.org or FOAF are currently maintained on version-controlled repositories, e.g. GitHub. Several efforts were made to support ontology versioning in the ontology development phases \cite{garijo2017widoco,halilaj2016vocol}.
In 2012 Poveda-Villalón et al.
\cite{poveda2012landscape} semi-automatically accessed and analyzed 256 ontologies included in LOV registry, showing that 23\% of surveyed ontologies were not available. In 2018 Fèrnandez-Lopez et al. \cite{fernandez2019ontologies} performed a similar investigation, showing that the percentage increased to 36\%.
\textbf{Benefits.} Relying on reused ontologies delegates to third-party ontology engineers the problem of dealing with ontology preservation, versioning, storage and evolution. If an ontology engineer opts for the reuse of ontologies that result from well established initiatives (e.g. FOAF, Schema.org) they might count on large communities to cope with both maintenance and preservation, along with the compliance with standards (e.g. serialisation syntaxes, modelling languages, etc). The latter are part, as the ontologies themselves, of the evolutionary process that affects any human artifact. For example, new serialisation formats may evolve or arise. Similarly, new logical patterns might be included in the reference modelling language (e.g. punning has been introduced in OWL 2, but it was not available in OWL 1).
\textbf{Gaps and challenges.} Aforementioned surveys show that there is an urge for vocabulary publishers to ensure long-term availability of ontologies and allow programmatic processing by catalogues. However, there is no standard solution to ensure long-term preservation of vocabularies on the Web \cite{baker2013requirements}. Free of charge proprietary repositories are currently the cheapest solution for sharing source code, but they lack some features that systems dedicated to distributed vocabulary development usually provide, e.g. integrated support for modeling, population and testing \cite{halilaj2016vocol}. Moreover, maintenance and availability are ensured by suppliers only, arising trust issues. As a matter of fact, ontologies produced by small enterprises and scholars in the context of short-term research projects are likely to suffer long-term preservation issues.
\subsection{Ontology integration}
\textbf{State of the art.} OI methods ensure different degrees of interoperability between reused and reusing ontologies depending on policies, design choices, and whether the process is manually or automatically performed.
Several frameworks for semi-automatic ontology aggregations have been proposed, such as architectures for automating ontology generation through ontology reuse \cite{lonsdale2010reusing}, ranking models for evaluating ontologies based on semantic matching \cite{park2011ontology}, as well as compatibility metrics for comparing and integrating ontologies \cite{trokanas2016ontology}, frameworks for ontology integration based on knowledge graphs integration and machine learning techniques to identify core terms to be integrated \cite{zhao2014ontology}, and methods for building ontologies from ontology patterns and vice-versa \cite{ruy2017reference}. Ontology integration based on ontology matching has been largely investigated in literature~\cite{Euzenat2011,Euzenat2013} and many solutions exist at the state of the art, such as~\cite{Jain2010,David2011}.
\textbf{Benefits.} Ontology matching relieves ontology engineers of the cognitive effort of identifying alignments manually and speeds up the ontology modelling process. Additionally, it allows ontology engineers who are not domain experts to design ontologies by dealing with fine-grained terminologies wrt the domain.
The ontology design process might benefit of multiple integrations. For example, most of the solutions at the state of the art associate more than one correspondence with external ontologies for each class or property the ontology engineer aims at integrating within the target ontology.
Finally, the requirement of reusing well established and largely adopted ontologies is transparently addressed by following this approach. In fact, ontology matching algorithms might be fed with repositories and catalogues referencing specific target ontologies that address the latter requirement.
\textbf{Gaps and challenges.} Integration activities may be error-prone. Ambiguities, inconsistencies, and heterogeneity in existing ontologies may affect results at different dimensions, such as syntactic heterogeneity, terminological heterogeneity, conceptual heterogeneity, and semiotic heterogeneity \cite{Euzenat2013}. Gangemi and Pisanelli \cite{gangemi1998ontology} previously described terminological issues of local ontologies in terms of semantic imprecision (e.g. relation range violation), ontological opaqueness (e.g. lack of motivation for choosing a certain predicate), and awkward linguistic naming conventions. Hence, human intervention is likely to be required not only in the search and the selection of ontologies, but also in the disambiguation and formalisation of integrated ontologies.
Secondly, ontology integration is not a one-time task. It may need to be applied several times, since integrated ontologies may change over time \cite{pinto2004ontologieshow}. According to the literature, it is not clear yet whether it is more cost-effective to build a new ontology from scratch that perfectly meets current needs than to try to rebuild and adapt existing ontologies.
\subsection{Ontology reuse implementation}
\textbf{State of the art.} Ontologies resulting from OR practice may reference and acknowledge original ontologies in several ways \cite{poveda2012landscape}, such as (1) importing the entire ontology, (2) reference ontology terms by means of annotation properties, or (3) acknowledging reused ontology in external alignment documents.
In the empirical survey conducted by Schaible et al.
\cite{schaible2014survey} on 79 LOD practitioners, the authors show that direct reuse of popular vocabularies is preferred than defining new terms and align to other vocabulary terms. In \cite{fernandez2019ontologies} the authors compare results of their analysis on OR trends in LOV ontologies to the previous work done by \cite{poveda2012landscape}. They show that W3C endorsed ontologies are reused with different nuances in 78\% of ontologies catalogued in LOV catalogue. More important, they notice that the evolution in OR practices from direct import of ontologies to the direct reuse of ontology terms has increased until approximately a half of the total re-users. OR practices in domain applications confirm the same trend. Experiments in the biomedical domain showed that OR between ontologies from BioPortal is limited, and presents a number of antipatterns \cite{kamdar2017systematic}. \cite{Ochs17} shows that the same ontologies available in BioPortal adopt several nuances of direct reuse, including specializations and extensions.
\textbf{Benefits.} On the one hand, \texttt{owl:imports} axiom is natively supported by OWL and its semantics is clear. Additionally, to the best of our knowledge, most ontology development frameworks (e.g. Prot\'{e}g\'{e}) support the axiom.
On the other hand, the use of annotation properties, e.g. \texttt{rdfs:isDefinedBy}, introduces more flexibility with respect to modularity (i.e. it is possible to declare which parts of an external ontology are reused rather than importing the whole ontology).
More recently, a few solutions have been proposed for implementing OR by relaxing the monolithic constraint of \texttt{owl:imports} and preserving formal semantics at the same time. For example, OPLa~\cite{Hitzler17opla} defines classes and properties for declaring which modules or patterns are reused within a target ontology. OPLa is supported by Prot\'{e}g\'{e} by means of a dedicated plug-in~\cite{Shimizu2018}.
\textbf{Gaps and challenges.} On the one hand, \texttt{owl:imports} does not allow to customise the import of ontologies (i.e. it only allows to import ontologies as a whole).
As reported in \cite{hammar2012}, being \texttt{owl:imports} transitive, concepts included in unintentionally imported ontologies may be irrelevant to and incompatible with the requirements of the local ontology.
Moreover, when importing ontologies into a new one, it may happen that the reused ontologies are not available over time, and their semantics is lost.
On the other hand, the flexibility given by annotation properties corresponds to less semantic rigour. As a matter of fact, none of the built-in annotation properties is explicitly meant for ontology reuse. That is, their adoption for implementing reuse does not address any formal semantics. Hence, none of the existing platforms for ontology design fully support them. In fact, when only terms belonging to external ontologies are imported in a new ontology, reasoning on the reused ontology constraints is not possible when reasoning on the reusing one. Secondly, if a reusing ontology includes contradictions with respect to a reused ontology, these cannot be automatically detected. Only when both ontologies are imported in a third ontology, contradictions may emerge \cite{fernandez2019ontologies}. Likewise, when external documents including alignments are provided, the same situation may happen if adopters do not include alignments in the reusing ontology.
\section{Case studies}\label{casestudies}
In this section we introduce two representative use cases exemplifying respectively direct reuse and hybrid reuse approaches.
\subsection{Direct reuse of ontologies: the OpenCitations Data Model}
The OpenCitations Data Model (OCDM) \cite{peroni2018opencitations} is based on the OpenCitations Ontology (OCO, \url{https://w3id.org/oc/ontology}), which aggregates terms from existing well-known ontologies. It was initially developed in 2016 by the OpenCitations organisation to describe data in the OpenCitations Corpus (\url{http://opencitations.net/corpus}). In recent years OpenCitations has developed other datasets, and the OCDM is currently adopted by several external projects that contribute to the growth of the model. The current version of OCDM takes such changes into account and describes a generic bibliographic dataset, making easier its adoption by third parties. Fig. \ref{oco-fig1} shows classes and properties of OCO.
\textbf{Ontology development methodology.}
The first version of the OCDM released in 2016 was developed by directly reusing (i.e. merging) terms from existing models. The latter include the SPAR Ontologies, a set of modular and orthogonal ontologies developed by using SAMOD \cite{peroni2016simplified}, an agile data-driven methodology for ontology development.
Within the context of a project recently funded by the Wellcome Trust, SAMOD was adopted to extend OCO with terms relevant to generic bibliographic datasets. The outcome includes motivating scenarios, competency questions, and a glossary of terms of all the new entities included in OCO.
\textbf{Ontology selection.} OCO reuses selected terms from popular domain ontologies, i.e. the SPAR Ontologies \cite{peroni2018spar}, and W3 endorsed ontologies, such as the Web Annotation Ontology (OA) \cite{sanderson2013open}, PROV-O \cite{lebo2013prov}, and VoID \cite{alexander2011describing}. The choice is motivated by several pragmatic factors, such as: ontology designers' background knowledge with respect to reused ontologies, the fact that potential stakeholders of OpenCitations datasets already use the same ontologies, and, in most of cases, the possibility to directly modify reused ontologies rather than creating new terms.
\begin{figure}
\includegraphics[width=\textwidth]{ocdm.png}
\caption{Overview of OCDM main classes and properties} \label{oco-fig1}
\end{figure}
\textbf{Ontology access and preservation.} OCDM is released under CC-BY license. Both OCO and reused SPAR ontologies are developed and maintained by the OpenCitations organisation. Long term preservation of such artefacts is ensured by the OpenCitations organisation, that has been recently selected by the Global Sustainability Coalition for Open Science Services (SCOSS, \url{https://scoss.org/}) for its funding programme that guarantees long term sustainability. Accessibility of other reused ontologies is ensured by W3 Consortium.
\textbf{Ontology integration.} The integration process of ontological terms gathered from aforementioned ontologies led to the creation of a new single ontology (i.e. OCO) with the purpose of aggregating only the terms that are relevant to the representation of motivating scenarios. Few extensions and specialisations of reused ontologies were necessary. Whenever applicable, changes were directly applied in the reused ontologies, e.g. new terms were added to SPAR Ontologies. Only in one case an initial request\footnote{See \url{https://lists.w3.org/Archives/Public/public-openannotation/2019Sep/}} for the addition of a new property to a W3 endorsed ontology (i.e. OA) was made. Meanwhile, a bespoke property \texttt{oco:hasAnnotation} has been created and declared as the inverse property of \texttt{oa:hasTarget}.
\textbf{Ontology reuse implementation.}
As aforementioned, OCO directly reuses existing ontology terms rather than importing ontologies. Provenance of reused terms is recorded by means of \texttt{rdfs:isDefinedBy} statements.
\subsection{Hybrid reuse of ontologies and patterns: the ArCo ontologies}
ArCo (Architecture of Knowledge, \url{https://w3id.org/arco}) is a project dedicated to the publication of the open knowledge graph (KG) of the Italian cultural heritage (CH) \cite{Carriero2019}, deriving from the General Catalogue of Italian Cultural Heritage maintained by the Italian Institute of the General Catalogue and Documentation (ICCD).
As part of the project, several ontologies for representing data about Italian cultural heritage were developed. A wide and complex domain as CH easily leads to a large ontology, and ArCo aims at modelling with a fine grain a wide range of concepts relating to 30 different types of cultural properties; this motivates the choice of a network of ontology modules, as thematically coherent sub-areas of the domain addressing a subset of requirements, rather than a monolithic design. A preliminary manual clustering of metadata in sets of topics resulted in 7 modules. The \texttt{arco} module (\url{https://w3id.org/arco/ontology/arco/}) is the root node of the network, addressing top-level distinctions and a general taxonomy, while the \texttt{core} module models general concepts, reused by the other modules (e.g. location, cultural events). Ontology modules are connected by means of import axioms.
\textbf{Ontology development methodology.} In order to cope with a huge and diverse domain, and to minimize the impact of changes in its incremental releases, ArCo adopts an agile and iterative, pattern-based and test-driven ontology development methodology, called eXtreme Design (XD) \cite{Blomqvist2016}.
The project's ontological commitment is represented by general constraints and competency questions (CQs). Requirements are provided in the form of small user stories, as scenarios and real use cases, by a growing community of customers, adopters and consumers via a Google Form, GitHub issues, mailing-list and meetups. This supports ArCo in continuously collecting new emerging requirements and extending the customer team beyond its main customer, i.e. ICCD.
These user stories are analysed and transformed into competency questions and constraints, and used for selecting suitable ODPs.
The iterative design process goes in parallel with testing and validation of ontology components against requirements, by verifying CQs coverage and model consistency.
\textbf{Ontology selection.} Modules mainly reuse ontology patterns available in the Ontology Design Patterns (ODPs) catalogue, e.g. Time-Indexed Situation, Sequence: indeed, ODPs play a central role in ArCo's design and are recommended by XD as small reusable solutions. Other ontologies have been selected because of their level of standardisation or community adoption, e.g. CIDOC-CRM, Cultural-ON, OntoPiA, or as extensions of the latter, e.g. OAEntry.
\textbf{Ontology access and preservation.} Both ontologies and data are distributed by ICCD (\url{https://github.com/ICCD-MiBACT/ArCo}) under a CC-BY 4.0 (Attribution-ShareAlike) license. A SPARQL endpoint and additional materials (software, documentation, test suit) are available.
Ontologies directly reused by ArCo are developed and maintained by the same team (ISTC-CNR).
\textbf{Ontology integration.} The ontology integration process led to the creation of a new set of ontologies, which integrate entities, fragments and patterns from existing ontologies and ontology design patterns.
Being ArCo an evolving project, it directly reuses only ontologies that are considered reference standards by the Italian Government, evolve rather slowly and involve ArCo's team.
When ontology terms are not directly reused, manual alignments have been performed so as to record which parts matched the project's ontological commitment.
\textbf{Ontology reuse implementation.}
Ontology terms are first designed according to the set of analysed requirements, in order to primarily formalise the semantics of the domain to be represented and address all requirements with the apt level of expressiveness. Secondly, terms from the resulting ontology are either refactored or aligned to terms of existing ontologies.
\begin{figure}
\includegraphics[width=\textwidth]{direct-indirect3.png}
\caption{Overview of ArCo ontology reuse approaches} \label{arco-fig1}
\end{figure}
Specifically, ArCo directly reuses Cultural-ON (\url{http://dati.beniculturali.it/cis/}), and OntoPiA (\url{https://w3id.org/italia/onto/FULL/}), a network of ontologies for top-level and cross-domain information.
Other ontologies and Ontology Design Patterns are reused as templates.
Alignments to external ontologies are encoded in separate files, along with documentation. Reused ODPs are annotated with the ontology design pattern annotation language OPLa (\url{http://ontologydesignpatterns.org/opla/})
which allows to express that new patterns are, for instance, a specialisation of existing patterns.
An example of OR policies is shown in Figure~\ref{arco-fig1}.
In detail, terms from OntoPiA ontology, e.g. the class \texttt{tiapit:TimeInterval} and its related properties, are imported in the model. The class \texttt{a-cat:CatalogueRecord} is defined, in a separate file\footnote{\url{https://github.com/ICCD-MiBACT/ArCo/blob/master/ArCo-release/ontologie/catalogue/catalogue-aligns.owl}}, as a subclass of the class \texttt{fabio:EntityMetadata} from the FaBiO ontology.
Likewise, terms from the Sequence pattern are reused as template for new terms, e.g. \texttt{a-cat:isPreviousVersionOf}. The alignment between the new pattern and the original pattern is recorded in files annotated by means of OPLa (i.e. \texttt{opla:specializationOfPattern}).
\section{Discussion and conclusion}\label{discussion}
In this section we discuss aspects that can support the selection of the best OR approach to be adopted at the beginning of a project. We discuss approaches adopted in the two use cases in the light of benefits and limits illustrated in Section \ref{challenges}.
Secondly, we motivate these ontology reuse approaches by using the list of attributes proposed by \cite{hammar17}, namely: implementation complexity, external reasoning dependencies, content overhead, reused content modifiability.
Finally, we draw conclusions on decision criteria and limits of this work.
\textbf{OS.} While OCDM mainly relies on popular and W3 endorsed ontologies, ArCo mainly relies on community standards and ontology design patterns. In both cases the designers have performed cognitive analysis based on a set of competency questions.
In OCDM, existing ontologies were deemed satisfying for representing the publishing domain. Stakeholders producing similar data agree on reusing the same ontologies for the task. In ArCo, direct reuse of existing ontologies presented obstacles, such as different levels of abstraction, as in CIDOC-CRM, axiomatic paucity, as in EDM, or limits in naming conventions. In this case, stakeholders (ArCo designers, the Italian Ministry of Cultural Heritage and the larger community involved), cannot afford direct reuse.
\textbf{OA.} In both scenarios, adopters preferred to rely either on W3 endorsed standards or on ontologies developed and maintained by the same organisation that developed the new ontology. As a matter of fact, trust in third-party ontology suppliers is a fundamental aspect in OS. Secondly, both projects aim at producing and maintaining knowledge graphs that are representative of a given domain, and their long-term preservation and trustworthiness are ensured by the institutions issuing those technologies.
\textbf{OI.} OCDM is based on OCO, a unique ontology including terms from diverse ontologies. ArCo is based on a network of ontology modules, with new terms using a new URI scheme, and appropriate alignments to existing ontologies when the semantics is the same, or more general/specific.
In the OCDM scenario, the modularity of reused ontologies is privileged, allowing adopters that are confident with well-known existing ontology definitions to retrieve faster information they seek for. Moreover, OCDM comes with a human-readable documented data model, where extensive data examples are provided as an aid. In the ArCo scenario, developers privilege the usability of new term definitions, which natively correspond to the intended conceptualisation and are under the control of the stakeholders, thus preventing any semantic shifting in external ontologies. Also in this case, there is ample documentation, Docker containers on GitHub to locally test or use ArCo data, and extensive data and query examples.
\textbf{OR.} OCDM aims at representing the publishing domain by directly reusing established ontologies, while ArCo attempts to grasp the broader cultural heritage domain that is differently represented in a vast amount of scattered projects. In the first scenario, the knowledge to be represented is stable, meaning that bibliographic data are usually factual, structured pieces of information and their representational requirements are clear. In the second scenario, ArCo is an open project: knowledge to be integrated keeps growing, and may change over time (both at instance and schema level), hence a tight control on the evolution of its semantics is needed.
\paragraph{Attributes of OR approaches.}
Wrt the list of attributes in \cite{hammar17}, OCDM OR policy requires a low implementation effort, by directly reusing ontology terms; on the contrary, ArCo mostly adopts a more expensive approach, that is copying the structure of the reused parts into the target ontology with new IRIs. While the ArCo approach prevents from resolving external references to derive all intended inferences, also supporting content modifiability, direct reuse depends on external ontologies and generally does not allow us to modify reused content: nevertheless, this does not affect OCDM consistency and refactoring, since external reasoning is not required and the evolution of reused ontologies is under control of the OCDM team. Finally, both OCDM and ArCo have low content overhead, since only relevant content is (in)directly reused.
\paragraph{Conclusive remarks.} We reckon few dimensions in OR practices were not completely addressed in previous works. In particular, four dimensions can contribute to define a shareable set of criteria for the decision about OR approaches. The dimensions can be summarised as follows:
\begin{itemize}
\item \textbf{Task-based OS.} Competency questions as a measure to acquire (positive or negative) requirements are currently the main intuitive, effective instrument to prevent bad design choices. This is widely accepted as a precondition for specifying conceptualisation in the form of ontologies. As such, a task-based OS guided by competency questions can justify OR approaches, followed by other common practices and motivations for OS. In addition, the community working on ontology design patterns has established good practices in matching competency questions to reusable modeling solutions, so enabling a pragmatic cognitive analysis.
\item \textbf{Trustworthiness of OA solutions.} Political and social argumentations influence OR, such as trustworthiness of ontology suppliers that ensure OA. Such aspects apply to suppliers of both reused and reusing ontologies, and need to be pushed at the forefront of semantic technologies, as ontology design is both engineering and social negotiation.
\item \textbf{Usability of OI results.} A trade-off between previous knowledge of ontology adopters, usability of ontology documentation, and means to facilitate query of data created according to new ontologies must be found so as to motivate OI strategies and results.
\item \textbf{Stability of knowledge that implements OR.} Domain knowledge keeps growing and evolving, and OR practices are inevitably affected. Clear evidence of the current state of the knowledge domain should be addressed at the beginning of a project in order to justify OR implementation strategies. Here a raised awareness of cognitive analysis practices used in ontology design projects is a major priority to enhance semantic interoperability.
\end{itemize}
Decision approaches and metrics should be provided in order to effectively support ontology designers when considering the dimensions discussed. Some of those aspects are not easily measurable, and cannot be automatically detected. We strongly believe that ``non-functional'' requirements are fundamental for the design of reusable ontologies and should be always addressed in ontology design methodologies, recommendations and evaluation.
\bibliographystyle{splncs04}
|
\section{Introduction}
In the field of modern celestial mechanics and dynamical astronomy, one of the most intriguing and important problems is the dynamics of few bodies, in particular being the case of a circularly restricted three body problem \citep{Szebehely67}. This problem has been applied in various fields in astronomy, like planetary dynamics, galactic and stellar cluster dynamics and even molecular dynamics. Currently, with the advent of LIGO and the detection of gravitational waves from binary black hole mergers \citep{Abbott2016a, Abbott2016b}, the investigation of such systems in strong gravitational fields have become a field of intense research once again. The black holes involved in these discoveries span a mass range of 10 M$_\odot$ to 100 M$_\odot$, and are all consistent to have initially formed from the death of massive stars.
There is also strong observational evidence that a different class of super massive black holes (SMBHs), with masses ranging from 10$^5$ to 10$^{10}$ M$_\odot$, are residing in almost all centres of galaxies \citep{Beckmann12}. It is expected that some of these SMBHs will pair up as binaries as their host galaxies merge \citep*{Begelman80}. In fact, there is ample evidence of several active galaxies with double nucleus \citep{Komossa03, Muller15}. It is also speculated that the eventual inspiral and merger of some of these SMBH binaries constitutes a prime gravitational wave source for the planned LISA observatory \citep{Amaro-Seoane12}. In addition, there is also increasing evidence that there are Kerr black hole binaries which are merging \citep{Gwak19, Ruffini16, Beckmann12} and are sources of gravitational radiation.
In such binary black hole mergers which also accrete, the investigation of the chaotic dynamics of test particles within accretion discs or inside the halo surrounding these compact objects has become a subject of prime importance \citep*{Levin00, Schnittman01, Cornish02, Cornish03, Hartl05, Gopakumar05, Wu07, Wu08, Wu10, Wu11, Wu15a, Wu15b, Zhong10a, Wang11, Li19, Mei13, Luo17, Huang14a, Huang14c, Huang16}. Some authors have also studied the numerical schemes and techniques which can be used for such non-linear, chaotic problems along with the dynamics of these systems \citep{Zhong10b, Wu15b, Luo17}. Investigations of such dynamics of charged particles moving under the influence of magnetic and strong gravitational fields of a single compact object have already been studied in some detail within the general relativistic framework \citep*{Kopacek10, Kopacek14, Kopacek15, Takahashi09, Kovar08, Kovar10}. Such studies have been extended to the motion of test particles under the influence of the relativistic gravitational field of accreting black holes \citep*{Semerak10, Semerak12, Semerak13, Witzany15, Vogt03} and also for motion under the influence of gravity produced by an extended body \citep*{Letelier97, Vieira99, Wu06a, deCastro11}.
On the other hand, escaping particles from dynamical systems has also been a subject under focus for decades. Especially the issue of escape in Hamiltonian systems is directly related to the problem of chaotic scattering which has been an active field of research over the last decade and it still remains an open area \citep*{Benet98, Benet96, Bleher90, Bleher89, Bleher98, Churchill75, Contoupoulos90, Contoupoulos92, Eckhardt88, Motter02, Ott93, Seoane06}. It is well known that some types of Hamiltonian systems have a finite energy of escape. For lower values of the particle energy, the equipotential surfaces of these systems are closed and therefore escape is impossible. For energies above the escape energy, these surfaces open and exit channels emerge through which particles can escape to infinity. There is a comprehensive body of work on such ``open'' or ``leaking'' Hamiltonian systems \citep*[e.g.][]{Barrio09, Contopoulos12, Ernst14, Kandrup99, Lai11, Navarro01, Siopis95a, Siopis95b, Siopis96, Zotos14a, Zotos14b, Zotos15a, Zotos15b, Zotos16a, Zotos17a}. However, it is needless to say that this list of citations is neither complete nor exhaustive. It is just indicative of the body of work that has happened in these fields and is still continuing.
The restricted three-body problem (RTBP) is an excellent example of such open Hamiltonian systems with escape \citep*[e.g.][]{Winter94a, Winter94b}. Over the last few decades, a large number of studies have been devoted to the classification of orbits in the RTBP. It all started with the pioneering works of \citet*{Nagler04, Nagler05} where initial conditions of orbits were classified as bounded, escaping or collisional. Moreover, bounded orbits were further classified into orbital families by taking into account the type of motion of the test particle around the primary sources. Such classifications have also been done in the context of planetary systems, Earth-Moon system and Saturn-Titan system \citep*{deAssis14, Zotos16a}. In this context, it is important to mention that a simplified modification of the RTBP is the Hill approximation which focuses on the vicinity of the secondary source \citep*[e.g.][]{Hill86, Petit86, Petit87, Steklain06, Steklain09}. This facilitates for the study of the motion of test particles in the neighborhood of the Lagrange (equilibrium) points $L_1$ and $L_2$. At this point it should be mentioned that the Hill approximation is valid only when the mass of the secondary is much smaller than the mass of the primary body. One can directly obtain the Hill model from the classical RTBP by translating the origin to the center of the secondary body and also by re-scaling the coordinates suitably. The Hill problem was proved to be non-integrable by \citet*{Meletlidou01}, and is chaotic, as shown by \citet*{Simo00}. Subsequently, thorough numerical investigations of this problem were performed by carrying out a systematic classification of the initial conditions of the orbits \citep*{Zotos17a}. More precisely, the initial conditions of the orbits were classified into four categories: (i) non escaping regular orbits; (ii) trapped chaotic orbits; (iii) escaping orbits; and (iv) collisional orbits. In addition, the issue of equilibrium points in circular restricted three body problem (CRTBP) has also been studied widely and in great detail (see \citet*{Henon97} and references there in). The discovery of the Trojan asteroids around the Lagrangian points L$_4$ and L$_5$ in the Sun-Jupiter system \citep*{Murray99}, and the recent observations of asteroids around L$_4$ for the Sun-Earth system \citep*{Connors11}, has added a great impetus to theoretical studies on the subject. Moreover, the dynamics of non-conservative RTBP have also been investigated extensively, like the case of CRTBP with gravitational radiation \citep*{Schnittman10}, an elliptic restricted three-body problem \citep*{Wang16} and that of a dissipative CRTBP with drag forces \citep*{Wang18}.
One of the first attempts at studying the relativistic CRTBP under the assumptions of low velocities and weak gravity was made by \citet*{Krefetz67} in the year 1967. He looked at the post-Newtonian equations for the first time using the Einstein-Infeld-Hoffmann (EIH) formalism \citep*{Einstein38}. Since then this problem has been studied by several authors where they have attempted to present the Lagrangian explicitly \citep*{Contopoulos76}. Some authors have also tried to explore the deviations to the Lagrangian points due to the post-Newtonian corrections \citep*{Maindl96}. In addition, analytical solutions were also attempted in the GR regime using the EIH approximation up-to the first order \citep*{Yamada10}. Recently, as one of the first studies of chaotic orbits in the post-Newtonian CRTBP, \citet*{Huang14a} explored the influence of the distance of separation between the two primaries. They observed that if the primary bodies are close enough, the post-Newtonian dynamics is qualitatively quite different, particularly where some Newtonian bounded orbits become unstable.
In more recent studies, several authors have formulated this problem using pseudo-Newtonian potentials developed for non-rotating Schwarzschild-like (Paczy\'{n}sky-Witta potential) \citep*{Paczynski80} and rotating Kerr-like primaries \citep*{Artemova96, Semerak99, Mukhopadhyay02} to avoid the complications of a post-Newtonian formulation. Subsequently, detailed studies of orbits and the dynamics of test particles around a single Schwarzschild primary and a binary system, as well as Kerr like primaries have been made in recent years with the idea of investigating the chaotic and unstable nature of orbits in the relativistic regime. In a very recent study, \citet*{Dubeibe16} used the Fodor-Hoenselaers-Perj\'es (FHP) procedure \citep*{Fodor89} (taking into account the corrections made by \citet*{Sotiriou04}) to derive an approximate potential for the gravitational field of two uncharged spin-less particles modeled as sources with multi-pole moment, $m$. In this work, they have explored the dynamics of a massless test particle using the Poincar\'{e} section and the Lyapunov exponent as indicators of chaos. As they have mentioned, this potential is not ad-hoc as other pseudo-Newtonian potentials but rather it is exactly derived from the multipolar structure of the sources. In our current study, we also follow a similar route and use the FHP procedure to derive the multipolar structure of a spinning binary system. Subsequently, we construct a beyond-Newtonian potential to imitate the gravitational effects of this system on a test particle in the CRTBP scheme.
The paper is organized as follows. In the next section, we present the formulation of the gravitational beyond-Newtonian potential of each Kerr-like source using the FHP procedure. Next, we present the Lagrangian and the equations of motion of a test particle in context to CRTBP. In the subsequent section, we present a detailed analysis of the Hill curves or the zero velocity surfaces as the system makes a gradual transition from the FHP beyond-Newtonian approximation to the classical regime through a parameter $\epsilon$ in the beyond-Newtonian potential. Here we also present a detailed analysis of the orbits and a discussion on the fixed points of this system along with their stability as a function of the parameter $\epsilon$. The classification of the nature of orbits is made using Poincar\'e surfaces of section and the variational method for the calculation of the largest Lyapunov exponent, as done by several previous authors. In the next section, we present a comparison between the dynamics of a test particle around a binary system of Schwarzchild and spinning primaries. Finally, in the last section we conclude with a summary of our main results and present certain new directions that we intend to investigate in the near future.
\section{Formulation of Beyond-Newtonian Potential for Kerr Binary}
The version of CRTBP we consider consists of two massive, spinning primaries with masses $\mathcal{M}_1$ and $\mathcal{M}_2$ and intrinsic angular momenta $a_1$ and $a_2$, at positions $X_1$ and $X_2$, respectively, describing a circular orbit in the $x-y$ plane about their common centre of mass (taken to be the origin $\mathcal{O}$). The centre-to-centre distance remains fixed and remains sufficiently far apart, while the orbital angular velocity is $\omega_0$. The aim is to set up the beyond-Newtonian potential (up to the first non-Newtonian term) for this CRTBP system and consequently write down the Euler-Lagrange equations of motion of a test particle under the influence of this potential. The schematic of the configuration is illustrated in figure (\ref{diagram}).
To simulate the dynamics of the CRTBP at hand, we employ the Fodor-Hoenselaers-Perj\'es (FHP) procedure to perform an expansion in the mass and rotation potential of each primary up to the first non-Newtonian term. This essentially generates first-order general relativistic effects, the dynamics of which is analyzed at length in the following section. The beyond-Newtonian potential for the system is then constructed by virtue of a superposition of the potentials corresponding to the two primaries, modelled to describe circular orbits around their common centre of mass. We then write down the Lagrangian and consequently the equations of motion for a test particle under the influence of such a potential.
\subsection{Beyond-Newtonian potential}
\label{sec:PNpot}
We shall now briefly outline the steps involved in the FHP procedure leading to the construction of the beyond-Newtonian potential for the problem at hand. The FHP algorithm involves the decomposition of the Einstein field equation in the so-called Ernst formalism. In this formalism, the field equations of GR are reduced to a pair of complex equations by virtue of introducing the complex potentials $\zeta$ and $\Psi$. These complex potentials are further defined in terms of two new potentials $\xi$ and $\varsigma$ through the relations
\begin{align}
\zeta = \frac{1 - \xi}{1 + \xi} \,, \quad \Psi = \frac{\varsigma}{1 + \xi} \,.
\label{1.01}
\end{align}
The field potentials satisfy \citep*{Ernst68I, Ernst68II}
\begin{align}
(\xi \xi^* - \varsigma \varsigma^* - 1) \, \nabla^2 \xi = 2 (\xi^* \nabla \xi - \varsigma^* \nabla \varsigma ) \cdot \nabla \xi \,, \label{1.02} \\
(\xi \xi^* - \varsigma \varsigma^* - 1) \, \nabla^2 \varsigma = 2 (\xi^* \nabla \xi - \varsigma^* \nabla \varsigma ) \cdot \nabla \varsigma \,.
\label{1.03}
\end{align}
The above set of equations are an alternative representation of the Einstein-Maxwell field equations. As a matter of fact, they could be interpreted as the generalization of Laplace's equation for the Papapetrou's metric describing the space-time around a stationary and axisymmetric source
\begin{align}
ds^2 = - F (dt - \omega d\phi)^2 + F^{-1} [e^{2 \gamma}(d \rho^2 + dz^2) + \rho^2 d \phi^2] \,,
\label{1.04}
\end{align}
where the metric coefficients $F$, $\omega$, and $\gamma$ depend only on the Weyl-Papapetrou co-ordinates $\rho$ and $z$. These metric functions can be reformulated in terms of the Ernst complex potentials \citep*{Sotiriou04} $\zeta$ and $\Psi$ and described by the associated Einstein-Maxwell field equations (\ref{1.02}) and (\ref{1.03}). \\
The new set of field potentials $\xi$ and $\varsigma$ are related to the classical gravitational and electromagnetic potentials in the following way
\begin{align}
\xi = \Phi_M + i \, \Phi_J \,, \quad \varsigma = \Phi_E + i \, \Phi_H \,,
\label{1.04}
\end{align}
where $\Phi_M, \Phi_J, \Phi_E,$ and $\Phi_H$ represent the mass, angular momentum, electrostatic and magnetic potentials, respectively. As our massive, spinning primaries do not possess electromagnetic fields, we set $\Phi_E = \Phi_H =0$, which from (\ref{1.01}) implies $\varsigma = \Psi = 0$. The seminal work of \citet*{Geroch70} and \citet*{Hansen74} allows us to determine the multipolar moments of asymptotically flat spacetimes. In this prescription, the induced 3-metric $h_{i j}$ is mapped by virtue of a conformal transformation $h_{i j} \rightarrow \tilde{h}_{i j} = \Omega^2(x) h_{i j}$ onto a conformal metric $\tilde{h}_{i j}$. This conformal factor $\Omega$ satisfies the conditions
\begin{align}
\Omega\big|_{\Lambda} = \tilde{D}_i \Omega\big|_{\Lambda} = 0,\quad \tilde{D}_i \tilde{D}_j \Omega\big|_{\Lambda} = 2 h_{i j}\big|_{\Lambda},
\end{align}
where $\tilde{D}$ denotes the covariant derivative on the induced surface and $\Lambda$ denotes the point added due to conformal compactification. Essentially, $\Omega$ transforms the potential $\xi$ into $\tilde{\xi} = \Omega^{-1/2} \xi$ with the explicit transformation being $\Omega = r^{\prime 2} = \rho^{\prime 2} + z^{\prime 2}$. The relation between the primed and unprimed Weyl-Papapetrou coordinates are
\begin{align}
\rho^{\prime} = \frac{\rho}{\rho^2 + z^2},\quad z^{\prime} = \frac{z}{\rho^2 + z^2} \,,
\end{align}
with $\phi$ remaining unchanged. This helps in mapping the infinity to the origin of the primed coordinates $(\rho^{\prime}, z^{\prime}) = (0,0)$. Besides, the potential $\tilde{\xi}$ can be expressed as a power series expansion in $\rho^{\prime}$ and $z^{\prime}$ as
\begin{align}
\tilde{\xi} = \sum_{i, j=0}^{\infty} a_{i j} \rho^{\prime\, i} z^{\prime\, j}
\end{align}
with the coefficients $a_{i j}$ determined by recursive relations presented explicitly in \cite{Sotiriou04}. Following this procedure, one can deduce approximate relations for the gravitational potential $\xi$, in terms of the parameters of the primary once its gravitational multiple moments $P_i$ are known. Thus, we apply this outlined prescription to a massive, spinning primary whose multipolar structure we take to be:
\begin{align}
P_0 = m\,, \quad P_1 = i m a\,, \quad P_i = 0 \,\quad \textrm{for} \,\, i \geq 2 \,,
\end{align}
such that $m$ and $a$ denote the mass and angular momentum of the source, respectively.
We now aim to set up the beyond-Newtonian potential (up-to the first non-Newtonian term) for the CRTBP system at hand and consequently write down the Euler-Lagrange equations of motion of a test particle under the influence of this potential. For clarity, we restate the conditions and assumptions of the CRTBP model we are trying to construct:
\begin{itemize}
\item The two primaries, with masses $\mathcal{M}_1$ and $\mathcal{M}_2$ and intrinsic angular momenta $a_1$ and $a_2$, at positions $X_1$ and $X_2$, respectively, describe a circular orbit about their common centre of mass (taken to be the origin $\mathcal{O}$). The centre-to-centre distance $r$ remains fixed and remains sufficiently far apart, while the orbital angular velocity is $\omega_0$.
\item A beyond-Newtonian potential describing the primaries is constructed assuming that the principle of superposition holds: that the total gravitational potential of the system is a linear sum of the mass and rotation potentials (up to first order effects) of the individual sources.
\item A test particle of mass $\mathcal{M}$, that is very small compared to the primaries, now moves under the effect of this beyond-Newtonian potential in the $z = 0$ orbital plane of the primaries. The motion of this test particle has no effect on the primaries whatsoever.
\item The convention $G = \mathcal{M} = \omega_0 = r = 1$ is used throughout the analysis hereon (further details on this choice of units has been discussed extensively in section (\ref{sec:dyn})).
\end{itemize}
In accordance with the above conditions and following the preceding discussion on the FHP formalism, we now construct the beyond-Newtonian potential $\Omega$ describing the primaries of our CRTBP model from the reconstructed potential $\xi$ describing a single source. Keeping explicitly the factors of $c$ to show the corresponding order-wise contributions, we have the beyond-Newtonian potential for our system:
\begin{align}
\Omega (x,y) &= -\sum_{i=1}^{2}\frac{\mathcal{M}_i}{r_i} + \frac{1}{2c^4} \sum_{i=1}^2 \frac{\mathcal{M}_i^3}{r_i^3} \nonumber \\
&+ \frac{1}{c^2}\sum_{i=1}^{2} \frac{\mathcal{M}_i a_i}{r_i^2} \cos \theta_i + \frac{1}{2c^4}\sum_{i=1}^2 \frac{\mathcal{M}_i a_i^2}{r_i^3}(3\cos^2 \theta_i - 1)
\label{pnp}
\end{align}
where the primaries are stationed at positions $X_1 = (x_1,0)$ and $X_2 = (x_2,0)$ respectively, and $r_{1,2} = \sqrt{(x - x_{1,2})^2 + y^2}$. We note that the first two terms of equation (\ref{pnp}) describe the mass potential and the next two terms represent the rotation potential of the binary system upto first order corrections respectively. Also, following the FHP procedure, we see the potential that is constructed is written in terms of powers of $1/c^2$. The 1st order corrections to the Newtonian potentials, both in mass as well as for rotation, are retained and the higher order terms are dropped since their contribution is smaller compared to the leading order (by appropriate factors of $1/c^2$). The term `beyond-Newtonian' is designated to these 1st order corrections to the Newtonian potentials that arise in our final form of the potential, as seen in equation~(\ref{pnp}).
Moreover, in order to observe the transition of the system from the Newtonian regime to a beyond-Newtonian one, we introduce a parameter $\epsilon$, such that,\[\frac{1}{c^2}\rightarrow\frac{1}{c^2}\epsilon\]
with $\epsilon\in[0,1]$ using the fact that $\frac{1}{c^2}\rightarrow0$ reduces equation~(\ref{pnp}) to the Newtonian case. That is, the $\epsilon = 0$ classical limit is essentially the Newtonian problem that models non-spinning binaries composed of weak gravitational sources as found in say, planetary systems and binary stars which are not in close contact with each other. On the other hand, the $\epsilon = 1$ beyond-Newtonian case models departures from Newtonian behaviour that can be found in compact spinning binaries constituted of strong gravitational sources, for example black-hole and compact binaries. The parameter $\epsilon$ can thereby be thought of as a knob that slowly ``turns on'' corrections (both in the mass and rotation potentials as seen from equation~(\ref{pnp})) to the Newtonian potential as we gradually go from the classical limit $\epsilon = 0$ to the beyond-Newtonian regime $\epsilon = 1$.
\section{Dynamics of a test particle}\label{sec:dyn}
In order to simplify the numerical simulation of the three dimensional system described in the section above, we confine ourselves to the plane of the two primaries. We adopt a modified version of the Szebehely convention to de-dimensionalize the problem. Numerous types of scaling transformations have had applications in literature \citep*{Huang14a, Huang14b, Su16}. For example, studies of chaotic dynamics of asteroids in planetary systems scale primaries to the solar mass. However, for our problem, the absolute masses of the two primaries are irrelevant and do not reveal any new physical information about the system. Therefore, with $\mathcal{M}_1+\mathcal{M}_2=\mathcal{M}$ and $a_1+a_2=a$, we define a dimensionless mass $\mu_1=\mathcal{M}_2/\mathcal{M}$ and dimensionless spin $\mu_2=a_2/a$. Applying the scaling relations described above, we enforce the sum of the masses of the two primaries and the distance between the two to be unity. This has been enforced by adopting geometrized units, $G=1$ and $c=1$, with distance and time now having the dimension of mass (this choice of units has been discussed in detail in the next paragraph). Additionally, this scaling also ensures that the sum of the spins of the primaries be unity. Thus, applying the above discussed scaling relations we obtain:
\begin{equation}
\begin{array}{lcl}
\mathcal{M}_1=1-\mu_1&;&\mathcal{M}_2=\mu_1 \\
a_1=1-\mu_2&;&a_{2}=\mu_2 \\
\end{array}
\end{equation}
At this point, it is worthwhile to note that different system of units have been used in literature for simplifying the respective problem, both analytically and numerically. The choice of units always mostly depend on the length scales, masses and the time-scales involved. As a result of this, the speed of light $c$ can assume different values. For example, in planetary systems, setting $G=1$, the unit of mass to be the sum of masses, the unit of distance to be the semi-major axis of the secondary body (which is set to unity) and using Kepler's second law, the speed of light assumes different values like $c$ = 22946.5 for the case of Sun-Jupiter, and $c$ = 10065.3 in the case of Sun-Earth \citep*{Lhotka15}. However, while studying the dynamics of test particles around compact objects under the circular restricted three body scheme (CRTBP) in post-Newtonian (PN) treatments \citep{Einstein38}, the speed of light $c$ surfaces as a parameter which measures the order of the PN contributions. For ease in numerical simulations, $c$=1 is later enforced and $a$, which is the separation between the parent bodies, becomes an important parameter for the first post-Newtonian (1-PN) order effect. Thus, this choice of unit and relevant scaling transformations facilitates the study of how the separation between the primaries affect the dynamics of this system \citep*{Huang14a}. Another variation to this post-Newtonian three body scheme was recently studied by \citet*{Dubeibe17} who used $c=10000$ in his calculations. To show this, one can use the Sun-Earth system as an example (for details refer to \citet*{Klavcka08}). It was shown here that this value of $c$, the choice of units and relevant scaling transformations, as opposed to $c=1$ in an earlier work by \citet*{Huang14a}, facilitates a better conservation of the Jacobi integral of motion numerically. This is due to the fact that the contributions of the higher order PN terms vary depending on the formulation and thus a truncation brings about a non-conservation of the Jacobi integral (discussed in detail later). Recently, \citet*{Deng20} used different values of $c$ to indicate perturbations from the PN contribution, which were used to find an optimal method for the calculation of eccentric anomaly.
However, in our study, the Jacobi integral of motion is a constant. Thus, our choice of the value of $c$ is to just facilitate the simplification of the system, both algebraically and numerically. As we had noted earlier, the beyond-Newtonian effects are scaled by a factor of $1/c^2$ which is taken care of by the introduction of the parameter $\epsilon$ in our system of units. Hence, $c=10000$ will scale down the beyond-Newtonian terms by a factor of $10^{-8}$, which can be compensated by suitably adjusting the range of $\epsilon$, since it is a free parameter in our system. Thus it can be concluded that the nature of the dynamics of the system will not be affected by the choice of the value of $c$, as has been verified by our simulations too.
The separation between the two primaries is then scaled as,
\begin{equation}
\begin{array}{lcl}
x_{1}=-\mu_1&;&x_{2}=1-\mu_1
\end{array}
\end{equation}
\begin{figure}
\includegraphics[scale=0.7]{diagram.png}
\caption{The configuration of the two primaries, $\mathbf{\mathcal{M}_1}$ and $\mathbf{\mathcal{M}_2}$, in the centre-of-mass frame which is rotating about the z-axis with angular frequency $\mathbf{\omega_0}$ ($=1$). A test particle with infinitesimal mass $\mathbf{P}$ is placed at an arbitrary position in the equatorial plane.}
\label{diagram}
\end{figure}
Thus, $\mu_{1},\mu_{2}\in[0,\frac{1}{2}]$ are the only two control parameters for the system. Applying the earlier described scaling and putting $c=1$, the potential becomes:
\begin{align}
\Omega(x,y)=&-\bigg(\frac{1-\mu_1}{r_1}+\frac{\mu_1}{r_2} \bigg)+\frac{1}{2}\epsilon^2\bigg(\frac{(1-\mu_1)^3}{r_1^3}+\frac{\mu_1^3}{r_2^3} \bigg) \nonumber \\
&+\epsilon\bigg(\frac{(1-\mu_1)(1-\mu_2)}{r_1^2}\cos\theta_1+\frac{\mu_1\mu_2}{r_2^2}\cos\theta_2\bigg) \nonumber \\
& +\frac{1}{2}\epsilon^2\bigg\{\frac{(1-\mu_1)(1-\mu_2)^2}{r_1^3}\bigg(3\cos^2\theta_1-1\bigg) \nonumber \\
& +\frac{\mu_1\mu_2^2}{r_2^3}\bigg(3\cos^2\theta_2-1\bigg)\bigg\}
\end{align}
The Lagrangian for the system may be constructed as follows:
\begin{equation}
\mathcal{L}=\frac{V^2+2A+R^2}{2}-\Omega(x,y)
\label{lag}
\end{equation}
where $V=\sqrt{\dot{x}^2+\dot{y}^2}$ represents the magnitude of the velocity of the test particle, $R=\sqrt{x^2+y^2}$ the position of the test particle with respect to the centre of mass in the non-inertial rotating frame and $A = \dot{y} x - \dot{x} y$. Thus, the Euler-Lagrange equations of motion are:
\begin{align}
\ddot{x} &= 2 \dot{y}+x-\bigg[\frac{(1-\mu_{1})}{r_{1}^3}\bigg( x+\mu_{1} \bigg)+\frac{\mu_{1}}{r_{2}^3}\bigg(x+\mu_{1}-1\bigg)\bigg] \nonumber \\
& -\epsilon\bigg\{ \frac{(1-\mu_{1})(1-\mu_{2})}{r_{1}^4}\bigg[y\sin\theta_{1}-2\cos\theta_{1}\bigg(x+\mu_{1}\bigg)\bigg] \nonumber \\
& +\frac{\mu_{1}\mu_{2}}{r_{2}^4}\bigg[y\sin\theta_{2}-2\cos\theta_{2}\bigg(x+\mu_{1}-1\bigg) \bigg] \bigg\} \nonumber \\
& -\frac{3}{2}\epsilon^2\bigg\{\frac{(1-\mu_{1})(1-\mu_{2})^2}{r_{1}^5}\bigg[y\sin2\theta_{1}-\bigg(3\cos^2\theta_{1}-1\bigg) \nonumber \\
& \bigg(x+\mu_{1}\bigg) \bigg] -\frac{(1-\mu_{1})^3}{r_{1}^5}\bigg(x+\mu_{1}\bigg)+\frac{\mu_{1}\mu_{2}^2}{r_{2}^5}\bigg[y\sin2\theta_{2} \nonumber \\
& -\bigg(3\cos^2\theta_{2}-1\bigg)\bigg(x+\mu_{1}-1\bigg) \bigg]-\frac{\mu_{1}^3}{r_2^5}\bigg(x+\mu_{1}-1\bigg) \bigg\}
\label{eqn:eqnOfMotionfor_x}
\end{align}
\begin{align}
\ddot{y} =& -2\dot{x}+y\bigg[1-\bigg( \frac{1-\mu_{1}}{r_{1}^3}+\frac{\mu_{1}}{r_{2}^3} \bigg)\bigg]+\epsilon\bigg\{\frac{(1-\mu_{1})(1-\mu_{2})}{r_{1}^4} \nonumber \\
& \bigg[\bigg(x+\mu_{1}\bigg)\sin\theta_{1}+2y\cos\theta_{1} \bigg]+\frac{\mu_{1}\mu_{2}}{r_{2}^4 } \bigg[\bigg(x+\mu_{1}-1\bigg) \nonumber \\
& \sin\theta_{2}+2y\cos\theta_{2} \bigg]\bigg\} +\frac{3}{2}\epsilon^2\bigg\{ \frac{(1-\mu_{1})(1-\mu_{2})^2}{r_{1}^5}\bigg[\bigg(x+\mu_{1}\bigg) \nonumber \\
& \sin2\theta_{1}+\bigg(3\cos^2\theta_{1}-1\bigg)y\bigg] +\frac{(1-\mu_{1})^3}{r_{1}^5}y+\frac{\mu_{1}\mu_{2}^2}{r_{2}^5} \nonumber\\
& \bigg[\bigg(x+\mu_{1}-1\bigg)\sin2\theta_{2}+\bigg(3\cos^2\theta_{2}-1\bigg)y \bigg] +\frac{\mu_{1}^3}{r_{2}^5}y\bigg\}
\label{eqn:eqnOfMotionfor_y}
\end{align}
where,
\begin{align*}
r_{1}&=\sqrt{(x+\mu_{1})^2+y^2} \\
r_{2}&=\sqrt{(x+\mu_{1}-1)^2+y^2}\\
\theta_{1}&=\tan^{-1}[y/(x+\mu_{1})] \\
\theta_{2}&=\tan^{-1}[y/(x+\mu_{1}-1)]
\end{align*}
The Jacobi integral for the above system is given by,
\begin{equation}
J(x,y,\dot{x},\dot{y})=(x^2+y^2)-2\Omega(x,y)-(\dot{x}^2+\dot{y}^2)=C_{j}
\label{jacobian}
\end{equation}
where $C_{j}$ is a constant of motion for the given system and is called the Jacobian constant.
Here we note that the Lagrangian for our system, as stated in equation~(\ref{lag}), has terms only up to the quadratic order in velocity $V$ of the test particle as a result of which the Jacobian constant (equation (\ref{jacobian})) is exactly derived. This is in contrast to the post-Newtonian (PN) framework where the Jacobian does not remain conserved and consequently limits the extent of dynamical studies. The reasoning behind this has to do with the relations between the PN Lagrangian and Hamiltonian approaches at the same PN order. Additionally, it also depends on the relations between the approximately truncated as well as the exactly non-truncated Euler-Lagrange equations of motion for this PN Lagrangian approach. The equivalence between the Lagrangian and Hamiltonian approaches at the same PN order was established in \citet*{Damour01, Damour02, deAndrade01, Levi14}. However, recent contradictions of the same have been discussed in \citet{Wu15a, Wu15b}; \citet*{Wang15, Chen16}; \citet{Huang16}. It has been shown by \citet{Li19, Li20} that the approximately truncated Euler-Lagrange equations of motion for this PN Lagrangian approach have different dynamical behaviours of order and chaos than its exactly non-truncated counterpart. As a result, the reasons why the Jacobian constant cannot be conserved in the PN approach is because (a) some higher-order PN terms are truncated when the Euler-Lagrange equations of motion are derived from this PN Lagrangian approach, and (b) some higher-order PN terms are still truncated when the Hamiltonian (corresponding to the Jacobian constant) is derived from this PN Lagrangian approach. If the truncated higher-order PN terms are large, as in the case of strong gravitational fields of compact objects, the Jacobian constant shows a poor accuracy; while it shows a better accuracy if the same truncated terms are comparatively smaller, as in the case of weak gravitational fields found in our Solar system. It should be expected that for our potential (equation (\ref{pnp})), the Lagrangian and Hamiltonian approaches at the same beyond-Newtonian order are not equivalent in general. This inequivalence should also be true for the approximately truncated as well as the exactly non-truncated Euler-Lagrangian equations of motion for our beyond-Newtonian Lagrangian approach. However, the equations of motion (\ref{eqn:eqnOfMotionfor_x}) and (\ref{eqn:eqnOfMotionfor_y}), the corresponding Hamiltonian and the Jacobian constant (\ref{jacobian}) are exactly derived and have no terms truncated from the beyond-Newtonian Lagrangian (equation (\ref{lag})) because it has no higher-order terms with respect to the test particle velocity $V$. As a result, the Jacobian constant, given by equation (\ref{jacobian}), is said to be exactly derived.
\begin{figure*}
\centering
\includegraphics[scale=0.5]{mu1__mu2_0001_HillCurve.png}
\caption{Evolution of the Hill curves for $\mu_{1}=\mu_{2}=0.001$ and $C_{j}=4.0$ with the parameter $\epsilon$. The white regions of the plot represents the points in the X-Y plane are energetically allowed, while the dotted regions are energetically disallowed, for the test particle whose Jacobian $C_{j}=4.0$. The larger black dot on the left represents the position of the mass $\mathcal{M}_1$ and the smaller black dot on the right represents the position of mass $\mathcal{M}_2$ in each of the plots.}
\label{fig:HillCurve0001WithEpsilon}
\includegraphics[scale=0.5]{mu1__mu2_05_HillCurve.png}
\caption{Evolution of the Hill curves for $\mu_{1}=\mu_{2}=0.5$ and $C_{j}=4.0$ with the parameter $\epsilon$. The white regions of the plot represents the points in the X-Y plane are energetically allowed, while the dotted regions are energetically disallowed, for the test particle whose Jacobian $C_{j}=4.0$. The black dot on the left represents the position of the mass $\mathcal{M}_1$ and the black dot on the right represents the position of mass $\mathcal{M}_2$ in each of the plots.}
\label{fig:HillCurve05WithEpsilon}
\end{figure*}
\subsection{Hill Curves}
The Hill curves or the zero-velocity curves for the system, for a set of chosen values of $C_{j}$, $\mu_{1}$, $\mu_{2}$ and $\epsilon$, divide the equatorial plane into regions where the motion of the body is energetically allowed and regions where the motion is energetically disallowed (for a discussion on zero-velocity curves refer to \citet*{Szebehely63} for a Newtonian CRTBP system and \citet*{Zotos18a} for a pseudo-Newtonian CRTBP with Schwarzschild like primaries). All points, where $(x^2+y^2)-2\Omega(x,y)-(\dot{x}+\dot{y}) > C_{j}$, are energetically allowed for the test particle, while all points, where $(x^2+y^2)-2\Omega(x,y)-(\dot{x}+\dot{y}) < C_{j}$ are energetically disallowed. The velocity of the test particle (as we shall calculate from equations (\ref{eqn:initial_conditions})) in the disallowed region is imaginary (will be calculated explicitly in the next subsection). The Hill curves of the system have an equation,
\begin{equation}
(x^2+y^2)-2\Omega(x,y)=C_{j} \,.
\end{equation}
Figures (\ref{fig:HillCurve0001WithEpsilon}) and (\ref{fig:HillCurve05WithEpsilon}) show the evolution of the Hill curves with the introduction of beyond-Newtonian effects for $\mu_{1}=\mu_{2}=0.001$ (or the biased-mass system) and $\mu_{1}=\mu_{2}=0.5$ (or the Copenhagen system) respectively. The beyond-Newtonian effects are introduced by increasing $\epsilon$ from $0.0$ to $1.0$ in steps of $0.1$. The equatorial plane is divided into three regions by the Hill curves -- a central region where the particle is energetically allowed but is bounded by the Hill curves, an unbounded energetically allowed region, and a disallowed region in-between them. Test particles with initial conditions in the unbound region may execute stable orbits around both the primaries or may escape to infinity, while test particles with initial positions in the enclosed and energetically allowed regions are `trapped' and cannot escape to infinity since they cannot cross the Hill curves. The energetically allowed regions are represented by white in figure (\ref{fig:HillCurve0001WithEpsilon}) and figure (\ref{fig:HillCurve05WithEpsilon}), while the dotted regions are energetically disallowed for the test particle. The two black dots represent the positions of the primaries $\mathcal{M}_1$ and $\mathcal{M}_2$ respectively.
For the biased-mass system, the potential due to mass $\mathcal{M}_1$ dominates the Hill curves. The introduction of beyond-Newtonian effects distorts the curves of the Newtonian system, such that for all values of $\epsilon\gtrsim0.0865$, no trapped circular orbits exist. For the Copenhagen system, the chosen value of $C_j$ corresponds to the energy at the first Lagrange point $L_1$. As $\epsilon$ increases, the contribution of the spin becomes apparent and the enclosed allowed region becomes smaller. Circular trapped orbits around both the primaries exist for small values of $\epsilon$. For $\epsilon>0.1248$, circular orbits no longer exist around the primary $\mathcal{M}_2$ while for $\epsilon>0.134$, circular orbits no longer exist around the primary $\mathcal{M}_1$. Thus, for both systems, we choose our initial conditions in the unbounded energetically allowed region for the sake of consistency of initial conditions for all values of $\epsilon$, $\mu_{1}$ and $\mu_{2}$.
\subsection{Orbits}\label{sec:orbits}
Using the six stepped, fifth-order Runge-Kutta method implemented with the Dortmund-Prince algorithm, the equations of motion equations (\ref{eqn:eqnOfMotionfor_x},\ref{eqn:eqnOfMotionfor_y}) are integrated using time step $\tau=10$ for $n=3000$ iterations. For a preliminary investigation of the system, the following initial conditions are considered (similar to \citet{Dubeibe16} which investigates orbits for a system with Schwarzschild like primaries): $x_{0}=[3.0,3.5,3.75,4.0,4.25,4.5,4.75,5.0,5.25,5.5,6.0]$, $y_{0}=0.0$ and $\dot{x}_{0}=0.0$, with $C_{j}=4$. The value of $\dot{y}_{0}(x_{0},y_{0},\dot{x}_{0})$ is calculated from the following equation:
\begin{align}
\dot{x}_{0}&=\frac{y_{0}}{r_{0}}\sqrt{(x^2+y^2)-2\Omega(x,y)-C_{j}} \nonumber \\
\dot{y}_{0}&=-\frac{x_{0}}{r_{0}}\sqrt{(x^2+y^2)-2\Omega(x,y)-C_{j}}
\label{eqn:initial_conditions}
\end{align}
where $r_{0}=\sqrt{x_{0}^2+y_{0}^2}$.
The orbits for a test particle for the biased mass and Copenhagen systems are investigated for $\epsilon\in[0,1]$ and the set of initial conditions mentioned in the paragraph above. Since the system is conservative, the Jacobi constant $C_{j}$ has to remain constant as the equations of motion are integrated.
The integrator used, being non-symplectic in nature, usually does not conserve the Jacobian. The use of such integrators for conservative systems have been well studied and multiple corrective methods, such as the velocity correction method (\citet*{Ma08}; \citet{Wang16, Wang18, Deng20}), have been developed for better accuracy. In Figure (\ref{velCorr0001}), we have shown a comparison of the relative error in the Jacobi constant $C_j$ with time for both the non-corrected and velocity corrected integrators. It is observed that the accuracy in the conservation of $C_j$ for the velocity corrected method ranges from $10^{-16}$ -- $10^{-14}$ for stable orbits and goes up to $10^{-8}$ for chaotic and sticky orbits at large times ($> 5 \times 10^3$ years), as has been pointed out in \citet{Wang16, Wang18}. We also observe that our non-corrected integrator has a fairly similar accuracy at the start. However, the growth in error is faster at late times and reaches values of $10^{-10}$ for stable orbits and goes up to $10^{-8}$ for chaotic and sticky orbits. Hence, we conclude that the non-corrected fifth-order Runge-Kutta method is also of reasonable accuracy for the relevant time-periods of our investigation.
\begin{figure*}
\begin{tabular}{ccc}
\includegraphics[scale=0.22]{velCorr1_2.png} &\includegraphics[scale=0.22]{velCorr2_2.png}&\includegraphics[scale=0.22]{velCorr3_2.png}
\end{tabular}
\caption{Plot of the log of the relative error in the Jacobi constant $C_j$ with log of time for the non-corrected Runge-Kutta (4,5) integrator using the Dormand-Prince algorithm (black) and velocity-corrected 4th order Runge-Kutta integrator (grey). The first plot from the left ($\mu=0.001$, $\epsilon=0.6$, $x_0=3.0$, $y_0=0.0$) shows the evolution of $C_j$ with time for a sticky initial condition, the centre plot ($\mu=0.5$, $\epsilon=0.1$, $x_0=5.0$, $y_0=0.0$) shows the evolution of $C_j$ with time for a stable initial condition, and the plot on the right ($\mu=0.5$, $\epsilon=0.7$, $x_0=5.5$, $y_0=0.0$) shows the evolution of $C_j$ with time for a chaotic initial condition.}
\label{velCorr0001}
\end{figure*}
By observing their evolution, the orbits may be categorized as regular, sticky or escaping. Orbits are said to be sticky if they show regular behavior for a long period of time before their chaotic nature manifest \citep*{Dvorak99} and escaping if the particle directly escapes from the system without executing any regular orbits \citep{Contoupoulos90, Contoupoulos92}. We classify the stability of the initial conditions based on the number of iterations for which the orbit of the particle is stable. If the test particle executes stable orbits for 3000 iterations, it is classified as regular. If the orbits are stable for at-least 100 iterations before they escape from the system, they are classified as sticky. If the test particle reaches a distance of 50 times the separation between the two primaries within 1000 iterations, they are said to be escaping. The third column of the table in Appendix (\ref{tab:Details0001}) and Appendix (\ref{tab:Details05}) records the type of orbit for the test particle given some initial conditions for the biased-mass and Copenhagen system respectively.
For the biased mass system, among the initial conditions considered, orbits for $x_{0}=[3.5,3.75,4.0,4.25,4.5]$ are stable for all values of $\epsilon$. Most initial conditions are either sticky or escaping for non-zero values of $\epsilon$. But the interesting initial conditions are the ones where the intermediate values of $\epsilon$ are the most chaotic. The initial conditions $x_0=[5.0,5.25,5.5,6.0]$ show such behavior. For the Copenhagen system, orbits for $x_0=[3.5,3.75,4.0,4.5,4.75]$ are stable for all values of $\epsilon$. The initial condition $x_0=4.25$ destabilizes for $\epsilon>0.4$, implying a region of chaotic initial conditions interjects stable initial conditions in the phase space. Orbits for $x_0=[5.0,5.25]$ are either sticky or escaping for all values of $\epsilon$ except $\epsilon=[0.0,0.1]$. A stable orbit for $x_0=5.5$ exists only for $\epsilon=0.0$, while no stable orbits exist for $x_0=6.0.$ for any value of $\epsilon$. This implies that regions of initial conditions allowing stable orbits shrink as $\epsilon$ increases for the Copenhagen system.
Contrary to expectation, $\epsilon=1.0$ does not result in the maximum number of sticky and escaping initial conditions in either of the systems. Instead, the intermediate values of $\epsilon$ have the most number of unstable initial conditions. For the biased mass system, $\epsilon=[0.2,0.4,0.6,0.7,0.8]$ have the least number of stable initial conditions, namely 6 out of the 11 investigated. $\epsilon=0.9$ has the least number of stable initial conditions for the Copenhagen system, namely 5 out of the 11 investigated. In contrast, $\epsilon=1.0$ has 9 and 6 initial conditions out of 11 for the biased mass and Copenhagen systems, respectively.
\begin{figure*}
\begin{tabular}{ccc}
\includegraphics[scale=0.26]{final_0001_0.png}&\includegraphics[scale=0.26]{final_0001_03.png}&\includegraphics[scale=0.26]{final_0001_05.png} \\
\includegraphics[scale=0.26]{final_0001_07.png}&\includegraphics[scale=0.26]{final_0001_1.png}&
\end{tabular}
\caption{Locations of some of the equilibrium points of the biased-mass system ($\mu_1=\mu_2=0.001$), marked by grey squares, on the intersection of $\partial\Omega/\partial x=0$, marked by the dashed line, and $\partial\Omega/\partial y=0$, marked by the solid line, for $\epsilon=[0.0,0.3,0.5,0.7,1.0]$. For $\epsilon\neq0$, there are two non-collinear equilibrium points on either sides of the more massive primary which could not be shown on the plots due to their close proximity to it. The smaller primary has three more collinear equilibrium points, one of which lies between the two primaries. These too could not be marked on the plots due to their proximity to the primary.}
\label{fixedPoint0001}
\end{figure*}
\begin{figure*}
\begin{tabular}{ccc}
\includegraphics[scale=0.26]{final_05_0.png}&\includegraphics[scale=0.26]{final_05_03.png}&\includegraphics[scale=0.26]{final_05_05.png} \\
\includegraphics[scale=0.26]{final_05_07.png}&\includegraphics[scale=0.26]{final_05_1.png}&
\end{tabular}
\caption{Locations of some of the equilibrium points of the Copenhagen system ($\mu_1=\mu_2=0.05$), marked by gray squares, on the intersection of $\partial\Omega/\partial x=0$, marked by the dashed line, and $\partial\Omega/\partial y=0$, marked by the solid line, for $\epsilon=[0.0,0.3,0.5,0.7,1.0]$. For $\epsilon\neq0$, there are two non-collinear equilibrium points on either sided of both primaries which could not be shown on the plots due to their close proximity to them.}
\label{fixedPoint05}
\end{figure*}
\subsection{Fixed Points and their stability}
For a system having equilibrium points (fixed points), the necessary and sufficient conditions are:
\begin{equation}
\dot{x}=\dot{y}=\ddot{x}=\ddot{y}=0
\end{equation}
Thus, the co-ordinates of the co-planar fixed points are determined by solving the following pair of partial differential equations (refer to equation~(\ref{pnp}) for the detailed expression of $\Omega(x,y)$) \citep*{Strogatz94}:
\begin{equation}
\frac{\partial\Omega(x,y)}{\partial x}=\frac{\partial\Omega(x,y)}{\partial y}=0
\label{omegaPartial}
\end{equation}
The intersection of the curves for equations~(\ref{omegaPartial}) for a set of values of $\mu_1$, $\mu_2$ and $\epsilon$ gives us a set of fixed points for the system. Figures~(\ref{fixedPoint0001}) and~(\ref{fixedPoint05}) show the positions of the fixed points for $\epsilon=[0.0,0.3,0.5,0.7,1.0]$ for the biased-mass and Copenhagen systems respectively. It is evident that for both systems, the number of fixed points is highly dependent on the value of $\epsilon$, the details of which are enlisted in separate tables in Appendices~(\ref{tab:FixedPnts0001}) and (\ref{tab:FixedPnts05}). A summary of the salient features of the fixed points with respect to $\epsilon$ is presented below:
\begin{itemize}
\item For $\epsilon=0$, both the biased-mass and Copenhagen systems reduce to their Newtonian counterparts. These systems have five fixed-points each, as expected.
\item For $\epsilon=0.3$, the biased-mass system has five fixed points while the Copenhagen system has nine.
\item For $\epsilon=0.5$, the biased mass system has nine fixed points while the Copenhagen system has thirteen. The less massive primary in the biased system has three collinear fixed points.
\item For both $\epsilon=0.7$ and $\epsilon=1.0$, the biased mass system has five fixed points. Only the collinear fixed points in either of the systems is beyond the less massive primary. However, the Copenhagen system has nine equilibrium points for both $\epsilon=0.7$ and $\epsilon=1.0$.
\item Finally, the more massive primary in the biased mass system as well as both the primaries in the Copenhagen system have two non-collinear equilibrium points very near to it for values of $\epsilon\geq 0.3$ (not shown in Figures (\ref{fixedPoint0001}) and (\ref{fixedPoint05}) since they fall very close to the primaries).
\end{itemize}
It is thus evident that the number of equilibrium points for both the biased-mass and the Copenhagen systems become maximum at intermediate values of $\epsilon$.
Now, moving on to the issue of stability of these fixed points, their linear stability may be determined
by Taylor expanding the system's equations of motion around the fixed point ($x_0\,\text{,} y_0$) upto first order. In the perturbation equations, the time-independent coefficient matrix of variations is identified as
\begin{equation}
A=\begin{bmatrix}
0& 0& 1& 0\\
0& 0& 0& 1\\
\frac{\partial^2\Omega_0}{\partial x^2}&\frac{\partial^2\Omega_0}{\partial x \partial y} &0 &2 \\
\frac{\partial^2\Omega_0}{\partial y \partial x}& \frac{\partial^2\Omega_0}{\partial y^2} & -2& 0
\end{bmatrix}
\label{charMatrix}
\end{equation}
where the subscript $0$, attached to the partial derivatives of second order of $\Omega$, denotes evaluation at the position of the equilibrium point ($x_0\,\text{,} y_0$). The necessary and sufficient condition that a fixed point is stable is that all the eigenvalues of matrix $A$ be purely imaginary. Applying this method to the fixed points, obtained by numerically solving equations~(\ref{omegaPartial}), we can conclude the following:
\begin{itemize}
\item For $\epsilon=0$, the collinear fixed points for both the biased-mass and the Copenhagen systems are unstable while the triangular fixed points are stable.
\item For $\epsilon=0.3$, none of the fixed points are stable for the biased mass system while two fixed points are stable for the Copenhagen system.
\item For $\epsilon=0.5$, one fixed point is stable for the biased mass system while two are stable for the Copenhagen system.
\item For $\epsilon=0.7$, no fixed point is stable for the biased mass system while one is stable for the Copenhagen system.
\item For $\epsilon=1.0$, no fixed point is stable for either of the systems.
\end{itemize}
Thus, the evolution and stability of the fixed points of the system under consideration show non-trivial evolution with the parameter $\epsilon$. However, the knowledge about the basins of convergence along with the libration points is of prime importance since the attracting domains reflect some of the most intrinsic properties of the dynamical system. This has been a topic of intense research in recent years for many different dynamical systems such as the Hill problem \citep*{Douskos10}, the four-body problem \citep*[e.g.][]{Baltagiannis11, Kumari14, Zotos17c} and the pseudo-Newtonian planar circular restricted three body problem \citep*{Zotos17b, Zotos17d, Zotos18b}. We plan to investigate these aspects for our beyond-Newtonian potential in detail as part of our future work.
\section{Poincar\'e Map of Section}
\label{sec:PS}
The Poincar\'e map, or the first return map, is a powerful and conventional tool for examining the motion of dynamical systems \citep{Tabor89, Parker89, Dubeibe16}. In order to construct the map, we evolve the system for 3000 iterations in time-steps of $\tau=10$ and plot the section of the orbit for $y=0.0, \dot{y}<0$. This is done for 11 initial conditions
$x_0=[3.0,3.5,3.75,4.0,4.25,4.5,4.75,5.0,5.25,5.5,6.0]$ and $y_0=0.0$ while increasing $\epsilon$ from 0.0 to 1.0 in steps of 0.1. Figure (\ref{fig:PS_0.001}) and Figure (\ref{fig:PS_0.5}) show the evolution of the Poincar\'e map for the biased-mass and and Copenhagen systems respectively. The Poincar\'e map of a system primarily shows two types of structures: concentric Kolmogorov-Arnold-Moser (KAM) tori which represent bounded, quasi-periodic motions and a sea of chaotic points surrounding such tori. At the centre of each island of concentric tori is a point which corresponds to a stable, periodic and resonant orbit \citep*{Gidea07, Broer10, Huang14a}. The extent of the sea of scattered points in comparison to islands of the tori provides a visual representation of the extent of chaos in the system.
For the biased-mass system, the initial conditions $x_{0}=[3.5,3.75,4.0,4.25,4.5]$ show KAM tori on their Poincar\'e maps for all values of $\epsilon$, implying quasi-periodic orbits. For $x_{0}=[4.75,5.0,5.25,5.5,6.0]$, the destruction of their KAM tori implies chaotic or sticky orbits, as was observed in subsection (\ref{sec:orbits}). For the Copenhagen system, the Poincar\'e maps $x_0=[3.5,3.75,4.0,4.5,4.75]$ show KAM tori for all values of $\epsilon$. The torus for $x_0=3.0$ breaks up only for $\epsilon=0.9$, while the tori for $x_0=5.5$ and $x_0=[5.0,5.25]$ break up for $\epsilon>0.0$ and $\epsilon>0.1$ respectively. No KAM tori appear for $x_0=6.0$ for any value of $\epsilon$, implying that the initial condition is chaotic for all values of $\epsilon$. Thus, the Poincar\'e maps for both the systems corroborate the observations presented in subsection (\ref{sec:orbits}).
\begin{figure*}
\begin{tabular}{ccc}
\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_0001_e=0_3000.png}&\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_0001_e=1_3000.png}&\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_0001_e=2_3000.png} \\
\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_0001_e=3_3000.png}&\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_0001_e=4_3000.png}&\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_0001_e=5_3000.png} \\\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_0001_e=6_3000.png}&\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_0001_e=7_3000.png}&\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_0001_e=8_3000.png} \\
\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_0001_e=9_3000.png}&\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_0001_e=10_3000.png}&\\
\end{tabular}
\caption{The Poincar\'e map for orbits with $x_0=[3.0,3.5,3.75,4.0,4.25,4.5,4.75,5.0,5.25,5.5,6.0]$ and $y_0=0.0$ as $\epsilon$ increases from 0 to 1 in steps of 0.1. The system is evolved for 3000 iterations in time-steps of $\tau=10$ and the Poincar\'e map of section for $y=0.0, \dot{y}>0.0$ is plotted for all 11 initial conditions for the biased-mass system ($\mu_{1}$=$\mu_{2}=0.001$).}
\label{fig:PS_0.001}
\end{figure*}
\begin{figure*}
\begin{tabular}{ccc}
\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_05_e=0_3000.png}&\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_05_e=1_3000.png}&\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_05_e=2_3000.png} \\
\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_05_e=3_3000.png}&\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_05_e=4_3000.png}&\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_05_e=5_3000.png} \\\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_05_e=6_3000.png}&\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_05_e=7_3000.png}&\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_05_e=8_3000.png} \\
\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_05_e=9_3000.png}&\includegraphics[scale=0.27]{mu2less_PS_mu1_mu2_05_e=10_3000.png}&\\
\end{tabular}
\caption{The Poincar\'e map for orbits with $x_0=[3.0,3.5,3.75,4.0,4.25,4.5,4.75,5.0,5.25,5.5,6.0]$ and $y_0=[0.0]$ as $\epsilon$ increases from 0 to 1 in steps of 0.1. The system is evolved for 3000 iterations in time-steps of $\tau=10$ and the Poincar\'e map of section for $y=0, \dot{y}<0$ is plotted for all 11 initial conditions for the Copenhagen system ($\mu_{1}$=$\mu_{2}=0.5$) }
\label{fig:PS_0.5}
\end{figure*}
\section{Lyapunov Characteristic Exponents}
A very popular indicator of chaos in dynamical systems is the calculation of the Lyapunov Characteristic Exponents (LCE), which has been extensively applied to the study of chaos in celestial dynamics especially in the context of the three-body problem (\citet{Gueron01, Dubeibe16, Dubeibe17}; \citet*{Wu03, Wu06b}). It is a measure of the exponential divergence of two neighbouring trajectories in phase space. The rate of separation of the two trajectories is dependent on the initial separation vector. For a pair of trajectories, the number of exponents for the system is equal to dimension of its phase space. However, the largest exponent dominates in the limit $t\to\infty$. The largest Lyapunov exponent, called the Maximal Lyapunov Exponent (MLE), is defined by,
\begin{equation}
\Lambda_{max}=\lim_{t\to\infty}\frac{1}{t}\log\frac{||\Upsilon(t)||}{||\Upsilon(0)||}
\label{eqn:LyapunovVar}
\end{equation}
where $\Upsilon(t)$ is the solution to the variational equations for the potential under consideration \citep*{Tancredi01}. Such a computation mechanism for the MLE is called the variational method and is the most accurate. However, for systems such as the one under consideration where computation of the variational equations are cumbersome, an alternative was introduced in \citet*{Benettin76}. The equation (\ref{eqn:LyapunovVar}) is thus replaced by the following:
\begin{equation}
\Lambda_{max}=\lim_{t\to\infty}\frac{1}{t}\log\frac{||\mathbf{\delta x}(t)||}{||\mathbf{\delta x}(0)||}
\label{eqn:Lyapunov}
\end{equation}
where, the deviation vector between the two trajectories is $\mathbf{\delta x(t)}$, with $\mathbf{\delta x(0)}\rightarrow0$. The mean rate of deviation of the two trajectories is given by:
\begin{align}
\Lambda_{max}=&\frac{1}{n\tau}\sum_{k=1}^n\log\frac{||\mathbf{\delta x}(k\tau)||}{||\mathbf{\delta x}(0)||}
\end{align}
\begin{figure}
\includegraphics[scale=0.2]{final_lya.png}
\caption{Plot of the Lyapunov Characteristic Exponent (LCE) vs $\epsilon$.}
\label{fig:LyapunovPlot}
\end{figure}
This method is called the two-particle method and is the one we utilize to calculate the MLE for each system. The result is accurate as long as the two trajectories are in the immediate neighbourhood of each other in phase space and the machine used for computation has enough precision. As concluded in the work by \citet{Tancredi01}, we have also taken the initial separation between the two trajectories to be $\mathbf{\delta x(0)}=10^{-8}$ and have integrated the system in double precision for $n=10^{5}$ iterations, each of time step $\tau=0.1$. Numerical integration diverges rapidly unless the deviation vectors are re-normalized periodically. The two trajectories are evolved separately and the deviation vector is re-normalized using the Gram-Schmidt re-normalization after each time step. To get a quantitative representation of the chaos in the system, the MLE is averaged over the entire phase space. But, as a preliminary investigation of system, we use the initial conditions: $x_{0}=[3.0,3.5,3.75,4.0,4.25,4.5,4.75,5.0,5.25,5.5,6.0]$, $y_{0}=0.0$ and $\dot{x}_{0}=0.0$, with $C_{j}=4.0$ for the biased-mass and Copenhagen systems. If trajectories are stable, the value of the MLE remains very small, usually less than $5\times10^{-4}$ (low value of MLE). But for chaotic trajectories, the deviations are exponential and the value of the MLE increases rapidly with time. After $10^5$ iterations, its value is usually greater than $5\times10^{-4}$ (high value of MLE). The MLE for initial conditions $x_0=[3.0,3.5,3.75,4.0,4.25,4.5,4.75,5.0,5.25,5.5,6.0]$ are calculated and averaged for each value of $\epsilon$ and is called the Lyapunov Characteristic Exponent (LCE) for the particular value of $\epsilon$ \citep*{Nag17}. The LCE provides a qualitative measure of the amount of chaos in the system, even for the few initial conditions chosen for the study (see \citet{Nag17} as an example). Figure (\ref{fig:LyapunovPlot}) is the plot of the LCE against $\epsilon$ for both the biased-mass and Copenhagen systems.
Both for the biased-mass system and the Copenhagen system, the total chaos in the system for small $\epsilon$ is low. For the biased-mass system, the LCE for all the initial conditions are $<5\times10^{-4}$ for values of $\epsilon=0.0$ indicating stable orbits. The same is true for the Copenhagen system, except for $x_0=6.0$ which gives an LCE of $4.22\times10^{-03}$. For the biased-mass system, the LCE for all initial conditions are $<5\times10^{-4}$ for $x_0=[3.5,3.75,4.0,4.5]$, implying stable orbits. Some initial conditions, like $x_0=[4.25,4.75,5.0,5.25,6.0]$ for the biased mass system and $x_0=5.5$ for the Copenhagen system, the system shows high values of LCE for intermediate values of $\epsilon$, but low values of LCE for higher values of $\epsilon$. The most interesting among these is the initial condition $x_0=6.0$, which shows low values of LCE only for $\epsilon=0.0$ and $1.0$. This reaffirms the conclusion drawn from the Poincar\'e maps that the chaos in the system is maximum for intermediate values of $\epsilon$. For the Copenhagen system, values of LCE for $x_0=[3.5,3.75,4.0,4.5,4.75]$ are low for all values of $\epsilon$. The initial condition $x_0=6.0$ show high values of LCE for values of $\epsilon$. The initial condition $x_0=5.5$ shows high values of LCE for all values of $\epsilon$ except for $\epsilon=0.0$.
Figure (\ref{fig:LyapunovPlot}) shows that the chaos in the system is low for both the biased-mass system and the Copenhagen system, as indicated by low values of the averaged LCE. Its value rises rapidly for the biased mass system but much slower for the Copenhagen system. Both the systems show maximum values of the averaged LCE for intermediate values of epsilon, which for the biased mass system is at $\epsilon=0.825$ and $\epsilon=0.525$ for the Copenhagen. This re-iterates the observations made from the orbital evolution and the Poincar\'e maps of the systems.
\begin{figure*}
\begin{tabular}{cc}
\includegraphics[scale=0.4]{0001kerr1.png}&\includegraphics[scale=0.4]{0001pseudokerr1.png} \\
\includegraphics[scale=0.4]{0001kerr5.png}&\includegraphics[scale=0.4]{0001pseudokerr5.png} \\
\includegraphics[scale=0.4]{0001kerr10.png}&\includegraphics[scale=0.4]{0001pseudokerr10.png} \\
\end{tabular}
\caption{Poincar\'e map of section for $y=0$ and $\dot{y}<0.0$ for the biased-mass systems (mass ratio of the primaries equals to 0.001) for different values of $\epsilon$. The figures on the left are maps for the system with Schwarzschild-like primaries while those on the right are for the system with Kerr-like primaries.}
\label{fig:kerr_pesudokerr0001}
\end{figure*}
\begin{figure*}
\begin{tabular}{cc}
\includegraphics[scale=0.4]{05kerr1.png}&\includegraphics[scale=0.4]{05pseudokerr1.png} \\
\includegraphics[scale=0.4]{05kerr5.png}&\includegraphics[scale=0.4]{05pseudokerr5.png} \\
\includegraphics[scale=0.4]{05kerr10.png}&\includegraphics[scale=0.4]{05pseudokerr10.png} \\
\end{tabular}
\caption{Poincar\'e map of section for $y=0.0$ and $\dot{y}<0.0$ for the Copenhagen systems (mass ratio of the primaries equals to 0.5) for different values of $\epsilon$. The figures on the left are maps for the system with Schwarzschild-like primaries while those on the right are for the system with Kerr-like primaries.}
\label{fig:kerr_pesudokerr05}
\end{figure*}
\section{Schwarzschild and Kerr primaries: a comparison}
In order to examine the effect of the spin of the primaries on the system, we present a comparison to a system with two Schwarzschild like primaries. Using the potential described in \citet{Dubeibe16} and \citet{Zotos17b}, we construct a set of Poincar\'e maps of section for $\epsilon=[0.1,0.5,1.0]$. We evolve each orbit for 3000 iterations in time-steps of $\tau=10$ and plot the section of the orbit for $y=0.0$ and $\dot{y}<0.0$. Figures (\ref{fig:kerr_pesudokerr0001}) and (\ref{fig:kerr_pesudokerr05}) show Poincar\'e maps of the biased-mass and Copenhagen systems respectively. For both the systems, the plots on the left are for the system with Schwarzschild like primaries, while those on the right are for the system with Kerr like primaries.
While for $\epsilon=0.0$ both systems reduce to the Newtonian CRTBP, it is apparent that even for small perturbations to the Newtonian system, as represented by $\epsilon=0.1$, the introduction of the spin destabilizes a number of initial conditions. For the biased-mass case, the Schwarzschild system shows all chosen initial conditions to be stable and quasi-periodic, with the Poincar\'e map showing KAM tori for all values of $\epsilon$. The Poincar\'e map for the Kerr system differs radically from its Schwarzschild counterpart even for $\epsilon=0.1$, showing a large sea of chaotic points surrounding an island of stable initial conditions. The island of stability grows smaller as $\epsilon$ is increased, as has already been discussed in section (\ref{sec:PS}). For the Copenhagen case, the Poincar\'e maps for both the Schwarzschild and Kerr systems feature an island of stability surrounded by a sea of chaos for all three values of $\epsilon$. The Poincar\'e maps for both the systems look alike, implying a similar number of stable initial conditions. As $\epsilon$ is increased, the number of stable initial conditions for the Kerr system decreases rapidly, as evident from the smaller islands of stability on the Poincar\'e maps of the system for $\epsilon=0.5$ and $\epsilon=1.0$. However, for the same values of $\epsilon$, the number of stable initial conditions for the Schwarzschild system remains approximately the same.
Thus, for both mass ratios, we observe that the introduction of spin in the CRTBP with Schwarzschild-like primaries destabilizes a number of initial conditions, with the amount of chaos in the system growing with increase in $\epsilon$.
\section{Conclusions}
In the present paper, we present a beyond-Newtonian potential for the planar circular restricted three-body problem with Kerr like primaries. This is achieved by using the Fodor-Hoenselaers-Perj\'es procedure to expand the Kerr metric and by retaining corrections up to the first non-Newtonian term. The system is conservative, with the Hamiltonian being time independent. The parameter $\epsilon\in[0.0,1.0]$ is introduced in order to facilitate the observation of the system as it transitions from the Newtonian to the beyond-Newtonian regime. The dynamics of a test particle in this potential for $\mu_{1}=\mu_{2}=0.001$ (or the biased-mass system) and for $\mu_{1}=\mu_{2}=0.5$ (or the Copenhagen system), are inspected for a Jacobi constant $C_{j}=4.0$. For an initial investigation of the system, orbits for a few selected initial conditions are plotted. A short analysis of the fixed points of the systems and their stability is undertaken. A purely Newtonian CRTBP system is known to have five Lagrange points, as seen for $\epsilon\,=\,0$ in our case. However, number of Lagrange points is not constant as the system transitions from the Newtonian to the beyond-Newtonian regime. It is observed that the number of fixed points strongly depends on the parameter $\epsilon$ as does their stability. Next, the stability of the orbits is also examined through the use of the Poincar\'e map of section for different values of $\epsilon$. The Poincar\'e maps for all non-zero values of $\epsilon$ show islands of stability constructed of concentric Kolmogorov-Arnold-Moser (KAM) tori, embedded in a sea of chaos.
Thus we note that the introduction of the parameter $\epsilon$ helps us to conclude that even small perturbations to the Newtonian CRTBP destabilizes the system for both the cases. If we track the evolution of the system keeping the Jacobian constant fixed, a stable orbit in the Newtonian system is observed to become either chaotic or sometimes even remain regular in the beyond-Newtonian limit. In the limits $\epsilon\,=\,0$ and $\epsilon\,=\,1$, the phase space is seen to be filled mostly with periodic orbits, rarely interspersed with chaotic ones. However, as $\epsilon$ departs even slightly from zero, trajectories that were stable in the Newtonian system become unstable. It is seen that in most of the cases (for a given set of initial conditions) whose phase space is bounded in the classical regime, correspond to unbounded trajectories in the non-Newtonian regime. This implies that both systems become largely unstable for intermediate values of $\epsilon$. The instability of the orbits can possibly be linked to the observed lack of stable fixed points in both the systems. This is also confirmed by the Lyapunov Characteristic Exponent, calculated for each value of $\epsilon$, which is in accordance to the conclusions made by several authors earlier for different systems \citep{Huang14a, Dubeibe16, Nag17}. In conclusion, we may say that even the smallest corrections to the Newtonian circular restricted three-body problem could drastically change the stability and the dynamics of the system.
In addition, we would like to note that an in-depth study of the phase space using more rapid indicators of chaos, like Fast Lyapunov Indicators (FLI) (\citet*{Froeschle97, Froeschle00}; \citet{Wu06b}), Small Alignment Index (SALI) \citep*{Skokos01} and General Alignment Index (GALI) \citep*{Skokos07} will facilitate a much more detailed analysis of the evolution of the Lagrange points of the proposed potential. Coupled with this, a detailed linear stability analysis of the Lagrange points as a function of the parameter $\epsilon$ and an analysis of the basins of convergence is expected to reveal more information about the attractors of the system. Further, we would also like to investigate the degree of equivalence of the potential constructed in our paper with the pseudo-Newtonian potential formulation of a binary with spinning primaries, for example that of a system modelled by the superposition of two Artemova potentials \citep{Artemova96}. This would allow us to reproduce features like the Innermost Stable Circular Orbit (ISCO), maximally stable orbits, and the horizon radius, in our chosen scalings and units. This would in turn facilitate the calculation of physically relevant distances, for example, the coordinates of fixed points for different values of $\epsilon$ and primary masses in real physical units, thereby allowing us to predict real astrophysical scenarios using our present model (for a recent example refer to \citet*{Yi20}). Thus, we would like to explore these issues in greater depth as part of our future work.
We also note that the current formalism is strictly valid for particles whose motion is restricted to the plane containing the primaries. However, a more general model for accreting particles should also include a study of the dynamics of such off-axis motion. Thus, we would like to direct our future studies to incorporate such effects for off-axis halo particles in a generalized beyond-Newtonian framework.
\section*{Acknowledgements}
The authors would like to thank Dr. Sankhasubhra Nag for his helpful suggestions and discussions. The authors would also like to thank Ms. Pratyusha Banerjee for her help with computational work. In addition, they would also like to acknowledge Dr. Tanaya Bhattacharyya for taking time out to go through the manuscript meticulously. Last but not the least, the authors would like to acknowledge the anonymous referee for his/her valuable comments and suggestions. We would like to dedicate this paper to all the warriors fighting the COVID-19 pandemic across the globe.
\section*{Data Availability}
No new data were generated or analysed in support of this research.
|
\section{Introduction}
\label{sec:introduction}
Histopathological image understanding has been revolutionized by recent machine learning advancements, especially deep learning ($\mathrm{DL}$) \cite{Bera2019,Serag2019}. $\mathrm{DL}$ has catered to increasing diagnostic throughput as well as a need for high predictive performance, reproducibility and objectivity.
However, such advantages come at the cost of a reduced transparency in decision-making processes~\cite{Holzinger2017,Tizhoosh2018,Hagele2020}. Considering the need for reasoning any clinical decision, it is imperative to enable the explainability of $\mathrm{DL}$ decisions to pathologists.
\begin{figure}[!t]
\centering
\centerline{\includegraphics[width=\linewidth]{figures/fig0.png}}
\caption{Sample explanations produced by pixel- and entity-based explainability techniques for a ductal carcinoma \emph{in situ} $\mathrm{RoI}$.}
\vspace{-1em}
\label{fig:teaser}
\end{figure}
Inspired by the explainability techniques (explainers) for $\mathrm{DL}$ model decisions on natural images \cite{Simonyan2013,Zeiler2014,Yosinski2015,Bach2015,Montavon2015,Selvaraju2017,Kindermans2017,Zintgraf2017, Chattopadhay2018,Kim2018}, several explainers have been implemented in digital pathology, such as feature attribution \cite{Korbar2017, Binder2018, Hagele2020}, concept attribution \cite{Graziani2020}, and attention-based learning \cite{Lu2020b}.
However, pixel-level explanations, exemplified in Figure~\ref{fig:teaser}, pose several notable issues, including:
(1)~a pixel-wise analysis disregards the notion of biological tissue entities, their topological distribution, and inter-entity interactions;
(2)~a typical patch-based $\mathrm{DL}$ processing and explainer fail to accommodate complete tumor macro-environment information; and
(3)~pixel-wise visual explanations
tend to be blurry.
Explainability in entity space is thus a natural choice to address the above issues. To that end, an entity graph representation is built for a histology image, where nodes and edges denote biological entities and inter-entity interactions followed by a Graph Neural Network ($\mathrm{GNN}$) \cite{Kipf2017,Xu2018b}. The choice of entities, such as cells \cite{Gunduz2004,Zhou2019,Pati2020}, tissues \cite{Pati2020} or others, can be task-dependent. Subsequently, explainers for graph-structured data \cite{Baldassarre2019,Pope2019,Ying2019} applied to the entity graphs highlight responsible entities for the concluded diagnosis, thereby generating intuitive explanations for pathologists.
In the presence of various graph explainers producing distinct explanations for an input, it is crucial to discern the explainer that best fits the explainability definition \cite{Arrieta2020}. In the context of computational pathology, explainability is defined as making the $\mathrm{DL}$ decisions understandable to pathologists \cite{Holzinger2017}. To this end, the qualitative evaluation of explainers' explanations by pathologists is the candid measure. However, it requires evaluations by task-specific expert pathologists, which is subjective, time-consuming, cumbersome, and expensive.
Additionally, though the explanations are intuitive, they do not relate to pathologist-understandable terminologies, \eg ``How big are the important nuclei?", ``How irregular are their shape?" etc., which toughens the comprehensive analysis.
These bottlenecks undermine not only any qualitative assessment but also quantitative metrics requiring user interactions~\cite{Mohseni2018}.
Furthermore, expressing the quantitative metrics in user-understandable terminologies \cite{Arrieta2020} is fundamental to achieve interpretability \cite{Doshivelez2017, Nguyen2020}.
To this end, the most popular quantitative metric, explainer \emph{fidelity}~\cite{Ribeiro2016,Dhurandhar2017,Samek2017,Hoffman2018,Mohseni2018,Pope2019}, is not satisfactory. Moreover, explainers intrinsically maintain high-\emph{fidelity}, \eg $\textsc{GnnExplainer}$~\cite{Ying2019} produces an explanation to match the $\mathrm{GNN}$'s prediction on the original graph.
Thus, we propose a set of novel user-independent quantitative metrics expressing pathologically-understandable \emph{concepts}. The proposed metrics are based on class separability statistics using such \emph{concepts}, and they are applicable in other domains by incorporating domain-specific \emph{concepts}.
We use the proposed metrics to evaluate three types of graph-explainers, (1) graph pruning: $\textsc{GnnExplainer}$ \cite{Ying2019, Jaume2020}, (2) gradient-based saliency: $\textsc{GraphGrad-CAM}$ \cite{Selvaraju2017,Pope2019}, $\textsc{GraphGrad-CAM++}$ \cite{Chattopadhay2018}, (3) layer-wise relevance propagation: $\textsc{GraphLRP}$~\cite{Bach2015,Montavon2015,Schwarzenberg2019}, for explaining Cell-Graphs \cite{Gunduz2004} in Breast Cancer Subtyping as shown in Figure \ref{fig:teaser}.
Our specific contributions in this work are:
\begin{itemize}
\item A set of novel quantitative metrics based on the statistics of class separability using domain-specific \emph{concepts} to characterize graph explainability techniques. To the best of our knowledge, our metrics are the first of their kind to quantify explainability based on domain-understandable terminologies;
\item Explainability in computational pathology using pathologically intuitive entity graphs;
\item Extensive qualitative and quantitative assessment of various graph explainability techniques in computational pathology, with a validation of the findings by expert pathologists.
\end{itemize}
\begin{figure*}[!t]
\centering
\centerline{\includegraphics[width=\textwidth]{figures/fig1.png}}
\caption{Overview of the proposed framework.
(a) presents pathologist, and entity-based (cell-graph + $\mathrm{GNN}$) diagnosis of a histology image.
(b) presents nuclei-level pathologically relevant \emph{concept} measure $D$, a post-hoc graph explainability technique to derive nuclei-level importance $\mathcal{I}$ for \emph{concepts} $\mathcal{C}$, measurable \emph{attributes} $\mathcal{A}_c$, and classes $\mathcal{T}$. $D$, $\mathcal{I}$ and prior pathological knowledge defining \emph{concepts'} relevance are utilized to propose a novel set of quantitative metrics to evaluate the explainer quality in pathologist-understandable terms.}
\label{fig:approach}
\end{figure*}
\section{Related work}
\textbf{Graphs in Digital Pathology:}
Graph-based tissue image analysis effectively describes a tissue environment by incorporating morphology, topology, and tissue components interactions.
To this end, cell-graph ($\mathrm{CG}$) is the most popular graph representation, where nodes and edges depict cells and cellular interactions \cite{Gunduz2004}. Cell morphology is embedded in the nodes via hand-crafted features~\cite{Gunduz2004,Zhou2019,Pati2020} or $\mathrm{DL}$ features \cite{Chen2019,Pati2021}. The graph topology is heuristically defined using k-Nearest Neighbors, probabilistic modeling, Waxman model etc. \cite{Sharma2015}
Subsequently, the $\mathrm{CG}$s are processed by classical machine learning \cite{Sharma2015,Sharma2016,Sharma2017} or $\mathrm{GNN}$ \cite{Zhou2019,Chen2019,Gadiya2020,Pati2020} to map the tissue structure to function relationship.
Recently, improved graph-representations using patches \cite{Aygunes2020}, tissue components \cite{Pati2020}, and hierarchical cell-to-tissue relations \cite{Pati2020} are proposed to enhance the structure-function mapping.
Other graph-based applications in computational pathology include cellular community detection \cite{Javed2020}, whole-slide image classification \cite{Zhao2020, Adnan2020} etc. Intuitively, a graph representation utilizes pathologically relevant entities to represent a tissue specimen, which allows pathologists to readily relate with the input, also enabling them to include any task-specific prior knowledge.
\textbf{Explainability in Digital Pathology:}
Explainability is an integral part of pathological diagnosis. Though $\mathrm{DL}$ solutions have achieved remarkable diagnostic performance, their lack of explainability is unacceptable in the medical community~\cite{Tizhoosh2018}. Recent studies have proposed visual explanations~\cite{Hagele2020} and salient regions~\cite{Korbar2017,Hagele2020} using feature-attribution techniques \cite{Selvaraju2017,Chattopadhay2018}.
Differently, concept-attribution technique~\cite{Graziani2020} evaluates the sensitivity of network output w.r.t.\ quantifiable image-level pathological \emph{concepts} in patches. Although such explanations are pathologist-friendly, image-level \emph{concepts} are neither fit nor meaningful for real-world large histology images that contain many localized concepts.
Furthermore, attention-based learning~\cite{Lu2020b}, and multimodal mapping between image and diagnostic report~\cite{Zhang2019} are devised to localize network attention.
However, the pixel-wise and patch-based processing in all the aforementioned techniques ignore biological entities' notion; thus, they are not easily understood by pathologists.
Separately, the earlier stated entity graph-based processing provides an intuitive platform for pathologists. However, research on explainability and visualization using entity graphs has been scarce: CGC-Net~\cite{Zhou2019} analyzes cluster assignment of nodes in $\mathrm{CG}$ to group them according to their appearance and tissue types. CGExplainer~\cite{Jaume2020} introduces a post-hoc graph-pruning explainer to identify decisive cells and interactions. Robust spatial filtering~\cite{Sureka2020} utilizes an attention-based $\mathrm{GNN}$ and node occlusion to highlight cell contributions.
No previous work has comprehensively analyzed and quantified graph explainers in computational pathology while expressing explanations in a pathologist-understandable form to the best of our knowledge.
This gap between the existing and desired explainability of $\mathrm{DL}$ outputs in digital pathology motivates our work herein.
\section{Method}
In this section, we present entity graph processing, explainability methods, and our proposed evaluation metrics. First, we transform a histology region-of-interest ($\mathrm{RoI}$) into a \emph{biological entity graph}. Second, we introduce a ``black-box" $\mathrm{GNN}$ that maps the \emph{entity graph} to a corresponding class label. Third, we employ a post-hoc graph explainer to generate explanations. Finally, we perform a qualitative and quantitative assessments of the generated explanations. An overview of the methodology is shown in Figure~\ref{fig:approach}.
\subsection{Entity graph notations}
We define an attributed undirected entity graph $G := (V, E, H)$ as a set of nodes $V$, edges $E$, and node attributes $H \in \mathbb{R}^{|V| \times d}$. $d$ denotes the number of attributes per node, and $|.|$ denotes set cardinality.
The graph topology is defined by a symmetric graph adjacency, $A \in \mathbb{R}^{|V| \times |V|}$, where $A_{u,v} = 1$ if $e_{uv} \in E$.
We denote the neighborhood of a node $v \in V$ as $\mathcal{N}(v) := \{u \in V \; | \; v \in V, \; e_{uv} \in E \;\}$.
We denote a set of graphs as $\mathcal{G}$.
\subsection{Entity graph construction} \label{sec:graph_construction}
Our methodology begins with transforming $\mathrm{RoI}$s into entity graphs. It ensures the method's inputs are pathologically interpretable, as the inputs consist of biologically-defined objects that pathologists can directly \emph{relate-to} and \emph{reason-with}. Thus, image-to-graph conversion moves from $\textit{uninterpretable}$ to $\textit{interpretable}$ input space.
In this work, we consider cells as entities, thereby transforming $\mathrm{RoI}$s into cell-graphs ($\mathrm{CG}$s). A $\mathrm{CG}$ nodes and edges capture the morphology of cells and cellular interactions. A $\mathrm{CG}$ topology acquires both tissue micro and macro-environment, which is crucial for characterizing cancer subtypes.
First, we detect nuclei in a $\mathrm{RoI}$ at 40$\times$ magnification using Hover-Net~\cite{Graham2019}, a nuclei segmentation algorithm pre-trained on MoNuSeg~\cite{Kumar2017}. We process patches of size 72$\times$72 around the nuclei by ResNet34 \cite{He2016} pre-trained on ImageNet~\cite{JiaDeng2009} to produce nuclei visual attributes. We further concatenate nuclei spatial attributes, $\ie$ nuclei centroids min-max normalized by $\mathrm{RoI}$ dimension.
The nuclei and their attributes (visual and spatial) define the nodes and node attributes of the $\mathrm{CG}$, respectively.
Following prior work~\cite{Pati2020}, we construct the $\mathrm{CG}$~topology by employing thresholded $\textit{k}$-Nearest Neighbors algorithm. We set $\textit{k}=5$, and prune the edges longer than $50$ pixels (12.5 $\mu$m).
The $\mathrm{CG}$-topology encodes how likely two nearby nuclei will interact~\cite{Francis1997}.
A $\mathrm{CG}$ example is presented in Figure~\ref{fig:teaser}.
\subsection{Entity graph learning} \label{sec:graph_learning}
Given $\mathcal{G}$, the set of $\mathrm{CG}$s, the aim is to infer the corresponding cancer subtypes. We use $\mathrm{GNN}$s~\cite{Scarselli2009, Defferrard2016, Kipf2017, Hamilton2017b, Velickovic2017, Ying2018, Gilmer2017}, the conceptual analogous of 2D convolution for graph-structured data, to classify the $\mathrm{CG}$s.
A $\mathrm{GNN}$ layer follows two steps: for each node $v \in V$, \begin{inparaenum}[(i)]
\item \emph{aggregation step}: the states of neighboring nodes, $\mathcal{N}(v)$, are aggregated via a differentiable and permutation-invariant operator to produce $a(v) \in \mathbb{R}^d$, then,
\item \emph{update step}: the state of $v$ is updated by combining the current node state $h(v) \in \mathbb{R}^d$ and the aggregated message $a(v)$ via another differentiable operator. \end{inparaenum}
After $L$ iterations, \ie the number of $\mathrm{GNN}$ layers, a \emph{readout step} is employed to merge all the node states via a differentiable and permutation-invariant function to result in a fixed-size graph embedding.
Finally, the graph embeddings are processed by a classifier to predict the class label.
In this work, we use a flavor of Graph Isomorphism Network (GIN)~\cite{Xu2018b},
that uses \emph{mean} and a \emph{multi-layer perceptron} (MLP) in the \emph{aggregation} and \emph{update} step respectively. Formally, we define a layer as,
\begin{align}
h(v)^{(l+1)} = \mbox{MLP}^{(l)} \Big(h(v)^{(l)} + \frac{1}{|\mathcal{N}(v)|} \sum_{u \in \mathcal{N}(v)} h(u)^{(l)} \Big)
\end{align}
where $h(v)$ denotes features of node $v$, and $l \in \{1, ..., L\}$. Our $\mathrm{GNN}$ consists of 3-GIN layers, with each layer including a 2-layer MLP. The dimension of latent node embeddings is fixed to 64 for all layers. We use \emph{mean} operation in \emph{readout step}, and feed the graph embedding to a 2-layer MLP classifier. The $\mathrm{GNN}$ is trained end-to-end by minimizing cross-entropy loss between predicted logits and target cancer subtypes.
We emphasize that the entity-based processing follows a pathologist's diagnostic procedure that identifies diagnostically relevant nuclei and analyzes cellular morphology and topology in a $\mathrm{RoI}$, as shown in Figure~\ref{fig:approach}.
\begin{figure*}[!t]
\centering
\centerline{\includegraphics[width=\linewidth]{figures/fig2.png}}
\caption{Overview of proposed quantitative assessment. (a) presents input dataset $\mathcal{D}$, and parameters \emph{concepts} $\mathcal{C}$, measurable \emph{attributes} $\mathcal{A}_c$, classes $\mathcal{T}$, and importance thresholds $\mathcal{K}$. For simplicity $|\mathcal{A}_c|=1, \forall c \in \mathcal{C}$ in this figure. (b) shows histogram probability densities for $\forall a \in \mathcal{A}_c, \forall k \in \mathcal{K}, \forall t \in \mathcal{T}$. (c) displays the algorithm for computing class separability score $S$. (d) presents the algorithm for computing the proposed class separability-based risk-weighted quantitative metrics.}
\label{fig:metric}
\end{figure*}
\subsection{Post-hoc graph explainer} \label{sec:explainer}
We generate an explanation per entity graph by employing post-hoc graph explainers. The explanations allow to evaluate the pathological relevance of black-box neural network reasoning. Specifically, we aim to evaluate the agreement between the pathologically relevant set of nuclei in a $\mathrm{RoI}$, and the explainer identified set of important nuclei, \ie nuclei driving the prediction, in corresponding $\mathrm{CG}$. In this work, we consider three types of graph explainers for explaining $\mathrm{CG}$s, which follow similar operational setting, \ie \begin{inparaenum}[(i)]
\item input data are attributed graphs,
\item a $\mathrm{GNN}$ is trained \emph{a priori} to classify the input data, and
\item each data point can be inferred independently to produce an explanation.
\end{inparaenum}
We present the graph explainers in the following sections and their detailed mathematical formulations in the Appendix.
\textbf{$\textsc{GraphLRP}$:} Layerwise relevance propagation (LRP)~\cite{Bach2015} propagates the output logits backward in the network using a set of propagation rules to quantify the positive contribution of input pixels for a certain prediction. Specifically, LRP assigns an importance score to each neuron such that the output logit relevance is preserved across layers.
While initially developed for explaining fully-connected layers, LRP can be extended to $\mathrm{GNN}$ by treating the $\mathrm{GNN}$ \emph{aggregation step} as a fully connected layer that projects the graph adjacency matrix on the node attributes as in \cite{Schwarzenberg2019}. LRP outputs per-node importance.
\textbf{$\textsc{GraphGrad-CAM}$:} $\textsc{Grad-CAM}$ \cite{Selvaraju2017} is a feature attribution approach designed for explaining $\mathrm{CNN}$s operating on images. It produces class activation explanation following two steps. First, it assigns weights to each channel of a convolutional layer $l$ by computing the gradient of the targeted output logit w.r.to each channel in layer $l$. Second, importance of the input elements are computed by the weighted combination of the forward activations at each channel in layer $l$.
The extension to $\mathrm{GNN}$ is straightforward~\cite{Pope2019}, and only requires to compute the gradient of the predicted logits w.r.to a $\mathrm{GNN}$ layer. Following prior work~\cite{Pope2019}, we take the average of node-level importance-maps obtained from all the $\mathrm{GNN}$ layers $l \in \{1, ..., L\}$ to produce smooth per-node importance.
\textbf{$\textsc{GraphGrad-CAM++}$:} $\textsc{Grad-CAM++}$ \cite{Chattopadhay2018} is an increment on $\textsc{Grad-CAM}$ by including spatial contributions into the channel-wise weight computation of a convolutional layer. The extension allows weighting the contribution by each spatial location at a layer for improved spatial localization. The spatial locations in a convolutional layer are analogous to the size of the graph in a $\mathrm{GNN}$ layer. With this additional consideration, we propose an extension of $\textsc{Grad-CAM++}$ to graph-structured data.
\textbf{$\textsc{GnnExplainer}$:} $\textsc{GnnExplainer}$~\cite{Ying2019, Jaume2020} is a graph pruning approach that aims to find a compact sub-graph $G_s \subset G$ such that mutual information between $G_s$ and $\mathrm{GNN}$ prediction of $G$ is maximized.
Sub-graph $G_s$ is regarded as the explanation for the input graph $G$. $\textsc{GnnExplainer}$ can be seen as a feature attribution technique with binarized node importances. To address the combinatorial nature of finding $G_s$, $\textsc{GnnExplainer}$ formulates it as an optimization problem that learns a mask to activate or deactivate parts of the graph. \cite{Jaume2020} reformulates the initial approach in \cite{Ying2019} to learn a mask over the nodes instead of edges. The approach in~\cite{Jaume2020} is better suited for pathology as the nodes, \ie biological entities, are more intuitive and substantial for disease diagnosis than heuristically-defined edges.
The optimization for an entity graph results in per-node importance.
\subsection{Quantitative metrics for graph explainability} \label{sec:metrics}
In the presence of several graph explainers producing distinct explanations for an input, it is imperative to discern the explainer that produces the most pathologically-aligned explanation. Considering the limitations of existing qualitative and quantitative measures presented in Section \ref{sec:introduction}, we propose a novel set of quantitative metrics based on class separability statistics using pathologically relevant \emph{concepts}. Intuitively, a good explainer should emphasize the relevant \emph{concepts} that maximize the class separation. Details of the metric evaluations are presented as follows.
\textbf{Input:} A graph explainer outputs an explanation, \ie node-level \emph{importance} $\mathcal{I}$, for an input $\mathrm{CG}$. To quantify a \emph{concept} $c \in \mathcal{C}$, $\mathcal{C}$ denoting the set of \emph{concepts}, we measure nuclear \emph{attributes} $a \in \mathcal{A}_{c}$ for each nucleus in $\mathrm{CG}$, \eg, for $c\!=\!$ \emph{nuclear shape}, we measure $\mathcal{A}_{c}\!=\!$ \{\emph{perimeter, roughness, eccentricity, circularity}\}.
We create a dataset $\mathcal{D} = \bigcup_{t \in \mathcal{T}} \mathcal{D}_t$, $\mathcal{T}$ denoting the set of cancer subtypes. We define $\mathcal{D}_t := \{(D^t_i, \mathcal{I}^t_i) | i = 1, \dots, N_t\} \forall t \in \mathcal{T}$, where $N_t$ is the number of CGs for tumor type $t$. $\mathcal{I}^t_i$ and $D^t_i$ are, respectively, the sorted importance matrix for a CG indexed by $i$ and corresponding node-level attribute matrix.
To perform inter-concept comparisons, we conduct \emph{attribute}-wise normalization across all $D^t_i \; \forall t,i$. In order to compare different explainers, we conduct $\mathrm{CG}$-wise normalization of $\mathcal{I}$. The structure of input dataset $\mathcal{D}$ is presented in Figure \ref{fig:approach}\textcolor{red}{(a)}.
Note that the notion of important nuclei vary
\begin{inparaenum}[(1)]
\item per-$\mathrm{CG}$ since the number of nodes vary across $\mathrm{CG}$s, and
\item per-explainer.
\end{inparaenum}
Hence, selecting a \emph{fixed} number of important nuclei per-$\mathrm{CG}$ and per-explainer is not meaningful. To overcome this issue, we assess different number of important nuclei $k \in \mathcal{K}$, selected based on node importances, per-$\mathrm{CG}$ and per-explainer.
In the following sections we will show how to aggregate the results for a given explainer.
\textbf{Histogram construction:}
Given the input dataset $\mathcal{D}$, and parameters $\mathcal{K}, \mathcal{C}, \mathcal{A}_c, \mathcal{T}$, we apply threshold $k \in \mathcal{K}$ on $\mathcal{I}_i^t, \forall t \in \mathcal{T}, \forall i \in N_t$ to select $\mathrm{CG}$-wise most important nuclei. The cancer subtype-wise selected set of nuclei data from $\mathcal{D}$ are used to construct histograms $H_t^{(k)}(a), \forall a \in \mathcal{A}_c$, $\forall c \in \mathcal{C}$ and $\forall t \in \mathcal{T}$. For histogram $H_t^{(k)}(a)$, bin-edges are decided by quantizing the complete range of \emph{attribute} $a$, \ie $\mathcal{D}(a)$, by a fixed step size. We convert each $H_t^{(k)}(a)$ into a probability density function.
Similarly, sets of histograms are constructed by applying different thresholds $k \in \mathcal{K}$. Sample histograms are shown in Figure \ref{fig:approach}\textcolor{red}{(b)}.
\textbf{Separability Score (\emph{S}):}
Given two classes $t_x, t_y \in \mathcal{T}$ and corresponding probability density functions $H_{t_x}^{(k)}(a)$ and $H_{t_y}^{(k)}(a)$, we compute \emph{class separability} $s_a^{(k)}(t_x, t_y)$ based on optimal transport as the Wasserstein distance between the two density functions.
We average $s_a^{(k)}(t_x, t_y)$ over all $a \in \mathcal{A}_c$ to obtain a score $s_c^{(k)}(t_x, t_y)$ for \emph{concept} $c$ and threshold $k$.
Finally, we compute the area-under-the-curve (AUC) over the threshold range $\mathcal{K}$ to get the aggregated class separability $S_{(t_x, t_y), c}$ for a \emph{concept} $c$.
The class separability score indicates the significance of \emph{concept} $c$ for the purpose of separating $t_x$ and $t_y$.
Thus, separability scores can be used to compare different \emph{concepts} and to identify relevant ones for differentiating $t_x$ and $t_y$.
A pseudo-algorithm is presented in Algorithm \ref{algorithm:metric}, and illustrated in Figure \ref{fig:approach}\textcolor{red}{(c)}.
A separability matrix $S \in \mathbb{R}^{\Omega \times |\mathcal{C}|}$ is built by computing class separability scores for all pair-wise classes, \ie $\forall \; (t_x, t_y) \in \Omega:= {|\mathcal{T}| \choose 2}$ and $\forall c \in \mathcal{C}$.
\textbf{Statistics of Separability Score:}
Since explainability is not uniquely defined, we include multiple metrics highlighting different facets.
We compute three separability statistics $\forall (t_x, t_y) \in \Omega$ using $S$ as given in Equation~\eqref{eqn:stats}, \ie
\begin{inparaenum}[(1)]
\item \emph{maximum}: the utmost separability,
\item \emph{average}: the expected separability. These two metrics encode (model+explainer)'s focus, \ie ``how much the black-box model implicitly uses the \emph{concepts} for class separability?"
\item \emph{correlation}: encodes the agreement between (model+explainer)'s focus and pathological prior $P$. $P \in \mathbb{R}^{\Omega \times |\mathcal{C}|}$ signifies the relevance $\forall c \in \mathcal{C}$ for differentiating $(t_x, t_y) \in \Omega$, \eg nuclear \emph{size} is highly relevant for classifying benign and malignant tumor as important nuclei in malignant are larger than important nuclei in benign.
\end{inparaenum}
\begin{align}
\label{eqn:stats}
\centering
\begin{aligned}
s_{\max}(t_x,t_y) &= \max_{c \in \mathcal{C}}S_{(t_x, t_y), c} \\
s_{\text{avg}}(t_x,t_y) &= \frac{1}{|C|} \sum_{c \in \mathcal{C}}S_{(t_x, t_y), c} \\
s_{\text{corr}}(t_x,t_y) &= \rho(S_{(t_x, t_y), c=1,..,|\mathcal{C}|}, P_{(t_x, t_y), c=1,..,|\mathcal{C}|})
\end{aligned}
\end{align}
where $\rho$ denotes Pearson correlation.
\emph{$s_{\max}$}, \emph{$s_{\text{avg}}$} $\!\in\!$ [0,$\rotatebox{90}{8}$) show separation between unnormalized class-histograms; and
\emph{$s_{\text{corr}}$} $\in$ [-1, 1] shows agreement between $S$ and $P$.
We build $S_{\max}$, $S_{\text{avg}}$ and $S_{\text{corr}}$ by computing Equation~\eqref{eqn:stats} $\forall (t_x, t_y) \in \Omega$.
Metrics' complementary may lead to relevant \emph{concepts} different to pathological understanding.
\iffalse
Specifically, for each class-pair $t_x, t_y \in \mathcal{T}$,
\begin{inparaenum}[(i)]
\item The maximum measures the maximum separability between $t_x$ and $t_y$ for any $c \in \mathcal{C}$.
\item The expectation measures the average separability between $t_x$ and $t_y$ for all $c \in \mathcal{C}$.
\item The correlation measures the correlation between concept-wise separability scores and prior pathological knowledge.
\end{inparaenum}
\fi
\textbf{Risk:}
We \emph{conceptually} introduce the notion of risk as a weight to indicate the cost of misclassifying a sample of class $t_x$, erroneously as class $t_y$ \cite{ThaiNghe2010, He2013}.
Indeed, misclassifying a malignant tumor as a benign tumor is riskier than misclassifying it as an atypical tumor.
Thus, we construct a risk vector $R \in \mathbb{R}^{\Omega}$. In this work, each entry in $R$ defines the symmetric risk of differentiating $t_x$ from $t_y$ measured as the number of class-hops needed to evolve from $t_x$ to $t_y$.
\textbf{Metrics:}
Finally, we propose three quantitative metrics based on class separability to assess an explainer quality. The metrics are computed as the risk weighted sum of the statistics of separability scores, \ie,
\begin{inparaenum}[(1)]
\item \emph{maximum separability} $S_{\max, R} := S_{\max} \odot R$,
\item \emph{average separability} $S_{\text{avg}, R} := S_{\text{avg}} \odot R$,
\item \emph{correlated separability} $S_{\text{corr}, R} := S_{\text{corr}} \odot R$, where $\odot$ defines the Hadamard product.
\end{inparaenum}
The first two metrics are pathologist-independent, and the third metric requires expert pathologists to impart the domain knowledge in the form of pathological prior $P$. Such prior can be defined individually by a pathologist or collectively by consensus of several pathologists, and it is independent of the algorithm generated explanations.
\begin{algorithm}[ht]
\KwIn{$\mathcal{D}=\{(D_i^t, \mathcal{I}_i^t) \}, t\in \mathcal{T}, i \in N_t$ }
\Parameter{$\mathcal{T}$, $\mathcal{C}$, $\mathcal{A}_c$, $\mathcal{K}$}
\KwResult{$S \in \mathbb{R}^{{|\mathcal{T}| \choose 2} \times |\mathcal{C}|}$}
\For(\tcp*[h]{go over concepts}){c in $\mathcal{C}$ }{
\For(\tcp*[h]{go over nuclei thresh}){k in $\mathcal{K}$}{
\For(\tcp*[h]{go over attributes}){a in $\mathcal{A}_c$}{
\For(\tcp*[h]{go over classes}){t in $\mathcal{T}$}{
$\mbox{var} \gets D_i^t(a)[:k]$ \tcp*[h]{sorted $I_i^t$}
$H_t^{(k)}(a) \gets \mbox{histogram}(\mbox{var})$
}
\For(\tcp*[h]{go over class pairs}){$(t_x,t_y)$ in ${|\mathcal{T}| \choose 2}$}{
$s_a^{(k)}(t_x, t_y) \gets d(H_{t_x}^{(k)}(a), H_{t_y}^{(k)}(a))$
}
}
$s_c^{(k)}(t_x, t_y) \gets \frac{1}{|\mathcal{A}_c|} \sum_{a \in \mathcal{A}_c} s_a^{(k)}(t_x, t_y)$
}
$S_{(t_x, t_y), c} \gets \mbox{AUC}_{k \in \mathcal{K}} (s_c^{(k)}(t_x, t_y))$
}
\caption{Class separability computation.}
\label{algorithm:metric}
\end{algorithm}
\iffalse
\begin{algorithm}[ht]
\SetAlgoLined
\KwIn{$\mathcal{D}=\{(D_i^t, \mathcal{I}_i^t) \}, t\in \mathcal{T}, i \in N_t$ }
\Parameter{$\mathcal{T}$, $\mathcal{C}$, $\mathcal{A}_c$, $\mathcal{K}$}
\KwResult{$S \in \mathbb{R}^{{|\mathcal{T}| \choose 2} \times |\mathcal{C}|}$}
\For(\tcp*[h]{go over concepts}){c in $\mathcal{C}$ }{
\For(\tcp*[h]{go over attributes}){a in $\mathcal{A}_c$}{
\For(\tcp*[h]{go over nuclei thresh}){k in $\mathcal{K}$}{
\For(\tcp*[h]{go over classes}){t in $\mathcal{T}$}{
$\mbox{var} \gets \{D_i^t(a)|I_i^t>\text{top}(k), \forall i \in N_t\}$\;
$H_t^{(k)}(a) \gets \mbox{histogram}(\mbox{var})$\;
}
\For(\tcp*[h]{go over class pairs}){$(t_x,t_y)$ in ${|\mathcal{T}| \choose 2}$}{
$s_a^{(k)}(t_x, t_y) \gets d(H_{t_x}^{(k)}(a), H_{t_y}^{(k)}(a))$
}
}
$S_{(t_x, t_y), a} \gets \mbox{AUC}_{k \in \mathcal{K}} (s_a^{(k)}(t_x, t_y))$\;
}
$S_{(t_x, t_y), c} \gets \frac{1}{|\mathcal{A}_c|} \sum_{a \in \mathcal{A}_c} S_{(t_x, t_y), a}$ \;
}
\caption{Class separability computation.}
\label{algorithm:metric}
\end{algorithm}
\fi
\section{Results}
This section describes the analysis of $\mathrm{CG}$ explainability for breast cancer subtyping. We evaluate three types of graph explainers and quantitatively analyze the explainer quality using the proposed class separability metrics.
\subsection{Dataset}
We experiment on BReAst Cancer Subtyping (BRACS), a large collection of breast tumor $\mathrm{RoI}$s \cite{Pati2020}. BRACS consists of 4391 $\mathrm{RoI}$s at 40$\times$ resolution from 325 H\&E stained breast carcinoma whole-slides.
The $\mathrm{RoI}$s are annotated by the consensus of three pathologists as,
\begin{inparaenum}[(1)]
\item Benign (B): normal, benign and usual ductal hyperplasia,
\item Atypical (A): flat epithelial atypia and atypical ductal hyperplasia, and
\item Malignant (M): ductal carcinoma \emph{in situ} and invasive.
\end{inparaenum}
The $\mathrm{RoI}$s consist of an average \#pixels=$3.9\pm4.3$ million, and average \#nuclei=$1468\pm1642$, and are stain normalized using \cite{Stanisavljevic2019}. The train, validation, and test splits are created at the whole-slide level, including 3163, 602, and 626 $\mathrm{RoI}$s.
\begin{figure*}[!t]
\centering
\centerline{\includegraphics[width=0.81\linewidth]{figures/fig3.png}}
\caption{Qualitative results. The rows represent the cancer subtypes, \ie Benign, Atypical and Malignant, and the columns represent the graph explainability techniques, \ie $\textsc{GnnExplainer}$, $\textsc{GraphGrad-CAM}$, $\textsc{GraphGrad-CAM++}$, and $\textsc{GraphLRP}$. Nuclei-level importance ranges from blue (the least important) to red (the most important).}
\label{fig:qualitative_results}
\end{figure*}
\subsection{Training}
We conducted our experiments using PyTorch~\cite{Paszke2019} and the Deep Graph Library (DGL)~\cite{Wang2019}.
The $\mathrm{GNN}$ architecture for $\mathrm{CG}$ classification is presented in Section~\ref{sec:graph_learning}.
The $\mathrm{CG}$ classifier was trained for 100 epochs using Adam optimizer~\cite{Kingma2015}, $10^{-3}$ learning rate and 16 batch size.
The best $\mathrm{CG}$-classifier achieved $74.2\%$ weighted F1-score on the test set for the three-class classification.
Average time for processing a 1K$\times$1K $\mathrm{RoI}$ on a NVIDIA P100 GPU is 2s for $\mathrm{CG}$ generation and 0.01s for $\mathrm{GNN}$ inference.
\subsection{Qualitative assessment}
Figure~\ref{fig:qualitative_results} presents explanations, \ie nuclei importance maps, from four studied graph explainers.
We observe that $\textsc{GraphGrad-CAM}$ and $\textsc{GraphGrad-CAM++}$ produce similar importance maps.
The $\textsc{GnnExplainer}$ generates almost binarized nuclei importances.
Interestingly, the gradient and pruning-based techniques consistently highlight similar regions.
Indeed, the approaches focus on relevant epithelial region and unfocus on stromal nuclei and lymphocytes outside the glands.
Differently, $\textsc{GraphLRP}$ produces less interpretable maps through high spatial localization (Figure~\ref{fig:qualitative_results}\textcolor{red}{(d)}) or less spatial localization (Figure~\ref{fig:qualitative_results}\textcolor{red}{(h,l)}).
Qualitative visual assessment of Figure~\ref{fig:qualitative_results} conclude that,
\begin{inparaenum}[(1)]
\item \emph{fidelity} preserving explainers result differently based on the underlying mechanism,
\item high \emph{fidelity} does not guarantee straightforward pathologist-understandable explanations,
\item qualitative assessment cannot rigorously compare explainers' quality, and
\item large-scale tedious pathological evaluation is inevitable to rank the explainers.
\end{inparaenum}
\iffalse
\fi
\begin{table*}[t]
\centering
\begin{tabular}{l|c|ccc|c|c|c|c}
\hline
\multicolumn{2}{l|}{Tasks $(\Omega)$} & B vs. A & B vs. M & A vs. M & \multicolumn{4}{c}{B vs. A vs. M} \\
\hline
\multicolumn{2}{l|}{Accuracy (in \%)} & $77.19$ & $90.29$ & $80.42$ & \multicolumn{4}{c}{$74.92$} \\
\hline
Explainer & \multicolumn{4}{c|}{Metric $\;\;\forall \; (t_x,t_y) \in \Omega$ ($\uparrow$)} & \multicolumn{2}{c|}{Agg. Metric w/o Risk ($\uparrow$)} & \multicolumn{2}{c}{Agg. Metric w/ Risk ($\uparrow$)} \\
\hline
$\textsc{GnnExplainer}$ & \parbox[t]{2mm}{\multirow{5}{*}[-0.4ex]{\rotatebox[origin=c]{90}{$s_{\max}(t_x, t_y)$}}}
& $\mathbf{3.26}$ & $\mathbf{6.24}$ & $\mathbf{3.48}$ &
\parbox[t]{2mm}{\multirow{5}{*}[-0.4ex]{\rotatebox[origin=c]{90}{$S_{\max}$}}}
& $\mathbf{12.98}$ &
\parbox[t]{2mm}{\multirow{5}{*}[-0.4ex]{\rotatebox[origin=c]{90}{$S_{\max,R}$}}}
& $\mathbf{19.22}$ \\
$\textsc{GraphGrad-CAM}$ & & $1.24$ & $4.41$ & $3.36$ & & $9.01$ & & $13.42$ \\
$\textsc{GraphGrad-CAM++}$ & & $1.27$ & $\underline{4.42}$ & $\underline{3.40}$ & & $\underline{9.09}$ & & $\underline{13.51}$ \\
$\textsc{GraphLRP}$ & & $\underline{2.33}$ & $2.46$ & $1.28$ & & $6.07$ & & $8.53$ \\
$\textsc{Random}$ & & $1.02$ & $1.26$ & $1.11$ & & $3.39$ & & $4.65$ \\
\hline
$\textsc{GnnExplainer}$ & \parbox[t]{2mm}{\multirow{5}{*}[-0.4ex]{\rotatebox[origin=c]{90}{$s_{\text{avg}}(t_x, t_y)$}}}
& $\mathbf{1.54}$ & $\mathbf{2.78}$ & $1.93$ &
\parbox[t]{2mm}{\multirow{5}{*}[-0.4ex]{\rotatebox[origin=c]{90}{$S_{\text{avg}}$}}}
& $\mathbf{6.25}$ &
\parbox[t]{2mm}{\multirow{5}{*}[-0.4ex]{\rotatebox[origin=c]{90}{$S_{\text{avg},R}$}}}
& $\mathbf{9.03}$ \\
$\textsc{GraphGrad-CAM}$ & & $1.15$ & $2.57$ & $\underline{2.08}$ & & $5.80$ & & $8.37$ \\
$\textsc{GraphGrad-CAM++}$ & & $1.18$ & $\underline{2.58}$ & $\mathbf{2.09}$& & $\underline{5.85}$ & & $\underline{8.43}$ \\
$\textsc{GraphLRP}$ & & $\underline{1.38}$ & $1.59$ & $1.47$ & & $4.44$ & & $6.03$ \\
$\textsc{Random}$ & & $1.05$ & $1.00$ & $0.95$ & & $3.00$ & & $4.00$ \\
\hline
$\textsc{GnnExplainer}$ & \parbox[t]{2mm}{\multirow{5}{*}[-0.4ex]{\rotatebox[origin=c]{90}{$s_{\text{corr}}(t_x, t_y)$}}}
& $-0.02$ & $0.36$ & $0.38$ &
\parbox[t]{2mm}{\multirow{5}{*}[-0.4ex]{\rotatebox[origin=c]{90}{$S_{\text{corr}}$}}}
& $0.72$ &
\parbox[t]{2mm}{\multirow{5}{*}[-0.4ex]{\rotatebox[origin=c]{90}{$S_{\text{corr},R}$}}}
& $1.08$ \\
$\textsc{GraphGrad-CAM}$ & & $\underline{-0.01}$ & $\underline{0.57}$ & $\underline{0.58}$ & & $\underline{1.14}$ & & $\underline{1.71}$ \\
$\textsc{GraphGrad-CAM++}$ & & $\mathbf{-0.01}$ & $\mathbf{0.58}$ & $\mathbf{0.59}$ & & $\mathbf{1.16}$ & & $\mathbf{1.74}$ \\
$\textsc{GraphLRP}$ & & $-0.15$ & $-0.49$ & $-0.23$ & & $-0.87$ & & $-1.36$ \\
$\textsc{Random}$ & & $-0.37$ & $-0.31$ & $-0.18$ & & $-0.86$ & & $-1.17$ \\
\hline
\end{tabular}
\caption{Quantitative assessment of graph explainers: $\textsc{GnnExplainer}$, $\textsc{GraphGrad-CAM}$, $\textsc{GraphGrad-CAM++}$ and $\textsc{GraphLRP}$, using proposed \emph{maximum, average}, and \emph{correlated separability} metrics. Results are provided for each pair-wise breast subtyping tasks, and are aggregated w/o and w/ risk weighting, \ie $S_{\max}$ and $S_{\max,R}$. The first and second best values are indicated in $\textbf{bold}$ and $\underline{\mbox{underline}}$.}
\label{tab:table1}
\end{table*}
\begin{table*}
\centering
\begin{tabular}{l|ccc|cc}
\hline
Concept \scriptsize{(Attributes)} \normalsize{/ Tasks ($\Omega$)} & B vs. A & B vs. M & A vs. M & w/o risk ($\uparrow$) & w/ risk ($\uparrow$) \\
\hline
Size \scriptsize{(area)} & $\mathbf{3.26}$ & $\mathbf{6.24}$ & $\mathbf{3.47}$ & $\mathbf{12.97}$ & $\mathbf{19.21}$ \\
Shape \scriptsize{(perimeter, roughness, eccentricity, circularity)} & $1.27$ & $2.23$ & $1.60$ & $5.10$ & $7.34$ \\
Shape variation \scriptsize{(shape factor)} & $0.69$ & $2.30$ & $1.99$ & $4.97$ & $7.28$ \\
Density \scriptsize{(mean density, std density)} & $1.01$ & $0.80$ & $0.52$ & $2.33$ & $3.14$ \\
Chromaticity \scriptsize{(GLCM contrast, homogeneity, ASM, entropy, variance)} & $\underline{1.44}$ & $\underline{2.31}$ & $\underline{2.07}$ & $\underline{5.82}$ & $\underline{8.13}$ \\
\hline
\emph{Average separability} ($\uparrow$) & $1.54$ & $2.78$ & $1.93$ & $6.25$ & $9.03$ \\
\hline
\end{tabular}
\caption{Quantification of \emph{concepts} for pair-wise and aggregated class separability in $\textsc{GnnExplainer}$. The first and second best values are indicated in $\textbf{bold}$ and $\underline{\mbox{underline}}$. The per-\emph{concept} \emph{attributes} are presented in the first column.}
\label{tab:table2}
\end{table*}
\subsection{Quantitative results}
For cancer subtyping, relevant \emph{concepts} are nuclear morphology and topology \cite{Rajbongshi2018,Kashyap2018,Nguyen2017,Allison2016}. Here, we focus on nuclear morphology, \ie $\mathcal{C}$ = \{\emph{size}, \emph{shape}, \emph{shape variation}, \emph{density}, \emph{chromaticity}\}.
Table~\ref{tab:table2} lists the \emph{attributes} $\mathcal{A}_c, \forall c \in \mathcal{C}$. In our experiments, we select $\mathcal{K} = \{5, 10, ..., 50 \}$ nuclei per $\mathrm{CG}$.
We further introduce a $\textsc{Random}$ explainer via \emph{random} nuclei selection strategy per $\mathrm{CG}$ to assess a lower bound per quantitative metric.
Table~\ref{tab:table1} presents the statistics of pair-wise class separability and aggregated separability w/ and w/o risk to assess the studied explainers quantitatively. Also, for each class pair $(t_x, t_y)$, we compute classification accuracy by using the $\mathrm{CG}$s of type $t_x$, $t_y$.
Noticeably, $\textsc{GnnExplainer}$ achieves the best \emph{maximum} and \emph{average separability} for majority of pair-wise classes. $\textsc{GraphGrad-CAM++}$ and $\textsc{GraphGrad-CAM}$ followed $\textsc{GnnExplainer}$ except for (B vs. A), where $\textsc{GraphLRP}$ outperforms them. All explainers outperform $\textsc{Random}$ which conveys that the quality of the explainers' explanations are better than random.
Notably, $\textsc{GraphGrad-CAM}$ and $\textsc{GraphGrad-CAM++}$ quantitatively perform very similarly, which is consistent with our qualitative analysis in Figure \ref{fig:qualitative_results}.
Interestingly, a positive correlation is observed between pair-wise class accuracies and \emph{average separability} for the explainers, \ie better classification leads to better \emph{concept} separability, and thus produces better explanations. Further, the observation does not hold for $\textsc{Random}$ generated explanations, which possesses undifferentiable \emph{average concept} separability.
To obtain pathological prior to compute \emph{correlation separability}, we consulted three pathologists to rank the \emph{concepts} in terms of their relevance for discriminating each pair of classes. For instance, given an atypical $\mathrm{RoI}$, we asked how important is nuclear \emph{shape} to classify the $\mathrm{RoI}$ as \emph{not} benign and \emph{not} malignant.
Acquired \emph{concept} ranks for each class pair are \emph{min-max} normalized to output prior matrix $P$.
We observe that $\textsc{GnnExplainer}$, $\textsc{GraphGrad-CAM}$ and $\textsc{GraphGrad-CAM++}$ have positive \emph{correlated separability} for (B vs. M), (A vs. M), and nearly zero values for (B vs. A). It shows that the explanations for (B vs. M) and (A vs. M) bear similar relevance of \emph{concepts} as the pathologists, and focus on a different relevance of \emph{concepts} for (B vs. A).
$\textsc{GraphGrad-CAM++}$ has the best overall agreement at the \emph{concept}-level with the pathologists, followed by $\textsc{GraphGrad-CAM}$ and $\textsc{GnnExplainer}$. $\textsc{Random}$ agrees significantly worse than the three explainers, and $\textsc{GraphLRP}$ has the least agreement.
Table~\ref{tab:table2} provides more insights by highlighting the per-\emph{concept} metrics of $\textsc{GnnExplainer}$. Nuclear \emph{size} is the most relevant \emph{concept}, followed by \emph{chromaticity} and \emph{shape variation}. Comparatively nuclear \emph{density} is the least relevant \emph{concept}.
\section{Conclusion}
In this work, we presented an approach for explaining black-box $\mathrm{DL}$ solutions in computational pathology. We advocated for biological entity-based analysis instead of conventional pixel-wise analysis, thus providing an intuitive space for pathological understanding.
We employed four graph explainability techniques, \ie graph pruning ($\textsc{GnnExplainer}$), gradient-based saliency ($\textsc{GraphGrad-CAM}$, $\textsc{GraphGrad-CAM++}$) and
layerwise relevance propagation ($\textsc{GraphLRP}$), to explain ``black-box" $\mathrm{GNN}$s processing the entity graphs.
We proposed a novel set of user-independent quantitative metrics expressing pathologically-understandable \emph{concepts} to evaluate the graph explainers, which relaxes the exhaustive qualitative assessment by expert pathologists.
Our analysis concludes that the explainer bearing the best class separability in terms of \emph{concepts} is $\textsc{GnnExplainer}$, followed by $\textsc{GraphGrad-CAM++}$ and $\textsc{GraphGrad-CAM}$. $\textsc{GraphLRP}$ is the worst explainer in this category while outperforming a randomly created explanation.
We observed that the explainer quality is directly proportional to the $\mathrm{GNN}$'s classification performance for a pair of classes.
Furthermore, $\textsc{GraphGrad-CAM++}$ produces explanations that best agrees with the pathologists in terms of \emph{concept} relevance, and objectively highlights the relevant set of \emph{concepts}.
Considering the expansion of entity graph-based processing, such as radiology, computation biology, satellite and natural images, graph explainability and their quantitative evaluation is crucial.
The proposed method encompassing domain-specific user-understandable terminologies can potentially be of great use in this direction. It is a meta-method that is applicable to other domains and tasks by incorporating relevant entities and corresponding \emph{concepts}.
For instance, with entity-graph nodes denoting car/body parts in Stanford Cars~\cite{krause2013}/ Human poses~\cite{andriluka2014}, and expert knowledge available on car-model/ activity, our method can infer relevant parts by quantifying
their agreement with experts.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction }\label{intro}
The Cucker-Smale model \cite{CS1,CS2} is a particles system which exhibits the so-called flocking phenomenon, namely the dynamical property of alignment of all the velocities and gathering of all the positions asymptotically when the time of evolution diverges. Its study has developed an intense activity these last years, see for example the intensive bibliography in the recent paper \cite{hakimzhang}.
It consists in the following vector field on $\mathbf{R}^{2dN}$
\begin{equation}\label{CS}
\left\{\begin{array}{rcl}
\dot {x_i}&=&v_i\\
\dot {v_i}&=&
\frac1N\sum\limits_{j=1}^N\psi(x_i-x_j)(v_j-v_i).
\end{array}
\right.
\hskip 1cm x_i,v_i\in\mathbf{R}^d,\
\ i=1,\dots,N
\end{equation}
where $\psi:\mathbf{R}^d\to\mathbf{R}$ is a bounded Lipschitz continuous positive nonincreasing function.
Associated to \eqref{CS}, the following Vlasov type kinetic equation on $\mathbf{R}^{2d}$ was introduced in \cite{ht}:
\begin{equation}\label{vlasovcs}
\partial_t\rho^t(x,v)+v\cdot\nabla_x\rho^t(x,v)+\nabla_v\cdot\left(\rho^t(x,v)\int_{\mathbf{R}^{2d}}\psi(x-y)(v-w)\rho^t(y,w)dydw \right)=0.
\end{equation}
So far the kinetic non-linear equation \eqref{vlasovcs} has been derived in a Lagrangian point of view, i.e. by following the trajectories solving the vector field \eqref{CS} in the so-called empirical measure
\begin{equation}\label{[defempmes}
\Pi_{(x_1,\dots,x_n;v_1,\dots,v_N)}(x,v):=\frac1N\sum_{i=1}^N
\delta(x-x_i)\delta(v-v_i).
\end{equation}
Indeed, let $\Phi^t$ be the flow generated by the system \eqref{CS} and let us define
\begin{equation}\label{defempmest}
\Pi^t_{(x_1,\dots,x_n;v_1,\dots,v_N)}:=\Pi_{\Phi^{-t}(x_1,\dots,x_n;v_1,\dots,v_N)}.
\end{equation}
One shows easily that $\Pi^t_{(x_1,\dots,x_n;v_1,\dots,v_N)}(x,v)$ solves \eqref{vlasovcs} for each $N$.
This point of view follows directly the Dobrushin way \cite{Dobrush}
of deriving the Vlasov equation for the large $N$ limit of Hamiltonian vector fields. The Cucker-Smale model \eqref{CS} and its Vlasov type associated equation \eqref{vlasovcs} have been extensively studied in \cite{hakimzhang} and we quote the following of their results: the function $\psi$ is supposed to be positive, bounded, uniformly Lipschitz continuous and
nonincreasing.
\begin{enumerate}
\item\label{item1}
the solutions of \eqref{CS} satisfy for all $t$
$$
\sup_{i,j\leq N}\|x_i(t)-x_j(t)\|\leq C \ \mbox{ and } \
\sup_{i,j\leq N}\|v_i(t)-v_j(t)\|\leq e^{-Dt}
$$
where $C$ and $D$ are two positive constants depending only on
$\sup_{i,j\leq N}\|x_i(0)-x_j(0)\|$ and
$\sup_{i,j\leq N}\|v_i(0)-v_j(0)\|$.
\item\label{item2}
let $\Pi_{(x_1,\dots,x_n;v_1,\dots,v_N)}\to\mu(0)$ as $N\to\infty$, then
$$
\overline{\lim_{N\to\infty}}\sup_{t\in\mathbf{R}}W_p(\Pi^t_{(x_1,\dots,x_n;v_1,\dots,v_N)}-\mu(t))=0
$$
where $W_p$ is the Wasserstein distance of exponent $p\in\mathbf{N}$ (see definition in Remark \ref{highorders} below).
\item\label{item3}
Let $\int_{\mathbf{R}^{2d}}d\mu(0)=1,\ \int_{\mathbf{R}^{2d}}vd\mu(0)=:\bar v,\ \int_{\mathbf{R}^{2d}}v^2d\mu(0)<\infty$. Then, for some $E,F>0$ and every $p$,
$$
\left(\int\|v-\bar v\|^pd\mu(t)\right)^{\frac1p}\leq Ee^{-Ft}.
$$
\end{enumerate}
\vskip 1cm
Recently, a more Eulerian way of deriving the Vlasov equation associated to Hamiltonian vector fields was introduced (\cite{gmp}, after \cite{gmr}). It consists in considering the movement of a generic particle solution to the Liouville equation associated to an Hamiltonian system, rather than considering the empirical measure associated to the Hamiltonian flow.
Transposed in the (non Hamiltonian) present situation, the method leads to the following.
We will consider the pushforward\footnote{We recall that the pushforward of a measure $\mu$ by a measurable function $\Phi$ is $\Phi\#\mu$ defined by $\int \varphi d(\Phi\#\mu):=\int (\varphi\circ f)d\mu$ for every measurable function $f$.} by the flow $\Phi^t$ associated to \eqref{CS} of a compactly supported, symmetric by permutations of the variables $N$-body probability density
$\rho^{in}_N$ on phase space $\mathbf{R}^{2dN}$.
In order to describe the movement of a ``generic" particle
in the limit of diverging number of particles,
we want to perform an average on the $N$ particles but one. This means that
we consider the first marginal of
\begin{equation}\label{defrhoNt}
\Phi^t\#\rho^{in}_N:=\rho^t_N
\end{equation}
that is
\begin{equation}\label{marg}
\rho^t_{N;1}(x,\xi):=\int_{\mathbf{R}^{2(d-1)N}}
\rho^t_N(x,\xi,x_2,\xi_2,\dots,x_N,\xi_N)dx_2d\xi_2\dots dx_Nd\xi_N.
\end{equation}
Then, when $\rho^{in}_N$ is factorized as $\rho^{in}_N=(\rho^{in})^{\otimes N}$, where $\rho^{in}$ is a probability measure on $\mathbf{R}^{2d}$, we will prove that the marginal of $\rho^t_N$ tends, as $N\to\infty$, to the solution of the effective non-linear Vlasov type equation \eqref{vlasovcs} with
$\rho^{t=0}=\rho^{in}\in L^1(\mathbf{R}^{2d},dxdv)$.
In fact, we can prove (see Remark \ref{highorders} below) that the marginals at every order $n$ of $\rho_N^t$ as defined by
\begin{equation}\label{margn}
\rho^t_{N;n}(x_1,\xi_1,\dots,x_n,\xi_n):=\int_{\mathbf{R}^{2d(N-n)}}
\rho^t_N(x_i,\xi_1,x_2,\xi_2,\dots,x_N,\xi_N)dx_{n+1}d\xi_{n+1}\dots dx_Nd\xi_N,
\end{equation}
tend, in Wasserstein topology of any exponent $p\geq 1$, to the $n$th tensorial power of the solution $\rho^t$ of the Vlasov equation, i.e. $\rho^t_{N;n}\to(\rho^t)^{\otimes n}$. Nevertheless, for sake of clarity of this short note, we will present in detail only the case $n=1$.
This result (for $n=1$), i.e. Theorem \ref{main} below, has to be put in correspondence with the item \ref{item2} of the results of \cite{hakimzhang} just described, but there are several differences. First, our results hold true for more general initial data than the empirical measures. Second, our result will not be uniform in time as in item \ref{item1}. Third, we will get an explicit rate of convergence in the asymptotic $N\to\infty$.
Our methods will also apply to systems of the general form
\begin{equation}\label{genCS}
\left\{\begin{array}{rcl}
\dot {x_i}&=&v_i\\
\dot {v_i}&=&
\frac1N\sum\limits_{j=1}^N\gamma(x_i-x_j,v_i-v_j)
\end{array}
\right.
\hskip 1cm x_i,v_i\in\mathbf{R}^d,\
\ i=1,\dots,N,
\end{equation}
where $\gamma(x,v):\mathbf{R}^{2d}\to\mathbf{R}^d$ is a Lipschitz continuous function, bounded in $(x,v)$, Corollary \ref{maingen} or bounded in $x$ and sublinear in $v$, Theorem \ref{mainsublin} and Corollary \ref{maingensublinest}.
In the following, we will denote by $\operatorname{Lip}(\gamma)_{(x,v)}$ the (local) Lipschitz constant of $\gamma$ at the point $(x,v)\in\mathbf{R}^{2d}$ and we will suppose, without loss of generality, that there exists $\gamma_0>0$ such that, for all $(x,v)\in\mathbf{R}^{2d}$,
\begin{eqnarray}
\gamma(x,v)&\leq&\gamma_0|v|\nonumber\\
\operatorname{Lip}(\gamma)_{(x,v)}&\leq&\gamma_0|v|\nonumber
\end{eqnarray}
where $|\cdot|$ denotes the Euclidean norm on $\mathbf{R}^d$.
We will show in Appendix \ref{dynestpar} the gobal existence of the solutions to \eqref{genCS}, with an exponential growth in time with respect to the initial conditions. Of course, uniqueness is a consequence of (the iteration in time of) the Cauchy-Lipschitz Theorem.
The Cucker-Smale vector fields satisfies this assumptions with $\gamma_0=\|\psi\|_{L^\infty}$.
Note that the following Vlasov type kinetic equation on $\mathbf{R}^d$ is obviously associated to \eqref{genCS}, in a natural way:
\begin{equation}\label{vlasovgencs}
\partial_t\rho^t(x,v)+v\cdot\nabla_x\rho^t(x,v)+\nabla_v\cdot\left(\rho^t(x,v)\int_{\mathbf{R}^{2d}}\gamma(x-y,v-w)\rho^t(y,w)dydw \right)=0.
\end{equation}
Let us immediately notice that our assumptions on $\gamma$ contain the two assumptions in Theorem 2.3 in \cite{prt}. The first one because of the Lipschitz and sublinearity properties of $\gamma$, and the second one thanks to the fact, see \cite{VillaniAMS,VillaniTOT}, that
$$
\sup_{|\operatorname{Lip}(\varphi)|\leq 1}|\int \varphi(\mu-\nu)|\leq W_1(\mu,\nu)
$$
where $W_1$ is the Wasserstein distance of exponent $1$, as defined in Remark \ref{highorders} in the present paper.
Therefore, thanks to the results of \cite{prt} (see also \cite{carr2}), there exists a unique continuous solution to \eqref{vlasovgencs} in $ C_0(\mathbf{R},\mathcal P_c(\mathbf{R}^{2d}))$, where
$\mathcal P_c(\mathbf{R}^{2d})$ is the space of compactly supported probability measures on $\mathbf{R}^{2d}$.
Our last result, Theorem \ref{inverse} and Remark \ref{margplusinv}, gives an insight of the ``inverse problem" of the mean-field limit in the case where the solution to the Vlasov kinetic equation \eqref{vlasovgencs} exhibits a flocking behavior of the form \eqref{suppcs} below, without any knowledge concerning the flocking behavior of the original particles system. Namely, when the support of the solution to the Vlasov equation remains of finite size in the variable $x$ and reduces asymptotically exponentially in time to a single point in the momentum variable $v$, we prove an approximate similar behavior for all the marginals of initial distributions
pushforwarded by the $N$-body dynamics, for $N$ large enough.
\vskip 0.5cm
Let us finish this introduction by quoting very few references among the huge number of works dedicated to the rigorous derivation of the mean-field limit of particles systems, after the pioneering work \cite{Dobrush} already mentioned: \cite{szn, carr1,jw2} for stochastic systems and using empirical measures, \cite{haliu,hakimzhang} (already mentioned) specifically for the Cucker-Smale model and using empirical measures too, \cite{gmr,gmp} (already mentioned too) for globally Lipschitz forces, \cite{jw1} for rough but bounded forces and finally \cite{gp2} for a derivation using the full hierarchy of equations satisfied by the marginals, for Hamiltonian systems with analytic potentials.
\section{The results}\label{result}
In this section we will state our main results, but let us start by recalling the definition of the second order Wasserstein distance $W_2$ (see \cite{ambrosiosavare,VillaniAMS,VillaniTOT}).
\begin{Def}[quadratic Wasserstein distance]\label{defwas}
The Wasserstein distance of order two between two probability measures $\mu,\nu$ on $\mathbf{R}^m$ with finite second moments is defined as
$$
W_2(\mu,\nu)^2
=
\inf_{\gamma\in\Gamma(\mu,\nu)}\int_{\mathbf{R}^m\times\mathbf{R}^m}
|x-y|^2\gamma(dx,dy)
$$
where $\Gamma(\mu,\nu)$ is the set of probability measures on $\mathbf{R}^m\times\mathbf{R}^m$ whose marginals on the two factors are $\mu$ and $\nu$.
That is to say that elements $\gamma$ of $\Gamma(\mu,\nu)$, called couplings or transportation maps or transport maps or just maps, depending of the authors, satisfy, for every test functions $a$ and $b$ in $C_0(\mathbf{R}^d)$,
$$
\int_{\mathbf{R}^{2m}}a(x)\gamma(dx,dy)=\int_{\mathbf{R}^m}a(x)\mu(dx),\ \ \
\int_{\mathbf{R}^{2m}}b(y)\gamma(dx,dy)=\int_{\mathbf{R}^m}b(y)\nu(dy).
$$
\end{Def}
The symmetry property in $\mu,\nu$ is obvious and the separability property is easily proven by taking the optimal coupling between $\mu$ and itself equal to
\begin{equation}\label{coupop}
\gamma=\mu\delta(x-y).
\end{equation}
Conversely, if $W_2(\mu,\nu)^2=0$, then $
\int_{\mathbf{R}^m\times\mathbf{R}^m}
|x-y|^2\gamma(dx,dy)=0$ for some $\gamma$ so that $x=y$ $\gamma$ a.e. and, for every Borel function $\varphi$,
$$
\int\varphi(x)\mu(dx)=\int\varphi(x)\gamma(dx,dy)=\int\varphi(y)\gamma(dx,dy)=
\int\varphi(y)\nu(dy)\Rightarrow\mu=\nu.
$$
The proof of the triangular inequality is much more involved, see again \cite{ambrosiosavare,VillaniAMS,VillaniTOT}.
\vskip 1cm
We can now state the main result of this note, proven in Section \ref{Proofmain}.
\begin{Thm}[Cucker-Smale model]\label{main}\
Let $\Phi^t$ be the flow generated by the system \eqref{CS}, $\psi$ bounded positive nonincreasing Lipschitz continuous, and let $\rho^t$ be the solution to \eqref{vlasovcs} with an initial condition $\rho^{in}\in L^1(\mathbf{R}^{2d})$ compactly supported.
Let moreover $\rho^t_{N;1}$ be the first marginal of $\rho^t_N:=\Phi^t\#({\rho^{in}})^{\otimes N}$, as defined in \eqref{marg}.
Then, for all $N>1,\ t\in \mathbf{R}$,
\begin{equation}\nonumber
W_2(\rho^t_{N;1},\rho^t)
\leq
4
\|\psi\|^2_\infty
(2|\bar v|+|supp[\rho^{in}]|)
\left(\frac{e^{Lt}-1}L\right)^{\frac12}
{N^{-\frac12}}
\end{equation}
with
\begin{equation}\nonumber
L:=2(1+8\|\psi\|^2_\infty\|v\|_{L^\infty(supp[\rho^{in}])}^2)\ and \ \bar v=\int v\rho^{in}dxdv,
\end{equation}
where $\|v\|_{L^\infty(supp[\rho^{in}])}:=\sup\limits_{(x,v)\in supp[\rho^{in}]}|v|$.
\end{Thm}
\begin{Rmk}\label{margplus}
There exists an equivalent result for higher orders marginals of $\rho(t)$ and other Monge-Kantorovich distances but we prefer in this note to concentrate on the case of first marginal and quadratic Wasserstein distance. The proof in the more general situation is very close to the one presented here. See Remark \ref{highorders} for some details.
\end{Rmk}
\vskip 1cm
Theorem \ref{main} is actually a corollary of the following more general result, proven in Section \ref{Proofmainsublin}.
\begin{Thm}[General Cucker-Smale model with general sublinear force]\label{mainsublin}\
Let $\Phi^t$ the flow defined by the system \eqref{genCS}, $\gamma(x,v)$ Lipschitz continuous bounded in $x$ and sublinear in $v$, and let $\rho^t$ be the solution to \eqref{vlasovgencs}, $\gamma$ bounded on $\mathbf{R}^{2d}$, with an initial condition $\rho_1^{in}\in L^1(\mathbf{R}^{2d})$ compactly supported.
Finally, let $\rho^t_{N;1}$ be the first marginal of $\rho^t_N:=\Phi^t\#({\rho^{in}})^{\otimes N}$, as defined in \eqref{marg}.
Then, for all $N>1, t\in \mathbf{R}$,
\begin{equation}\nonumber
W_2(\rho^t_{N;1},\rho^t)
\leq
C(t)N^{-\frac12}
\end{equation}
with
\begin{eqnarray}\nonumber
C(t)&:=&
\left(
4
\int_0^t
\sup_{(x,v),(x',v')\in supp[\rho^t]}|\gamma(x-x',v-v')|^2
e^{\int_s^tL (u)du}ds
\right)^{\frac12}\,
,\nonumber\\
{\beta}
L(u)&:=&
2(1+2
\min{(\sup_{\substack{i,l=1,\dots,N\\
(Y,\Xi)\in supp[\rho^u_N]}}\operatorname{Lip}{(\gamma)}^2_{(y_i-y_l,\xi_i-\xi_l)},\sup_{\substack{
(x,v),(x',v')\in supp[\rho^u]}}\operatorname{Lip}{(\gamma)}^2_{(x-x',v-v')})})
).\nonumber
\end{eqnarray}
\end{Thm}
The following result is elementarly derived from Theorem \ref{mainsublin}.
\begin{Cor}[General Cucker-Smale model with bounded global Lipschitz force]\label{maingen}\
Let $\rho^t$ be the solution to \eqref{vlasovgencs}, $\gamma$ globaly Lipschitz and bounded on $\mathbf{R}^{2d}$, with an initial condition $\rho^{in}\in L^1(\mathbf{R}^{2d})$ and $\Phi^t$ the flow deinfed by the system \eqref{genCS}. Finally, let $\rho^t_{N;1}$ be the first marginal of $\rho^t_N:=\Phi^t\#({\rho^{in}})^{\otimes N}$, as defined in \eqref{marg}.
Then, for all $N>1, t\in \mathbf{R}$,
\begin{equation}\nonumber
W_2(\rho^t_{N;1},\rho^t)
\leq
C(t)N^{-\frac12}
\end{equation}
with
\begin{equation}\nonumber
C(t):=\left(4\|\gamma\|_\infty
\frac{e^{\Lambda t}-1}{\Lambda}\right)^{\frac12}
\ \ \ \ \Lambda:=2(1+2\operatorname{Lip}(\gamma)^2).
\end{equation}
\end{Cor}
Note that no need of compacity of the support of $\rho^{in}$ is needed any more in Corollary \ref{maingen}.
Theorem \ref{mainsublin} gives a precise estimate involving, through the expression of the functions $C(t)$ and $L(t)$, the knowledge of the size of the support of the initial data propagated by the particle flow $\Phi(t)$ driven by \eqref{vlasovgencs} and the kinetic flow induced by \eqref{vlasovgencs}. This information might be given by the explicit models, i.e. the function $\gamma$, as it is the case for the Cucker-Smale model (see also the very end of this section).
In the general case, the boundness in space and sublinearity in velocities of $\gamma$ allow to control the increasing of the flow $\Phi(t)$, and leads to our next result, whose proof is given in Section \ref{proofcor2} below.
\begin{Cor}\label{maingensublinest}
Under the same hypothesis as in Theorem \ref{mainsublin} we have that, for all $N>1, t\in \mathbf{R}$,
\begin{equation}\nonumber
W_2(\rho^t_{N;1},\rho^t)
\leq
C(t)N^{-\frac12}
\end{equation}
with
\begin{eqnarray}
C(t)
&=&
2\gamma_0(\|v\|_{L^\infty(supp[\rho^{in}])}+
\|v\|_{L^1(supp[\rho^{in}])})\nonumber\\
&&\times \left( 2e^{\left(e^{4\gamma_0t}4\gamma_0
(\|v\|_{L^\infty(supp[\rho^{in}])}+
\|v\|_{L^1(supp[\rho^{in}])})^2\right)}e^{4t}(e^{4\gamma_0t}-1)
\right)^{\frac12}.
\end{eqnarray}
\end{Cor}
\vskip 1cm
Let us finish this section of results by a simple remark.
Flocking properties of the solution of the Vlasov kinetic equation \eqref{vlasovcs} can be derived from the corresponding properties of the particle system \eqref{CS}, as in \cite{haliu,hakimzhang}. But they can also be derived by a direct PDE study of \eqref{vlasovcs}, as in \cite{prt}. This suggest a kind of inverse questioning: suppose one determines some flocking properties for the solution of a general kinetic Vlasov equation, e.g. \eqref{vlasovgencs}. Does this infer on flocking properties for the corresponding particle system, e.g. \eqref{genCS}? Our Corollary \ref{maingensublinest} gives some insight on this problem, as it tells us quantitatively how close is the solution of the kinetic equation (and its tensorial powers) to the marginals of any orders of the pushforward obey the corresponding particle flow of a general $N$ particle density.
More precisely, suppose that the solution $\rho^t$ to \eqref{vlasovgencs} satisfies the following property:
\begin{equation}\label{suppcs}
supp[\rho^t]\subset B(\bar x+t\bar v,X)\times B(\bar v,Ve^{-\alpha t})
\end{equation}
for some $(\bar x,\bar v)\in\mathbf{R}^{2d}$ and some positive constants $ X, V,\alpha$. Here $B(\bar w,W)$ designates the ball of center $\bar w$ and radius $W$ in $\mathbf{R}^d$.
This implies easily that $L(u)$ in Theorem \ref{mainsublin} can be easily estimated by $$
L(u)\leq 2(1+8\gamma_0^2(\bar v^2+V^2e^{-2\alpha u}))
\leq 2(1+8\gamma_0^2(\bar v^2+V^2))
$$
and, therefore, in the same Theorem,
\begin{equation}\label{cdet}
C(t)\leq
4\gamma_0(\bar v^2+V^2)\frac{e^{2(1+8(\bar v^2+V^2))t}-1}{2(1+8(\bar v^2+V^2)}.
\end{equation}
Note that in \eqref{cdet}, $C(t)$ has an exponential growth, and not a double exponential one as in Corollary \ref{maingensublinest}. In particular, since \eqref{suppcs} holds true for the Cucker-Smale models by Theorem 3.1, equation $(3.2)$, in \cite{prt}, \eqref{cdet} provides an alternative proof of Theorem \ref{main} with (slightly) different values of the constants involved in its statement.
Of course the first marginal $\rho_{N;1}^t$ of the pushforward of $(\rho^{in})^{\otimes N}$ by the flows induced by the system \eqref{genCS} has no reason for being compactly supported, as we did not impose anything on the particle flow. Nevertheless, the following result provides for $\rho_{N;1}^t$ a weak version of the support property \eqref{suppcs}.
\begin{Thm}\label{inverse}
With the same notations and hypothesis as in Theorem \ref{mainsublin}, let us suppose moreover that $\rho^t$ satisfies \eqref{suppcs}.
For all $t,\epsilon>0$ let us define
$$
N_{t,\epsilon}:=\left(\frac{C(t)}\epsilon\right)^2,
$$ where $C(t)$ is the function defined by \eqref{cdet}.
Then, for every $N\geq N_{t,\epsilon}$ and every Lipschitz function $\varphi$ on $\mathbf{R}^{2d}$ of Lipschitz constant smaller than $1$,
$$
\int\limits_{
\mathbf{R}^{2d}\setminus
B(\bar x+t\bar v,X)\times B(\bar v,Ve^{-\alpha t})}
\varphi(x,v)\rho_{N;1}^tdxdv\leq \epsilon.
$$
\end{Thm}
\begin{Rmk}\label{margplusinv}
Note that, by Remark \ref{margplus}, the same type of result is also true for marginals of any order, by cooking up a new value of $N_{t,\epsilon}$, i.e. $C(t)$, using the constants given in Remark \ref{margplusinv} below:
for all $n=1,\dots,N,\ t,\ \epsilon>0$, there exists $ N_{t,\epsilon,n}$ such that, for every $N\geq N_{t,\epsilon,n}$ and every Lipschitz function $\varphi$ on $\mathbf{R}^{2dn}$ of Lipschitz constant smaller than $1$,
$$
\int\limits_{
\mathbf{R}^{2dn}\setminus
(B(\bar x+t\bar v,X)\times B(\bar v,Ve^{-\alpha t}))^{\times n}}
\varphi(x,v)\rho_{N;n}^tdxdv\leq \epsilon.
$$
\end{Rmk}
\begin{proof}
It is an immediate consequence of Theorem \ref{mainsublin} or Theorem \ref{maingensublinest}, and the fact that the Wasserstein distance induces a metric for the weak topology in the sense that
$$
\sup_{\operatorname{Lip}{(\varphi)}\leq 1}\int(\mu-\nu)\varphi(x,v)dxdv\leq W_2(\mu.\nu)
$$
for every probability measures $\mu$ and $\nu$
\end{proof}
\section{Proofs}\label{proofs}
\subsection{Preliminaries}\label{mprelem}
Let us first recall the general situation we are dealing with, in order also to fix the notations.
We consider on $\mathbf{R}^{2dN}$ the following Cucker-Smale type vector field
\begin{eqnarray}\label{eq1}
\dot {x_i}&=&v_i\\
\dot {v_i}&=&G_i(
X,V),
\ i=1,\dots,N\nonumber
\end{eqnarray}
where
\begin{equation}\label{defG}
G_i(
X,V)
=\frac1N\sum_{j=1}^N\gamma(x_i-x_j,v_i-v_j).
\end{equation}
Here the function $\gamma(x,v):\mathbf{R}^{2d}\to\mathbf{R}^d$
is a Lipschitz, bounded in $x$ and sublinear in $v$, continuous function., such that $\gamma(x,v),\operatorname{Lip}(\gamma)_{(x,v)}\leq \gamma_0|v|$.
We used the notation $X=(x_1,\dots,x_N),V=(v_1,\dots,v_N)$.
In fact, we are rather interested in the Liouville equation associated to \eqref{eq1} \cite{ht}, namely
\begin{equation}\label{liouvnd}
\partial_t\rho^t_N+V\cdot\nabla_X\rho^t_N+
\sum_{i=1}^N\nabla_{v_i}.(G_i\rho^t_N)=0,\ \rho^{t=0}_N=\rho^{in}_N
\end{equation}
with $\rho^{in}_N\in\mathcal P(\mathbf{R}^{2dN})$.
Although the argument is standard, let us recall why the solution of \eqref{liouvnd} is equal to the pushforward of $\rho^{in}_N$ by the flow $\Phi_t$ generated by \eqref{eq1}: integrating $\rho^{in}_N$ against a test $\varphi$ function composed by $\Phi^t$ gives
\begin{equation}\label{liouv3}
\partial_t\int\varphi(\Phi^t(X,V))\rho^{in}_N(X,V)dXdV
=\int\varphi(X,V)\partial_t(\Phi^t\#\rho^{in}_N(X,V))dXdV.
\end{equation}
On the other side
\begin{eqnarray}\label{liouv4}
\partial_t\int\varphi(\Phi^t(X,V))\rho^{in}_N(X,V)dXdV
&=&\int (\dot \Phi^t\cdot\nabla_{(X,V)}\varphi)(\Phi^t(X,V))\rho^{in}_N(X,V)dXdV\nonumber\\
&=&\int((V,G)\cdot\nabla_{(X,V)}\varphi)(\Phi^t(X,V)))\rho^{in}_N(X,V)dXdV\nonumber\\
&=&\int((V,G)\cdot\nabla_{(X,V)}\varphi)(X,V))(\Phi^t\#\rho^{in}_N(X,V)dXdV\nonumber\\
&=&
-\int\varphi(X,V)\nabla_{(X,V)}\cdot\big((V,G)\Phi^t\#\rho^{in}_N(X,V)\big)dXdV,\label{retruc2}
\end{eqnarray}
so that \eqref{liouv3} and \eqref{retruc2} implies that $\Phi^t\#\rho^{in}_N$ solves \eqref{liouvnd}.
We want to prove that the marginals of $\rho^t_N$ tend, as $N\to\infty$, to
the solution of a Vlasov type equation.
Let us recall that such Vlasov-type equation associated to \eqref{liouvnd}, introduced in \cite{ht} for the Cucker-Smale model, reads
\begin{equation}\label{vlasnd}
\partial_t
\rho^t(x,v)
+v\cdot\nabla_x\rho^t(x,v)+
\nabla_v\cdot (G_{\rho^t}\rho^t(x,v))=0
,\ \ \
\rho^{t=0}=\rho^{in}_1\in L^1(\mathbf{R}^{2d},dxdv),
\end{equation}
with
\begin{equation}\label{defgrho}
G_{\rho}(x,v)=\int_{\mathbf{R}^{2d}}\gamma(x-y,v-w)\rho(y,w)dydw.
\end{equation}
We can now prove the main results of this paper.
\subsection{Proof of Theorem \ref{mainsublin}}\label{Proofmainsublin}
The proof will be articulated around the four lemmas which follow.
We will denote $X=(x_1,\dots,x_N),V=(v_1,\dots,v_N),Y=(y_1,\dots,y_N), \Xi=(\xi_1,\dots,\xi_N)$.
Let $\pi^{in}_N$ be defined by
$$
\pi^{in}_N(Y,\Xi,X,V):=(\rho^{in})^{\otimes N}(X,V)\delta(X-Y)\delta(V-\Xi).
$$
Obviously $\pi_N^{in}\in\Pi((\rho^{in})^{\otimes N},(\rho^{in})^{\otimes N})$. Moreover, as mentioned before,
\begin{equation}\label{inop}
\int_{\mathbf{R}^{2dN}\times \mathbf{R}^{2dN}}
(|Y-X|^2+|\Xi-V|^2)
\pi^{in}_N(dY,d\Xi,dX,dV)=0,
\end{equation}
so that $\pi^{in}_N$ is an optimal coupling between $(\rho^{in})^{\otimes N}$ and itself.
The following first Lemma will be one of the keys of the proof of Theorem \ref{main}. It consists in considering in evolving a coupling $\pi^{in}_N$ of two initial conditions of the Liouville \eqref{liouvnd} and Vlasov \eqref{vlasnd} equations by the two dynamics of each factor of $\pi^{in}_N$.
\begin{Lem}\label{lemone}
Let $\pi_N(t)$ be the unique (measure) solution to the following linear transport equation
\begin{equation}\label{lintrans}
\partial_t\pi_N+V\cdot\nabla_X\pi_N+\Xi\cdot\nabla_Y\pi_N
+
\sum_{i=1}^N\left(\nabla_{\xi_i}\cdot(G_i(Y,\Xi)\pi_N)+\nabla_{v_i}\cdot(G_{\rho^t}(x_i,v_i))\pi_N)\right)=0
\end{equation}
with $\pi_N(0)=\pi_N^{in}$.
Then, for all $t\in\mathbf{R}$, $\pi_N(t)$ is a coupling between $\rho^t_N$ and $(\rho^t)^{\otimes N}$.
\end{Lem}
\begin{proof}
By taking the two marginals of the two sides of the equality, one gets that they satisfy the two Liouville and Vlasov equations. The result is then obtained by uniqueness of the solutions of both equations.
\end{proof}
\begin{Lem}\label{lemtwo}
Let
\begin{eqnarray}
D_N(t)&:=&\int\frac1N
\sum_{j=1}^N(|x_j-y_j|^2+|v_j-\xi_j|^2)d\pi_N(t)\nonumber\\
&=&
\frac1N\int
((X-Y)^2+(V-\Xi)^2)
d\pi_N(t)\,.\nonumber
\end{eqnarray}
Then
$$
\begin{aligned}
\frac{dD_N}{dt}\le{\bf L}(t) D_N+\frac{1}N\sum_{j=1}^N\int\left|
G_{\rho^t}(x_i,v_i)-G_i(X,V)
\right|^2
{(\rho^t)^{\otimes N}}dXdV
&\,,
\end{aligned}
$$
with
\begin{eqnarray}
{\bf L}(t):=&&
\label{defl}\\
2(1+2
\min{(\sup_{\substack{i,l=1,\dots,N\\
X,V,Y,\Xi\in supp[\pi_N(t)]}}\operatorname{Lip}{(\gamma)}^2_{(x_i-x_l,v_i-v_l)},\sup_{\substack{i,l=1,\dots,N\\
X,V,Y,\Xi\in supp[\pi_N(t)]}}\operatorname{Lip}{(\gamma)}^2_{(y_i-y_l,\xi_i-\xi_l)}})&&
\nonumber
).
\end{eqnarray}
\end{Lem}
\begin{proof}
We first notice that
$$
\frac{dD_N}{dt}
=\frac2N\int
\left((V-\Xi).(X-Y)+
\sum_{i=1}^N(v_i-\xi_i).(G_{\rho^t}(x_i,v_i)-G_i(Y,\Xi))\right)d\pi_N.
$$
Using $2uv\leq u^2+v^2$ we get
\begin{eqnarray}\label{trucenplume}
\frac{dD_N}{dt}
&\leq&\frac1{N}\int\left((X-Y)^2+2(V-\Xi)^2)
+
\sum_{i=1}^N
|G_i(Y,\Xi)-G_{\rho^t}(x_i,v_i)|^2\right)d\pi_N
\nonumber\\
&\leq&
2D_N(t)
+\frac1{N}\int\left(
\sum_{i=1}^N
|G_i(Y,\Xi)-G_{\rho^t}(x_i,v_i)|^2\right)d\pi_N.
\end{eqnarray}
Let us
add to $G_i(Y,\Xi)-G_{\rho^t}(x_i,v_i)$
the null term
$
\big(G_i(X,V)-G_i(X,V)\big)
$
so that
$$
|G_i(Y,\Xi)-G_{\rho^t}(x_i,v_i)|^2\leq
2\big(|G_i(Y,\Xi)-G_i(X,V)|^2+|G_i(X,V)-G_{\rho^t}(x_i,v_i)|^2\big)
$$
Let us first estimate
\begin{eqnarray}
&&|G_i(Y,\Xi)-G_i(X,V)|^2\leq
\min{(\operatorname{Lip}(G_i)_{(X,V)}^2,\operatorname{Lip}(G_i)_{(Y,\Xi)}^2)}\big(|X-Y|^2+|V-\Xi|^2\big)\nonumber.
\end{eqnarray}
By \eqref{defG}, we have
\begin{eqnarray}
&&\min{(\operatorname{Lip}(G_i)_{(X,V)}^2,\operatorname{Lip}(G_i)_{(Y,\Xi)}^2)}\nonumber\\
&\leq&
\min{(\sup_{l=1,\dots,N}\operatorname{Lip}{(\gamma(x_i-x_l,v_i-v_l)}^2,\sup_{l=1,\dots,N}\operatorname{Lip}{(\gamma(y_i-y_l,\xi_i-\xi_l)}^2)}.\nonumber
\end{eqnarray}
Therefore, by convexity,
\begin{eqnarray}
&&\frac2{N}\int\sum_{i=1}^N
|G_i(Y,\Xi)-G_i(X,V)|^2
\pi_N(dX,dV,dY,d\Xi)\nonumber\\
&\leq &
4\min{(\sup_{\substack{i,l=1,\dots,N\\
X,V,Y,\Xi\in supp[\pi_N(t)]}}\operatorname{Lip}{(\gamma(x_i-x_l,v_i-v_l)}^2,\sup_{\substack{i,l=1,\dots,N\\
X,V,Y,\Xi\in supp[\pi_N(t)]}}\operatorname{Lip}{(\gamma(y_i-y_l,\xi_i-\xi_l)}^2)}
\nonumber\\
&&\times D_N(t).\nonumber
\end{eqnarray}
And the lemma follows by \eqref{trucenplume}.
\end{proof}
\vskip 1cm
It remains to estimate in \eqref{trucenplume} the term
\begin{eqnarray}
&&\frac{1}N\int
\sum\limits_{i=1}^N
|G_i(X,V)-G_{\rho^t}(x_i,v_i)|^2
\pi^t_N(dX,dV,dY,d\Xi)\nonumber\\
&=&
\frac{1}N\int
\sum\limits_{i=1}^N
|G_i(X,V)-G_{\rho^t}(x_i,v_i)|^2
(\rho^t)^{\otimes N}dXdV.\nonumber
\end{eqnarray}
The following result is a variant of Lemma 3.3 in \cite{gmp} with the special value $p=2$ and $d$ replaced by $2d$.
\begin{Lem}\label{lemthree}
Let
$\rho$ be a compactly supported probability density on
$\mathbf{R}^{2d}$
and let $F$ be a locally bounded vector field on $\mathbf{R}^{2d}$.
For each $j=1,\ldots,N$, one has
\begin{eqnarray}\label{eqlemthree}
&&\int\left|F\star\rho(x_j,v_j)-\frac1N\sum_{k=1}^NF(x_j-x_k,v_j-v_k)\right|^{2}\rho^{\otimes N}dXdV\\
&&\le
\frac{4}{N}\sup_{\substack{
(x,v),(x',v')\in supp[\rho]}}|F(x-x',v-v')|^2.\nonumber
\end{eqnarray}
\end{Lem}
\begin{proof}
Let us denote, for $x_j,x,v_j,v\in\mathbf{R}^d, j=1,\dots,N$,
\begin{equation}\label{defnu}
\nu_{x_j,v_j}(x,v):=
F\star\rho(x_j,v_j)-F(x_j-x,v_j-v)
\end{equation}
(note that $F\star\rho(x_j,v_j)$ doesn't depend on $(x,v)$).
One has (let us remind the notation $X:=(x_1,\dots,x_N),\ V:=(v_1,\dots,v_N)$)
\begin{eqnarray}
&&\int\left|F\star\rho(x_j,v_j)-\frac1N\sum_{k=1}^NF(x_j-x_k,v_j-v_k)\right|^{2}\rho^{\otimes N}dXdV\nonumber\\
&=&
\int\left|\frac1N\sum_{k=1}^N\nu_{x_j,v_j}(x_k,v_k)\right|^{2}\rho^{\otimes N}dXdV\nonumber\\
&=&
\frac1{N^2}
\sum_{k,l=1,\dots N}\int \nu_{x_j,v_j}(x_k,v_k)\nu_{x_j,v_j}(x_l,v_l)\rho^{\otimes N}dXdV\nonumber\\
&=&
\frac1{N^2}
\sum_{k\neq l\neq j\neq k}\int\left(\int \nu_{x_j,v_j}(x_k,v_k)\rho(x_k,v_k)dx_kdv_k\right)\nu_{x_j,v_j}(x_l,v_l)\rho(x_l,v_l)\rho(x_j,v_j)dx_ldv_ldx_jdv_j\nonumber\\
&+&
\frac1{N^2}
\sum_{k\neq l= j}\int\left(\int \nu_{x_j,v_j}(x_k,v_k)\rho(x_k,v_k)dx_kdv_k\right)\nu_{x_j,v_j}(x_j,v_j)\rho(x_j,v_j)dx_jdv_j\nonumber\\
&+&
\frac1{N^2}\sum_{k\neq j}
\int \nu_{x_j,v_j}(x_k,v_k)^2\rho(x_k,v_k)dx_kdv_k\rho(x_j,v_j)dx_jdv_j\nonumber\\
&+&
\frac1{N^2}
\int \nu_{x_j,v_j}(x_j,v_j)^2\rho(x_j,v_j)dx_jdv_j\nonumber\\
&\leq &\frac{N+1}{N^2}\sup_{(x,v),(x',v')\in supp[\rho]}\nu_{x',v'}(x,v)^2\nonumber
\end{eqnarray}
since, by \eqref{defnu},
$
\int\nu_{x',v'}(x,v)\rho(x,v)dxdv=0.
$ for all $x',v'\in \mathbf{R}^d$, and $\int \rho(x,v)dxdv=1$.
\vskip 0.5cm
By \eqref{defnu} again,
$$|\nu_{x',v'}(x,v)|\leq 2\sup\limits_{\substack{
(x,v)\in supp[\rho]}}|F(x'-x,v'-v)|
\leq
2\sup\limits_{\substack{
(x,v),(x',v')\in supp[\rho]}}|F(x'-x,v'-v)|
$$ and the lemma is proved.
\end{proof}
Lemma \ref{lemthree} with $F(x,v)=\gamma(x,v)$ together with Lemma \ref{lemtwo} gives immediately that
$$
\begin{aligned}
\frac{dD_N}{dt}\le L (t) D_N+\frac4N
\sup_{(x,v),(x',v')\in supp[\rho^t]}|\gamma(x-x',v-v')|^2
\end{aligned}
$$
and, by Gronwall's inequality,
\begin{equation}\label{IneqD}
D_N(t)\le
\frac4N \int_0^
\sup_{(x,v),(x',v')\in supp[\rho^t]}|\gamma(x-x',v-v')|^2
e^{\int_s^tL (u)du}ds\,
\end{equation}
since $D_N(0)=0$ by \eqref{inop}.
Let us denote by $(\pi_N(t))_1$ the measure on $\mathbf{R}^{2d}\times \mathbf{R}^{2d}$ defined, for every test function $\varphi(x_1,v_1;y_1,\xi_1)$, by
$$
\int _{\mathbf{R}^{2dN}\times\mathbf{R}^{2dN}}\varphi\pi_N(dX,dV;dYd\Xi)
=
\int_{\mathbf{R}^{2d}\times\mathbf{R}^{2d}}\varphi(x_1,v_1;y_1,\xi_1)(\pi_N(t))_1(dx_1,dv_1;dy_1,d\xi_1).
$$
We now notice the following
straightforward fact.
\begin{Lem}\label{lemfour}
$(\pi_N(t))_1$ is a coupling between $\rho^t_{N;1}$ and $\rho^t$.
\end{Lem}
Let us note that $\pi_N^{in}$ is obviously symmetric by permutation of the phase-space variables, that is
\begin{equation}\label{SympiN0}
T_{\sigma}\#\pi_N^{in}=\pi_N^{in}\,,\qquad\hbox{ for each }{\sigma}\in\mathfrak{S}_N\,,
\end{equation}
where $\mathfrak{S}_N$ is the group of permutations of $N$ elements and
$$
\begin{aligned}
T_{\sigma}(x_1,v_1,\ldots,x_N,v_N,y_1,\xi_1,\ldots,y_N,\xi_N)&
\\
=(x_{{\sigma}(1)},v_{{\sigma}(1)},\ldots,x_{{\sigma}(N)},v_{{\sigma}(N)},y_{{\sigma}(1)},\xi_{{\sigma}(1)},\ldots,y_{{\sigma}(N)},\xi_{{\sigma}(N)})&\,.
\end{aligned}
$$
Therefore, $\pi_N(t)$ is also symmetric by permutations for all $t\in\mathbf{R}$, as being the unique solution to the equation \eqref{lintrans}, being itself, by construction, symmetric by permutations, .
Consequently, one has easily that
\begin{equation}\label{MinDNp}
D_N(t)=\int(|x_1-y_1|^2+|v_1-\xi_1|^2)d(\pi_N(t))_1
\end{equation}
and Lemma \ref{lemfour} immediately implies that
\begin{equation}
D_N(t)\ge W_2(\rho^t_{N;1},\rho^t)^2\,,
\end{equation}
so that, by \eqref{IneqD},
\begin{equation}\label{oufff}
W_2(\rho^t_{N;1},\rho^t)^2
\leq
\frac4N \int_0^
\sup_{(x,v),(x',v')\in supp[\rho^t]}|\gamma(x-x',v-v')|^2
e^{\int_s^tL (u)du}ds.
\end{equation}
Finally, recalling that
$$
\pi^{in}_N(Y,\Xi,X,V):=(\rho^{in})^{\otimes N}(X,V)\delta(X-Y)\delta(V-\Xi),
$$
one gets immediately that
\begin{eqnarray}
(Y,\Xi;X,V)\in supp[\pi_N(t)]&\Rightarrow& (Y,\Xi)\in supp[\rho^t_N]\mbox{ and }\nonumber\\
&\Rightarrow& (X,V)\in supp[(\rho^t)^{\otimes _N}]
\Leftrightarrow (x_i,v_i)\in supp[\rho^t],\ i=1,\dots,N.\nonumber
\end{eqnarray}
Therefore, after \eqref{defl}
\begin{eqnarray}
{\bf L}(t)\leq L(t):=&&
\label{defldef}\\
2(1+2
\min{(\sup_{\substack{i,l=1,\dots,N\\
(Y,\Xi)\in supp[\rho^t_N]}}\operatorname{Lip}{(\gamma)}^2_{(y_i-y_l,\xi_i-\xi_l)},\sup_{\substack{
(x,v),(x',v')\in supp[\rho^t]}}\operatorname{Lip}{(\gamma)}^2_{(x-x',v-v')})})&&
).\nonumber
\end{eqnarray}
Theorem \ref{mainsublin} is proven.
\begin{Rmk}[Higher order Wasserstein and marginals]\label{highorders}
As we wanted to leap this note as short as possible, we expressed our results only for the first marginal $\rho_{N;1}$ in the $2$-Wasserstein topology, but the method developed in \cite{gmp} allows as well, with the same kind of modification than the ones used before in this section, to the higher cases.
Let us remind, for $p\geq 1$, the definition of $W_p(\mu,\nu)$ for two positive measures $\mu,\nu$ (cf. Definition \ref{defwas}
$$
W_p(\mu,\nu)^p=\inf_{\gamma\in\Gamma(\mu,\nu)}\int_{\mathbf{R}^m\times\mathbf{R}^m}
|x-y|^p\gamma(dx,dy),
$$
and, for any probability measure $\rho$ on $\mathbf{R}^{2dN}$ and $n=1,\dots,N$, the definition of the $n$th marginal $\rho_{N;n}$ defined on $\mathbf{R}^{2d{n}}$
\begin{eqnarray}
&&\rho_{N;n}(x_1,\dots,x_{n},v_1,\dots,v_{n}):=\nonumber\\
&&
\int_{\mathbf{R}^{2d(N-n)}}\rho((x_1,\dots,x_{n}, x_{n+1},\dots,x_N,v_1,\dots,v_{n},v_{n+1},\dots,v_N)dx_{n+1}\dots dx_Ndv_{n+1}\dots dv_N.\nonumber
\end{eqnarray}
One gets, for each $p\geq 1$, $N\geq 1$ and $n=1,\dots,N$,
$$
\frac1nW_p(\rho^t_{N;n},(\rho^t)^{\otimes n})^p\leq D_{p,n}(t)N^{-\min{(p/2,1)}}
$$
with
$$
D_{p,n}(t)= 2^{2p}\max{(1,p-1)}([p/2]+1)\int_0^t
\sup_{\substack{k,l=1,\dots,N\\(Y,\Xi)\in supp[\rho^t_N]}}|\gamma(y_k-y_l,\xi_k-\xi_l)|^
pe^{
2\max{(1,p-1)}\int_s^t
L(u)
)du}ds
$$
where
$$
L(u)=
1+2^{p-1}
\sup_{\substack{(x,v),(y,\xi)\in supp[\rho^t]}}\operatorname{Lip}(\gamma)_{(x-y,v-\xi)}^
p.
$$
The main changes in the proof are the use of the Young inequality
$$
puv^{p-1}\leq u^p+(p-1)v^p\leq\max(1,p-1)(u^p+v^p)
$$
instead of $2uv\leq u^2+v^2$ before \eqref{trucenplume}, the convexity of $|\cdot|^p$ for $p\geq 1$ and a variant of Lemma \ref{lemthree}, similar to Lemma 3.3 in \cite{gmp}, which reads
\begin{eqnarray}
&&\int\left|F\star\rho(x_j,v_j)-\frac1N\sum_{k=1}^NF(x_j-x_k,v_i-v_j)\right|^{p}
\rho^{\otimes N}dXdV\nonumber\\
&\leq&
\frac{2[p/2]+2}{N^{\min(p/2,1)}}
\sup_{\substack{k,l1,\dots,N\\(Y,\Xi)\in supp[\rho^t_N]}}|F(y_k-y_l,\xi_k-\xi_l)|^{p}.\nonumber
\end{eqnarray}
Finally, the statement of Lemma \ref{lemfour} becomes now easily that
$\pi_N(t)_n$ is a coupling between $\rho^t_{N;n}$ and $(\rho^t)^{\otimes n}$, where $\pi_N(t)_n$ is defined through by
\begin{eqnarray}
&&\int _{\mathbf{R}^{2dN}\times\mathbf{R}^{2dN}}\varphi\pi_N(dX,dV;dYd\Xi)
=
\int_{\mathbf{R}^{2d}\times\mathbf{R}^{2d}}\varphi
\pi_N(t)_n(d(x,v;y,\xi)_1\dots d(x,v;y,\xi)_n)\nonumber
\end{eqnarray}
for every test function $\varphi((x,v;y,\xi)_1,\dots,(x,v;y,\xi)_n)$,
\end{Rmk}
\subsection{Proof of Corollary \ref{maingensublinest}}\label{proofcor2}
We get to estimates $L(t)$ and $C(t)$
as given in Theorem \ref{mainsublin} out of the estimates established in Appendices \ref{dynestpar} and \ref{dynestpkin}.
Let us recall that
\begin{eqnarray}\nonumber
L(t)&:=&
2(1+2
\min{(\sup_{\substack{i,l=1,\dots,N\\
(Y,\Xi)\in supp[\rho^t_N]}}\operatorname{Lip}{(\gamma)}^2_{(y_i-y_l,\xi_i-\xi_l)},\sup_{\substack{
(x,v),(x',v')\in supp[\rho^t]}}\operatorname{Lip}{(\gamma)}^2_{(x-x',v-v')})})
).\nonumber\\
&\leq&
2(1+2
\sup_{\substack{
(x,v),(x',v')\in supp[\rho^t]}}\operatorname{Lip}{(\gamma)}^2_{(x-x',v-v')}
).\nonumber\\&\leq&
2(1+2\gamma_0^2
\sup_{\substack{
(x,v),(x',v')\in supp[\rho^t]}}|v-v'|^2)\nonumber\\
&\leq &
2(1+8\gamma_0^2\sup_{\substack{
(x,v)\in supp[\rho^t]}}|v|^2)\nonumber\\
&\leq &
2(1+8\gamma_0^2\sup_{\substack{
(x,v)\in supp[\rho^{in}]}}|\Phi_v(t)(x,v)|^2).\nonumber
\end{eqnarray}
Thanks to \eqref{estiphi}, we get
\begin{eqnarray}
L(t)
&\leq&
2(1+16\gamma_0^2e^{4\gamma_0t}(\|v\|_{L^\infty(supp[\rho^{in}])}+
\|v\|_{L^1(supp[\rho^{in}])})^2)\nonumber
\end{eqnarray}
By the same argument, we get
\begin{eqnarray}
C(t)^2&:=&
4
\int_0^t
\sup_{(x,v),(x',v')\in supp[\rho^t]}|\gamma(x-x',v-v')|^2
e^{\int_s^tL (u)du}ds
,\nonumber\\
&\leq&
32\gamma_0^2
(\|v\|_{L^\infty(supp[\rho^{in}])}+
\|v\|_{L^1(supp[\rho^{in}])})^2\nonumber\\
&&
\times \int_0^t e^{4(\gamma_0s+t-s)}e^{16\gamma_0^2(\|v\|_{L^\infty(supp[\rho^{in}])}+
\|v\|_{L^1(supp[\rho^{in}])})^2(e^{4\gamma_0t}-e^{4\gamma_0s})/4\gamma_0}ds
\nonumber\\
&\leq &
8\gamma_0^2(\|v\|_{L^\infty(supp[\rho^{in}])}+
\|v\|_{L^1(supp[\rho^{in}])})^2e^{e^{4\gamma_0t}4\gamma_0
(\|v\|_{L^\infty(supp[\rho^{in}])}+
\|v\|_{L^1(supp[\rho^{in}])})^2}e^{4t}(e^{4\gamma_0t}-1).\nonumber
\end{eqnarray}
Corollary \ref{maingensublinest} is proven.
\subsection{Proof of Theorem \ref{main}}\label{Proofmain}
In order to prove Theorem \ref{main}, we need to estimate, in the Cucker-Smale particular case, that is when
$$
\gamma(x,v)=\psi(x)v,
$$
the two quantities
\begin{eqnarray}\nonumber
C(t)&:=&
\left(
4
\int_0^t
\sup_{(x,v),(x',v')\in supp[\rho^t]}|\gamma(x-x',v-v')|^2
e^{\int_s^tL (u)du}ds
\right)^{\frac12}\,
\nonumber\\
&\leq&
2\|\psi\|_\infty\left(
\int_0^t
\sup_{(x,v),(x',v')\in supp[\rho^t]}|v-v'|^2
e^{\int_s^tL (u)du}ds
\right)^{\frac12}.\,
\nonumber\\
L(t)&:=&
2(1+2
\min{(\sup_{\substack{i,l=1,\dots,N\\
(Y,\Xi)\in supp[\rho^t_N]}}\operatorname{Lip}{(\gamma)}^2_{(y_i-y_l,\xi_i-\xi_l)},\sup_{\substack{
(x,v),(x',v')\in supp[\rho^t]}}\operatorname{Lip}{(\gamma)}^2_{(x-x',v-v')})})
)\nonumber\\
&\leq&
2(1+2\sup_{\substack{i,l=1,\dots,N\\
(Y,\Xi)\in supp[\rho^t_N]}}\operatorname{Lip}{(\gamma)}^2_{(y_i-y_l,\xi_i-\xi_l)})\nonumber\\
&\leq &
2(1+2\|\psi\|^2_\infty\sup_{\substack{i,l=1,\dots,N\\
(Y,\Xi)\in supp[\rho^t_N]}}
|\xi_i-\xi_l|^2).\nonumber
\end{eqnarray}
We will need just a very little part of the stability results expressed by Ha, Kim and Zhang in \cite{hakimzhang}, namely the following inequality:
$$
\frac d{dt}\sup
_{\substack{i,l=1,\dots,N\\(X,V)\in supp[\rho_N^{in}]}}|v_k(t)-v_l(t)|\leq 0,\ \forall t.
$$
Indeed, formula $(8)$ in Lemma 2.2 in \cite{hakimzhang} stipulates that $\tfrac d{dt}\mathcal D_V(t)\leq 0$, where $\mathcal D_V(t)$, as defined in Corollary $1$ of \cite{hakimzhang}, is precisely $ \sup\limits
_{\substack{i,l=1,\dots,N\\(X,V)\in supp[\rho_N^{in}]}}|v_k(t)-v_l(t)|$ in the case $p=2$ of the definition (4) of $\mathcal D_V(0):=\mathcal D_V$ in \cite{hakimzhang}.
This inequality leads naturally to
$$
\sup_{\substack{i,l=1,\dots,N\\(X,V)\in supp[\rho_N^t]}}|v_i-v_l|^2
\leq 4\|v\|_{L^\infty(supp[\rho^{in}])}^2,
$$
which implies
$$
L(t)\leq 2(1+8\|\psi\|^2_\infty\|v\|_{L^\infty(supp[\rho^{in}])}^2):=L.
$$
We will estimate $\sup\limits_{\substack{(x,v),(y,\xi)\in supp[\rho^t]}}|v-\xi|^2 \leq 2\sup\limits_{\substack{(x,v),(y,\xi)\in supp[\rho^t]}}(|v|^2+|\xi|^2)$ thanks to Lemma 3.2 in \cite{prt} which stipulates that
$$
supp_v[\rho^t]\subset B(\bar v, V(t))
$$
with $\bar v=\int v\rho^{in}dxdv$ and $\frac d{dt}V(t)\leq 0$.
Therefore $\|v\|_{L^\infty(supp[\rho^t])}\leq |\bar v|+V(0)$ so that, since one can take $V(0)=|\bar v|+|supp[\rho^{in}]|$,
$$
\sup\limits_{\substack{(x,v),(x'v')\in supp[\rho^t]}}|v-v'|^2
\leq
2\|\psi\|^2_\infty|(2|\bar v|+|supp[\rho^{in}]|)^2.
$$ and
\begin{eqnarray}
C(t)
&\leq&
2\|\psi\|^2_\infty\left(
2\int_0^t
(2|\bar v|+|supp[\rho^{in}]|)^2
e^{L(t-s)}ds
\right)^{\frac12}\nonumber\\
&=&
2\|\psi\|^2_\infty\left(
2
(2|\bar v|+|supp[\rho^{in}]|)^2
\frac{e^{Lt}-1}L
\right)^{\frac12}.\nonumber
\end{eqnarray}
Theorem \ref{main} is proved.
\begin{appendix}
\section{Dynamical estimates for general Cucker-Smale particle systems}\label{dynestpar}
In this section, we give global estimates on the flow $\Phi^N(t)$ generated by \eqref{genCS}.
We have, for each $i=1,\dots,N$,
\begin{eqnarray}
\frac d{dt}|v_i|&\leq&
|\dot v_i|\nonumber\\
&\leq&
\frac1N \sum_{j=1}^N
|\gamma(x_i-v_j,v_i-v_j)|\leq
\frac1N \sum_{j=1}^N
\gamma_0|v_i-v_j|\nonumber\\
&\leq&
\frac1N \sum_{j=1}^N\gamma_0(|v_j|+|v_i|),\label{crossfingers}
\end{eqnarray}
so that
$$
\frac d{dt}\sum_{i=1}^N|v_i|
\leq2\gamma_0 \sum_{i=1}^N|v_i|
$$
and, by Gronwall inequality,
\begin{equation}\label{crucial}
\sum_{i=1}^N|v_i(t)|\leq \sum_{i=1}^N|v_i(0)|e^{2\gamma_0t}.
\end{equation}
Turning back to \eqref{crossfingers}, we get by \eqref{crucial}, for each $i=1,\dots,N$,
\begin{eqnarray}
\frac d{dt}(|v_i|)&\leq& \gamma_0|v_i|+\frac{\gamma_0}N\sum_{j=1}^N|v_j|\leq\gamma_0|v_i|+\gamma_0e^{2\gamma_0t}\frac1N\sum_{j=1}^N|v_j(0)|\nonumber\\
&\leq&
\gamma_0|v_i|+\gamma_0e^{2\gamma_0t}\max_{j=1,\dots,N}|v_j(0)|.\nonumber
\end{eqnarray}
Therefore, again by Gronwall Lemma and uniformly in $N$,
\begin{eqnarray}\label{estpart}
|v_i(t)|&\leq& e^{\gamma_0t}\big(
|v_i(0)+\gamma_0\max_{j=1,\dots,N}|v_j(0)|\frac{e^{\gamma_0t}-1}{\gamma_0}\big)\nonumber\\
&\leq& \max_{j=1,\dots,N}|v_j(0)|e^{2\gamma_0t},\ \ \ i=1,\dots,N.
\end{eqnarray}
Finally, \eqref{genCS} gives immediatly that
$$
||x_1(t)|-|x_i(0)||
\leq
\max_{j=1,\dots,N}|v_j(0)|\frac{e^{2\gamma_0t}-1}{2\gamma_0},\ \ \ i=1,\dots,N.
$$
\section{Dynamical estimates for general Cucker-Smale kinetic systems}\label{dynestpkin}
Let us recall that, according to Theorem 2.3 in \cite{prt}, there exists a diffeomorphism $\Phi(t)$ on $\mathbf{R}^{2d}$ such that the solution $\rho^t$ of \eqref{vlasovgencs} is given by
$$
\rho(t)=\Phi(t)\#\rho^{in}.
$$
Moreover, $\Phi(t)$ solves the system
$$
\dot\Phi(t)
:=
\binom{\dot\Phi_x(t)}{\dot\Phi_v(t)}=\binom{\Phi_v(t)}{\big(\gamma *\rho^t\big)(\Phi(t)):=\int \gamma(\Phi_x(t)-y,\Phi_v(t)-\xi)\rho^t(y,\xi)dxd\xi}.
$$
Therefore,
\begin{equation}\label{avantlemme}\frac d{dt} |\Phi_v(t)|
\leq |\dot\Phi_v(t)|
\leq
\gamma_0(\|\rho^t
\|_1|\Phi_v(t)|+\int|\xi|\rho^t(y,\xi)dyd\xi).
\end{equation}
Obviously $\|\rho^t
\|_{L^1}=\|\rho^{in}
\|_{L^1}=1$.
\begin{Lem}\label{crucialvlasov}
$$
\int|
v|
\rho^tdxdv\leq e^{2\gamma_0t}\int|
v|
\rho^{in}dxdv.
$$
\end{Lem}
\begin{proof}
\begin{eqnarray}
\frac d{dt}\int |v|\rho^tdxdv
&
\leq&|
\int |v|
\dot \rho^tdxdv|\nonumber\\
&\leq&
|\int |v|
\nabla_v(\int\gamma(x-y.v-\xi)\rho^t(y,\xi)\rho^t(x,v)dxdvdyd\xi|\nonumber\\
&\leq&
\int|\frac v{|v|} \gamma_0((v|+|\xi|)\rho^t(y,\xi)\rho^t(x,v)dxdvdyd\x
\leq
2\gamma_0\int|
v|
\rho^tdxdv.\nonumber
\end{eqnarray}
The result follows by Gronwall inequality.
\end{proof}
Thanks to Lemma \ref{crucialvlasov}, \eqref{avantlemme} becomes
$$
\frac d{dt} |\Phi_v(t)|
\leq
\gamma_0|\Phi_v(t)|+\gamma_0e^{2\gamma_0t}\int|v|\rho^{in}dxdv,
$$
and, by Gronwall Lemma, we have
\begin{eqnarray}\label{estiphi}
|\Phi_v(t)(x,v)|
&\leq&
e^{\gamma_0t}(|v|
+
(e^{\gamma_0t}-1)\int|v|\rho^{in}dxdv)\nonumber\\
&\leq &
e^{2\gamma_0t}(\|v\|_{L^\infty(supp[\rho^{in}])}+
\|v\|_{L^1(supp[\rho^{in}])})
.
\end{eqnarray}
\end{appendix}
\vskip 1cm
\textbf{Acknowledgements.} The work of Thierry Paul was partly supported by LIA LYSM (co-funded by AMU, CNRS, ECM and INdAM) and IAC (Istituto per le Applicazioni del Calcolo "Mauro Picone") from CNR.
\vskip 1cm
|
\section*{\refname}}
\def\section*{\refname}{\section*{\refname}}
\DeclareMathOperator{\Tr}{Tr}
\begin{document}
\title{Glueballs in the Klebanov-Strassler Theory: Pseudoscalars vs Scalars}
\author{Cornélio Rodrigues Filho}\email{<EMAIL>}
\affiliation{Departamento de F\'\i sica Teorica e Experimental, Universidade Federal do Rio Grande do Norte, Campus Universit\'ario, Lagoa Nova, Natal-RN 59078-970, Brazil}
\begin{abstract}
\begin{center}
\textbf{Abstract}
\end{center}
We discuss the $0^{+-}$ singlet sector of glueballs in the Klebanov-Strassler theory. We report the results of a numerical study of the linearized equations in the Klebanov-Strasller background and make a comparison with the spectrum of the scalar sector. While for four towers of the total six towers of massive pseudoscalar states our results match the spectrum of the corresponding towers of scalars, the values for the remaining two towers diverge with those of the scalars. We discuss possible interpretations of this divergence.
\end{abstract}
\date{\today}
\keywords{Gauge/Gravity Duality, Klebanov-Strassler Theory, Singlet Glueballs, Pseudoscalars vs Scalars.}
\maketitle
\section{Introduction}
The example of the gauge/gravity duality \cite{Maldacena:1997re,Gubser:1998bc,Witten:1998qj} proposed by Klebanov and Strassler (KS) \cite{Klebanov:2000hb} consists of the dual description of the $\mathcal{N}=1$ supersymmetric non-conformal gauge theory by a solution of the type IIB supergravity equations. For reviews of this duality see \cite{Herzog:2001xk,Herzog:2002ih}. Some important features of the KS theory include the chiral symmetry breaking and an unconventional renormalization group (RG) flow \cite{Strassler:2005qs}, which among other things leads to color confinement. The confinement in this model is observed as follows: in the KS theory the $SU(M+N)\times SU(N)$ gauge group undergoes a cascading flow to a strongly coupled $\mathcal{N}=1$ supersymmetric $SU(M)$ gauge theory, which exhibits confinement. In the infrared regime one can separate the pure $\mathcal{N}=1$ Supersymmetric Yang-Mills sector within the full theory and study its low energy states (glueballs) using classical supergravity approximation. The holographic methods have a big advantage over standard field theory approach, since the latter does not have analytical access to the information about glueballs.
In the last years a series of works \cite{Caceres:2000qe,Berg:2005pd,Berg:2006xy,Dymarsky:2006hn,Benna:2006ib,Dymarsky:2007zs,Benna:2007mb,Dymarsky:2008wd,Gordeli:2009nw,Gordeli:2013jea,Melnikov:2020cxe} have used the KS theory to estimate the behavior of masses of the glueballs for pure $\mathcal{N}=1$ theory. In these works the glueball states that are singlets under the global $SU(2)\times SU(2)$ symmetry of the KS theory \cite{Klebanov:2000hb} were considered. The holographic techniques replace the computation of the two-point correlation functions in the field theory by the analysis of the linearized equations of classical supergravity \cite{Csaki:1998qr,Brower:2000rp}. For the analysis beyond the singlet regime see \cite{Elander:2009bm,Elander:2010wd,Elander:2012yh,Elander:2014ola,Elander:2017cle,Elander:2017hyr,Elander:2018aub}.
As an alternative to holography, the spectrum of the glueballs can be estimated from lattice calculations \cite{Morningstar:1999rf,Teper:1998kw,Chen:2005mg,Lucini:2004my,Lucini:2010nv,Holligan:2019lma,Athenodorou:2020ani,Gregory:2012hu}. The predictions of the two methods can be compared.
The main goal of this paper is the discussion of the spectrum of the $0^{-+}$ (pseudoscalar) glueballs. The linearized equations and preliminary results of the spectrum calculation were recently presented in \cite{Melnikov:2020cxe}. Here we provide some details of the numerical analysis and suggest an alternative interpretation of the results to that presented in \cite{Melnikov:2020cxe}. In particular, we discuss the possibility that the results on the pseudoscalar glueballs indicate that the spectrum of the scalars known from \cite{Berg:2005pd,Berg:2006xy} must be corrected.
This letter is organized as follows. In section \ref{glue} we present a short discussion of the derivation of the pseudoscalar glueballs. We also discuss the consistency checks of the results. In section \ref{hvsl} we apply the numerical approach to estimate the spectrum of this sector and make a comparison with the spectrum of the scalar glueballs. In section \ref{conclusion} we present our conclusions and some final remarks.
\section{Pseudoscalar Glueballs}
\label{glue}
The Klebanov-Strassler theory \cite{Klebanov:2000hb} provides a possible setup to analyze the spectrum of glueballs, which are expected to constitute the low energy spectrum of $\mathcal{N}=1$ Yang-Mills Theory. The glueballs are bound states mainly composed of gluons. Their classification is given in terms of the $J^{PC}$ quantum numbers, where $J$ is the spin, $P$ is the parity and $C$ is the charge conjugation quantum numbers. The action of these discrete symmetries in the KS theory was discussed in some detail in \cite{Melnikov:2020cxe}. Although, classical supergravity approximation can not be used to capture the glueball states with high spin, the structure of the low spin spectrum already contains some interesting information.
For our purposes we are interested in the masses of glueballs, which are singlets under the global $SU(2)\times SU(2)$ symmetry \cite{Berg:2005pd,Berg:2006xy,Dymarsky:2006hn,Benna:2006ib,Dymarsky:2007zs,Benna:2007mb,Dymarsky:2008wd,Gordeli:2009nw,Gordeli:2013jea,Melnikov:2020cxe} of the KS theory. It is these states that we expect to match the low energy limit of the pure $\mathcal{N}=1$ SYM. Specifically, we are interested in the $0^{-+}$ glueballs that were recently described in \cite{Melnikov:2020cxe}. In this section we review the derivation of the equations for the pseudoscalars.
The KS background is a type IIB supergravity solution with a warped deformed conifold metric \cite{Candelas:1989js} of $AdS^5\times T^{11}$ topology. The metric is supported by fluxes of the $F_5$, $F_3$ and $H_3$ forms ($F_5$ is a self-dual R-R 5-form, $F_3$ is the R-R 3-form and $H_3$ is the NS-NS 3-from from the bosonic sector of the type IIB supergravity). The solution also assumes that $C=\Phi=0$, which are the dilaton and R-R axial scalar, respectively.
For reviews of the gravity solution of KS theory see \cite{Herzog:2001xk,Herzog:2002ih}.
The pseudoscalar sector can be described by the following fluctuations of the type IIB supergravity background \cite{Klebanov:2000hb}:
\begin{widetext}
\begin{eqnarray}
\delta (ds^2_{T^{1,1}})&=&B(g^1\cdot g^4-g^2\cdot g^3),\label{dg13}\\
\delta (ds^2_5) &=& (*_4{d a}+A d\tau ) \wedge g^5,\label{daA}\\
\delta C&=&C, \label{C}\\
\delta C_{2}&=&C_{2}^{-}(g^{1}\wedge g^{2}-g^{3}\wedge
g^{4})+\\ \nonumber &&+C_{2}^{+}(g^{1}\wedge g^{2}+g^{3}\wedge g^{4}),\label{cpcm}\\
\delta B_{2}&=&B_{2} \left( g^{1}\wedge g^{3}+g^{2}\wedge
g^{4}\right)\label{b2},\\
\nonumber
\delta F_{5} & =& \frac{lG^{55}}{2}\Bigg\{\Big[\partial_{\mu}(a +\phi_{1})dx^{\mu}+( A+\phi _{2}) d\tau\Big] \wedge g^{1}\wedge g^{2}\wedge g^{3}\wedge
g^{4} -
\sqrt{-G}( G^{11}G^{33})^{2}\Big[h^{1/2} \ast_4d( a -\phi _{1}) \wedge d\tau +\\
&&+
(A -\phi _{2})
G^{55}d^{4}x\wedge g^{5}
\Big] - h^{1/2}\sqrt{-G}
G^{11}G^{33}( G^{55})^{2}*_4d\phi
_{3} \wedge dg^{5}+\partial_{\mu}\phi _{3}dx^{\mu}\wedge d\tau
\wedge g^{5}\wedge dg^{5}\Bigg\}.
\label{df5}
\end{eqnarray}
\end{widetext}
Here we consider fluctuations of the metric, R-R axial scalar, R-R 2-form potential, NS-NS 2-form potential and R-R 5-form potential, introducing a set of ten unknown functions of the coordinates $x^\mu\equiv\vec{x}=(t,x,y,z)$ and $\tau$ ($AdS_5$ radial variable),
\begin{equation}
a\,,A\,,B\,,B_2\,,C_2^-\,,C_2^+\,,\phi_1\,,\phi_2\,,\phi_3.
\end{equation}
Note that \eqref{df5} is constructed in such a way to guarantee the self-duality condition, $ F_5=*F_5$. Functions $\phi_2$ and $\phi_3$ can be eliminated through the Bianchi identity for $F_5$.
Perturbations \eqref{dg13}-\eqref{df5} produce a coupled set of eight ordinary differential equations \eqref{eqf3}-\eqref{eqA} (seven of which are of second order and one is of first order). Function $A$ enters the equations without second derivatives. In total there are eight unknown modes for eight ODE's.
The system of the linearized equations looks rather complicated, so a number of consistency checks were carried out to ensure their validity. The first check is the gauge invariance of the system. (See the action of the gauge transformations in equation \eqref{gauge} below.)
Another non-trivial check is imposed by supersymmetry arguments. As predicted in \cite{Gordeli:2009nw,Gordeli:2013jea}, the pseudoscalar sector has to be characterized by six physical modes. Indeed, this is confirmed by the equations. As mentioned in \cite{Melnikov:2020cxe}, not all equations are linear independent so that it is possible to reduce the full system to a set of six equations and modes via the following arguments. First, by gauge freedom we can set $a=0$ and, as a consequence of the linear dependence, one can drop the second order ODE related to $a''$, \eqref{eqa}. Second, we use the first order ODE, \eqref{eqA}, to write the mode $A$ in terms of the remaining ones. Finally, we substitute the algebraic expression for $A$ in the remaining six equations and find the expected result.
However, this particular system has shown to be problematic in the numerical analysis. A more appropriate choice is to set $A=0$ and work with seven second order ODE's \eqref{eqf3}-\eqref{eqa}. First order ODE \eqref{eqA} imposes a set of constraints which the remaining modes must obey. These constraints project onto the six physical modes.
In table \ref{tab:dims} we show the physical modes together with their conformal scaling dimensions obtained from the analysis of the asymptotic expansion at $\tau\to\infty$. The fact that the dimensions are integer and match the expectations from the supersymmetry analysis (see \cite{Melnikov:2020cxe}) provides another check of the consistency of the system of equations.
\begin{table}[htb]
\begin{center}
\begin{tabular}{l|cccccc}
\hline
Mode & $\phi_3$ & $P_2$ & $Q_2$ & $C_2^+$ & $C$ & $B$ \\ \hline
Dimension, $\Delta$ & $5$ & $3$ & $7$ & $4$ & $4$ & $3$ \\
\hline
\end{tabular}
\end{center}
\caption{Physical modes and the corresponding scaling dimensions of the dual operators of the pseudoscalar sector. Here $P_2=B_2-C_2^-$ and $Q_2=B_2+C_2^-$ .}
\label{tab:dims}
\end{table}
The last consistency check will be discussed in the next section, when we are going to use numerical calculations to estimate the spectrum of the pseudoscalar glueballs. In addition we will compare our results with those ones for the scalar sector \cite{Berg:2006xy}.
\section{Numerical Results}\label{hvsl}
Let us begin this section explaining the numerical approach employed to solve the system of equations described in the last section (also see equations \eqref{eqf3}-\eqref{eqA} in the appendix). A common method to solve the eigenvalue problem numerically is the shooting method, which involves replacing the boundary problem with the initial value one on either of the boundaries of the system $\tau\to 0$ (IR), or $\tau\to \infty$ (UV). However, this technique can not be applied directly to coupled systems of equations. Instead, one uses the generalization that is called Midpoint Determinant Method (MDM) \cite{Berg:2006xy}, which combines both the IR and UV initial problems. In the MDM one builds a $2n\times 2n$ quadratic matrix for $n$ fields from the UV and IR numerical solutions for the fields and their derivatives at some intermediate point $\tau_{mid}$, which is a point where the solutions are matched, i. e.,
\begin{equation}\label{deta}
\gamma=
\left(
\begin{array}{cc}
x_{IR} & x_{UV}\\
\partial x_{IR} & \partial x_{UV}
\end{array}\right)
_{\tau = \tau_{mid}}.
\end{equation}
Here $x_i$ is the set of fields, $i=1,\ldots,n$, and $\partial x_i$ is the set of derivatives. The eigenvalues are given by the condition $\det\gamma=0$, so that $\gamma$ is a function of $\tilde{m}^2$. To be precise as $\det\gamma$ is an oscillatory function of $\tilde{m}^2$ one finds the eigenvalues looking for the loci, where the determinant changes of sign.
To run the MDM it is necessary to specify five parameters, they are $\tau_{IR}$ and $\tau_{UV}$ the IR and UV cut-offs respectively, middle point $\tau_{mid}$, the step $\delta \tau$ and $n_{imposed}$. The latter as explained in \cite{Berg:2006xy} is used to control the behavior of the numerical solutions when some modes fluctuate dramatically faster than the other ones. After that, one constructs the analytic solutions in the IR and UV ends to be fed as initial conditions.
As we have mentioned the pseudoscalar sector is described by six physical modes and a convenient way to express this sector is through seven second ODE's where we have to add a nonphysical mode. Even in the presence of this mode we expect to separate the six physical families of eigenvalues.
We now move on to discuss the results obtained by application of the MDM procedure. In our analysis we calculate a $14\times 14$ matrix with the following parameters: $\tau_{IR}=0.1\,,\tau_{UV}=20\,,\tau_{mid}=1\,,\delta \tau=0.01\,,n_{imposed}=500$. We use the boundary conditions obtained in \cite{Melnikov:2020cxe}. In order to compare our results with those found in \cite{Berg:2006xy} we adopt their mass normalization, i.e, $m^2=0.9409\tilde{m}^2$.
Before we present the results of the numerical analysis we would like to mention that we also investigated the $12\times 12$ matrix scenario, with only six physical modes. In this case we computed the $12\times 12$ matrix dropping different modes to build it, but this approach does not provide consistent results. In particular, it does not capture the expected six towers of eigenvalues. For $14\times 14$ case,
the results offer a good consistency, independent from the variation of the parameters.
Finally, we provide the results found for the MDM analysis. In table \ref{tab:mid7} we show the lower values of $m^2$, in table \ref{tab:full} we summarize the first 65 values of $m^2$ and in addition we also include the results from the scalar sector computed in \cite{Berg:2006xy} to perform a comparison between the results of both sectors latter. The heavier part of table \ref{tab:full} is shown in figure \ref{fig:my_label}, our results are in red and the dots in vertical are the results of \cite{Berg:2006xy}.
\begin{table}[b]
\begin{center}
\begin{tabular}[c]{|l|l|l|l|l|l|l|l|l|}
\hline
\bf{$n$} & \bf{1} & \bf{2} & \bf{3} & \bf{4} & \bf{5} & \bf{6} \\\hline
$m^2$ & 0.273 & 0.513 & { 0.946} & { 1.38} & 1.67 & { 2.09} \\\hline
\bf{$n$} & \bf{7} & \bf{8} & \bf{9} &\bf{10} &\bf{11} &\bf{12} \\ \hline
${m}^2$ & 2.34 & { 2.73} & 3.33 & { 3.63} & 4.24 & 4.43 \\\hline
\bf{$n$} & \bf{13} & \bf{14} & \bf{15} & \bf{16} & \bf{17} & \bf{18} \\ \hline
$m^2$ & { 4.96} & 5.44 & 5.63 & { 6.25} & 6.63 & {6.96} \\ \hline
\end{tabular}
\captionof{table}{The lowest values of $m^2 $ found by MDM analysis.}\label{tab:mid7}
\end{center}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=0.47\textwidth]{heavyt.pdf}
\caption{The heavier values of $m^2$. Red line corresponds to our values for MDM analysis and the dashed lines are the values of \cite{Berg:2006xy}.}\label{fig:my_label}
\end{figure}
From figure \ref{fig:my_label} we can extract some interesting results. First, by supersymmetry arguments the $0^{-+}$ and $0^{++}$ sectors must share the same spectrum. In \cite{Berg:2006xy} seven families of eigenvalues were obtained, but one of these towers of eigenvalues was found in \cite{Gordeli:2009nw} to be a superpartner of a $1^{++}$ glueball. Then, as can be seen in figure \ref{fig:my_label} there is a periodicity among the eigenvalues so that as expected we are able to separate the spectrum of the pseudoscalar glueballs into six towers of eigenvalues in such a way that it gives us another consistency test. However, we also see that only four towers have good matches with the results of \cite{Berg:2006xy}. The remaining two towers show a very distinct behavior from the $0^{++}$ values. So, this apparent tension between the $0^{-+}$ and $0^{++}$ spectra brings up some questions.
The first question is about the validity of the linearized equations. We believe that their derivation is correct and we showed that they pass a number of consistency checks. It is more natural to expect that the problem is in the numerical analysis. In our case we also used the same MDM method as in \cite{Berg:2006xy} to compute the spectrum. The results of $0^{++}$ were independently checked in \cite{Elander:2014ola}. Therefore, in principle we are led to conclude that the $0^{++}$ spectrum is the correct one. On the other hand, we have not been able to improve our numerical analysis in such a controllable way that the convergence to the spectrum of~\cite{Berg:2006xy} is improved for the two inconsistent modes. In particular, table \ref{tab:check} demonstrates that the corresponding eigenvalues, like the value $m^2=17.68$, are rather stable with respect to the variation of the parameters of the numerical analysis.
Let us give some additional considerations in support of our spectrum. First, we note that our values give better quadratic fits for the low eigenvalues, as compared to the eigenvalues of \cite{Berg:2006xy} (for other glueballs in the KS background the quadratic fit usually works quite well),
\begin{eqnarray}
m^2&\approx&0.269 n^2+1.041 n+1.062\,, \label{fit1}\\
m^2&\approx&0.273 n^2-0.148 n+0.172\,,\\
m^2&\approx& 0.270 n^2 + 0.250 n +0.115\,,\\
m^2&\approx&0.268 n^2+0.775 n+0.233\,,\\
m^2&\approx&0.272 n^2+1.693 n+2.46\,,\\
m^2&\approx&0.272 n^2+1.975 n+3.383. \label{fit6}
\end{eqnarray}
These fits match well even the lowest states of table \ref{tab:full}, which is not the case of the fits obtained in~\cite{Berg:2006xy}. To illustrate this we plot in figure \ref{fig:fits} our fits and those of \cite{Berg:2006xy}. Note that the fits of \cite{Berg:2006xy} do not capture several values of $m^2$. Besides that, there are several degenerate states as seen in figure \ref{fig:fits} B. Our fits, on the other hand, exhibit a good match with the values of $m^2$ in table \ref{tab:full}. We also do not observe degenerate eigenvalues.
Yet another interesting consideration comes from the comparison with the lattice results.
The lowest states of $0^{-+}$ glueball in table \ref{tab:mid7} better matches the position of this glueball in the lattice calculations \cite{Morningstar:1999rf,Teper:1998kw,Chen:2005mg,Lucini:2004my,Lucini:2010nv,Holligan:2019lma,Athenodorou:2020ani,Gregory:2012hu}. To be precise, an exceptional match occurs when we use the lattice results for $SU(\infty)$ extrapolation \cite{Lucini:2004my}, which is the regime where we expect the supergravity approximation to be valid. In this case our result captures the fundamental state and its excitation, see table \ref{tab:lattice}.
\begin{table}[t]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\hline
$\tau_{IR}$ & $\tau_{UV}$ & $\tau_{mid}$& $\delta \tau$ & $n_{imposed}$& $m^2$ \\ \hline
0.005 & 15 & 1 & 0.01& 500 & 17.6838\\ \hline
0.005 & 20 & 1 & 0.001& 500 & 17.6861 \\ \hline
0.01 & 22 & 4 & 0.01& 400 & 17.6858 \\ \hline
0.1 & 20 & 1 & 0.01& 500 & 17.6782 \\ \hline
0.1 & 20 & 1 & 0.001& 500 & 17.6810 \\ \hline
0.2 & 18 & 2 & 0.01& 400 & 17.6709 \\ \hline
\end{tabular}
\end{center}
\caption{Values of $m^2 \approx 17.68$. for different chooses of the parameters of MDM.} \label{tab:check}
\end{table}
\begin{table}[t]
\begin{center}
\begin{tabular}{l|cccccc}
\hline
Ratio & \cite{Lucini:2004my} & This Work & \cite{Berg:2006xy} \\ \hline
$m_{0^{++}}/m_{2^{++}}$ & $0.689$ & $0.700$ & $0.640$ \\ \hline
$m_{0^{++*}}/m_{2^{++}}$ & $1.264$ & $1.264$& $1.249$ \\
\hline
$m_{\lambda\lambda}/m_{2^{++}}$ & -- & $0.511$ & $0.421$ \\
\hline
$m_{\lambda\lambda^\ast}/m_{2^{++}}$ & -- & $0.952$ & $0.894$ \\
\hline
\end{tabular}
\end{center}
\caption{Ratio of masses of the glueballs. As usual we normalize the spectrum by $m_{2^{++}}$. For holographic results we use the $m_{2^{++}}$ of \cite{Berg:2006xy}.}
\label{tab:lattice}
\end{table}
\section{Conclusion and Final Remarks}
\label{conclusion}
In this paper we have discussed the pseudoscalar glueballs of the KS theory. We have explained the derivation of the linearized supergravity equations that compute the spectrum and reviewed several consistency checks. In the numerical analysis it was possible to identify six families of eigenvalues. Four of those have good matches with the spectrum of scalar glueballs. The fact that the two remaining towers do not match with the remaining scalars indicates that results of the numerical approach in either $0^{-+}$ or $0^{++}$ sectors are incorrect. Here we have argued that the results of our numerical analysis are consistent. We also observed that our results show more expected behavior than those in the scalar sector, as demonstrated by fits \eqref{fit1}-\eqref{fit6}, figure~\ref{fig:my_label} and table \ref{tab:lattice}. In particular, our fits allow to reliably disentangle all six modes.
Another interesting conclusion that can be made based on our spectrum is the position of the lightest superpartner states in the $\mathcal{N}=1$ SYM theory, as compared to the lattice predictions of the purely bosonic sector. The separation of the six towers allows to conclude that the lightest state is likely to be the fermion bilinear $\lambda\lambda$. Its mass ratio with the $2^{++}$ glueball, shown in table~\ref{tab:lattice}, shows that it is unlikely the scalar $\Tr F^{\mu\nu}F_{\mu\nu}$. Moreover, the next excited state in the same tower, which corresponds to the value $m_{\lambda\lambda^{*}}/m_{2^{++}}=0.952$, does not have a good match in the bosonic sector as well, so for the six lightest states we find the hierarchy
\begin{equation}\nonumber
m_{\lambda\lambda}<m_{0^{++}}< m_{2^{++}} < m_{0^{-+}}<m_{\lambda\lambda^\ast} < m_{0^{++\ast}}
\end{equation}
Finally, we note that the $0^{++}$ and $0^{-+}$ sectors were derived in a different way. In \cite{Berg:2006xy} the derivation was done through a sigma model approach and in \cite{Melnikov:2020cxe} as well as in \cite{Dymarsky:2006hn,Benna:2006ib,Dymarsky:2007zs,Benna:2007mb,Dymarsky:2008wd,Gordeli:2009nw} it was done through the type IIB supergravity derivation. In a future work we plan to check the sigma model equations in the $0^{++}$ sector through the direct linearization of the supergravity equations and derive the sigma model equations for the pseudoscalars. We hope that this will allow us to settle the divergence issue.
\acknowledgements{The author would like to thank D. Melnikov for collaboration and useful comments on the manuscript of this letter. I would also like to thank Daniel Elander and Maurizio Piai for useful correspondence. I thank the financial support by Brazilian Ministry of Education. This research was developed within the project of the Brazilian agency CNPq, process 433935/2018-9.}
|
\section{Introduction \& Motivation} \label{s:intro}
Due to rapid growth of mobile devices in recent years, many mobile applications have become popular, such as image recognition\cite{6648564}, virtual reality (VR)\cite{biocca2013communication}, and augmented reality (AR)\cite{azuma1997survey}, which are generally computationally intensive. However, due to the heavy duty of computation tasks, it would be very energy-consuming to be completed by the mobile device itself. On the other hand, cloud computing can save the mobile device from computation resources, but will lead to long service delay since the cloud computing center is always based remotely from the mobile device.
In order to solve this dilemma, mobile edge computing (MEC) offers a promising solution, which implements {an edge server in the vicinity of the mobile device and permits the mobile device to offload its heavy computation task to the edge server \cite{8016573}.}
Regarding to offloading the computation task \footnote{{When the task offloading is mentioned in this paper, it actually means to offload the input data of a computation task (which would be explained in Section \ref{s:model_task}). To avoid misleading, this course is uniformed to be {\it task offloading} throughout this paper.}}
from one mobile device to the edge server, {binary and partial offloading are mainly investigated. For binary offloading, the computation task is not separable. Hence it is completed locally at the mobile device or offloaded to and then computed at the edge server.
For partial offloading, it is assumed that} the computation task can be separated into two sub-tasks by any fraction, with one sub-task completed at the mobile device and the other one offloaded to and then computed in parallel at the edge server
\cite{8016573}.
{
Within the framework of either binary or partial offloading, the mobile device's energy consumption and task completion delay are the two most significant cost metrics in literature, which can be highly reduced by optimizing the task offloading decision, the allocation of communication and computation resources \cite{7879258,wang2016mobile,8488502,7842160,8387798,7762913}.
Specifically, task offloading decision determines how much data is left for local computations and offloading to the edge server, respectively \cite{7879258}; }
the communication resource involves not only mobile device transmit power for task offloading \cite{wang2016mobile,8488502}, but also the bandwidth \cite{7842160} or time slot duration \cite{8387798,7762913} for every mobile device when multiple devices are accessing the edge server; the computation resource primarily indicates edge server or mobile device's CPU frequency (which represents the computation capability) over time.
The general research goal is to minimize energy consumption per mobile device while promising the computation task completed within a tolerable delay, by jointly optimizing task offloading decision, the allocation of communication and computation resources.
{
On the other hand, many important MEC applications for 5G \cite{Add_survey}, such as AR {\cite{azuma1997survey}}, video stream processing in immerse media {\cite{6648564}}, and deep neural network (DNN) driven intelligence applications including target recognition or anomaly detection {\cite{SIG-039}}, involve the type of computation task with multiple indivisible sub-tasks conforming to sequential dependency. To be specific, the whole computation task is comprised of multiple sub-tasks and every sub-task is executable only after completing its precedent sub-task, which is regarded as a {\it sequential task} in the following. Since a sequential task cannot be decomposed into two sub-tasks running in parallel, the strategies based on partial offloading in literature do not work anymore. Furthermore, although the strategies based on binary offloading in literature can help to generate a feasible solution for task offloading decision if we see the sequential task as a whole, energy consumption and time completion delay is not fully reduced since it can be further cut down when we choose to offload from some proper sub-task. Hence there exist model mismatch and performance degradation if the developed strategies based on partial offloading or binary offloading are utilized for our focused sequential task.}
{
In terms of task offloading for sequential tasks,
given that the edge server has more powerful computation capability over the mobile device, it would be sensible to offload its computation task to the edge server as early as possible, i.e., to offload the sub-task at an early stage. However, due to the randomness of channel gain and the inhomogeneous input data amount associated with every sub-task, {a} long delay would be lead to in case of large input data amount and poor channel condition, no matter the offloaded sub-task is {at an early stage} or not. Hence it is not always good to offload the sub-task at an early stage, and it is interesting to study the best occasion to offload, which corresponds to a task offloading decision problem.}
{
Although there is some MEC literature working on the task offloading decision problem for the sequential task \cite{6849257,6846368}, which would be surveyed in Section \ref{s:related_work}, they do not explore other ways to reduce mobile device's energy consumption and task completion delay. It would be beneficial to further perform joint allocation of communication and computation resources since both of them can affect the above two types of cost metrics.
However, due to the entanglement of multiple types of variables which will lead to the non-convexity of the investigated problem and unpredictable random fluctuation of channel state, it is challenging to solve the focused problem directly. Additionally, this challenging problem cannot be answered well by the developed strategies based on partial or binary offloading due to the aforementioned model mismatch and performance degradation.
}
{In this work, we solve the task offloading decision problem together with the allocation of communication and computation resources for sequential task.}
Specifically, {both slow and fast fading channels are considered.}
Under these two considered channel models, task offloading strategy, communication resource, and/or computation resource are jointly optimized so as to minimize the energy consumption of the mobile device, while guaranteeing the task computation latency. The contributions are summarized as follows.
\begin{itemize}
\item For slow fading channel, channel coherence time is long enough and the channel state keeps unchanged over the whole process of task completion, no matter it happens at local or at the edge server.
The task offloading strategy, transmit power when task offloading is performed by the mobile device, and the CPU frequency at local for computing every assigned sub-task at the mobile device are jointly optimized, {which corresponds to a mixed-integer non-convex optimization problem and is challenging to find the global optimal solution.
To tackle this challenge, the task offloading decision is fixed in the first step, in which case the problem for optimizing the rest variables turns to be a continuous one but is still non-convex.
In the second step, with some variable replacement and mathematical manipulations, the non-convex optimization problem is transformed to be a convex one equivalently, whose global solution is achievable.
In the third step, to further reduce computation complexity for solving the transformed convex optimization problem, it is decomposed into two levels. In the lower level, a closed-form optimal solution is derived by exploring its Karush-Kuhn-Tucker (KKT) condition. Then the upper-level problem can be simplified into a single-variable convex optimization problem, whose optimal solution can be easily found via the Golden search method.
In the last step, the optimal task offloading decision can be found by selecting the one with minimal energy consumption.
Through the above operations, we find the global optimal solution with low complexity for the original formulated mixed-integer non-convex optimization problem.}
\item For fast fading channels, the channel state varies frequently. The channel state when performing task offloading may fluctuate not only when the mobile device decides to offload data but also in the process of offloading the data of one sub-task. In this case, channel state information is non-predictable {and online policy is desired, which is challenging to work out. To address this challenge, a stochastic optimization problem is formulated}, which optimizes the instant decision on whether to offload subsequent sub-task, how much time is consumed for task offloading, and the transmit power for every fading block in the process of task offloading. {To find its optimal solution}, by exploring the special property of minimal expected energy consumption for task offloading, how much time should be consumed for task offloading is answered in the first step. {This facilitates the optimizing of the rest variables, whose global optimal solution is achieved by utilizing the method of dynamic programming. With the derived solution,
only one comparison with a value in a table (generated through offline computing) is required when performing the online policy, which is computationally efficient.}
What is even further and interesting, through theoretical analysis, we find that the dependency of mobile device energy consumption on instant channel state will fade away when the channel coherence time approaches zero. {In this case, the original online policy dwells into a simple offline policy.
Through the above operations, we do not only find out the optimal online policy with low running complexity, but also prove that the derived online policy could converge to a simple offline policy when channel coherence time is short, which further reduces the computation complexity.}
\end{itemize}
The rest of the paper is organized as follows.
In Section \ref{s:related_work}, related literature is surveyed.
Section \ref{s:model} presents the system model and formulates the research problems under slow fading channel and fast fading channel.
In Section \ref{s:prob_solution}, the optimal solution for the formulated research problems under slow fading channel and fast fading channel are presented respectively, followed by the discussion when channel coherence time approaches zero. Numerical results are given in Section \ref{s:numerical_results} and concluding remarks are presented in Section \ref{s:conclusion}.
\section{Related works} \label{s:related_work}
In literature, plenty of works have studied binary offloading or partial offloading.
Due to the limit of space, we only survey the most representative works here, which includes the reference papers \cite{6574874,wang2016mobile,8387798,7762913,Add_NOMA_full_Ding,Add_NOMA_full_Peng,Add_NOMA_partial_Wu,8537962,Add_energy_harvesting,gu2020mobile}.
{
For binary offloading, \cite{6574874} proposes to take the action with less energy consumption between the options of fully local computing and fully offloading for a single-user MEC system.
\cite{wang2016mobile} investigates partial offloading problem together with the communication and computation resource allocation for a single-user MEC system.
\cite{8387798} and \cite{7762913} studies the joint resource allocation and partial offloading for a time division multiple access (TDMA) and orthogonal frequency division multiple access (OFDMA) MEC system, respectively. Since non-orthogonal multiple access (NOMA) can achieve higher spectrum efficiency over orthogonal multiple access (OMA), it is also taken into account for task offloading. With the setup of fully offloading, the benefit of NOMA for a MEC system is firstly analyzed and verified in \cite{Add_NOMA_full_Ding} and the resource allocation problem in a NOMA-aided MEC system is investigated in \cite{Add_NOMA_full_Peng}. For partial offloading, \cite{Add_NOMA_partial_Wu} studies task offloading and resource allocation jointly with decoding order fixed, while \cite{8537962} further optimizes the decoding order. Besides partial offloading, binary offloading, together with resource allocation, is also investigated in a NOMA-MEC system \cite{Add_NOMA_partial_Wu}.
In addition to the multiple access technique, the idea of energy-neutral is also brought into the MEC system, with every mobile device driven by energy harvesting technique \cite{Add_energy_harvesting,gu2020mobile}.
For energy harvesting MEC system,
partial offloading is investigated with the mobile device supported by randomly arrived environmental energy \cite{Add_energy_harvesting}, while both partial offloading and binary offloading are explored with the mobile device supported by wireless-power-transfer technique in \cite{gu2020mobile}.
}
Compared with the rich literature focusing on binary offloading or partial offloading, research works taking the sequential task model into account are very limited, \cite{6849257,6846368,8854339}.
In \cite{6849257}, task completion delay is minimized for a single mobile device.
Not only link rate but also the computation capability at the mobile device and edge server are supposed to be stable.
Hence only task offloading strategy is needed to be investigated.
One-climb policy is disclosed to be optimal, in which the mobile device only offloads its computation task to the edge server once and only one segment of sub-tasks is computed at the edge server if ever.
With such a framework, \cite{6849257} figures out which segment of sub-tasks should be selected for the edge server
to achieve the minimal task completion delay.
In \cite{6846368}, a single mobile devices is considered and its expected energy consumption is minimized while limiting expected or outage probability of task completion delay.
Three types of channel conditions are investigated, which are given as static channel, independently and identically distributed (i.i.d) block fading channel, and Markovian stochastic channel.
Similar to \cite{6849257}, only task offloading decision is explored and one-climb policy is proven to be optimal for all the investigated cases.
In \cite{8854339}, a MEC system with two mobile devices is considered under static channel condition, both of which have a sequential task to compute.
Differently, one intermediate sub-task of the second mobile device requires the last sub-task's output of the first mobile device. Weighted sum of these two mobile devices' energy consumption and task completion delay is aimed to be minimized by optimizing not only the task offloading decision but also the CPU frequency and transmit power for offloading of these two mobile devices. With the task offloading decision given, the optimal solution of all the other variables is found. In terms of the task offloading decision, the one-climb policy is also proven to be optimal.
This research also investigates the MEC system under the sequential task model adopted in \cite{6849257} and \cite{6846368} for both slow fading channels and fast fading channels.
{
However, different from \cite{6849257} and \cite{6846368}, both of which solely focus on task offloading policy, we investigate the joint optimization of task offloading decision, communication resource (including mobile device's transmit power in every fading block) allocation, and/or computation resource allocation (including the allocated CPU for completing every sub-task).}
Although \cite{8854339} also investigates the joint design of task offloading decision and resource allocation,
{its working procedure involves inter-user task dependency, which is different from ours and the one in \cite{6849257} and \cite{6846368}. The research results published in \cite{8854339} cannot address the questions raised in this paper}.
\section{System Model and Problem Formulation}\label{s:model}
{Consider a MEC system with one edge server and multiple mobile devices. Every mobile device is allocated with one dedicated channel for task offloading and one dedicated CPU core at the edge server for computing. The bandwidth of the dedicated channel is $W$ and the dedicated CPU core at the edge server has a frequency of $f_e$
\footnote{Considering the incoming and outgoing of mobile devices in the coverage of the edge server, it would be computationally prohibitive if the edge server performs joint allocation of communication and computation resources among the connected mobile devices every once the set of connected mobile devices alters, which generally corresponds to a mixed-integer optimization problem. The dedicated allocation of spectrum and computation capability to every mobile device enable every mobile device to make its task offloading decision individually and separately, which can save the edge server from heavy and frequent optimization.}
.
Without loss of generality and for the ease of presentation, we can study the MEC system by investigating one mobile device and the associated CPU core at the edge server, which is abbreviated to be one edge server for the ease of presentation, in the following.}
The mobile device has a computation task to complete. The task may be computationally intensive and has to be completed within time $T_{\text{th}}$. To complete the computation task in time and to save mobile device energy consumption, the mobile device can offload all or part of the computation task to the edge server on the allocated spectrum. In order to capture a good opportunity for task offloading, instant channel gain between the mobile device and edge server will be measured before the mobile device makes the decision to offload. After the edge server completes the computation, it will return computational results to the mobile device, which is often of small data size and requires little time to be transmited from the edge server to the mobile device \cite{8387798,7762913,8537962}.
\subsection{Computation Task Model} \label{s:model_task}
For the computation task, a sequentially dependent task model is considered.
Specifically, the mobile application is assumed to be composed of $N$ sub-tasks, denoted as $\phi_1$, $\phi_2$, ..., $\phi_N$ respectively, the indices of which constitutes the set $\mathcal{N} \triangleq \{1, 2, ..., N\}$. These $N$ sub-tasks has to be completed sequentially, i.e., one can only compute the task of $\phi_i$ after completing the task of $\phi_{i-1}$, for $i=2, 3, ..., N$.
Every sub-task can be characterized by two parameters. For sub-task $i\in \mathcal{N}$, the associated parameter pair is $(l_i, d_i)$, where $l_i$ indicates the amount of computation (which is in the unit of CPU cycles) and $d_i$ describes the amount of input data size (which is in the unit of nats for the ease of presentation). Note that with a sequentially dependent setup for the task model, the input data of sub-task $\phi_i$, which has a size of $d_i$, is also the output data of sub-task $\phi_{i-1}$.
With the selection of the sequentially dependent task model, the mobile device will first compute some sub-tasks at local and then hand over the computing of the rest of sub-tasks to the edge server.
\subsection{Computation Model}
In terms of computation, it is composed of local computation both at the mobile device and edge server. Suppose the computation capability (which is also called CPU frequency) of the mobile device is $f_l \in [0, f_{\max}]$.
{Then the running power of the mobile device CPU is $k_0
f_l^3$ where $k_0$ is a physical parameter depending on the CPU's architecture according to \cite{BurdProcessor,6846368}.
Hence to complete a computation task with the computation amount being $l$, the time consumed is $l/f_l$ and the consumed energy is $k_0 f_l^2 l$.}
{
To save energy consumption of {the} mobile device considering that it is energy sensitive and its running power is largely determined by CPU frequency,
$f_l$ is assumed to be adjustable over disjoint sub-tasks. }
For sub-task $i \in \mathcal{N}$, suppose the associated CPU frequency is $f_i$, then to complete a sub-task, say $\phi_i$, the associated time consumption would be
\begin{small}
\begin{equation} \label{e:time_local_comp}
t_m (i) = \frac{l_i}{f_i}, \forall i \in \mathcal{N},
\end{equation}
\end{small}
and the associated energy consumption would be
\begin{small}
\begin{equation} \label{e:energy_local_comp}
e_m (i) = k_0 \times l_i \times f_i^2, \forall i \in \mathcal{N}.
\end{equation}
\end{small}
The computation capability of the edge server, i.e., the CPU frequency, is given as $f_e$ where $f_e > f_{\max}$. According to the previous discussion, to complete a sub-task, say $\phi_j$, the time consumption would be
\begin{small}
\begin{equation} \label{e:time_edge_comp}
t_e (j) = \frac{l_j}{f_e}, \forall j \in \mathcal{N}.
\end{equation}
\end{small}
\subsection{Communication Model} \label{s:communication_model}
The communication activity in the system encompasses uploading the data of some sub-task from the mobile device to the edge server, and downloading the data of computational results from the edge server to the mobile device.
In terms of data downloading, the data {size} of computational results is always small. In addition, the edge server has {a} strong power supply and can transmit information with {a} high data rate. Hence the time consumption for data downloading is omitted. This assumption is broadly adopted \cite{7762913,gu2020mobile}.
In terms of task offloading, if the mobile device decides to stop local computing by sub-task $\phi_{n-1}$, then needs to offload the data for computing sub-task $\phi_n$ to the edge server, which has an amount of $d_n$ nats. In the process of offloading these $d_n$ nats, the associated channel gain from the mobile device to the edge server is denoted as $h_n$ for $n\in \mathcal{N}$.
When the timespan of task offloading is less than the channel coherence time, which is denoted as $\tau$, the $h_n$ would be a single random variable. Otherwise, the $h_n$ should be split into multiple independently distributed fading blocks, which can be denoted as $h_n = \left(h_{n,1}, h_{n,2}, h_{n,M(n)}\right)^T$. Detailed discussion on this point will be offered in the rest part of this subsection. Note that for the ease of presentation, the $h_n$ for $n \in \mathcal{N}$ throughout this paper is actually the normalized channel gain, which represents the received SNR at the edge server when the transmit power of the mobile device is 1. In addition, $h_n$ or $h_{n,m}$ are all supposed to be i.i.d.
By comparing the time-scale of task offloading and sub-task computation at local with the channel coherence time $\tau$, two categories of channel models can be expected.
\subsubsection{Slow fading channel} \label{s:communication_model_one}
In this type of channel model, channel coherence time would be longer than $T_{\text{th}}$.
In this case, $h_i$ does not change with $i$ for $i \in \mathcal{N}$ in the whole process of completing the computation task, no matter the sub-task is being processed locally, offloaded, or processed at the edge server. Denote the unchanged $h_i$ for $i\in \mathcal{N}$ as $h$, then the mobile device only needs to perform one measurement of $h$ before it starts to execute the computation task.
With such an assumption, denote the transmit power of the mobile device as $P_t$ when it is executing task offloading, then the data transmission rate between the mobile device and edge server can be written as $r= W \ln \left(1 + P_t h\right)$. In order to offload $d_n$ nats when the mobile device stops local computing by sub-task $\phi_{n-1}$, the time consumption is
\begin{small}
\begin{equation} \label{e:time_offload_one_fading}
T_{\text{off}}^{\text{I}}(n) = \frac{d_n}{W \ln \left(1 + P_t h\right)}, \forall n \in \mathcal{N}
\end{equation}
\end{small}
and the energy consumption can be written as
\begin{small}
\begin{equation} \label{e:energy_offload_one_fading}
E_{\text{off}}^{\text{I}}(n) = \frac{d_n P_t}{W \ln \left(1 + P_t h\right)}, \forall n \in \mathcal{N}.
\end{equation}
\end{small}
\subsubsection{Fast fading channel} \label{s:communication_model_multiple}
In this type of channel model, channel coherence time $\tau$ is much shorter {than not only the time scale of computing any sub-task but also the time scale of offloading the input data of any sub-task}.
Therefore, for $n \in \mathcal{N}$, the task offloading time window, whose duration is $T(n)$, may contain multiple (say $M(n)$) fading blocks, which are indexed as $(n, 1)$, $(n, 2)$, ..., $(n, M(n))$, respectively.
In this case, $h_n$ is a vector rather a scalar, which is composed of independently distributed $h_{n,1}$, $h_{n,2}$, ..., $h_{n, M(n)}$, for $n \in \mathcal{N}$. In addition, we suppose the random variables $h_{n,1}$, $h_{n,2}$, ..., $h_{n, M(n)}$ are also subject to a common distribution function $p(\cdot)$ with a sample space $\mathcal{H}$.
Within $T_n$, there are $M(n)$ fading blocks. Define $t_{n,m}$ as the time duration of $m$th fading block for $n \in \mathcal{N}$, $m \in \{1, 2, ..., M(n)\}$, then there is
$t_{n,m} = \tau, \text{if~} m < M(n)$, and $t_{n,m} = t, \text{if~} m = M(n)$,
where $t\in (0, \tau]$ characterizes the time duration of the ending fading block for task offloading, which may be less than $\tau$ since $T(n)$ may not be exactly integer multiple of channel coherence time $\tau$.
Hence $T_n$ satisfies
$T_n = (M(n)-1)\tau + t, \forall n \in \mathcal{N}$.
When the mobile device decides to offload the input data for computing sub-task $\phi_n$, it will experience $M(n)$ fading blocks.
In one fading block whose time duration is $t_{n,m}$ and channel gain is $h_{n,m}$, then to offload $d$ nats, according to Shannon capacity formula, the associated energy consumption is
$e(d, h_{n,m}, t_{n,m}) = {\left(e^{\frac{d}{Wt_{n,m}}} - 1\right)t_{n,m}}/{h_{n,m}}$.
Define $Q_{n,m}(d, h; M(n),t)$ as the minimum amount of energy that the mobile device will consume for task offloading over the rest $(M(n) - m + 1)$ fading blocks under the following four conditions:
1) the channel gain at the beginning of fading block $(n, m)$, i.e., $h_{n,m}$, is measured to be $h$;
2) there are $d$ data nats to offload;
3) the total number of fading blocks for the task offloading is $M(n)$;
4) the last fading block, i.e, fading block $(n,M(n))$, has a time duration of $t$.
Define $Q_{n,m}(d; M(n), t)$ as the minimum amount of energy that the mobile device will consume for the task offloading over the rest $(M(n) - m + 1)$ fading blocks when the mobile device has no knowledge of channel gain $h_{n,m}$ at the beginning of fading block $(n, m)$ and the above mentioned conditions 2, 3, and 4 are satisfied.
With such a definition, when the mobile device finishes computing sub-task $\phi_{n-1}$, it will first measure the channel gain $h_{n,1}$, then evaluate the function $Q_{n,1}(d_n, h_{n,1}; M(n), t)$, and finally make the decision on whether to offload the data for computing sub-task $\phi_n$ based on the evaluated $Q_{n,1}(d_n, h_{n,1}; M(n), t)$. How to make the decision according to $Q_{n,1}(d_n, h_{n,1}; M(n), t)$ will be discussed in Section \ref{OS3}.
According to the discussion in \cite{gu2020mobile}, the functions of $Q_{n,m}(d, h; M(n), t)$ and $Q_{n,m}(d; M(n), t)$ for $m\in \{1, 2,..., M(n)\}$ and $n\in \mathcal{N}$ can be calculated iteratively in the following way.
\begin{small}
\begin{equation} \label{e:Q_int}
{Q_{n,m}}\left(d; M(n), t\right)
= \int_{0}^{\infty} {{Q_{n,m}}\left(d,{h_{n,m}; M(n), t}
\right)p\left( h_{n,m}\right)d{h_{n,m}}} , \forall m \in \left\{ {1,...,M(n)} \right\}, \forall n \in \mathcal{N},
\end{equation}
\end{small}
\begin{small}
\begin{equation} \label{e:Q_iterative}
\begin{split}
& Q_{n,m}(d,{h_{n,m}}; M(n), t) \\
= & \mathop {\min }\limits_{0 \leq {d_{n,m}} \leq d} \bigg( e(
{{d_{n,m}},{h_{n,m}}}, t_{n,m}) +
\int_{0}^{\infty} {{Q_{n,m+1}}\left( {d - d_{n,m},{h_{n,m+1}}; M(n), t} \right)
p\left( {{h_{n, m+1}}} \right)d{h_{n,m+1}}} \bigg), \\
= & \mathop {\min }\limits_{0 \leq {d_{n,m}} \leq d} \bigg( e(
{{d_{n,m}},{h_{n,m}}}, t_{n,m}) +
{{Q_{n,m+1}}\left( d - d_{n,m}; M(n), t \right)} \bigg),
\forall m \in \{1, 2,...,M(n)-1\}, \forall n \in \mathcal{N},
\end{split}
\end{equation}
\end{small}
The $d_{n,m}$ in (\ref{e:Q_iterative}) represents the amount of data to be offloaded within fading block $(n,m)$,
and
\begin{small}
\begin{equation} \label{e:Q_init}
{Q_{n,M(n)}}\left({d,h_{n,M(n)}}; M(n), t\right) = e\left( {d,{h_{n,M(n)}},t_{n,M(n)}} \right)
= \frac{\left({e^{\frac{d}{tW}}} - 1\right) t}{{h_{n,M(n)}}}, \forall n \in \mathcal{N}.
\end{equation}
\end{small}
\subsection{Problem Formulation}\label{s:p}
With the two channel models introduced in Section \ref{s:communication_model}, our target is to minimize the total energy consumption of the mobile device while guaranteeing the computation task is completed within time $T_{\text{th}}$. Specifically, the associated optimization problems are given as follows.
\subsubsection{Slow fading channel} \label{s:prob_for_I}
In this case, the total energy consumption of the mobile device is composed of two parts: 1) The energy consumption for local computing; 2) The energy consumption for task offloading.
Hence combining the expressions in (\ref{e:energy_local_comp}) and (\ref{e:energy_offload_one_fading}), the total energy consumption if the mobile device decides to offload data exactly after completing sub-task $\phi_{n-1}$ can be written as
\begin{small}
\begin{equation} \label{e:energy_consum_one_slot}
E_1(n) = \frac{{{d_n}
{P_t}}}{{W \ln \left( {1 + {P_t} h} \right)}} +
\sum\limits_{i = 1}^{n{\rm{ - 1}}} {{l_i} f_i^2 k_0}, \forall n \in \mathcal{N},
\end{equation}
\end{small}
where the second term of the right-hand side of (\ref{e:energy_consum_one_slot}) represents the energy consumption of the mobile device for computing sub-tasks $\phi_1$, $\phi_2$, ..., $\phi_{n-1}$ at local.
For the time consumption to complete the mobile application, if the mobile device decides to offload data exactly after completing sub-task $\phi_{n-1}$, by recalling (\ref{e:time_local_comp}), (\ref{e:time_edge_comp}), and (\ref{e:time_offload_one_fading}), the total time consumption, denoted as $T_1(n)$, can be written as
\begin{small}
\begin{equation} \label{e:time_consum_one_slot}
T_1(n) = \sum\limits_{i = 1}^{n{\rm{ -
1}}}{\frac{l_i}{f_i}} + \frac{d_n}{W \ln \left({1+P_t h}\right)} + \sum\limits_{i = n}^N {\frac{l_i}{f_e}}, \forall n \in \mathcal{N}.
\end{equation}
\end{small}
Then we need to select the stopping time $n\in \mathcal{N}$ and optimize $P_t$ and $f_i$ for $i\in \{1,2,...,n-1\}$, so as to minimize the energy consumption in (\ref{e:energy_consum_one_slot}) while requiring the $T_1(n)$ to be no larger than $T_{\text{th}}$. Specifically, to minimize the total energy consumption of the mobile device, the following optimization problem needs to be solved
\footnote{{Since the running time for working out the optimal solution of Problem \ref{s:p:constant_gain} by the mobile device is ignorable compared with the time scale of $T_{\text{th}}$, which will be demonstrated in Section \ref{s:num_slow}, it is not necessary to take into account this part of time consumption in (\ref{p1_1}). Even if it has to be accounted for, the $T_{\text{th}}$ in constraint (\ref{p1_1}) can be replaced with $\left(T_{\text{th}} - \Delta_c\right)$, where $\Delta_c$ represents the running time for solving the optimization problem. A similar discussion also follows for constraint (\ref{e:s:p:multiple_deadline}) in Problem \ref{s:p:multiple}}.}
\begin{prob}\label{s:p:constant_gain}
\begin{small}
\begin{subequations}
\begin{align}
\mathop{\min} \limits_{P_t,n,\{f_i\}} & \quad \frac{d_n
P_t}{W \ln \left( {1 + {P_t} h} \right)} +
\sum\limits_{i = 1}^{n{\rm{ - 1}}} {{l_i} f_i^2 k_0} \nonumber \\
\text{s.t.} & \quad
\sum\limits_{i = 1}^{n - 1} \frac{l_i}{f_i} + \frac{d_n}{W \ln \left({1+P_t h}\right)} + \sum\limits_{i = n}^N {\frac{l_i}{f_e}} \le T_{\text{th}} , \label{p1_1} \\
& \quad 0 \leq {f_i} \le {f_{\max}} , \forall i \in \{1,2,...,n-1\}; \quad P_t \geq 0; \quad n \in \mathcal{N}.
\end{align}
\end{subequations}
\end{small}
\end{prob}
\subsubsection{Fast fading channel} \label{s:prob_for_III}
In this case, \\$h_{1,1}, ..., h_{1,M(1)}, h_{2,1}, ..., h_{2,M(2)}, ..., h_{N,1}, ..., h_{N,M(N)}$ are all i.i.d random variables. The mobile device cannot determine when to offload data in advance, due to the randomness of channel gains in forthcoming fading blocks. Therefore, online policy is required to be designed.
When the mobile device finishes the computing of sub-task $\phi_{n-1}$ at local, it can measure the channel gain $h_{n,1}$ and then makes the decision on whether to offload the input data of sub-task $\phi_n$ right now.
Hence before the mobile device determines to offload the computing data of sub-task $\phi_n$, it has the knowledge of ${h}_1, {h}_2, ..., {h}_{n-1}, h_{n,1}$.
In this case the stopping rule $n$ can be expressed as $n\left({h}_1,...,{h}_{n-1}, h_{n,1}\right)$ or $n\left({h}_1 \rightarrow h_{n,1}\right)$ for brevity.
Without confusion and for the ease of presentation, we will interchange the use of $n$ and $n\left({h}_1 \rightarrow h_{n,1}\right)$ in the following discussion for this scenario.
In terms of $f_i$ for $i\in \{1, 2, ..., n-1\}$, from the perspective of the mobile device, it is unclear when it will switch to task offloading, not to mention how to dynamically adjust $f_i$ over $i\in \{1,2,...,n-1\}$. To be equal with every sub-task, the mobile device will fix $f_i$ at $f_l \in [0, f_{\max}]$ for $i\in \{1, 2, ..., n-1\}$.
In this scenario, the time delay for completing the whole mobile application can be written as
\begin{small}
\begin{equation}
T_3(n, M(n), t) = \sum\limits_{i = 1}^{n -
1}{\frac{l_i}{f_l}} + \left(M(n)-1\right)\tau + t + \sum\limits_{i = n}^N {\frac{l_i}{f_e}}, \forall n \in \mathcal{N}.
\end{equation}
\end{small}
Then to meet the deadline requirement, there should be
$T_3(n, M(n), t) \leq T_{\text{th}}$.
The incurred energy consumption when the mobile device decides to stop local computing by sub-task $\phi_{n-1}$ can be written as
\begin{small}
\begin{equation}
E_3(n, M(n), t)
= Q_{n,1}(d_n, h_{n,1}; M(n), t) +
\sum\limits_{i = 1}^{n- 1} {{l_i} f_l^2 k_0}, \forall n \in \mathcal{N},
\end{equation}
\end{small}
then to minimize the expected total energy consumption of the mobile device, the associated optimization problem can be formulated as
\begin{prob}\label{s:p:multiple}
\begin{small}
\begin{subequations}
\begin{align}
\mathop {\min }\limits_{\substack{n({h}_1\rightarrow h_{n,1}), M(n), t}} \quad & \mathbb{E}_{{h}_1, ..., {h}_N} \left\{E_3(n, M(n), t)\right\} \nonumber \\
\text{s.t.} \quad & n({h}_1\rightarrow h_{n,1}) \in \mathcal{N}, \\
& T_3(n, M(n), t) \leq T_{\text{th}}, \label{e:s:p:multiple_deadline}\\
& M(n) \in \mathcal{I}; 0 < t \leq \tau,
\end{align}
\end{subequations}
\end{small}
\end{prob}
where $\mathcal{I}$ is the integer set.
In Problem \ref{s:p:multiple}, it should be noticed that not only $n({h}_1\rightarrow h_{n,M(n)})$ but also $M(n)$ and $t$ are all online policies depending on ${h}_1, {h}_2, ..., {h}_{n-1}, h_{n,1}$, and the objective function represents the expected energy consumption $E_3(n, M(n), t)$ over the possible realizations of ${h}_1, {h}_2, ..., {h}_N$.
\section{Optimal Solution} \label{s:prob_solution}
\subsection{Optimal Solution of the Task Offloading and Resource Allocation Problem in Slow Fading Channels} \label{OS1}
In this subsection, the problem of task offloading and resource allocation in slow fading channels, which corresponds to Problem \ref{s:p:constant_gain}, would be solved.
Problem \ref{s:p:constant_gain} involves the optimizing of variables $n\in \mathcal{N}$, $P_t$, and $f_i$ for $i\in \{1, 2, ..., n-1\}$, which contains both integer and continuous variables, and thus cannot be solved directly.
To find the optimal solution of Problem \ref{s:p:constant_gain}, we decompose it into two levels. In the {lower} level, $n$ is fixed while the rest variables $P_t$ and $f_i$ for $i\in \{1, 2, ..., n-1\}$ are optimized. The associated optimization problem can be written as
\begin{prob}\label{p:I_lower}
\begin{small}
\begin{subequations}
\begin{align}
E_{I}(n) \triangleq
\mathop{\min} \limits_{P_t,\{f_i\}} \quad & \frac{d_n
P_t}{W \ln \left( {1 + {P_t} h} \right)} +
\sum\limits_{i = 1}^{n- 1} {k_0 l_i f_i^2} \nonumber \\
\text{s.t.} &
\sum\limits_{i = 1}^{n - 1} \frac{l_i}{f_i} + \frac{d_n}{W \ln \left({1+P_t h}\right)} + \sum\limits_{i = n}^N {\frac{l_i}{f_e}} \le T_{\text{th}} , \label{e:I_lower_active_inequality} \\
& 0 \leq {f_i} \le {f_{\max}} , \forall i \in \{1,2,...,n-1\}; P_t \geq 0.
\end{align}
\end{subequations}
\end{small}
\end{prob}
In the upper level, optimal $n$ should be found so as to solve the following optimization problem
\begin{prob} \label{p:I_upper}
\begin{small}
\begin{equation}
\mathop{\min} \limits_{n \in \mathcal{N}} \quad E_{I}(n)
\end{equation}
\end{small}
\end{prob}
It can be checked that the upper level problem, Problem \ref{p:I_upper}, is equivalent with the original optimization problem, Problem \ref{s:p:constant_gain}. In the following, the lower level problem, Problem \ref{p:I_lower} will be solved first.
Problem \ref{p:I_lower} is a non-convex optimization problem, since the function $\frac{d_n P_t}{W \log \left(1+ P_t h\right)}$ in the objective function of Problem \ref{p:I_lower} is non-convex with $P_t$. To simplify the solving of Problem \ref{p:I_lower}, we define an assistant variable $\tau_t = \frac{d_t}{W\log\left(1+P_t h\right)}$, which represents the time consumption for offloading the input data of sub-task $\phi_n$. By expressing $P_t$ with $\tau_t$, i.e., $P_t = \left(e^{\frac{d_t}{W \tau_t}} - 1\right)\frac{1}{h}$, Problem \ref{p:I_lower} turns to be the following optimization problem
\begin{prob}\label{p:I_lower_convex}
\begin{small}
\begin{subequations}
\begin{align}
E_{I}(n) \triangleq
\mathop{\min} \limits_{\tau_t,\{f_i\}} & \quad \tau_t \left(e^{\frac{d_n}{W \tau_t}} - 1\right) \frac{1}{h} +
\sum\limits_{i = 1}^{n- 1} {k_0 l_i f_i^2} \nonumber \\
\text{s.t.} & \quad
\sum\limits_{i = 1}^{n - 1} \frac{l_i}{f_i} + \tau_t + \sum\limits_{i = n}^N {\frac{l_i}{f_e}} \le T_{\text{th}} , \label{e:I_lower_convex_active_inequality} \\
& \quad 0 \leq {f_i} \le {f_{\max}} , \forall i \in \{1,2,...,n-1\}, \label{e:I_lower_convex_f_interval} \\
& \quad \tau_t \geq 0.
\end{align}
\end{subequations}
\end{small}
\end{prob}
For Problem \ref{p:I_lower_convex}, it can be checked that the function $\tau_t \left(e^{\frac{d_n}{W \tau_t}} - 1\right) \frac{1}{h}$ is convex with $\tau_t$ since its second order derivative with $\tau_t$ is larger than 0. In addition, both the function ${k_0 l_i f_i^2} $ and the function $\frac{l_i}{f_i} $ are convex with $f_i$ for $i \in \{1, 2, ..., n-1\}$. Hence Problem \ref{p:I_lower_convex} is a convex optimization problem, which can be solved by traditional methods, such as the elliptical method or interior method \cite{boyd_vandenberghe_2004}.
To further simplify the solving of Problem \ref{p:I_lower_convex}, we target to design a simple yet optimal solution for Problem \ref{p:I_lower_convex}.
To explore the special property of the optimal solution of Problem \ref{p:I_lower_convex}, with a given $\tau_t$, we investigate the following optimization problem
\begin{prob}\label{p:I_lower_lower}
\begin{small}
\begin{subequations}
\begin{align}
E_{I}^{\text{loc}}\left( {\tau_t, n} \right) \triangleq &\mathop {\min }\limits_{\{f_i\}} \sum\limits_{i = 1}^{n-1} {k_0 f_i^2{l_i}} \nonumber \\
\text{s.t.} \quad &0 \leq {f_i} \le {f_{\max}}, \forall i \in \{1, 2, ..., n-1\}, \label{e:l_lower_lower_box}\\
&\sum\limits_{i=1}^{n-1}{\frac{l_i}{f_i}} \le
T_{\text{th}}- \tau_t - \sum\limits_{i = n}^N {\frac{l_i}{f_e}} \label{e:I_lower_lower_f_cons}.
\end{align}
\end{subequations}
\end{small}
\end{prob}
In Problem \ref{p:I_lower_convex}, since $f_i \leq f_{\max}$ for $i \in \{1, 2, ..., n-1\}$ according to (\ref{e:I_lower_convex_f_interval}), $\tau_t$ has to be no larger than $\left(T_{\text{th}} - \sum_{i=1}^{n-1}\frac{l_i}{f_{\max}} - \sum_{i=n}^{N} \frac{l_i}{f_e}\right)$ according to (\ref{e:I_lower_convex_active_inequality}).
With this implicit constraint on $\tau_t$ and the definition of $E_I^{\text{loc}}(\tau_t, n)$ in Problem \ref{p:I_lower_lower}, Problem \ref{p:I_lower_convex} is equivalent with the following optimization problem
\begin{prob} \label{p:I_lower_upper}
\begin{small}
\begin{subequations}
\begin{align}
E_{I} \left( n \right) = \mathop{\min} \limits_{\tau_t} & \quad \tau_t \left(e^{\frac{d_n}{W \tau_t}} - 1\right) \frac{1}{h} + E_{I}^{\text{loc}}\left( {\tau_t, n} \right) \nonumber \\
\text{s.t.} & \quad 0 \leq \tau_t \leq T_{\text{th}} - \sum_{i=1}^{n-1}\frac{l_i}{f_{\max}} - \sum_{i=n}^{N} \frac{l_i}{f_e}.
\end{align}
\end{subequations}
\end{small}
\end{prob}
Problem \ref{p:I_lower_lower} is convex and satisfies the Slater's condition, then the Karush-Kuhn-Tucker (KKT) condition can serves as a sufficient and necessary condition for the optimal solution
\cite{boyd_vandenberghe_2004}. Specifically, the Lagrange function of Problem \ref{p:I_lower_lower} can be written as
\begin{small}
\begin{equation}
\begin{array}{ll}
L(\{f_i\}, \lambda, \{\mu_i\}, \{\nu_i\}) = \sum_{i=1}^{n-1} k_0 l_i f_i^2 - \sum_{i=1}^{n-1} \mu_i f_i
+ \lambda \left(\sum\limits_{i = 1}^{n - 1} \frac{l_i}{f_i} + \tau_t + \sum\limits_{i = n}^N {\frac{l_i}{f_e}} - T_{\text{th}}\right) + \sum_{i=1}^{n-1} \nu_i \left(f_i - f_{\max}\right)
\end{array}
\end{equation}
\end{small}
where $\lambda \geq 0$ is the Lagrange multiplier associated with constraint (\ref{e:I_lower_lower_f_cons}), $\mu_i \geq 0$ and $\nu_i\geq 0$ are the Lagrange multipliers associated with constraints $f_i \geq 0$ and $f_i \leq f_{\max}$ in (\ref{e:l_lower_lower_box}) for $i \in \{1, 2, ..., n-1\}$, respectively.
Then the KKT condition can be listed as in (\ref{e:KKT})
\begin{small}
\begin{subequations} \label{e:KKT}
\begin{align}
2k_0 f_i l_i - \lambda \frac{l_i}{f_i^2} - \mu_i + \nu_i = 0, \forall i \in \{1, 2, ..., n-1\} \label{e:KKT_Lagrange}\\
\sum\limits_{i = 1}^{n - 1} \frac{l_i}{f_i} + \tau_t + \sum\limits_{i = n}^N {\frac{l_i}{f_e}} \le T_{\text{th}} \label{e:KKT_sum_cons}\\
0 \leq {f_i} \le {f_{\max}} , \forall i \in \{1,2,...,n-1\} \label{e:KKT_box_cons}\\
\lambda \left(\sum\limits_{i = 1}^{n - 1} \frac{l_i}{f_i} + \tau_t + \sum\limits_{i = n}^N {\frac{l_i}{f_e}} - T_{\text{th}}\right) = 0 \label{e:KKT_slack_sum_cons}\\
\mu_i f_i, \forall i \in \{1, 2, ..., n-1\} \label{e:KKT_slack_lower_bound}\\
\nu_i \left(f_i - f_{\max}\right), \forall i \in \{1, 2, ..., n\} \label{e:KKT_slack_upper_bound}\\
\lambda \geq 0 \label{e:KKT_lambda}\\
\nu_i \geq 0, \mu_i \geq 0, \forall i \in \{1, 2, ..., n-1\} \label{e:KKT_mu_nu}
\end{align}
\end{subequations}
\end{small}
With KKT condition in (\ref{e:KKT}), the optimal solution of $f_i$ for $i\in \{1, 2, ..., n-1\}$ can be derived, which is given in the following lemma.
\begin{thm} \label{lem:I_optimal_f}
The optimal solution of Problem \ref{p:I_lower_lower} is $f_i = \frac{\sum_{i=1}^{n-1} l_i}{\left(T_{\text{th}} - \sum_{i=n}^{N} \frac{l_i}{f_e} - \tau_t\right)}$ for $i\in \{1, 2, ..., n-1\}$ when $\tau_t \in \left(0, T_{\text{th}} - \sum_{i=1}^{n-1} \frac{l_i}{f_{\max}} - \sum_{i=n}^{N}\frac{l_i}{f_e}\right]$.
\end{thm}
\begin{IEEEproof}
Please refer to Appendix \ref{app:l_optimal_f}.
\end{IEEEproof}
To this end, the optimal solution of Problem \ref{p:I_lower_lower} has been found, and its minimal achievable cost function can be written as
$E_I^{\text{loc}}(\tau_t, n) = {k_0 \left(\sum_{i=1}^{n-1} l_i \right)^3}/{\left(T_{\text{th}} - \sum_{i=n}^{N} {l_i}/{f_e} - \tau_t\right)^2}$
which is actually a convex function with $\tau_t$. Next we go back to solve Problem \ref{p:I_lower_upper}.
Given the expression of $E_I^{\text{loc}}(\tau_t, n)$, it can be easily seen that for $\tau \in [0, T_{\text{th}} - \sum_{i=1}^{n-1}\frac{l_i}{f_{\max}} - \sum_{i=n}^{N} \frac{l_i}{f_e}]$, the objective function of Problem \ref{p:I_lower_upper} is a convex function. Hence Problem \ref{p:I_lower_upper} is a one-dimensional convex optimization problem, whose optimal solution can be found by the Golden search method \cite{press2007numerical}. To this end, Problem \ref{p:I_lower_upper}, which is also equivalent to Problem \ref{p:I_lower}, has been solved optimally.
Notice that our presented method only requires one dimension numerical search for the optimal $\tau_t$ while the optimal solutions of all the other variables, i.e., $f_i$ for $i\in \{1, 2, ..., n-1\}$, are returned in closed-form. This procedure only involves a computation complexity of $O\left(\log\left(\frac{T_{\text{th}} - \sum_{i=1}^{n-1}\frac{l_i}{f_{\max}} - \sum_{i=n}^{N} \frac{l_i}{f_e}}{\delta}\right)\right)$, where $\delta$ is the tolerance. Compared with traditional numerical methods,
which has a computation complexity $O(n^3)$ \cite{boyd_vandenberghe_2004}, our presented method can largely reduce the computation complexity.
In terms of solving Problem \ref{p:I_upper}, by evaluating $E_I(n)$ for $n \in \mathcal{N}$, the $n$ that associates the minimal $E_I(n)$ can be recognized as the optimal solution of Problem \ref{p:I_upper}, and the optimal solution of Problem \ref{s:p:constant_gain} as well. At the end of this subsection, the whole procedure for solving Problem \ref{s:p:constant_gain} is summarized in the following algorithm.
\begin{algorithm}[H]
\caption{The procedure for solving Problem \ref{s:p:constant_gain}.}
\begin{algorithmic}[1]\label{a:I}
\FOR {$n = 1:N$}
\STATE {With the given $n$, use the Golden section method to find the optimal $\tau_t$ for Problem \ref{p:I_lower}, which is denoted as $\tau_t^*(n)$, and record the associated minimal achievable cost function $E_I(n)$.}
\ENDFOR
\STATE {Find $n^* = \mathop{\arg\min} \limits_{n \in \mathcal{N}} E_I(n)$.}
\STATE {Output $n^*$ as the optimal solution of $n$ for Problem \ref{s:p:constant_gain}, $\left(e^{\frac{d_{n^*}}{W \tau_t^*(n^*)}}-1\right)\frac{1}{h}$ as the optimal $P_t$ of Problem \ref{s:p:constant_gain}, $\frac{\sum_{i=1}^{n-1} l_i}{\left(T_{\text{th}} - \sum_{i=n}^{N} \frac{l_i}{f_e} - \tau_t^*(n^*)\right)}$ as the optimal $f_i$ of Problem \ref{s:p:constant_gain} for $i = 1, 2, ..., n^*-1$.}
\end{algorithmic}
\end{algorithm}
\begin{remark} \label{r:mark_1}
In Algorithm \ref{a:I}, the computation complexity to evaluate $E_I(n)$ can be written as $O\left(\log\left(\left({T_{\text{th}} - \sum_{i=1}^{n-1}{l_i}/{f_{\max}} - \sum_{i=n}^{N} {l_i}/{f_e}}\right)/{\delta}\right)\right)$. Hence the whole computation complexity to find the optimal solution of Problem \ref{s:p:constant_gain} is $O\left(\sum_{n=1}^{N} \log\left(\left({T_{\text{th}} - \sum_{i=1}^{n-1} {l_i}/{f_{\max}} - \sum_{i=n}^{N} {l_i}/{f_e}}\right)/{\delta}\right)\right)$.
\end{remark}
\subsection{Optimal Solution of the Task Offloading and Resource Allocation Problem in Fast Fading Channels} \label{OS3}
In this subsection, the problem of task offloading and resource allocation in fast fading channels, which corresponds to Problem \ref{s:p:multiple}, would be solved.
In Problem \ref{s:p:multiple}, not only $n\left({h}_1 \rightarrow h_{n, 1} \right)$ but also $M(n)$ and $t$ are required to be solved. To solve Problem \ref{s:p:multiple}, the case with $n\left({h}_1 \rightarrow h_{n, 1} \right)$ given is firstly investigated.
With $n\left({h}_1 \rightarrow h_{n, 1} \right)$ given, define $Q(n, h_{n,1})$ as the minimal achievable cost of the following optimization problem
\begin{prob}\label{o:p3:1}
\begin{small}
\begin{subequations}
\begin{align}
Q(n,h_{n,1}) \triangleq \mathop {\min }\limits_{M(n),t} \quad & Q_{n, 1}(d_n,h_{n,1}; M(n),t) \nonumber \\
\text{s.t.} \quad & T_3(n, M(n), t) \leq T_{\text{th}}, \label{e:Q_n_T_3}\\
& M(n) \in \mathcal{I}; 0 < t \leq \tau. \label{e:Q_n_t}
\end{align}
\end{subequations}
\end{small}
\end{prob}
Then Problem \ref{s:p:multiple} is equivalent with Problem \ref{s:p:multiple_two_level}, which is given as
\begin{prob}\label{s:p:multiple_two_level}
\begin{small}
\begin{subequations}
\begin{align}
\mathop {\min }\limits_{n} \quad & \mathbb{E}_{{h}_1, {h}_2, ..., {h}_N} \left\{Q(n, h_{n,1}) + \sum\limits_{i = 1}^{n- 1} {{l_i} f_l^2 k_0} \right\} \nonumber \\
\text{s.t.} \quad & n({h}_1\rightarrow h_{n,1}) \in \mathcal{N}.
\end{align}
\end{subequations}
\end{small}
\end{prob}
Note that in Problem \ref{s:p:multiple_two_level}, the $n$ and $n({h}_1\rightarrow h_{n,1})$ are used interchangeably for the ease of presentation. In the following part of this subsection, this kind of simplification is also utilized, which will not lead to confusion.
We start by solving Problem \ref{o:p3:1}. For Problem \ref{o:p3:1}, to find its optimal solution, the following lemmas can be expected.
\begin{lem}\label{lem:OS3:1-1}
With $n$, $m$, $h$, and $d$ given, where $n\in \mathcal{N}$, $m \in \{1, 2, ..., M(n)\}$, and $d \geq 0$, there is $Q_{n,m}(d,h; M(n), t_1) \geq Q_{n,m}(d,h; M(n), t_2)$ for $0 \leq t_1 \leq t_2 \leq \tau$.
\end{lem}
\begin{IEEEproof}
Please refer to Appendix \ref{app:OS3:1-1}.
\end{IEEEproof}
\begin{lem}\label{lem:OS3:1-2}
With $n$ and $d$ given, where $n\in \mathcal{N}$ and $d\geq 0$,
for $m_1 \leq m_2$,
there is $Q_{n,1}(d,h; m_1, \tau) \geq Q_{n,1}(d,h; m_2, \tau)$.
\end{lem}
\begin{IEEEproof}
Please refer to Appendix \ref{app:OS3:1-2}.
\end{IEEEproof}
Lemma \ref{lem:OS3:1-1} implies that the function $Q_{n,m}(d,h; M(n), t)$ is a non-increasing function with $t$ for $t \in [0, \tau]$. Note that the original domain of definition for $t$ is $(0, \tau]$, while the case $t=0$ is considered in Lemma \ref{lem:OS3:1-1}. It should be noticed that $Q_{n,m}(d,h;M(n),t) = Q_{n,m}(d,h; M(n)-1, \tau)$ when $t=0$. Hence the case that $t=0$ is not weird. What is even further, the case that $t=0$ can facilitate the proof for Theorem \ref{lem:OS3:thm}.
Lemma \ref{lem:OS3:1-2} shows that the function $Q_{n,1}(d,h; m', \tau)$ is a non-increasing function with $m'$.
With the help of Lemma \ref{lem:OS3:1-1} and Lemma \ref{lem:OS3:1-2}, Theorem \ref{lem:OS3:thm} can be expected, which can help to generate the optimal solution of Problem \ref{o:p3:1}.
\begin{thm} \label{lem:OS3:thm}
To solve Problem \ref{o:p3:1}, it is optimal to select $M(n)\in \mathcal{I}$ and $t \in (0, \tau]$ such that
$T_3(n, M(n), t) = T_{\text{th}}$.
\end{thm}
\begin{IEEEproof}
Please refer to Appendix \ref{app:OS3:thm}.
\end{IEEEproof}
With Theorem \ref{lem:OS3:thm}, the optimal $M(n)$ and $t$ of Problem \ref{o:p3:1} have been found for a given $n$, which is defined as $M^*(n)$ and $t^*(n)$ respectively. In the next step, we need to solve Problem \ref{s:p:multiple_two_level}. For Problem \ref{s:p:multiple_two_level}, the method of dynamic programming can be utilized to find the optimal solution \cite{10.5555/526593}.
Specifically,
define
\begin{small}
\begin{equation} \label{e:Z_end}
Z_N(h_{N,1}) = \min\left(Q_{N,1}(d_N, h_{N,1}; M^*(N), t^*(n)), k_0 l_n f_l^2\right),
\end{equation}
\end{small}
\begin{small}
\begin{equation} \label{e:Z_n_int}
Z_n(h_{n,1}) = \min \left(Q_{n,1}(d_n, h_{n,1}; M^*(n), t^*(n)), k_0 l_n f_l^2 + Z_{n+1}\right),
\forall n \in \mathcal{N} \setminus \{N\}.
\end{equation}
\end{small}
and
\begin{small}
\begin{equation} \label{e:Z_n_iter}
Z_n = \int_{0}^{\infty} Z_n(h_{n,1}) p(h_{n,1}) d h_{n,1}, \forall n \in \mathcal{N}.
\end{equation}
\end{small}
It can be speculated that $Z_n(h_{n,1})$ is the minimal expected forthcoming energy consumption of the mobile device when the mobile device has just completed the local computing of sub-task $\phi_{n-1}$ and has measured the channel gain of $h_{n,1}$ for $n \in \mathcal{N}$.
In addition, $Z_n$ is the minimal expected forthcoming energy consumption of the mobile device when the mobile device has completed the local computing of sub-task $\phi_{n-1}$ but has no knowledge of $h_{n,1}$ for $n\in \mathcal{N}$.
For the ease of presentation, we also define $Z_{N+1}=0$. Then (\ref{e:Z_n_int}) will also be effective for $n=N$.
To achieve minimal energy consumption, the mobile device should decide whether to offload the input data of the next sub-task after it completes local computing of every sub-task, until it has made the decision to stop local computing. Suppose the mobile device has just completed the local computing of sub-task $\phi_{n-1}$ and has measured the channel gain $h_{n,1}$, then it will stop local computing and starts task offloading of sub-task $\phi_{n}$ if $Q_{n,1}(d_n, h_{n,1}; M^*(n), t^*(n)) < \left(k_0 l_n f_l^2 + Z_{n+1}\right)$ and continue local computing otherwise.
In summary, the solving procedure of Problem \ref{s:p:multiple} can be summarized in Algorithm \ref{a:3} as follows.
\begin{remark} \label{r:mark_2}
In terms of computation complexity of Algorithm \ref{a:3}, exponential complexity would be involved in the procedure of iterative computing for $Q_{n,m}(d,h; m, t)$. However, this part of computation belongs to the offline computing, which can be completed by the edge server and does not impose a computational burden on the mobile device. With a table of $Q_{n,m}(d,h; m, t)$ stored at the mobile device, it only needs to make a comparison between two values in each step, whose computation complexity is ignorable.
\end{remark}
\begin{algorithm}[H]
\caption{The procedure for solving Problem \ref{s:p:multiple}.}
\begin{algorithmic}[1]\label{a:3}
\STATE{Offline Part:}
\STATE{Utilizing (\ref{e:Q_int}), (\ref{e:Q_iterative}), (\ref{e:Q_init}) and with the aid of Theorem \ref{lem:OS3:thm} to calculate $Q_{n,m}(d_n, h_{n,1}; M(n)^*, t^*)$ for $n \in \mathcal{N}$ and $m \in \{1, 2, ..., M^*(n)\}$. Mark $Q_{n,1}(d_n, h_{n,1}; M(n)^*, t^*)$ as $Q_{n,1}(d_n, h_{n,1})$ for $n\in \mathcal{N}$. Utilizing (\ref{e:Z_end}), (\ref{e:Z_n_int}) and (\ref{e:Z_n_iter}) to calculate $Z_{n}(h_{n,1})$ and $Z_n$ for $n \in \mathcal{N}$.
Store the calculated $Z_n$, and $Z_n(h_{n,1})$, $Q_{n,m}(d_n)$, $Q_{n,m}(d_n, h_{n,1})$ and the calculated optimal $d_{n,m}$ for solving the optimization problem in (\ref{e:Q_iterative}) at the mobile device before it starts computing the whole task.
Store the calculated $Z_n$, and $Z_n(h_{n,1})$ for $n \in \mathcal{N}$ and $m\in \{1, 2, ..., M^*(n)\}$ at the mobile device.}
\STATE{Online Part:}
\WHILE{$n \leq N$ and the mobile device has not stopped local computing}
\IF{$ Q_{n,1}(d_n, h_{n,1}; M^*(n), t^*(n)) < \left(k_0 l_n f_l^2 + Z_{n+1}\right)$}
\STATE{Stop local computing and upload sub-task $\phi_n$ to the edge server. In the procedure of task offloading, adopt the stored optimal solution of $d_{n,m}$ for solving (\ref{e:Q_iterative}) to realize the expected energy consumption $Q_{n,1}(d_n, h_{n,1}; M^*(n), t^*(n))$. }
\ELSE
\STATE{Complete the computing of sub-task $\phi_{n}$ at local.}
\ENDIF
\STATE{$ n = n + 1$}
\ENDWHILE
\end{algorithmic}
\end{algorithm}
\subsection{Further Discussion for Fast Fading Channels} \label{s:further_discussion}
In fast fading channels, the $\tau$ may be very small compared with the time scale for offloading the input data of one sub-task. To get some insight into this case, we will analyze the trend when $\tau$ approaches zero in the following. Suppose the mobile device selects to offload the data for computing sub-task $\phi_n$. As $\tau$ converges to zero, $M(n)$ will converge to infinity and the $t$ can be taken as $\tau$ without loss of generality. In this case, looking into the expression of $Q_{n,1}(d, h_{n,1}; M(n), \tau)$, there is
\begin{small}
\begin{equation} \label{e:Q_n_1_extreme}
Q_{n,1}(d, h_{n,1}; M(n), \tau) =
{
\mathop{\min} \limits_{0 \leq d_{n,1} \leq d} \left(e(d_{n,1}, h_{n,1}, \tau) + Q_{n,2}(d- d_{n,1}; M(n), \tau) \right).
}
\end{equation}
\end{small}
As $\tau$ approaches zero, the optimal $d_{n,1}$ for achieving $Q_{n,1}(d, h_{n,1}; M(n), t)$, denoted as $d^*_{n,1}$, would also converge to zero. Otherwise an infinite amount of data would be offloaded within time $T_n$. Hence there is $Q_{n,2}(d- d^*_{n,1}; M(n), t)=Q_{n,2}(d; M(n), t)$.
Combing with the fact that $Q_{n,1}(d, h_{n,1}; M(n), \tau) \leq Q_{n,2}(d; M(n), \tau)$ according to (\ref{e:Q_com_m}) and $e(d^*_{n,1}, h_{n,1}, t_{n,1})\geq 0$. It can be easily derived that $e(d^*_{n,1}, h_{n,1}, t_{n,1}) = 0$.
Therefore, there is
\begin{small}
\begin{equation} \label{e:Q_equi_inf}
Q_{n,1}(d, h_{n,1}; M(n), t) = Q_{n,1}(d; M(n), t) = Q_{n,2}(d; M(n), t).
\end{equation}
\end{small}
The equation (\ref{e:Q_equi_inf}) indicates that there is no difference in the expected energy consumption for task offloading no matter how much the $h_{n,1}$ is.
In this case, we only need to evaluate the minimal expected energy consumption, i.e., $Q_{n,1}(d; M(n), t)$, rather than measure the $h_{n,1}$, before deciding to offload the sub-task $\phi_n$.
However, since $M(n)$ is approaching infinity, it is not practical to calculate $Q_{n,1}(d; M(n), t)$ by the iterative method presented in (\ref{e:Q_iterative}).
To overcome this problem, recalling the fact that there is an infinite number of independently faded blocks within time $T_n$, it is reasonable to assume that the process of $h_{n,1}, h_{n,2}, ..., h_{n, M(n)}$ are subject to an ergodic process.
Therefore, within time $T_n$, the amount of data for task offloading can be written as $\mathbb{E}\{W\log\left(1 + p_n(h)\right)\}T_n $, where $p_n(h)$ represents mobile device transmit power when instant channel state is $h$, and the expected energy consumption can be written as $\mathbb{E}\{p_n(h)\} T_n$.
Then the energy consumption minimization problem for offloading $d_n$ nats can be reformulated as follows
\begin{prob}\label{s:p:ergodic1}
\begin{small}
\begin{subequations}
\begin{align}
\mathop{\min} \limits_{\{p_n(h)| h \in \mathcal{H}\}} & \quad \mathbb{E} \left\{p_n(h) \right\} T_n \nonumber \\
\text{s.t.} & \quad
\mathbb{E} \left\{W\log \left(1 + p_n(h)\right)\right\} T_n\geq d_n; p_n(h) \geq 0, \forall h \in \mathcal{H},
\end{align}
\end{subequations}
\end{small}
\end{prob}
which is equivalent with the following optimization problem
\begin{prob}\label{s:p:ergodic2}
\begin{small}
\begin{subequations}
\begin{align}
\mathop{\min} \limits_{\{p_n(h)| h \in \mathcal{H}\}} & \quad \mathbb{E} \left\{p_n(h) \right\} \nonumber \\
\text{s.t.} & \quad
\mathbb{E} \left\{ \log \left(1 + p_n(h)\right)\right\} \geq \frac{d_n}{T_nW}, \label{s:p:ergodic2_offload_data}\\
& \quad p_n(h) \geq 0, \forall h \in \mathcal{H}.
\end{align}
\end{subequations}
\end{small}
\end{prob}
Problem \ref{s:p:ergodic2} is a convex optimization problem and falls into the category of stochastic optimization problem.
The Lagrangian of Problem \ref{s:p:ergodic2} can be written as
$\mathcal{L}\left(\{p_n(h)\}, \zeta \right) = \mathbb{E} \{p_n(h) \}- \zeta \left(\mathbb{E} \{ \log\left(1 + p_n(h) h \right)\} - \frac{d_n}{T_n W}\right)$,
where $\zeta$ is the non-negative Lagrange multiplier associated with the constraint (\ref{s:p:ergodic2_offload_data}).
Then the Lagrange dual function can be written as
\begin{small}
\begin{equation}
D(\zeta) = \mathop {\min} \limits_{\{p_n(h) \geq 0 | h \in \mathcal{H}\}} \mathcal{L}\left(\{p_n(h)\}, \zeta \right).
\end{equation}
\end{small}
Given the fact that Problem \ref{s:p:ergodic2} is a convex optimization problem and the Slater condition is satisfied \cite{boyd_vandenberghe_2004}, the maximum of the function $D(\zeta)$ over $\zeta \geq 0$ is also the minimum of Problem \ref{s:p:ergodic2}.
To find the maximum of $D(\zeta)$ over $\zeta \geq 0$, a sub-gradient method can be utilized \cite{bertsekas2003convex}, which updates the $\zeta$ in the following way until the $\zeta$ converges
\begin{small}
\begin{equation} \label{e:multiplier_update}
\zeta(i+1) = \left(\zeta(i) - a(i) \left( \frac{d_n}{T_nW} - \mathbb{E} \{\log (1 + p_n(h,i) h )\} \right)\right)^+.
\end{equation}
\end{small}
In (\ref{e:multiplier_update}), $\zeta(i)$ is the Lagrange multiplier $\zeta$ at $i$th iteration,
$p_n(h,i)$ is the optimal $p_n(h)$ to achieve $D(\zeta(i))$ at $i$th iteration for $h\in \mathcal{H}$,
$a(t)$ is the positive step size at $i$th iteration,
and $(x)^+ = \max (x, 0)$. It should be noticed that $a(t)$ should satisfy $\sum_{i=0}^\infty a(i) = \infty$ and $\sum_{i=0}^{\infty} a(t)^2 < \infty$. One candidate $a(i)$ is $a(i)=1/i$.
For a given $\zeta$, to achieve $D(\zeta)$, the optimal $p_n(h)$ of the following optimization problem for every realization of $h$ should be derived
\begin{prob} \label{p:Lagrange_one_realization}
\begin{small}
\begin{subequations}
\begin{align}
\mathop{\min} \limits_{p_n(h)} \quad & p_n(h) - \zeta \log\left( 1 + p_n(h)\right) + \zeta \frac{d_n}{T_n W} \\
\text{s.t.} \quad & p_n(h) \geq 0.
\end{align}
\end{subequations}
\end{small}
\end{prob}
For Problem \ref{p:Lagrange_one_realization}, the optimal $p_n(h)$ can be easily obtained by setting the derivative of its objective function over $p_n(h)$ to be zero when $p_n(h) \geq 0$, which can be presented as
$p_n(h) = \left(\zeta - \frac{1}{h}\right)^+$.
To this end, Problem \ref{s:p:ergodic1} has been solved and how much energy would be consumed for offloading $d_n$ nats when the length of one fading block $\tau$ goes to zero has been answered.
Note that in this case, the minimal energy consumption for offloading $d_n$ nats is a fixed number, which can be denoted as $Q(n)$, rather than a random number depending on $h_{n,1}$ anymore.
Hence the optimal $n$ can be easily found among $\{1, 2, ..., N\}$ by selecting the $n$ with the minimal associated total energy consumption, which can be written as $\sum_{i=1}^n l_i f_l^2 k_0 + Q(n)$.
\begin{remark} \label{r:mark_3}
According to the discussion in this subsection, it can be found that as $\tau$ decreases to zero, the online policy derived in Section \ref{OS3}, which depends on the instant $h_{n,1}$ in the $n$th step, trends to converge to an offline policy that can predetermine the optimal $n$ before the mobile device starts to run the whole computation task.
Compared with online policy, offline policy can save the calculation of $Q_{n,m}(d, h; M(n), t)$ for every $m\in \{1, ..., M(n)\}$ and $h\in \mathcal{H}$, which will involve a lot of computation.
Hence when $\tau$ is relatively small compared with the time scale for task offloading, offline policy discussed
in this subsection would be preferred.
\end{remark}
\section{Numerical Results} \label{s:numerical_results}
In this section, numerical results are presented. Default system parameters are set as follows.
By referring to the setup in \cite{6846368}, a computation task with 10 sub-tasks is adopted.
The input data size $d_n$ of these 10 sub-tasks are given as $\{36, 22, 30, 6, 47, 30, 5, 47, 14, 49\}$ k bits, and the workload ${l_n}$ for $n\in \mathcal{N}$ are selected to be
$\{7, 30, 25, 16, 32, 15, 37, 44, 24, 40\}$ M CPU cycles.
Similar to \cite{7442079}, $k_0$ is set as $10^{-28}$.
The channel coherence time, i.e., the length of one fading block, $\tau$ is set as 20ms.
The deadline for completing the whole computation task $T_{\text{th}}$ is selected to be 350ms.
\footnote{This time delay is necessary by evaluating the computation capability of the edge server and the workload of the investigated computation task.}.
The CPU frequency of the edge server $f_e=3$ GHz.
The maximal computation capability of the mobile device $f_{\max}=500$ MHz.
The computation capability of the mobile device in case of fast fading channels $f_l$ is also set as 500 MHz \cite{6846368}.
The bandwidth $W=1$ MHz.
In the case of fast fading channels, the channel is Rayleigh distributed, which means that the channel gain $h_{n,m}$ is subjected to an exponential distribution.
The mean of $h_{n,m}$ is set as 50.
{All the simulations are run on a desktop with the CPU of Intel i7-7700K working at a frequency of 4.2 GHz.}
\subsection{Slow Fading Channels} \label{s:num_slow}
{Fig. \ref{NR1-1} compares the mobile device's energy consumption between our proposed method in Section \ref{OS1} and the methods in \cite{6846368}
and \cite{6574874}, which represent existing strategies dealing with sequential task and binary offloading respectively, under various edge server's CPU frequency $f_e$ \footnote{{The other literature dealing with the sequential task, including \cite{6849257} and \cite{8854339}, are not compared due to similarity with the method in \cite{6846368} or model mismatch. }}.
For the method in \cite{6846368}, the transmit power, CPU frequency at the mobile device and the edge server are not optimized, so we adopt its default setting\cite{6846368}.
For the method in \cite{6574874}, we utilize its energy consumption model to generate offloading decision in the framework of binary offloading.
It can be observed that our proposed method always outperforms the methods in \cite{6846368} and \cite{6574874}.
This result can be also found in Figs. \ref{NR1-2}, \ref{NR2-1}, and \ref{NR2-2}.}
These results prove the effectiveness of our proposed method.
Additionally, it can be seen that
the energy consumption under the method in \cite{6846368} goes down like a ``staircase''.
This can be explained as follows: The transmit power and CPU frequency are fixed in \cite{6846368}. Hence the energy consumption can be only reduced when the mobile device finds a better occasion to offload (by deciding to offload since another sub-task) for the method in \cite{6846368}.
Similar trends can be also found in Fig. \ref{NR1-2} due to the same reason, which will not be discussed anymore in the part of Fig. \ref{NR1-2}.
It can be also seen that as the CPU frequency of the edge server $f_e$ grows, the associated energy consumption trends to decrease for the proposed method in Section \ref{OS1}.
This is because the increasing of $f_e$ can contribute to the enlargement of Problem \ref{s:p:constant_gain}'s feasible region.
Hence the energy consumption under our proposed method, which is also the minimum of Problem \ref{s:p:constant_gain}, would decrease.
{
Fig. \ref{NR1-2} plots mobile device's energy consumption with our proposed method in Section \ref{OS1} and the methods in \cite{6846368} {and \cite{6574874}} versus the deadline for completing the whole computation task $T_{\text{th}}$, when the channel gain $h$ is set as 40 and 60, respectively.}
From Fig. \ref{NR1-2}, we can observe that for a fixed channel gain, the mobile device's minimal energy consumption decreases as $T_{\text{th}}$ grows. This is due to the fact that the increase of $T_{\text{th}}$ can help to relax the feasible region of Problem \ref{s:p:constant_gain}. Hence the mobile device's minimal energy consumption, which is also the minimum of Problem \ref{s:p:constant_gain}, would decrease.
{
Fig. \ref{NR1-3} plots the running time when the number of sub-tasks $N$ varies. When $N>10$, we rerun the default computation task. It can be seen that the average running time to work out Problem \ref{s:p:constant_gain} is around 3.0 ms, which is ignorable compared with the scale of $T_{\text{th}}$.
This is attributed to the fact that one Golden search, which is simple to realize, is enough to find out the optimal solution of Problem \ref{s:p:constant_gain}, as shown in this paper. Thanks to the fast-growing manufacturing process of CPU in recent years, the computation capability of one mobile device nowadays does not lose to our desktop running the simulation. Hence we can claim the computation burden for working out the optimal solution of Problem \ref{s:p:constant_gain}
at the mobile device is ignorable when evaluating the time consumption for completing the computation task.
}
\begin{figure}
\centering
\subfloat[Energy consumption versus $f_e$.]{
\label{NR1-1}
\includegraphics[width=0.6\figwidth]{figures/r11-141.pdf}}
\subfloat[Energy consumption versus $T_{\text{th}}$.]{
\label{NR1-2}
\includegraphics[width=0.6\figwidth]{figures/r12-141.pdf}}
\subfloat[Running time vs. Numbers of sub-tasks $N$.]{
\label{NR1-3}
\includegraphics[width=0.63\figwidth]{figures/r13-time.pdf}}
\caption{Performance analysis under slow fading channel.}
\label{f:performance_slow_fading}
\end{figure}
\subsection{Fast Fading Channels} \label{s:num_fast}
\begin{figure}
\centering
\subfloat[Energy consumption versus $f_e$.]{
\label{NR2-1}
\includegraphics[width=0.3\textwidth]{figures/r21-141.pdf}}
\subfloat[Energy consumption versus $T_{\text{th}}$.]{
\label{NR2-2}
\includegraphics[width=0.29\textwidth]{figures/r22-141.pdf}}
\subfloat[Running time vs. Numbers of sub-tasks $N$.]{
\label{NR2-3}
\includegraphics[width=0.31\textwidth]{figures/r23-time.pdf}}
\caption{Performance analysis under fast fading channel.}
\label{f:performance_fast_fading}
\end{figure}
{
Fig. \ref{NR2-1} illustrates mobile device's expected energy consumption {for our proposed method in Section \ref{OS3} and the methods in \cite{6846368} and \cite{6574874}} when edge server's CPU frequency $f_{e}$ varies.}
In the legend of Fig. \ref{NR2-1}, ''$k=1.0$" and"$k=0.96$'' indicate the scaling factor imposed on the input data size of every sub-task.
In terms of the trends of the methods in \cite{6846368} and \cite{6574874}, the energy consumption for the method in \cite{6846368} still goes down like a ``staircase'', which can be explained as the same way for Fig. \ref{NR1-1}, while the energy consumption for the method in \cite{6574874} gradually decrease with $f_e$.
The reason behind the behavior of the method in \cite{6574874} can be explained as follows.
In \cite{6574874}, a table is also built between expected energy consumption and data amount for offloading, which represents both continuity and convexity for the expected energy consumption with respect to the data amount for offloading as we have in our proposed method. Therefore a similar trend with our proposed method can be seen for the method in \cite{6574874}. On the other hand, the table building procedure does not consider the optimization of transmit power and will surely lead to higher energy consumption compared with our proposed method. Hence the method in \cite{6574874} will always have weaker performance than our proposed method. Similar trends can be also found in Fig. \ref{NR2-2} based on the same reason, which is not further discussed in the part of Fig. \ref{NR2-2}.
It can be also observed that mobile device's expected energy consumption decreases as $f_e$ grows, which can be explained in the same way as the case in slow fading channel.
{Fig. \ref{NR2-2} plots mobile device's expected energy consumption for our proposed method in Section \ref{OS3} and the methods in \cite{6846368} and \cite{6574874} when $T_{\text{th}}$ varies. It can be observed that mobile device's expected energy consumption for our proposed method will decrease with $T_{\text{th}}$ when the mean of channel gain is set as 40 and 60, respectively.}
This can be explained as follows. Enlarged $T_{\text{th}}$ can help to relax the feasible region of
Problem \ref{s:p:multiple}. Hence the expected energy consumption, which is also the minimum of Problem \ref{s:p:multiple}, would decrease.
{
Fig. \ref{NR2-3} illustrates the average running time of online computing for solving Problem \ref{s:p:multiple} as the number of sub-tasks $N$ varies. When $N>10$, we rerun the default computation task. It can be observed that the running time is at the scale of 0.14 ms, while the offline computing time is at the scale of 1793.47 ms.
As discussed in Remark \ref{r:mark_2}, offline computing can be completed in advance by the edge server. Hence the computation burden for the mobile device, which merely comes from online computing, is ignorable (considering the rough equivalence of the computation capability between the mobile device and our simulating desktop as discussed in Section \ref{s:num_slow}) when evaluating the time consumption for solving Problem \ref{s:p:multiple}.
}
\begin{figure}
\centering
\subfloat[Verification of Lemma \ref{lem:OS3:1-1} and Lemma
\ref{lem:OS3:1-2}.]{
\label{NR3-1}
\includegraphics[width=0.3\textwidth]{figures/r31.pdf}}
\subfloat[Verification of the discussion in \ref{s:further_discussion}.]{
\label{NR3-2}
\includegraphics[width=0.3\textwidth]{figures/r32.pdf}}
\caption{Verification of lemmas and discussion.}
\label{f:verification}
\end{figure}
Fig. \ref{NR3-1} plots the function $Q_{n,1}(d,h; m, t)$ with $d = 1.2 \times 10^6$
bits when the duration of the last fading block $t$ varies from 5 ms to 20 ms.
The channel gain of the first fading block $h$ is set as 60.
It can be observed that the $Q_{n,1}(d,h; m, t)$ degrades as the duration of the last block $t$ increases when the number of fading blocks is fixed, which verifies the claim in Lemma \ref{lem:OS3:1-1}.
It can be also checked that the $Q_{n,1}(d,h; m, t)$ will decrease with the number of fading blocks when the duration of the last fading block $t$ is fixed, which verifies the claim in Lemma \ref{lem:OS3:1-2}.
{
In Fig. \ref{NR3-2}, the minimal expected energy consumption for offloading $d_n = 1.2 \times 10^6$
bits, i.e., \\
$Q_{n,1}(d_n, h_{n,1}; M(n), \tau)$, under various selection of $h_{n,1}$ is plotted versus channel coherence time $\tau$ when $n=1$. The $T_n$ is set as $840$ms.}
As a comparison, the minimal $\mathbb{E}\{p_n(h) T_n\}$ by solving Problem \ref{s:p:ergodic1} is also plotted. It can be observed that as $\tau$ decreases, $Q_{n,1}(d, h_{n,1}; M(n), \tau)$, no matter what $h_{n,1}$ is, trends to converge to the value of minimal $\mathbb{E}\{p_n(h) T_n\}$. This verifies the discussion in Section \ref{s:further_discussion}.
\section{Conclusion} \label{s:conclusion}
In this paper, we considered the MEC system with sequentially dependent computation tasks under both slow and fast fading channels. To minimize mobile device energy consumption, the optimization of task offloading strategy, communication resource allocation, and computation resource allocation are performed. For slow fading channel, with task offloading decision given, a simple and optimal solution of CPU frequency for local computing and transmit power for task offloading were found by dividing the associated optimization problem into two levels. {The Golden search method is shown to be able to find the optimal solution in the upper level and a closed-form solution is derived in the lower level.}
Then the optimal task offloading decision can be easily found by one-dimensional search. For fast fading channels, an optimal online task offloading policy in response to the instant channel state was derived. It was also disclosed that the derived online policy trends to be an offline policy as the channel coherence time approaches zero, which can help to save the computation complexity when the channel coherence time is small enough. Our research results could provide helpful insights for a MEC system
{running the applications with sequential task.}
\begin{appendices}
\section{Proof of Theorem \ref{lem:I_optimal_f}} \label{app:l_optimal_f}
Since $f_i \in [0, f_{\max}]$ for $i \in \{1, 2, ..., n-1\}$, the discussion will be unfolded from three possibles cases: 1) $f_i = 0$; 2) $0<f_i< f_{\max}$; 3) $f_i = f_{\max}$. For the first case, it is evident that the optimal $f_i$ for Problem \ref{p:I_lower_lower} cannot be zero for $i \in \{1, 2, ..., n-1\}$. Otherwise, the objective function of Problem \ref{p:I_lower_lower} would be infinite. Hence the case that $f_i=0$ for $i\in \{1, 2, ..., n-1\}$ can be precluded.
Define $\mathcal{A}_1(n-1) = \{i | 0<f_i< f_{\max}, i \in \{1, 2, ..., n-1\}\}$ and $\mathcal{A}_2(n-1) = \{i | f_i = f_{\max}, i \in \{1, 2, ..., n-1\}\}$, then there is $\mathcal{A}_1(n-1) \cup \mathcal{A}_2(n-1) = \{1, 2, ..., n-1\}$.
With regard to the second case, we can first derive that $\mu_i=\nu_i=0$ for $i \in \mathcal{A}_1(n-1)$ according to (\ref{e:KKT_slack_lower_bound}) and (\ref{e:KKT_slack_upper_bound}). Therefore, for $i\in \mathcal{A}_1(n-1)$, Eqn. (\ref{e:KKT_Lagrange}) dwells into
$2 k_0 f_i l_i = \lambda \frac{l_i}{f_i^2}, \forall i \in \mathcal{A}_1(n-1)$,
which further indicates that
\begin{small}
\begin{equation} \label{e:I_lem_f_case_II_Lagrange}
f_i = \left(\frac{\lambda}{2k_0}\right)^{\frac{1}{3}}, \forall i \in \mathcal{A}_1(n-1).
\end{equation}
\end{small}
In other words, for $i \in \mathcal{A}_1(n-1)$, all the $f_i$s are equal.
With regard to the third case, it can be also derived that $\mu_i=0$ for $i\in \mathcal{A}_2(n-1)$ according to (\ref{e:KKT_slack_lower_bound}). Then Eqn. (\ref{e:KKT_Lagrange}) turns to
$2 k_0 f_{i} l_i = \lambda \frac{l_i}{f_{i}^2} - \nu_i, \forall i \in \mathcal{A}_2(n-1)$,
which can be further written as
\begin{small}
\begin{equation} \label{e:I_lem_f_case_III_Lagrange}
\lambda = 2k_0 f_{\max}^3 + \frac{f_{\max}^2\nu_i}{l_i}, \forall i \in \mathcal{A}_2(n-1)
\end{equation}
\end{small}
since $f_i= f_{\max}$ for $i \in \mathcal{A}_2(n-1)$.
On the other hand, Eqn. (\ref{e:I_lem_f_case_II_Lagrange}) indicates that
\begin{small}
\begin{equation} \label{e:I_lem_f_case_II_Lagrange_trans}
\lambda = 2 k_0 f_i^3, \forall i \in \mathcal{A}_1(n-1).
\end{equation}
\end{small}
Combing (\ref{e:I_lem_f_case_III_Lagrange}), (\ref{e:I_lem_f_case_II_Lagrange_trans}), and the facts that $f_i < f_{\max}$ for $i \in \mathcal{A}_1(n-1)$ and $\nu_j \geq 0$ for $j \in \mathcal{A}_2(n-1)$, there is
$\lambda = 2 k_0 f_i^3 < 2 k_0 f_{\max}^3 \leq 2k_0 f_{\max}^3 + \frac{f_{\max}^2\nu_j}{l_j} = \lambda$,
which contradicts the fact the $\lambda=\lambda$.
Hence the case that $\mathcal{A}_1(n-1)\neq \emptyset$ and $\mathcal{A}_2(n-1)\neq \emptyset$ cannot coexist.
In other words, only two possible cases for the optimal solution of Problem \ref{p:I_lower_lower} could happen: 1) $\mathcal{A}_1(n-1) = \{1, 2, ..., n-1\}$ while $\mathcal{A}_2(n-1) = \emptyset $ ; 2) $\mathcal{A}_2(n-1) = \{1, 2, ..., n-1\}$ while $\mathcal{A}_1(n-1) = \emptyset $. It should be noticed that all the $f_i$s are equal in every possible case.
On the other hand, we notice that the objective function of Problem \ref{p:I_lower_lower} is monotonically increasing function with $f_i$ while the left-hand side of constraint (\ref{e:I_lower_lower_f_cons}) is decreasing function with $f_i$ for $i\in \{1, 2, ..., n-1\}$. Hence to minimize the objective function of Problem \ref{p:I_lower_lower}, it would be optimal to decrease $f_i$ as much as possible,
which will lead the activeness of constraint (\ref{e:I_lower_lower_f_cons}), i.e., the holding the following equality
\begin{small}
\begin{equation} \label{e:I_f_cons_equality}
\sum\limits_{i=1}^{n-1}{\frac{l_i}{f_i}} =
T_{\text{th}}- \tau_t - \sum\limits_{i = n}^N {\frac{l_i}{f_e}}.
\end{equation}
\end{small}
As a summary, when $\tau_t = T_{\text{th}} - \sum_{i=1}^{n-1} \frac{l_i}{f_{\max}} - \sum_{i=n}^{N}\frac{l_i}{f_e}$, (\ref{e:I_f_cons_equality}) and the fact that $f_i$ are all equal can lead to the solution $f_i = \frac{\sum_{i=1}^{n-1} l_i}{\left(T_{\text{th}} - \sum_{i=n}^{N} \frac{l_i}{f_e} - \tau_t\right)} = f_{\max}$ for $i\in \{1, 2, ..., n-1\}$, which corresponds to the case that $\mathcal{A}_1(n-1) = \{1, 2, ..., n-1\}$ and $\mathcal{A}_2(n-1) = \emptyset$; when $0 < \tau_t < T_{\text{th}} - \sum_{i=1}^{n-1} \frac{l_i}{f_{\max}} - \sum_{i=n}^{N}\frac{l_i}{f_e}$, (\ref{e:I_f_cons_equality}) and the fact that $f_i$ are all equal can lead to the solution
$f_i = \frac{\sum_{i=1}^{n-1} l_i}{\left(T_{\text{th}} - \sum_{i=n}^{N} \frac{l_i}{f_e} - \tau_t\right)}$ for $i\in \{1, 2, ..., n-1\}$, which corresponds to the case that $\mathcal{A}_2(n-1) = \{1, 2, ..., n-1\}$ and $\mathcal{A}_1(n-1) = \emptyset$.
\section{Proof of Lemma \ref{lem:OS3:1-1}} \label{app:OS3:1-1}
We complete this proof via induction method.
In the first step, the case that $m= M(n)$ is investigated.
When $m=M(n)$, there is one fading block left for task offloading. In this case, there is $Q_{n,M(n)}(d,h; M(n), t) = e(d, h, t)$.
For the function $e(d, h, t)$, the first-order derivative of $e(d, h, t)$ with $t$ can be given as
\begin{small}
\begin{equation} \label{e:diff_e_t}
\frac{d e(d, h, t)}{dt} = \frac{1}{h}\left(e^{\frac{d}{t}}-1-\frac{d}{t}\times
e^{\frac{d}{t}}\right)
\end{equation}
\end{small}
Define $x \triangleq d/t \geq 0$, the term $\left(e^{\frac{d}{t}}-1-\frac{d}{t}\times
e^{\frac{d}{t}}\right)$ in the right-hand side of (\ref{e:diff_e_t}) can be written as
$g(x) = e^x - 1 - xe^x$, whose first-order derivative $g'(x) = -xe^x \leq 0$. Then there is
$g(x) \leq g(0) = 0 $ for $x\geq 0$, which proves that $\frac{de(d,h,t)}{dt} \leq 0$.
Hence for $0 \leq t_1 \leq t_2 \leq \tau$, there would be
$e(d, h, t_1) \geq e(d, h, t_2)$.
In the second step, we need to prove the holding of $Q_{n,m}(d, h; M(n), t_1) \geq Q_{n,m}(d, h; M(n), t_2)$ given the condition that $Q_{n, m+1}(d, h; M(n), t_1) \geq Q_{n,m+1}(d, h; M(n), t_2)$ for $0 \leq t_1 \leq t_2 \leq \tau$ and any $d$ and $h$ value.
Denote $d_1^*$ and $d_2^*$ as the optimal $d_{n,m}$ to achieve $Q_{n, m}(d, h; M(n), t_1)$ and
$Q_{n, m}(d, h; M(n), t_2)$ respectively, where the definition of $d_{n,m}$ can be found in (\ref{e:Q_iterative}), then there is
\begin{small}
\begin{equation} \label{e:lem3-1-e1}
\begin{split}
Q_{n, m}(d, h; M(n), t_1)
=& e(d_1^*, h, \tau) + Q_{n, m+1}(d-d_1^*; M(n), t_1)
\geq e(d_1^*, h, \tau) +
Q_{n, m+1}(d-d_1^*; M(n), t_2) \\
\geq &e(d_2^*, h, \tau) +
Q_{n,m+1}(d-d_2^*; M(n), t_2 )
= Q_{n, m}(d, h; M(n), t_2).
\end{split}
\end{equation}
\end{small}
The first inequality of (\ref{e:lem3-1-e1}) comes from the fact that $Q_{n,m+1}(d-d_1^*; M(n), t_1) \geq Q_{n,m+1}(d - d_1^*; M(n), t_2)$ is a given condition.
The second inequality holds since $d_2^*$, rather than $d_1^*$, is the optimal solution of $d_{n,m}$ to achieve $Q_{n, m}(d, h; M(n), t_2)$.
\section{Proof of Lemma \ref{lem:OS3:1-2}} \label{app:OS3:1-2}
We first prove that $Q_{n,1}(d; m-1, \tau) \geq Q_{n,1}(d; m, \tau)$.
Suppose $d^*_m$ is the optimal solution of $d_{n,1}$ to achieve $Q_{n,1}(d, h; m, \tau)$, where the definition of $d_{n,1}$ is given in (\ref{e:Q_iterative}). Then there would be
\begin{small}
\begin{equation} \label{e:Q_com_m}
\begin{array}{ll}
Q_{n,1}(d, h; m, \tau)
& = e(d^*_m, h, \tau) + Q_{n,2}(d - d^*_m; m, \tau)
{ \leq e(0, h, \tau) + Q_{n,2}(d; m, \tau) } \\
& { = Q_{n,2}(d; m, \tau) }
{= Q_{n,1}(d; m-1, \tau)}
\end{array}
\end{equation}
\end{small}
Hence according to (\ref{e:Q_com_m}), there is \footnote{
{
The inequality in (\ref{e:Q_com_m_final}) can be also envisioned intuitively: The function $Q_{n,1}(d; m, \tau)$ actually represents the minimum expected energy consumption for offloading $d$ nats over $(m-1)$ fading blocks, which would be surely no larger than $Q_{n,1}(d; m-1, \tau)$, which represents the minimum expected energy consumption for offloading the same amount of data over a shorter time interval.
}}
\begin{small}
\begin{equation} \label{e:Q_com_m_final}
\begin{array}{ll}
Q_{n,1}(d; m, \tau)
& = \int_{0}^{\infty} Q_{n,1}(d, h_{n,1}; m, \tau) p(h_{n,1}) dh_{n,1}
\leq \int_{0}^{\infty} Q_{n,1}(d; m-1, \tau) p(h_{n,1}) dh_{n,1}
= Q_{n,1}(d; m-1, \tau).
\end{array}
\end{equation}
\end{small}
With the conclusion that $Q_{n,1}(d; m, \tau) \leq Q_{n,1}(d; m-1, \tau)$ derived in (\ref{e:Q_com_m_final}) and it can be further found that
\begin{small}
\begin{equation} \label{e:Q_com_m_h}
\begin{array}{ll}
Q_{n,1}(d, h; m, \tau)
& = e(d^*_m, h, \tau) + Q_{n,1}(d - d^*_m; m-1, \tau)
\leq e(d^*_{m-1}, h, \tau) + Q_{n,1}(d - d^*_{m-1}; m-1, \tau) \\
& \leq e(d^*_{m-1}, h, \tau) + Q_{n,1}(d - d^*_{m-1}; m-2, \tau)
= Q_{n,1}(d, h; m-1, \tau).
\end{array}
\end{equation}
\end{small}
Then with (\ref{e:Q_com_m_h}), it can be derived that for $m_1 \leq m_2$, there is
\begin{small}
\begin{equation}
\begin{array}{ll}
Q_{n,1}(d, h; m_1, \tau) & \geq Q_{n,1}(d, h; m_1+1, \tau) \geq \cdots \geq Q_{n,1}(d, h; m_2, \tau).
\end{array}
\end{equation}
\end{small}
\section{Proof of Theorem \ref{lem:OS3:thm}} \label{app:OS3:thm}
Define $M^*(n)$ and $t^*$ as the $M(n)$ and $t$ such that the constraint (\ref{e:Q_n_T_3}) is active, i.e., $T_3(n, M(n), t) = T_{\text{th}}$, respectively.
To prove this lemma, we need to prove that
$Q_{n,1}(d_n, h; M^*(n), t^*) \leq Q_{n,1}(d_n, h; M'(n), t')$
for any $M'(n) \in \mathcal{I}$ and $t' \in (0, \tau]$ such that
$T_3(n, M'(n), t') \leq T_{\text{th}}$.
Since $T_3(n, M^*(n), t^*) = T_{\text{th}}$ and $T_3(n, M'(n), t') \leq T_{\text{th}}$, there is
\begin{small}
\begin{equation} \label{e:M_compare}
\left(M^*(n) - 1\right) \tau + t^* \geq \left(M'(n) - 1\right) \tau + t' .
\end{equation}
\end{small}
Combing the fact that both $t^*$ and $t'$ falls into the interval $(0, \tau]$, then it can be concluded that
$M^*(n) \geq M'(n)$.
Then the discussion can be unfolded from the following two cases:
1). $M'(n) = M^*(n)$. In this case, there is $t^* > t'$ according to (\ref{e:M_compare}). Then according to Lemma \ref{lem:OS3:1-1}, there is $Q_{n,1}(d_n, h; M^*(n), t^*) \leq Q_{n,1}(d_n, h; M'(n), t')$.
2). $M'(n) < M^*(n)$. In this case, it is hard to say $t^*> t'$ or $t^* \leq t'$, but there is $M'(n) \leq \left(M^*(n) - 1\right)$. By utilizing Lemma \ref{lem:OS3:1-1} and Lemma \ref{lem:OS3:1-2}, there is
\begin{small}
\begin{equation}
\begin{array}{ll}
Q_{n, 1}(d_n, h; M'(n), t')
& \geq Q_{n, 1}(d_n, h; M'(n), \tau) \geq Q_{n,1}(d_n, h; M^*(n)-1, \tau) \\
& = Q_{n,1}(d_n, h; M^*(n), 0) \geq Q_{n,1}(d_n, h; M^*(n), t^*).
\end{array}
\end{equation}
\end{small}
\end{appendices}
\bibliographystyle{IEEEtran}
|
\section{\label{sec:level1}Introduction}
Ever since people realized that quantum computers are much more powerful than their classical counterparts\cite{Feynman1982,SHOR1994Algorithms}, researchers have been working hard on the problem for years, which has been making considerable progress and leads to a dramatic increase in the complexity of controllable quantum systems, from prototyped small quantum devices to machines with thousands of qubits (a typical example is the D-wave quantum annealing system \cite{Boix}). Undoubtedly, in order to reduce the impact of dissipation or decoherence that is always present in such complex systems, people continue to work on time-optimal quantum control, which is trying to minimize the necessary traversing time required to reach a quantum target (namely, a quantum state or a unitary operation) and proves to be important for building efficient gates in quantum computing architectures.
Therefore, a lot of research about the quantum speed limit (QSL) has been reported, which gives a lower bound of the time needed by a quantum system to travel from the initial to the final state and is estimated with the average energy or its variance~\cite{Campo2013,Deffner2013,Taddei2013,Mirkin2016}.
In addition, time optimal control provides a physical framework\cite{Schulte2005} for defining complexity of quantum algorithms, probably superior to traditional concepts~\cite{QuanInf}. For example Nielsen et al. proposed a new quantum computing criterion by using the Riemannian geometry, namely transforming the problem of finding an optimal quantum trajectory into a geometric one in the space of Hamiltonians\cite{Nielsen2006sci}, and revealed that the complexity of a quantum gate is related to the problem of optimal control~\cite{Nielsen2006pre}.
In a series of work about QSL~\cite{Campo2013,Deffner2013,Taddei2013,Mirkin2016}, the estimation of minimum evolution time is quite general and applicable to many different situations, in which, however the optimal evolution path is not given explicitly. In the literature, the GRAPE algorithm and the Krotov algorithm are very popular for locating optimal evolution path. In the GRAPE algorithm~\cite{Khaneja2005,Maximov2008}, the initial and the target state evolve forward and backward respectively for the same amount of time to obtain two state sequences and the control variables are updated by iteratively minimizing the difference between the two sequences. The GRAPE algorithm relieves the limitation in the number of control variables and shows high flexibility. However, the convergence is slowing down when the difference is close to some extreme values. Also, improper selection of step size often leads to a tortuous search route. In order to overcome this difficulty, the Krotov method~\cite{Maday2003,Maximov2008,Zhu1998,Shi1990} introduces a second term to penalize the undesired effects. A key feature of the algorithm is that all information is utilized at each time step which often consistently reduces the objective functional and accelerates the iteration. However, when parameters are too small, its convergence rate deteriorates, and sometimes it becomes even worse that the algorithm turns unstable.
Another way to find some time-optimal solution and its precise characterization in case of different constraints on the Hamiltonian is through invoking interesting theories such as Pontryagin maximum principle and the geometry of the unitary group~\cite{KN2,Boozer2012,Hegerfeldt2013}.
A significant piece of work was reported by Carlini~\cite{Carlini2006}, who derived a quantum brachistochrone equation (QBE) with variational method which is widely used and plays a crucial role~\cite{Meiss1992,Bunimovich1995,Lan2004,Ghoussoub2007,Dong2014,Wang2018} in classical systems. The scheme produces a Schrodinger equation together with the optimal control strategy under an energy constraint. Except for some special cases~\cite{Carlini2007,Carlini2008,Carlini2014}, analytic solutions are hard to obtain and hence numerical solution is the only resort. A scheme based on a geometric point of view is designed recently~\cite{Xiaoting2015,Xiaoting2017}, which depicts the quantum brachistochrone paths as geodesics on the constraining manifold and solves the problem by solving a family of geodesic equations.
However, there is an infinite number of geodesic families present even for locally time-optimal solutions, which makes the scheme less convenient to use. Another drawback is the complexity of the commutators introduced by different terms in the Hamiltonian to the Euler-Lagrange equation (see Eq.(2) in Ref.\cite{Xiaoting2015}). As a result, the computation load for multiple qubits could get very high and hardly be applied in practice. The problem of how to reach an effective dimension reduction and achieve fast maneuver of quantum bits is worth further exploration.
In this paper, we uncover the source of computational complexity from the perspective of a variational principle, and in the process propose an alternative set of equations with much improved efficiency, especially for multiple bit with limited control parameters. We first slightly extend the formulation in Ref.~\cite{Carlini2006} and derive all the necessary equations for later convenience. By introducing a set of new variables, the redundancy originated from the commutation relation between operators gets eliminated, which enables a very efficient computation of optimal paths in the presence of multiple qubits.
A relaxation scheme is the designed and implemented to solve the problem involving quantum entanglement when interaction between different bits comes into paly. With several continuous symmetries identified in the Lagrangian formulation, it is possible to remove the annoying neutral directions in the numerical computation. In all the examples we tried, the simplest optimal paths seem to always bear the time-reversal symmetry if the initial and target state are related by a symmetry of the Hamiltonian.
The paper is organized as follows. In Sec.~\ref{sec:1}, a Lagrangian formation of the QBE is reproduced with an introduction of uncontrollable terms (the coefficients of which are fixed values). All the necessary variational equations are derived and written in convenient vector forms. The solution of these equations is checked analytically and verified numerically in Sec.~\ref{sec:rotation}, in the absence of interaction. To deal with multiple qubits with interaction, a new set of equations are written down in Sec.~\ref{sec:numerical} by change of variables from the original set, which achieves a great reduction of complexity. The evolution of the entanglement on an optimal path is always unimodal, which is demonstrated in all the examples in Sec.~\ref{sec:example}. The results are summarized in the final section.
\section{the variational principle}
\label{sec:1}
Without loss of generality, the Hamiltonian of a quantum system that is of interest could be written as:
\begin{equation}\label{eq:H}
H=\sum_{j=1}^m \xi_j(t) A_j\,,
\end{equation}
where $\xi_j(t)$'s are real numbers and $A_j$'s are Hermitian operators that satisfy $Tr A_j=0$ and $Tr(A_jA_k)=\delta_{jk}~(j,k \leq m)$.
The traceless Hamiltonian could be divided into two parts: the controllable part $\bm{\mathcal A}=\sum\limits_{j=1}^n \xi_j(t) A_j $ with each $\xi_j (j=1,...,n)$ subject to external control and the uncontrollable part $\bm{\mathcal B}=\sum\limits_{j=n+1}^m \xi_j(t) A_j$ with $\xi_j$'s being constant or changing in a pre-determined way, often used to model interaction between qubits.
For later convenience we use $\bm u$ and $\bm v$ to label coefficients of the controllable part $\bm u=(\xi_1\,,\xi_2\,,\cdots\,,\xi_n\,,0\,,\cdots\,,0)^t$ and the uncontrollable part $\bm v=(0\,,0\,,\cdots\,,\xi_{n+1}\,,\xi_{n+2}\,,\cdots\,,\xi_m)^t$, and hence $\bm \xi=\bm u + \bm v$
. In this work, we build a formulation applicable to the general case which includes both components.
\subsection{\label{sec:level2}A variational scheme for the QBE}
The action integral is written as~\cite{Carlini2006}:
\begin{widetext}
\begin{equation}
S(\psi\,,H\,,\phi\,,\lambda,,\lambda')= \displaystyle{\int} dt \dfrac{\sqrt{ \langle \dot{\psi}|(1-P)|\dot{\psi}\rangle}}{\Delta E}+\big(i\langle \dot{\phi}|\psi\rangle+\langle\phi|H|\psi\rangle+c.c.\big)+
\lambda(\sum\limits_{j=1}^m\frac{\xi_j^2}{2}-\omega^2) +\sum\limits_{j=n+1}^m\lambda^{'}_j (\xi_j-Q_j)\
\,,\label{eq:S}
\end{equation}
\end{widetext}
where $P(t)=|\psi(t)\rangle\langle\psi(t)|$ is the projection to the state $|\psi(t)\rangle$ which is a normalized wavefunction defined in the configuration space, $\phi$ is an auxiliary wavefunction and $c.c.$ denotes complex conjugates. $(\Delta E)^2 \equiv \langle\psi|H^2|\psi\rangle-\langle\psi|H|\psi\rangle^2$ is the energy variance. Parameters $\lambda$ and $\lambda_j^{'}$ are Lagrange multipliers and all $\{\lambda'_j\}_{j=1,2,...,n}$ are set to zero, $\omega$ is a given constant, the $Q_j$'s denote the coupling constants in the Hamiltonian and the Planck’s constant $\hbar$ is chosen to be 1 for simplicity.
Now we take the variation of the action Eq.~(\ref{eq:S}) to obtain the equation of motion and all the constraints imposed to the variables.
\paragraph{}
The variation with respect to $\phi$ gives
\begin{equation}\label{Eq1}
i|\dot{\psi}\rangle=H|\psi\rangle\,,
\end{equation}
which is the usual Schrödinger equation. In the projective space $\mathbb{C}P^{n-1}$, the Fubini-Study line element is written as $ds^2=\langle d\psi|(1-P)|d\psi\rangle$, which measures the displacement in the angular direction. Here, it is easy to check that $ ds^2=(\Delta E)^2 dt^2$. Hence, the first part of Eq.~(\ref{eq:S}) refers to the total time of the process and the other parts embody differential or other constraints.
\paragraph{}
Here, we take the variation with respect to $\lambda$ and $\lambda'_j$ separately, resulting in
$\frac{1}{2}\sum\limits_{j=1}^m\xi_j^2=\omega^2$, which defines the size of the Hamiltonian as a finite energy constraint, and $\xi_j=Q_j$~($n<j \leq m$), which depicts the influence of generic interactions.
\paragraph{}
The variation with respect to $\psi$ gives
\begin{equation}\label{Eq4}
i\dfrac{d}{dt}\bigg[\dfrac{H-\langle H\rangle}{2(\Delta E)^2}\bigg]|\psi\rangle-i|\dot{\phi}\rangle+H|\phi\rangle=0 \,,
\end{equation}
where $\langle \cdot \rangle$ refers to the average with respect to $|\psi\rangle$. Eqs.~$(\ref{Eq1})$ and $(\ref{Eq4})$ are seen in the literature\cite{Carlini2006} but the equations below appear new.
\paragraph{}
The variation with respect to $H$ could now be taken as multivariate change with respect to $\{\xi_j\}_{j=1,2,...m}$, resulting in
\begin{equation}\label{Eq6}
D_j=\frac{1}{2\Delta E^2}\sum\limits_k F_{jk} \xi_k -\lambda \xi_j -\lambda^{'}_j ~~~~(\text{with}~ \lambda^{'}_j=0 ~\text{for}~ j\le n),
\end{equation}
where $F_{jk}=\langle A_jA_k+A_kA_j\rangle-2 \langle A_j\rangle\langle A_k\rangle$, $ D_j=\langle \phi|A_j|\psi\rangle +c.c.$ and $\sum\limits_{k}$ stands for $\sum_{k=1}^{m}$ which is also the case in all the following equations. The detailed derivation of Eq~(\ref{Eq6}) has been relegated to Appendix~\ref{sec:A}.
Next, we derive the equations of motion for $\lambda$ and $\{\xi_j\}_{j=1,...,n} $.
A multiplication of both sides of Eq.~($\ref{Eq6}$) with $\{\xi_j\}_{j=1,...,n} $, the controllable components, and then a summation over j results in
\begin{equation}\label{Eq7}
\lambda=\dfrac{1-\bm D\cdot \bm\xi-\sum\limits_{j=n+1}^m \bigg(\sum\limits_k \dfrac{\xi_jF_{jk}\xi_k}{2\Delta E^2} - D_j\xi_j\bigg)}{\omega'^2} \,,
\end{equation}
where $\bm D\cdot \bm\xi=\sum\limits_{j=1}^{m} D_j \xi_j$ and $\omega'^2 = \sum\limits_{j=1}^{n}\xi_j^2$ is constant (the constraint on the controllable part of the Hamiltonian). Noticing that $\{\dot \xi_j\}_{j=n+1,...m}\equiv 0 $, we could obtain:
\begin{equation}\label{Eq8}
\dot{\lambda}=\dfrac{\sum\limits_{j=n+1}^{m}\bigg\{-\dfrac{\xi_j}{2\Delta E^2}\bigg[\sum\limits_k (\dot{F_{jk}}\xi_k+F_{jk}\dot{\xi_k})
-\dfrac{\tilde{\bm F}\cdot \dot{\bm\xi}}{\Delta E^2}\tilde{F_j}\bigg]+\dot{D_j}\xi_j\bigg\}}{\omega'^2}\,,
\end{equation}
where $\tilde{\bm F}=\bm\xi \cdot F$ and $\tilde{F_j}=\sum_k F_{jk}\xi_k$. Here we used the fact that $ \dfrac{d}{dt}\bm D\cdot\bm\xi=0$ (details in Appendix~\ref{sec:C}), $\bm\xi\cdot \bm C= \bm C\cdot\bm\xi=0 $ (details in Appendix~\ref{sec:D}) and
\begin{equation}\label{Eq9}
\dfrac{d}{dt}D_j=\dfrac{1}{2\Delta E^2}(\sum\limits_k F_{jk} \dot{\xi}_k-\dfrac{\tilde{\bm F}\cdot\dot{\bm\xi}}{\Delta E^2}\tilde{F_j})+\sum\limits_{k}C_{jk}D_k\,,
\end{equation}
where $C_{jk} $ satisfies $[H,A_j]=-i\sum_k C_{jk} A_k $, with $C_{jk}$ being a real number. For a detailed derivation of Eq.~($\ref{Eq9}$), please check Appendix B. Feeding Eq.~(\ref{Eq9}) into Eq.~($\ref{Eq8}$) results in
\begin{align}\label{Eq10}
\dot{\lambda}= & \dfrac{\sum\limits_{j=n+1}^{m}\sum\limits_{k=1}^{m} \xi_jC_{jk}(D_k-\frac{\tilde{F}_k}{2\Delta E^2})}{\omega'^2}\nonumber \\
=& -\dfrac{ \bm{v\cdot C \cdot \lambda}' }{\omega'^2} \,,
\end{align}
with $\dot{\bm F}=\bm C\cdot \bm F +\bm F\cdot \bm C^{T}$ (Appendix \ref{sec:E}). Taking the time derivative of both sides of Eq.~(\ref{Eq6}), and together with Eq.~(\ref{Eq9}), we get
\begin{equation}\label{Eq11}
\dot{\bm u}=- \dfrac{\bm C\cdot \bm D}{\lambda}+\dfrac{\bm C\cdot F\cdot \bm\xi
}{2\lambda\Delta E^2}+\dfrac{(\bm{v\cdot C \cdot \lambda}')}{\lambda\omega'^2}\bm u\,.
\end{equation}
With Eq.~(\ref{Eq9}), Eq.~(\ref{Eq10}), Eq.~(\ref{Eq11}) and the Schrödinger equation Eq.~(\ref{Eq1}), the optimal path can be calculated now.
Here let's discuss a special case with $H\in\mathcal{A} $, namely $\bm \xi=\bm u$, which means that all parts are controllable and thus $\bm v =0$.
It is obvious that this case is simple since
\begin{align}
D_j&=\frac{1}{2\Delta E^2}\sum\limits_k F_{jk} u_k -\lambda u_j, \label{Eq12}\\
\dfrac{d}{dt}D_j&=\dfrac{1}{2\Delta E^2}(\sum\limits_k F_{jk} \dot{u}_k-\dfrac{\tilde{\bm F}\cdot\dot{\bm u}}{\Delta E^2}\tilde{F_j})+\sum\limits_{k}C_{jk}D_k .\label{Eq13}
\end{align}
Hence Eq.(\ref{Eq10}) implies that $\dot\lambda=0 $. Eq.~(\ref{Eq11}) could be reduced to
\begin{equation}\label{Eq14}
\dot{\bm u}=- \dfrac{\bm C\cdot \bm D}{\lambda}+\dfrac{\bm C\cdot \bm F\cdot \bm u
}{2\lambda\Delta E^2}.
\end{equation}
Considering $\bm u\cdot \bm C=0$ (Appendix~\ref{sec:D}) and feeding Eq.~(\ref{Eq12}) into Eq.~(\ref{Eq14}) leads to a trivial solution $\dot {\bm u}=0 $, which indicates that $\bm u$ is a constant vector and the evolution corresponds to a simple rotation with a fixed axis on the Bloch sphere.
\subsection{Gauge symmetries in the Lagrangian}
\label{sec:symm}
In this section, we discuss the gauge symmetries in the Lagrangian Eq.~(\ref{eq:S}), which create neutral directions in the evolution and may bring complication in numerical calculation. First, consider a transform $\hat{U}_1: \phi \to \phi+ia\psi$, $a\in \mathcal{R}$, with which the Lagrangian becomes
\begin{equation}\label{Eq15}
S'=S+\int dt ~(a\langle \dot{\psi}|\psi\rangle-ia\langle\psi|H|\psi\rangle+c.c.)
\,,
\end{equation}
The expression $\langle\dot{\psi}|\psi\rangle+c.c$ is a total differentiation, which could be integrated out. With the Hermiticity of $H$, the extra integral in Eq.~(\ref{Eq15}) vanishes and hence $\hat{U}_1$ is a symmetry transform. Next, we consider another transform $\hat{U} _2:\phi\rightarrow\dfrac{\lambda+b}{\lambda}\phi,~\lambda\rightarrow\lambda+b,~\bm\lambda'\rightarrow\dfrac{\lambda+b}{\lambda}\bm\lambda',~ b\in \mathcal{R}$ which sends Eq.~($\ref{eq:S}$) to
\begin{align}\label{Eq16}
S(\psi,H,\phi,\lambda,\lambda')&=\int dt \dfrac{\sqrt{\langle \dot{\psi}|(1-P)|\dot{\psi}\rangle}}{\Delta E}\nonumber\\
&+\dfrac{\lambda+b}{\lambda}\left[ (i\langle \dot{\phi}|\psi\rangle+\langle\phi|H|\psi\rangle+c.c.)
+\lambda(\sum_{j=1}^{m}\frac{\xi_j^2}{2}-\omega^2)+\sum_{j=n+1}^{m} \lambda'_j (\xi_j-Q_j) \right]
\,,
\end{align}
which only changes the values of the invisible Lagrange multipliers $\phi,~\lambda,~\bm\lambda'$ while the equation for the observables $\psi,~\bm u$ remain intact. Actually, it is most obvious in the simplified formulation Eq.~(\ref{eq:evolution2}) below, where $\bm\Omega,~\bm{u},~\dot{\bm\Omega},~\dot{\bm u}$ keep invariant under the transform $\hat{U}_2$. Hence $\hat{U}_2$ is a symmetry transform.
\section{The evolution of Hamiltonian as a rotation }
\label{sec:rotation}
The equation $\bm\xi\cdot\bm C =0 $ tell us that $\bm v\cdot\bm C =-\bm u \cdot \bm C $. Substituting the Eq.~(\ref{Eq6}) into Eq.~(\ref{Eq11}) results in
\begin{equation}
\dot{\bm u}=(I-P_{u})\cdot\dfrac{\bm C\cdot\bm\lambda'}{\lambda}
\,,\label{eq:dua}
\end{equation}
where $I$ is the identity matrix and $P_{u}$ is the projection operator $\dfrac{|\bm{u}\rangle\langle \bm u|}{|\bm u|^2}$ ($\omega'^2 = \sum\limits_{j=1}^{n}\xi_j^2=\sum\limits_{j=1}^{n}u_j^2=|\bm u|^2$). Here we use the Dirac notation to emphasize the projection, whereas $\bm u$ is actually real. With the fact that $\bm{\dot{v}}=0$, it is easy to check that $\bm\xi\cdot \bm{\dot{\xi}}=\bm u\cdot \dot{\bm u}=0$.
And the uncontrollable part of the time derivative of Eq.~(\ref{Eq6}) similarly gives the evolution of $\bm\lambda'$
\begin{equation}
\dot{\bm\lambda}'=
\bm C\cdot\bm\lambda'
+\dfrac{\bm v\cdot \bm C\cdot\bm\lambda'}{\omega'^2}\bm v=(I+\dfrac{|\bm v|^2}{|\bm u|^2}P_{\bm v})\cdot \bm C \cdot\bm\lambda'
.\label{eq:dlambda'}
\end{equation}
where $P_{v}=\dfrac{|\bm v\rangle\langle \bm v|}{|\bm v|^2}$ and $\dfrac{|\bm v|^2}{|\bm u|^2}$ is a real constant.
After,
defining $\bm{\Omega}=-\dfrac{2}{\sqrt{Tr I}}\dfrac{\bm{\lambda'}}{\lambda}$, the angular velocity of the rotation of $\bm{u}$ can be described with $\bm\Omega$. Furthermore, the QBE problem can be characterized with a simpler set of equations
\begin{equation}
\begin{aligned}
\dot{\bm u}&=-\dfrac{\sqrt{Tr I}}{2}(I-P_{u})\cdot \bm C\cdot\bm\Omega \\
\dot{\bm \Omega}&=(I+\dfrac{|\bm v|^2}{|\bm u|^2}P_{\bm v})\cdot \bm C\cdot\bm\Omega-\dfrac{\sqrt{Tr I}}{2\omega'^2}(\bm v\cdot \bm C\cdot\bm\Omega)\bm\Omega
\end{aligned}
\label{eq:evolution2}
\end{equation}
along with the Schrödinger equation $\dot{\psi}=-iH\psi$.
According to Appendix \ref{sec:D}, $\bm C$ can be conveniently determined, which in specific circumstances may provide possible routes to analytic solutions. As a specifical case, in the case without entanglement, the phase space can be divided into separate Bloch spheres of qubits.
Then, according to Eq.~(\ref{eq:evolution2}), the equation of a single qubit in 3-dimensional space is appreciably simplified to $\dot{\bm u}=\bm \Omega\times \bm\xi-\hat{\bm u}\cdot(\bm\Omega\times \bm\xi)~\hat{\bm u}=\bm\Omega\times \bm u$. Meanwhile, the uncontrollable part is generally limited to no more than one dimension per qubit, or there will be no solution for most boundary conditions. As a result, $\bm\Omega$ is in the direction of $\bm\lambda'$ which is along $\bm v$ in the case of three dimensions. Here the cross product holds for each $SO(3)$ Bloch sphere, which indicates that $\bm u$ is rotating with the angular velocity $\bm\Omega$.
As $\bm\Omega, \bm v$ is in the same direction for a single qubit. After multiplying $\bm\Omega$ with Eq.~(\ref{eq:evolution2}) we see that $|\bm\Omega|$, the rotation speed of $\bm u$, is a constant vector. In fact, in this case $\bm\lambda'=0$ and thus $\bm\Omega=0$, $\dot{\bm u}=0$ as given in the previous section.
It is easy to see that the spin vector $\langle\bm\sigma\rangle$ ($\sigma_{j=\{x,y,z\}}$ are the Pauli matrices) is rotating around $\bm u$, and the evolution of wave function is analytically computed with rotation matrices.
\subsection{Analytical solution of a single qubit}\label{sec:1bit_ana}
The evolution path of $\psi$ can be analytically addressed, as a rotation with the angular velocity $\bm\omega_\mathrm{eff}$ around an axis $\bm u$ rotating itself at $\bm\Omega$.
The effective angular velocity $\bm\omega_\mathrm{eff}$ of spin vector is determined with
\begin{equation}
\bm\omega_\mathrm{eff}=2\bm B=\dfrac{2}{\sqrt{TrI}}\bm u=\sqrt{2}\bm u
\,.
\end{equation}
where $B$ is an effective magnetic field and the Hamiltonian $H=\bm\sigma\cdot\bm B=\sum\limits_j A_ju_j$. This is derived from the fact that the SU(2) generator of a rotation with $\Delta\phi$ is $e^{i\bm\sigma\frac{\Delta\phi}{2}}$. In the frame of reference rotating with $\bm\Omega$, the motion of $\psi$ is viewed as a fixed-axis rotation with the angular velocity $\bm\Omega'\!=\bm\omega_\mathrm{eff}\!-\bm\Omega$.
For instance, consider the numerically simulated case~\cite{Carlini2006} where $\psi(0)=|+x\rangle$, $\psi_{target}=|-x\rangle$, $\bm B(0)$ is $\omega$ in the $-y$ direction and $\bm\Omega$ lies in the $z$ direction with $Tr(H\sigma_z)=0$. Explicitly, $\langle\bm\sigma\rangle(0)=(1,0,0),
\bm B(0)=(0,-\omega,0),
\bm\Omega\equiv(0,0,\Omega)$.
Now the evolution turns out to be simply a rotation around $\bm\Omega'$ followed by a reversal rotation around $\bm\Omega$ back to the static frame. The rotation matrix $R(\phi,\theta,\Psi)=R_{\bm\Omega}^{-1}(\phi)R_{\bm\Omega'}(\theta)=R_z^{-1}(\phi)R_x^{-1}(\Psi)R_y(\theta)R_x(\Psi)$ (see Appendix~$\ref{sec:G}$ ). Here $(\phi,\theta,\Psi)$ is valued $(\Omega t,\Omega' t,\mathrm{tan^{-1}}\dfrac{\Omega}{\omega_\mathrm{eff}})$, where $\Omega'=\sqrt{\Omega^2+\omega_\mathrm{eff}^2}$. Then it is easy to check that
\begin{equation}
\langle\bm\sigma\rangle(t)=R(\phi,\theta,\Psi)\cdot \langle\bm\sigma\rangle(0)=
\begin{pmatrix}
&~\mathrm{cos}\!~\Omega t\!~\mathrm{cos}\!~\Omega't+\dfrac{\Omega}{\Omega'}~\mathrm{sin}\!~\Omega t\!~\mathrm{sin}\!~\Omega't\\
&-\mathrm{sin}\!~\Omega t\!~\mathrm{cos}\!~\Omega't+\dfrac{\Omega}{\Omega'}~\mathrm{cos}\!~\Omega t\!~\mathrm{sin}\!~\Omega't\\
&\dfrac{\omega_\mathrm{eff}}{\Omega'}\mathrm{sin}\!~\Omega't
\end{pmatrix}
\,,
\end{equation}
which gives a family of analytic solutions if we choose $(\Omega T, \Omega' T, \dfrac{\Omega}{\omega_{\mathbf{eff}}})=(k\pi, l\pi, \dfrac{k}{\sqrt{l^2-k^2}})$, where $l, k$ are non-negative integers and $l+k$ is odd. The case $k=0$ corresponds to the geodesic solution with no anisotropic contraints and cases of $k\not=0$ correspond to the oscillating solutions with $l-1$ nodes, and setting L= $l-1$ for later convenience. The oscillating curve in FIG.~\ref{Fig2} plots the solution with $k=1,L=1$.
\subsection{Analytical solution for multiple qubits}
First we give an example of two qubits with $H=B_x^{(1)}\sigma_x^{(1)}+B_y^{(1)}\sigma_y^{(1)}+B_x^{(2)}\sigma_x^{(2)}+B_y^{(2)}\sigma_y^{(2)}$.
Similarly, our purpose is to flip both qubits (where $\psi(0)^{(1)}=|+x\rangle$ and $\psi_{target}^{(1)}=|-x\rangle$ for qubit one, $\psi(0)^{(2)}=|+y\rangle$ and $\psi_{target}^{(2)}=|-y\rangle$ for qubit two). Obviously, these two qubits are both reversing their states, and because $x-$ and $y-$ directions are totally equivalent there qubits should share the same energy at the ground state(namely $\omega_{\mathbf{eff}}^{(1)}=\omega_{\mathbf{eff}}^{(2)}$). Fig.~\ref{Fig3} shows a local optimal solution of the constrained two-qubits system with $L=1$. $\bm\Omega^{(1)}=(0,0,-\dfrac{2}{\sqrt 3}\omega),~\bm\Omega^{(2)}=(0,-\dfrac{2}{\sqrt 3}\omega,0)$. It is appearent that the evolution of $\bm B$ turns out to be a rotation as expected.
Actually, the local solutions can be derived analytically as long as the entangled terms are missing in the Hamiltonian. According to the rotation picture, the analytical approach can be applied to non-entangled many-qubit systems with an arbitrarily large number of qubits. Here in Fig.~\ref{Fig3} the solution corresponds to $(k^{(1)},L^{(1)})=(k^{(2)},L^{(2)})=(1,1)$ as the oscillating solution in Fig.~\ref{Fig2}, so they naturally share the same geometric property.
Later on, these solutions would be good initial guesses in the case of weak entanglement (see \ref{sec:relaxation}).
\section{An Accelerated Approach to the QBE}
\label{sec:numerical}
In practice, there are usually a large number of uncontrollable terms in the Hamiltonian. Sometimes for the convenience of computation and design, almost all the interaction terms are made to vanish. At this moment the dimension of control parameters is appreciably low, however, the computation complexity of the conventional numerical methods remains as high. To address and eliminate the redundant complexity, here we propose an alternative approach, reintroducing the Lagrange multiplier $\phi$ previously replaced by $\bm D$.
In general situations we see that the tremendous complexity is mainly brought by the factor $\bm C\cdot \bm D$, since the commutation relation~(\ref{b6}) will incur a large number of or infinitely many more $A_k$'s and thus $D_k$'s in the QBE equations when there are multiple interacting qubits. It is frustrating that most of the previous researches~\cite{Carlini2006,Xiaoting2015} are confronted with similar problems but no special attention has been paid to this problem. Here, we introduce an effective deduction scheme based on the fact that $\bm C\cdot\bm D$ is immediately derivable from $\phi$. Technically, consider
\begin{equation}\label{phi2}
\phi'=\phi-\dfrac{1}{2\Delta E^2}(H-\langle H\rangle)\psi,
\end{equation}
then Eq.~(\ref{Eq6}) turns to
\begin{equation}\label{eq:D2}
D'_j=\langle\phi'|A_j|\psi\rangle+c.c.=
-\lambda \xi_j-\lambda'_j, ~~~~(\text{with}~\lambda'_j=0~\text{for}~j\le n).
\end{equation}
Thus ${\bm C\cdot D'_j}={\bm C\cdot\langle\phi'|A_j|\psi\rangle}+c.c.$.
Furthermore, it is much easier to evolve $\phi'$ instead of $\phi$ since Eq.~(\ref{Eq4}) yields
\begin{equation}\label{eq:dphi}
i\dot{\phi}-H\phi=\frac{i}{2\Delta E^2}[(\dot{H}-\langle\dot{H}\rangle)\psi-\dfrac{\tilde{\bm F}\cdot\dot{\bm\xi}}{\Delta E^2}(H-\langle H\rangle)\psi]\\ \,,
\end{equation}
which gives
\begin{equation}\label{eq:dphi'}
\begin{aligned}
i\dot{\phi}&'=H\phi'\,.\\
\end{aligned}
\end{equation}
In analogy to previous derivations in Sec.~\ref{sec:rotation}, it is easy to check that Eq.~(\ref{eq:evolution2}) now holds in an alternative form
\begin{equation}\label{eq:evo2}
\begin{aligned}
\dot{\bm u}&=(I-P_{u})\cdot\dfrac{\bm C\cdot\bm\lambda'}{\lambda},\\
i\dot{\phi}&'=H\phi',\\
\dot{\lambda}&=-\dfrac{ \bm{v\cdot C \cdot \lambda}' }{2\omega^2}, \\
\dot{\psi}&=-iH\psi \,,\\
\end{aligned}
\end{equation}
where from Eq.~(\ref{eq:D2}), $\bm C\cdot\bm\lambda'= -\bm C\cdot\bm D'$ with $\sum\limits_{j}C_{ij}D'_j={\langle\phi'|[H,A_i]|\psi\rangle}+c.c.$, which could be evaluated directly with the computed states $\phi^{\prime}$ and $\psi$ and the known $H$ and $A_i$'s.
In conclusion, the whole QBE problem is fully described by Eq.~(\ref{eq:evo2}). Terms like $\bm C\cdot\bm F\cdot\bm\xi/2\Delta E^2$ in $\bm C\cdot\bm D$ are eliminated now because of the reintroduction of the state variable $\phi'$, and hence the previous large number of equations needed to evaluate $\bm C\cdot\bm D$ are simply replaced by the evolution of $\phi'$. Since the number of degrees of freedom (DOF) of vector $\phi’$ is much smaller than that of matrix $\bm C$, the computation will be significantly accelerated.
Based on Eq.~(\ref{Eq6}),~(\ref{eq:D2}) and the fact that $Tr H=0$ (i.e.~$\xi_0=0$), we immediately obtain
\begin{equation}\label{Eq17}
\begin{aligned}
2Re(\langle\phi|\psi\rangle) &=D_0=\sum_j\dfrac{F_{0j}}{2\Delta E^2}\xi_j-\lambda\xi_0\equiv 0\,,\\
Re(\langle\phi'|\psi\rangle) &=D'_0=-\lambda\xi_0\equiv0\,,
\end{aligned}
\end{equation}
where $F_{0j}=0$ and $A_0=\sigma_0$, which is the unit matrix.
Hence the component of $\phi$ or $\phi'$, being parallel to $\psi$, vanishes. Notice that $\bm u$, in the absence of uncontrollable components, is parallel to $Re(\langle\phi'|\bm\sigma|\psi\rangle)$ according to Eq.~(\ref{eq:D2}). So $\bm u\cdot\langle\bm\sigma\rangle \propto Re(\langle\phi'|\bm\sigma|\psi\rangle)\cdot\langle\bm\sigma\rangle=0$ (proved in Appendix~\ref{sec:F}) for every qubit.
Now a two-point boundary value problem with Eq.(\ref{eq:evo2}) is left for us to solve. To begin with, we count the number of independent variables. The number of different variables in Eq.~(\ref{eq:evo2}) is $N=N_u+N_{\phi'}+N_\lambda+N_\psi+N_{\lambda'}$, while the number of the constraints is $N'=N_u+N_{\lambda'} + 2N_{\psi}+1$ given by Eq.(\ref{eq:D2}) (giving $N_u+N_{\lambda'}$), the energy constraint $\sum\limits_{j=1}^n\xi_j^2=\omega'^2$, the prescrible initial and final states $\psi(0)=\psi_0$ and $\psi(T)=\psi_{target}$ (giving $2N_{\psi}$), then the number of independent variables is $N-N'=0$ (where $N_\lambda=1$ and $N_{\phi'}=N_{\psi}$).
In the current paper the \emph{shooting method} (\ref{sec:shoot}) and the idea of \emph{relaxation} play a major role in finding the optimal path when Hamiltonian $H$ contains interaction terms. We give some details below.
\subsection{Shooting Method}
\label{sec:shoot}
The shooting method aims to implement the multidimensional Newton-Raphson method to locate zeros of $N_{target}$ functions
with $N_{target}$ varibales, which are obtained by integrating $N$ differential equations in our examples.
For a general dynamical system, the ordinary differential equation (ODEs) is written as
\begin{equation}\label{v1}
\dot{\bm x}=\bm v(\bm x(t)),
\end{equation}
where $\bm x\in \mathbb R^d$, $t \in \mathbb R$, and the state of the system obtained by integrating Eq.(\ref{v1}) at time $t$ is $\bm x(t)= f^t(\bm x)$. The corresponding Jacobian matrix is $J(\bm x,t)=\frac{\partial\bm x(t)}{\partial\bm x(0)}$, which could be obtained by integrating
\begin{equation}\label{J}
\frac{dJ}{dt}=AJ,~~~A_{ij}=\frac{\partial{\bm{v}_i}}{\partial{\bm{x}_j}}, ~~~\text{with }J(\bm x,0)=\bm 1.
\end{equation}
Thus, the initial displacement $\delta \bm x(0)$ becomes $\delta \bm x(T)$ at time $T$, namely
\begin{equation}
\delta \bm x(T)=J(\bm x,T) \cdot \delta \bm x(0)
\,.\end{equation}
Define a discrepancy vector $\bm X=\bm x(T)-\bm x_{target}$. To procure a vanishing discrepancy, the correction $\Delta x_0$ should satisfy
\begin{equation*}
J\cdot \Delta\bm x_0=-\bm X,
\end{equation*}
to reach a better approximation
\begin{equation}\label{dx}
\bm x_{0,new}=\bm x_{0,old}+\Delta \bm x_{0}
\,.\end{equation}
Typically, the Jacobian matrix $J$ is numerically evaluated by
\begin{equation}
J_{ij}=\dfrac{\bm X_i(\bm x_{0,1},\ldots,\bm x_{0,j}+\Delta \bm x_{0,j},\ldots,T)-\bm X_i(\bm x_{0,1},\ldots,\bm x_{0,j},\ldots,T)}{\Delta\bm x_{0,j}}\,.
\end{equation}
In the current case, the adjustable part is the free parameter $\phi$ and the discrepancy is $\psi (T)-\psi_{target}$.
Let's slow down here and have a discussion on possible additional constraints and the treatment of the symmetries mentioned earlier. Firstly, based on Eq.(\ref{Eq17}), $Re(\langle\delta\phi'|\psi\rangle)=0$,
which with the help of the symmetry transformation $\hat{U}_1$ could always be satisfied by taking the difference $\phi^{\prime}_{0,new}-ia\psi$, where $ia\psi=|\psi\rangle\langle\psi|\phi'_{0,new}\rangle$. The unitary evolution of $\psi$ will produce a singularity for the Jacobian matrix $J$. To cope with this problem, our strategy is to totally ignore this eigen-direction since it corresponds to irrelevant symmetry directions. The constraint $\sum_{j=1}^{n}\xi_j^2=\omega'^2$ could be garanteed by the symmetry $\hat{U}_2$ since $\lambda_0$ and thus $b$ could be obtained by substituting $\phi^{\prime}_{0,new}$ into Eq.~(\ref{eq:D2}) with $j \le n$ when starting a new iteration.
It is likely that the discrepancy may significantly fluctuate in the first few iterations. So it is important to properly choose the target hyperplane for the shooting on which the target state lies.
The final state $\psi(T)$ is adjusted to 'hit' the target state $\psi_{target}$ by adjusting the initial position Eq.(\ref{dx}) and the flight time $T$.
\subsection{Multiple Qubits: `Relaxation'}
\label{sec:relaxation}
When it comes to large-scale qubit systems, the shooting method per se can be too costly and need a good trial solution that is sufficiently close to the exact solution. In order to overcome those shortcomings we suggest a `relaxation' method to meet the boundary conditions step by step. First one stars from the case without interaction, solving the problem with separate qubits.
With interaction present, it is possible to solve the problem progressively. First, we invite part of the interactions back and use the non-interacting solution as the initial condition ({\em i.e.},taking the first-step coupling constant $J=0.01$ in our examples), so a new solution in case of the weak interaction is obtained, which can be used as the initial condition for the case with a little bit stronger interaction. Repeat this process until the full interaction is restored and an optimal path is then obtained. This is a relaxation process since the effective Hamiltonian in the computation is gradually relaxed to the desired one.
\begin{figure}[b]
\includegraphics[width=7cm]{Fig1}
\caption{\label{Fig1} Evolution of Qubit 1 on a Bloch sphere. The black line represents the state $\psi(t)$ that evolves from the initial state (purple point on the left) to the final state (blue points on the right) on the Bloch sphere. The green grid inside the sphere marks the equatorial plane, and the red part is the area swept by the magnetic field vector. Similar representations are used in later diagrams.}
\includegraphics[width=7cm]{Fig2a
\includegraphics[width=7cm]{Fig2b}
\caption{\label{Fig2} The evolution of $\psi(t)$ of a single qubit with $\omega_{eff}=2|\bm B|=2 $ and L=1 (left), 2 (right) (namely, the black path intersects the equatorial plane L times apart from the initial and final states).}
\end{figure}
\section{examples}
\label{sec:example}
\subsection{One-Qubit Bloch Sphere}
\label{sec:1bit}
So far most of the practical quantum control systems are based on spin magnetic moment. Here we first try to search for local optimal paths of single-qubit evolution on the Bloch sphere. The goal is to flip the spin state of a single Fermion from $|+x\rangle$ to $|-x\rangle$.
Assuming $\hbar=1$, the Hamiltonian in the isotropic condition is
\begin{equation*}
H=-\bm\mu_S\cdot\bm B=-\dfrac{\gamma}{2}\bm\sigma\cdot\bm B=-\dfrac{\gamma}{2}(B_x\sigma_x+B_y\sigma_y+B_z\sigma_z)
\,,\end{equation*}
where $\gamma=g\dfrac{q}{2m}$ is the gyromagnetic ratio. For electrons $\gamma\approx -2$ then simply $H=\bm B\cdot\bm\sigma$.
From previous discussions, we know that $\dot{\bm u}=0$, i.e. $\dot{\bm B}=0$. So the optimal path is naturally a geodesic curve (see Fig.~\ref{Fig1}, which shows one of the many geodesic solutions). The simulated paths are displaced as well in Fig.~\ref{Fig1}. The red line on the equatorial plane (green grid inside the sphere) indicates the evolution of the external magnetic field $\bm B$, which does not change throughout the evolution in the case.
In principle, the constraints $ Tr(H\sigma_j)_{j=x,y,z}=0$ allow oscillating solutions with $\dot {\bm B}\not=0$.
Let's take the case of $H=B_x\sigma_x+B_y\sigma_y, ~B_z\equiv 0$ as an example.
In numerical computation, we use the shooting method to identify the initial value $\phi'_0$ for the prescribed boundary condition $\psi(0)=|+x\rangle,~\psi(T)=|-x\rangle$.
Fig.~\ref{Fig2} shows the resulting optimal path which oscillates with one node on the equator, where $\bm\Omega=(0,0,-\dfrac{2}{\sqrt{3}})$. The evolution of $\bm B$ (the area swept by the red line in Fig.~\ref{Fig2}) corroborates the theory of rotation in Sec.~\ref{sec:rotation}. Of course, it is not hard to get a longer locally optimal path with more than one node, which we will not show here.
\subsection{Two-Qubit Bloch Sphere}\label{sec:2bitw}
\begin{figure*}[]
\includegraphics[width=16cm]{Fig3
\caption{\label{Fig3} An optimal path with $L=1$, $\omega^{(1)}_{eff}=2$, $\omega^{(2)}_{eff}=2$ connecting the initial state $|+x\rangle^{(1)},|+y\rangle^{(2)}$ to the final state $|-x\rangle^{(1)},|-y\rangle^{(2)}$ of Qubits 1 and 2 plotted on two bloch spheres.}
\end{figure*}
In the case of two qubits without interaction, two bloch spheres may be used to represent a non-entangled state and the geodesic solution is simple according to the discussion in Secion~\ref{sec:1bit} above. Here, we focus on the oscillating solutions under the anisotropic constraint featuring $ B_z=0$. We give an example with
\begin{equation}
H=B_x^{(1)}\sigma_x^{(1)}+B_y^{(1)}\sigma_y^{(1)}+B_x^{(2)}\sigma_x^{(2)}+B_y^{(2)}\sigma_y^{(2)}.
\end{equation}
Fig.~\ref{Fig3} depicts the local optimal path obtained by our method, which oscillates with $L=1$. Also, we can obtain longer paths, not shown here.
\begin{figure*}[]
\includegraphics[width=4.3cm]{Fig4a}
\includegraphics[width=4.3cm]{Fig4b}
\includegraphics[width=4.3cm]{Fig4c}
\includegraphics[width=4.3cm]{Fig4d}
\caption{\label{Fig4} The time evolution of the entanglement Ep of $\psi(t)$ for different $J_{yy}$ and L=1(a), 3(b), 2(c), 4(d). t/T represents the normalized time where T is the total time. The initial state are $|+x\rangle^{(1)},|+y\rangle^{(2)}$ and the final state are $|-x\rangle^{(1)},|-y\rangle^{(2)}$ for Qubits 1 and 2}
\end{figure*}
\begin{figure}[]
\includegraphics[width=7cm]{Fig11}
\caption{\label{Fig11} The evolution of the entanglement $E_p(t)$ of wavefunction $\psi(t)$ from the initial state $\psi(0)=(0.5, 0.5i, 0.5, 0.5i)^T $ to the final state(-0.216558-0.450669i, 0.528077+0.256587i, 0.070204+0.391403i, -0.294949+0.400223i) with J=0.65, 0.9, 1 and L=1.
}
\end{figure}
\begin{figure*}[]
\includegraphics[width=6cm]{Fig12a}
\includegraphics[width=6cm]{Fig12b}
\includegraphics[width=6cm]{Fig12c}
\includegraphics[width=6cm]{Fig12d}
\caption{\label{Fig12} The evolution and the convergence. The evolution of the observed measurement $\langle\pmb{\sigma}(t)\rangle$ (a) and the control magnetic field $\bm{B}$(t) (b), where $\langle\sigma(t)_m^{(l)}\rangle=\langle\psi(t)|\sigma(t)_m^{(l)}|\psi(t)\rangle~(m=x, y, z; l = 1, 2)$ represents the observed value of the spin in the m-direction on the $l~th$ bit. The convergence of our method (c) and the GRAPE method (d), where Error $d_1=|\psi(T)-\psi_{target}|$ and $d_2=|1-\langle\psi_{target}|\psi(T)\rangle|$, N is the number of iteration steps and J=1, L=1.}
\end{figure*}
\begin{figure}[]
\includegraphics[width=6cm]{Fig13}
\caption{\label{Fig13} The Fubini-Study distance F(t) between $\psi(t)$ and $\psi_{target}$ with $\omega =2.8284,~2.7803,~2.6926,~2.5554,~2.3601$. The distance between neighboring states along a trajectory is defined by using the Fubini-Study line element $ds=\sqrt{\langle d\psi|(1-P)|d\psi\rangle}$, where $d\psi$ is the change of the state and $P=|\psi><\psi|$ is a projection operator
}
\end{figure}
\begin{figure*}[]
\includegraphics[width=6cm]{Fig14a}
\includegraphics[width=6cm]{Fig14b}
\caption{\label{Fig14} The evolution of the wave function $\psi(t)$ with $J=0.2$ and $L=1$. $\text{Re}(\psi(t))$ and $\text{Im}(\psi(t))$ represent real part (a) and imaginary part (b) of wave function $\psi(t)$ respectively. The notation $|000\rangle \sim|111\rangle$ in the legend represents the eight components of $\psi(t)$.}
\end{figure*}
\subsection{Two or Three Qubits with Interaction}
\label{sec:2bitw/o}
In the case of a single qubit or double qubits without interaction, it is easy to obtain the solutions analytically, which are very regular.
For two and three qubits with interaction, however, it is very difficult to obtain such an analytical solution, and thus we resort to the numerical method discussed in~\ref{sec:shoot}.In this section, we give three different cases are investigated together with a general description of the consequences brought by possible entanglement in the wave functions.
Case 1, the initial and final spin states are the same as in section~\ref{sec:2bitw}, but there is an extra interaction term in the Hamiltonian
\begin{equation}\label{Hj}
H=B^{(1)}_x\sigma^{(1)}_x + B ^{(1)}_y\sigma^{(1)}_y+B ^{(2)}_x\sigma ^{(2)}_x + B ^{(2)}_y\sigma ^{(2)}_y+J_{yy}\sigma^{(1)}_y\sigma^{(2)}_y
\end{equation}
where $J_{yy}=1 $ is a fixed interaction constant.
In the numerical computation, in order to utilize a reasonable initial condition, we gradually increase the value of $J_{yy} $ from 0.01 to 1 with a step-size 0.01.
After the interaction term $J_{yy} $ is introduced, the wave function $\psi(t)$ is no longer separable, but becomes entangled.
Following the work in the literature~\cite{Wootters1998,Charles1996}, we may define an entanglement index $E_p$ for the two subsystems A and B, $E_p(|\psi\rangle_{AB})\equiv -Tr_A(\rho_A log_2\rho_A)=-Tr_B(\rho_B log_2\rho_B)$, where $\rho_A=Tr_B(|\psi\rangle_{AB}\langle\psi|)$ is the reduced density matrix of the pure state $|\psi\rangle_{AB}$ over subsystem A, and $\rho_B$ has a similar definition.
Fig.~\ref{Fig4}(a-d) depict the entanglement of $\psi(t)$ with L=1, 3, 2, 4, which have a common feature: the entanglement of the wave function connecting the two separable states (the initial and target states) increases first, reaches a maximum and then decreases back to zero. The entanglement evolution profile for L=1, 3 (Fig.~\ref{Fig4} a and b) is symmetric and the maximum is exactly at the symmetry point, while for L =2, 4 (Fig.~\ref{Fig4} c and d), the profile is skewed. Given a suitable initial value of $\phi'_0 $, our numerical method converges exponentially when approaching an optimal path connecting the initial and target states (See Fig.~\ref{Fig12}(c)).
Next, we keep the initial state unchanged and set the target state as an entangled one with $E_p=0.6061$. The wavefunction $\psi(t)$ connecting the two states is computed with J=1, 0.9, 0.65, the entanglement evolution profile of which is depicted in Fig.~\ref{Fig11}.
Case 2, we consider a two-qubit model, which is used in~\cite{Xiaoting2015}, with the following Hamiltonian
\begin{equation}\label{case3}
H=\sum\limits_{l,m}B_m^{(l)}\sigma_m^{(l)} +J\sum\limits_{m}\sigma_m^{(1)}\sigma_m^{(2)},
\end{equation}
where $\sigma_m^{(l)}~\text{with}~(m=x,y,z,l=1,2)$ are the Pauli matrices for the $l^{th}$ qubit. the initial and final spin states are the same as in section~\ref{sec:2bitw} and the coupling constant J=1. The results obtained by our method are portrayed in Fig.~\ref{Fig12}(a) and (b), where (a) depicts the components of the observable and (b) shows that the control magnetic field B(t).
Next, we compare the efficiency of finding such a connecting wavefunction among different methods, such as the GRAPE and the “Geodesic-search”~\cite{Xiaoting2015}. On the same machine and programming language, our method finds such a connection wavefunction and the modulation magnetic field (shown in Fig.~\ref{Fig12}(b)) in less than one minute. The GRAPE needs about 3.3 minutes and the “Geodesic-search” needs about 18 minutes. Fig.~\ref{Fig12}(c) and (d) compare the convergence upon iteration between our method and the GRAPE for the same convergence criterion $10^{-5}$. Obviously, although the computation with our method oscillates at the initial stage, it converges much more quickly and displays good stability as well. However, the convergence of the GRAPE is quite slow (with more than 500 iterations), which explains its relatively long computation time. For more complex multi-qubit systems, we believe that these advantages will become more obvious.
In this case, the optimal time is computed as $T=\frac{\sqrt{N}\pi\hbar}{\omega}$ for a quantum state evolving between two end states of a diameter on the Bloch sphere, which is consistent with the lower bound computed by Margolus and Levitin~\cite{Margolus1998,Levitin2009}. The Fubini-Study distance to the target state seems to decrease linearly with time as shown in Fig.~\ref{Fig13}, and we note that with the decrease of the energy $\omega$, the traveling time T increases.
Case 3, we study a model with three qubits for which the Hamiltonian is
\begin{equation}\label{H3}
H=\sum\limits_{l,m}B_m^{(l)}\sigma_m^{(l)}+ J(\sigma_y^{(1)}\sigma_y^{(2)}+\sigma_y^{(2)}\sigma_y^{(3)})
\end{equation}
where $\sigma_m^{(l)}~\text{with}~(m=x,y,z,l=1,2,3)$ are the Pauli matrices for the $l^{th}$ qubit. We choose the initial state $|+x\rangle^{(1)}\otimes|+y\rangle^{(2)}\otimes|+x\rangle^{(3)}$ and the final state $|-x\rangle^{(1)}\otimes|-y\rangle^{(2)}\otimes|-x\rangle^{(3)}$ . The obtained optimal wavefunction $\psi(t)$ connecting them with the interaction strength J=0.2 and L=1 is depicted in Fig.~\ref{Fig14}.
\begin{figure*}[]
\includegraphics[width=4.3cm]{Fig6a}
\includegraphics[height=3.5cm,width=4.3cm]{Fig6b}
\includegraphics[width=4.3cm]{Fig6c}
\includegraphics[width=4.3cm]{Fig6d
\caption{\label{Fig6} The wave function $\psi(t)$ for L=3 and $J_{yy}=1$. The red points are invariant under the symmetry operation F. $|00>,|01>,|10> $ and $|11>$ are the four components of $\psi(t) $, ‘Re’ and ‘Im’ represents the real and image part.}
\end{figure*}
\subsection{ The Symmetry Operation Discussion}
\begin{figure*}[]
\includegraphics[height=3cm,width=3.6cm]{Fig9a}
\includegraphics[height=3cm,width=3.6cm]{Fig9b}
\includegraphics[height=3cm,width=3.6cm]{Fig9c}
\includegraphics[height=3cm,width=3.6cm]{Fig9d}
\includegraphics[height=3cm,width=3.6cm]{Fig8a}
\includegraphics[height=3cm,width=3.6cm]{Fig8b}
\includegraphics[height=3cm,width=3.6cm]{Fig8c}
\includegraphics[height=3cm,width=3.6cm]{Fig8d
\caption{\label{Fig8} The function $\bm B(t)$ for L=2 ($a\sim d$) with $J_{yy}=0.05$ and L=3 ($e\sim h$)with $J_{yy}=0.2$. The black lines plot the controlled part $\bm B(t)$ for qubit one and the red lines plot the controlled part $\bm B(t)$ for the second qubit. The results of $c,~d,~g$ and $h$ were obtained for the transformed Hamiltonian and states $H'=F^T H F$, $\psi_0'=F\psi_0$, $\psi_{f}'=F\psi_{f}$.}
\end{figure*}
\begin{figure}[]
\includegraphics[width=6cm]{Fig7}
\caption{\label{Fig7} The entanglement distribution of $\psi(t)$ (the black line) and $\psi'(t)$ (the red line) with $J_{yy}=0.05$ and L=3, $\psi'(t) $ is the symmetry counterpart of $\psi(t) $ under the action $F$.}
\end{figure}
In this section, we check a symmetry operation of the Hamiltonian Eq.~(\ref{Hj}) and ferret out possible mechanism underlying the phenomenon discussed in Part~\ref{sec:2bitw/o} above.
It is interesting to observe a discrete symmetry operation in Hamiltonian~(\ref{Hj}) associated with
\begin{equation}\label{ary}
F=
\left[
\begin{matrix}
~~0 & ~~0 & ~~0 & ~~i \\
~~0 & ~~0 & ~~i & ~~0 \\
~~0 & -i & ~~0 & ~~0 \\
-i & ~~0 & ~~0 & ~~0
\end{matrix}
\right]
\end{equation}
which connects the initial state $\psi_0 $ and the target state $\psi_f$, {\em i.e.} $F\psi_0=-\psi_f $ in the current case. Obviously, the magnetic field $\bm B(t)$ in the Hamiltonian is artificially controlled, which should be changed accordingly (by $F^THF$) to complete the symmetry operation, as confirmed in Fig.~\ref{Fig8}.
For the case of L=1,3, especially, the middle state $\psi_{mid}$ (at t=0.5) remains invariant under this symmetry operation, namely $F\psi_{mid}=\psi_{mid}$. In fact, to search the symmetry, a clue could be directly taken from Fig.~\ref{Fig6}, where the red points represent the middle state $\psi_{mid}= (0.09483-0.08982i,~ -0.69486-0.02312i,~ -0.02309+0.69443i,~ -0.08993-0.09488i)^t$. It is easy to check that the time course of the wave function $\psi_{t}$ is also symmetric about the middle state $\psi_{mid}$. From the view of the controlled part $\bm B(t)$ of $H$, take the cases of L=2 and 3 as an example. Fig.\ref{Fig8} $a,~b~(L=2)$ and $e,~f~(L=3)$ depict the time course of $\bm B(t)$ from $\psi_0$ to $\psi_{f}$ while Fig.\ref{Fig8} $c,~d~(L=2)$ and $g,~h~(L=3)$ plot the time course of $\bm B(t)$ from $\psi_0'=F\psi_0$ to $\psi_{f}'=F\psi_{f}$ with the transformed Hamiltonian $H'= F^T H F$. Apparently, Fig.\ref{Fig8} $e$ and $f$ coincide with $g$ and $h$ respectively up to a minus sign in the relevant component. In addition, the evolution of $B^{(1)}_x$ and $B^{(2)}_y$ is axisymmteric with respect to $t=0.5$ and the evolution profiles of $B^{(1)}_y$ and $B^{(2)}_x$ have a center symmetry with respect to (0.5, 0). All the above indicates that the maximum of $E_p$ is associated with a fixed point of the operation $F$ for the case L=1, 3.
In the cases L=2, 4, the symmetry in $\psi(t)$ and $\bm B(t)$ discussed previously disappear. Taking L=2 for an example, Fig.\ref{Fig8} $a$ and $b$ do not coincide with Fig.\ref{Fig8} $c$ and $d$ and the maximum value of $E_p$ is not at t=0.5. On the other hand, as the symmetry partner $\psi'_{t}$ of the wave function $\psi_{t}$ is obtained with the symmetry operation $F$, even if the entanglement evolution of $\psi'_{t}$ (see Fig.~\ref{Fig7}, the red line) is asymmetric, those of $\psi'(t)$ and $\psi(t)$ are axially symmetric with respect to each other about t=0.5.
\begin{figure}[]
\includegraphics[width=7cm]{Fig10}
\caption{\label{Fig10} The entanglement evolution of $\psi(t)$ with $L=1$ and $J_{yy}=0.4$ } in Eq.~(\ref{Hj}), the angle between $\psi_0$ and $\psi_f$ is from $\pi$ to $0.7\pi$.
\end{figure}
In the case where $\psi_0$ and $\psi_f$ do not flip to each other, {\em e.g.} the angle between $\psi_0$ and $\psi_f$ is less than $\pi$, we then have $F\psi_0\neq-\psi_f$ (take the case $L=1$ as an example), which indicates that the entanglement evolution is asymmetric (shown in Fig.~\ref{Fig10}). However, the maximum entanglement still exists which is unique but not at $t=0.5$.
\section{SUMMARY AND DISCUSSION}
In this work, we derive a general set of differential equations (Eq.~(\ref{eq:evo2})) for an optimal quantum control, which is valid for single or multiple qubits with or without interaction. In the derivation and application of Eq.~(\ref{eq:evo2}), we discussed the following aspects:
First, previous numerical methods conventionally involve a large number of physically irrelevant variables derived from the commutation relation with system Hamiltonians. We thus introduce a new set of observables to eliminate the physically irrelevant part, whereby an accelerated computation becomes possible. As a result, the efficiency of computation are boosted significantly. With the help of a ‘relaxation’ idea, the drawback of the shooting method, of requiring a good initial guess, is surpassed by slowly pushing the trial solution to the correct one with a gradual restoration of the interaction. Second, we discussed symmetries of the Lagrangian formalism of the QBE. In the numerical calculation, these entail unwanted degenerate directions in the Jacobian matrix, which may cause serious trouble in the solution of the resulting boundary value problem.
Finally, groups of analytic solutions are given as rotations of qubits on Bloch spheres in the absence of interaction, which may be used as the starting point of the above relaxation scheme.
In an application of the new scheme to the case of two qubits with interaction, the solution to the QBE displays a unimodal evolution profile of the entanglement. If the initial and final state transform to each other under a symmetry operation of the Hamiltonian. The evolution of the optimal path either bears a related symmetry or has a symmetry partner. More explicitly, for the even oscillation number symmetry disappears in the profile but the symmetry operation gives another optimal path -- the symmetry partner of the original solution.
With current formalism, we are able to design optimal control strategy for multi-qubits with possibly complex interaction.
However, the evolution is not unique and many local optimal paths exist which could be indexed by the number of oscillations of the paths in simple cases. It could be interesting to classify these paths and check their bifurcation routes when system parameters change. On the other hand, this work may provide experimental researchers a practical computation tool for quantum information regulation, since an optimal control strategy of the magnetic field is readily designed based on our scheme.
\acknowledgments
This work was supported by the National Natural Science Foundation of China under Grants No. 11775035, and also by the Fundamental Research Funds for the Central Universities with Contract No.2019XD-A10.
|
\section{Introduction}
\label{sec:intro}
Since the discovery of the $J/\Psi$ charmonium in 1974\cite{Aubert,Augustin}, the quarkonium production has attracted much attention experimentally and theoretically as it provides an excellent way to investigate both the perturbation and non-perturbation regimes of quantum chromodynamics. In recent years, the measurements released by the ALICE Collaboration\cite{Aamodt}, CMS Collaboration\cite{Silvestre}, ATLAS Collaboration\cite{Aad} and LHCb Collaboration\cite{Aaij} provide an unique opportunity to explore the heavy quarkonium production mechanism. On the theoretical aspect, several approaches have been proposed to calculate the production of heavy quarkonium, such as the color singlet mechanism (CSM)\cite{Petrelli,Braaten,Einhorn,Chang,Carlson}, the color dipole mechanism\cite{Song,Kharzeev,Goncalves,Qiu}, the color-evaporation model (CEM)\cite{Godbole,Fujii}, $k_T$-factorization method\cite{Gribov,Kniehl1,Baranov1} and the NRQCD\cite{Bodwin,Mathews,He1,He2,Baranov2,Kang1,Kang2,Butenschoen,Han}. Among these approaches, the most widely used methods for heavy quarkonia production are CSM and NRQCD.
To describe the heavy quarkonium production, a key ingredient is the nonperturbative evolution of an intermediate $Q\bar{Q}$ pair into the final quarkonium. The CSM is the first model to describe this nonperturbative evolution. The CSM was successful in describing quarkonium production at low energy\cite{Schuler}. However, there are considerable discrepancies between CSM predictions and the experimental data at Tevatron. The CSM prediction of inclusive $J/\Psi$ hadroproduction underestimates the experimental data by more than one order of magnitude \cite{Abe}. For inclusive $\Psi(2S)$ hadroproduction, the discrepancy is even larger, reaching a factor of 50\cite{Abe}. The reason for these discrepancies is the coupled quantum number of the heavy quark pair is assumed to be in a color singlet in the CSM. To solve this puzzle, Bodwin, Braaten and Lepage proposed the NRQCD mechanism by introducing the color octet process\cite{Bodwin}. In the NRQCD framework, the quarkonium can evolve from color singlet intermediate heavy quark pair, as well as color octet heavy quark pair, which greatly fill the gap between the CSM predictions and experimental measurements\cite{Kniehl2}. The theoretical results about quarkonium hadroproduction from NRQCD are also in good agreement with the experimental measurements of $\chi_{cJ}$ and $\Upsilon(1S)$\cite{Kramer}. As NRQCD factorization method has achieved great success in description heavy quarkonuim hadroproduction, many efforts have been done to investigate the heavy quarkonuim photoproduction with NRQCD.
Photoproduction process is a vital process in electron-proton deep inelastic scattering (DIS) and high energy ultra-peripheral hadron collisions, since it offers significant information about the interaction mechanisms\cite{Cai1}. The $J/\Psi$ photoproduction at HERA has been studied in the framework of the NRQCD by taking into account the color octet contributions. It was found that the color octet matrix elements, which are obtained by fitting the hadroproduction data at the Tevatron, overestimate the HERA data an order of magnitude\cite{Ko}. This excess is disappeared when the higher order effects are included\cite{Kniehl2}. The quarkonium photoproduction cross section at LHC has been investigated by using the CSM in Ref.\cite{Goncalves2}. Meanwhile, the NRQCD has been used to study the inclusive diffractive quarkonium photoproduction at LHC by combining with the resolved pomeron model\cite{Goncalves3}. Later on, the $\eta_c$ production has been performed in inclusive and diffractive processes with the NRQCD model in Ref.\cite{Goncalves4}. Although considerable efforts have been made to improve the performance of the theoretical models, the theoretical calculations can match with the experimental data under certain uncertainties, the mechanism for heavy quarkonium production is still far from the deep understanding. Therefore, a comprehensive analysis of the contributions from different subprocesses may be helpful for understanding the heavy quarkonium production mechanism, such as $g+g\rightarrow Q\bar{Q}+g$\cite{Yang}.
In our previous study, the inclusive diffractive heavy quarkonium photoproduction has been investigated by taking into account the $g+g\rightarrow Q\bar{Q}+g$\cite{Yang}. We found that the resolved photoproduction processes have a significant contribution to the heavy quarkonium production. In this paper, the quark subprocesses, $\gamma + q \rightarrow Q\bar{Q}+q$, $g + q \rightarrow Q\bar{Q}+q$, $q + \bar{q} \rightarrow Q\bar{Q}+g$, are included into the direct and resolved photoproduction processes on top of our previous studies in Ref.\cite{Yang}. We find that the contributions from the quark involved subprocesses can reach to $8\%$ in the rapidity distribution and $6\%$ in the transverse momentum distribution.
\section{Inclusive diffractive quarkonium photoproduction in NRQCD }
In this section, we firstly introduce the mechanism of the diffractive heavy quarkonium photoproduciton based on the NRQCD combined with the resolved pomeron model. Then we give the total cross section for heavy quarkonium produciton in direct and resolved photoproduction processes. The total cross section and the rapidity distribution of the heavy quarkonium photoproduction in $pp$, $pPb$ and $PbPb$ collisions are presented at the end of this section.
\subsection{The mechanism of the diffractive quarkonium photoproduciton}
\label{sec:nrqcd}
The NRQCD is one of the most widely used theoretical tool to describe the quarkonium production\cite{Bodwin,Mathews,He1,He2,Baranov2,Kang1,Kang2,Butenschoen,Han}. In the NRQCD formalim, the cross section for the production of a heavy quarkonium $H$ can be factorized as\cite{Hao}
\begin{equation}
\sigma(ab\rightarrow H+X)=\sum_n\sigma(ab\rightarrow Q\overline{Q}[n]+X)<\mathcal{O}_{[n]}^H> ,
\label{nrqcd}
\end{equation}
where $\sigma(ab\rightarrow Q\overline{Q}[n]+X)$ is the short distance process and $<O_{[n]}^H>$ is the long distance process. In the short distance process, a pair of quark-antiquark is produced at collision point, and the quark-antiquark forms a special system $Q\overline{Q}[n]$. In the long distance process, the special system form a bound state, and eventually the bound state evolutes into the final quarkonium through soft gluon emission. In the NRQCD formulism, the short distance matrix elements can be calculated perturbatively, but the long distance matrix is a non perturbative process and it is usually given by the lattice QCD calculation or extracting from experimental data.
In this work, we combine the NRQCD factorization approach with resolved pomeron model to investigate the inclusive diffractive heavy quarkonium photoproduction in hadron-hadron collisions. The photoproduction has direct and resolved two type processes according to the scattering mechanism of the photon. A schematic diagrams are shown in Fig.\ref{fig1}. In the direct photoproduction processes (Fig.\ref{fig1} (left panel)), the photons emitted by the hadron ($A$) interact with the partons (quarks and gluons) from the resolved pomerons in the hadron ($B$) to produce final particle $H$. In the resolved photoproduction processes (Fig.\ref{fig1} (right panel)), due to the Heisenberg uncertainty principle the energetic photons from the energetic hadron ($A$) can fluctuate into partons. Then these partons interact with the partons from the resolved pomerons to form the final heavy quarkonium.
\begin{figure}[h]
\setlength{\unitlength}{1.5cm}
\begin{center}
\epsfig{file=1, width=12cm,height=5cm}
\end{center}
\caption{The schematic diagrams of inclusive diffractive heavy quarkonium direct (left panel) and resolved (right panel) photoproduction in the resolved pomeron model.}
\label{fig1}
\end{figure}
\subsection{The cross section of the direct photoproduciton}
\label{sec:dir}
According to the NRQCD formulism, the total cross section can be factorized into the partonic distribution function and differential cross section. Therefore, the total cross section of direct photoproduction process $\gamma + B\rightarrow H + X$ can be written as
\begin{equation}
\sigma_{dir}(\gamma+B\rightarrow H+X)=\int dzdp^2_T\frac{x_{dir}f_{b/B}(x_{dir},Q^2)}{z(1-z)}\frac{d\sigma}{dt}(\gamma+b\rightarrow H+X),
\label{totdir}
\end{equation}
where $f_{b/B}(x_{dir},Q^2)$ is the diffractive parton distribution\cite{Ingelman}. The $x_{dir}$ in Eq.(\ref{totdir}) is the momentum fraction of hadron carried by the quark or gluon and can be written as
\begin{equation}
x_{dir}=\frac{p^2_T+M^2(1-z)}{W^2_{\gamma h}z(1-z)},
\label{xdir}
\end{equation}
where $p_T$ is the transverse momentum of the quarkonium and $M$ is the mass of the quarkonium. The $W_{\gamma h}$ in Eq.(\ref{xdir}) is the photon-hadron center-of-mass energy
\begin{equation}
W_{\gamma h}=\sqrt{2\omega\sqrt{s}},
\label{cms}
\end{equation}
where $\sqrt{s}$ represents the center-of-mass energy of the $AB$ collision system and $z$ is the fraction of the photon energy. The $\omega$ in Eq.(\ref{cms}) is the photon energy in the proton-proton collision system
\begin{equation}
\omega=\frac{M}{2}\exp(\pm y).
\label{ey}
\end{equation}
Note that the fraction of photon energy carried by the quarkonium in Eq.(\ref{totdir}) is integrated over the range\cite{Aaron},
\begin{equation}
0.3\le z\le 0.9.
\end{equation}
\subsection{The cross section of resolved photoproduction}
\label{res}
In the resolved photoproduction process, the energetic photons fluctuate into quarks and gluons which then interact with the partons from the pomeron. According to the factorization formalism, the cross section for the resolved photoproduction process is given by
\begin{equation}
\sigma_{res}(a+B \rightarrow H+X)=\int dzdp^2_Tdx\frac{x_{res}f_{a/\gamma}(x_{res},Q^2)f_{b/B}(x,Q^2)}{z(1-\frac{z}{x_{res}})}\frac{d\sigma}{dt}(a+b\rightarrow H+X),
\label{totres}
\end{equation}
where $f_{a/\gamma}(x_{res},Q^2)$ is the parton distribution function of the hadron-like photon\cite{Gluck}, and $x_{res}$ is the momentum fraction of photon carried by the quark or gluon
\begin{equation}
x_{res}=\frac{xp^2_T+M^2(x-z)}{W^2_{\gamma p}z(1-\frac{z}{x})}.
\label{res}
\end{equation}
In the early studies, the heavy quarkonium was investigated in the CSM\cite{Goncalves2,Berger,Baier1,Krmer}, in which the $Q\overline{Q}$ produced by the partonic process is color-singlet. In direct and resolved photoproduction, the color-singlet processes are given by
\begin{equation}
\gamma+g\rightarrow Q\overline{Q}[^3S_1,1]+g,
\label{res9}
\end{equation}
\begin{equation}
g+g\rightarrow Q\overline{Q}[^3S_1,1]+g.
\label{res10}
\end{equation}
The contribution of $\gamma g$ process to heavy quarkonium photoproduction was studied in Ref.\cite{Goncalves2,Goncalves3} and the contribution of $gg$ process was investigated in our previous studies in Ref.\cite{Yang}. Although the theoretical calculations can match with the experimental data under certain uncertainties, there are many sources that contribute to the theoretical uncertainties, such as the uncertainties in long-distance matrix elements, scale, and parton distribution functions\cite{Yang}. To get a comprehensive understanding about the mechanism of heavy quarkonium production, an analysis of contribution from the quark involved subprocesses is essential. In fact that the quarks are also active partons in hadron-like photons and pomerons, so one has to include the contributions from quark processes in order to improve our understanding of the heavy quarkonium production mechanism. In this work, we take into account the quark processes to investigate the inclusive diffractive heavy quarkonium photoproduction by combining the NRQCD model with resolved pomeron model. We call this approach as quark improved NRQCD model. When the quark contributions are considered, the following processes should be included
\begin{equation}
\gamma+q\rightarrow Q\overline{Q}[^3S_1,1]+q,
\label{res11}
\end{equation}
\begin{equation}
g+q\rightarrow Q\overline{Q}[^3S_1,1]+q,
\label{res12}
\end{equation}
\begin{equation}
q+\overline{q}\rightarrow Q\overline{Q}[^3S_1,1]+g,
\label{res13}
\end{equation}
where the above three subprocesses can be obtained in Ref.\cite{Klasen}.
We know that the CSM can be used to describe $J/\Psi$ production in $\gamma p$ or $ep$ collisions. However, the theoretical prediction based on the CSM underestimates the cross section of inclusive $J/\Psi$ production measured in $pp$ collisions\cite{Abe}, and this inconsistency cannot be attributed to different energy scales, the heavy quark mass or parton distribution function\cite{Cano-Coloma}. To solve this difficulty, Bodwin, Braaten, and Lepage\cite{Bodwin} developed the Color-Octet mechanism (COM) based on NRQCD which allows a systematic calculation of inclusive heavy quarkonium production cross sections. In the COM, the processes can be expressed as follows\cite{Klasen}
\begin{equation}
\gamma+g\rightarrow Q\overline{Q}[^3S_0, ^3S_1, ^3P_{0,1,2},8]+g,
\label{res14}
\end{equation}
\begin{equation}
g+g\rightarrow Q\overline{Q}[^1S_0, ^3S_1, ^3P_{0,1,2},8]+g,
\label{res15}
\end{equation}
\begin{equation}
\gamma+q\rightarrow Q\overline{Q}[^3S_0, ^3S_1, ^3P_{0,1,2},8]+q,
\label{res16}
\end{equation}
\begin{equation}
g+q\rightarrow Q\overline{Q}[^1S_0, ^3S_1, ^3P_{0,1,2},8]+q,
\label{res17}
\end{equation}
\begin{equation}
q+\overline{q}\rightarrow Q\overline{Q}[^1S_0, ^3S_1, ^3P_{0,1,2},8]+g.
\label{res18}
\end{equation}
Using Eqs.(\ref{res9})-(\ref{res18}), we can get the contributions of heavy quarkonium in the direct and resolved photoproduction processes.
\subsection{The total cross section and the rapidity distribution of the heavy quarkonium photoproduction}
\label{sec:tot}
In this subsection, we will firstly discuss the total cross section of heavy quarkonium production. Then we give the rapidity distribution of the quarkonium photoproduction on the basis of total cross section. The total cross section for the heavy quarkonium in the photoproduction process is given by\cite{Cai2}
\begin{equation}
\sigma_{tot}(A+ B\rightarrow A\otimes H+X\otimes B) =\int d\omega\frac{dN_{\gamma/A}(\omega)}{d\omega}\sigma_{\gamma B}\rightarrow HX\otimes B\\
+\int d\omega\frac{dN_{\gamma/B}(\omega)}{d\omega}\sigma_{\gamma A}\rightarrow HX\otimes A,
\label{totcs}
\end{equation}
where $\otimes$ denotes the existence of a rapidity gap in the final state and $\omega$ denotes the energy of photon. The $A$ and $B$ denote a hadron or nucleus. For $pp$ collision, using the relationship between rapidity and energy in Eq.(\ref{ey}), the rapidity distribution can be written as
\begin{equation}
\frac{d\sigma_{tot}^{pp\rightarrow pH+Xp}}{dy}=[\omega\frac{dN^p_\gamma}{d\omega}\sigma^{\gamma p\rightarrow HXp}]_{\omega=\omega_l}+[\omega\frac{dN^p_\gamma}{d\omega}\sigma^{\gamma p\rightarrow HXp}]_{\omega=\omega_r},
\label{totpp}
\end{equation}
where the subscrips $l(r)$ denote photon flux from the right(left) proton.
The equivalent photon flux $\frac{dN_\gamma^p}{d\omega}$ of the relativistic proton in Eq.(\ref{totpp}) is given by\cite{Klein,Drees}
\begin{equation}
\frac{dN_\gamma^p}{d\omega}=\frac{\alpha_{em}}{2\pi\omega}[1+(1-\frac{2\omega}{\sqrt{s}})^2](\ln\Omega-\frac{11}{6}+\frac{3}{\Omega}-\frac{3}{2\Omega^2}+\frac{1}{3\Omega^3}),
\label{tot}
\end{equation}
with
\begin{equation}
\Omega=1+[(0.71\rm {GeV}^2)/Q_{min}^2],
\label{tot}
\end{equation}
where $Q_{min}^2 = \omega^2/[\gamma_L^2(1-2\omega/\sqrt{s})]\approx(\omega/\gamma_L)^2$ and $\gamma_L$ is the lorentz factor.
To calculate the rapidity distribution in Eq.(\ref{totpp}), we need to specify the diffractive parton distribution $f_{b/p}(x, Q^2)$, which can be expressed as the convolution of the pomeron flux in the proton $f_\mathcal{P}^P(x_\mathcal{P})$ and parton distribution in the pomeron\cite{Ingelman}
\begin{equation}
f_{b/p}(x,Q^2)=\int_x^1\frac{dx_\mathcal{P}}{x_\mathcal{P}}f^p_\mathcal{P}(x_\mathcal{P})g_\mathcal{P}(\frac{x}{x_\mathcal{P}},Q^2),
\label{pdfp}
\end{equation}
where the pomeron fluxes is given by
\begin{equation}
f_\mathcal{P}^p(x_\mathcal{P})=\int_{t_{min}}^{t_{max}}dtf_\mathcal{P}^p(x_\mathcal{P},t)=\int_{t_{min}}^{t_{max}}\frac{\lambda e^{\beta t}}{x_\mathcal{P}^{2\alpha_{\mathcal{P}}(t)-1}}dt,
\label{tot}
\end{equation}
with
\begin{equation}
t_{max}=-m_p^2x_\mathcal{P}^2/(1-x_\mathcal{P}),
\label{res}
\end{equation}
\begin{equation}
t_{min}=-1\,\rm{GeV}^2.
\label{tot}
\end{equation}
In Eq.(\ref{pdfp}), $g_\mathcal{P}(\frac{x}{x_\mathcal{P}}, Q^2)$ is the parton distribution in pomeron, which is given by a parameterized formula in Ref.\cite{Aktas}.
For $pPb$ collisions, the rapidity distribution is
\begin{equation}
\frac{d\sigma_{tot}^{pPb\rightarrow pHXPb}}{dy}=[\omega\frac{dN^{Pb}_\gamma}{d\omega}\sigma^{\gamma Pb\rightarrow HXp}]_{\omega=\omega_l}+[\omega\frac{dN_\gamma^p}{d\omega}\sigma^{\gamma p\rightarrow HXPb}]_{\omega=\omega_r},
\label{totppb}
\end{equation}
where $\frac{dN_\gamma^{Pb}}{d\omega}$ denotes the photon flux of lead\cite{Papageorgiu}
\begin{equation}
\frac{dN_\gamma^{Pb}}{d\omega}=\frac{2Z^2\alpha_{em}}{\pi\omega}[\xi K_0(\xi)K_1(\xi)-\frac{\xi^2}{2}(K_1^2(\xi)-K_0^2(\xi))],
\label{tot}
\end{equation}
with $\xi = \omega(R_p + R_{Pb})/\gamma_L$. The $K_{0}$ and $K_{1}$ are the modified Bessel functions and $Z$ is atomic number. Since the photon flux of the nuclei is larger than the photon flux of the proton, the rapidity distribution should be obvious asymmetry with a great peak at the lead side in the $pPb$ collision. The diffractive parton distribution of the lead can be expressed as follows\cite{Basso}
\begin{equation}
f_{b/Pb}(x_\mathcal{P})=\int_{t_{min}}^{t_{max}}dtf_\mathcal{P}^{Pb}(x_\mathcal{P},t)=R_gA^2\int_{t_{min}}^{t_{max}}\frac{\lambda e^{\beta t}}{x_\mathcal{P}^{2\alpha_{\mathcal{P}}(t)-1}}dtF_A^2(t),
\label{tot}
\end{equation}
where $R_{g}$ is the suppression factor associated to the nuclear shadowing and $F_A(t)$ is the nuclear form factor.
In addition to the rapidity distribution of heavy quarkonium production in $pp$ and $pPb$ collisions, we also study the heavy quarkonium photoproduction in the $PbPb$ collision by including the quark involved subprocesses as mentioned above. The rapidity distribution of $PbPb$ collision can be written as
\begin{equation}
\frac{d\sigma_{tot}^{PbPb\rightarrow PbHXPb}}{dy}=[\omega\frac{dN^{Pb}_\gamma}{d\omega}\sigma^{\gamma Pb\rightarrow HXPb}]_{\omega=\omega_l}+[\omega\frac{dN_\gamma^{Pb}}{d\omega}\sigma^{\gamma Pb\rightarrow HXPb}]_{\omega=\omega_r}.
\label{totpbpb}
\end{equation}
\section{Results}
\label{sec:numsol}
In this section, we firstly calculate the total cross section distribution of the inclusive $J/\Psi$, $\Psi(2S)$ and $\Upsilon(1S)$ production to explore the contribution of the quark subprocesses, and compare our theoretical calculations with the $J/\Psi$ experimental data from H1 Collaboration\cite{Aaron}. Then we use the quark improved NRQCD model to make predictions of rapidity and transverse momentum distributions of the inclusive diffractive $J/\Psi$, $\Psi(2S)$ and $\Upsilon(1S)$ photoproduction in $pp$, $pPb$ and $PbPb$ collisions.
In our studies, the masses of charm quark and bottom quark are taken 1.5 GeV and 4.5 GeV, respectively. The minimum value of the transverse momentum is set to 1 GeV. We set the factorization scale as $Q^2 = P^2_T + (M/2)^2$. In addition, the differential cross section of the partonic processes is calculated in Ref.\cite{Klasen}, and the long-distance matrix elements are listed in Table.1\cite{Yu}.
\begin{table}[h]
\begin{tabular}{cccc}
\toprule
$(\mathrm{GeV}^3)$ &$J/\Psi$ &$\Psi(2S)$ &$\Upsilon(1S)$\\
\hline
$<\mathcal{O}[^3S_1^{[1]}]>$& 1.2& 0.76&10.9\\
$<\mathcal{O}[^3S_0^{[8]}]>$& $0.018\pm0.0087$&$0.0080\pm 0.0067$&$0.121\pm0.00400$ \\
$<\mathcal{O}[^3S_1^{[8]}]>$& $0.0013\pm0.0013$& $0.00330\pm0.0021$& $0.0477\pm0.0334$\\
$<\mathcal{O}[^3P_0^{[8]}]>$& $(0.018\pm0.0087)m_c^2$&$(0.0080\pm 0.0067)m_c^2$&$(0.121\pm0.00400)m_b^2$ \\
$<\mathcal{O}[^3P_1^{[8]}]>$& $3\times(0.018\pm0.0087)m_c^2$&$3\times(0.0080\pm 0.0067)m_c^2$&- \\
$<\mathcal{O}[^3P_2^{[8]}]>$& $5\times(0.018\pm0.0087)m_c^2$&$5\times(0.0080\pm 0.0067)m_c^2$&- \\
\hline
\end{tabular}
\caption{Values of long-distance matrix elements for $J/\Psi$, $\Psi(2S)$ and $\Upsilon(1S)$.}
\end{table}
\begin{figure}[b]
\setlength{\unitlength}{1.5cm}
\begin{center}
\epsfig{file=jps, width=8cm, height=10cm,angle=270}
\end{center}
\caption{The energy dependence of the cross section for the $J/\Psi$ photoproduction in inclusive $\gamma p$ interactions. Data are from H1 Collaboration\cite{Aaron}.}
\label{fig2}
\end{figure}
In Fig.\ref{fig2}, we give the theoretical results for energy dependence of the inclusive $J/\Psi$ total cross section calculated by Eqs.(\ref{totdir}) and (\ref{totres}) with a parton direct from the proton and compare with the experimental data. The dash purple line is the contribution from the $\gamma g$ direct photoproduction process. The dotted green line is the result from the $gg$ resolved photoproduction process. The dot-dash orange line is the contributions from the processes of quarks($\gamma q$, $qq$, $qg$), which are introduced, for the first time, into theoretical calculations of the heavy quarkonium photoproduction in the NRQCD framework. The solid red line is the total cross section. As one can see that the quark improved NRQCD model can give a rather successful description of the $J/\Psi$ photoproduction data from H1 Collaboration, which indicates that the quark involved three subprocesses have a significant contributions.
Using Eqs.(\ref{totdir}) and (\ref{totres}), the predictions of the inclusive total cross section for $\Psi(2S)$ and $\Upsilon(1S)$ are computed and shown in Fig.\ref{fig3}. We can see that the quark involved three subprocesses also have a significant contributions to the $\Psi(2S)$ and $\Upsilon(1S)$ productions. From Figs.\ref{fig2} and \ref{fig3}, we can see that in the low region of center-of-mass (or large $x$), the contributions of the quark involved three subprocesses are larger than the ones from $gg$ process, since the quark distribution is larger than the gluon distribution function in the low energy region.
\begin{figure}[t]
\setlength{\unitlength}{1.5cm}
\begin{center}
\epsfig{file=psi, width=6cm,height=8cm,angle=270}
\epsfig{file=upsilon, width=6cm,height=8cm,angle=270}
\end{center}
\caption{Predictions for the energy dependence of the cross section for the $\Psi(2S)$ and $\Upsilon(1S)$ photoproduction in inclusive $\gamma p$ interaction.}
\label{fig3}
\end{figure}
\begin{figure}[t]
\setlength{\unitlength}{1.5cm}
\begin{center}
\epsfig{file=jpp, width=4cm,height=5cm,angle=270}
\epsfig{file=ppp, width=4cm,height=5cm,angle=270}
\epsfig{file=upp, width=4cm,height=5cm,angle=270}
\end{center}
\caption{Rapidity distributions for the inclusive diffractive $J/\Psi$, $\Psi(2S)$ and $\Upsilon(1S)$ photoproduction in $pp$ collisions at $\sqrt{s}$ = 5.02 TeV.}
\label{fig4}
\end{figure}
From the inclusive $J/\Psi$, $\Psi(2S)$ and $\Upsilon(1S)$ photoproduction at HERA energies, one can see that the quark involved three subprocesses have a significant contributions to the total cross section. It is reasonable to believe that the three subprocesses mentioned above are also important in the heavy quarkonium production at LHC energies. So, we use the quark improved NRQCD model to study the heavy quarkonium photoproduction at LHC energies. Via using Eq.(\ref{totpp}), we calculate the predictions of the rapidity distribution for the inclusive diffractive $J/\Psi$, $\Psi(2S)$ and $\Upsilon(1S)$ in $pp$ collision at $\sqrt{s} = 5.02$ TeV, the relevant numerical results are shown in Fig.\ref{fig4}. We estimate the percentage of contributions from the quark involved three subprocesses, and find that the three subprocesses can reach to $8\%$ of the total rapidity distribution. Thus, one needs to include their contributions into the heavy quarkonium production for an accurate description of the experimental data.
\begin{figure}[t]
\setlength{\unitlength}{1.5cm}
\begin{center}
\epsfig{file=jpa, width=4cm,height=5cm,angle=270}
\epsfig{file=ppa, width=4cm,height=5cm,angle=270}
\epsfig{file=upa, width=4cm,height=5cm,angle=270}
\end{center}
\caption{Rapidity distributions for the inclusive diffractive $J/\Psi$, $\Psi(2S)$ and $\Upsilon(1S)$ photoproduction in $pPb$ collision at $\sqrt{s}$ = 5.02 TeV.}
\label{fig5}
\end{figure}
\begin{figure}[t]
\setlength{\unitlength}{1.5cm}
\begin{center}
\epsfig{file=jaa, width=4cm,height=5cm,angle=270}
\epsfig{file=paa, width=4cm,height=5cm,angle=270}
\epsfig{file=uaa, width=4cm,height=5cm,angle=270}
\end{center}
\caption{Rapidity distributions for the inclusive diffractive $J/\Psi$, $\Psi(2S)$ and $\Upsilon(1S)$ photoproduction in $PbPb$ collision at $\sqrt{s}$ = 5.02 TeV.}
\label{fig6}
\end{figure}
In the LHC energies, $pPb$ and $PbPb$ collisions have also attracted much attention. So we calculate the predictions of rapidity distributions for these two collision processes via Eqs.(\ref{totppb}) and (\ref{totpbpb}), respectively. We present the numerical calculations of the rapidity distributions for the inclusive diffractive $J/\Psi$, $\Psi(2S)$ and $\Upsilon(1S)$ photoproduciton in $pPb$ collision at $\sqrt{s} = 5.02$ TeV in Fig.\ref{fig5}. As expected, in $pPb$ collision the rapidity distributions are asymmetric, with the peak at the lead side, since the photon spectrum of a nucleus is much larger than a proton.
The rapidity distributions for the inclusive diffractive $J/\Psi$, $\Psi(2S)$ and $\Upsilon(1S)$ photoproduciton in $PbPb$ collision at $\sqrt{s} = 5.02$ TeV are shown in Fig.\ref{fig6}. As in the $pp$ collision, the total rapidity distributions are symmetric about the midrapidity. However, the rapidity distributions are larger than the $pp$ collision due to the enhancement of the photon flux.
\begin{figure}[b]
\setlength{\unitlength}{1.5cm}
\begin{center}
\epsfig{file=jppdpt, width=4cm,height=5cm,angle=270}
\epsfig{file=pppdpt, width=4cm,height=5cm,angle=270}
\epsfig{file=uppdpt, width=4cm,height=5cm,angle=270}
\end{center}
\caption{Transverse momentum distributions for the inclusive diffractive $J/\Psi$, $\Psi(2S)$ and $\Upsilon(1S)$ photoproduction at central rapidities($y = 0$) in $pp$ collisions at $\sqrt{s}$ = 5.02 TeV.}
\label{fig7}
\end{figure}
\begin{figure}[t]
\setlength{\unitlength}{1.5cm}
\begin{center}
\epsfig{file=jpadpt, width=4cm,height=5cm,angle=270}
\epsfig{file=ppadpt, width=4cm,height=5cm,angle=270}
\epsfig{file=upadpt, width=4cm,height=5cm,angle=270}
\end{center}
\caption{Transverse momentum distributions for the inclusive diffractive $J/\Psi$, $\Psi(2S)$ and $\Upsilon(1S)$ photoproduction at central rapidities($y = 0$) in $pPb$ collisions at $\sqrt{s}$ = 5.02 TeV.}
\label{fig8}
\end{figure}
\begin{figure}[t]
\setlength{\unitlength}{1.5cm}
\begin{center}
\epsfig{file=jaadpt, width=4cm,height=5cm,angle=270}
\epsfig{file=paadpt, width=4cm,height=5cm,angle=270}
\epsfig{file=uaadpt, width=4cm,height=5cm,angle=270}
\end{center}
\caption{Transverse momentum distributions for the inclusive diffractive $J/\Psi$, $\Psi(2S)$ and $\Upsilon(1S)$ photoproduction at central rapidities($y = 0$) in $PbPb$ collisions at $\sqrt{s}$ = 5.02 TeV.}
\label{fig9}
\end{figure}
In Figs.\ref{fig7}, \ref{fig8} and \ref{fig9}, we present the predictions of the transverse momentum distribution for the inclusive diffractive $J/\Psi$, $\Psi(2S)$ and $\Upsilon(1S)$ photoproduction at central rapidities ($y$ = 0) in $pp$, $pPb$ and $PbPb$ collisions at $\sqrt{s} = 5.02$ TeV. As expected that the transverse momentum distributions of $\Upsilon(1S)$ are smaller than that of $J/\Psi$ and $\Psi(2S)$, due to the fact that the mass of $\Upsilon(1S)$ is larger than charmonium. Throughout the calculations, we find that the quark involved three subprocesses make a significant contributions to the heavy quarkounium inclusive diffractive transverse momentum distributions, since the percentage of the contributions from these three subprocesses can reach to $6\%$.
\section{Summary}
\label{sec:sum}
We extend our previous model, which is based on the NRQCD factorization formlism, to include quark involved three subprocesses to study the heavy quarkonium production. We call this approach as quark improved NRQCD model. The quark improved NRQCD model is use to calculate the total cross section of inclusive $J/\Psi$ production at HERA. By comparing our theoretical calculations of the $J/\Psi$ production with the experimental data, it finds that the contributions from the the three subprocesses mentioned above play a significant role in an accurate description of the heavy quarkonium production. Then we use the quark improved NRQCD model to study heavy quarkonium photoproduction at the LHC energies by combining with the resolved pomeron model. We make the predictions of the inclusive diffractive rapidity and transverse momentum distributions of $J/\Psi$, $\Psi(2S)$ and $\Upsilon(1S)$ in $pp$, $pPb$ and $PbPb$ collisions. We calculate the percentage of the quark involved three subprocesses in heavy quarkonium photoproduction, it finds that their contributions can reach to $8\%$ and $6\%$ at LHC energies, respectively. These results show that the quark involved three subprocesses may offer an insight to understand the underlying mechanism of heavy quarkonium production.
\begin{acknowledgments}
This work is supported by the National Natural Science Foundation of China under Grant Nos.11765005, 11305040, 11947119 and 11847152; the Fund of Science and Technology Department of Guizhou Province under Grant Nos.[2018]1023, and [2019]5653; the Education Department of Guizhou Province under Grant No.KY[2017]004; the National Key Research and Development Program of China under Grant No.2018YFE0104700, and Grant No.CCNU18ZDPY04.
\end{acknowledgments}
|
\section{Introduction}
Road casualties have been described as a global epidemic, representing the leading cause of death among young people worldwide (see, e.g. \textcite{mackay_traffic_1972, nantulya_neglected_2002}).
Car crashes and other types of collisions are responsible for more than 1 million deaths each year (1,250,000 in 2015, 17 deaths per 100,000 people), as reported by \textcite{worldhealthorganization_global_2018}.
In high income nations, such as the United Kingdom (UK), the roads are safer than the global average, but car crashes are still the cause of untold suffering.
According to the statistics published by the UK's Department for Transport (DfT) in the \textit{Annual report on Road Casualties in Great Britain} \parencite{dft_report}, approximately 153,000 road traffic collisions resulting in casualties were recorded in 2019, 5\% lower than 2018 and the lowest level since records began, in 1979.
Nevertheless, the DfT estimates that approximately 33,648 people were killed or seriously injured (KSI) in 2019, and while this number is slightly lower than in 2018, the decaying rate has been getting lower and lower starting from 2010.
These figures are worrisome considering that car occupant fatality rates are particularly high in the 17-24 age band \parencite[p. 17]{dft_report}.
To tackle the flattening trend in the KSI rate over the past decades, a range of interventions are needed, and analytical approaches can help prioritise them.
This paper presents a statistical model to identify street sections with anomalously high car crashes rates, and to support police responses and cost-effective investment in traffic calming measures \parencite{pacts_roads_2020}.
Statistical models of road crashes have become more advanced over time.
In the 1990s, models tended to consider only the discrete and heterogeneous nature of the data \parencite{miaou1993modeling, miaou1994relationship, shankar1995effect}, omitting spatial characteristics.
More recent statistical models of crash data include consideration of crash location in two-dimensional space, with three main advantages for road safety research \parencite{el2009urban}.
First, consideration of space allows estimating appropriate measures of risk (such as expected counts, rates or probabilities) at different levels of resolution and the subsequent ranking of geographical areas to support local interventions.
Second, spatial dependence can be a surrogate for unknown, potentially unmeasured (or unmeasurable) covariates; adjusting for geographic location can reduce model misspecification \parencite{dubin1988estimation, cressie1993statistics}.
Third, the spatial dimension can be used to take advantage of autocorrelation in the relevant variables, borrowing strength from neighbouring sites and improving model parameter estimation.
Road crash datasets, which are typically available on a single accident basis, can be spatially aggregated in two main ways: administrative zones (such as cantons, census wards, or regions) or street network features (either as contiguous segments or divided into corridors and intersections).
In both cases, the spatial support is a lattice, i.e. a countable collection of geometrical units (polygons or lines, respectively), possibly supplemented by a neighbourhood structure.
Several papers addressed the statistical modelling of crash frequencies at the areal level, based on available zoning systems in the study region (see, e.g., \textcite{miaou2003roadway, aguero2006spatial, noland2004spatially, boulieri2017space}).
The second approach has gained in popularity in recent years, with a number of papers analysing road crash events aggregated to the street level (see, e.g., \textcite{miaou2005bayesian, aguero2008analysis, wang2009impact}), as detailed in several review papers \parencite{lord2010statistical, savolainen2011statistical, ziakopoulos2020review}.
In reference to street level data, we note that there has been a recent surge of research for spatial point patterns living on networks \parencite{rakshit2019fast, cronie2020inhomogeneous, baddeley2020analysing}.
Both zone and network level approaches have advantages, notably computational requirements for the former and spatially disaggregated results for the latter.
Given that computational resources are less of a constraint in the 2020s than they were in previous decades, and the fact that it is the nature of roads (not zones) that is responsible for crashes, we argue that road segments are the more appropriate aggregation units for the analysis of road crash data.
Network analysis can be used to bring attention to specific segments, and, for these reasons, the models presented in the next sections were developed considering a network lattice.
Aggregation, e.g. number of crashes per road segment, enables comparison between different road segments.
However, spatial aggregation also leads to a well-known problem in geographical analysis, the Modifiable Areal Unit Problem (MAUP), firstly described in \textcite{openshaw1981modifiable}: the size of the spatial units impacts on the statistical analysis, influencing, and possibly biasing, modelling choices and results.
Hence, conclusions drawn at one scale of spatial aggregation might not necessarily hold at another scale or be somehow different.
The MAUP has been mainly ignored in the road safety literature and, as reported by \textcite{xu2018modifiable} and \textcite[p. 21]{ziakopoulos2020review}, it is mentioned only in a handful of recent papers \parencite{ukkusuri2012role, abdel2013geographical, zhai2019influence, briz2019investigation}, which explore the impact of changing the areal zoning system (e.g. TAZ, block groups and census tracts) on parameter estimates, significance and hotspot detection.
Only one early paper \parencite{thomas1996spatial} could be found exploring the impacts of the MAUP on road crash data, albeit only in terms of summary statistics of aggregated counts.
To assess the MAUP effect on network data modelling, we employed an algorithm to modify the structure of a road network, merging contiguous segments in the same corridor preserving the geometrical properties of the network \parencite{R-dodgr}.
Then, we compared the results obtained with the two different network configurations.
To the best of our knowledge, this is the first attempt at exploring and estimating the presence and the magnitude of MAUP in models that consider a network lattice.
Finally, we note that systems of collision classification present a multivariate nature \parencite{kirk_implications_2020}.
The occurrences of different severity degrees can be correlated to each other, and their spatial dynamics can be potentially interdependent.
Hence, it is necessary to account for correlations between crashes counts at different levels of severity.
We consider two types of accidents: \textit{slight} and \textit{severe}.
The severe class is very sparse in the dataset at hand, hence modelling both types of accidents simultaneously allows to borrow strength from the existing correlations and improves estimates.
We underline that the methodology for classifying the severity level of a car crash in the UK has been modified starting from 2016, adopting the injury-based systems called \texttt{CRASH} and \texttt{COPA} \parencite{CRASH}.
All police forces are gradually adopting these new reporting systems in England, and the Office for National Statistics (ONS) developed a logistic regression model to correct the severity levels between different years and classification systems.
The data used in this paper have been adjusted using the procedures developed by ONS \parencite[p. 38-41]{dft_report}.
Following ideas introduced in \textcite{barua2014full}, we consider a range of competing models, developed in a full hierarchical Bayesian paradigm.
This approach allows one to encompass complex structures of spatial dependence in a quite natural way.
Spatially structured random effects are defined using both Intrinsic Multivariate Conditional Auto-regressive (IMCAR) and Proper Multivariate Conditional Auto-regressive (PMCAR) priors \parencite{besag1974spatial, mardia1988multi, martinez2019disease, palmiperales2019bayesian}.
\textcite{lord2010statistical, savolainen2011statistical, ziakopoulos2020review} review other examples of multivariate car crashes models on a network.
The case study is the metropolitan area of Leeds (population 800,000) in North England.
We accessed Ordnance Survey data on major roads (3661 segments, total length 450 km), creating a spatial network substantially larger than previous studies, many of which report findings on only a few roads, with \textcite{borgoni2020assessing} representing a notable exception albeit with a simpler spatial structure.
We present results for an entire metropolitan area, approximating more closely the level at which road policing activities and investment in road safety interventions are prioritised.
The scale of the case study presented several computational challenges, and, in terms of Bayesian parameter estimation, we used the computationally efficient Integrated Nested Laplace Approximation (INLA) approach instead of Markov chain Monte Carlo (MCMC) sampling \parencite{INLA1, INLA2}.
The rest of the paper is organised as follows.
In Section~\ref{sec:data}, the data sources are described.
In Section~\ref{sec:methodology}, the statistical methodology adopted in this paper is discussed in detail.
In Section~\ref{sec:results}, the main results of the paper are presented whereas, model criticism and further model discussion, such as MAUP analysis, are provided in Section~\ref{sec:criticism}.
Conclusions, in Section~\ref{sec:conclusions}, end the paper.
\section{Data}
\label{sec:data}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/msoav2.eps}
\caption{The grey polygons show the MSOAs in West-Yorkshire region, while the dark-green area highlights the city of Leeds. The inset map locates the position of the study-area with respect to England.}
\label{fig:msoa}
\end{figure}
The datasets analysed in this paper came from several different sources and required a number of preprocessing steps before they could be made into a structure suitable for a statistical analysis.
The study region was defined as the Middle Super Output Area (MSOA) zones within the local authority of Leeds.
The City of Leeds was selected because it is a car-dependent city with a large network of major roads that approach the city centre (the city was dubbed the 'motorway city of the 70s') and would therefore be expected to be a place where road safety could be improved.
Leeds is part of West Yorkshire and accounts for approximately 40\% of all car crashes in the region.
Origin-destination data from the 2011 UK Census were used to estimate traffic volumes, to provide an estimate of exposure, with traffic volumes used as part of the denominator of the statistical models presented in Section~\ref{sec:methodology}.
The road network was obtained from Ordnance Survey, covering all major roads in Leeds.
We matched the network and the MSOAs using an overlay operation.
We associated all car crashes that occurred in the city of Leeds from 2011 to 2018 with the nearest point on the road network, counting the occurrences in each street segment.
Finally, a set of socio-economic variables obtained from 2011 UK Census data were included in the statistical models as fixed effects.
\subsubsection*{MSOA zones}
There are 6,791 MSOAs in England, 299 of which belong to the West Yorkshire region and 107 of which constitute Leeds.
These were accessed from the github-page\footnote{URL: \url{https://github.com/npct/pct-outputs-regional-R/tree/master/commute}, last access on 06/2020} of Propensity Cycle Tool \parencite{lovelace_propensity_2017}.
The MSOAs represent the starting point for all the following steps, and they are mapped in Figure~\ref{fig:msoa} as grey polygons for the West-Yorkshire, and as dark-green polygons for the City of Leeds.
The inset map is used to locate the study-area in the British territory.
\subsubsection*{Traffic flow}
The \textit{traffic flow} data represent the \textit{commuting journeys} from home to workplace using several modes of transport, such as train, bus, bike and motorcycle.
The data were collected during the 2011 Census at the individual level, and then aggregated at the MSOA level.
The UK Data Service shares the flow data through the \texttt{WICID} interface as cross-tables reporting the flows between all pairs of a predefined set of MSOAs \parencite{WICID}.
We considered the commuting flows in the region of Leeds for all possible modes of transport.
Figure~\ref{fig:raw_flows} shows a random sample of $1000$ \textit{traffic flows} (out of 10,536 in total) between the centroids of the MSOAs in Leeds, coloured according to the number of daily commuters.
Raw WICID data, however, ignore that people may travel to their workplace through several MSOAs.
For this reason, we calculated a new traffic measure using the following procedure.
Starting from the MSOAs, we defined a graph where the vertices are the centroids of each area, and the edges connect neighbouring areas.
Then, we estimated the shortest path for all commuting journeys downloaded from WICID and assigned to each MSOA a value that is equal to the number of all raw traffic measures going through the area.
These values represent the new traffic measures and are displayed in Figure~\ref{fig:new_flows}.
A similar approach was also adopted in \textcite{boulieri2017space}, and we refer to the references therein for more details.
The raw data flows and the MSOAs polygons were downloaded using the R package \texttt{pct} \parencite{pct}.
\subsubsection*{Road network}
\begin{figure}
\centering
\subfloat[Random sample of one thousand Origin-Destination data representing the daily flows between MSOAs in Leeds. \label{fig:raw_flows}]{
\includegraphics[width=0.47\linewidth]{figures/flowsv2.eps}
}
\hspace{0.01\linewidth}
\subfloat[Estimates of the new traffic measure. The white star represents Leeds City Centre. \label{fig:new_flows}]{
\includegraphics[width=0.47\linewidth]{figures/trafficv2.eps}
}
\caption{Raw and modified traffic flows in the area of Leeds. The map on the right highlights several contiguous MSOAs that correspond to the arterial thoroughfares that are used to reach the City Centre.}
\label{fig:flows}
\end{figure}
The \textit{road network} was built using data downloaded from Ordnance Survey (OS), an agency that provides digital maps and other services for location-based products \parencite{OS}.
We downloaded the \texttt{Vector OpenMap Local} data for a geographical region covering Leeds, selected the \texttt{Roads} and \texttt{Tunnels} layers, and filtered the streets that belong to the city.
Ordnance Survey represents all the streets of a road network as the union of a finite set of segments, and it includes additional fields such as the road name or the street type.
These segments represent the elementary units for the statistical analysis described in Section~\ref{sec:methodology}.
The road network downloaded from OS is composed of approximately 50,000 segments.
The OS network was simplified using the following procedure.
We first selected only the major roads, such as the \textit{Motorways}, \textit{Primary Roads} and \textit{A roads}.
They represent 3,668 segments, i.e. less than 10\% of the total road network, but more than 50\% of all car crashes registered during 2011-2018 occurred in their proximity.
The output of this procedure is a road network composed by a big cluster of connected streets, displayed in Figure~\ref{fig:show_network_crashes}, and several isolated segments of small groups of road segments (which are also called \textit{islands}), created by the exclusion of their links to the other roads.
These small clusters can be problematic from a modelling perspective since they produce a not-fully-connected network (see \textcite{hodges2003precision, FRENISTERRANTINO201825} and the properties of ICAR and MICAR distributions explained in Section~\ref{sec:methodology}), so we implemented an algorithm to further simplify the road network and remove them.
This algorithm is based on the dual representation of a road network as a geographical entity, composed by points and lines, and a graph object, with nodes and edges \parencite{porta2006network, marshall2018street, gilardi_lovelace_padgham_2020}.
More precisely, we created a graph whose vertices correspond to the street segments of the road network, and we defined an edge for each pair of spatial units sharing a point at their boundaries.
This graph uniquely determines a (sparse) adjacency matrix amongst the spatial units (i.e. the road segments) that summarises the graph dimension of the road network.
We sketched a toy example in Figure~\ref{fig:W_matrix}, representing the idea behind the dual representation of a road network and the definition of the adjacency matrix.
\begin{figure}
\begin{minipage}[c]{0.35\textwidth}
\centering
\includegraphics[scale=.35]{figures/W_matrix} \end{minipage}
\hspace{0.1\textwidth}
\begin{minipage}[c]{.55\textwidth}
\[
\begin{blockarray}{ccccccc}
& \textcolor{col1}{1}
& \textcolor{col2}{2}
& \textcolor{col3}{3}
& \textcolor{col4}{4}
& \textcolor{col5}{5}
& \textcolor{col6}{6} \\
\begin{block}{c(cccccc)}
\textcolor{col1}{1}
& \cdot & \cdot & \cdot & \cdot & 1 & 1 \\
\textcolor{col2}{2} & \cdot & \cdot & 1 & 1 & \cdot & \cdot \\
\textcolor{col3}{3} & \cdot & 1 & \cdot & \cdot & \cdot & 1 \\
\textcolor{col4}{4} & \cdot & 1 & \cdot & \cdot & 1 & \cdot \\
\textcolor{col5}{5} & 1 & \cdot & \cdot & 1 & \cdot & 1 \\
\textcolor{col6}{6} & 1 & 1 & 1 & \cdot & 1 & \cdot \\
\end{block}
\end{blockarray}
\]
\end{minipage}
\caption{Graphical example showing the dual nature of a road network. Left: map showing the geographical dimension. Each segment is coloured and labelled using a different ID and colour respectively. Right: adjacency matrix of the graph associated with the road network. Each vertex corresponds to a segment whereas an edge connects two vertices if they share one boundary point. For example, segments $1$ and $2$ are not neighbours since they do not share any point at the boundaries, even if they intersect each other. This situation may occur at bridges or overpasses.}
\label{fig:W_matrix}
\end{figure}
Using the graph and the adjacency matrix, we excluded all road segments that did not belong to the main cluster.
In particular, we dropped 7 segments spread across different parts of the city with a total length of approximately 380m (out of 450km).
That also implied that we removed from the analysis 3 car crashes that occurred in those segments (see below).
However, given the extremely small fraction of records discarded from the dataset, we expect that excluding these observations does not influence the final results.
It should be stressed that this procedure creates a fully connected network, which has relevant benefits on the rank-deficiency problem of the ICAR models described in Section~\ref{sec:methodology}.
In the end, the road network is composed of 3661 units, and it is shown in Figure~\ref{fig:show_network_crashes}, where the segments are coloured according to their road types.
The segments have different lengths, ranging from 0.1m to 2597m, with an average value of 118m (\textit{sd} = 178m). As explained in the next Section, these lengths are used in the exposure parameter of the statistical models in order to guarantee a comparable rate amongst the network units.
Finally, since the street network and the MSOAs are spatially misaligned, they were matched using an overlay operation: each road segment was assigned to the MSOA that intersects the largest fraction of the segment.
This procedure allows us to assign a traffic estimate to each road segment, which will be used in the exposure parameter (along with the segments' lengths) in the statistical models considered below.
\subsubsection*{Road traffic collision data}
\begin{figure}
\centering
\subfloat[Road network and car crashes in Leeds. \label{fig:show_network_crashes}]{
\includegraphics[width=0.47\linewidth]{figures/network_crashesv2.eps}
}
\hspace{0.01\linewidth}
\subfloat[Choropleth map of severe counts. \label{fig:show_serious_counts}]{
\includegraphics[width=0.47\linewidth]{figures/severe_counts.eps}
}
\caption{The map on the left represents the road network in Leeds. Each segment is coloured according to its OS classification. The black dots represent the car crashes. On the right, we report a choropleth map displaying severe car crashes counts.}
\label{fig:network_crashes}
\end{figure}
We analysed all road traffic collisions that occurred between the 1st of January 2011 and the 31st of December 2018 in the MSOAs pertaining to the City of Leeds, which involved at least one car, personal injuries, occurred on public roads and became known to the Police forces within thirty days of the occurrence.
The geographical coordinates of the crashes are provided at ~10m or less resolution in the UK’s official coordinate reference system (the Ordnance Survey National Grid, EPSG code 27700) depending upon the year to which the data refer to \textcite{Dep4Trans2011}. Hence, we adopted this value as a threshold to account for the potential misalignment between the event locations and the network and excluded all those events that occurred farther than ten meters from the closest segment in the simplified road network since they might be related to other streets.
The data were downloaded from the UK's official road traffic casualty database, called \textit{STATS19}, using the homonym R package \parencite{stats19}.
As mentioned above, we focussed on car crashes that occurred in a public highway and involved personal injuries.
In general, however, there is no obligation for people to report all personal injury accidents to the police and, for this reason, a proportion of non-fatal and no-injury casualties remain unknown to the police \parencite{dft_report}.
Hence, we acknowledge that the counts of slight accidents considered hereinafter may suffer under-reporting to some extent.
We refer to \textcite[Sec 2.1]{savolainen2011statistical} and references therein for more details.
The final sample is composed of 5,862 events, and they are reported as black dots in Figure~\ref{fig:show_network_crashes}.
Then, we projected all crashes to the nearest point on the road network, and we counted the number of \textit{slight} or \textit{severe} occurrences on all street segments.
We decided to ignore the temporal dimension since \textit{severe} crashes counts present an extreme sparsity, with more than 80\% of zero counts during 2011-2018.
Moreover, 40\% of all segments registered no car crashes during the study period, while another 40\% reported two or more crashes.
These numbers highlight a common temporal trend between the eight years, and we refer the interested reader to the supplementary material for a space-time representation.
The map in Figure~\ref{fig:show_serious_counts} shows the spatial distribution of \textit{severe} crashes counts.
\subsubsection*{Socio-economic covariates}
The statistical models described in Sections~\ref{sec:methodology} and~\ref{sec:results} include two socio-economic covariates that were obtained from the 2011 UK Census data and downloaded from Nomis website\footnote{URL: \url{https://www.nomisweb.co.uk/}. Last access: 06/2021.}.
In particular, we consider the population density (given by the ratio of the number of inhabitants in a given region and its area in squared metres) and the employment rate (given by the ratio of employed population between 16 and 64 and total population between 16 and 64).
Both covariates were obtained at the LSOA level (a more detailed aggregation level than MSOA) and, for the same reasoning as before, they were matched with the street segments using an overlay operation.
\section{Statistical methodology}
\label{sec:methodology}
We first focus on the definition of a three-level hierarchical model structure which is shared among all the alternative specifications considered below.
Then, we introduce two baseline models that serve as benchmarks and starting points for the other specifications.
Thereafter, four different extensions to the baseline models are introduced.
Finally, some techniques used for model comparison are discussed.
The common theme behind all the seven alternatives is the presence of spatially structured and unstructured multivariate random effects.
Let $Y_{ij}, \ i = 1, \dots, n,\ $ represent the number of car crashes that occurred in the $i$-th road segment with severity level $j, \ j = 1, \dots, J$.
In this paper we consider two possible severity levels, a car crash being either \textit{severe}, $j = 1$, or \textit{slight}, $j = 2$.
In the first stage of the hierarchy, we assume that
\[
Y_{ij} | \lambda_{ij} \sim \text{Poisson} \left(E_i\lambda_{ij}\right),
\]
where $E_i$ is an exposure parameter and $\lambda_{ij}$ represents the car crashes rate in the $i$th road segment for severity level $j$.
As mentioned in the previous section, the exposure parameter, $E_i$, is given by the product of two quantities \parencite{wang2009impact}: the segment's length and the estimate of traffic flow (see Section~\ref{sec:data}).
The exposure accounts for the fact that a longer street segment has a higher collision risk than a shorter one, guaranteeing that comparable rates amongst units (segments) are preserved.
At the same time, the traffic flow estimates allow different segments of the network to be “weighted” differently, being more exposed to accidents those segments with a higher traffic flow, all the rest being fixed.
At the second stage of the hierarchical model, a log-linear structure on $\lambda_{ij}$ is specified.
We assume that
\[
\log\left(\lambda_{ij}\right) = \beta_{0j} + \sum_{m = 1}^{M}\beta_{mj}X_{ijm} + \theta_{ij} + \phi_{ij},
\]
where $\beta_{0j}$ represents a severity-specific intercept, $\left\lbrace\beta_{mj}\right\rbrace_{m = 1}^{M}$ is a set of coefficients, $(X_{ij1}, \dots, X_{ijM})$ is a collection of $M$ covariates, $\phi_{ij}$ is a spatially structured random effect and $\theta_{ij}$ represents a normally distributed error component.
The third stage that completes the hierarchical model is the specification of prior and hyperprior distributions. We assigned a vague $N(0, 1000)$ prior to $\beta_{mj}, \ m = 0, \dots M$.
The two random effects, namely $\theta_{ij}$ and $\phi_{ij}$, represent the unstructured and structured spatial components and are defined differently in different models as discussed below.
Hereafter, we follow the notation used in \textcite{martinez2019disease}.
\subsection{Baseline models: independent spatial and unstructured effects}
The two baseline models are defined considering multivariate spatial and unstructured random effects with independent components.
More precisely, a bivariate Gaussian prior with independent components is assigned to $\left(\theta_{i1}, \theta_{i2}\right)$ for both baseline models:
\begin{equation}
\left(\theta_{i1}, \theta_{i2}\right)\sim N_2\left(\bm{0}, \begin{bmatrix}
\sigma^{2}_{\theta_{1}} & 0 \\
0 & \sigma^2_{\theta_{2}}
\end{bmatrix}\right), \ \quad i = 1, \dots, n.
\label{eq:def_indep_iid}
\end{equation}
We assigned a Gamma hyperprior with parameters $1$ (shape) and $0.00005$ (inverse scale) to the inverse of $\sigma^2_{\theta_{1}}$ and $\sigma^2_{\theta_{2}}$, i.e. the precisions.
The spatially structured term in the first baseline model was defined using an \textit{Independent Intrinsic Multivariate Conditional Auto-regressive} (IIMCAR) prior, whereas, for the second model, we adopted an \textit{Independent Proper Multivariate Conditional Auto-regressive} (IPMCAR) prior.
The IIMCAR and IPMCAR distributions are briefly introduced hereafter, starting from their classical univariate counterparts, namely the ICAR and PCAR distributions.
Univariate spatial random effects are traditionally modelled using a prior that belongs to the family of \textit{Conditional Auto-regressive} (CAR) distributions \parencite{besag1974spatial}.
Given a random vector $\bm{\phi} = (\phi_1, \dots, \phi_n)$, the \textit{Intrinsic Conditional Auto-Regressive} (ICAR) distribution, which is a particular case of the CAR family, is usually defined through a set of conditional distributions \parencite{besag1995conditional}:
\begin{equation}
\phi_i | \lbrace \phi_{i'}, i' \in \partial_i \rbrace; \sigma^2 \sim N\left(m_i^{-1}\sum_{i' \in \partial_i}\phi_{i'}, \frac{\sigma^2}{m_i}\right), \ i = 1, \dots, n,
\label{eq:def_conditional_ICAR}
\end{equation}
where $\partial_i$ and $m_i$ denote, respectively, the indices and the cardinality of the set of neighbours for spatial unit $i$.
These quantities are defined through a sparse binary symmetric neighbourhood matrix $\bm{W}$ with dimensions $n \times n$ that summarises the spatial relationships in the region of study.
We built it taking advantage of the dual representation of a road network as a spatial and a graph object (see \textcite{porta2006network} and Section~\ref{sec:data}).
More precisely, $\bm{W}$ is the adjacency matrix of a graph whose vertices correspond to the street segments of the road network and the edges identify a shared point at the boundaries of two spatial units.
This procedure defines a \textit{First Order} neighbourhood matrix.
\textit{Second} and \textit{Third Order} neighbourhood matrices are defined iteratively in the same way.
It is possible to prove that the prior defined by~\eqref{eq:def_conditional_ICAR} suffers from rank-deficiency problems, that are usually fixed by imposing a set of sum-to-zero constraints on the vector $\bm{\phi}$, one for each group of connected segments in the graph of the road network \parencite{hodges2003precision}.
In this paper, we deal with a fully connected road network (see the pre-processing procedures detailed in Section~\ref{sec:data}), so we always had to fix only one set of constraints.
The \textit{Proper Conditional Auto-Regressive} (PCAR) distribution is another member of the CAR family and it is usually defined as follows:
\begin{equation}
\phi_i | \lbrace \phi_{i'}, i' \in \partial_i \rbrace; \sigma^2, \rho \sim N\left(\rho\left(m_i^{-1}\sum_{i' \in \partial_i}\phi_{i'}\right), \frac{\sigma^2}{m_i}\right), \ i = 1, \dots, n,
\label{eq:def_conditional_PCAR}
\end{equation}
where $\partial_i$ and $m_i$ are defined as for the ICAR distribution and $\rho$ is a parameter controlling the strength of spatial dependence, usually called \textit{spatial autoregression coefficient} \parencite{cressie1993statistics}.
It is possible to prove that the joint distribution defined by~\eqref{eq:def_conditional_PCAR} is proper if $\abs{\rho} < 1$, hence there is no need to set any sum-to-zero constraint in this case.
The ICAR prior can be seen as a limit case of the PCAR distribution with $\rho \to 1$, analogously to the relationship between Auto-Regressive and Random-Walk models in time series models \parencite{botella2013spatial}.
The family of \textit{Multivariate Conditional Auto-regressive} (MCAR) distributions was firstly introduced by \textcite{mardia1988multi}, extending the ideas of \textcite{besag1974spatial} to the multivariate case.
Given a random matrix $\bm{\Phi} = (\phi_{ij})$, which is defined for $i = 1, \dots, n$ units and $j = 1, \dots, J$ levels, the \textit{Intrinsic Multivariate Conditional Auto-regressive} (IMCAR) distribution is a particular case of the MCAR family, defined through a set of multivariate conditional distributions \parencite{martinez2019disease}:
\begin{equation}
\bm{\Phi}_{i\cdot} | \text{vec}\left(\bm{\Phi}_{-i\cdot}\right); \Omega \sim N_{J}\left(m_i^{-1}\sum_{i' \in \partial_i}\bm{\Phi}_{i'\cdot}^{T}; m_i^{-1}\Omega^{-1}\right).
\label{eq:def_conditional_IMCAR}
\end{equation}
The terms $\bm{\Phi}_{i\cdot}$ and $\bm{\Phi}_{-i\cdot}$ denote, respectively, the $i$th row of $\bm{\Phi}$ and the matrix obtained by excluding the $i$th row from $\bm{\Phi}$.
The $\text{vec}$ operator is used for row-binding the columns of a matrix, meaning that $\text{vec}\left(\bm{\Phi}_{-i\cdot}\right) = \left(\bm{\Phi}_{-i1}^{T}, \dots, \bm{\Phi}_{-iJ}^{T}\right)^{T}$.
The elements $m_i$ and $\partial_i$ are defined as before, through the adjacency matrix $\bm{W}$ of the graph associated to the road network, and they represent the spatial dimension of the IMCAR distribution.
The $J \times J$ precision matrix $\Omega$ is used to model the associations between pairs of levels in the same road segment $i$, and it acts as a multivariate extension of the parameter $\sigma^2$ in equation~\eqref{eq:def_conditional_ICAR}.
This distribution suffers from the same rank-deficiency problems as its univariate counterpart, which are usually solved by imposing appropriate sum-to-zero constraints.
The number of restrictions is equal to the number of clusters in the graph of the road network times the number of levels in the multivariate setting.
The pre-processing operations that we performed on the network data (see Section~\ref{sec:data}) imply that we always have to set only $J$ sum-to-zero constraints.
The IIMCAR distribution is a particular case of~\eqref{eq:def_conditional_IMCAR}, which is obtained by setting $\Omega^{-1} = \text{diag}(\sigma^2_{\phi_1}, \dots, \sigma^2_{\phi_J})$.
More precisely, if we assume $J = 2$ as we do in this paper, then IIMCAR is defined by the following set of multivariate conditional distributions:
\begin{equation}
\bm{\Phi}_{i\cdot} | \text{vec}\left(\bm{\Phi}_{-i\cdot}\right); \sigma_{\phi_1}^2, \sigma_{\phi_2}^2 \sim N_{2}\left(m_i^{-1}\sum_{i' \in \partial_i}\bm{\Phi}_{i'\cdot}^{T}; m_i^{-1}
\begin{bmatrix}
\sigma_{\phi_1}^2 & 0 \\
0 & \sigma_{\phi_2}^2
\end{bmatrix}\right).
\label{eq:def_conditional_IIMCAR}
\end{equation}
In equation~\eqref{eq:def_conditional_IIMCAR} we are assuming independence between the $2$ levels, and this implies that the IIMCAR distribution is equivalent to two independent ICAR distributions, one for each level.
Analogously to the univariate case, the \textit{Proper Multivariate Conditional Auto-regressive} (PMCAR) distribution is a particular case of the MCAR family characterised by the following set of multivariate conditional distributions:
\begin{equation}
\bm{\Phi}_{i\cdot} | \text{vec}\left(\bm{\Phi}_{-i\cdot}\right); \rho, \Omega \sim N_{J}\left(m_i^{-1}\rho\sum_{i' \in \partial_i}\bm{\Phi}_{i'\cdot}^{T}; m_i^{-1}\Omega^{-1}\right)
\label{eq:def_conditional_PMCAR}.
\end{equation}
The strength of the spatial dependence is controlled by $\rho$ (as for the univariate PCAR distribution) and all the other parameters are defined as before.
It can be proved that the joint distribution defined by equation~\eqref{eq:def_conditional_PMCAR} is proper if $\abs{\rho} < 1$, although we restricted ourself to $\rho \in (0, 1)$ to avoid some counter-intuitive behaviour of the PMCAR distribution \parencite{wall2004close, miaou2005bayesian}.
The IPMCAR distribution is defined as a particular case of equation~\eqref{eq:def_conditional_PMCAR} with $\Omega^{-1} = \text{diag}(\sigma^2_{\phi_1}, \dots, \sigma^2_{\phi_J})$.
More precisely, if we assume $J = 2$, then IPMCAR is defined through the following set of multivariate conditional distribution:
\begin{equation}
\bm{\Phi}_{i\cdot} | \text{vec}\left(\bm{\Phi}_{-i\cdot}\right); \rho, \sigma^2_{\phi_1}, \sigma^2_{\phi_2} \sim N_{2}\left(m_i^{-1}\rho\sum_{i' \in \partial_i}\bm{\Phi}_{i'\cdot}^{T}; m_i^{-1}\begin{bmatrix}
\sigma^2_{\phi_1} & 0 \\
0 & \sigma^2_{\phi_2}
\end{bmatrix}\right).
\label{eq:def_conditional_IPMCAR}
\end{equation}
For the same reasoning as in equation~\eqref{eq:def_conditional_IIMCAR}, the IPMCAR distribution is equivalent to $J$ independent PCAR distributions.
Now we can characterise the random effects for the two baseline models.
The first model was defined by considering unstructured random effects with a bivariate independent Gaussian prior \eqref{eq:def_indep_iid}, and spatial random effects with an IIMCAR prior \eqref{eq:def_conditional_IIMCAR}.
The second one was defined analogously to the first baseline model, but assuming an IPMCAR distribution for the spatial random effects \eqref{eq:def_conditional_IPMCAR}.
These models assume independence between the two levels both in the spatial and unstructured components, so they were used as benchmarks.
In the next sections, we will also refer to the two baseline models using, respectively, the codes (A) and (B).
We assigned an improper prior to $\sigma^2_{1}$ and $\sigma^{2}_{2}$, the variances in $\Omega$, defined on $\mathbb{R}^{+}$, and a $\text{Uniform}(0, 1)$ prior to $\rho$.
Hereafter we introduce two increasingly complex sets of extensions that generalise the baseline models.
The first one is characterised by the removal of the independence assumption from the spatially structured random effects, whereas, in the second set of extensions, we also relax the independence assumption from the unstructured random effects.
\subsection{Model extensions}
\subsubsection*{First set of extensions}
Starting from the baselines, we defined two new models replacing the IIMCAR and IPMCAR priors with their non-independent multivariate counterparts, the generic IMCAR and PMCAR defined above.
If we assume $J = 2$, then the variance-covariance matrix $\Omega^{-1}$ in \eqref{eq:def_conditional_IMCAR} and \eqref{eq:def_conditional_PMCAR} can be written as
\[
\Omega ^ {-1} = \begin{bmatrix}
\sigma^2_{\phi_1} & \rho_{\phi}\sigma_{\phi_1}\sigma_{\phi_2} \\
\rho_{\phi}\sigma_{\phi_1}\sigma_{\phi_2} & \sigma^2_{\phi_2}
\end{bmatrix},
\]
where $\sigma_1^2$ and $\sigma_2^2$ represent the conditional variances and $\rho_{\phi}$ represents the correlation coefficient between the two levels in the same spatial unit.
These models represent a generalisation of the baselines since we are now taking into account the correlations between different levels in the same road segment.
We will also refer to them using, respectively, the codes (C) and (D).
Following \textcite{palmiperales2019bayesian}, we assigned a Wishart hyperprior to $\Omega^{-1}$ with parameters $2$ and $\bm{I}_{2}$, i.e. the identity matrix of size two.
The prior distributions on the unstructured random effects were left unchanged with respect to the baselines.
\subsubsection*{Second set of extensions}
In these models, the independence assumption of the spatially unstructured random effects is removed.
More precisely, assuming $J = 2$, we assign a generic bivariate Gaussian prior to the unstructured random effects:
\[
\left(\theta_{i1}, \theta_{i2}\right)\sim N_2\left(\bm{0}, \begin{bmatrix}
\sigma^{2}_{\theta_{1}} & \rho_{\theta}\sigma_{\theta_{1}}\sigma_{\theta_{2}} \\
\rho_{\theta}\sigma_{\theta_1}\sigma_{\theta_2} & \sigma^2_{\theta_{2}}
\end{bmatrix}\right).
\]
Parameters $\sigma_{\theta_{1}}^2$ and $\sigma_{\theta_{2}}^2$ represent the marginal variances of the unstructured random error, whereas $\rho_{\theta}$ represents their correlation.
These models will also be identified using the codes (E) and (F).
We assigned a Wishart hyperprior to the variance-covariance matrix with parameters $2$ and $\bm{I}_2$.
Finally, following the ideas introduced in \textcite{gelfand2003proper}, we also tested an extension of model (F), named model (G), which is characterised by a generalisation of the PMCAR distribution that introduces a separate spatial autoregression coefficient, $\rho_{j}$, for each level in $\bm{\Phi}$.
We found that this extension did not improve over model (F).
Hence, we will not add more details here and refer the interested reader to the supplementary materials.
The prior distributions adopted for the random effects in the two baselines and their extensions are summarised in Table~\ref{tab:summary_models}.
We also included the IDs that will be used to identify each model in subsequent Tables and Sections.
\begin{table}
\centering
\caption{Summary of the prior distributions assigned to the random effects in the models introduced in Section~\ref{sec:methodology}.}
\begin{tabular}{llll}
\toprule
ID & Model & Unstructured Effect & Spatial effect\\
\midrule
(A) & Baseline 1 & Independent Gaussian & Independent IMCAR \\
(B) & Baseline 2 & Independent Gaussian & Independent PMCAR \\
(C) & Extension 1 - Model 1 & Independent Gaussian & IMCAR \\
(D) & Extension 1 - Model 2 & Independent Gaussian & PMCAR \\
(E) & Extension 2 - Model 1 & Correlated Gaussian & IMCAR \\
(F) & Extension 2 - Model 2 & Correlated Gaussian & PMCAR \\
\bottomrule
\end{tabular}
\label{tab:summary_models}
\end{table}
\subsection{Model comparison}
The models proposed in the previous paragraphs were compared using Deviance Information Criterion (DIC) \parencite{spiegelhalter2002bayesian} and Watanabe–Akaike Information Criterion (WAIC) \parencite{gelman2014understanding, watanabe2010asymptotic}.
These criteria represent a measure for the adequacy of a model, penalised by the number of effective parameters.
In both cases, the lower is the value of the index, the better is the fitting of the model.
\section{Results}
\label{sec:results}
We estimated the models previously described using the software \texttt{INLA} \parencite{INLA1, INLA2, gomez2020bayesian}, interfaced through the homonymous R package \parencite{RSoftware}.
We used the \textit{Simplified Laplace} strategy for approximating the posterior marginals and the \textit{Central Composite Design} strategy for determining the integration points.
The code behind the definition of multivariate ICAR and PCAR random effects is defined in the package \texttt{INLAMSM} \parencite{palmiperales2019bayesian}.
It took approximately 30 - 45 minutes to estimate each model using a virtual machine with an Intel Xeon E5-2690 v3 processor, six cores, and 32GB of RAM.
\subsubsection*{Fixed effects}
The models listed in Table~\ref{tab:summary_models} share a common structure for the fixed effect component, with a severity-specific intercept and a set of covariates representing some social and physical characteristics of the road segments.
In particular, we considered five severity-specific covariates, namely the two socio-economic variables mentioned in Section~\ref{sec:data}, a dummy variable recording whether a road segment lays in a dual carriageway street or not, the road type (either \textit{Motorway}, \textit{Primary Road} or \textit{A Road}, according to OS definition), and the edge betweenness centrality measure, which reflects the number of shortest paths traversing each segment \parencite{kolaczyk2014statistical}.
The last covariate can be considered as a proxy for the average vehicle miles traveled (VMT) \parencite{briz2019estimating}, the population density can be thought of as a surrogate for pedestrian and cycling volumes, and the employment rates may represent an indirect measure of income.
Finally, to improve the stability of INLA algorithms, we scaled all numerical variables to zero mean and unit variance.
Table~\ref{tab:fixed_effects} shows the posterior means and standard deviations for the fixed effects.
We first notice that the estimates are stable among the models.
The intercept for severe car crashes, $\beta_{01}$, is found slightly smaller than $\beta_{02}$.
This is not surprising since severe accidents are rarer than the slight ones.
The coefficients of edge betweenness centrality measures are found close to zero for all models, and their 95\% credible interval (not reported in the table) always include the value zero.
The Road type parameters represent relative differences with respect to the reference category (i.e. \textit{A Roads}), hence \textit{Motorways} are found less prone to severe and slight car crashes than \textit{A roads}.
A similar finding was also reported by \textcite{boulieri2017space} for UK data.
An analogous interpretation applies to \textit{Primary Roads}.
In agreement with other studies (see, for instance \parencite{HUANG201710} and references therein), we found that population density significantly correlates to both slight and severe accidents, although its effect is stronger for the former type.
On the other hand, the second socio-economic variable was not found significant in any of the outcomes considered in this paper.
Unfortunately, more direct income measures or poverty indicators are not available and are difficult to construct at such detailed spatial resolution.
More work is needed to to assess the relationship between road safety and socio-economic inequalities.
Finally, dual carriageway roads have been found significantly less prone to slight car accidents, whereas no impact has been found for severe car crashes.
An analogous result was also reported by \textcite{star-rating-UK}.
\setlength{\tabcolsep}{1pt}
\begin{sidewaystable}
\centering
\caption{Estimates for the posterior means and standard deviations, in round brackets, of the fixed effects included in the models described in Table~\ref{tab:summary_models}.}
{
\small
\begin{tabular}{l|SSSSSSS|SSSSSSS}
\toprule
\multirow{2}{*}{ID} & \multicolumn{7}{c}{Severe Crashes} & \multicolumn{7}{c}{Slight Crashes}\\
& {$\beta_{01}$} & {Betw.} & {Motorways} & {Prim. Roads} & {Ratio Empl.} & {Pop. Dens.} & {Dual Carr.} & {$\beta_{02}$} & {Betw.} & {Motorways} & {Prim. Roads} & {Ratio Empl.} & {Pop. Dens.} & {Dual Carr.} \\
\midrule
\multirow{2}{*}{(A)}
& -14.229 & 0.011 & -0.850 & 0.327 & 0.009 & 0.181 & 0.040 &
-12.679 & -0.004 & -0.009 & 0.656 & -0.005 & 0.265 & -0.318 \\
& (0.092) & (0.058) & (0.180) & (0.148) & (0.061) & (0.053) & (0.110) &
(0.061) & (0.036) & (0.117) & (0.101) & (0.043) & (0.040) & (0.072) \\
\multirow{2}{*}{(B)}
& -14.348 & -0.015 & -0.834 & 0.336 & 0.026 & 0.157 & 0.031 &
-12.700 & -0.023 & -0.023 & 0.630 & 0.011 & 0.258 & -0.324 \\
& (0.153) & (0.059) & (0.187) & (0.150) & (0.063) & (0.055) & (0.114) &
(0.132) & (0.036) & (0.123) & (0.104) & (0.045) & (0.042) & 0.073 \\
\multirow{2}{*}{(C)}
& -14.245 & -0.051 & -0.737 & 0.397 & 0.011 & 0.168 & -0.023
& -12.665 & -0.031 & 0.011 & 0.636&-0.005&0.241&-0.317
\\
&(0.092)&(0.058)&(0.185)&(0.149)&(0.063)&(0.056)&(0.108)
&(0.062)&(0.036)&(0.121)&(0.104)&(0.045)&(0.041)&(0.071)
\\
\multirow{2}{*}{(D)}
&-14.354&-0.069&-0.697&0.516&0.008&0.115&-0.111
&-12.662&-0.038&-0.077&0.58&0.022&0.243&-0.3
\\
&(0.12)&(0.053)&(0.181)&(0.131)&(0.059)&(0.055)&(0.107)
&(0.098)&(0.036)&(0.124)&(0.101)&(0.045)&(0.042)&(0.072)
\\
\multirow{2}{*}{(E)}
&-14.285&-0.005&-0.805&0.426&-0.007&0.209&-0.006
&-12.665&-0.012&-0.021&0.639&-0.004&0.261&-0.299
\\
&(0.092)&(0.057)&(0.183)&(0.143)&(0.062)&(0.056)&(0.11)
&(0.061)&(0.036)&(0.117)&(0.101)&(0.043)&(0.04)&(0.071)
\\
\multirow{2}{*}{(F)}
& -14.357&-0.043&-0.758&0.473&0.009&0.155&-0.053
& -12.677&-0.026&-0.055&0.599&0.015&0.256&-0.299
\\
&(0.138)&(0.055)&(0.183)&(0.136)&(0.06)&(0.056)&(0.11)
&(0.119)&(0.036)&(0.121)&(0.102)&(0.044)&(0.041)&(0.072)
\\
\bottomrule
\end{tabular}
}
\setlength{\tabcolsep}{6pt}
\vspace{1.5cm}
\label{tab:fixed_effects}
\caption{Estimates for the posterior means and standard deviations, in round brackets, of hyperparameters included in the models described in Table~\ref{tab:summary_models}.}
\begin{tabular}{l|SSSSSSS}
\toprule {ID} & {$\sigma^2_{\theta_1}$} & {$\sigma^2_{\theta_2}$} & {$\rho_\theta$} & {$\rho$} & {$\sigma^2_{\phi_1}$} & {$\sigma^2_{\phi_2}$} & {$\rho_\phi$} \\
\midrule
\multirow{2}{*}{(A)} & 0.0001 & 0.759 & & & 0.152 & 0.127 & \\
& (0.0003) & (0.049) & & & (0.028) & (0.019) & \\
\multirow{2}{*}{(B)} & 0.0001 & 0.578 & & 0.996 & 0.373 & 0.347 & \\
& (0.0003) & (0.047) & & (0.0008) & (0.052) & (0.045) & \\
\multirow{2}{*}{(C)} & 0.0001 & 0.594 & & & 0.308 & 0.249 & 0.898 \\
& (0.0003) & (0.048) & & & (0.047) & (0.035) & (0.021) \\
\multirow{2}{*}{(D)} & 0.0001 & 0.376 & & 0.986 & 0.747 & 0.690 & 0.904 \\
& (0.0003) & (0.054) & & (0.005) & (0.121) & (0.108) & (0.019) \\
\multirow{2}{*}{(E)} & 0.589 & 0.752 & 0.413 & & 0.132 & 0.134 & 0.789 \\
& (0.081) & (0.046) & (0.013) & & (0.020) & (0.017) & (0.042) \\
\multirow{2}{*}{(F)} & 0.484 & 0.643 & 0.405 & 0.997 & 0.276 & 0.269 & 0.829 \\
& (0.081) & (0.046) & (0.016) & (0.001) & (0.038) & (0.033) & (0.030) \\
\bottomrule
\end{tabular}
\label{tab:random_effects}
\end{sidewaystable}
\subsubsection*{Random effects}
The posterior means and standard deviations for all hyperparameters are reported in Table~\ref{tab:random_effects}, which reflects the models nested structure, also summarised in Table~\ref{tab:summary_models}.
We started from two baselines, (A) and (B), with independent random effects, and generalised them until model (F), that presents multiple autocorrelation parameters between the two severity levels.
Models from (A) to (D), which assume independent unstructured random effects, exhibit a degenerate posterior distribution of $\sigma_{\theta_{1}}^2$, i.e. the variance of severe random component, and this is possibly due to the severe car crashes sparseness.
This problem gets mitigated once the correlation parameter between the two severity levels is included in the model, suggesting that the estimation procedure benefits from the inclusion of a multivariate structure that allows borrowing strength from less rare events.
The estimates of $\sigma_{\theta_{2}}^2$ and $\rho_{\theta}$ are stable among the models, and the correlation parameter is estimated as high as $0.40$, suggesting a positive and mildly strong relationship between the two random components.
The posterior means for hyperparameters $\rho$ in models (B), (D), and (F), are always very close to one, which is not uncommon for this type of models \parencite{carlin2003hierarchical}.
The estimates of the posterior distributions for the two conditional variances, $\sigma_{\phi_1}^2$ and $\sigma_{\phi_2}^2$, are found less stable compared to the unstructured errors.
The credible intervals of the two hyperparameters overlap in all models, indicating a similar spatial structure between the two kinds of severities.
The posterior mean of $\rho_\phi$, the correlation coefficient between the two severity levels, is found approximately equal to 0.9 (models C and D) and 0.8 (models E and F), indicating a strong multivariate nature for the spatial random component.
These results suggest that car crash data have a complex latent structure being the severity levels strongly correlated, and the spatially structured and unstructured effects statistically relevant.
\subsubsection*{Model comparisons}
We compared the models listed in Table~\ref{tab:summary_models} using DIC and WAIC criteria.
The results are reported in Table~\ref{tab:DIC}.
We first notice that PMCAR models (i.e. (B), (D) and (F)) are found to perform always better than their Intrinsic counterparts in terms of goodness of fit.
They are somewhat unexplored in the road safety literature on spatial networks, \textcite{miaou2005bayesian} being the only paper we found that analyse the importance of a spatial autocorrelation parameter.
However, our results suggest that PCAR distribution and its generalisations should deserve more attention.
Moving from (A) to (F) the model performance improves, indicating one more time the benefits of considering a correlated multivariate structure for the spatial and the unstructured components.
In particular, model (F) is the best one according to both criteria; hence, hereafter, we focus on this model.
\begin{table}
\centering
\caption{Estimates of DIC and WAIC values for the models described in Section~\ref{sec:methodology}. The columns \textit{Balanced Accuracy-Severe} and \textit{Balanced Accuracy - Slight} will be explained in Section~\ref{sec:criticism}.}
\begin{tabular}{lrrrr}
\toprule ID & DIC & WAIC & Balanced Accuracy - Severe & Balanced Accuracy - Slight \\
\midrule
(A) & 14462.56 & 14474.89 & 0.631 & 0.718 \\
(B) & 14408.06 & 14433.77 & 0.635 & 0.716 \\
(C) & 14269.79 & 14296.06 & 0.645 & 0.718 \\
(D) & 14154.09 & 14167.64 & 0.661 & 0.717 \\
(E) & 14122.46 & 14106.69 & 0.662 & 0.720 \\
(F) & 14103.44 & 14086.46 & 0.675 & 0.720 \\
\bottomrule
\end{tabular}
\label{tab:DIC}
\end{table}
\subsubsection*{Car crashes rates}
\begin{figure}
\centering
\subfloat[Severe Car Crashes. \label{fig:pred_serious}]{
\includegraphics[width=\linewidth]{figures/pred_severe.eps}
}
\hspace{0.01\linewidth}
\subfloat[Slight Car Crashes. \label{fig:pred_slight}]{
\includegraphics[width=\linewidth]{figures/pred_slight.eps}
}
\caption{Maps representing the posterior means for severe and slight car crashes rates, estimated using model (F). The colours go from red (higher quantiles) to blue (lower quantiles). The black star indicates the city centre while the inset maps highlight the road network in its proximity.}
\label{fig:pred}
\end{figure}
Figure~\ref{fig:pred} displays the posterior means of car accident rates, $\lambda_{ij}$, estimated using model (F) both for severe and slight crashes.
The colours of the road segments were generated by dividing the predicted values of each severity level into ten classes based on a set of quantiles ranging from red (highest quantile) to blue (lowest quantiles).
In both cases, the highest values correspond to, approximately, one severe or ten slight car crashes every kilometre and every thousand of daily commuters (i.e. the two quantities that define the offset).
Moreover, the inset maps highlight a few roads close to Leeds city centre (denoted by a black star).
The four maps show similar patterns, but some roads in the southern part of the city (especially M621) look more prone to severe car crashes.
The city of Leeds appeared to be divided into several areas.
The northern and north-eastern part of the city are associated with lower car accident rates compared to other suburbs.
The areas located in north-west, east, and south of the city centre seem to be associated with the highest levels of car crashes rates, especially severe ones.
This is probably linked with some congested arterial thoroughfares reaching the centre.
Finally, we note that the roads closer to the city centre are the safest part of the city network, as it is clearly shown by the two inset maps in Figure~\ref{fig:pred}.
\section{Model criticism and sensitivity analysis}
\label{sec:criticism}
DIC and WAIC criteria were never intended to be absolute measures of model fit, and they cannot be used for \textit{Model Criticism}.
Hence, we tested the adequacy of model (F) using two strategies.
\subsection{First strategy for criticism}
The classical criterion for criticism of a Bayesian hierarchical model is the Probability Integral Transform \parencite{marshall2003approximate, held2010posterior}, typically adjusted in case of a discrete response variable (such as car crashes counts) using a continuity correction.
Unfortunately, these adjustments do not seem to work appropriately when modelling sparse count data, such as severe crashes, since the correction is not adequate.
We refer the interested reader to the supplementary material for more details.
Therefore, hereafter, we followed a different strategy.
We binned the observed and predicted counts into two classes: \textit{Zero} and \textit{One or more} car crashes.
Then, we built a confusion matrix and evaluated the model performance via some accuracy measures that are summarised in Table~\ref{tab:confusion_matrix}.
A similar procedure for sparse count data was also presented in \textcite{ma2017multivariate}.
We decided to adopt one as a threshold to dichotomise the variables since more than 80\% of road segments registered no severe car crash during 2011-2018.
However, the algorithm proposed here can be extended to three or more classes, defined using a set of different thresholds (such as \textit{Zero}, \textit{One}, and \textit{Two or more} road crashes).
The \textit{accuracy} measure, usually adopted for evaluating the predictive performance of a model, is typically biased and overly-optimistic in case of unbalanced classes (such as \textit{Zero} and \textit{One or more} severe car crashes per road segment), since, even in the worst case, it is as high as the percentage of observations in the more frequent class \parencite{5597285}.
The \textit{balanced accuracy}, firstly introduced by \textcite{5597285}, is defined as the average of Sensitivity and Specificity, and it overcomes this drawback since it represents an average between the predictive performances on each class.
\begin{table}
\centering
\hspace{-2cm}
\caption{Left: Confusion matrix showing the observed and predicted counts, binned in two classes. The rows represent the actual counts while the columns are the predicted counts. Right: Definition of accuracy measures.}
\begin{tabular}{cccc}
\multicolumn{2}{c}{} & Zero & One or more \\
\multicolumn{2}{l}{Zero} & A & B \\
\multicolumn{2}{l}{One or More} & C & D \\
\end{tabular}
\qquad
\qquad
\begin{tabular}{ll}
Sensitivity & = $\frac{A}{A + B}$ \\[3pt]
Specificity & = $\frac{D}{C + D}$ \\[3pt]
Accuracy & = $\frac{A + D}{A + B + C + D}$ \\[3pt]
Balanced Accuracy & = $\frac{1}{2}\left(\frac{A}{A + B} + \frac{D}{C + D}\right)$
\end{tabular}
\label{tab:confusion_matrix}
\end{table}
The output of a Bayesian hierarchical model is an estimate of the posterior distribution of predicted values, while the procedure reported in the previous paragraph can only be applied to binary data.
For this reason, we simulated $n$ Poisson random variables (one for each road segment) with mean equal to the mean of each posterior distribution.
Then, we binned the observed and sampled counts into two classes, i.e. Zero and One or more car crashes, and we compared the two values, obtaining a single estimate of balanced accuracy.
Its distribution was finally approximated by repeating this procedure $N = 5000$ times.
Moreover, we calculated several quantiles of the posterior distribution of each predicted value, and we run the same steps as before, sampling from a Poisson distribution with mean equal to each of those quantiles.
Lastly, being severe and slight car crashes potentially quite different processes, this algorithm was applied independently for the two severity levels.
We reported in the supplementary material the pseudo-code for running this procedure, whereas results are displayed in Figure~\ref{fig:balancedaccuracyserious} (severe cashes) and Figure~\ref{fig:balancedaccuracyslight} (slight crashes).
In both cases, the red curve represents the distribution of balance accuracy obtained by a binary classification based on the posterior means, whereas the other curves represent the same distribution obtained using the set of quantiles.
\begin{figure}
\centering
\subfloat[Severe Crashes \label{fig:balancedaccuracyserious}]{
\includegraphics[width=0.47\linewidth, trim={0 0 0 0.735cm}, clip]{figures/balanced_accuracy_severe.eps}
}
\hspace{0.01\linewidth}
\subfloat[Slight Crashes \label{fig:balancedaccuracyslight}]{
\includegraphics[width=0.47\linewidth, trim={0 0 0 0.735cm}, clip]{figures/balanced_accuracy_slightv2.eps}
}
\caption{Distribution of balanced accuracy for severe crashes (left) and slight crashes (right), considering a binary classification using the posterior mean and a set of quantiles. The red curve represents the mean.}
\label{fig:balancedaccuracy}
\end{figure}
It looks like the optimal threshold for binary classification of severe car crashes is given by the $0.975$-quantile, where the balanced accuracy distribution is concentrated around $0.675$.
The optimal threshold for binary classification of slight car crashes is given by the median, and the distribution of balanced accuracy is centred around $0.72$.
These plots remark the differences between the two severity levels in terms of sparsity, suggesting the adoption of a higher quantile for the prediction of the more sparse events.
However, using the appropriate cut-off(s), Model (F) seems to perform reasonably well in both cases with slightly better performance for slight car crashes.
We also explored the goodness of fit of model (F) estimating the balanced accuracy measure separately for the three road classes and the two road carriageway types.
We always obtained results similar to Figures~\ref{fig:balancedaccuracyserious} and~\ref{fig:balancedaccuracyslight}.
Moreover, we did not observe any relevant difference among the road typologies, highlighting that model (F) successfully predicts traffic collisions in all tested situations.
The third column in Table~\ref{tab:DIC} summarises the estimates of Balanced Accuracy obtained for each model in Table~\ref{tab:summary_models} under the best scenario (i.e. the optimal quantile) for severe crashes, while the fourth column summarises the same quantities for slight crashes.
We notice that the accuracy improves every time a new correlation term is included into a model, particularly for severe car accidents, which is the rarer car crash typology.
In fact, considering that improving risk estimation of very rare events is one of the main reasons why one may want to adopt a multivariate model in the first place, our results seem to suggest that the approach proposed in this paper represents a reasonable way to investigate road collision dynamics.
\subsection{Second strategy for criticism}
Following the results illustrated before, we estimated the $0.975$-quantile of $\lambda_{i1}$ (severe crashes) and the median of $\lambda_{i2}$ (slight crashes), and we multiplied them by the corresponding offset values, i.e. $E_i$.
Then, we created a sequence of histograms of predicted values, grouped by the observed counts categorised in four levels: 0, 1, 2, and 3 or more.
The results are summarised in Figure~\ref{fig:posterior_means_serious} and Figure~\ref{fig:posterior_means_slight}.
Both graphs show a good agreement between predicted and observed number of crashes, since the distributions corresponding to higher observed counts progressively move more and more to the right.
Moreover, Figure~\ref{fig:posterior_means_serious} shows the importance of our previous analysis and the pitfalls of predicting severe car crashes counts using the posterior means.
\begin{figure}
\centering
\subfloat[Severe Crashes \label{fig:posterior_means_serious}]{
\includegraphics[width=0.47\linewidth]{figures/posterior_means_serious}
}
\hspace{0.01\linewidth}
\subfloat[Slight Crashes \label{fig:posterior_means_slight}]{
\includegraphics[width=0.47\linewidth]{figures/posterior_means_slight}
}
\caption{Histogram of posterior $0.975$-quantile (left) and posterior median (right), grouped by the corresponding observed counts. Other means "Three or more."}
\label{fig:posterior_means}
\end{figure}
\subsection{Sensitivity analysis and the modifiable areal unit problem}
Finally, we performed a sensitivity analysis evaluating the robustness of model (F) under different specifications for 1) the hyperprior distributions, 2) the adjacency matrix, and 3) the definition of the segments in the road network.
The models described in Section~\ref{sec:methodology} considered a Wishart hyperprior for the precision matrix $\bm{\Omega}$ with rank equal to $2$ and scale matrix equal to $\bm{I}_2$.
We repeated the analysis using more vague and more informative Wishart distributions, setting the scale matrix equal to $\text{diag}(2, 2)$ and $\text{diag}(0.5, 0.5)$.
We did not find any noticeable differences amongst alternative specifications.
Hence results are not reported hereafter, but we refer the interested readers to the supplementary material.
We compared different definitions for the adjacency matrix $\bm{W}$, testing second and third order neighbours and distance-based spatial neighbours, considering the first order neighbours in case two adjacent road segments are longer than the given threshold.
Also in this case, we did not find any noticeable differences as far as the estimation of the fixed effects is concerned, whereas only small differences were found in the posterior distributions of the random effects (especially for $\sigma_{\phi_1}^2$ and $\sigma_{\phi_2}^2$ when we considered a spatial adjacency matrix with a threshold equal to $500m$).
However, worse DIC and WAIC values were found for models using alternative definitions of $\bm{W}$ matrix, and we refer to the supplementary material for more details.
Similar findings are also reported by \textcite{aguero2008analysis, wang2016macro, alarifi2018exploring}.
Finally, we explored the influence of a particular configuration of the network segments on our results.
In fact, the location of the vertices (and, hence, the edges) in a road network created with OS data is essentially arbitrary (although some minimal consistency requirements must be satisfied, see \textcite{karduni2016protocol, gilardi_lovelace_padgham_2020}), which implies that there is no unique and unambiguous way of defining the lengths and relative positions of the road segments.
We, therefore, considered an alternative network configuration reshaping and contracting the road network using an algorithm implemented in \textcite{R-dodgr}.
This algorithm manipulates a network by excluding all \textit{redundant} vertices, i.e. those vertices that connect two contiguous segments without any other intersection \parencite{R-dodgr}.
A toy example representing the ideas behind the contraction of a road network is sketched in Figure~\ref{fig:contract}.
The red dots in Figure~\ref{fig:redundant} represent redundant vertices since they can be removed without tampering the shape or the routability of the network, meaning that excluding those vertices does not add any new cluster to the graph structure.
The goal is to remove all redundant vertices and merge the corresponding edges, creating a graph which looks identical to the original one but with fewer edges.
Figure~\ref{fig:contracted} shows the results of the contraction operations applied to the toy network sketched in Figure~\ref{fig:redundant}.
We can see that the redundant vertices were removed, combining the road segments that touched them.
\begin{figure}
\centering
\subfloat[Road Network with redundant vertices (in red) \label{fig:redundant}]{
\includegraphics[width=0.47\linewidth]{figures/redundantv2.eps}
}
\hspace{0.01\linewidth}
\subfloat[Contracted road network \label{fig:contracted}]{
\includegraphics[width=0.47\linewidth]{figures/contractedv2.eps}
}
\caption{Sketching of the algorithm used for contracting the road network. Red points on the left represent redundant vertices.}
\label{fig:contract}
\end{figure}
\section{Discussion and Conclusions}
\label{sec:conclusions}
This paper investigated the spatial distribution of road crashes in a major city using Bayesian methods for road network analysis.
The relationship between crashes of different severity levels, either \textit{slight} or \textit{severe}, were modelled using a range of multivariate models to explore their spatial dynamics.
Key to the approach was constraining crash locations to the city's road network, a one-dimensional linear network composed of segments representing a spatial lattice.
We tested a range of multivariate hierarchical models with different random effects, and we found that the best model (according to DIC, WAIC and \textit{balanced accuracy} criteria) includes a multivariate spatially unstructured random effect and a multivariate spatially structured PMCAR random effect.
The physical and social environment of the street segments were considered in the model specification.
Our results, summarised in Tables~\ref{tab:fixed_effects} and~\ref{tab:random_effects}, suggest that population density is positively correlated with severe and slight crashes, although the effect is stronger for the latter case.
On the other hand, the employment rate does not seem related with car crash occurrences, highlighting the importance of testing alternative proxies for poverty or income.
As far as street category is concerned, Primary roads have been found safer than A roads both for severe and slight car crashes, whereas Motorways are significantly less prone to severe car accidents than A roads.
No significant effect is found for slight accidents.
The betweenness centrality does not seem to influence slight or severe accidents; this is an unexpected result which may deserve further investigation, possibly with alternative proxies to measure the VMT.
Finally, dual carriageway roads have been found significantly less prone to slight car accidents, whereas no impact has been found for severe car crashes.
Other potentially relevant physical characteristics might be the traffic speed, the line width, the presence of speed limits or junctions, the slope and the curvature of the road segments.
Unfortunately, these variables were not available to us at the time when this paper was written.
However, given the detailed network adopted in this paper, which is composed of thousands of short segments, we believe that some of those variables may be of limited relevance due to the short length of many segments.
These unmeasured effects are accounted, to some extent, by the second order parameters of the random effects included in our models.
Concerning this last point, we underline that the unstructured random effects are supposed to reflect the uncertainty due to scarce sampling information (which may occur in small spatial domains like road segments) as well as the differences among the segments, such as faults in the pavements, crosses or junctions.
These characteristics act as local shocks on the car crash occurrences and, although unrelated in space, they act both on severe and slight car accidents.
Hence, the corresponding random effects are expected to be correlated and the parameter $\rho_\theta$ is meant to measure this correlation.
Similarly, spatially structured random effects are expected to account for the impact of unmeasured variables that have regularities is space (i.e. meteorological conditions or structural characteristics of the roads not available in our analysis).
These variables act both on severe and slight car accidents and parameter $\rho_\phi$ is expected to measure the correlation between spatially structured random effects due to these components.
Both correlations are found relevant in our data, although the correlation between spatially structured effects is found stronger than the other.
Hence, we argue they both should be considered when modelling car crashes data.
The relevant interactions between the two severity levels allow one level to borrow strength from the other and improve the estimation of the risk associated with each piece of the city road network, especially for the rarer event.
We evaluated the sensitivity of our modelling approach to different hyperprior specifications and adjacency configurations of the components of the lattice network, showing that the statistical model presents substantial robustness in this respect.
We finally considered the impact of MAUP when modelling data collected on a spatial network.
An algorithm was proposed in the paper to assess the magnitude of MAUP effect in the estimates and model predictions.
As mentioned above, differently from several previous studies that considered the MAUP for various areal partitions of the spatial region of interest, we found that our results are quite robust under an alternative configuration of the road network.
This can be related to the fact that road networks have a physical meaning, hence they are expected to suffer MAUP less.
Nevertheless, further research, possibly in different fields, is definitely necessary to better understand the impact of MAUP on network lattice data.
Finally, we remark that using areal units as spatial support ignores the fact that car crashes cannot occur outside the road network, and ascribes an estimated risk to all the streets in that polygon, whether or not they are actually exposed to it.
Differently, using a lattice based on a network structure draws attention to limited and specific parts of the spatial support.
Hence, adopting a more appropriate spatial disaggregation can be fundamental for local authorities to plan actions (such as to install a new traffic light, add or remove roundabouts or enforce police control) to mitigate this risk where it is found too high.
The ideas presented in this paper could be extended in several directions.
A first step forward could be focused towards the development of a spatio-temporal extension of model (F), following the suggestions in \textcite{miaou2005bayesian, wang2011predicting, boulieri2017space, ma2017multivariate}.
We point out, however, that this is not straightforward (and, to the best of our knowledge, it was pursued only by \textcite{ma2017multivariate} using a single road divided into a few segments) given the extreme sparse spatio-temporal nature of severe car crashes on a metropolitan road network.
Indeed, for the dataset at hand, more than 95\% of all car crashes registered no fatal or serious car accident for any given year, something that could require a different methodological approach.
The procedure for MAUP detection could also be improved by developing new routines for testing alternative algorithms for network reshaping and contraction, which were first developed for areal data in the field of geography (see, e.g., \textcite{xu2018modifiable} and references therein).
An additional improvement to the approach could involve the development of spatial or spatio-temporal theoretical point pattern models for car crashes on networks \parencite{baddeley2020analysing}.
This approach represents a flexible and powerful way to investigate the spatial dynamics of random events on a graph support and may provide a tool that largely circumvents the MAUP problem.
However, whereas it is moderately straightforward to include covariates in the bivariate lattice models suggested in this paper, the theory of bivariate point pattern models on networks and the inclusion of spatial covariates in this framework would require substantial methodological development.
In addition, most of the work in point pattern modelling is non-parametric or semiparametric in nature, whereas the approach adopted in the present paper and, to same extent, in the analysis spatial lattice data, is grounded in the Bayesian framework.
In this sense, the two approaches are complementary when modelling spatial or spatio-temporal dynamics of crash data.
A bridge between the two would be the log-Gaussian Cox processes, where a stochastic component is included in the intensity function to deal with the unexplained spatial variation \parencite{Molleretal1998}.
Dealing with an intensity function governed by a spatial random field, however, raises challenging methodological problems since defining a proper covariance function on graph spatial support is not straightforward.
Considering this issue is beyond the scope of the present paper; hence, we do not discuss it further and leave it for future research.
It is clear that more research is needed to evaluate the full range of possible models for identifying crash `hot spots' and highlight segments on the network.
The approach presented in this paper demonstrates the potential of network-based approaches to work at city scales for flexible and robust estimates of crash rates, down to the road segment level, providing a foundation for further work in the field.
\section*{Code and data availability statement}
The data and the code that were used to estimate models from (A) to (G) and the model on the contracted network can be obtained at the following link: \url{https://github.com/agila5/multivariate-analysis-car-crashes}.
\section*{acknowledgements}
Contains OS data © Crown copyright [and database right] 2020.
J. Mateu is partially suported by funds PID2019-107392RB-I00 and AICO/2019/198.
We greatly acknowledge the DEMS Data Science Lab for supporting this work by providing computational resources.
\printbibliography
\end{document}
|
\section{Introduction}
\par{\bf Code-based Cryptography.} Many cryptosystems as public-key encryption schemes \cite{M78,A11,MTSB12}, authentication protocols \cite{S93} or pseudorandom generators \cite{FS96} are built relying on the hardness of finding the closest codeword, a task called {\em decoding}. In the case of a random linear code, which is the standard case, this problem can be expressed as follows
\begin{definition}[$\textup{$\mathsf{DP}$}(q,n,k,t)$] The decoding problem with parameters $q,n,k,t\in\mathbb{N}$ is defined as:
\begin{itemize}
\item Given: $(\mat{G},\vec{u} \mat{G} + \ev)$
where $\mat{G} \in \mathbb{F}_{q}^{k \times n}$ and $\vec{u} \in \F_q^{k}$ are sampled uniformly at random over their domain and $\ev \in \F_q^{n}$ over the words of weight $t$,
\item Find: $\ev$
\end{itemize}
\end{definition}
This problem really corresponds to decode the code $\ensuremath{\mathscr{C}}$ which is the $k$-dimensional vector space generated by the rows of $\mat{G}$:
\begin{equation}\label{eq:def_code}
\ensuremath{\mathscr{C}} \stackrel{\text{def}}{=} \{\vec{u} \mat{G}:\;\vec{u} \in \F_q^k\},
\end{equation}
\textit{i.e.} we are given the noisy codeword $\vec{c} + \ev$ where $\vec{c}$ belongs to $\ensuremath{\mathscr{C}}$ and we are asked to find the error $\ev$ (or what amounts to the same, the original codeword $\vec{c}$).
This problem for random codes has been studied for a long time and despite many efforts on this issue, the best algorithms are exponential in the codelength $n$ in the regime where $t$, $k$ and $n-k$ are all linear in $n$.
Usually this decoding problem is considered in the regime where the code rate $R \stackrel{\text{def}}{=} \frac{k}{n}$ is fixed in $(0,1)$ and $q=2$, but there are also other interesting parameters for cryptographic applications. For instance, the Learning Parity with Noise problem (LPN) corresponds to $\textup{$\mathsf{DP}$}(q,n,k,t)$ where $n$ is the number of samples, $k$ the length of the secret while the error is sampled according to a Bernoulli distribution of fixed rate $t/n$. As the number of samples in LPN is unlimited, this problem really corresponds to decoding a code of rate arbitrarily close to $0$.
While the security of many code-based cryptosystems relies on the hardness of the decoding problem, it can also be based on finding a ``short'' codeword (as in \cite{MTSB12} or in \cite{AHIKV17,BLVW19,YZWGL19} to build collision resistant hash functions), a problem which is stated as follows.
\begin{definition}[short codeword problem $\scodeP(q,n,k,w)$] Let $q,n,k,w\in\mathbb{N}$. The short codeword problem with parameters $q,n,k,w$ is defined as follows:
\begin{itemize}
\item Given: $\mat{H} \in\mathbb{F}_{q}^{(n-k)\times n}$ which is sampled uniformly at random,
\item Find: $\vec{c}\in\mathbb{F}_{q}^{n}$ such that $\mat{H} \transpose{\vec{c}} = \mathbf{0}$ and the weight of $\vec{c}$ belongs to $(0,w]$
\end{itemize}
\end{definition}
Here we are looking for a non-zero codeword $\vec{c}$ of weight $ \leq w$ in the $k$-dimensional code $\ensuremath{\mathscr{C}}$ defined by the so-called parity-check matrix $\mat{H}$, namely:
$$
\ensuremath{\mathscr{C}} \stackrel{\text{def}}{=} \left\{ \vec{c}\in\F_q^{n} \mbox{ : } \mat{H}\transpose{\vec{c}} = \mathbf{0} \right\}.
$$
Decoding and looking for short codewords are problems that have been conjectured for a long time to be extremely close. They have been studied for a long time \cite{P62,S88,D89,MMT11,BJMM12,MO15,BM18}, and for instance in the regime of parameters where the rate $R=\frac{k}{n}$ is fixed in $(0,1)$, the best algorithms for solving them are the same (namely Information Set Decoding). A reduction from decoding to the problem of finding short codewords is known but in an LPN context \cite{AHIKV17,BLVW19,YZWGL19}. However, even in an LPN context, no reduction is known in the other direction. These problems can be viewed in some sense as a code version of the LWE and SIS problems respectively in lattice based cryptography \cite{R09}. A breakthrough there was to obtain a quantum reduction from SIS to LWE \cite{R05,SSTX09}. Our contribution in this article is precisely to give the code based version of this reduction, namely a quantum reduction from finding short codewords to decoding. This problem was open for quite some time. To simplify the statements, we will state it in the regime of parameters where the rate $R$ is fixed in $(0,1)$, but actually it also works in the LPN setting (but needs to be adapted in several places where we use exponential bounds in $n$).
\par{\bf Parameter range for $\decP$ and $\scodeP$.} An important parameter for the reduction is the distance decoding $t$.
The largest value of $t$ for which the decoding problem is ensured to have a unique solution is equal to $\left\lfloor \frac{d_{\textup{min}}-1}{2} \right\rfloor$ where $d_{\textup{min}} \stackrel{\text{def}}{=} \min \{d(\vec{c},\vec{c}'): \;\vec{c} \in \ensuremath{\mathscr{C}}, \; \vec{c}' \in \ensuremath{\mathscr{C}}, \;\vec{c} \neq \vec{c}'\}$ is the minimum distance of $\ensuremath{\mathscr{C}}$ (which depends of course on the metric $d(\cdot,\cdot)$ that is considered). Standard probabilistic arguments can be used to show that the minimum distance of a random linear code (code $\ensuremath{\mathscr{C}}$ obtained as in \eqref{eq:def_code} by a generator matrix $\mat{G}$ chosen uniformly at random in $\F_q^{k \times n}$) is with very high probability equal, up to an additive constant, to the {\em Gilbert-Varshamov distance} $d_{\textup{GV}}(n,k)$ (or simply $d_{\textup{GV}}$ if there is no ambiguity). It is defined (for all metrics) for a code of dimension $k$ and length $n$, as the largest integer $t$ for which
\begin{equation}
q^k \cdot B_t \leq q^n
\end{equation}
where $B_t$ is the size of a ball of radius $t$. For the Hamming metric we have
\begin{eqnarray*}
\delta_{\textup{GV}}(n,k) &=& h_q^{-1}\left(1-\frac{k}{n}\right)+\OO{\frac{1}{n}}\;\text{ where}\\
h_q(x) & \stackrel{\text{def}}{=} & -x\log_q\left( \frac{x}{q-1} \right) -(1-x) \log_q(1-x) \;\text{ and $h_q^{-1}$ its inverse ranging over $[0,\frac{q-1}{q}]$}
\end{eqnarray*}
and $\delta_{\textup{GV}}(n,k)$ is the normalized Gilbert-Varshamov distance defined as $\delta_{\textup{GV}}(n,k) \stackrel{\text{def}}{=}\frac{d_{\textup{GV}}(n,k)}{n}$.
This Gilbert-Varshamov distance turns out to quantify also the region where we {\em typically} have unique decoding. More precisely, it turns out that the same probabilistic arguments also show that the solution of the decoding problem is unique with probability $1 - 2^{-\Omega(n)}$ as long as $t \leq (1-\varepsilon)d_{\textup{GV}}(n,k)$ as $n$ goes to infinity for fixed positive $\varepsilon$.
Whereas the best algorithms for solving the decoding have exponential complexity in $n$ as soon as $t$ is linear in $n$ and the code rate
$R \stackrel{\text{def}}{=} \frac{k}{n}$ is bounded away from $0$ and $1$, this is not true for the short codeword problem which becomes easy when the weight $w$ is above a certain range. The reason for this, is that it is easy to produce codewords of small weight by using the fact that the code is a vector space of dimension $k$. Thus we can just produce codewords with $k-1$ entries equal to $0$ by solving a linear system which gives good candidates for having small weight.
It is straightforward that this strategy produces in polynomial time, for instance with the Hamming metric, codewords of weight $\approx \omega_{\textup{easy}}(n,k) n$ where
\begin{equation}\label{eq:weasy}
\omega_{\textup{easy}}(n,k) \stackrel{\text{def}}{=}
\frac{q-1}{q}(1-k/n)
\end{equation}
Obtaining larger weights is also readily obtained by choosing only part of the $k-1$ entries to be equal to $0$. It should be noted that below $\omega_{\textup{easy}}(n,k)$ the best known algorithms for solving this problem have all exponential complexity for a fixed rate $R$ and a fixed ratio $\omega = \frac{w}{n}$.
\par{\bf Regev's quantum reduction strategy adapted to coding theory.}
In \cite{R05} (see also the extended version \cite{R09}) Regev showed how to transform a
random oracle solving the decoding problem in a lattice into a quantum algorithm outputting a rather small
vector in the dual lattice. Our aim is to show here that the natural translation of this approach in coding theory gives an algorithm that outputs a rather small vector in the dual code. Roughly speaking Regev's approach relies on a fundamental result about the Fourier transform.
\begin{proposition}\label{prop:constant_on_coset_and_dual}
Consider an Abelian group $G$ and a function $f : G \mapsto \mathbb{C}$ that is constant on the cosets of a subgroup $H$ of $G$. Then the Fourier transform $\reallywidehat{f}$ is constant on the dual subgroup $H^\perp$.
\end{proposition}
Arguably this innocent looking fact (together with the fact that the Fourier transform can be performed in polylog time when the group $G$ is Abelian) is the key to several remarkable quantum algorithms solving in polynomial time the period finding in a vectorial Boolean function \cite{S94d}, the factoring problem \cite{S94a} or the discrete logarithm problem \cite{S94a}. All of these problems can be rephrased in terms of the hidden Abelian subgroup problem, where one is given such a function $f$ that is constant (and distinct) on the cosets of an unknown subgroup $H$ and one is asked to recover $H$. This is achieved by :\\
(i) creating the uniform superposition $\frac{1}{\sqrt{|G|}} \sum_{x \in G}\ket{x}\ket{f(x)}$, \\
(ii) measuring the second register and discarding it, yielding a quantum state of the form $\frac{1}{\sqrt{|H|}} \sum_{h \in H} \ket{x+h}$, \\
(iii) applying the Fourier transform to it yielding a superposition of elements in the dual subgroup $H^\perp$ (and therefore gaining information on $H$ in this way).
Proposition \ref{prop:constant_on_coset_and_dual} is used in a similar way in Regev's reduction. Translating Regev's reduction in coding theory would use this framework by considering that
the linear code $\ensuremath{\mathscr{C}}$ we want to decode plays the role of the aforementioned $H$. From now on we will assume that this code is of dimension $k$ and length $n$ over $\F_q$. The algorithm would basically look as follows for reducing the search of small codewords in the dual code $\ensuremath{\mathscr{C}}^\perp= \{\vec{c}^\perp \in \F_q^n: \vec{c} \cdot \vec{c}^\perp = 0,\;\forall \vec{c}^\perp \in \ensuremath{\mathscr{C}}\}$ (where $\vec{x} \cdot \vec{y} =\sum_{i=1}^n x_i y_i$ is the standard inner product in $\F_q^n$) to decoding errors of weight $t$ in $\ensuremath{\mathscr{C}}$.
\begin{itemize}
\item[Step 1.] Use a quantized version of the decoding algorithm to prepare the state
$$\frac{1}{\sqrt{Z}}\sum_{\vec{c} \in \ensuremath{\mathscr{C}},\ev } \pi_{\ev} \ket{\vec{c} + \ev}$$
where $Z$ is a normalizing constant and $(|\pi_{\ev}|^2)_{\ev}$ is a probability distribution on errors that concentrates around the weight $t$ we are able to decode. This is done \\
(i) by preparing first a superposition of codewords and errors
$$\frac{1}{\sqrt{Z}} \sum_{\vec{c} \in \ensuremath{\mathscr{C}}} \sum_{\ev \in \F_q^n} \pi_{\ev} \ket{\vec{c}}\ket{\ev},$$
(ii) then
adding the second register to the first one to get the entangled state
$$\frac{1}{\sqrt{Z}} \sum_{\vec{c} \in \ensuremath{\mathscr{C}}} \sum_{\ev \in \F_q^n} \pi_{\ev} \ket{\vec{c} + \ev}\ket{\ev}$$
(iii) disentangle it by a quantized version of the decoding algorithm which from $\vec{c} + \ev$ recovers $\ev$ and subtracts it from the second register to get the state
$$\frac{1}{\sqrt{Z}} \sum_{\vec{c} \in \ensuremath{\mathscr{C}}} \sum_{\ev \in \F_q^n} \pi_{\ev} \ket{\vec{c}+\ev}\ket{\vec{0}}$$
\item[Step 2.] Apply the quantum Fourier transform on $\F_q^n$ to obtain a superposition of elements $\vec{c}^\perp$ in the dual code
$$\sum_{\vec{c}^\perp \in \ensuremath{\mathscr{C}}^\perp} \alpha_{\vec{c}^\perp}\ket{\vec{c}^\perp}.$$
\item[Step 3.] Measure the register to output $\vec{c}^\perp$ of rather small norm in $\ensuremath{\mathscr{C}}^\perp$.
\end{itemize}
The second point is a direct consequence of Proposition \ref{prop:constant_on_coset_and_dual}.
The last point raises the issue of whether or not the Fourier transform concentrates the weight
output by this algorithm on weights $t'$ for which finding a codeword in $\ensuremath{\mathscr{C}}^\perp$ is not known to be easy, as is the case for Regev's reduction on lattices equipped with the Euclidean metric.
\par{\bf On the difficulty of translating Regev's reduction to the Hamming metric.} This thread of research has already been pursued in the binary case by Yilei Chen \cite{C20a} and later on in \cite{CV20}, where basically the following approach was taken. The natural analogue of the Gaussian noise model used in Regev's reduction \cite{R09} in the case of the Hamming metric is the i.i.d. Bernoulli model on each coordinate, i.e.
$|\pi_\ev|^2=(p/(q-1))^{|\ev|}(1-p)^{n-|\ev|}$ where $|\ev|$ stands for the Hamming weight of $\ev$, $n$ the length of $\ev$ and $p$ the parameter of the Bernoulli noise. Both distributions can be expressed in terms of the heat kernel operator (the usual Laplacian in the case of the Gaussian noise and the discrete Laplacian for the case of $\F_q^n$), in both cases the Fourier transform yields a dual noise which is itself Gaussian or Bernoulli based and the quantum state corresponding to the error is a product state which simplifies a great deal the computation. However, it has been realized that this natural approach hits a wall. The problem is the following: we begin to choose the parameter $p$ of the Bernoulli noise, so that the typical weight $pn$ of an error $\ev$ is equal to or slightly below the weight $t$ we can decode. It turns out that the most likely weight we measure at Step 3 is always zero if we want to have a chance that the dual codeword we measure has normalized weight $< \omega_{\textup{easy}}$, i.e. is in the regime where there is a chance that it is difficult to produce such words. In other words, the straightforward application of Regev's approach to coding theory fails to give a useful reduction.
We give in Section \ref{app:failed_attempt} of the appendix an explanation for the failure of this approach.
It can be summarized by saying that the Bernoulli noise model is not enough concentrated on the typical weight $p n$. Intuitively what is going wrong in the case of the Bernoulli noise model, can be explained by bringing in the quantum state
$$ \ket{\pi} \stackrel{\text{def}}{=} \sum_{\ev \in \F_q^n} \pi_{\ev} \ket{\ev},
$$
which represents in some sense the noise we add to the codeword.
In the case of an i.i.d Bernoulli noise of parameter $p$ (i.e. a $q$-ary symmetric channel of crossover probability $p$) we have
$$
\ket{\pi}= \sum_{\ev \in \F_q^n} (1-p)^{\frac{n - |\ev|}{2}} \left(\frac{p}{q-1}\right)^{\frac{|\ev|}{2}}\ket{\ev} = \left( \sqrt{1-p} \ket{0} + \sum_{\alpha \in \F_q^*} \sqrt{\frac{p}{q-1}} \ket{\alpha}\right)^{\otimes n}
$$
Applying the quantum Fourier transform to this state yields a quantum state
\begin{eqnarray*} \QFT{\pi} &= &\sum_{\ev \in \F_q^n} (1-p^\perp)^{\frac{n - |\ev|}{2}} \left(\frac{p^\perp}{q-1}\right)^{\frac{|\ev|}{2}}\ket{\ev} = \left( \sqrt{1-p} \ket{0} + \sum_{\alpha \in \F_q^*} \sqrt{\frac{p^\perp}{q-1}} \ket{\alpha}\right)^{\otimes n}\\
\text{where } p^\perp & \stackrel{\text{def}}{=} &\frac{\left(\sqrt{(q-1)(1-p)}-\sqrt{p}\right)^2}{q}
\end{eqnarray*}
The issue is to understand which is the most likely weight we get when we measure the quantum state at Step 3. This should be the integer $w$ which maximizes
the probability $p_w$ to measure a dual codeword $\vec{c}^\perp$ of weight $w$ which is equal to (see Lemma \ref{lem:measure} in Section \ref{app:th:main} of the appendix)
\begin{equation}
\label{eq:pw}
p_w = \frac{q^{2k}}{Z} \sum_{\vec{c}^\perp \in \ensuremath{\mathscr{C}}^\perp: |\vec{c}^\perp|=w}\left|\widehat{\pi}_{\vec{c}^\perp}\right|^2
\end{equation}
where $\QFT{\pi}=\sum_{\ev \in \F_q^n} \widehat{\pi}_{\ev} \ket{\ev}$. In our case, $\widehat{\pi}$ is a radial function (namely $\widehat{\pi}_{\vec{e}} = f(|\vec{e}|)$ for some function $f$). Therefore we have (for $|\vec{e}| = w$):
$
p_w = \frac{q^{n+k}}{Z} \; \frac{N_{w}^{\perp}}{q^{n-k}} \; |\widehat{\pi}_{\ev}|^{2}
$
where $N_{w}^{\perp}$ is the number of codewords of weight $w$ in $\ensuremath{\mathscr{C}}^{\perp}$.
The density of codewords of Hamming weight $w$ in a random code $\ensuremath{\mathscr{C}}^\perp$ is about the same as the density of elements of weight $w$ in the whole space $\mathbb{F}_q^n$, namely $\frac{1}{q^{n-k}} \; N_{w}^{\perp} \approx \frac{(q-1)^{w}\binom{n}{w}}{q^{n}}$. However there is one notable exception, namely when $w=0$ where the density of codewords is $\frac{1}{q^{n-k}}$ instead of $\frac{1}{q^n}$. In other words, if $\widehat{\pi}_{\vec{0}}$ is too big, it is in $0$ that $p_w$ is maximal.
For $w>0$, we expect that the maximum of $w$ is attained for weights where the probability of measuring a weight $w$ is maximal when we measure directly the quantum state $\QFT{\pi}$, in this case the most likely outcome is a weight $\approx p^\perp n$. We miss to measure this quantity since in our case $|\widehat{\pi}_{\vec{0}} |^2$ is just too large. It is readily seen that
$$\widehat{\pi}_\vec{0} = \frac{\braket{\pi}{\vec{1}}}{\sqrt{q^n}}$$
(where $\un$ is the all one vector and $\braket{}{}$ the hermitian product) and that more or less the best we can do to minimize this quantity if we stick to (i) radial probability distributions $|\pi_\ev|^2$ (i.e. depending only on the weight of $\ev$ --which is a natural constraint), (ii) distributions that concentrate almost all their probability around weight $t$,
and (iii) nonnegative $\pi_{\ev}$'s\footnote{This restriction simplifies significantly the proof in several crucial places. Furthermore if we ask ourselves whether in the Bernoulli case we can improve upon the choice of $\pi_{\ev}$ by keeping the equality $|\pi_{\ev}|^2=(p/(q-1))^{|\ev|} (1-p)^{n-|\ev|}$ and choosing $\pi_{\ev}$ in the complex numbers, it turns out that the choice of positivity for the $\pi_{\ev}$ that was made leads to a dual distribution $\QFT{\pi}$ which is concentrated on the smallest possible weight, namely $p^\perp n$ here.} is to actually concentrate the whole distribution on the weight $t$.
\par{\bf Our approach.}
This is precisely what we have done by choosing
$$
\ket{\pi} = \sum_{\ev : |\ev|=t} \frac{1}{\sqrt{S_t}} \ket{\ev}
$$
where $S_t$ is the cardinality of the sphere of radius $t$ in the Hamming metric, namely $(q-1)^t \binom{n}{t}$. Making this choice actually complicates rather significantly the reasoning. Understanding which weight $w$ maximizes $p_w$ is much more involved (it involves in particular rather delicate properties of Krawtchouk polynomials). However, it will turn out that when $\omega \stackrel{\text{def}}{=} \frac{w}{n}$ lies in a whole interval starting precisely at $\tau^\perp \stackrel{\text{def}}{=} \frac{\left(\sqrt{(q-1)(1-\tau)}-\sqrt{\tau}\right)^2}{q}$ where
$\tau \stackrel{\text{def}}{=} \frac{t}{n}$, we have many points where $p_w$ is actually $\frac{1}{\mathsf{poly}(n)}$. The weight distribution is in this case not really concentrated on a single value but is spread out on a large interval (but with the property that there are weights $w$ close to $\tau^\perp n$ for which $p_w = \frac{1}{\mathsf{poly}(n)}$).
In other words, the previous Bernoulli noise model (with probability $\tau$) now captures with its most likely weight outcome $\tau^\perp n$ when we measure $\QFT{\pi}$ directly, weights which are rather likely to be output at Step 3
(and $0$ is now not anymore the most likely outcome).
Roughly speaking with our approach, we transform through the quantum Fourier transform a decoding algorithm correcting $ \tau n$ errors into an algorithm outputting with non-negligible probability words of weight $\approx \tau^\perp n$ in the dual code. If $\tau^\perp$ is below $\omega_{\textup{easy}}(n-k,n) n$ (here we want to find short codewords in the dual code $\ensuremath{\mathscr{C}}^\perp$ which is of dimension $n-k$) then this would yield a useful reduction.
$\tau^\perp$ is clearly a decreasing function of $\tau$ and the issue is now whether or not there exists a $\tau < \delta_{\textup{GV}}(n,k)$ (this is the biggest value for which we can hope that decoding is successful with probability $1-o(1)$) such that $\tau^\perp < \omega_{\textup{easy}}(n,n-k)$. It turns out that in many cases we have to choose $\tau > \delta_{\textup{GV}}(n,k)/2$ meaning that we are not in the regime where the decoding has necessarily at most one solution. This complicates somehow the proof of the reduction since with a quantized version of the decoding algorithm we will not be able to produce at Step 1 the state $\frac{1}{Z}\sum_{\vec{c} \in \ensuremath{\mathscr{C}},\ev } \pi_{\ev} \ket{\vec{c} + \ev}$ (since decoding fails for some $\ev$) but we will be able to show that as long as $\tau < \delta_{\textup{GV}}(n,k)$ we will get a state close to this one. This will be enough for our purpose.
Moreover by building upon the proof technique of \cite{SSTX09} we can show a reduction which is more relevant to cryptography. We consider that we have a decoding algorithm which is only successful for some potentially very small probability $\varepsilon$ and we want to turn it into an algorithm outputting a word
of weight $\approx \tau^\perp n$ in the dual code $\ensuremath{\mathscr{C}}^\perp$ with some probability $\mathsf{poly}(\varepsilon)$. The ``ideal'' version of the algorithm that we have presented before (where we assume that we are always successful with our decoding algorithm) only describes a state that we get at Step 2 which is not completely orthogonal (the scalar product is bounded from below by a quantity $\mathsf{poly}(\varepsilon)$) to the ``real'' state after applying this approximate decoding process + quantum Fourier step. If we were to measure this state directly at that point we would not be sure to measure with probability $\mathsf{poly}(\varepsilon)$ a word of weight $\approx \tau^\perp n$. To ensure this, we have to apply a quantum amplification step that would produce in the ideal case a state which is concentrated on a certain weight $w$ close to $\tau^\perp n$. It is based on the fact that for a random code we know with probability $1 - 2^{-\Omega(n)}$ with exponential precision the probability of measuring a weight $w$ close to $\tau^\perp n$. Putting all these ingredients together we are able to prove the following result.
\begin{theorem}[informal]
The short codeword problem $\scodeP(q,n,n-k,w)$ reduces to the decoding problem $\textup{$\mathsf{DP}$}(q,n,k,t)$ for
$w = \tau^\perp n + O(1)$ where
\begin{eqnarray*}
\tau &\stackrel{\text{def}}{=} &\frac{t}{n}\\
\tau^\perp & \stackrel{\text{def}}{=} & \frac{\left(\sqrt{(q-1)(1-\tau)}-\sqrt{\tau}\right)^2}{q}
\end{eqnarray*}
\end{theorem}
It will turn out that for $q=2$ (see Section \ref{sec:useful}) we can find for any rate $R=\frac{k}{n}$ in $(0,1)$ a
$t < d_{\textup{GV}}(n,k)$ for which the corresponding $w$ is below $\omega_{\textup{easy}}(n,n-k) n$ (the reduction is useful in this case). Unfortunately, this is not true anymore when $q \geq 5$, where there is always a range for $R$ for which $w$ is above $\omega_{\textup{easy}}(n,n-k) n$ and this for any choice of $t < d_{\textup{GV}}(n,k)$: the reduction becomes useless in this case. Roughly speaking, when $q$ grows, the Hamming metric gets coarser (we have only $n+1$ different values for the metric on $\F_q^n$, whereas the size of the ambient space gets bigger) and this reflects in the fact that the range of values of $R$ where this reduction is useful gets smaller. The whole approach that we have followed here (properly choosing the error distribution, if needed go beyond the unique decoding radius for decoding, and apply a subsequent amplification step if needed) can of course be adapted to other metrics. It is easy for instance to apply it for the rank metric \cite{DRT21} which becomes increasingly popular in code-based cryptography, see for instance \cite{ABDGHRTZABBBO19,AABBBDGZCH19,BCGMM19,BGHM20}. This metric is even coarser: on $\F_q^{m \times n}$ there are only $1+\min(m,n)$ different values for the metric. In this case, it can be verified that the reduction is always useless (i.e. reduces to weights which are always easy to produce for a random linear code). However, it should be interesting to investigate it for metrics like the Lee metric (more or less the $L_1$ norm version of the Euclidean metric on $\mathbb{Z}_q^n$) which has also begun to find its way in code-based cryptography \cite{HW20} and should have a behavior closer to the Euclidean metric if the size of the alphabet grows with the code length.
\par{\bf Notation.}
For $a$ and $b$ integers with $a \leq b$, we denote by $\llbracket a,b \rrbracket$ the set of integers $\{a,a+1,\dots,b\}$.
Vectors are in {\em row notation} and they will be written with bold letters (such as $\ev$). Uppercase bold letters are used to denote matrices (such as $\mat{H}$). ${\mathcal S}_t$ is the sphere of radius $t$ around $0$ in $\F_q^n$ (for a metric $|\cdot|$ that will be clear from the context) and $S_t$ is its cardinality. $\mathsf{poly}(n)$ denotes a quantity which is an $\OO{n^a}$ for some constant $a$.
\section{Notation}
\section{Quantum Reduction from Sampling Short Codewords to Decoding}
\subsection{A general result}
We assume here that we have a probabilistic algorithm ${\mathcal A}$ that solves (sometimes) the decoding problem at distance $t$. Its inputs are a generator matrix $\mat{G}\in \F_q^{k \times n}$ of a code $\ensuremath{\mathscr{C}} \subseteq \F_q^{n}$
(\textit{i.e:} $\ensuremath{\mathscr{C}} = \{\vec{u} \mat{G}: \vec{u} \in \F_q^k\}$) and a noisy codeword $\vec{c}+\ev$ where
$\vec{c}$ belongs to $\ensuremath{\mathscr{C}}$. We denote by $\vec{w} \in \{0,1\}^{\ell}$ the internal coins of ${\mathcal A}$. It outputs with a certain probability $\varepsilon$, the ``right'' $\ev$ when being fed with $\vec{c} +\ev$ where $\vec{c}$ is chosen uniformly at random in $\ensuremath{\mathscr{C}}$ and $\ev$ is chosen uniformly at random among the errors of weight $t$:
$$
\varepsilon \stackrel{\text{def}}{=} \mathbb{P}_{\mat{G},\vec{c},\ev,\vec{w}}\left( {\mathcal A}(\mat{G},\vec{c}+\ev,\vec{w}) = \ev \right).
$$
We can implement ${\mathcal A}$ quantumly in the following way: it maps the quantum state $\ket{\ev}\ket{\vec{c}+\ev}\ket{\vec{w}}$ to $\ket{\ev - {\mathcal A}(\mat{G},\vec{c}+\ev,\vec{w})}\ket{\vec{c}+\ev}\ket{\vec{w}}$.
The quantum reduction starts by building the initial superposition
$$
\ket{\psi_0} \stackrel{\text{def}}{=} \frac{1}{\sqrt{2^\ell q^k}} \sum_{\ev \in \F_q^n} \sum_{\vec{c} \in \ensuremath{\mathscr{C}}} \sum_{\vec{w} \in \mathbb{F}_2^\ell} \pi_{\ev} \ket{\ev}\ket{\vec{c}}\ket{\vec{w}}
$$
where $\ket{\pi }\stackrel{\text{def}}{=} \sum_{\ev \in \F_q^n} \pi_{\ev} \ket{\ev}$ is some quantum superposition of errors.
The quantum algorithm which gives the reduction can be described as follows.
\begin{center}
{\bf Algorithm of the quantum reduction.}\\
\noindent\fbox{
\parbox{\textwidth}{
\begin{align}
&\text{Initial state preparation} & = &\;\;\;\frac{1}{\sqrt{2^\ell q^k}} \sum_{\ev \in \F_q^n}\sum_{\vec{c} \in \ensuremath{\mathscr{C}}} \sum_{\vec{w} \in \mathbb{F}_2^\ell} \pi_{\ev} \ket{\ev} \ket{\vec{c}}\ket{\vec{w}} \nonumber \\
&\text{adding $\ev$ to $\vec{c}$:} & \mapsto &\;\;\; \frac{1}{\sqrt{2^\ell q^k}} \sum_{\ev \in \F_q^n} \sum_{\vec{c} \in \ensuremath{\mathscr{C}}} \sum_{\vec{w} \in \mathbb{F}_2^\ell} \pi_{\ev} \ket{\ev}\ket{\vec{c} +\ev}\ket{\vec{w}} \nonumber \\
&\text{applying ${\mathcal A}$:} & \stackrel{{\mathcal A}}{\mapsto} &\;\;\; \frac{1}{\sqrt{2^\ell q^k}} \sum_{\ev \in \F_q^n} \sum_{\vec{c} \in \ensuremath{\mathscr{C}}} \sum_{\vec{w} \in \mathbb{F}_2^\ell} \pi_{\ev} \ket{\ev - {\mathcal A}(\mat{G},\vec{c}+\ev,\vec{w})} \ket{\vec{c}+\ev}\ket{\vec{w}} \label{eq:A} \\
&\text{QFT on the $2$nd register:} & \mapsto &\;\;\; \ket{\psi_{\Ac}^{\textup{QFT}}} \label{eq:QFT}\\
&\text{Amplification of amplitudes:} & \mapsto &\;\;\; \ket{\psi_{\Ac}^{\textup{Ampl}}} \label{eq:AAmpl}\\
&\text{measuring the whole state:} & \mapsto &\;\;\; \ket{\ev} \ket{\vec{c}^\perp}\ket{\vec{w}} \label{eq:measure}
\end{align}}}
\end{center}
We will now give a general theorem about an algorithm ${\mathcal A}$ of this kind and will show that it succeeds with probability $\mathsf{poly}(\varepsilon)$ to output a codeword of the dual code $\ensuremath{\mathscr{C}}^\perp$ of some weight $u$ by using only a polynomial number of calls to ${\mathcal A}$ when certain conditions are met.
\begin{restatable}{theorem}{thmain} \label{th:main}
Assume that $\ket{\pi}=\sum_{\ev \in \F_q^n} \ket{\ev}$ is radial and nonnegative, i.e $\pi_\ev = f(|\ev|)$ for some function $f$ and $\pi_\ev \geq 0$ for all $\ev \in \F_q^n$. Let $p_t \stackrel{\text{def}}{=} \sum_{\ev: |\ev|=t}=S_t f(t)^2$. $\widehat{\ket{\pi}} = \sum_{\ev\in\F_q^{n}} \widehat{\pi}_{\ev}\ket{\ev}$ is radial too and we let $f^{\perp}(u) = \widehat{\pi}_{\vec{e}}$ where $\vec{e}$ is any element of $\F_q^{n}$ of Hamming weight $u$.
Furthermore, assume that :
\begin{equation*}
\frac{\braket{\pi}{\vec{1}}^2}{q^{n-k}} = 2^{-\Omega(n)}, \quad \frac{q^{k}}{S_{u}} = 2^{-\Omega(n)} \quad \mbox{and} \quad S_{u}|f^{\perp}(u)|^{2} = \Omega\left(\frac{1}{\mathsf{poly}(n)}\right)
\;\;\text{for some $1 \leq u \leq n$}.
\end{equation*}
with $\ket{\vec{1}}$ being the (unnormalized) superposition of all errors : $\ket{\vec{1}} \stackrel{\text{def}}{=} \sum_{\ev \in \F_q^n} \ket{\ev}$.
Suppose that there exists an algorithm ${\mathcal A}$ solving the decoding problem with success probability $\varepsilon$.
Then, there exists a quantum algorithm making only a polynomial number of calls to ${\mathcal A}$ and to additional elementary $1$ or $2$ qubit gates which takes as input a generator matrix $\mat{G} \in \F_q^{k \times n}$ of $\ensuremath{\mathscr{C}}$ and outputs a codeword of weight $u$ in $\ensuremath{\mathscr{C}}^\perp$ with probability bigger than $\frac{p_{t}^{2}\varepsilon^{3}}{16} - O(p_{t}^{4}\varepsilon^{5}) - 2^{- \Omega(n)} - \OO{q^{-\min(k,n-k)}}$.
\end{restatable}
\begin{remark}
This theorem is stated here for the Hamming metric, but actually it can be applied to {\em any} metric for which the Fourier transform is radially preserving: it also applies for instance to the rank metric.
\end{remark}
This theorem will follow from a sequence of lemmas. Roughly speaking the proof uses the following steps.
\begin{itemize}
\item[Step 1.]
The first one explains that after applying ${\mathcal A}$ in the previous reduction we get a state
$$
\ket{\psi_{\Ac}} \stackrel{\text{def}}{=} \frac{1}{\sqrt{2^{\ell} q^{k}}}\sum_{\ev \in \F_q^n} \sum_{\vec{c}\in \ensuremath{\mathscr{C}}}\sum_{\vec{w} \in \mathbb{F}_2^{\ell}}\pi_{\ev} \ket{\ev - {\mathcal A}(\mat{G},\vec{c}+\ev,\vec{w})}\ket{\vec{c}+\ev}\ket{\vec{w}}
$$ which is sufficiently close to the ``disentangled'' state
\begin{equation}\label{eq:psiIdeal}
\ket{\psi_{\textup{ideal}}} \stackrel{\text{def}}{=} \frac{1}{\sqrt{Z}}\sum_{\ev \in \F_q^n} \sum_{\vec{c}\in\ensuremath{\mathscr{C}}}\sum_{\vec{w} \in \mathbb{F}_2^{\ell}}\pi_{\ev} \ket{\zero{n}} \ket{\vec{c}+\ev}\ket{\vec{w}}
\end{equation}
where $Z$ is a normalizing constant ensuring that the quantum state is indeed valid (\textit{i.e.} is of norm $1$).
\item[Step 2.] We then analyze the effect of the Fourier transform on the ``ideal state'' $\ket{\psi_{\textup{ideal}}}$, which gives a quantum state $\ket{\psi_{\textup{ideal}}^{\textup{QFT}}}$, and we study a subsequent measure of it. We namely prove that measuring it would output a codeword $\vec{c}^\perp$ in $\ensuremath{\mathscr{C}}^\perp$ of weight $u$ with probability $\frac{1}{\mathsf{poly}(n)}$.
\item[Step 3.]
The last amplitude step is a unitary transform ${\mathcal U}$ which when applied to $\ket{\psi_{\textup{ideal}}^{\textup{QFT}}}$ would output a state $\ket{\psi_{\text{ideal}}^{\textup{Ampl}}}$ whose norm is concentrated up to a negligible $2^{-\Omega(n)}$ term on codewords $\vec{c}^\perp$ of weight $u$. We prove in this step how this can be achieved by making only a polynomial number of calls to ${\mathcal A}$ under the assumptions of Theorem \ref{th:main} and finish the proof by proving that if we apply ${\mathcal U}$ to $\ket{\psi_{\Ac}^{\textup{QFT}}}$ then the conclusion of the theorem follows.
\end{itemize}
The reason why we use an amplification step is that because of the assumption on the success probability of our decoding algorithm we can only say that $\ket{\psi_{\Ac}^{\textup{QFT}}}$ is at trace distance $\leq 1-\eta$ of $\ket{\psi_{\textup{ideal}}^{\textup{QFT}}}$ for some $\eta > 0$ depending polynomially on $\varepsilon$. This implies that the distribution of outcomes after measurement between these states will be at statistical distance $\leq 1-\eta$ but even if measuring $\ket{\psi_{\textup{ideal}}^{\textup{QFT}}}$ would yield
a codeword $\vec{c}^\perp$ in $\ensuremath{\mathscr{C}}^\perp$ of weight $u$ with probability $\frac{1}{\mathsf{poly}(n)}$, measuring $\ket{\psi_{\Ac}^{\textup{QFT}}}$ might give such a codeword with probability $0$. This is not true anymore if we were to reason now on $\ket{\psi_{\text{ideal}}^{\textup{Ampl}}}$. It is at trace distance $\leq 1- \eta$ from
$\ket{\psi_{\Ac}^{\textup{Ampl}}}$, but now measuring $\ket{\psi_{\Ac}^{\textup{Ampl}}}$ yields necessarily a codeword $\vec{c}^\perp$ in $\ensuremath{\mathscr{C}}^\perp$ of weight $u$ with probability $ \geq \eta - 2^{\Omega(n)}$ since the norm of
$\ket{\psi_{\text{ideal}}^{\textup{Ampl}}}$ is exponentially concentrated around such codewords and the trace distance between the outcomes of measuring $\ket{\psi_{\Ac}^{\textup{Ampl}}}$ and $\ket{\psi_{\text{ideal}}^{\textup{Ampl}}}$ is also $\leq 1 - \eta$.
The detailed proofs of these steps are given in Section \ref{app:th:main} of the appendix.
\subsection{Application to the Hamming metric}
The assumptions of Theorem \ref{th:main} will be satisfied for the Hamming metric for a weight $u$ close to $\tau^{\perp} n$ and we will be able to prove that
\begin{theorem}\label{theo:Hamming}
Suppose that there exists an algorithm ${\mathcal A}$ solving with success probability $\varepsilon$ the decoding problem at Hamming distance $1 \leq t \stackrel{\text{def}}{=} \tau n \leq \min\left(\frac{n}{q},d_{\textup{GV}}(n,k)(1-\delta)\right)$ for some $\delta > 0$.
Then, there exists a quantum algorithm making only a polynomial number of calls to ${\mathcal A}$ and to additional elementary $1$ or $2$ qubit gates which takes as input $\mat{G}\in\F_q^{k\times n}$, a generator matrix of $\ensuremath{\mathscr{C}}$, and which outputs $\vec{c}^{\perp}\in \ensuremath{\mathscr{C}}^{\perp}$ of weight $u \stackrel{\text{def}}{=} (\tau^{\perp} + o(1))n$ with probability (over a uniform choice of $\mat{G}$)
bigger than $\frac{\varepsilon^{3}}{16} - O(\varepsilon^{5}) - 2^{-\Omega(n)}$ where:
\begin{equation*}
\tau^{\perp} \stackrel{\text{def}}{=} \frac{1}{q}\left( \sqrt{(q-1)(1-\tau)} - \sqrt{\tau}\right)^{2}
\end{equation*}
\end{theorem}
This theorem is proved in Section \ref{app:theo:Hamming} of the appendix.
\begin{remark}
The term $\frac{n}{q}$ which appears in the upper-bound for the range of values for which we can apply our reduction is of no importance in the most important case, namely in the binary case ($q=2$) and for larger values of $q$, $[1, \frac{n}{q}]$ contains a significant part of the values of $t$ for which the reduction is useful (see Section \ref{sec:useful}). This term $\frac{n}{q}$ comes from the way we estimated Krawtchouk polynomials in the interval formed by their first and last zeros. We could have used \cite{IS98} but this would involve lengthy computations (but would improve the $\frac{n}{q}$ term to $\frac{q-1}{q} n$). However the method we used, relying on a generalization of \cite[Prop. 25]{KS21} avoids a lot of computations and is much shorter.
\end{remark}
\section{About the usefulness of our reduction.}\label{sec:useful}
\begin{figure}[!h]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{TauPerpRF2.pdf}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{TauPerpRF7.pdf}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{TauPerpRF57.pdf}
\end{subfigure}
\caption{$\tau^{\perp}$ as function of $\tau$ for a rate $R = 1/2$.}
\label{fig:TauPerpRF}
\end{figure}
It is now interesting to look at the parameters for which this reduction is useful. In the example of Fig. \ref{fig:TauPerpRF}, when $q=2$, when we take a code of rate $R=\frac{1}{2}$, we can see that $\tau^\perp$ is always above $\omega_{easy}$ when $\tau < \delta_{GV}(R)/2$, hence the necessity of going beyond the unique decoding radius. An important point- at some point when $\tau$ goes beyond $\delta_{GV}(R)/2$, $\tau^\perp$ reaches $\omega_{easy}$ and goes below as $\tau$ goes to $\delta_{GV}(R)$, meaning that we are able with our reduction to get a codeword of a weight in the range where the problem is not known to be easy (above $\delta_{GV}(1-R)$ which corresponds to the smallest possible relative weight where there might be a solution and below $\omega_{easy}$). As $q$ grows, it appears that the range of values $\tau$ for which the reduction is useful shrinks. We can see that by taking a look at the cases where $q=7$ or $q=57$. In the latter, for example, the reduction is never useful.
\begin{figure}[!h]
\centering
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[width=\textwidth]{TauPerpR2.pdf}
\end{subfigure}
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[width=\textwidth]{TauPerpR7.pdf}
\end{subfigure}
\begin{subfigure}[b]{0.49\textwidth}
\includegraphics[width=\textwidth]{TauPerpR57.pdf}
\end{subfigure}
\caption{Optimal $\tau^{\perp}$ as function of $R$.}
\label{fig:TauPerpR}
\end{figure}
Now, if we plot $\tau^\perp$ against $R$ when $\tau = \delta_{GV}(R)$ (Fig. \ref{fig:TauPerpR}, when $q=2$), the range of values for which the short codeword problem is hard becomes more apparent. In fact, it is always between $\delta_{GV}(1-R)$ (below which there is no non-trivial codeword) and $\omega_{easy}$ (above which the problem is easy). As $q$ grows, when we plot $\tau^\perp$ for $q=7$ and $q=57$, we see that the range of values of $R$ for which $\tau^\perp$ is in the hard zone of the problem shrinks.
\section{Proof of Theorem \ref{th:main}}\label{app:th:main}
\subsection{Step 1.}
We are going to measure the distance between quantum states as in \cite{SSTX09} by using the trace distance that is closely connected to the (classical) statistical distance. The trace distance is defined as follows for pure quantum states:
\begin{equation}\label{eq:def_dist_tr}
D_{\textup{tr}}(\ket{\phi},\ket{\psi}) \stackrel{\text{def}}{=} \sqrt{1 -\left|\braket{\phi}{\psi}\right|^2}
\end{equation}
Two important properties of this distance are:
\begin{itemize}
\item[(i)] It can never increase after a quantum evolution \cite[\S9,Th. 9.1]{NC16};
\item[(ii)] The pair of probability distributions $(p_m,q_m)$ of the measurement outcome $m$ of any quantum measurement performed on the pair
of states $(\ket{\phi},\ket{\psi})$ satisfies \cite[\S9,Th. 9.2]{NC16}
\begin{equation}\label{eq:dstat_dtr}
D_{\textup{stat}}(p_m,q_m) \leq D_{\textup{tr}}(\ket{\phi},\ket{\psi})
\end{equation}
where $D_{\textup{stat}}$ is the statistical distance (also called the total variation distance) between two probability distributions. It is defined by:
$$
D_{\textup{stat}}(p,q) \stackrel{\text{def}}{=} \frac{1}{2} \sum_{x \in {\mathcal X}} |p(x)-q(x)|
$$
where $p$ and $q$ are two discrete probability distributions on ${\mathcal X}$.
\end{itemize}
In our case, the trace distance between $\ket{\psi_{\Ac}}$ and $\ket{\psi_{\textup{ideal}}}$ is upper-bounded in the following lemma.
\begin{lemma}\label{lem:dist_algo_ideal}
Let $\varepsilon_\mat{G}$ be the probability that ${\mathcal A}$ returns the right error $\ev$ when the input matrix is $\mat{G}$, i.e.
$$
\varepsilon_\mat{G} \stackrel{\text{def}}{=} \mathbb{P}_{\vec{c},\ev,\vec{w}}\left( {\mathcal A}(\mat{G},\vec{c}+\ev,\vec{w}) = \ev \right).
$$
We have
$$
D_{\textup{tr}}(\ket{\psi_{\Ac}},\ket{\psi_{\textup{ideal}}}) \leq \sqrt{1 - \frac{2^\ell q^k p_t^2}{Z} \varepsilon_\mat{G}^2 }
$$
\end{lemma}
\begin{proof}
Let ${\mathcal G}$ be the set of $(\vec{c},\ev,\vec{w})$'s that correspond to inputs to ${\mathcal A}$ that are correctly decoded:
$$
{\mathcal G} \stackrel{\text{def}}{=} \left\{ (\vec{c},\ev,\vec{w}) \in \ensuremath{\mathscr{C}} \times {\mathcal S}_t \times \mathbb{F}_2^{\ell} : {\mathcal A}(\mat{G},\vec{c}+\ev,\vec{w}) = \ev \right\}.
$$
We clearly have ($\pi_{\vec{e}}\geq 0$):
\begin{eqnarray*}
\braket{\psi_{\Ac}}{\psi_{\text{ideal}}} &\geq &\frac{1}{\sqrt{2^\ell q^k Z }} \sum_{(\vec{c},\ev,\vec{w})\in {\mathcal G}} \pi_{\ev}^2 \\
& = & \sqrt{\frac{2^\ell q^k}{Z}} S_t f(t)^2 \frac{\#{\mathcal G} }{2^\ell q^k S_t} \\
& = & \sqrt{\frac{2^\ell q^k}{Z}}p_t\; \varepsilon_\mat{G}.
\end{eqnarray*}
\end{proof}
All the probabilistic results of this section are easier to prove if instead of choosing a code $\ensuremath{\mathscr{C}}$ by picking uniformly at random a generator matrix $\mat{G}$ for it (\textit{i.e.} $\ensuremath{\mathscr{C}} \stackrel{\text{def}}{=} \{\vec{u} \mat{G}: \;\vec{u} \in \F_q^k\})$ we change slightly the probabilistic model by picking uniformly at random a parity-check matrix $\mat{H} \in \F_q^{(n-k)\times n}$ for it, \textit{i.e.}
$$\ensuremath{\mathscr{C}} =\{\vec{x} \in \F_q^n : \mat{H} \transpose{\vec{x}} = \mathbf{0}\}.
$$
We will denote $\mathbb{P}_{\mat{G}}$ and $\mathbb{P}_{\mat{H}}$ respectively the probabilities in the initial model and the probabilities in the new model.
The two probability distributions are closely related: the first model always produces linear codes of dimension $\leq k$ and codes of dimension $=k$ with probability $1-\OO{q^{-(n-k)}}$ whereas the second model always produces linear codes of dimension $\geq k$ and codes of dimension $=k$ with probability $1-\OO{q^{-k}}$. This relationship is expressed by the following lemma.
\begin{lemma}\label{lem:GvsH}
Let ${\pazocal E}$ be an ensemble of linear codes of length $n$ in $\F_q$. We have
$$
\mathbb{P}_{\mat{G}}({\pazocal E}) \leq \mathbb{P}_{\mat{H}}({\pazocal E})+\OO{q^{-\min(k,n-k)}}.
$$
\end{lemma}
With this new probabilistic model we can easily upper-bound the probability that $Z$ is bigger than $2^{\ell}q^k(1+\eta)$ for any $\eta > 0$.
\begin{lemma}
\label{lem:Z}
Let $\eta > 0$. We have:
$$
\mathbb{P}_{\mat{G}}(Z > 2^{\ell}q^k(1+\eta)) \leq \frac{1}{\eta}\; \frac{\braket{\pi}{\vec{1}}^2}{q^{n-k}} +\OO{q^{-\min(k,n-k)}}.
$$
\end{lemma}
\begin{proof}
This is obtained through Markov's inequality by computing $\mathbb{E}_\mat{H}(Z)$. We namely have
\begin{eqnarray*}
Z & = & \norm{\sum_{\vec{c} \in \ensuremath{\mathscr{C}},\ev \in \F_q^n,\vec{w} \in \mathbb{F}_2^\ell} \pi_{\ev} \ket{\zero{n}}\ket{\vec{c}+\ev}\ket{\vec{w}}}^2\\
& = & 2^\ell \norm{\sum_{\vec{c} \in \ensuremath{\mathscr{C}},\ev \in \F_q^n}\pi_{\ev} \ket{\vec{c}+\ev}}^2\\
& = & 2^\ell \left( q^k \sum_{\ev \in \F_q^n} \pi_{\ev}^2 + \sum_{{\substack{(\vec{c},\ev) \neq (\vec{c}',\ev') :\\ \vec{c}+\ev=\vec{c}'+\ev'}}} \pi_{\ev} \pi_{\ev'}\right)\\
& = & 2^\ell q^k \left( 1+ \sum_{\ev \neq \ev' : \mat{H} \transpose{\left(\ev-\ev'\right)}= \mathbf{0}} \pi_{\ev} \pi_{\ev'} \right)
\end{eqnarray*}
where $\mat{H}$ is an arbitrary-parity check matrix for $\ensuremath{\mathscr{C}}$. Let
$$X \stackrel{\text{def}}{=} \sum_{\ev \neq \ev' : \mat{H} \transpose{\left(\ev-\ev'\right)}=0} \pi_{\ev} \pi_{\ev'}.$$
The point of the probabilistic model where the parity-check matrix $\mat{H}$ is chosen uniformly at random is that for non-zero element $\vec{x} \in \F_q^n$ we have
$$
\mathbb{P}_{\mat{H}}(\vec{x} \in \ensuremath{\mathscr{C}}) = \mathbb{P}_\mat{H}(\mat{H} \transpose{\vec{x}}=0) = \frac{1}{q^{n-k}}.
$$ From this we deduce
\begin{eqnarray*}
\mathbb{E}_\mat{H}(X) & = & \sum_{\ev \neq \ev'} \pi_{\ev} \pi_{\ev'} \mathbb{P}_{\mat{H}}((\ev - \ev') \in \ensuremath{\mathscr{C}}) \\
& = & \sum_{\ev \neq \ev'} \frac{\pi_{\ev} \pi_{\ev'} }{q^{n-k}} \\
& \leq & \sum_{t,t'} \frac{f(t)f(t')\;S_t \;S_{t'}}{q^{n-k}}\\
& = & \frac{\left(\sum_t f(t)\;S_t \right)^2}{q^{n-k}}\\
& = & \frac{\braket{\pi}{\vec{1}}^2}{q^{n-k}}.
\end{eqnarray*}
Therefore,
\begin{eqnarray*}
\mathbb{P}_{\mat{G}}(Z > 2^{\ell}q^k(1+\eta)) & = & \mathbb{P}_{\mat{G}}(X>\eta) \\
& \leq & \mathbb{P}_{\mat{H}}(X>\eta) + \OO{q^{-\min(k,n-k)}} \quad\text{(by Lemma \ref{lem:GvsH})}\\
& \leq & \frac{1}{\eta}\; \mathbb{E}_{\mat{H}}(X) + \OO{q^{-\min(k,n-k)}} \quad\text{(Markov inequality) }\\
& \leq & \frac{1}{\eta} \; \frac{\braket{\pi}{\vec{1}}^2}{q^{n-k}} + \OO{q^{-\min(k,n-k)}}.
\end{eqnarray*}
\end{proof}
By putting Lemmas \ref{lem:dist_algo_ideal} and \ref{lem:Z} (with $\eta = 1$) together we immediately obtain the following proposition.
\begin{proposition}\label{prop:step1}
With probability greater than $ 1-\frac{\braket{\pi}{\vec{1}}^2}{q^{n-k}} -\OO{q^{-\min(k,n-k)}}$ over the choices of $\mat{G}$ we have:
$$
D_{\textup{tr}}(\ket{\psi_{\Ac}},\ket{\psi_{\textup{ideal}}}) \leq \sqrt{1 - \frac{p_t^2}{2} \varepsilon_\mat{G}^2 }.
$$
\end{proposition}
\subsection{Step 2.}
Recall that the quantum Fourier transform $\QFT{\psi}$ of a state $\ket{\psi} \stackrel{\text{def}}{=} \sum_{\vec{x} \in \F_q^n} \alpha_{\vec{x}} \ket{\vec{x}}$ is defined by using the characters $\chi_{\vec{y}}$ of the additive group $\F_q^n$ (there are as many characters as there are elements in $\F_q^n$ and we assume that the characteristic of $\F_q$ is the prime $p$ with
$q = p^s$)
\begin{eqnarray*}
\chi_{\vec{y}}(\vec{x}) & \stackrel{\text{def}}{=} & e^{\frac{2i \pi \Tr(\vec{x} \cdot \vec{y})}{p}}\;\;\text{where}\\
\vec{x} \cdot \vec{y} & \stackrel{\text{def}}{=} & \sum_{i=1}^n x_i y_i \;\;\text{with $\vec{x}=(x_i)_{i=1}^n$ and $\vec{y}=(y_i)_{i=1}^n$}\\
\Tr(a) & \stackrel{\text{def}}{=} & a + a^p + a^{p^2} + \cdot + a^{p^{s-1}}\\
\QFT{\psi} &\stackrel{\text{def}}{=} &\frac{1}{\sqrt{q^n}} \sum_{\vec{x}, \vec{y} \in \F_q^n} \alpha_{\vec{x}}\chi_{\vec{y}}(\vec{x}) \ket{\vec{y}}
\end{eqnarray*}
The dual code $\ensuremath{\mathscr{C}}^\perp$ of a linear code $\ensuremath{\mathscr{C}}$ over $\F_q$ is easily seen to be defined equivalently from the inner product $\vec{x} \cdot \vec{y}$ or from the characters as follows
\begin{eqnarray*}
\ensuremath{\mathscr{C}}^\perp & \stackrel{\text{def}}{=} & \{\vec{y} \in \F_q^n: \forall \vec{c} \in \ensuremath{\mathscr{C}}, \; \chi_{\vec{y}}(\vec{c})=1\}\\
& = & \{\vec{y} \in \F_q^n: \forall \vec{c} \in \ensuremath{\mathscr{C}},\;\vec{y} \cdot \vec{c} =0\}.
\end{eqnarray*}
If we apply the unitary $\mat{U}$ corresponding to the quantum Fourier transform on the second register of
$\ket{\psi_{\textup{ideal}}}$ (given in Equation \eqref{eq:psiIdeal}) it is readily seen that we obtain:
$$
\ket{\psi_{\textup{ideal}}^{\textup{QFT}}} \stackrel{\text{def}}{=} \left(\mat{Id} \otimes \mat{U} \otimes \mat{Id} \right)\ket{\psi_{\textup{ideal}}} = \frac{q^k}{\sqrt{Z}} \sum_{\vec{c}^\perp \in \ensuremath{\mathscr{C}}^\perp}
\sum_{\vec{w} \in \mathbb{F}_2^\ell} \reallywidehat{\pi}_{\vec{c}^\perp} \ket{\zero{n}}\ket{\vec{c}^\perp}\ket{\vec{w}},
$$
where $\QFT{\pi} = \sum_{\ev \in \F_q^n} \widehat{\pi}_{\ev} \ket{\ev}$ is the quantum Fourier transform of $\ket{\pi}$.
\begin{lemma}\label{lem:measure}
If the Fourier transform is radially preserving, meaning that it transforms a radial function into a radial function, then after measuring $\ket{\psi_{\textup{ideal}}^{\textup{QFT}}}$ we obtain a state $\ket{\zero{n}}\ket{\vec{c}^\perp}\ket{\vec{w}}$ with $\vec{c}^\perp\in\ensuremath{\mathscr{C}}^\perp$ of weight $u$ with probability
$\frac{2^\ell q^{2k}}{Z}N^\perp_u \left|f^\perp(u)\right|^2$
where $f^\perp(u) \stackrel{\text{def}}{=} \reallywidehat{\pi}_{\ev}$ for an arbitrary $\ev$ of weight $u$ and $N^\perp_u$ is the number of codewords of weight $u$ in $\ensuremath{\mathscr{C}}^\perp$.
\end{lemma}
\begin{proof}
For $\ev \in \F_q^n$, let
$$
\ket{\vec{1}_{\ensuremath{\mathscr{C}}+\ev}} \stackrel{\text{def}}{=} \sum_{\vec{c} \in \ensuremath{\mathscr{C}}} \ket{\vec{c}+\ev}.
$$
We have
\begin{eqnarray}
\QFT{\vec{1}_{\ensuremath{\mathscr{C}}+\ev}}& = & \sum_{\vec{c} \in \ensuremath{\mathscr{C}}} \frac{1}{\sqrt{q^n}} \sum_{\vec{y} \in \F_q^n}
\chi_\vec{y}(\vec{c}+\ev) \ket{\vec{y}} \nonumber\\
& = & \frac{1}{\sqrt{q^n}} \sum_{\vec{y} \in \F_q^n} \chi_{\vec{y}}(\ev) \sum_{\vec{c} \in \ensuremath{\mathscr{C}}}
\chi_{\vec{y}}(\vec{c}) \ket{\vec{y}}\nonumber \\
& = & \frac{q^k}{\sqrt{q^n}} \sum_{\vec{c}^\perp \in \ensuremath{\mathscr{C}}^\perp} \chi_{\vec{c}^\perp}(\ev) \ket{\vec{c}^\perp}\;\;\text{ (since $\sum_{\vec{c} \in \ensuremath{\mathscr{C}}}
\chi_\vec{y}(\vec{c})=0$ if $\vec{y} \notin \ensuremath{\mathscr{C}}^\perp$ and $q^k$ else)}\label{eq:QFT_coset}
\end{eqnarray}
Therefore
\begin{eqnarray*}
\QFT{\mathop{\sum}\limits_{\ev \in \F_q^n,\vec{c} \in \ensuremath{\mathscr{C}}}\pi_{\ev} \ket{\vec{c}+\ev}} &= & \sum_{\ev \in \F_q^n} \pi_{\ev} \QFT{\vec{1}_{\ensuremath{\mathscr{C}}+\ev}}\\
& = & \frac{q^k}{\sqrt{q^n}} \sum_{\ev \in \F_q^n}\pi_{\ev} \sum_{\vec{c}^\perp \in \ensuremath{\mathscr{C}}^\perp} \chi_{\vec{c}^\perp}(\ev) \ket{\vec{c}^\perp}\\
& = & q^k \sum_{\vec{c}^\perp \in \ensuremath{\mathscr{C}}^\perp} \frac{1}{\sqrt{q^n}}\sum_{\ev \in \F_q^n}\pi_{\ev} \chi_{\vec{c}^\perp}(\ev) \ket{\vec{c}^\perp}\\
& = & q^k \sum_{\vec{c}^\perp \in \ensuremath{\mathscr{C}}^\perp} \reallywidehat{\pi}_{\vec{c}^\perp} \ket{\vec{c}^\perp},
\end{eqnarray*}
The lemma directly follows from this last equation.
\end{proof}
\subsection{Step 3.}
For this we first need to have a good estimation of $\ket{\psi_{\textup{ideal}}^{\textup{QFT}}}$'s amplitudes.
This will be a consequence of the following lemma.
\begin{lemma}\label{lem:Nperp}
If the generator matrix $\mat{G}$ of a code $\ensuremath{\mathscr{C}}$ is chosen uniformly at random in $\F_q^{k \times n}$ then the number $N^\perp_u$ of codewords of weight $u$ in $\ensuremath{\mathscr{C}}^\perp$ satisfies
$$\mathbb{P}\left(\left| N^\perp_u - \frac{S_{u}}{q^{k}} \right| \geq \left( \frac{S_{u}}{q^{k}}\right)^{3/4} \right) \leq (q-1)\; \sqrt{\frac{q^k}{S_u}}.$$
\end{lemma}
\begin{proof}\label{lemma:BT}
Let $\mathds{1}_{\vec{x}}$ be the indicator function of the event ``$\vec{x}\in\ensuremath{\mathscr{C}}^\perp$''. By definition,
\begin{equation}
N_u^\perp = \sum_{\vec{x}\in {\mathcal S}_u}\mathds{1}_{x}
\end{equation}
It is clear that $\mathbb{E}(\mathds{1}_\vec{x})=\mathbb{P}(\vec{x} \in \ensuremath{\mathscr{C}}^\perp)= \frac{1}{q^{k}}$. Therefore, $\mathbb{E}(N_u^\perp) = \frac{S_u}{q^k}$.
By using Bienaym\'{e}-Tchebychev’s inequality, we obtain:
\begin{align}
\mathbb{P}\left(\left|N^\perp_u -\frac {S_u}{q^k}\right| \geq a\right) &\leq \frac{\mathbf{Var}(N_{u}^{\perp})}{a^{2}} \nonumber \\
&= \frac{1}{a^{2}}\left( {\sum_{\vec{x}\in {\mathcal S}_u}\mathbf{Var}(\mathds{1}_{\vec{x}}) + \sum_{\substack{\vec{x},\vec{y}\in {\mathcal S}_{u}\\ \vec{x}\neq \vec{y} }} \mathbb{E}(\mathds{1}_{\vec{x}}\mathds{1}_{\vec{y}}) - \mathbb{E}(\mathds{1}_{\vec{x}})\mathbb{E}(\mathds{1}_{\vec{y}}) } \right)\nonumber \\
&\leq \frac{1}{a^2}\left( \sum_{\vec{x} \in {\mathcal S}_u} \mathbb{E}(\mathds{1}_\vec{x})+ \sum_{\substack{\vec{x},\vec{y}\in {\mathcal S}_u\\ \vec{x}\neq \vec{y} }} \mathbb{E}(\mathds{1}_{\vec{x}}\mathds{1}_{\vec{y}}) - \mathbb{E}(\mathds{1}_{\vec{x}})\mathbb{E}(\mathds{1}_{\vec{y}}) \right) \nonumber\\
&=\frac{1}{a^{2}}\left( \frac{S_u}{q^k}+ \sum_{\substack{\vec{x},\vec{y}\in {\mathcal S}_u\\ \vec{x}\neq \vec{y} }} \mathbb{E}(\mathds{1}_{\vec{x}}\mathds{1}_{\vec{y}}) - \mathbb{E}(\mathds{1}_{\vec{x}})\mathbb{E}(\mathds{1}_{\vec{y}}) \right) \label{eq:varD}
\end{align}
where we used that $\mathbf{Var}(\mathds{1}_{\vec{x}}) \leq \mathbb{E}(\mathds{1}_{\vec{x}}^{2}) = \mathbb{E}(\mathds{1}_{\vec{x}})$. Let us now upper-bound the second term of the inequality. It is readily verified that:
\begin{equation*}
\mathbb{E}(\mathds{1}_{\vec{x}}\mathds{1}_{\vec{y}}) = \left\{
\begin{array}{ll}
\nicefrac{1}{q^{k}}& \mbox{ if } \vec{x} \mbox{ and } \vec{y} \mbox{ are colinear} \\
\nicefrac{1}{q^{2k}} & \mbox{ otherwise.}
\end{array}
\right.
\end{equation*}
Therefore, we deduce that:
\begin{align}
\sum_{\substack{\vec{x},\vec{y}\in {\mathcal S}_u\\ \vec{x}\neq \vec{y} }} \mathbb{E}(\mathds{1}_{\vec{x}}\mathds{1}_{\vec{y}}) - \mathbb{E}(\mathds{1}_{\vec{x}})\mathbb{E}(\mathds{1}_{\vec{y}}) &= \sum_{\vec{x}\in {\mathcal S}_u}\sum_{\substack{\vec{y}\in {\mathcal S}_u\setminus\vec{x}: \\ \text{ colinear to }\vec{x} }} \frac{1}{q^{k}} - \frac{1}{q^{2k}} \nonumber\\
&\leq \sum_{\vec{x}\in {\mathcal S}_u}\sum_{\substack{\vec{y}\in {\mathcal S}_u\setminus\vec{x}: \\ \text{ colinear to }\vec{x} }} \frac{1}{q^{k}} \nonumber \\
&\leq \frac{(q-2) S_u}{q^{k}} \label{eq:finVar}
\end{align}
It gives by plugging \eqref{eq:finVar} in \eqref{eq:varD} :
\begin{align*}
\mathbb{P}\left(\left|N^\perp_u -\frac {S_u}{q^k}\right| \geq a\right) &\leq \frac{1}{a^{2}}\left( \frac{S_u}{q^k}+ \frac{(q-2)S_u}{q^k}\right)\\
&= \frac{(q-1)S_u}{a^{2}q^k}
\end{align*}
which concludes the proof by choosing $a=\left( \frac{S_u}{q^k} \right)^{3/4}$.
\end{proof}
The point of this lemma is that it will turn out that (where $d_{\textup{GV}}^{+}(n,n-k)$ denotes the largest integer $t$ such that $q^{n-k} \cdot S_t \geq q^n$) we will choose a weight $u$ such that it is above $d_{\textup{GV}}(n,n-k)$ and below $d_{\textup{GV}}^{+}(n,n-k)$ and then for many metrics
the term $\frac{q^k}{S_u}$ is exponentially small.
Therefore, the probability of measuring $\vec{c}^{\perp}\in \ensuremath{\mathscr{C}}^{\perp}$ of weight $u$ after measuring $\ket{\psi_{\textup{ideal}}^{\textup{QFT}}}$ is known up to an exponentially small factor.
Recall now that quantum amplitude amplification techniques (with a classical additional tweak) \cite[\S 2.1]{BHMT02} enable to turn a quantum algorithm, that performs no measurement and which succeeds with probability $p$, into a quantum algorithm that succeeds with probability {\em exactly} one in roughly $\sqrt{p}$ iterations. The same result algorithms carries over almost verbatim to give an algorithm working with probability very close to $1$ if $p$ is only known with good precision. Let us give here
the corresponding statement with the corresponding proof for the reader's convenience.
\begin{lemma}\label{lemma:amplQ}
Let $\mathcal{B}$ be any quantum algorithm that performs no measurement.
We suppose that measuring the output of $\mathcal{B}$ gives a solution
with probability $p$ (in this case we say that $\mathcal{B}$ succeeds). Furthermore, we suppose that $p$ is unknown but we know $q = \Omega\left( \frac{1}{\mathsf{poly}(n)}\right)$ such that:
\begin{equation}
p \in ((1-\delta)q,(1+\delta)q) .
\end{equation}
Then, there exists a quantum algorithm that runs a $\mathsf{poly}(n)$ number of times $\mathcal{B}$ and $\mathcal{B}^{-1}$ (and uses $\mathsf{poly}(n)$ other gates) such that measuring the output of it gives
a solution
with probability $\geq 1 - \mathsf{poly}(n) O(\delta^{2})$.
\end{lemma}
\begin{proof} From any quantum algorithm that succeeds with probability $a$, it is easy to build (by adding a qubit and a rotation) a new quantum algorithm that succeeds with probability $\alpha a $ for any chosen $\alpha\in[0,1]$ \cite[\S 2.1]{BHMT02}.
By using this tweak, we can construct an algorithm ${\mathcal B}'$ outputting a solution of ${\mathcal B}$ with probability $p' = \alpha p$ for an $\alpha \in [0,1]$ that we will choose later on.
Let $\theta \stackrel{\text{def}}{=} \arcsin \sqrt{ \alpha p}$. By
hypothesis the output of $\mathcal{B}'$ is equal to:
$$
\sin \theta\ket{G} + \cos \theta \ket{B}
$$
where $\ket{G}$ denotes the quantum superposition of solutions and $\ket{B}$ a state orthogonal to it.
Now by making $T$ steps of amplification, we obtain the following quantum state:
$$
\sin((2T+1)\theta) \ket{G} + \cos((2T+1)\theta) \ket{B}.
$$
The probability of measuring a solution is given by $\sin^{2}((2T+1)\theta)$.
We choose $\alpha$ as the largest $\alpha \in [0,1]$ such that
\begin{equation}\label{eq:kZ}
T \stackrel{\text{def}}{=} \frac{\pi}{4 \rho}-\frac{1}{2} \in \mathbb{Z^+}
\quad \text{where $\rho \stackrel{\text{def}}{=} \arcsin \sqrt{\alpha q}$.}
\end{equation}
Clearly $T=\mathsf{poly}(n)$.
Let us compute now the success probability $p_{\textup{succ}}$ after $T$ steps of amplification. We have the following computation:
\begin{align}
p_{\textup{succ}}
&= \sin^{2}((2T+1)\theta) \nonumber \\
&= \sin^{2}\left((2T+1)\rho + (2T+1)(\theta-\rho)\right) \nonumber \\
&= \cos^{2}\left( (2T+1)(\theta-\rho) \right) \quad (\mbox{by Equation \eqref{eq:kZ}}) \nonumber \\
&\geq 1 - \left( (2T+1)(\theta-\rho) \right)^{2} \nonumber\\
&\geq 1 - \mathsf{poly}(n)(\theta-\rho)^{2} \nonumber \\
& \geq 1 - \mathsf{poly}(n) \delta^2. \nonumber
\end{align}
\end{proof}
We will make the following assumption in the whole section from now on.
\begin{assumption}\label{eq:hypot}
The error distribution $\pi = (\pi_{\vec{e}})_{\vec{e}\in\F_q^{n}}$ and $1 \leq u \leq n$ verify:
\begin{equation*}
\frac{\braket{\pi}{\vec{1}}^2}{q^{n-k}} = 2^{-\Omega(n)}, \quad \frac{q^{k}}{S_{u}} = 2^{-\Omega(n)} \quad \mbox{and} \quad S_{u}|f^{\perp}(u)|^{2} = \Omega\left(\frac{1}{\mathsf{poly}(n)}\right).
\end{equation*}
\end{assumption}
We have now the following proposition.
\begin{proposition}\label{prop:step2}
If the Fourier transform is radially preserving, then under Assumption \ref{eq:hypot} it exists a quantum algorithm such that when starting from $\ket{\psi_{\textup{ideal}}}$, then for a proportion $\geq 1- \beta(\pi) $ of matrices $\mat{G}$, the probability of obtaining a codeword $\vec{c}^\perp$ of weight $u$ in $\ensuremath{\mathscr{C}}^\perp$ when measuring is greater than or equal to $1 - \alpha(\pi)$
where:
\begin{eqnarray*}
\alpha(\pi) &\stackrel{\text{def}}{=}& \widetilde{O}\left( \left( \frac{q^{k}}{S_{u}} \right)^{1/4} + \sqrt{\frac{\braket{\pi}{\vec{1}}^{2}}{q^{n-k}}} \right)^{2} \\
\beta(\pi) &\stackrel{\text{def}}{=}& (q-1)\sqrt{\frac{q^{k}}{S_{u}}} + \sqrt{\frac{\braket{\pi}{\vec{1}}^2}{q^{n-k}}} +\OO{q^{-\min(k,n-k)}}
\end{eqnarray*}
\end{proposition}
\begin{proof} Let $\mathcal{B}$ be the quantum algorithm starting from $\ket{\psi}$ which computes $(\mat{Id}\otimes \mat{U} \otimes \mat{Id})\ket{\psi}$. This algorithm succeeds when measuring a dual codeword $\vec{c}^{\perp}\in\ensuremath{\mathscr{C}}^{\perp}$ of weight $u$. When starting with $\ket{\psi_{\textup{ideal}}}$, the probability of success of $\mathcal{B}$ is equal to $\frac{2^\ell q^{2k} N^\perp_{u}}{Z}|f^\perp(u)|^2$ by Lemma \ref{lem:measure}. Let,
$$
\mathcal{E} \stackrel{\text{def}}{=} \left\{ \mat{G} \in \F_q^{k\times n} \mbox{ : } Z > 2^{\ell}q^{k}\left(1+\sqrt{\frac{\braket{\pi}{\vec{1}}^2}{q^{n-k}}}\right) \mbox{ or } \left| N^\perp_{u} - \frac{S_{u}}{q^{k}} \right| \geq \left( \frac{S_{u}}{q^{k}} \right)^{3/4} \right\}.
$$
By Lemmas \ref{lem:Z} and \ref{lem:Nperp} we have that $\mathbb{P}(\mat{G} \in \mathcal{E}) \leq \beta(\pi) = (q-1)\sqrt{\frac{q^{k}}{S_{u}}} + \sqrt{\frac{\braket{\pi}{\vec{1}}^2}{q^{n-k}}} +\OO{q^{-\min(k,n-k)}}$. Therefore, for a proportion $\geq 1 - \beta(\pi)$ of codes (over matrices $\mat{G}$):
\begin{enumerate}
\item[$(i)$] $Z \leq 2^{\ell}q^{k}\left(1+ \sqrt{\frac{\braket{\pi}{\vec{1}}^2}{q^{n-k}}}\right)$ and $Z \geq 2^{\ell}q^k$ (this is true for any $\mat{G}$ as $\pi_{\ev} \geq 0$ for any $\ev$),
\item[$(ii)$] $\left| \frac{q^{k}N_{u}^{\perp}}{S_{u}} - 1 \right| \leq \left( \frac{q^{k}}{S_{u}} \right)^{1/4}$.
\end{enumerate}
Therefore, we have for a proportion $\geq 1 - \beta(\pi)$ of codes:
\begin{equation}\label{eq:encProb}
S_{u} |f^{\perp}(u)|^{2} \; \left( \frac{1 - \left(\frac{q^{k}}{S_{u}}\right)^{1/4}}{1+\sqrt{\frac{\braket{\pi}{\vec{1}}^2}{q^{n-k}}} } \right)\leq \frac{2^\ell q^{2k} N^\perp_{u}}{Z}|f^\perp(u)|^2 \leq S_{u} |f^{\perp}(u)|^{2}\left(1 + \left(\frac{q^{k}}{S_{u}}\right)^{1/4} \right).
\end{equation}
Now,
$$
\frac{1 - \left(\frac{q^{k}}{S_{u}}\right)^{1/4}}{1+\sqrt{\frac{\braket{\pi}{\vec{1}}^2}{q^{n-k}}} } \geq 1 - \left(\frac{q^{k}}{S_{u}}\right)^{1/4} - \sqrt{\frac{\braket{\pi}{\vec{1}}^2}{q^{n-k}}}
$$
Therefore, by plugging this in Equation \eqref{eq:encProb} we have for a proportion $\geq 1- \beta(\pi)$ of codes:
\begin{equation}
\frac{2^\ell q^{2k} N^\perp_{u}}{Z}|f^\perp(u)|^2 \in S_{u} |f^{\perp}(u)|^{2}\;(1-\delta,1+\delta) \quad \mbox{where} \quad \delta \stackrel{\text{def}}{=} \left(\frac{q^{k}}{S_{u}}\right)^{1/4} + \sqrt{\frac{\braket{\pi}{\vec{1}}^2}{q^{n-k}}}
\end{equation}
By Assumption \eqref{eq:hypot}, $S_{u}|f^{\perp}(u)|^{2}= \Omega\left(\frac{1}{\mathsf{poly}(n)}\right)$, $\delta = 2^{-\Omega(n)}$ while $\frac{2^\ell q^{2k} N^\perp_{u}}{Z}|f^\perp(u)|^2$ denotes the success probability of $\mathcal{B}$. Therefore to conclude the proof we apply Lemma \ref{lemma:amplQ}.
\end{proof}
\subsection{Proof of Theorem \ref{th:main} }\label{proof:main}
{\bf \noindent Notation.} From now on, $\ket{\psi_{\text{ideal}}^{\textup{Ampl}}}$ ({\em resp.} $\ket{\psi_{\Ac}^{\textup{Ampl}}}$) will denote the quantum state after applying the quantum amplification algorithm of Proposition \ref{prop:step2} on $\ket{\psi_{\textup{ideal}}^{\textup{QFT}}}$ ({\em resp.} $\ket{\psi_{\Ac}^{\textup{QFT}}}$).
With this notation we are ready now to prove Theorem \ref{th:main} which we now recall
\thmain*
We start the proof by proving the following point
\begin{lemma}\label{lem:good}
Call ${\mathcal G}$ the set of ``good matrices'' $\mat{G} \in \F_q^{k \times n}$ that satisfy at the same time:\\
(i) $\varepsilon_{\mat{G}} \geq \varepsilon/2$,\\
(ii) $Z \leq 2^{\ell+1}q^k$\\
The proportion of good matrices is at least $\varepsilon/2 - \delta(\pi)$ where
$\delta(\pi)\stackrel{\text{def}}{=}
\frac{\braket{\pi}{\vec{1}}^2}{q^{n-k}} +\OO{q^{-\min(k,n-k)}}$.
\end{lemma}
\begin{proof} By definition,
$$
\varepsilon = \frac{1}{q^{kn}}\sum_{\mat{G}\in\F_q^{k\times n}} \varepsilon_{\mat{G}}.
$$
Let ${\mathcal B}$ be the set of matrices $\mat{G}$ that are not good, namely for which $(a)$ $\varepsilon_{\mat{G}} < \varepsilon/2$ or $(b)$ $Z > 2^{\ell + 1}q^{k}$. By Lemma \ref{lem:Z}, the density of matrices verifying $(b)$ is smaller than $\delta(\pi)$. Therefore,
$$
\varepsilon \leq \frac{1}{q^{kn}}\sum_{\mat{G}\notin {\mathcal B}} 1 + \delta(\pi)\; \frac{\varepsilon}{2} \leq \frac{1}{q^{kn}}\sum_{\mat{G}\notin {\mathcal B}} 1 + \delta(\pi) + \frac{\varepsilon}{2}
$$
which concludes the proof.
\end{proof}
We use this lemma to prove that the statistical distance between the distributions of weights $\abs{\vec{c}^\perp}$ we obtain by measuring $\ket{\psi_{\Ac}^{\textup{Ampl}}}$ can not be to far away from the distribution of weights when we measure the state $\ket{\psi_{\text{ideal}}^{\textup{Ampl}}}$:
\begin{lemma}\label{lem:nottofaraway}
Let $P$, respectively $Q$, be the distribution of the weights $|\vec{c}^\perp|$ of the state $\ket{\ev}\ket{\vec{c}^\perp}\ket{\vec{w}}$ obtained by measuring the state $\ket{\psi_{\Ac}^{\textup{Ampl}}}$, respectively $\ket{\psi_{\text{ideal}}^{\textup{Ampl}}}$. We have
$$
D_{\textup{stat}}(P,Q) \leq 1- \frac{p_t^2 \varepsilon^3}{16} + \OO{p_t^4 \varepsilon^5}+ \delta(\pi).
$$
\end{lemma}
\begin{proof}
Let
\begin{eqnarray*}
P(u|\mat{G}) & \stackrel{\text{def}}{=} & \mathbb{P}_{\vec{c},\ev,\vec{w}}\left(\text{measuring $\ket{\vec{c}^\perp}$ of weight $u$ in the $2$nd register of $\ket{\psi_{\Ac}^{\textup{Ampl}}}$ for a code choice $\mat{G}$} \right)\\
Q(u|\mat{G}) & \stackrel{\text{def}}{=} & \mathbb{P}_{\vec{c},\ev,\vec{w}}\left(\text{measuring $\ket{\vec{c}^\perp}$ of weight $u$ in the $2$nd register of $\ket{\psi_{\text{ideal}}^{\textup{Ampl}}}$ for a code choice $\mat{G}$} \right)
\end{eqnarray*}
We start the proof by noticing that
\begin{eqnarray*}
D_{\textup{stat}}(P,Q) & = & \frac{1}{2} \sum_u \left|P(u)-Q(u)\right| = \frac{1}{2} \sum_u \left| \sum_{\mat{G} \in \F_q^{k \times n}} \frac{1}{q^{kn}} \left(P(u|\mat{G}) - Q(u|\mat{G})\right)\right| \\
& \leq & \frac{1}{q^{kn}}\sum_{\mat{G} \in \F_q^{k \times n}} \frac{1}{2} \sum_u \left| P(u|\mat{G}) - Q(u|\mat{G}) \right|\\
&= &\frac{1}{q^{kn}} \sum_{\mat{G} \in \F_q^{k \times n}} D_{\textup{stat}}\left( P(u|\mat{G}) ,Q(u|\mat{G}) \right)\\
& \leq & \sum_{\mat{G} \in {\mathcal G}} \frac{D_{\textup{stat}}\left( P(u|\mat{G}) ,Q(u|\mat{G}) \right)}{q^{kn}} + \sum_{\mat{G} \notin {\mathcal G}} \frac{D_{\textup{stat}}\left( P(u|\mat{G}) ,Q(u|\mat{G}) \right)}{q^{kn}}\\
& \leq & \sum_{\mat{G} \in {\mathcal G}} \frac{D_{\textup{tr}}\left( \ket{\psi_{\Ac}},\ket{\psi_{\textup{ideal}}} \right)}{q^{kn}} + \sum_{\mat{G} \notin {\mathcal G}} \frac{1}{q^{kn}} \\
& \leq & \sum_{\mat{G} \in {\mathcal G}} \frac{\sqrt{1 - \frac{p_t^2 \varepsilon^2}{4}}}{q^{kn}} + \sum_{\mat{G} \notin {\mathcal G}} \frac{1}{q^{kn}}\\
& = & \sqrt{1 - \frac{p_t^2 \varepsilon^2}{4}} \mathbb{P}(\mat{G} \in {\mathcal G}) + \mathbb{P}(\mat{G} \notin {\mathcal G}) \\
& \leq &\left(\varepsilon/2 -\delta(\pi)\right)\left( 1- \frac{p_t^2 \varepsilon^2}{8} +\OO{p_t^4 \varepsilon^4} \right)+ 1 - \varepsilon/2 + \delta(\pi)\\
& \leq & 1- \frac{p_t^2 \varepsilon^3}{16} + \OO{p_t^4 \varepsilon^5}+ \delta(\pi).
\end{eqnarray*}
\end{proof}
We are now ready to prove Theorem \ref{th:main}.
\begin{proof}[Proof of Theorem \ref{th:main}]
By Proposition \ref{prop:step2} we know that
$$
Q(u)
\geq 1-\alpha(\pi))(1-\beta(\pi)) \geq 1 - \alpha(\pi) - \beta(\pi).
$$
But now we have the following computation,
\begin{align*}
P(u) &\geq Q(u) - D_{\textup{stat}}(P,Q) \\
&\geq 1 - \alpha(\pi) - \beta(\pi) - 1 + \frac{p_{t}^{2}\varepsilon^{3}}{16} - O(p_{t}^{4}\varepsilon^{5}) - \delta(\pi) \\
&= \frac{p_{t}^{2}\varepsilon^{3}}{16} - O(p_{t}^{4}\varepsilon^{5}) - \alpha(\pi) - \beta(\pi) - \delta(\pi)
\end{align*}
which concludes the proof by definition of $\alpha(\pi)$, $\beta(\pi)$ and $\delta(\pi)$.
\end{proof}
\section{Proof of Theorem \ref{theo:Hamming}}\label{app:theo:Hamming}
Recall that we have chosen $\ket{\pi}$ as
\begin{equation}\label{eq:piUnif}
\ket{\pi} = \frac{1}{\sqrt{\binom{n}{t}(q-1)^{t}}} \sum_{\ev \in \F_q^n:|\ev|=t} \ket{\ev}.
\end{equation}
This time, as long as $t$ is below $(1-\delta)d_{\textup{GV}}(n,k)$ for an arbitrary $\delta >0$ the term $\frac{\langle \pi,\vec{1} \rangle^{2}}{q^{n-k}}$ is exponentially small in $n$ as shown by
\begin{lemma}\label{lemma:GV} If $t \leq (1-\delta)d_{\textup{GV}}(n,k)$ then
$$
\frac{\langle \pi,\vec{1}\rangle^{2}}{q^{n-k}} = \frac{S_{t}}{q^{n-k}} = q^{\alpha(R,\delta) n(1+o(1))} \quad \mbox{where} \quad \alpha(R,\delta) <0.
$$
\end{lemma}
\begin{proof}
Recall that the size $B_t$ of the Hamming ball of radius $t$ is of the form
$$
B_t = q^{n\;h_q(\tau)(1+o(1))}
$$
where $h_q(x)= -(1-x)\log_q(1-x) - x \log_q\left( \frac{x}{q-1}\right)$ and $\tau = t/n$. We obtain from this
\begin{eqnarray*}
\frac{\langle \pi,\vec{1} \rangle^{2}}{q^{n-k}} &= & \frac{S_t}{q^{n-k}} \\
& \leq & \frac{B_t}{B_{d_{\textup{GV}}}} \qquad\text{(since $S_t \leq B_t$ and $B_{d_{\textup{GV}}} \leq q^{n-k}$)}\\
& \leq & q^{n(h_q(\tau)-h_q(\delta_{\textup{GV}})+o(1))}\\
& \leq & q^{n(h_q((1-\delta)\delta_{\textup{GV}})-h_q(\delta_{\textup{GV}})+o(1))}
\end{eqnarray*}
with $\delta_{\textup{GV}} \stackrel{\text{def}}{=} d_{\textup{GV}}(n,k)/n$. We finish the proof by noticing that
$h_q((1-\delta)\delta_{\textup{GV}})-h_q(\delta_{\textup{GV}})<0$.
\end{proof}
The Fourier transform $\QFT{\pi}$ of $\ket{\pi}$ can be expressed in terms of Krawtchouk polynomials as follows.
\begin{lemma}\label{lem:Krawtchouk}
$$
\QFT{\pi} = \frac{1}{\sqrt{q^{n}}} \sumFqn{y} \frac{K_{t}(\wt{y})}{\sqrt{\binom{n}{t}(q-1)^{t}}} \ket{\vec{y}}
$$
where
$$
K_{t}(X;q,n) \stackrel{\text{def}}{=} \sum_{j=0}^{t} (-1)^j \binom{X}{j} \binom{n-X}{t-j} (q-1)^{t-j}
$$
is the Krawtchouk polynomial of order $n$, parameter $q$ and degree $t\in \llbracket 0,n\rrbracket$
\footnote{To simplify notation, since $q$ and $n$ are clear here from the context: they are respectively the field size and the length of the codes we consider, we will drop the dependency in $q$ and $n$ and simply write
$K_t(X)$.}.
\end{lemma}
\begin{proof}
By definition of the Fourier transform we have
\begin{equation}
\label{eq:qft}
\QFT{\pi} = \frac{1}{\sqrt{q^{n}}}\sum_{\vec{y} \in \F_q^n} \frac{1}{\sqrt{\binom{n}{t}(q-1)^{t}}}
\sum_{\substack{\ev\in\F_q^{n}\\|\ev|=t}} \chi_\vec{y}(\ev) \ket{\vec{y}}.
\end{equation}
The sum of characters $\sum_{\substack{\ev\in\F_q^{n}\\|\ev|=t}} \chi_\vec{y}(\ev)$ can be expressed as a Krawtchouk polynomial evaluation (see for instance Lemma 5.3.1 in \cite[§5.3]{L99}):
$$
\sum_{\substack{\ev\in\F_q^{n}\\|\ev|=t}} \chi_\vec{y}(\ev) = K_{t}(\wt{y}).
$$
To finish the proof we just have to substitute the character sum for this expression in \eqref{eq:qft}.
\end{proof}
Therefore, for the error distribution $\pi$ we have:
\begin{equation}
S_{u}|f^{\perp}(u)|^{2} = \frac{1}{q^{n}} \; \binom{n}{u}(q-1)^{u}\; \frac{K_{t}(u)^{2}}{(q-1)^{t}\binom{n}{t}}.
\end{equation}
To apply Theorem \ref{th:main} to the Hamming case it remains now to understand how the Krawtchouk polynomial evaluations behave. In particular, if we find some $u$ such that:
$$
S_{u}|f^{\perp}(u)|^{2} = \Omega\left(\frac{1}{\mathsf{poly}(n)}\right) \quad \mbox{and} \quad \frac{q^{k}}{S_{u}} = 2^{-\Omega(n)},
$$
we would obtain a reduction from finding a dual codeword of weight $u$ to decoding at distance $t$. For our reduction to be meaningful we need to find the smallest $u$ as possible. It is roughly given by the first root of $K_{t}$ as we prove in what follows.
Let us start by some basic facts on roots of Krawtchouk polynomials.
\begin{enumerate}
\item For $1 \leq t \leq n$, the polynomial $K_{t}$ has $t$ distinct real roots on $(0,n)$.
\item Let $x_{1}$ be the first root of $K_{t}$. If $1 \leq t \leq n(q-1)/q$, we have \cite[Equation $(128)$]{L95}:
\begin{equation}\label{x_1}
\frac{x_{1}}{n} = \tau^{\perp} + o(1).
\end{equation}
\item The distance between any two consecutive roots of $K_{t}$ is an $o(n)$
\end{enumerate}
\begin{restatable}{proposition}{propoKraw}\label{propo:Kraw}
Between any two consecutive roots of $K_{t}$, where $1 \leq t \leq \frac{n}{q}$, there exists $u$ such that:
\begin{equation}
\frac{1}{q^{n}} \; \binom{n}{u}(q-1)^{u}\; \frac{K_{t}(u)^{2}}{(q-1)^{t}\binom{n}{t}} = \Omega\left( \frac{1}{n^{5}} \right)
\end{equation}
\end{restatable}
The proof of this proposition is given in Appendix \ref{app:propo:Kraw}. It is a generalization for any $q$ of \cite[Corollary 26]{KS21} which corresponds to the case where $q=2$.
We are now ready to prove Theorem \ref{theo:Hamming}. We are going to apply Theorem \ref{th:main} with the error distribution $\pi$ that we just introduced. Our proof essentially consists in verifying each hypothesis of Theorem \ref{th:main}.
\begin{proof}[Proof of Theorem \ref{theo:Hamming}] First, $\widehat{\ket{\pi}}$ is clearly radial (its expression is given in Lemma \ref{lem:Krawtchouk}).
By Lemma \ref{lemma:GV}, $\frac{\langle \pi,\vec{1}\rangle^{2}}{q^{n-k}} = \frac{S_{t}}{q^{n-k}} = q^{-\Omega(n)}$ as $t \leq (1-\delta)d_{\textup{GV}}(n,k)$ by assumption of Theorem \ref{theo:Hamming}.
Now, by Proposition \ref{propo:Kraw}, there exists $u$ between the two first consecutive roots of $K_{t}$ such that $S_{u}|f^{\perp}(u)|^{2} = \frac{1}{q^{n}} \; \binom{n}{u}(q-1)^{u}\; \frac{K_{t}(u)^{2}}{(q-1)^{t}\binom{n}{t}} = \Omega\left(\frac{1}{n^{5}}\right)$.
We have $|u - x_{1}| = o(n)$ as the distance between any two consecutive roots of $K_{t}$ is an $o(n)$. Therefore, by Equation \eqref{x_1}: $\frac{u}{n} = \tau^{\perp} + o(1)$. Furthermore we deduce that $\log_{2} q^{k}/S_{u} = n\left( R - h_{q}(\tau^{\perp}) + o(1)\right)$.
Now it can be verified that $h_{q}(\tau^{\perp}) >R$ for $R\in(0,1)$ and therefore that $q^{k}/S_u$ is negligible which concludes that proof.
\end{proof}
\section{A first failed attempt}\label{app:failed_attempt}
To apply Theorem \ref{th:main}, a natural choice for performing the reduction of searching short codewords in $\ensuremath{\mathscr{C}}^\perp$ to decoding $t$ errors in a linear code $\ensuremath{\mathscr{C}} \subseteq \F_q^n$ would be to choose a quantum state
$\ket{\pi}$ that at the same time
\begin{itemize}
\item[(i)] concentrates most of its norm on vectors of weight $\approx t$,
\item[(ii)] is radially symmetric,
\item[(iii)] would ideally be a separable quantum state $\ket{\pi} = \ket{\psi}^{\otimes n}$ which would simplify the computation of the Fourier transform a great deal.
\end{itemize}
All these requirements would lead to define the quantum state $\ket{\pi}$ as the separable state
$$
\ket{\pi^\text{try}} \stackrel{\text{def}}{=} \left( \sqrt{1-\tau}\ket{0} + \sqrt{\tau/(q-1)}
\sumFq{\alpha} \ket{\alpha} \right)^{\otimes n}
$$
Since $\ket{\pi^\text{try}}$ has also the following form
$$\ket{\pi^\text{try}}=\sumFqn{e}
(1-\tau)^{\frac{n-|\ev|}{2}}(\tau/(q-1))^{\frac{|\ev|}{2}} \ket{\ev},
$$
measuring this state $\ket{\pi^\text{try}}$ really mimics the error we have in a $q$-ary symmetric channel of crossover $\tau$, i.e.
\begin{equation}\label{eq:mu}
\mu_\tau(\ev) \stackrel{\text{def}}{=} \mathbb{P}(\text{measure outputs $\ev$}) = (1-\tau)^{n-|\ev|} (\tau/(q-1))^{|\ev|}
\end{equation}
where $\ev$ is any vector in $\F_q^n$.
It is straightforward to compute the Fourier transform of this state to obtain
\begin{eqnarray}
\widehat{\ket{\pi^\text{try}}} &= & \left( \sqrt{1-\tau^\perp}\ket{0} + \sqrt{\tau^\perp/(q-1)}
\sumFq{\alpha} \ket{\alpha} \right)^{\otimes n}\nonumber
\end{eqnarray}
In other words, the Fourier transform maps the quantum state ``representing'' the $q$-ary symmetric channel of crossover probability $\tau$ to a quantum state ``representing'' the $q$-ary symmetric channel of crossover probability $\tau^\perp$. This suggests that the quantum reduction outlined earlier reduces searching for codewords of weight $\approx t^\perp \stackrel{\text{def}}{=} \tau^\perp n$ to decoding $t$ errors in $\ensuremath{\mathscr{C}}$. Unfortunately, the fundamental quantity appearing in Theorem \ref{th:main} which is $\frac{\langle \pi^\text{try},\vec{1} \rangle^{2}}{q^{n-k}}$ is not negligible at all. Indeed we observe that
\begin{eqnarray*}
\frac{\langle \pi^\text{try},\vec{1} \rangle^{2}}{q^{n-k}}& =& \frac{q^k}{q^n} \left |\sum_{\vec{y} \in \F_q^n} \chi_{\vec{y}}(\mathbf{0}) \pi^{\text{try}}_\vec{y} \right|^2 \\
&=& q^{k}|f^{\perp}(0)|^2 \\
&= & q^{k}(1-\tau^{\perp})^{n}
\end{eqnarray*}
It can be verified that there is no way to choose $\tau$ such that at the same time:\\
(i) $\tau n \leq d_{\textup{GV}}(n,k)$ (otherwise there is no hope to decode correctly most of the time)\\
(ii) $\tau^\perp \leq \omega_{\textup{easy}}$ (otherwise finding codewords in $\ensuremath{\mathscr{C}}^\perp$ of weight $\tau^\perp n$ is easy)\\
(iii) $q^{k}(1-\tau^{\perp})^{n}=o(1)$.
The reason of this behavior can be traced back to the fact that the quantity $\frac{\langle \pi^\text{try},\vec{1} \rangle^{2}}{q^{n-k}}=q^{k}(1-\tau^{\perp})^{n}$ is just too big. Notice that
$$\frac{\langle \pi^\text{try},\vec{1} \rangle^{2}}{q^{n}}=(1-H^2(\mu_\tau,U))^2$$
where $\mu_\tau$ is the probability distribution on $\F_q^n$ obtained from measuring $\ket{\pi}$ as defined in \eqref{eq:mu} (\textit{i.e.} corresponding to the error distribution of a $q$-ary symmetric channel of crossover probability $\tau$), $U$ is the uniform distribution on $\F_q^n$ and $H(\pv,\vec{q})$ stands for the Hellinger distance between two discrete probabilities $p$ and $q$ defined over a same probability space:
$$
H(p,q) \stackrel{\text{def}}{=} \sqrt{1-\sum_i \sqrt{p_i,q_i}}.
$$
In other words the distribution of $\mu_\tau$ is too much spread out and we need a distribution which is much more concentrated around the weight $t$ for which we assume to have a decoder for $\ensuremath{\mathscr{C}}$.
\section{Proof of Proposition \ref{propo:Kraw}}\label{app:propo:Kraw}
Our aim in this section is to prove the following proposition
\propoKraw*
This proposition relies on \cite[Proposition $25$]{KS21} which is a general result about orthogonal polynomial for a positive discrete $\mu$ over $\{0,\dots,n\}$. Let $(P_{0},\dots,P_{n})$ be the family of orthogonal polynomials with respect to the inner product $\langle f,g \rangle \stackrel{\text{def}}{=} \sum_{i=0}^{n} f(i)g(i)\mu(i)$.
\begin{proposition}[{\cite[Proposition $25$]{KS21}}]\label{propo:KS} Let $s > 0$. Let the roots of $P_{s}$ be $y_1 < \dots < y_{s}$. Assume that $y_1 \geq 1$ and that $y_{s} \leq n-1$, and that the distance between any two consecutive roots is at least $2$. Assume also that the ratios $\frac{\mu(j)}{\mu(j+1)}$ and their inverses are uniformly bounded by some $K > 0$.
Then,
for any $1 \leq k \leq s-1$ the $\ell_{2}$ norm of $P_{s}$\footnote{The $\ell_{2}$ norm of $P_{s}$ is defined as $|P_{s}|_{2} \stackrel{\text{def}}{=} \sqrt{\sum_{i=0}^{n}P_{s}(i)^{2}\mu(i)}$.} is attained between $y_{k}$ and $y_{k+1}$, up to a factor of at most $O(\sqrt{K}n^{2}$).
\end{proposition}
The point is that Krawtchouk polynomials $K_{t}$'s are orthogonal polynomials (over $\{0,\dots,n\}$) for the measure $\mu(j) = \frac{(q-1)^{j}\binom{n}{j}}{q^{n}}$. Their $\ell_{2}$-norm is given by:
\begin{equation}
\sqrt{\sum_{i=0}^{n} K_{t}^{2}(i)\mu(i)} = \sqrt{(q-1)^{t}\binom{n}{t}}.
\end{equation}
Furthermore their smallest ({\em resp.} largest) root is $\geq 1$ ({\em resp. $\leq n-1$}), at least for $n$ large enough (see \cite[Corollary 6.1]{L95}). Therefore to prove Proposition \ref{propo:Kraw} (by just applying Proposition \ref{propo:KS}) it just remains to prove that the roots of $K_{t}$ are at distance $\geq 2$.
\begin{lemma}\label{lemma:distRoot}
When $1 \leq t \leq \frac{n}{q}$, the roots of $K_{t}$ are at distance $\geq 2$.
\end{lemma}
This lemma will be a consequence of the following theorem.
\begin{theorem}[{\cite[Theorem 2]{K03}}] \label{theo:Krasikov}
Let $P(x)$ be a discrete orthogonal polynomial, corresponding to an orthogonality measure supported on a subset of integers. Suppose that $P$ satisfies
\begin{equation}
P(x+1) = b(x)P(x) - c(x)P(x-1)
\end{equation}
and has all its roots $x_i$ in the open interval $I$. Then for any $i$, $|x_i - x_{i-1}| \ge 1$ provided $c(x) > 0$ for $x \in I$. If in addition, $b(x) > 0$ on $I$, then for any $i$, $|x_i - x_{i-1}| \ge 2$.
\end{theorem}
\begin{proof}[Proof of Lemma \ref{lemma:distRoot}] We are going to apply Theorem \ref{theo:Krasikov}. First, Krawtchouk polynomials verify the following equation:
\begin{equation}\label{eq:diff_eq}
(q-1)(n-x)K_t(x+1) = ((q-1)(n-x)+x-qt)K_t(x) - xK_t(x-1)
\end{equation}
All roots of $K_{t}$ lie in the interval $(0,n)$. Let,
\begin{equation*}
\forall x \in (0,n), \quad
\begin{cases}
b(x) &\stackrel{\text{def}}{=} \frac{(q-1)(n-x)+x-qt}{(q-1)(n-x)} \\
c(x) &\stackrel{\text{def}}{=} \frac{x}{(q-1)(n-x)}
\end{cases}
\end{equation*}
Clearly $c(x) > 0$ and $b(x) > 0$ (as $q \geq 2$ and $t \leq n/q$) on $(0,n)$ which concludes the proof of this lemma by applying Theorem \ref{theo:Krasikov}.
\end{proof}
We are now ready to prove Proposition \ref{propo:Kraw}.
\begin{proof}[Proof of Proposition \ref{propo:Kraw}] The roots of $K_{t}$ are at distance $\geq 2$ and its smallest ({\em resp.} largest) root is $\geq 1$ ({\em resp.} $\leq n-1$). Furthermore, $\frac{\mu(j)}{\mu(j+1)} = \frac{1}{q-1}\frac{j+1}{n-j}$. Therefore, the ratios $\frac{\mu(j)}{\mu(j+1)}$ and their inverses are uniformly bounded by $(q-1)n$. By applying Proposition \ref{propo:KS}, between each consecutive roots of $K_{t}$, there exists $u$ such that
$$
K_{t}(u)^{2}\mu(u) = |K_{t}|_{2}^{2} \; \Omega\left(\frac{1}{(q-1)n\; n^{4}}\right) \; \quad \Longrightarrow \quad K_{t}(u)^{2} \frac{\binom{n}{u}(q-1)^{u}}{q^{n}} = \binom{n}{t}(q-1)^{t} \;\Omega\left(\frac{1}{n^{5}}\right)
$$
which concludes the proof.
\end{proof}
|
\section{Introduction}
Due to the need for social distancing led by the COVID-19 pandemic, people and organizations have had to rely on digital technologies to stay connected and work remotely~\cite{pandey2020impact}. This has resulted in a surge in the usage of online conferencing tools.
Most organizations rely on these tools to conduct day-to-day business while people rely on them to connect with their family members and friends in these challenging times.
As a result, it is becoming more important than ever to provide high-quality speech audio under various household noise conditions.
In recent years, deep neural networks have shown great potential for single-channel speech enhancement (SE) (or noise-suppression)~\cite{weninger2015speech,hu2020dccrn,yin2020phasen,kinoshita2020improving,choi2018phaseaware,tang2020joint,isik2020poconet}. Although these models substantially remove background noise, most of them degrade the performance of downstream tasks such as automatic speech recognition (ASR) performance significantly, as modern commercial multi-condition trained ASR systems can usually recognize original noisy speech~\cite{kinoshita2020improving} well and the SE models introduces unseen distortions that are particularly harmful to ASR. When both live captioning and high-quality audio are needed, a common solution is for the local client to send both an enhanced signal for communication and an unaltered signal for transcription. Clearly, there is a benefit of creating a speech enhancement system that can improve speech quality without compromising the ASR accuracy.
In this paper, we propose a framework for optimizing speech enhancement models for both communication and transcription quality by leveraging pre-trained ASR models. The framework aims to build an SE model that achieves superior ASR performance while retaining the same speech quality as an SE model trained solely for SE objectives.
Our training framework alternately performs two steps: SE-step and ASR-step. In the SE-step, the model is trained with parallel data created by artificially mixing clean speech with noise files as with conventional supervised SE training. For the ASR-step, we use a realistic ASR model trained on a large amount of data.
The training audio is passed through the SE network and fed to the ASR network to calculate an ASR-based loss.
In both steps, only the SE network is updated while the ASR model parameters remain unchanged. This training scheme allows us to leverage real noisy recordings which do not have the corresponding clean speech signals to optimize the SE network. We evaluate our framework by using both real and synthetic data under various conditions with respect to signal-to-noise ratios (SNRs), types of noise, and recording scenarios. The experimental results show that our proposed framework improves the word error rate (WER) by 11.82\% over the state-of-the-art causal DCCRN model~\cite{hu2020dccrn} for
real recordings. Besides, we conduct ablation studies by changing the ASR model, the data for the ASR-step, and the schedule of the two training steps.
We use different ASR systems for the training and evaluation to show the generalization capability of the proposed framework.
\section{Related Work}
\textbf{Speech Enhancement:} There are regression-based and masking-based approaches for speech enhancement using neural network models~\cite{zhao2018convolutional,hu2020dccrn}. Regression-based approaches try to predict a clean speech signal or its time-frequency representation (TF) from the noisy speech input. Masking-based methods try to estimate a TF mask from the noisy input and apply the predicted mask to the same input to obtain the clean signal.
Various architectures, along with many objective functions, were proposed for solving the SE problem. We focus on the most recent and promising work.
Choi et al.~\cite{choi2018phaseaware} proposed a deep complex U-net (DCUNET) that uses real-valued convolution operations on the real and imaginary parts of the input speech STFT. They took advantages of the U-net structure and deep complex networks~\cite{deepcomplex}, which was shown to be useful for SE~\cite{pascual2017segan}. They employed a scale-invariant signal-to-noise ratio (SI-SNR) calculated in the time domain to optimize their model parameters. Hu et al.~\cite{hu2020dccrn} built upon the DCUNET and introduced a deep complex convolution recurrent network (DCCRN). They introduced complex long-short term memory (LSTM) layers in the bottleneck layer and complex batch normalization. The resulting model was significantly faster and had fewer parameters than DCUNET.
Kinoshita et al.~\cite{kinoshita2020improving} considered using a convolutional time-domain audio separation network (Conv-TasNet) for SE and proposed Denoising-TasNet, which directly processes a speech waveform with 1D convolutions. Their results show that the time-domain SE approach achieved a relative WER improvement of more than 30\%
for a robust ASR back-end. However, we could not observe WER improvements in our preliminary evaluation with time-domain approaches using a production-grade ASR back-end trained from both clean and noisy audio. Hence, we do not consider this approach in this work.
\textbf{ASR Multi-task Training: }
There are prior studies that employ joint or multi-task training for the front-end and back-end~\cite{seltzer2004likelihood,chen2015speech,giri2015improving,pironkov2016multi,xiao2016deep,meng2017deep,sainath2017multichannel,minhua2019frequency,subramanian2019speech,watanabe20202020,wang2020exploring}. However, they concerned only about the ASR accuracy and did not pay close attention to the SE quality nor analyzed the trade-off between the two tasks.
In contrast, we focus on the front-end and optimize it for both tasks.
\section{Method}
This section elaborates on our proposed framework.
We first present the proposed multi-task training framework using SE and ASR.
In the subsections that follow, we describe the pieces that constitute the proposed framework, i.e.,
our SE model as well as the SE and ASR loss functions to be used.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figures/sysoverview.pdf}
\vspace{-0.7cm}
\caption{Proposed multi-task training framework. $\hat{S}$ and $\hat{Y}$ denote enhanced signal and predicted transcription, respectively.}
\vspace{-0.5cm}
\label{fig:sysoverview}
\end{figure}
\subsection{Multi-Task Training} \label{sec:multi-task}
\vspace{-0.3em}
The idea behind the proposed method is to optimize an SE model for both noise suppression and ASR.
Multi-task training is usually performed by using training samples that have supervision signals (i.e., reference signals and word labels) for all the tasks to be considered.
We consider a scenario where each training sample has a supervision signal for only either SE or ASR.
This is because, in practice, most of the ASR datasets do not contain training samples with reference clean speech signals that could be used for supervised SE training. Also, some SE datasets do not come with human transcriptions.
To cope with this scenario, we update the SE model parameters by alternately performing the following two steps:
\begin{description}[style=unboxed,leftmargin=0cm]
\item[SE-step:] We mix clean speech and noise samples on the fly and use the noisy and clean speech pairs to evaluate a signal difference-based loss function. The SE model parameter gradients with respect to this loss function are computed to update the model parameters.
\item[ASR-step:] Noisy training samples in a mini-batch are fed to the SE network. The generated enhanced signals are input to the ASR network. We evaluate a loss function by comparing the ASR model output and the reference transcriptions. The loss is back-propagated all the way down to the SE network, and only the SE model parameters are updated.
This is because our objective is to find SE model parameter values that would work for existing well-trained ASR systems, and therefore we do not want the ASR network to adapt to the characteristics of the SE model.
\end{description}
Figure~\ref{fig:sysoverview} shows the diagram of the proposed framework.
The two-step approach allows us to take advantage of the real noisy speech samples that only have reference transcriptions for the SE model training.
At each training iteration, the update step to be used is chosen randomly from a Bernoulli distribution.
We refer to the probability of choosing the SE-step as the ``SE-step probability''.
Before performing the multi-task training, the SE model parameters are pre-trained on the SE training dataset.
\subsection{Speech Enhancement Model} \label{sec:se_model}
\vspace{-0.3em}
We employ DCCRN~\cite{hu2020dccrn} as our front-end model because it achieved the best SE performance in our preliminary test, while we also provide results using DCUNET to examine the generalization capability of the proposed framework.
DCCRN applies an encoder-decoder architecture with two LSTM layers in between. Instead of using conventional 2D convolutional/deconvolutional layers (conv2D/deconv2D), DCCRN builds on the DCUNET~\cite{choi2018phaseaware} that employs complex conv2D/deconv2D followed by complex batch normalization in the encoder and decoder blocks. Besides, DCCRN employs complex LSTM layers instead of conventional LSTM layers. Furthermore, it contains U-Net style skip connections (concatenation) from the encoder to the decoder as in DCUNET.
The DCCRN model takes the real and imaginary parts of a noisy spectrogram as input. It estimates a complex ratio mask (CRM) and applies it to the noisy speech. The masked signal is converted back to the time domain with ISTFT. The CRM can be defined as follows:
\begin{equation}\label{stargan}
\begin{split}
CRM = \frac{Y_{r}S_{r}+Y_{i}S_{i}}{Y^{2}_r+Y^{2}_i} + j\frac{Y_{r}S_{i}-Y_{i}S_{r}}{Y^{2}_r+Y^{2}_i}
\end{split},
\end{equation}
where $Y_r$ and $Y_i$ denote the real and imaginary parts of the noisy spectrogram, respectively while $S_r$ and $S_i$ denote the real and imaginary parts of the clean spectrogram, respectively.
We consider both non-causal and causal DCCRN configurations. The non-causal model can look ahead and employs bidirectional LSTM layers. The causal model can only process the current and previous frames and employs unidirectional LSTM layers and causal padding for conv2D/deconv2D layers. The latter is suitable for most applications.
For further details of DCCRN, we refer the reader to \cite{hu2020dccrn}.
\begin{table*}[ht!]
\centering
\caption{Evaluation results for seed SE models and their multi-task (MT) trained versions. SE-step probability for MT training was set at $0.5$ except for the Best WER MT case, where it was set at $0.0$.}
\vspace{-0.3cm}
\label{tab:main_results}
\begin{tabular}{l|ccccc|cc}
\hline \hline
\multicolumn{1}{c|}{\multirow{2}{*}{\textbf{System}}} & \multicolumn{5}{c|}{\textbf{Simulation Data}} & \multicolumn{2}{c}{\textbf{Real Data}} \\
\multicolumn{1}{c|}{} & \textbf{WER} & \textbf{pMOS} & \textbf{PESQ} & \textbf{SDR} & \textbf{STOI} & \textbf{WER} & \textbf{pMOS} \\ \hline
Noisy Speech (No enhancement) & 24.02 & 2.81 & 1.42 & 5.02 & 0.82 & 19.47 & 2.91 \\ \hline
DCUNet-causal-seed & 26.36 & 3.54 & 2.23 & 11.51 & 0.88 & 23.28 & 3.18 \\
DCUNet-causal-MT & 25.41 & 3.51 & 2.23 & 11.30 & 0.88 & 21.24 & 3.16 \\ \hline
DCCRN-non-causal-seed & 22.33 & 3.74 & 2.57 & 13.02 & 0.91 & 20.63 & 3.33 \\
DCCRN-non-causal-MT & 21.92 & 3.69 & 2.59 & 12.90 & 0.91 & 19.33 & 3.30 \\ \hline
DCCRN-causal-seed & 25.20 & 3.58 & 2.36 & 12.07 & 0.89 & 22.84 & 3.20 \\
DCCRN-causal-MT (Best WER Model) & 22.31 & 3.19 & 1.90 & 7.82 & 0.86 & 18.98 & 3.09 \\
DCCRN-causal-MT & 23.34 & 3.48 & 2.31 & 11.70 & 0.89 & 20.14 & 3.17 \\ \hline
\end{tabular}
\vspace{-0.4cm}
\end{table*}
\subsection{Loss Function for SE-Step} \label{sec:se_loss}
\vspace{-0.3em}
For the loss function of the SE-step and the SE model pre-training, we use the PHASEN loss function~\cite{yin2020phasen}.
It outperformed alternative loss functions, such as SI-SNR and power-spectrogram mean squared errors, in our preliminary tests.
The PHASEN loss comprises two parts: amplitude $L_a$ and phase-aware $L_p$ losses. The definition is as follows:
\begin{equation}\label{phasenloss}
\resizebox{.9\hsize}{!}{$\mathcal{L} =
\left||S|^p-|\hat{S}|^p\right|^2 + \left||S|^pe^{j\varphi (S)}-|\hat{S}|^pe^{j\varphi (\hat{S})}\right|^2$},
\end{equation}
where $S$ and $\hat{S}$ are the estimated and reference (i.e., clean) spectrograms, respectively. Hyper-parameter $p$ is a spectral compression factor and is set to 0.3. Operator $\varphi$ calculates the argument of a complex number.
\subsection{Loss Function for ASR-Step} \label{sec:asr_model}
\vspace{-0.3em}
Our back-end model used for the ASR-step training
is based on a sequence-to-sequence (Seq2Seq) model using an attention-based encoder-decoder structure \cite{wang2020exploring,chan2016listen}. We use the Seq2Seq model because it is simpler than a hybrid ASR system and thus facilitates multi-task training.
The model estimates text sequence $C = (c_1, c_2, \cdots)$ from STFT $X$ that is generated by the SE model.
We integrate STFT, log-mel filterbank energy extraction, and global mean-variance normalization into the ASR network to allow the gradients to pass through to the SE model.
In the ASR-step, the SE model parameters are updated to minimize the cross entropy loss between $C$ and reference label $R=\{r_1,...,r_N, r_{N+1}=\langle eos\rangle\}$, where $N$ is the length of the reference sequence $R$. Special symbol $\langle eos\rangle$ indicates the sequence end.
See Sections 2.3 and 4.4 of \cite{wang2020exploring} for our Seq2Seq ASR model.
\section{Experiments}
We conducted ASR and SE experiments to evaluate the proposed multi-task training framework under realistic settings.
\subsection{Datasets}
\vspace{-0.3em}
\textbf{Training Data for SE:}
\label{sec:train_data_se}
We utilized a large-scale and high-quality simulated dataset described in~\cite{braun2021efficient}, which includes around 1,000 hours of paired speech samples\footnote{We thank Sebastian Braun, Hannes Gamper, Chandan K.A. Reddy, and Ivan Tashev from Microsoft Research for providing us with the pre-mixed speech enhancement dataset and the pMOS tool.}. As a clean speech corpus, the dataset collects 544 hours of speech recordings with high mean opinion score (MOS) values from the LibriVox corpus~\cite{kearns2014librivox}. The mixtures are created using 247 hours of non-stationary noise recordings from the Audioset+Freesound~\cite{gemmeke2017audio,fonseca2017freesound} (187 hours), internal noise recordings (65 hours), and colored stationary noise (1 hour) as noise sources. In addition, the clean speech in each mixture is convolved with an acoustic room impulse response (RIR) sampled from 7,000 measured and simulated responses. See \cite{braun2021efficient} for details of this dataset. The data are available publicly, except for the 65 hours of the internal noise recordings\footnote{\url{https://github.com/microsoft/DNS-Challenge}}.
\noindent\textbf{Training Data for ASR: }
\label{sec:train_data_asr}
We trained our Seq2Seq ASR model based on
64 million anonymized and transcribed English utterances, totaling 75K hours.
\noindent\textbf{Multi-task training Data: }
We used different training data for the SE and ASR-steps as described earlier. For the ASR-step, we used a subset of 75K-hour transcribed data. Section \ref{sec:in_out_domain_data} explores considerations for selecting the subset, such as
in-domain vs. out-of-domain, and including vs. excluding simulated data.
Note that the 75K-hour data for the ASR-step also contained augmented/simulated data and that these simulated data were different from the SE training data.
\noindent\textbf{Evaluation Data: }
We used both simulated and real test data.
The simulated test set comprised 60 hours of simulated audio with SNRs ranging from -10 dB to 30 dB.
The simulation was performed as with
the training data while clean speech signals were taken from LibriSpeech train-clean-100~\cite{LibriSpeech} and convolved with RIRs generated by using different configurations. We added both Gaussian and non-stationary noise.
The latter was generated by convolving noise recordings from SoundBible+Freesound (10 hours in total) with simulated RIRs.
For the real test data, we employed two sets of noisy audio. The first set consisted of 18 hours of data recorded in an acoustically configurable audio lab where high-fidelity spatial noise sounds were played back from eight loudspeakers. The second set comprised 18 hours of meeting recordings. This test set contained various types of natural noise sounds. Furthermore, we included a clean speech test set consisting of 7803 words to measure the distortion introduced by the SE model.
\begin{figure*}[t!]
\centering
\centerline{\includegraphics[width=\textwidth]{figures/ablation.pdf}}
\vspace{-0.4cm}
\caption{(a) Performance comparison by using different ASR back-end models. ``Strong,'' ``Medium,'' and ``Weak'' in different colors represent three Seq2Seq models that have different ASR accuracies for the validation set. Each colored dot represents a checkpoint obtained every 5K iterations from multi-task training. (b) Impact of different ASR-step training datasets. Note that they are subsets of the ASR back-end model training data. ``In-domain'' means the training data is within a similar application scenario with the evaluation set. (c) Impact of SE-step probability on pMOS and WER.}
\vspace{-0.4cm}
\label{fig:ablation}
\end{figure*}
\subsection{Implementation Details}
\vspace{-0.3em}
Our framework was implemented in PyTorch~\cite{paszke2019pytorch}.
The seed (i.e., pre-trained) SE model was trained for 50 epochs with a batch size of 96 using 4 NVIDIA V100 GPUs.
In practice, we cannot fine-tune the SE model for every single back-end, which varies from time to time. Therefore, we used different back-end ASR models for training and evaluation, respectively. For ASR evaluation, a high-performance online hybrid ASR model was employed \cite{li2020high,li2019improving}. For multi-task training, the input feature for the offline Seq2Seq ASR model was 240-dimension log mel filterbanks, stacked by 3 frames with each frame having 10 msec. Global mean and variance normalization was applied before feeding the features to the ASR model encoder. We used 32K mixed-unit with $\langle space\rangle$ symbol between words as recognition units \cite{li2018advancing}. Label-smoothed cross-entropy loss \cite{chorowski2016towards} was applied for training.
\subsection{Discussion}
\vspace{-0.3em}
We evaluated our models using PESQ~\cite{rix2001perceptual}, STOI~\cite{taal2011algorithm}, SDR~\cite{raffel2014mir_eval}, and pMOS~\cite{gamper2019intrusive} metrics, where pMOS is a neural-network-based non-intrusive MOS estimator that shows high correlations with the human MOS ratings without requiring reference signals. Table~\ref{tab:main_results} shows the evaluation results of the seed models and their multi-task trained versions for both the simulation and real recordings. Except for the non-causal DCCRN model for the simulated test set, the seed SE models degraded the ASR performance compared with the original noisy signals. Applying the multi-task training to the seed models consistently improved the WER to a varying but significant degree. The real recording results show that the multi-task training improved the WER by 8.76\% (23.28 to 21.24), 6.30\% (20.63 to 19.33), and 11.82 \% (22.84 to 20.14) for DCUNet, DCCRN-non-causal, and DCCRN-causal, respectively.
These WER improvements were obtained with little SE quality degradation. For the real recordings, the pMOS values decreased only very marginally from 3.18 to 3.16, from 3.33 to 3.30, and from 3.20 to 3.17 for DCUNet, DCCRN-non-causal, and DCCRN-causal, respectively.
If we did not interleave the SE-steps between the ASR-steps by setting the SE-step probability at 0.0, the WER for the DCCRN-causal was further improved to outperform the WER for the noisy signals. However, this also sacrificed the SE quality to some extent, reducing pMOS from 3.20 to 3.09.
A similar trend was also observed for the
simulation set. There seems to be a trade-off between the ASR and SE quality.
An optimal operation point can be chosen by adjusting the SE-step probability based on the application needs. We further investigate this in Section 4.3.3.
\subsubsection{Impact of back-end ASR models}
\vspace{-0.3em}
Figure \ref{fig:ablation} (a) shows how the SE and ASR performance changed depending on the ASR back-end model used for the multi-task training.
We used three Seq2Seq models with different model structures, which are denoted as ``strong'' (green), ``medium'' (yellow), and ``weak'' (blue) based on a WER ranking calculated for an ASR validation set.
The result shows that
stronger ASR back-end models were more effective in closing the WER gap from the ``No Enhancement'' setting while preserving the SE improvement.
Note that the ``strong'' ASR model was trained on the 75K-hour data containing various noisy signals. This suggests that it is important to use a powerful back-end model instead of a noise-sensitive model trained only on clean signals.
\subsubsection{Impact of training data for ASR-step}\label{sec:in_out_domain_data}
\vspace{-0.3em}
Figure \ref{fig:ablation} (b) shows the impact that different ASR-step training sets had on the SE model's performance. The ``strong'' back-end model was used.
The result shows that the models trained on 1302-hour in-domain data (yellow) and those trained on its 224-hour random subset (red) performed equally well.
This indicates that it is sufficient to use a relatively small amount of in-domain data.
Meanwhile, combining simulated data (green) benefited the ASR performance compared with using only real data (red), which suggests the importance of acoustic diversity in terms of noise and reverberation conditions.
It was also observed that
using out-of-domain data (blue) degraded the pMOS score compared with using a similar amount of in-domain data (green) while they led to similar WERs.
\subsubsection{Impact of SE-step probability}
\vspace{-0.3em}
We also investigated how the SE model's performance was impacted by the SE-step probability value, which controls how often the SE-step is performed in the multi-task training,
We used the ``strong'' Seq2Seq back-end model and ``224-hour in-domain with simulation'' training data for ASR-step.
Figure \ref{fig:ablation} (c) reveals the trade-off between pMOS and WER with the variation of ``SE-step probability.'' Performing the ASR-step more frequently resulted in ASR performance improvement at the expense of pMOS compared with the seed DCCRN model.
When only the ASR-step was performed by setting the SE-step probability at 0, the WER surpassed that of ``No Enhancement'' condition, which substantially compromised the pMOS score.
For a general use case, we would prefer a moderate SE-step probability such as $0.5$ for serving human listening and live captioning, which is shown in the last row (DCCRN-causal-MT) of Table.\ref{tab:main_results}.
\section{Conclusions}
In this paper, we proposed a multi-task training framework for deep learning-based SE models to make them more friendly to ASR systems. In the proposed approach, we first pre-trained the SE and ASR models separately. Then, we froze the ASR model parameters and started the multi-task training, where we interleaved two update steps: the SE-step and the ASR-step. The SE-step is the same as the conventional supervised SE training using noisy and clean speech pairs. For the ASR-step, we first enhanced the signals with the SE model, then fed the enhanced signals to the ASR model and backpropagated the cross-entropy loss. We controlled the frequency of each step with a probability parameter. The experimental results showed that our framework improved the ASR performance with small or little SE quality degradation. The ASR model used for the multi-task training was different from that used for evaluation, indicating that the resultant SE model can generalize to different back-end conditions.
We also presented ablation study results that would guide the development of the proposed approach.
In this paper,
we opted to use a mix of public and private data to reflect the scale and complexity of real usage scenarios. We hope our work can encourage the research community to establish an open experimental platform for large-scale SE and ASR investigation.
\newpage
\bibliographystyle{IEEEtran}
\section{Introduction}
This template can be found on the conference website. Templates are provided for Microsoft Word\textregistered, and \LaTeX. However, we highly recommend using \LaTeX when preparing your submission. Information for full paper submission is available on the conference website.
\section{Page layout and style}
Authors should observe the following rules for page layout. A highly recommended way to meet these requirements is to use a given template (Microsoft Word\textregistered\ or \LaTeX) and check details against the corresponding example PDF file. Given templates, Microsoft Word\textregistered\ or \LaTeX, can be adapted/imported easily in other software such as LibreOffice, Apple Pages, Lua\LaTeX, and Xe\LaTeX, but please be careful to match the layout of the provided PDF example.
\subsection{Basic layout features}
\begin{itemize}
\item Proceedings will be printed in DIN A4 format. Authors must submit their papers in DIN A4 format.
\item Two columns are used except for the title section and for large figures that may need a full page width.
\item Left and right margin are 20 mm each.
\item Column width is 80 mm.
\item Spacing between columns is 10 mm.
\item Top margin is 25 mm (except for the first page which is 30 mm to the title top).
\item Bottom margin is 35 mm.
\item Text height (without headers and footers) is maximum 235 mm.
\item Headers and footers must be left empty.
\item Check indentations and spacings by comparing to this example file (in PDF).
\end{itemize}
\subsubsection{Headings}
Section headings are centered in boldface with the first word capitalized and the rest of the heading in lower case. Sub- headings appear like major headings, except they start at the left margin in the column. Sub-sub-headings appear like sub-headings, except they are in italics and not boldface. See the examples in this file. No more than 3 levels of headings should be used.
\subsection{Text font}
Times or Times Roman font is used for the main text. Font size in the main text must be 9 points, and in the References section 8 points. Other font types may be used if needed for special purposes. It is VERY IMPORTANT that while making the final PDF file, you embed all used fonts! To embed the fonts, you may use the following instructions:
\begin{enumerate}
\item For Windows users, the bullzip printer can convert any PDF to have embedded and subsetted fonts.
\item For Linux/Mac users, you may use \\
pdftops file.pdf\\
pstopdf -dPDFSETTINGS=/prepress file.pdf
\end{enumerate}
\LaTeX users: users should use Adobe Type 1 fonts such as Times or Times Roman. These are used automatically by the INTERSPEECH2021.sty style file. Authors must not use Type 3 (bitmap) fonts.
\subsection{Figures}
All figures must be centered on the column (or page, if the figure spans both columns). Figure captions should follow each figure and have the format given in Figure~\ref{fig:speech_production}.
Figures should be preferably line drawings. If they contain gray levels or colors, they should be checked to print well on a high-quality non-color laser printer.
Graphics (i.\,e., illustrations, figures) must not use stipple fill patterns because they will not reproduce properly in Adobe PDF. Please use only SOLID FILL COLORS.
Figures which span 2 columns (i.\,e., occupy full page width) must be placed at the top or bottom of the page.
\subsection{Tables}
An example of a table is shown in Table~\ref{tab:example}. The caption text must be above the table.
\begin{table}[th]
\caption{This is an example of a table}
\label{tab:example}
\centering
\begin{tabular}{ r@{}l r }
\toprule
\multicolumn{2}{c}{\textbf{Ratio}} &
\multicolumn{1}{c}{\textbf{Decibels}} \\
\midrule
$1$ & $/10$ & $-20$~~~ \\
$1$ & $/1$ & $0$~~~ \\
$2$ & $/1$ & $\approx 6$~~~ \\
$3.16$ & $/1$ & $10$~~~ \\
$10$ & $/1$ & $20$~~~ \\
$100$ & $/1$ & $40$~~~ \\
$1000$ & $/1$ & $60$~~~ \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Equations}
Equations should be placed on separate lines and numbered. Examples of equations are given below. Particularly,
\begin{equation}
x(t) = s(f_\omega(t))
\label{eq1}
\end{equation}
where \(f_\omega(t)\) is a special warping function
\begin{equation}
f_\omega(t) = \frac{1}{2 \pi j} \oint_C
\frac{\nu^{-1k} \mathrm{d} \nu}
{(1-\beta\nu^{-1})(\nu^{-1}-\beta)}
\label{eq2}
\end{equation}
A residue theorem states that
\begin{equation}
\oint_C F(z)\,\mathrm{d}z = 2 \pi j \sum_k \mathrm{Res}[F(z),p_k]
\label{eq3}
\end{equation}
Applying (\ref{eq3}) to (\ref{eq1}), it is straightforward to see that
\begin{equation}
1 + 1 = \pi
\label{eq4}
\end{equation}
Finally we have proven the secret theorem of all speech sciences. No more math is needed to show how useful the result is!
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figure.pdf}
\caption{Schematic diagram of speech production.}
\label{fig:speech_production}
\end{figure}
\subsection{Information for Word users only}
For ease of formatting, please use the styles listed in Table 2. The styles are defined in this template file and are shown in the order in which they would be used when writing a paper. When the heading styles in Table 2 are used, section numbers are no longer required to be typed in because they will be automatically numbered by Word. Similarly, reference items will be automatically numbered by Word when the ``Reference'' style is used.
\begin{table}[t]
\caption{Main predefined styles in Word}
\label{tab:word_styles}
\centering
\begin{tabular}{ll}
\toprule
\textbf{Style Name} & \textbf{Entities in a Paper} \\
\midrule
Title & Title \\
Author & Author name \\
Affiliation & Author affiliation \\
Email & Email address \\
AbstractHeading & Abstract section heading \\
Body Text & First paragraph in abstract \\
Body Text Next & Following paragraphs in abstract \\
Index & Index terms \\
1. Heading 1 & 1\textsuperscript{st} level section heading \\
1.1 Heading 2 & 2\textsuperscript{nd} level section heading \\
1.1.1 Heading 3 & 3\textsuperscript{rd} level section heading \\
Body Text & First paragraph in section \\
Body Text Next & Following paragraphs in section \\
Figure Caption & Figure caption \\
Table Caption & Table caption \\
Equation & Equations \\
\textbullet\ List Bullet & Bulleted lists \\\relax
[1] Reference & References \\
\bottomrule
\end{tabular}
\end{table}
If your Word document contains equations, you must not save your Word document from ``.docx'' to ``.doc'' because when doing so, Word will convert all equations to images of unacceptably low resolution.
\subsection{Hyperlinks}
For technical reasons, the proceedings editor will strip all active links from the papers during processing. Hyperlinks can be included in your paper, if written in full, e.\,g.\ ``http://www.foo.com/index.html''. The link text must be all black.
Please make sure that they present no problems in printing to paper.
\subsection{Multimedia files}
The INTERSPEECH organizing committee offers the possibility to submit multimedia files. These files are meant for audio-visual illustrations that cannot be conveyed in text, tables and graphs. Just like you would when including graphics, make sure that you have sufficient author rights to the multimedia materials that you submit for publication. The proceedings media will NOT contain readers or players, so be sure to use widely accepted file formats, such as MPEG, Windows WAVE PCM (.wav) or Windows Media Video (.wmv) using standard codecs.
Your multimedia files must be submitted in a single ZIP file for each separate paper. Within the ZIP file you can use folders and filenames to help organize the multimedia files. In the ZIP file you should include a TEXT or HTML index file which describes the purpose and significance of each multimedia file. From within the manuscript, refer to a multimedia illustration by its filename. Use short file names without blanks for clarity.
The ZIP file you submit will be included as-is in the proceedings media and will be linked to your paper in the navigation interface of the proceedings. Causal Productions (the publisher) and the conference committee will not check or change the contents of your ZIP file.
Users of the proceedings who wish to access your multimedia files will click the link to the ZIP file which will then be opened by the operating system of their computer. Access to the contents of the ZIP file will be governed entirely by the operating system of the user's computer.
\subsection{Page numbering}
Final page numbers will be added later to the document electronically. \emph{Do not make any footers or headers!}
\subsection{References}
The reference format is the standard IEEE one. References should be numbered in order of appearance, for example \cite{Davis80-COP}, \cite{Rabiner89-ATO}, \cite[pp.\ 417--422]{Hastie09-TEO}, and \cite{YourName21-XXX}.
\subsection{Abstract}
The total length of the abstract is limited to 200 words. The abstract included in your paper and the one you enter during web-based submission must be identical. Avoid non-ASCII characters or symbols as they may not display correctly in the abstract book.
\subsection{Author affiliation}
Please list country names as part of the affiliation for each country.
\subsection{Number of authors in the author list}
The maximum number of authors in the author list is twenty. If the number of contributing authors is more than twenty, they should be listed in a footnote or in acknowledgement section, as appropriate.
\subsection{Submitted files}
Authors are requested to submit PDF files of their manuscripts. You can use commercially available tools or for instance http://www.pdfforge.org/products/pdfcreator. The PDF file should comply with the following requirements: (a) there must be no PASSWORD protection on the PDF file at all; (b) all fonts must be embedded; and (c) the file must be text searchable (do CTRL-F and try to find a common word such as ``the''). The proceedings editors (Causal Productions) will contact authors of non-complying files to obtain a replacement. In order not to endanger the preparation of the proceedings, papers for which a replacement is not provided in a timely manner will be withdrawn.
\section{Discussion}
This is the discussion. This is the discussion. This is the discussion. Is there any discussion?
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Cras consequat mollis odio, nec venenatis enim auctor sed. Integer tincidunt fringilla lectus eget condimentum. In eget sapien id eros dapibus interdum vel ac quam. Aenean vitae rutrum erat. Aenean et risus pharetra, lacinia augue ut, fermentum ante. Integer dui arcu, interdum at ornare a, faucibus quis est. Mauris quis quam felis. Etiam pulvinar massa et turpis lacinia, eu posuere mi iaculis. Fusce at velit quis leo dignissim porttitor.
Fusce ut nunc eu sapien venenatis finibus a vel ligula. Pellentesque habitant morbi tristique senectus et netus et malesuada fames ac turpis egestas. Ut quam eros, volutpat at gravida consectetur, rutrum ut leo. Aenean cursus euismod feugiat. Cras hendrerit, ligula eu feugiat malesuada, neque turpis auctor lacus, sit amet accumsan neque orci a quam. Mauris suscipit ultrices mattis. Nulla at interdum metus, id pharetra diam. Curabitur at vestibulum sem, sed elementum massa. Donec iaculis et arcu ut rutrum. Fusce gravida, mauris porta volutpat eleifend, enim mauris eleifend orci, eu ultrices leo purus vitae metus. In pretium dolor ut magna dictum, at imperdiet lectus porta.
Quisque mollis lectus id risus pretium mattis. Morbi scelerisque posuere est, id efficitur urna luctus non. Praesent quam lacus, facilisis id ante eu, vehicula maximus ex. Nullam mollis in arcu vitae efficitur. Aliquam molestie eleifend ante, in pretium velit ultrices ac. Etiam laoreet nec sem non pulvinar. Integer ligula felis, interdum non lacus id, malesuada imperdiet turpis.
Aenean sit amet volutpat nisi. Aliquam eu erat quis tortor ultrices laoreet. Vivamus fermentum semper metus, non faucibus libero euismod vitae. Sed efficitur porta congue. Aenean in faucibus nisi. Donec suscipit augue vitae orci consequat, sit amet aliquet felis varius. Duis efficitur lacinia dolor sit amet lobortis. Curabitur erat sapien, molestie nec nisi eu, dignissim accumsan ipsum. Fusce id nibh nec risus dictum posuere in ac magna. Donec malesuada massa sed erat lacinia cursus. Suspendisse ornare augue nec volutpat consequat.
Vestibulum et vulputate nisi, a malesuada mi. Nam pellentesque arcu sapien, at placerat odio imperdiet ut. Curabitur nec venenatis tellus, vel aliquet nisi. Curabitur vel ligula sit amet metus auctor pretium. Nullam nulla mi, blandit a mattis id, vulputate sit amet enim. Proin mollis fringilla dictum. Proin lacinia orci purus.
Curabitur porttitor bibendum dolor, nec consectetur sapien pulvinar id. Donec eleifend, est vel dignissim pretium, tortor augue euismod nunc, id fermentum erat felis ac neque. Morbi id lectus ultricies, rutrum justo eu, sollicitudin risus. Suspendisse lobortis efficitur nisi sit amet pellentesque. Ut eget augue at mi aliquet mattis. Proin et feugiat erat, sit amet sodales eros. Integer sed elit quis est mattis ullamcorper. Pellentesque lectus nisi, vulputate a imperdiet tincidunt, auctor nec orci. Pellentesque sagittis nisl orci, vitae placerat massa lacinia nec. Sed egestas magna sed augue sollicitudin luctus. Praesent interdum bibendum tortor, eu porta purus. Aliquam convallis velit id mi fermentum, sed ornare eros cursus.
Quisque congue leo a fringilla pharetra. Phasellus sed tempor est, sed auctor purus. Morbi vel lacus ullamcorper, auctor mauris id, pulvinar lorem. Suspendisse potenti. Nam porta, purus non eleifend bibendum, erat metus pellentesque elit, non luctus nibh nunc ornare nisl. Sed rutrum lacinia nisi ac suscipit. Curabitur non blandit augue. Integer viverra, ipsum vel molestie euismod, sem quam tempus massa, eget efficitur ante turpis non metus. Quisque efficitur posuere velit in iaculis. Cras imperdiet varius urna vitae vestibulum. Donec accumsan eget nisi sed pellentesque. Vestibulum id quam ut urna volutpat ullamcorper gravida sit amet libero. Aliquam bibendum, ligula vitae porta malesuada, arcu diam congue erat, a pharetra diam sem vulputate tortor. Etiam luctus iaculis leo cursus tristique.
Mauris mattis sem dolor, sit amet ullamcorper arcu tincidunt ac. Vestibulum at blandit tortor. Quisque bibendum congue leo, vitae eleifend massa. Vestibulum vitae odio elit. Lorem ipsum dolor sit amet, consectetur adipiscing elit. Ut sagittis quam vel felis ornare, in gravida felis tempor. Donec molestie dui quis leo venenatis blandit. Nunc sit amet finibus metus. Cras ut lectus ex.
Suspendisse commodo libero vel leo tincidunt, a tempus mauris porta. Integer varius eros ac sapien lacinia vehicula. Donec porttitor, lacus faucibus rhoncus venenatis, neque quam imperdiet nunc, id consectetur metus purus quis sapien. Phasellus interdum nulla vel euismod posuere. Vestibulum finibus magna vel finibus mollis. Curabitur mollis turpis tortor, hendrerit vulputate justo egestas quis. Nam dignissim luctus leo non elementum. Phasellus a metus at leo malesuada bibendum. Mauris quis eleifend magna, nec vehicula ex. Donec venenatis urna fermentum commodo vehicula. Ut mattis scelerisque aliquam. Vivamus pulvinar erat metus, id tempus mi vulputate quis.
Fusce lobortis a urna eget blandit. Vivamus in eleifend neque, at sollicitudin lectus. Quisque faucibus egestas lorem, in commodo diam maximus eu. Morbi finibus ante ac felis porttitor euismod. Donec lobortis aliquam ipsum sit amet luctus. Cum sociis natoque penatibus et magnis dis parturient montes, nascetur ridiculus mus. Etiam rutrum neque sapien, eget luctus turpis iaculis pulvinar. Duis quis pulvinar nunc, nec bibendum ligula. Phasellus suscipit sagittis lacus molestie laoreet. Pellentesque lacus diam, tincidunt a aliquam vitae, aliquet non justo.
Etiam lectus lacus, commodo eget consectetur eget, auctor vitae leo. Praesent vitae erat in diam blandit semper vitae et eros. Maecenas auctor pharetra nibh eget egestas. Donec accumsan ut risus eget rhoncus. Nam placerat, erat sit amet gravida mollis, purus arcu accumsan diam, tempus pharetra risus mi ac sapien. Ut et tortor porta, pulvinar elit vitae, tempor mi. Nam interdum, nisl non pharetra molestie, turpis neque commodo ligula, sit amet pretium nisl nibh quis ante. Quisque et ex eget velit lobortis suscipit. Integer aliquam finibus molestie. Sed pellentesque neque eu turpis aliquet, mattis ornare enim finibus. In hac habitasse platea dictumst.
Integer congue quis justo a posuere. Quisque porta, ante et dignissim suscipit, arcu mauris ultrices libero, nec sollicitudin purus lacus a enim. Aliquam feugiat eget lacus molestie sodales. Duis blandit placerat nunc, et venenatis turpis dictum vel. Nulla facilisi. Nam ullamcorper, tellus eu posuere mattis, arcu lacus dictum nulla, vel mattis nisi sem posuere tellus. Etiam quis eros condimentum lectus lobortis eleifend. In ex lacus, sodales scelerisque egestas ac, aliquam nec purus. Nunc sit amet magna non libero ullamcorper dictum. Phasellus porta faucibus tempus. Praesent blandit tortor sed tellus ornare consectetur. Sed sed nisi id neque porta varius eu eu velit. Curabitur varius convallis justo id facilisis. Mauris auctor velit nec aliquam cursus.
Integer suscipit scelerisque leo sed faucibus. Ut commodo nulla luctus diam posuere egestas. Integer ut augue ac velit ullamcorper tempus. Pellentesque in mi rhoncus, sodales sem quis, commodo sem. Aenean dapibus euismod diam id rhoncus. Nullam vehicula placerat eros consectetur luctus. Aliquam auctor ipsum vitae egestas imperdiet. Ut nulla lacus, imperdiet quis urna vel, ornare imperdiet tortor. Mauris nec diam ac nunc laoreet volutpat at id turpis. Nulla eu neque a risus feugiat iaculis ac vel risus. Ut tempus elementum lorem eget porta. Nullam et ullamcorper urna. Cum sociis natoque penatibus et magnis dis parturient montes, nascetur ridiculus mus. Phasellus eget dui vitae nulla hendrerit ultrices quis rutrum leo.
Proin consectetur lacus sit amet eleifend varius. Etiam eu blandit risus. Curabitur pellentesque urna sed dolor congue mattis. Vestibulum ut velit posuere, feugiat leo a, dignissim massa. Proin eu nulla risus. Fusce luctus bibendum est, sit amet venenatis nisi finibus non. Donec ultricies ornare nunc at lobortis. Nam a auctor metus. Vestibulum pretium condimentum turpis ac mattis. Curabitur semper sagittis rhoncus. Duis molestie facilisis mattis. Sed pharetra lorem id tortor efficitur, sed maximus leo posuere. Quisque suscipit molestie convallis. Duis imperdiet placerat congue. Morbi placerat, velit ut tempor porta, ex nisi imperdiet purus, non feugiat ex velit nec nulla.
Phasellus mattis at erat eget lobortis. Vivamus sodales odio non erat luctus faucibus. Curabitur aliquam luctus nulla quis consectetur. Fusce vulputate finibus vulputate. Cras at condimentum massa. Duis vestibulum ipsum ac tortor lobortis fringilla. Cras non neque at nunc pellentesque mollis. Sed quis erat mauris. Ut dapibus sem lectus, quis imperdiet diam bibendum et. Maecenas quis venenatis ante. Nunc blandit a risus sed scelerisque. Praesent cursus est sit amet nisi tempus, quis placerat libero rutrum. Phasellus lacinia nisi quis consequat mollis. Phasellus sagittis aliquam lacus.
Sed dolor quam, posuere nec nunc eget, feugiat lobortis ligula. Fusce lacinia fermentum dolor, luctus dapibus ex venenatis feugiat. In hac habitasse platea dictumst. Nullam vitae ligula dignissim, interdum turpis quis, tincidunt metus. Nullam in nisl vitae mauris egestas porta. Nam fringilla aliquet sapien, non dapibus nunc sollicitudin id. Nunc hendrerit felis et vehicula consequat. Praesent varius libero id volutpat iaculis. Aliquam vel dui imperdiet, pharetra augue sed, iaculis nulla. Quisque mollis orci nec odio eleifend, eget laoreet nunc feugiat. Cras varius tortor a fringilla gravida.
Sed posuere erat eu dolor consequat euismod. Donec imperdiet, tellus nec convallis commodo, lorem sem lobortis purus, a lacinia massa ipsum vitae nisl. Vivamus auctor tellus in urna iaculis luctus. In dui nibh, posuere a erat a, lobortis finibus nulla. Sed vel suscipit nisi. Nunc eget nibh risus. Sed posuere tempus eleifend. Nullam ac lacinia ligula, ut blandit erat. In at est sed turpis consequat rutrum. Nunc eget lectus venenatis, convallis ligula non, mollis orci. Aliquam sit amet ligula turpis. Sed finibus laoreet elit nec molestie.
Mauris in nisi et neque euismod aliquam ut eget felis. Sed eget dictum tellus, finibus rutrum nibh. Fusce placerat augue a faucibus semper. Nam sed nisl ligula. Vivamus ante augue, faucibus at risus ut, hendrerit viverra risus. Etiam justo eros, dignissim a nunc sed, porta mollis erat. Fusce in sem accumsan, laoreet nibh porta, porttitor orci. Donec venenatis vehicula ante eget dictum.
Aliquam fermentum a metus pellentesque cursus. Vivamus eleifend ultricies tellus, vel scelerisque diam. Sed hendrerit est at elit suscipit placerat. Praesent nec pretium erat. Quisque blandit nunc in felis bibendum consequat.
\section{Conclusions}
Authors must proofread their PDF file prior to submission to ensure it is correct. Authors should not rely on proofreading the Word file. Please proofread the PDF file before it is submitted.
\section{Acknowledgements}
The ISCA Board would like to thank the organizing committees of the past INTERSPEECH conferences for their help and for kindly providing the template files. \\
Note to authors: Authors should not use logos in the acknowledgement section; rather authors should acknowledge corporations by naming them only.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{intro}
The increasing accuracy of \textit{first-principles} calculations of heat transport, along with the growing availability of computing resources, has enabled the systematic prediction of lattice thermal conductivity in semiconductors and two-dimensional (2D) materials~\cite{lindsay2019perspective,lindsay2018survey,mingo2014ab}. If key theoretical developments have solidified our understanding of phonon dynamics at the nanoscales, the fast progress of this field has been unarguably aided by the release of open-source packages. For example, \verb!ShengBTE!~\cite{li2014shengbte} implements the iterative solution of the Boltzmann transport equation (BTE)~\cite{omini1995iterative}, with force constants computed by \textit{first-principles}~\cite{broido2007intrinsic,esfarjani2011heat}. The tool can be primarily used for either bulk materials or nanowires~\cite{li2012thermal}. Later iterations include ~\verb!AlmaBTE!~\cite{Carrete2017AlmaBTEMaterials}, which, among several improvements, features a Monte Carlo solver for superlattices, and \verb!FourPhonons!~\cite{han2021fourphonon,feng2016quantum}, a tool that, as the name suggests, can handle four-phonon scattering. Another notable package is~\verb!Phono3Py!~\cite{phono3py}, where the BTE is solved both within the relaxation time approximation (RTA) and including the full scattering operator~\cite{Chaput2013DirectEquation}. Albeit not strictly required, these tools are mostly used in tandem with super-cell approaches. When a density functional perturbation theory (DFPT) is available for a given material, it is often convenient to compute directly the Fourier-transformed force constants, without resorting to supercells; a prominent package in this space is \verb!d3q!~\cite{paulatto2013anharmonic}, implemented on top of \verb!QUANTUM ESPRESSO!~\cite{giannozzi2009quantum}. This software also includes the possibility of simulating thin films~\cite{paulatto2020thermal} and coherence effects~\cite{simoncelli2019unified}. In a similar effort, the recently released~\verb!Kaldo!\cite{barbalinardo2020efficient} combines the Green-Kubo model and the BTE to take into account material disorder~\cite{isaeva2019modeling}.
While the dissemination of these software manifests the increasing excitement around phonon related applications, their focus is primarily on bulk materials or simple geometries. In fact, most phonon dynamics simulations in complex geometries are currently performed by MonteCarlo solvers~\cite{landon2014deviational,peraud2011efficient}, with a recent implementation provided by \verb!MCBTE!~\cite{pathak2021mcbte}. Thus, an open-source tool that solves space-dependent heat transport deterministically, i.e. in the same spirit as in bulk calculations, is still pending. We fill this gap by introducing \verb!OpenBTE!, a solver for the \textit{first-principles}, multidimensional phonon BTE. Key capabilities include the calculation of the temperature and heat flux maps, as well as the mode-resolve effective thermal conductivity. \textit{Ab-initio} harmonic and anharmonic properties are delegated to external software. While flexible and straightforward to expand, \verb!OpenBTE! currently implements the anisotropic-mean-free-path BTE (aMFP-BTE)~\cite{romano2021}, a method based on the interpolation of the phonon populations onto the vectorial MFP space. This method has shown a 50x speed up with respect to the mode-resolve BTE for thick Si membranes, while not compromising on the accuracy. Furthermore, it scales with constant time with respect to the number of phonon branches, therefore enabling fast simulations of size effects in complex-unit-cell materials, such as bismuth telluride.
The paper is organized as follows. First, we review the model and the assumptions therein, then we provide details on the implemented algorithm and the software's main modules. An example of a porous Si membrane will follow, and final remarks conclude the paper. We also provides technical implementation details in the appendices. Enabling fast simulations of heat transport in nanostructures, \verb!OpenBTE! sets out to accelerate the development of thermal-related applications, such as thermal management and thermoelectrics. \verb!OpenBTE! is released under the \verb!GPL-3.0! license and currently hosted on GitHub~\cite{openbte}.
\section{Model}
\label{model}
The steady-state, linearized phonon BTE in the temperature formulation reads~\cite{romano2021}
\begin{equation}\label{bte}
- C_\mu \mathbf{v}_\mu \cdot\nabla \Delta T_\mu(\mathbf{r}) = \sum_\nu W_{\mu\nu} \Delta T_\nu(\mathbf{r}),
\end{equation}
where $\mu,\nu$ collectively indicate phonon branch and wave vector, running up to $N_p$ and $N_q$, respectively. The unknowns $\Delta T_\mu (\mathbf{r})$ are the phonon pseudo-temperatures (referred to hereafter simply as temperatures), connected to the deviational non-equilibrium phonon population via $\Delta T_\mu(\mathbf{r}) = \Delta n_\mu(\mathbf{r}) \hbar \omega_\mu C_\mu^{-1}$. The mode-specific heat capacity is $C_\mu = k_B \left(\eta_\mu/ \sinh{\eta_\mu} \right)^2$, with $\eta_\mu=\hbar \omega_\mu/k_B/T/2$. The terms $\hbar \omega_\mu$ and $\mathbf{v}_\mu$ are the phonon frequencies and the group velocities, respectively. In this formulation, the thermal flux is given by $\mathbf{J}(\mathbf{r}) = \left(N_q \mathcal{V}\right)^{-1}\sum_\mu\Delta T_\mu (\mathbf{r}) \mathbf{S}_\mu$, where $\mathcal{V}$ is the volume of the unit cell, and $\mathbf{S}_\mu = C_\mu \mathbf{v}_\mu$. The term $W_{\mu\nu}$ is the energy form of the scattering operator~\cite{romano2020phonon}. While we plan to make available the implementation of Eq.~\ref{bte}, currently only the relaxation-time-approximation (RTA) is supported; within the RTA, the RHS of Eq.~\ref{bte} becomes $C_\mu/\tau_\mu\left[\Delta T_\mu(\mathbf{r}) - \Delta T(\mathbf{r})\right]$, leading to
\begin{equation}\label{rta}
-\mathbf{v}_\mu \cdot \nabla \Delta T_\mu(\mathbf{r}) = \frac{\Delta T_\mu(\mathbf{r}) - \Delta T(\mathbf{r}) }{\tau_\mu},
\end{equation}
where $\tau_\mu$ is the relaxation time, and $\Delta T(\mathbf{r})$ is the pseudo lattice temperature; this term is evaluated so that energy conservation is satisfied, i.e. $\nabla \cdot \mathbf{J}(\mathbf{r}) = 0$, yielding $\Delta T(\mathbf{r}) = \left[\sum_\nu C_\nu/\tau_\nu \right]^{-1} \sum_\mu C_\mu/\tau_\mu \Delta T_\mu(\mathbf{r})$~\cite{Hua2014TransportConduction}.
The simulation domain is a cuboid of size $L_x$, $L_y$ and $L_z$ with periodic boundary conditions applied throughout by default. A difference of temperature $\Delta T_{\mathrm{ext}}$ is also enforced along a chosen Cartesian axis. At the walls of the pores, energy conservation leads to $\sum_\nu \Delta T_\nu \mathbf{S}_\mu \cdot\mathbf{\hat{n}} = 0$, which can be split into incoming and outgoing flux, i.e. $\sum_\mu\Delta T_\mu \mathrm{ReLu}(\mathbf{S}_\mu \cdot\mathbf{\hat{n}}) = -\sum_\mu \Delta T_\mu \mathrm{ReLu}(-\mathbf{S}_\mu \cdot\mathbf{\hat{n}})$; note that we define $\mathrm{ReLu}(x) = x \Theta(x)$, with $\Theta(x)$ being the Heaviside function. The boundary condition is imposed to phonons leaving the surface; generally, we have $\Delta T_\mu^{-} = \sum_\nu R_{\mu\nu}\Delta T_\nu^+$, where $-$ ($+$) are for outgoing (incoming) phonons, and $R_{\mu\nu}$ is a mode-resolved bouncing matrix. However, currently only a simplified model is available, with which we assume that phonons thermalize to a single temperature~\cite{landon2014deviational}; this choice corresponds to $R_{\mu\nu} = \left[ \sum_k \mathrm{ReLu}(\mathbf{S}_k \cdot\mathbf{\hat{n}})\right]^{-1} \mathrm{ReLu}(\mathbf{S}_\mu \cdot\mathbf{\hat{n}})$. Lastly, the effective thermal conductivity, assuming we have imposed the temperature gradient along $x$, reads
\begin{equation}\label{keff}
\kappa^{\mathrm{eff}} = - \frac{L_x}{\Delta T_{\mathrm{ext}}} \sum_\mu \int_{A}\frac{d\mathbf{S}}{A} \mathbf{J}_\mu\cdot \mathbf{\hat{n}}.
\end{equation}
Equation~\ref{rta}, when solved iteratively, requires inverting as many linear systems as the number of phonons modes, an endeavour that may easily become prohibitive. To ameliorate the computational effort, we instead implement the recently introduced anisotropic-MFP-BTE (aMFP-BTE), a model based on interpolating the phonon temperatures onto the vectorial MFP space~\cite{romano2021}. In practice, instead of solving Eq.~\ref{rta} $N_pN_q$ times at each iteration, we solve it for a set of vectorial MFPs, $\mathbf{F}_{ml}$, located on a spherical grid. The indices $m$ and $l$ label the MFP $\Lambda_m$ (up to $N_\Lambda$) and phonon directions $\mathbf{\hat{s}}_l$ (up to $N_\Omega$), respectively. As $N_\Lambda N_\Omega << N_p N_q$, while not compromising on accuracy, this scheme results in a much faster runtime with respect to the case with mode-resolved resolution, i.e. no interpolation.
\section{Implementation}
As derived in~\ref{amfpbte}, the discretization of aMFP-BTE model in real and momentum space yields the following iterative linear system
\begin{eqnarray}\label{iter}
\mathbf{A}_{ml} \mathbf{\Delta T}_{ml}^{(n)} =\sum_{m'l'} \mathbf{S}_{ml}^{m'l'} \mathbf{\Delta T}_{m'l'}^{(n-1)} + \mathbf{P}_{ml},
\end{eqnarray}
where $\mathbf{A}_{ml}$ is the stiffness matrix for a given MFP and direction, $\mathbf{S}_{ml}^{m'l'}$ is associated to the lattice temperature and adiabatic boundary conditions, and $\mathbf{P}_{ml}$ represents the perturbation. Once Eq.~\ref{iter} converges, the effective thermal conductivity is computed as $\kappa_{\mathrm{eff}}=\kappa_{\mathrm{bal}}/2 + \sum_{ml}\mathbf{K}_{ml}^T \cdot \mathbf{\Delta T}_{ml}$, where $\kappa_{\mathrm{bal}}$ is the ballistic thermal conductivity, and $\mathbf{K}_{ml}$ is associated to the thermal flux.
The first guess to Eq.~\ref{iter} is given by the standard heat conduction equation. While it is straightforward to discretize the diffusive equation in orthogonal grids, it becomes convoluted for unstructured meshes. In \verb!OpenBTE!, we use the approach described in~\cite{murthy2002numerical}, where the non-orthogonal contribution of the flux is treated explicitly, i.e. it depends on previous solutions. The resulting system, analogously to Eq.~\ref{iter}, is an iterative linear system,
\begin{equation}\label{fourier_iter}
\mathbf{A} \mathbf{\Delta T}^{(n)}= \mathbf{S} \mathbf{\Delta T}^{(n-1)} + \mathbf{P},
\end{equation}
where $\mathbf{S}$ arises from the non-orthogonality of the mesh. Note that $\mathbf{S}_{ml}^{m'l'}$ in Eq.~\ref{iter} and $\mathbf{S}$ in Eq.~\ref{fourier_iter} are not related. The same applies with the terms sharing the same basename in these two equations. The derivation of Eq.~\ref{fourier_iter} is given in~\ref{fourier}.
Equation~\ref{iter} entails, for each iteration $n$, the inversion of $N_\Lambda N_\Omega$ matrices, whose size is the number of volumes in the computational domain. To enhance the computational efficiency, three distinct strategies have been put in place. First, by noting that the stiffness matrix does not depend on $n$, we are able to compute LU factorizations for the first iteration, and reuse them until convergence. Another major speed up is achieved by vectorizing the assembly of the stiffness matrices, while also exploiting their sparsity. Lastly, for each $n$, Eq.~\ref{iter} consists of a set of independent problems, thus it can be trivially parallelized~\cite{ali2014large}; to this end, we assign a set of phonon directions to each core ($\Omega_c$), while the indices $m$ run serially. The implementations of these three methods rely on \verb!SCIPY!, \verb!NUMPY!~\cite{harris2020array} and \verb!MPI4PY!~\cite{dalcin2011parallel}, respectively.
When vectorization and parallelization are both used, we may save memory by noting that, as shown in~\ref{amfpbte}, the stiffness matrices have the form $\mathbf{A}_{ml} = \mathbf{I} + \Lambda_m \mathbf{A}_l$; thus vectorization can be done only for $\mathbf{A}_l$, while $\mathbf{A}_{ml}$ is assembled \textit{on the fly} right before LU factorization. Similar arguments hold for $\mathbf{P}_{ml}$ and $\mathbf{S}_{ml}^{m'l'}$. The overall methodology is summerized in Algorithm~\ref{algo}, where, for readability, only $\mathbf{A}_{ml}$ has been expanded upon.
\begin{algorithm}[H]
\caption{Workflow for thermal conductivity calculations. The term $\Omega_c$ is the set of solid angles assigned to core $c$.}
\label{algo}
\begin{algorithmic}[1]
\State $\mathbf{\Delta T}_{ml}^{(0)}\leftarrow$ Fourier's law.
\While{error $>1e^{-3}$}
\State $\kappa^{\mathrm{eff},(n+1)} = \kappa_{\mathrm{bal}}/2$
\For{l in $\{\Omega_c\}$}
\For{m = 1:$N_{\Lambda}$}
\State $ \mathbf{B} = \sum_{m'l'} \mathbf{S}_{ml}^{m'l'} \mathbf{\Delta T}_{m'l'}^{(n-1)} + \mathbf{P}_{ml}$
\State $ \mathbf{\Delta T}_{ml}^{(n)} \leftarrow \left(\mathbf{I} + \Lambda_m \mathbf{A}_l \right)\mathbf{\Delta T}_{ml}^{(n)} = \mathbf{B} $
\State $\kappa^{\mathrm{eff},(n+1)} += \mathbf{K}_{ml}^T \cdot \mathbf{\Delta T}_{ml} $
\EndFor
\EndFor
\State error $\gets |(\kappa^{\mathrm{eff},(n+1)} - \kappa^{\mathrm{eff},(n)})|/\kappa^{\mathrm{eff},(n+1)}$
\State $n \gets n +1$
\EndWhile\label{euclidendwhile}
\end{algorithmic}
\end{algorithm}
\section{Workflow}
The easiest way to run \verb!OpenBTE! is to use it as a module in \verb!Python!. A sketch for the workflow is illustrated in Fig.~\ref{fig11}. The core module is \verb!Solver!, which runs both the heat conduction equation as a first guess to Eq.~\ref{iter} and all subsequent iterations, outlined in Algorithm~\ref{algo}. Before invoking it, though, it is necessary to define the material's internal boundaries using the \verb!Geometry! module, and bulk-related data via the \verb!Material! module. The latter is the interface to external packages. Once the simulation is finalized, results can be plotted using the \verb!Plot! module. Note that \verb!OpenBTE! follows a pure functional approach, where data is simply stored as dictionaries. Each module is expanded upon in the next sections.
\begin{figure}
{\includegraphics[width=0.9\textwidth]{figure_1.png}}
\cprotect\caption{The software architecture of \verb!OpenBTE!. The key components are the \verb!Material!, \verb!Geometry!, \verb!Solver! and \verb!Plot! modules. Currently, the code is interfaced to \verb!AlmaBTE! and \verb!Phono3Py!, which computes first-principles data. Mesh generation is delegated to \verb!GMSH!, and output maps can be visualized with \verb!Paraview!.\label{fig11}}
\end{figure}
\section{Material}
The \verb!Material! module converts mode-resolved bulk data, $C_\mu$, $\tau_\mu$ and $\mathbf{v}_\mu$, into the spherically-resolved quantities $\mathbf{G}_{ml}$ and $C_{ml}$. The input files, \verb!rta.npz!, are created by postprocessing data obtained by external packages. Currently, \verb!OpenBTE! provides interfaces to \verb!AlmbaBTE!~\cite{Carrete2017AlmaBTEMaterials} and \verb!Phono3Py!~\cite{phono3py}. A small set of precomputed \verb!rta.npz! is also provided. The file created by \verb!Material! is stored into \verb!material.npz! and ready to be used by \verb!Solver!. Depending on the geometry, this module has two models:
\begin{itemize}
\item \verb!rta3D!: this model must be used for geometries with $l_z > 0$. In this case, the angular interpolation of $\mathbf{F}_\mu$ is performed in both polar and azimuthal angles.
\item \verb!rta2DSym!: when $l_z$ is not specified, the actual simulation domain is 2D, and this model must be used. Note that when the base material is 3D, using this option is equivalent to considering an infinite thickness. In fact, in this case, the mode-resolved $\mathbf{F}_\mu$ are mapped into $\mathbf{F}_{tk}=\Lambda_t \cos(\phi_k)$, where $\Lambda_t$ implicitly includes the MFP and the azimuthal angle, and $\phi_k$ is the polar angle on the \textit{x-y} plane~\cite{romano2021}.
\end{itemize}
The model as well as the discretization grid can be chosen with
\begin{minted}[breaklines]{python}
from openbte import Material
Material(filename='rta_Si_300',model='rta2DSym',n_phi=48)
\end{minted}
where, by default, the material file is taken from the database, whose entries are updated online. In this case, we choose to discretize the polar angle into 48 slices.
\section{Geometry module}
The \verb!Geometry! module allows the user to build an arbitrarily patterned nanomaterial. It interfaces with \verb!GMSH!~\cite{geuzaine2009gmsh} to create an unstructured Delaney mesh, and creates the file \verb!geometry.npz!. While we plan to handle user-provided meshes, \verb!OpenBTE! currently supports 2D membranes with finite and infinite thickness. The size of the unit cell is specified by \verb!lx! and \verb!ly!, while the thickness is defined by \verb!lz!. The direction of the applied gradient is indicated with \verb!direction! (default is \verb!x!). Periodic boundary conditions are applied along all directions; however, adiabatic surfaces can be imposed along specific directions with \verb!Periodic=[True,False,True]!, where each component declares whether a surface is periodic. Currently, there are two geometry models, as expanded upon in the next sections.
\subsection{Lattice model}
This model aids the creation of 2D periodic porous materials with pores having either predefined shapes or custom ones. In analogy to crystals, a porous material is defined as a set of unit vectors, which in our case are $\mathbf{a}_1 = l_x \mathbf{\hat{x}}$ and $\mathbf{a}_2 = l_y \mathbf{\hat{y}}$, and a base. The latter is a list of \textit{x-y} pairs, each of them being one pore. Here is an example with a single pore in the origin
\begin{minted}[breaklines]{python}
from openbte import Geometry
Geometry(model='lattice',base=[[0,0]],shape='circle')
\end{minted}
In the example above, we chose a simple shape, a circle; however, it is possible also to supply a user-defined shape using \verb!shape=custom!. In such a case, the keyword \verb!shape_functions! must be provided, optionally along with \verb!shape_options!.
For example, as shown in Fig.~\ref{fig:shape}a-b and c, we can define a structure having two pores, each of them described by the same custom shape.
\subsection{Custom model}
Using the option \verb!model=custom!, it is possible to define an arbitrary patterning by providing a list of polygons, with the periodicity being automatically guaranteed. This option significantly widens up the number of possible geometries at the expense of more input by the user. In spirit of \textit{subtractive manufacturing}, each polygon can be seen as a region where the material is carved out. An example of custom model is depicted in Fig.\ref{fig:shape}-d.
\begin{figure}
\begin{center}
{\includegraphics[width=1\textwidth]{shape.png}}
\cprotect\caption{a) An example of structure using the keyword-value pair \verb!shape=custom! for the \verb!lattice! model of the \verb!Geometry! module. b) the unit cell comprising two pores at designed positions. Note that when a pore crosses a boundary it gets repeated. c) singe-pore view of material. The coordinates of the shape are defined via the \verb!shape_function! option. d) An example of \verb!custom! shape obtained by carving out the base material with user-defined polygons. \label{fig:shape} }
\end{center}
\end{figure}
\section{Plot Module}
Simulations results can be plotted via the \verb!Plot! module. Currently, the following features are implemented:
\begin{itemize}
\item \verb!vtu!: the temperature and flux maps are stored in a \verb!vtu! file and ready to be opened by \verb!Paraview!~\cite{ahrens2005paraview}.
\item \verb!maps!: the maps are shown as web-ready, interactive structures.
\item \verb!kappa_mode!: this model computes the mode-resolved thermal conductivity from Eq~\ref{mode-resolved}, and stores in a file.
\item \verb!line!: this option, which currently is implemented only for 2D systems, allows for line plots, using linear interpolation.
\end{itemize}
\section{Thermal transport in porous Si}
\begin{figure*}[t!]
\subfloat[\label{fig1a}]
{\includegraphics[width=0.5\textwidth]{figure_2a.png}}
\hfill
\subfloat[\label{fig1b}]
{\includegraphics[width=0.5\textwidth]{figure_2b.png}}
\quad
\subfloat[\label{fig1c}]
{\includegraphics[width=0.5\textwidth]{figure_2c.png}}
\hfill
\subfloat[\label{fig1d}]
{\includegraphics[width=0.5\textwidth]{figure_2d.png}}
\caption[]{a) Thermal transport reduction when using Fourier's law, for regular circular pores and for a user-define shape; the Eucken-Maxwell model is also plotted. In the inset, the magnitude of the thermal map for the user-defined structure. The porosity is 0.1 and $L$ = 50 nm. b) a cut of the temperature maps, computed from both the BTE and Fourier's law, along the $x$-axis. c) The effective thermal conductivity, $\kappa^{\mathrm{eff}}$, of the user-defined structure with thickness $t$ = 10 nm, and for different porosities. In the inset, the case with porosities = 0.1 and 0.17 are shown. d) The mode-resolved thermal conductivity, $\kappa^{\mathrm{eff}}_\mu$, versus the MFP, $|F_{\mu,x}|$, for the cases $t$ = 10 nm and $t = \infty$.}
\end{figure*}
In some cases, it is convenient to use Fourier's law to estimate macroscopic geometric effects. In this first part, we thus aim to assess the reliability of the diffusive solver, implemented in \verb!OpenBTE!. To this end, we consider a membrane with circular pores and infinite thickness, for which $\kappa_\mathrm{fourier}/\kappa_{\mathrm{bulk}} = r$ can be evaluated analytically. In fact, in this case, the thermal transport suppression is given by the Eucken-Maxwell model $r = (1-\phi)/(1+\phi) $~\cite{hasselman1987effective}, $\phi$ being the porosity. In Fig.~\ref{fig1a}, we compare this model with the results from \verb!OpenBTE! for different porosities, obtaining excellent agreement ($<1\%$). We note that by using the option \verb!only_fourier=True! it is possible to run only the Fourier's model. Next, we compute $r$ for a more complicated structure, comprising two concatenated sub-lattices of pores with different user-provided shapes; this structure is generated using the \verb!shape=custom! option from the \verb!lattice! model, a two-pore base, and different options for each pore. As shown in Fig.~\ref{fig1a}, we obtain a much larger heat transport suppression due to the longest path heat must travel through the material.
Let us now compute $\kappa^{\mathrm{eff}}$ using the BTE, encoded in Eq.~\ref{iter}. For the chosen porosity of $0.1$ and $L$ = 50nm, we obtain ~17.4 Wm$^{-1}$K$^{-1}$, well below the corresponding macroscopic value~99.4 Wm$^{-1}$K$^{-1}$. In Fig.~\ref{fig1b}, we plot a cut of the temperature maps, both for the BTE and Fourier cases, along the $x$-axis. As expected, they have different trends, the BTE one being flatter due to ballistic transport. When a finite thickness is added, phonon undergo stronger suppression; in fact, for thickness $t$=10 nm and porosity 0.1, we obtain $\kappa^{\mathrm{eff}} = 8.9$ W m$^{-1}$K$^{-1}$, roughly a half of the value with $t = \infty$. In Fig.~\ref{fig1c}, $\kappa^{\mathrm{eff}}$ for different porosities is plotted. From the insets, we note that, as the porosity becomes larger, the heat flux peaks around the phonon bottleneck, e.g. the areas between the pores. Lastly, the mode-resolved thermal conductivities for $t$ = 10 nm and $\infty$ are plotted in Fig.~\ref{fig1d}; analogously to Ref.~\cite{romano2021}, we can appreciate the stronger suppression for large MFPs for the case with finite thickness.
\section{Conclusion}
We have presented \verb!OpenBTE!, a software for computing deterministically space-dependent heat transport in 2D and 3D structures. The efficient and modern implementation of the underlying algorithms enable accurate simulations both in the cloud and on common laptops, thus democratizing this type of simulations. Future work includes going beyond the RTA of the BTE and adding transient dynamics. GPU support and differentiability are also on the roadmap. Furthermore, we plan to develop a hybrid Fourier/BTE solver, which has the potential to dramatically reduce runtimes for the cases with 3D structures. Filling an important gap in the current offer of publicly available phonon transport solvers, \verb!OpenBTE! sets out to guide experiments on nanoscale heat transport and facilitate high-throughput prediction of thermal properties of nanostructures.
\section{Acknowledgments}
Research was partially supported by the Solid-State SolarThermal Energy Conversion Center (S3TEC), an Energy Frontier Research Center funded by the U.S. Department of Energy (DOE), Office of Science, Basic Energy Sciences (BES), under Award No. DESC0001. The author thanks all the early users, who helped improving \verb!OpenBTE!.
|
\section{Introduction}
\label{sec:intro}
Stochastic gradient methods including stochastic gradient descent (SGD)~\citep{SGD-1951} and its accelerated variants (e.g., SGD with momentum \citep{polyak1964some,sutskever2013importance}) have become the algorithmic workhorse in much of machine learning. The step-size (learning rate) is the most important hyper-parameter for controlling the speed at which gradient-based methods converge to stationarity. For problems with multiple local minima, the step-size also affects which local optimum the optimization process converges to. It therefore needs to be both well-designed and well-tuned to make SGD and its variants effective in practice.
\iffalse
Comments: delete this paragraph ?
\textcolor{red}{Numerous strategies have been proposed to set the step-size, either based on theory \citep{Moulines-Bach2011, hazan2014beyond,jain2019making} or empirical evidence~\citep{Adam, loshchilov2016sgdr,lang2019using}. Generally speaking, if the gradient is subject to additive noise, a constant step-size can promote a fast converge \citep{gower2019sgd,shi2020learning} to the ball around a stationary point. The convergence of SGD with a large step-size, as long as the iterates remain stable, is faster than a smaller step-size but gives rise to a higher residual error~\citep{shi2020learning,wang2021convergence}. The most successful step-size in practice, especially for deep learning, is the ``constant and then drop'' policy~\citep{imagenet, he2016deep, ge2019step,yuan2019stagewise,liu2020improved, wang2021convergence}. In this paper, we call the ``constant and then drop by a constant'' policy \emph{step-decay}. The policy initially applies a relatively large step-size to achieve a faster convergence and then decrease the step-size to increase the final solution accuracy. }
\fi
In the deep learning literature, cyclical step-sizes \citep{loshchilov2016sgdr, smith2017cyclical} and non-monotonic schedules \citep{keskar2015nonmonotone, sinewave, seong2018towards} have attracted strong recent interest, with significant benefits
for non-convex problems with poor local minima or saddle points~\citep{seong2018towards}.
Popular
cyclical schedules include the cosine step-size (cosine with restart) \citep{loshchilov2016sgdr} and the triangular policy~\citep{smith2017cyclical}, which have become the default choices in some deep learning libraries, e.g., PyTorch and TensorFlow (cf. lr\_scheduler.CyclicLR and CosineAnnealingLR). However, non-monotonic policies are much more complex to analyze than decaying ones, and theoretical results for these non-monotonic policies are scarce. This motivates us to focus on a bandwidth step-size framework, in which
\begin{align}
m\delta(t) &\leq \eta^t \leq M \delta(t) \label{eqn:bnd_def}
\end{align}
for some boundary function $\delta(t)$ and positive constants $m$ and $M$.
This framework allows for non-monotonic step-sizes and covers most of the situations discussed above. In particular, it includes the cosine~\citep{loshchilov2016sgdr}, triangular~\citep{smith2017cyclical}, sine wave~\citep{sinewave} step-sizes as special cases. The framework provides a uniform convergence rate guarantee for all step-size policies which remain in the band (\ref{eqn:bnd_def}). This gives a lot of freedom to design novel step-sizes schedules with improved practical performance without loosing track of their theoretical convergence guarantee.
Note that the framework also allows adaptive step-size policies, as long as the step-size is guaranteed to belong to a band on the form (\ref{eqn:bnd_def}). For example, the trust-region-ish algorithms by~\citet{curtis2019stochastic} uses a careful step normalization procedure to adapt the learning rate based on the gradient norm. It is not easy to derive a convergence rate guarantee for such a complex policy, but it can be captured by the bandwidth step-size framework and therefore inherits its theoretical convergence properties. In short, the bandwidth step-size framework is an alternative, and often easier, way to get the worst-case theoretical guarantees for complex step-size policies such as \cite{curtis2019stochastic} and other alarm conditions. A more detailed discussion is given in Section~\ref{rem:trust-region}.
The generic bandwidth framework has recently been proposed by~\citet{wang2021convergence2}, but they only analyzed strongly convex problems. We believe that more significant potential lies in the non-convex regime. For non-convex problems, non-monotonic step-sizes have distinct advantages, helping iterates to escape local minima and producing final iterates of high quality. In the paper, we demonstrate this point on both a simple toy example and on large-scale neural network training tasks.
Our main contribution is a sequence of non-asymptotic convergence results for the bandwidth step-size on non-convex optimization problems, based on the popular ``constant and then drop'' step-size schedules~\citep{imagenet, he2016deep,hazan2014beyond,ge2019step, wang2021convergence}. This allows non-monotonic variations both within each (inner) stage and between stages.
\subsection{Contributions}\label{sec:contribution}
Inspired by the strong potential of non-monotonic step-size schedules demonstrated above, we extend the bandwidth-based step-size framework to \emph{``constant and then drop''} (multi-stage) profiles, where the bands stay constant throughout each stage and drops between stages. We provide convergence guarantees for both SGD and its momentum variant (SGDM) on non-convex problems. Specifically,
\begin{itemize}
\item We establish worst-case theoretical guarantees for SGD with bandwidth step-size on smooth nonconvex problems. We \textbf{(i)} derive an optimal rate for SGD under a bandwidth step-size with $\delta(t)=1/\sqrt{t}$; \textbf{(ii)} and achieve optimal and near-optimal rates for step-decay (constant and then drop by a constant), improving the
results by~\citet{wang2021convergence}.
\item We provide worst-case theoretical guarantees for SGDM with bandwidth-based step-decay step-size in the smooth nonconvex setting. To the best of our knowledge, this is the first results that provide optimal (Theorem \ref{thm:mom:stepdecay:2}) and near-optimal (Theorem \ref{thm:mom:stepdecay:0}) results for momentum with \emph{step-decay} step-sizes. Moreover, our results significantly improve the convergence results from \cite{liu2020improved} (see Remark \ref{rem:liu}).
\item
Our analysis results also provide state-of-art theoretical guarantees for \emph{cosine}~\citep{loshchilov2016sgdr} and \emph{triangular}~\citep{smith2017cyclical} step-sizes~ if their boundary functions are within our bands. {Especially, we improve the result of \citet{li2020exponential} for cosine step-size and achieve a state-of-art rate (see Remark \ref{rem:cyclical}). Moreover, our results first provide the convergence guarantees for triangular step-size~\citep{smith2017cyclical}. }
\item We propose novel, possibly non-monotonic, step-size schedules (e.g., step-decay with linear-mode and cosine-mode) based on the bandwidth-based framework and demonstrate their efficacy on several large-scale neural network training tasks.
\end{itemize}
\subsection{Related Work}\label{sec:related:work}
This subsection reviews the theoretical development of the SGD algorithm and its momentum variant in the smooth non-convex setting, with a special focus on different step-size policies.
{\bf SGD for nonconvex problems} The first non-asymptotic convergence of SGD to a stationary point of a general smooth non-convex function was established in~\cite{ghadimi2013stochastic}. The authors proved that a constant step-size ${\mathcal O}(1/\sqrt{T})$ attains a convergence rate of $ \mathcal{O}(1/\sqrt{T})$, {where $T$ is the iteration budget}.
To the best of our knowledge, this rate
is not improvable and was proven to be tight up to a constant without additional assumptions~\citep{drori2020complexity}. For the $1/\sqrt{t}$ decay step-size, an $\mathcal{O}(\ln T/\sqrt{T})$ rate can be easily obtained from \citep{ghadimi2013stochastic}. This rate can be improved to the optimal by selecting a random iterate using weights proportional to the inverse of the step-size~\citep{wang2021convergence}.
The sampling rule in \citep{wang2021convergence} depends on the step-size and is easily applicable to different step-size policies. Thus, in this paper, we choose a similar sampling rule as \citep{wang2021convergence} to favor the later iterates when selecting the output for SGD and its momentum variant.
{\bf Step-decay step-sizes} Recently, the theoretical performance of step-decay or stagewise strategies has attracted an increasing attention due to their excellent practical performance~\citep{yuan2019stagewise,ge2019step,chen2018universal,li2020exponential,wang2021convergence}.
For a class of least-squares problems, \cite{ge2019step} established a near-optimal $\mathcal{O}(\ln T/T)$ rate for the step-decay step-size (cut by 2 every $T/\log_2(T)$ iterates) and showed that step-decay can perform better than the polynomial decay step-size. Stochastic optimization methods with stagewise step-sizes decaying as $1/t$ were analyzed in~\cite{chen2018universal}.
A near-optimal rate for the continuous version of step-decay, called exp-decay, as well as for cosine decay step-sizes under the Polyak-L\'ojasiewicz (PL) condition and a general smooth assumption were established in~\cite{li2020exponential}. However, in the smooth case, to achieve such results for exponential and cosine decay step-sizes, the initial step-size is required to be bounded by $\mathcal{O}(1/\sqrt{T})$. This is obviously impractical when the number of iterations $T$ is large. Near-optimal rates (up to $\ln T$) of SGD with step-decay step-size in several general settings including strongly convex, convex and smooth (non-convex) problems were proved in
\cite{wang2021convergence}. They also removed the restriction on the initial step-size for exponential decay step-sizes. Empirical evidences have been given in~\citep{wang2021convergence2} that bandwidth-based strategies can improve the performance of the step-decay step-size on some large scale neural network tasks. However, no theoretical guarantees for non-convex problems were given.
{\bf SGD with momentum on nonconvex problems} The momentum variant of SGD (SGDM) has been widely used in deep neural networks~\citep{imagenet,sutskever2013importance,he2016deep,zagoruyko2016wide}. Due to its practical success on neural networks, its theoretical performance is now attracting a lot of interest, especially for nonconvex problems~\citep{unified_momentum,gadat2018stochastic,chen2018universal,gitman2019understanding,mai2020convergence,liu2020improved,defazio2020understanding}. Under the assumption of bounded gradients, \cite{unified_momentum} proposed a unified analysis framework for stochastic momentum methods and proved an optimal $\mathcal{O}(1/\sqrt{T})$ rate under constant step-sizes. A similar result for the Nesterov-accelerated variant was established in~\cite{ghadimi2016accelerated}. However, studies related to the multi-stage performance of SGD with momentum is lacking and far from being complete. Reference~\cite{chen2018universal} considers a momentum method with a stagewise step-size, but the method is a proximal point algorithm with extra averaging between stages, and not the widely used momentum SGD considered here.
More recently, \cite{liu2020improved} established the convergence for multi-stage SGDM and provided empirical evidence to show that multi-stage SGDM is faster. However, their results require an inverse relationship between stage length and step-size which limits the initial stage length or step-size. A detailed comparison with \citep{liu2020improved} will be given in Section \ref{sec:4:mom} (see Remark~\ref{rem:liu}).
\textbf{Organization:} The rest of this paper is organized as follows. Notations and basic definitions are introduced in Section~\ref{sec:prelimi}. Our novel theoretical results for SGD and its momentum variant (SGDM) under bandwidth-based step-sizes are introduced in Sections \ref{sec:3:sgd} and \ref{sec:4:mom}, respectively. Numerical experiments are presented and reported in Section~\ref{sec:numerical}. Finally, Section~\ref{sec:conclusion} concludes the paper.
\section{Problem Set-Up}
\label{sec:prelimi}
\iffalse
{\color{blue}
\begin{itemize}
\item should we reserve index $t$ for stage? \textcolor{red}{Yes, it is better to keep $t$ as stage index. }
\item should we already here say that we consider multi-stage SGD? \textcolor{red}{Yes, I agree. We can directly introduce multi-stage stochastic gradient method here. Then we do not need to repeat it for momentum. }
\item We can write the assumptions a little more compact, and also perhaps also in a little more traditional way
\item One thing that is a little bit inconsistent is the time-dependence on the step-size and boundary functions. I know that you probably want to keep a similar notation as in your earlier work, but I would have preferred to have $\eta_t$ rather than $\eta(t)$, and perhaps also change the indexing for $\delta_1(t)$ and $\delta_2(t)$ (but this is less important, since they do not appear so often). One more comment: you say that $\delta_1(t)$ and $\delta_t(2)$ should be monotonic, but do not restric them to be decreasing (so they could be monotonically increasing\dots) \textcolor{red}{Yes, we can use the notation $\eta_t$ instead of $\eta(t)$. How to make $\delta_1(t)$ and $\delta_2(t)$ look nicer ? In the previous version, I use $d\delta_1(t)/dt \leq 0$ and $d\delta_2(t)/dt \leq 0$.}
\end{itemize}
}
\fi
We study the following, possibly non-convex, stochastic optimization problem
\begin{equation}\label{P1}
\min_{x\in \mathbb{R}^d} f(x) = \mathbb{E}_{\xi \sim \Xi }[ f(x; \xi)]
\end{equation}
where $\xi$ is a random variable drawn from some (unknown) probability distribution $\Xi$ and $f(x;\xi)$ is the instantaneous loss function over the variable $x \in \mathbb{R}^d$. We consider stochastic gradient methods that generate iterates $x^t$ according to
\begin{align}
x^{t+1} = x^t - \eta^t d^t
\end{align}
where $\eta^t$ is the step-size and $d^t$ the search direction (e.g., $d^t = \nabla f(x^t;\xi)$ for SGD). We assume that
there are constants $m>0$ and $M\geq m$, and two functions $n(t)$ and $\delta(t)$: $\mathbb{R} \rightarrow \mathbb{R} $ such that such that
\begin{equation*}
\eta^t = n(t)\delta(t), \,\forall \, t \geq 1,
\end{equation*}
where $n(t) \in [m, M]$ and $\delta(t)$ is monotonically decreasing function satisfying $\delta(1)=1$.
Note that even though the boundary function $\delta(t)$ is monotonic, the step-size itself is not restricted to be. Throughout the paper, we make the following assumptions:
\begin{assumption}\label{assumpt:func} The loss function $f$ satisfies
$\Vert \nabla f(x)-\nabla f(y)\Vert \leq L\Vert x-y\Vert$ for every $x,y \in \mbox{dom}\, (f)$.
\end{assumption}
\begin{assumption}\label{assump:gradient-oracle} For any input vector $x$, the stochastic gradient oracle $\mathcal{O}$ returns a vector $g$ such that
(a) $\mathbb{E}[\left\| g - \nabla f(x) \right\|^2 ] \leq \rho \left\| \nabla f(x) \right\|^2 + \sigma$ where $\rho \geq 0$ and $\sigma \geq 0$; (b) $\mathbb{E}[\left\| g \right\|^2] \leq G^2$.
\end{assumption}
\section{Non-asymptotic Convergence of SGD with Bandwidth-based Step-Size }\label{sec:3:sgd}
\iffalse
{\color{blue}
\begin{itemize}
\item in the first paragraph, we do not perhaps need to say that they include polynomial decay and step-decay, but rather emphasize that it is a rather general class of bandwidth-based step-sizes, which supports multi-stage SGD variants.
\end{itemize}
}
\fi
In this section, we provide the first non-asymptotic convergence guarantees for SGD with bandwidth-based step-sizes on smooth non-convex problems. The results consider a general family of bandwidth-based step-sizes which includes the classical multi-stage SGD as a special case.
Algorithm~\ref{alg:sgd} details our bandwidth-based version of the popular ``constant and then drop'' policy for SGD. Here, the boundary function $\delta(t)$
is adjusted in an outer stage, and the length of each stage $S_t$ is allowed to vary. Similar to~\cite{wang2021convergence}, the output distribution depends on the inverse of $\delta(t)$, hence puts more weight on the final iterates. By considering specific combinations of $\delta(t)$ and $S_t$, this framework allows us to analyze several important multi-stage SGD algorithms, including those with constant, polynomial-decay and step-decay step-sizes. Many interesting results on polynomial-decay step-size {(e.g., $\delta(t) = 1/\sqrt{t}$, we called it $1/\sqrt{t}$-band)} are given in Appendix \ref{subsec:poly:sgd}.
\begin{algorithm}[t]
\caption{SGD with Bandwidth-based Step-Size}\label{alg:sgd}
\begin{algorithmic}[1]
\STATE \textbf{Input:} initial point $x_1^1$, \# iterations $T$, \# stages $N$, stage length $\left\lbrace S_t \right\rbrace_{t=1}^{N}$ such that $\sum_{t=1}^{N}S_t = T$, {the sequences $\left\lbrace \delta(t)\right\rbrace_{t=1}^{N} $ and $\left\lbrace\left\lbrace n(t, i)\right\rbrace_{i=1}^{S_t}\right\rbrace_{t=1}^N \in [m, M]$ with $0 < m \leq M$}
\FOR{ $t = 1: N$ }
\FOR{ $i = 1: S_t$}
\STATE
Query a stochastic gradient oracle $\mathcal{O} $ at $x_i^t$ to get a vector $g_i^t$ such that $\mathbb{E}[g_i^t \mid \mathcal{F}_i^t] = \nabla f(x_i^t)$\footnotemark{}
\\\vspace{0.01in}
\STATE Update step-size $\eta_i^t = n(t,i)\delta(t)$
\STATE $x_{i+1}^t = x_i^t - \eta_i^t g_i^t$
\ENDFOR
\STATE $x_{1}^{t+1} = x_{S_t+1}^{t}$
\ENDFOR
\STATE {\bf Return:} $\hat{x}_T$ is uniformly chosen from $\left\lbrace x_{1}^{t^{\ast}}, x_{2}^{t^{\ast}}, \cdots, x_{S_{t^{\ast}}}^{t^{\ast}} \right\rbrace$, where the integer $t^{\ast} $ is chosen from $\left\lbrace 1, 2,\cdots, N\right\rbrace$ with probability $P_t = \delta^{-1}(t)/(\sum_{l=1}^{N}\delta^{-1}(l))$
\end{algorithmic}
\end{algorithm}
\footnotetext{We use $\mathcal{F}_i^t$ to denote $\sigma$-algebra formed by all the random information before current iterate $x_i^t$ and $x_i^t \in \mathcal{F}_i^t$. }
\iffalse
\added{deleted this Lemma ?}
\begin{lemma}\label{lem:sgd}
Suppose that Assumption~\ref{assumpt:func} and Assumption \ref{assump:gradient-oracle}(a) hold. If we run the Algorithm \ref{alg:sgd} with $T > 1$ and $\eta_i^t \leq 1/((\rho+1) L)$, we have
\begin{align}\label{eqn:unifiedBound}
\mathbb{E}[\left\|\nabla f(\hat{x}_T)\right\|^2] \leq \frac{1}{\displaystyle \sum_{t=1}^N S_t \delta^{-1}(t)}\left(\sum_{t=1}^{N}\frac{2(\mathbb{E}[f(x_1^{t})] - \mathbb{E}[f(x_{1}^{t+1})])}{m\delta(t)^2} + \frac{M^2L\sigma}{m} T\right).
\end{align}
\end{lemma}
\fi
\iffalse
\deleted{A key challenge in the analysis comes in estimating the first sum inside the parentheses in the right-hand side of (\ref{eqn:unifiedBound}). In the stochastic setting, the expected function value is not necessarily decreasing (as it is in the deterministic case~\citep{cartis2010complexity}). \textcolor{red}{But it is reasonable to have an upper bound that $\mathbb{E}[f(x_1^t) - f^{\ast}] \leq \Delta_0$ for each outer-stage $t \geq 1$, where $f^{\ast} = \min_{z} f(z)$. We believe that it is a fair assumption to make, and it has never been violated in our numerical experiments. Say something: \cite{hazan2015beyond}.
} }
\added{
A key challenge in the analysis comes in estimating the first sum inside the parentheses in the right-hand side of (\ref{eqn:unifiedBound}). In the stochastic setting, the expected function value is not necessarily decreasing (as it is in the deterministic case~\citep{cartis2010complexity}). We require that the expectation of the function value at each iterate $\mathbb{E}[f(x_1^t)$ is uniformly upper bounded. As shown by \cite{shi2020learning}, for any step-size (but not very large for example $\eta < 1/L$), then the function value at the iteration trajectory can be controlled by the initial state at the iterate $x_1^1$ which exponentially decays plus an error term (in the continue time). We believe that it is a fair assumption to make. Nevertheless, this assumption has been commonly used in any papers \cite{}, and it has never been violated in our numerical experiments.}
\fi
\subsection{Convergence Under Bandwidth Step-Decay Step-Size}\label{subsec:stepdecay:sgd}
Another
important step-size is Step-Decay (``constant and then drop by a constant''), which is popular and widely used in practice, e.g. for neural network training \citep{imagenet,he2016deep}. In this subsection, we analyze bandwidth step-sizes that include step-decay as a special case.
For Step-Decay, the stage length $S_t$ is typically a hyper-parameter selected by experience. We first analyze a bandwidth version of the algorithm analyzed in~[Theorem~3.2]\citep{wang2021convergence}, namely Algorithm~\ref{alg:sgd} with $N =(\log_{\alpha} T)/2$ ($\alpha > 1$) outer loops, each with a constant length of with $S_t = 2T/\log_{\alpha} T$. The logarithmic dependence of $N$ on $T$ leads to a small number of stages in practice, and was demonstrated to perform well in deep neural network tasks \citep{wang2021convergence}.
\begin{theorem}\label{thm:band:stepdecay}
Under Assumptions \ref{assumpt:func} and ~\ref{assump:gradient-oracle}(a), and {assume that there exists a constant $\Delta_0>0$ such that $\mathbb{E}[f(x_1^t) - f^{\ast}] \leq \Delta_0$ for each $t \geq 1$} where $f^{\ast} = \min f(x)$, if we run Algorithm \ref{alg:sgd} with $T > 1$, $\eta_i^t \leq 1/((\rho+1) L)$, $N= (\log_{\alpha}T)/2$, $S_t = 2T/\log_{\alpha} T$, and $\delta(t)=1/\alpha^{t-1}$ for $1\leq t\leq N$, where $\alpha>1$ then
\begin{align*}
\mathbb{E}[\left\|\nabla f(\hat{x}_T)\right\|^2] & \leq \left(\frac{4\Delta_0}{\alpha m} + \frac{M^2L\sigma}{2m}\right)\frac{(\alpha-1)}{\ln\alpha}\cdot \frac{\ln T}{\sqrt{T}}.
\end{align*}
\end{theorem}
Theorem \ref{thm:band:stepdecay} establishes a near-optimal (up to $\ln T$) rate for the step-decay bandwidth scheme which matches the result achieved at its boundaries i.e., $\eta_i^t=m\delta(t)$ or $\eta_i^t=M\delta(t)$ \citep{wang2021convergence}. As the next theorem shows, this guarantee can be improved by appropriate tuning of the stage length $S_t$.
\begin{remark}(Justification of uniformly bounds on the function values)
\label{rem:func:bound}
In Theorem \ref{thm:band:stepdecay}, we require that the expectation of the function value at each outer iterate $\mathbb{E}[f(x_1^t)]$ is uniformly upper bounded. As shown by \cite{shi2020learning}, the function values at the iterates of SGD can be controlled (bounded) by the initial state provided the step-size is bounded by $1/L$. So the assumption is fair if the initial state is settled. Nevertheless, this assumption (or its stronger version that the objective function is bounded) is commonly used or implied in optimization~\citep{hazan2015beyond, Yang-2019,pmlr-v115-xu20b, xu2019stochastic} and statistic machine learning \citep{vapnik1998statistical,cortes2019relative} , and it has never been violated in our numerical experiments.
\end{remark}
\begin{theorem}\label{thm:band:stepdecay:optimal}
Under Assumptions \ref{assumpt:func} and ~\ref{assump:gradient-oracle}(a), and assume that there exists a constant $\Delta_0>0$ such that $\mathbb{E}[f(x_1^t) - f^{\ast}] \leq \Delta_0$ for each $t \geq 1$, if we run Algorithm \ref{alg:sgd} with $T > 1$, $\eta_i^t \leq 1/((\rho+1) L)$, $S_0 = \sqrt{T}$, $S_t = S_0\alpha^{(t-1)}$ and $\delta(t) = 1/\alpha^{t-1}$ where $\alpha>1$, then
\begin{align*}
\mathbb{E}[\left\|\nabla f(\hat{x}_T)\right\|^2] & \leq \frac{1+\frac{2}{\alpha-1} }{T^{\frac{3}{2}}}\left( \frac{2(\sqrt{T}+1)^2 \Delta_0}{m} + \frac{M^2L\sigma}{m} \cdot T \right).
\end{align*}
\end{theorem}
{\bf Optimal rate for step-decay step-size} The theorem shows that if the stage length $S_t$ increases exponentially, and the length of the first stage is set appropriately, then we can achieve an optimal $\mathcal{O}(1/\sqrt{T})$ rate for the bandwidth step-decay step-size in the non-convex case. If $M=m$, which means that the bandwidth scheme degenerates to the step-decay type step-size, Theorem \ref{thm:band:stepdecay:optimal} removes the logarithmic term present in the results of \cite{wang2021convergence}. To the best of our knowledge, this is the first result that demonstrates that vanilla SGD with step-decay step-sizes can achieve the optimal rate for general non-convex problems. The numerical performance of the two step-size schedules in Theorems \ref{thm:band:stepdecay} and \ref{thm:band:stepdecay:optimal} are reported in Figure~\ref{fig:sgd:stagelength}.
{\bf Benefits of Theorems 3.5 vs the references of \cite{hazan2014beyond,yuan2019stagewise} } Another commonly used step-decay scheme in theory which halves the step-size after each stage and then doubles the length of each stage (e.g., \citep{hazan2014beyond,yuan2019stagewise}). In~\cite{hazan2014beyond}, which considers strongly convex problem, the initial stage is very short, $S_1=4$, while the analysis in~\cite{yuan2019stagewise} for PL functions use an inverse relation between stage length and step-size, which means that a longer initial stage length requires a smaller stepsize.
In contrast to these references, Theorem \ref{thm:band:stepdecay:optimal} considers a step-decay with a long first stage, $S_1=\sqrt{T}$, which allows to benefit from a large constant step-size for more iterations.
\begin{remark}({\bf Guarantees for cyclical step-sizes})\label{rem:cyclical}
In \citep{loshchilov2016sgdr}, the authors decay the step-size with cosine annealing and use $\eta_{\min}^t < \eta_{\max}^t$ to control the range of the step-size. If
$m\delta(t) \leq \eta_{\min}^t, \eta_{\max}^t \leq M\delta(t)$, then our results provide convergence guarantees for their step-size. To achieve a near-optimal rate,~\citet[Theorem 4]{li2020exponential} need to use an initial step-size that is smaller than
$\mathcal{O}(1/\sqrt{T})$ which is obviously impractical.
In contrast,
we allow the cosine step-size to start from a relatively large step-size and then gradually decay (see Theorem \ref{thm:band:stepdecay}) and also improve the convergence rate to be optimal (Theorem~\ref{thm:band:stepdecay:optimal}).
A triangular cyclical step-size is proposed by \cite{smith2017cyclical} which is varied around the two boundaries that drop by a constant after a few iterations. Our analysis first provides theoretical guarantees (e.g., Theorems \ref{thm:band:stepdecay} and \ref{thm:band:stepdecay:optimal}) also for this step-size. The details are shown in Appendix~\ref{suppl:cyclical}.
\begin{figure}[h]
\centering
\includegraphics[width=0.3\textwidth,height=1.5in]{cosine.pdf}
\includegraphics[width=0.3\textwidth, height=1.5in]{triangular.pdf}
\caption{Cosine step-size (left) and triangular step-size (right)}
\label{fig:example:cylical}
\end{figure}
\end{remark}
\section{Non-asymptotic Convergence of SGDM Under Bandwidth-based Step-Size }\label{sec:4:mom}
In this section, we establish the first non-asymptotic convergence properties of SGD with momentum (SGDM) under the bandwidth-based step-size on smooth nonconvex problems.
In this scheme, the inner iterations in Step~6 of Algorithm~\ref{alg:sgd} are essentially replaced by
\begin{align}
\label{alg:mom:1} v_{i+1}^{t} & = \beta v_i^t + (1-\beta) g_i^t \\
\label{alg:mom:2} x_{i+1}^t & = x_i^t - \eta_i^t v_{i+1}^t
\end{align}
for $\beta \in (0,1)$.
We refer to Algorithm~\ref{alg:mom} in Appendix~\ref{sec:appendix:mom} for a more detailed description.
As in many studies of momentum-methods (e.g.~\cite{ghadimi15heavyball,unified_momentum,liu2020improved,mai2020convergence}), we establish an iterate relationship on the form
$
\mathbb{E}[W^{t+1}] \leq \mathbb{E}[W^t] - c_0 \eta \mathbb{E}[e^t] + c_1\eta^2$,
where $W^t$ is a Lyapunov function, $e^t$ is a performance measure (here, $e^t = \left\|\nabla f(\cdot)\right\|^2$), $\eta$ is the step-size and $c_0$ and $c_1$ are constants.
However, due to the time-dependent and possibly non-monotonic bandwidth-based step-size, we cannot use the Lyapunov functions suggested in~\cite{unified_momentum,liu2020improved} but rely on the following non-trivial construction:
\iffalse
Before presenting the convergence of SGDM, we introduce an extra variable $z_i^{t} = \frac{x_i^t}{1-\beta} - \frac{\beta}{1-\beta}x_{i-1}^t$ for each $i\geq 1, t\geq 1$ as \cite{unified_momentum,liu2020improved}. However, the bandwidth-based step-size $\eta_i^t$ is time dependent and also possibly non-monotonic, so the commonly used equality $x_{i+1}^t = x_i^t - \eta g_i^t + \beta(x_i^t - x_{i-1}^t)$ \cite{unified_momentum} or $z_{i+1}^t = z_i^t - \eta^t g_i^t$ (see lemma 3 of \cite{liu2020improved}) does not hold in our analysis. This highly increases the difficulty for the analysis.
Next we design a non-trivial function $W_i^t$ which measures the averaged progress of Algorithm \ref{alg:mom} for each iterate. The Lyapunov function built in Lemma \ref{lem:mom:wit2} are motivated by series of lemmas, one can turn to Appendix \ref{sec:appendix:mom} for more details.
\fi
\begin{lemma}\label{lem:mom:wit2}
Suppose that Assumption \ref{assumpt:func} and Assumption \ref{assump:gradient-oracle}(b) hold. Let $z_i^t=(1-\beta)^{-1}(x_{i}^t-\beta x_{i-1}^t)$ and assume that there exists a constant $\Delta_0$ such that $\mathbb{E}[f(x_i^t) - f^{\ast}] \leq \Delta_0$ for $t, i \geq 1$ and the step-size in each stage is monotonically decreasing. Define the function $W_{i+1}^t$
\begin{align*}
W_{i+1}^t = \frac{f(z_{i+1}^t) - f^{\ast}}{\eta_i^t} + \frac{r\left\| x_{i+1}^t - x_i^t \right\|^2}{\eta_i^t} + 2r [f(x_{i+1}^t)-f^{\ast}],
\end{align*}
where $r=\frac{\beta L}{2(1-\beta^2)(1-\beta)^2}$. Then, if $\eta_i^t \leq 1/L$, for any $t$ and $i \geq 2$, we have
\begin{equation}\label{lem:mom:wit:key}
\mathbb{E}[W_{i+1}^t \mid \mathcal{F}_i^t] \leq W_i^t + A_1\left(\frac{1}{\eta_i^t} - \frac{1}{\eta_{i-1}^t}\right) - \left\|\nabla f(x_i^t)\right\|^2 + \eta_i^t \cdot B_1 G^2.
\end{equation}
where $A_1= \frac{\beta \Delta_0}{1-\beta} +\Delta_z+ \frac{rG^2}{L^2}$, $B_1=r(1-\beta)(2-\beta) + \frac{L}{2(1-\beta)^2}$, and $\Delta_z=\frac{\Delta_0}{1-\beta} +\frac{\beta G^2}{2(1-\beta)^2L}$.
\end{lemma}
Note that even though the step-size is assumed to be monotonically decreasing in each stage, it may be increased between stages, leading to a globally non-monotonic step-size.
The proposed bandwidth-based step-sizes (e.g., step-decay with linear or cosine modes) in the numerical experiments and the cosine annealing policy proposed in \citep{loshchilov2016sgdr} all satisfy this condition. Note that, unlike~\citep{mai2020convergence,liu2020improved}, the momentum parameter $\beta$ does not rely on the step-size, but can be chosen freely in the interval $(0,1)$. In particular, our analysis supports the common choice of $\beta = 0.9$ used as default in many deep learning libraries~\citep{imagenet,he2016deep}. Similar to Remark \ref{rem:func:bound}, the function value of the iterates for momentum can also be controlled (bounded) by the initial state given $\eta_i^t \leq 1/L$; see~\citep{shi2021-momentum}. Therefore, we believe our assumptions are reasonable.
If we restrict the analysis to a single stage, $N=1$, the lemma allows to recover the optimal rate for SGDM under the step-size $\eta_i^t=\eta_0/\sqrt{T}$~\citep{unified_momentum, mai2020convergence, liu2020improved,defazio2020understanding} and to prove, for the first time, an optimal ${\mathcal O}(1/\sqrt{T})$ rate for SGDM under the $1/\sqrt{i}$ stepsize. These results are formalized in Appendix~\ref{suppl:mom:1sqrt}.
\iffalse
In the stream case $N=1$ includes the commonly constant step-size and polynomial step-size. We first consider the constant step-size e.g., $\eta_i^t = \eta_0/\sqrt{T}$, the result in Lemma \ref{lem:mom:wit2} can recover the optimal rate of SGDM~\citep{unified_momentum, mai2020convergence, liu2020improved,defazio2020understanding}. Applying (\ref{lem:mom:wit:key}) from $i=1$ to $T$ ($t=1$), we can achieve the first optimal $\mathcal{O}(1/\sqrt{T})$ rate at an average for SGDM under the $1/\sqrt{i}$ decaying step-size.
\fi
\subsection{Convergence of SGDM for Bandwidth Step-Decay Step-Size}
We now show the convergence complexity of SGDM with the bandwidth step-decay step-size. Here \emph{step-decay} means that the bandwidth limits are divided by a constant after some iterations.
We first consider the total number of iterations $T$ to be given, the stage length $S_t$ to be constant, and the number of stages $N$ as a hyper-parameter.
\begin{theorem}\label{thm:mom:stepdecay:0}
Assume the same setting as Lemma \ref{lem:mom:wit2}. If given the total number of iterations $T \geq 1$, $N \geq 1$, $S_t = S = T/N$, $\delta(t) = 1/\alpha^{t-1}$ for each $1\leq t \leq N$ and $\alpha >1$, then
\begin{align}\label{eqn:mom:stepdecay:0}
\mathbb{E}[\left\|\nabla f(\hat{x}_T) \right\|^2] \leq \frac{ W_1^1 \cdot N}{T\alpha^{N-1}} + \left(\alpha C_0 + C_2\right)\cdot \frac{N}{T} + \frac{\left(\Delta_z + C_1\right)}{m }\cdot\frac{N\alpha^N}{T} + M B_1 G^2 \cdot\frac{N }{\alpha^{N-1}}
\end{align}
where $C_0 = r(\frac{G^2}{L} + 2\Delta_0)$, $C_1 =A_1+\Delta_z+\frac{\Delta_0}{1-\beta} $, and $C_2 =C_0 + A_2G^2 $, $A_2 =1+\frac{\beta}{2(1-\beta)^2}$, and $W_1^1$, $A_1$, $B_1$, $\Delta_z$, and $r$ are defined in Lemma \ref{lem:mom:wit2}. Furthermore, if $N=(\log_{\alpha}T)/2$ and $S_t = 2T/\log_{\alpha} T$ for each $1\leq t \leq N$ where $\alpha > 1$, we have
\begin{align}\label{thm:mom:stepdecay}
\mathbb{E}[\left\| \nabla f(\hat{x}_T)\right\|^2] & \leq \frac{\alpha W_1^1}{2\ln \alpha}\frac{\ln T}{T^{3/2}} +\frac{(\alpha C_0+C_2)}{2\ln \alpha}\frac{\ln T}{T} + \frac{(\Delta_z+C_1)}{2m\ln\alpha }\frac{\ln T}{\sqrt{T}} + \frac{\alpha M B_1G^2}{2\ln \alpha} \frac{\ln T }{\sqrt{T}}.
\end{align}
\end{theorem}
When $N=1$, $m \leq \eta_i^t \leq M$ and the bound (\ref{eqn:mom:stepdecay:0}) reduces to $\mathbb{E}[\left\|\nabla f(\hat{x}_T) \right\|^2] \leq \mathcal{O}(\frac{1}{T} + \frac{1}{mT} + M)$. If, in particular, $m$ and $ M $ are of order $\mathcal{O}(1/\sqrt{T})$, then we can derive the optimal convergence for constant bandwidth step-sizes, comparable to the literature for constant step-sizes~\citep{unified_momentum,liu2020improved,mai2020convergence,defazio2020understanding}.
It is not easy to explicitly minimize the right-hand-side of (\ref{eqn:mom:stepdecay:0}) with respect to $N$. However, $N=(\log_{\alpha} T)/2$ attempts to balance the last two terms and appears to be a good choice in practice. The theorem (see (\ref{thm:mom:stepdecay}) establishes an $\mathcal{O}(\ln T/\sqrt{T})$ rate under step-decay bandwidth step-size. If $M=m$, which means that the step-size follows the boundary functions, we get a near-optimal (upto $\ln T$) rate for stochastic momentum with a step-decay step-size on nonconvex problems. We believe that this is the first near-optimal rate for stochastic momentum with step-decay step-size. The next result shows how an exponentially increasing stage-length allow to sharpen this guarantee even further.
\begin{theorem}\label{thm:mom:stepdecay:2}
Suppose the same setting as Lemma \ref{lem:mom:wit2}. Consider Algorithm \ref{alg:mom}, if the functions $\delta(t)=1/\alpha^{t-1}$ with $\alpha > 1$, $S_t = S_0\alpha^{t-1}$ with $S_0=\sqrt{T}$, we have
\begin{align*}
\mathbb{E}[\left\| \nabla f(\hat{x}_T)\right\|^2] \leq \mathcal{O}\left(\frac{W_1^1}{T^{3/2}} +\frac{C_0}{T} + \frac{\Delta_z}{m\sqrt{T}} + \frac{C_1}{m\sqrt{T}} + \frac{M B_1 G^2}{\sqrt{T}} \right).
\end{align*}
\end{theorem}
The stage length $S_t$ in Theorem \ref{thm:mom:stepdecay:2} increases exponentially from $S_1=\sqrt{T}$ over $N=\log_{\alpha}((\alpha-1)\sqrt{T}+1)$ stages, resulting in an $\mathcal{O}(1/\sqrt{T})$ optimal rate for SGDM under the bandwidth-based step-decay scheme. This removes the $\ln T$ term of Theorem \ref{thm:mom:stepdecay}. To the best of our knowledge, this work is the first that is able to achieve an optimal rate for stochastic momentum with step-decay step-size in a general non-convex setting.
\begin{remark}\label{rem:liu}({\bf Better convergence than \cite{liu2020improved}})
We notice that reference \cite{liu2020improved} analyzes multi-stage momentum and obtains the bound
\begin{align}
\mathbb{E}[\left\| \nabla f(\tilde{x})\right\|^2] \leq \mathcal{O}\left(\frac{f(x_1)-f^{\ast}}{N} + \frac{\sigma L\sum_{t=1}^{N} \eta^t}{N}\right). \label{eqn:liubound}
\end{align}
Here, $\tilde{x}$
is a uniformly sampled iterate (unlike our results, which favour later iterates) and $N$ is the number of stages. The result uses a time-varying momentum parameter, whose value is determined by the step-size $\eta^t$, and also assumes an inverse relationship between the step-size and stage-length, i.e. that $\eta^t S_t$ is constant. Hence, $N$ is of ${\mathcal O}(\log_{\alpha} T)$ and the convergence guarantee in (\ref{eqn:liubound}) is of ${\mathcal O}(1/\log_{\alpha} T)$, which is far worse than the rate of Theorem \ref{thm:mom:stepdecay:0} and the optimal rate of Theorem \ref{thm:mom:stepdecay:2}.
\end{remark}
\iffalse
\begin{remark}
\label{rem:liu}
We notice that reference \cite{liu2020improved} analyzes an $N$-stage momentum and obtains the following results for a uniformly sampled output $\tilde{x}$: $\mathbb{E}[\left\| \nabla f(\tilde{x})\right\|^2] \leq \mathcal{O}\left(\frac{f(x_1)-f^{\ast}}{N} + \frac{\sigma L\sum_{t=1}^{N} \eta^t}{N}\right)$. Unlike the uniformly sampling, we apply an increasing weighted sampling rule which ensures that the final iterate is more highly selected. Moreover, they require that step-size and momentum weight $\beta$ satisfies that $\frac{\beta_t\eta^t}{1-\beta_t} = R_1$ with a constant $R_1$ (this ensures that $z_i^t = x_i^t - R_1 v_i^t$), and the step-size and the stage length $\eta^t S_t = R_2$ with a constant $R_2$, but these are not necessary in our results.
In our analysis, the momentum parameter $\beta \in (1/2,1)$ is able to be freely set and the step-size is possibly non-monotonic. If we ignore the requirements of \citep{liu2020improved} and adopt the step-decay schemes of Theorems \ref{thm:mom:stepdecay} and \ref{thm:mom:stepdecay:2}, the best rates obtained from \citep{liu2020improved} on both cases are $\mathcal{O}(1/N)$. Recall that number of stages $N = \mathcal{O}(\log_{\alpha} T)$ in Theorems \ref{thm:mom:stepdecay} and \ref{thm:mom:stepdecay:2}. So the results derived from \citep{liu2020improved} are far worse than the near-optimal rate of Theorem \ref{thm:mom:stepdecay} and an optimal rate of Theorem \ref{thm:mom:stepdecay:2}.
\end{remark}
\fi
\section{Numerical Experiments}\label{sec:numerical}
\iffalse
\red{Xiaoyu: I agree, it is hard to make the notations consistent for the two different kinds of bandwidth step-sizes. Maybe we can avoid to use $t$ or $i$ in the figures, but explain it clear in the text ? We can change the names for the four modes like poly1 decay, poly2 decay, linear decay, cosine decay ? Or we also can remove ``decay" ? to make the name shorter.}
\fi
In this section, we design and evaluate several specific step-size policies that belong to the bandwidth-based family. We consider SGD with and without momentum, and compare their performances on neural network training tasks on the CIFAR10 and CIFAR100 datasets.
\subsection{Baselines and Parameter Selection for the Bandwidth Step-Sizes}
\label{num:description:stepsize}
The bandwidth framework allows for a unified and streamlined (worst-case) analysis of all step-size policies that lie in the corresponding band. Within this family, the band gives a lot of freedom in crating innovative step-size policies with additional advantages. In particular, we will design a number of step-size policies that add periodic perturbations to a baseline step-size, attempting to both avoid bad local minima and to improve the local convergence properties.
The step-decay bandwidth step-sizes divide the total number of iterations into a small number of stages, in which the boundary functions are constant. The width of the band are determined by the constants $m$ and $M$. We will explore step-sizes that add a decreasing perturbation within each stage, starting at the upper band at the beginning of the stage, ending at the lower bound at the end of the stage, and decaying as $1/i$, $1/\sqrt{i}$, linearly or according to a cosine function. As baseline, we consider the step-decay step-size that follows the lower boundary function $m\delta(t)$. To use the same maximum value for the bandwidth step-sizes, we do not add any perturbation in the first stage; cf. Figure~\ref{stepsize:band}.
For the $1/\sqrt{t}$-band, on the other hand, stages correspond to epochs
and perturbing the step-size within a stage would be too frequent and lead to bias. Rather, we choose to add similar perturbations as for the step-decay band, but adjust the perturbation between stages. In our experiments, the two step-size policies perform roughly the same number of periods of perturbations over the training set. As baseline, we consider the step-size $\eta_i^t=m/\sqrt{t}$.
In all experiments, the hyper-parameters (e.g., $m$ and $M$) have been determined using grid search, see Section~\ref{cifar:suppl} for details.
\subsection{Bandwidth Schedule Helps to Avoid Local Minima }\label{sec:toy:example}
To demonstrate the potential benefits of bandwidth-based non-monotonic step-size schedules, we consider the toy example (see Section \ref{toy:suppl} for details and further results) from~\citep{shi2020learning}, which is non-convex and has four local minima\footnote{Notation: $\circled{1}-\circled{3}$ denote the local minima at top left, top right and bottom left, respectively;
and $\circled{4}$ denotes the global minimum at bottom right.}; see Figure~\ref{fig:contour:bandwidth}. We then compare the final iterates of SGD with constant step-sizes (both large and small), step-decay, and a bandwidth-based step-decay step size which we call \emph{linear-mode} (illustrated in Figure~\ref{stepsize:band}).
As shown in Figure~\ref{fig:contour:bandwidth}, a large constant step-size more easily escapes the bad local minima to approach the global minimum at $(0.7, -0.7)$
than a small constant step-size. However, with a large constant step-size, the final iterates are scattered and end up far from the global minimum, which also has been observed in Figure 5 of \citep{shi2020learning}. Therefore, we have to reduce the step-size at some points to reduce the error. This is exactly the intuition of step-decay step-size. As shown in Figure \ref{fig:contour:bandwidth}, the scatter plots of SGD with \emph{step-decay} (red) and \emph{step-decay with linear-mode} (green) are more concentrated around the global minimum than the constant step-sizes.
To quantify the ability of different step-sizes to avoid the local minima, Table~\ref{tab:rates} reports the percentage of the final iterates under the different step-size policies that are close to each minima.
We can see that the ability of the step-decay policy (named \emph{baseline}) to escape the local minima is slightly worse than the large constant step-size, but Figure~\ref{fig:contour:bandwidth} shows that the variance of the near-optimal iterates is reduced significantly. In a similar way, we can see that \emph{linear-mode} not only improves the ability to escape the local minima, but also produces final iterates that are more concentrated around the global optimum. Hence, it appears (at least in this example) that non-monotonic step-size schedules allow SGD to escape local minima and produce final iterates of high quality.
\begin{figure}[t]
\begin{minipage}{0.6\linewidth}
\centering
\captionsetup{width=0.9\linewidth}
\includegraphics[width=0.48\textwidth]{fig_bandwidth_stepsize.jpg} \hfill
\includegraphics[width=0.48\textwidth]{plot_func_toy.jpg}
\captionof{figure}{The scatter plots (left); function value distribution around global minima (middle)}
\label{fig:contour:bandwidth}
\end{minipage} \hfill
\begin{minipage}{0.35\linewidth}
\captionof{table}{The percentage (\%) of the final iterates (10000 runs) close to each local minima}
\label{tab:rates}
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}[width=1\linewidth]{@{}ccccc@{}}
\toprule
\multirow{2}{*}{} & \multicolumn{2}{c}{constant} & \multicolumn{2}{c}{step-decay} \\ \cline{2-5}
& small\textcolor{c1}{$\,\blacksquare$} & large\textcolor{c3}{$\,\blacksquare$} & baseline\textcolor{c2}{$\,\blacksquare$} & linear\textcolor{c4}{$\,\blacksquare$} \\ \midrule
$\circled{1}$ & $\bm{29.61}$ & 0.12 & 0.40 & 0.14\\
$\circled{2}$ & 24.66 & 3.45 & 6.89 & 2.95 \\
$\circled{3}$ & 25.13 & 3.28 & 7.55 & 2.99 \\
$ \circled{4}$ & 20.60 & $\bm{93.15}$ & $\bm{85.16}$ & $\bm{93.92}$ \\
\bottomrule
\end{tabular}}
\end{minipage}
\end{figure}
\subsection{Numerical Results on CIFAR10 and CIFAR100}\label{numerical:cifar}
To illustrate the practical performance of the bandwidth-based step-sizes, we choose the well-known CIFAR10 and CIFAR100~\citep{KrizWeb} image classification datasets. We consider the benchmark experiments of CIFAR10 on ResNet-18~\citep{he2016deep} and CIFAR100 on a $28\times 10$ wide residual network (WRN-28-10)~\citep{zagoruyko2016wide}, respectively. All the experiments are repeated 5 times to eliminate the influence of randomness.
We begin by evaluating our step-sizes for SGD. The left column of Figure \ref{fig:sgd:mom}
present the results of the $1/\sqrt{t}$-band step-sizes on the two datasets.
As shown in Figure \ref{stepsize:band}, these stepsizes
are all non-monotonic. The sudden increase in the step-size leads to a corresponding cliff-like reduction in accuracy followed by a recovery phase that consistently ends up at a better performance than in the previous stage.
The three $1/\sqrt{t}$-band step-sizes achieve significant improvements compared to their baseline ($\eta_i^t=m/\sqrt{t}$), in terms of both test loss and test accuracy. Moreover, the linear-mode performs the best compared to other polynomial decaying modes.
Then, the results of SGD with step-decay band (described in Section \ref{num:description:stepsize} or see Figure~\ref{stepsize:band} in Appendix) on CIFAR10 and CIFAR100 are given in Figure \ref{fig:sgd:mom} (middle column), respectively.
At the final stage, the bandwidth step-sizes improve both test loss (see Figure \ref{fig:sgd:mom:testloss} in Appendix) and test accuracy compared to the baseline. In particular, the cosine-mode performs the best on this problem. In the second stage, baseline methods have a sharp boost. Our guess is that the noise accumulates quickly under a relatively large constant step-size. But this phenomenon is only temporary. When we drop the step-size in the third stage, the performance improves.
Next, we evaluate the performance of step-decay bandwidth step-sizes on SGDM. The results are reported in Figure \ref{fig:sgd:mom} (right column). The first observation from Figure~\ref{fig:sgd:mom} (right column) is that the step-decay bandwidth step-sizes also work well for SGDM, and that again, the cosine-mode performs better than the others. Another interesting observation is that the performance of vanilla SGD with cosine-mode (red) in Figure~\ref{fig:sgd:mom} is comparable to (even better than) SGDM with the baseline step-decay step-size (black) in Figure~\ref{fig:sgd:mom}. A similar conclusion can also be made on CIFAR100.
\begin{figure}[t]
\centering
\begin{minipage}{1\textwidth}
\centering
\includegraphics[width=0.32\textwidth,height=1.8in]{1sqrt_band_cifar10_accuracy_2.pdf}
\includegraphics[width=0.32\textwidth,height=1.8in]{sgd_stepdecay_band_cifar10_accuracy_2.pdf}
\includegraphics[width=0.32\textwidth,height=1.8in]{mom_stepdecay_band_cifar10_accuracy_2.pdf}
\end{minipage}
\\
\begin{minipage}{1\textwidth}
\centering
\includegraphics[width=0.32\textwidth,height=1.8in]{1sqrt_band_cifar100_accuracy_2.pdf}
\includegraphics[width=0.32\textwidth,height=1.8in]{sgd_stepdecay_band_cifar100_accuracy_2.pdf}
\includegraphics[width=0.32\textwidth,height=1.8in]{mom_stepdecay_band_cifar100_accuracy_2.pdf}
\end{minipage}
\caption{The test accuracy of $1/\sqrt{t}$-band (left column), Step-decay-band for SGD (middle column), step-decay band for SGDM (right column)}
\label{fig:sgd:mom}
\end{figure}
\section{Conclusion}\label{sec:conclusion}
We have studied a general family of bandwidth step-sizes for non-convex optimization. The family specifies a globally decaying band in which the actual step-size is allowed to vary, and includes both stage-wise and continuously decaying step-size policies as special cases. We have derived convergence rate guarantees for SGD and SGDM under all step-size policies in two important classes of bandwidth step-sizes ($1/\sqrt{t}$ and step-decay), some of which are optimal. Our results provide theoretical guarantees for several popular ``cyclical'' step-sizes~\citep{loshchilov2016sgdr,smith2017cyclical}, as long as they are tuned to lie within our bands. We have also designed a number of novel step-sizes that add periodic perturbations to the global trend in order to avoid bad local minima and to improve the local convergence properties. These step-sizes were shown to have superior practical performance in neural network training tasks on the CIFAR data set.
\iffalse
We establish the worst-case convergence guarantees for SGD and its momentum variant (SGDM) under the bandwidth-based step-size on smooth nonconvex problems, and some of the results are optimal. Our results provide theoretical guarantees for other popular ``cyclical step-size"~\citep{loshchilov2016sgdr,smith2017cyclical} if their boundaries lie within our bands. We proposed some monotonic step-sizes based on the $1/\sqrt{t}$-band and step-decay band, and illustrated their superior performance for training of large-scale deep neural networks. In Section \ref{sec:numerical},
\fi
In the analysis of SGDM, we assume that the stochastic gradient is bounded (see Assumption \ref{assump:gradient-oracle}(b)). It is interesting to see how to relax this assumption in some special cases, for example, when the step-size is constant throughout each stage. It would also be interesting to see if the bandwidth framework could be specialized to a more narrow class of step-sizes, for which we can provide even stronger convergence rates.
|
\section{Compression Experiments}\label{sec:EXPERIMENTS}
We validate the PBA algorithm experimentally by comparing the performance
of a PBA-derived fixed-rate compressor against the performance of baseline fixed-rate
compressors. The code of our implementation can be found at \url{https://github.com/SourbhBh/PBA}. As we noted in the previous section, although variable-rate codes are more commonplace in practice, fixed-rate codes do offer some advantages over their more
general counterparts:
\begin{enumerate}
\item In applications where a train of source
realizations are compressed sequentially, fixed-rated
coding allows for simple concatenation of the compressed
representations. Maintaining synchrony between the encoder
and decoder is simpler than with variable-rate codes.
\item In applications where a dataset of source realizations
are individually compressed, fixed-rate coding allows for random access
of data points from the compressed representation.
\item In streaming in which a sequence of realizations will
be streamed, bandwidth provisioning is
simplified when the bit-rate is constant over time.
\end{enumerate}
Fixed-rate compressors exist for specialized sources such as
speech~\cite{McCreeB95, SchroederA85} and audio more
generally~\cite{Vorbis}. We
consider a general-purpose, learned, fixed-rate compressor
derived from PBA and the following two quantization operations.
The first, $Q_{CD}(a,\sigma^2,U,x)$\footnote{``CD'' stands for
``clamped dithered.''}
accepts the hyperparameter $a$, a variance estimate $\sigma^2$, a
dither realization $U$, and the scalar source realization to be
compressed, $x$, and outputs (a binary representation of) the
nearest point to $x$ in the set
\begin{equation}
\label{eq:quantizeset}
\left\{i + U : i \in \mathbb{Z} \ \text{and} \ i + U
\in \left(-\frac{\Gamma}{2},\frac{\Gamma}{2}\right] \right\}\newest{,}
\end{equation}
where
\begin{equation}
\Gamma = 2^{\lfloor \frac{1}{2} \log_2 (4a^2 \sigma^2 + 1) \rfloor }.
\end{equation}
This evidently requires $\log_2 \Gamma$ bits. The second
function, $Q_{CD}'(a^2,\sigma^2,U,b)$, where $b$ is a binary
string of length $\log_2 \Gamma$, maps the binary
representation $b$ to the point in~(\ref{eq:quantizeset}).
These quantization routines are applied separately to each
latent component. The $\sigma^2$ parameters are
determined during training. The dither $U$ is chosen
uniformly over the set $[-1/2,1/2]$, independently for
each component. We assume that $U$ is chosen pseudorandomly
from a fixed seed that is known to both the encoder and
the decoder. As such, it does not need to be explicitly
communicated. For our experiments, we fix the $a$ parameter
at $15$ and hard code this both at the encoder and at
the decoder. We found that this choice balances the
dual goals of minimizing the excess distortion due to
the clamping quantized points to the interval $(\Gamma/2,\Gamma/2]$
and minimizing the rate.
PBA compression proceeds by applying Algorithm~\ref{alg:PBA}
to a training set to determine the matrices $\bm{W}$ and
$\bm{T}$. The variance estimates $\sigma_1^2,\ldots,\sigma_d^2$
for the $d$ latent variances are chosen as the empirical
variances on the training set and are hard-coded in the encoder
and decoder. Given a data point $\bm{x}$,
the encoded representation is the concatenation of the bit strings
$b_1, \ldots, b_d$, where
$$
b_i = Q_{CD}(a^2,\sigma^2_i,U_i,\bm{w}_i^\top x),
$$
The decoder parses the received bits into $b_1, \ldots, b_d$.
and computes the latent reconstruction $\hat\bm{y}$, where
$$
\hat\bm{y}_i = Q_{CD}'(a^2,\sigma^2_i,U_i,b_i),
$$
The reconstruction is then $\bm{T}\hat\bm{y}$.
We evaluate the PBA compressor on MNIST \cite{LecunBBH98}, CIFAR-10 \cite{Krizhevsky09}, MIT Faces Dataset, Free Spoken Digit Dataset (FSDD) \cite{Zohar} and a synthetic Gaussian dataset. The synthetic Gaussian dataset is generated from a diagonal covariance matrix obtained from the eigenvalues of the Faces Dataset.
We compare our algorithms primarily using mean-squared error since our theoretical analysis uses mean squared error as the distortion metric. Our plots display Signal-to-Noise ratios (SNRs) for ease of interpretation. For image datasets, we also compare our algorithms using the Structural Similarity (SSIM) or the Multi-scale Strctural Similarity (MS-SSIM) metrics when applicable \cite{WangBSS04}. We also consider errors on downstream tasks, specifically classification, as a distortion measure.
For all datasets, we compare the performance of the PBA compressor
against baseline scheme derived from PCA that uses $Q_{CD}$
and $Q_{CD}'$.
The PCA-based scheme sends some of the principal components
essentially losslessly, and no information about the others.
Specifically, in the context of our framework,
for any given $k$, we choose the first $k$ columns
of $\bm{W}$ to be aligned with the first $k$ principal components
of the dataset; the remaining columns are zero. Each nonzero column
is scaled such that its Euclidean length multiplied by the eigenvalue has all the significant digits. This is done so that at high rates, the quantization procedure sends the $k$ principal components losslessly. The quantization
and decoder operations are as in the PBA-based scheme; in particular
the $a^2$ parameter is as specified above. By varying $k$, we trade off rate and distortion.
\subsection{SNR Performance}
\begin{figure}[hbt]
\centering
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PCA_1.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PCA_2.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PCA_3.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PCA_4.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PCA_5.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PBA_1.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PBA_2.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PBA_3.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PBA_4.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PBA_5.png}
\endminipage\hfill
\caption{Reconstructions at different bits/pixel values for PCA (top) and PBA (bottom)}\label{fig:reconstructions_pca_pba}
\end{figure}
\begin{figure}[hbt]
\centering
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_SNR_PBAPCAJPEG.png}
\endminipage\hfill
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_SNR_PBAPCAJPEG.png}
\endminipage\hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/FACES_SNR_PBAPCAJPEG.png}
\endminipage \hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/AUDIO_SNR.png}
\endminipage \hfill
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_SNR_PBAPCAJPEG_UltraZoom.png}
\endminipage\hfill
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_SNR_PBAPCA_UltraZoom.png}
\endminipage\hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/FACES_SNR_PBAPCAJPEG_UltraZoom.png}
\endminipage \hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/AUDIO_SNR_Zoom.png}
\endminipage
\caption{SNR/pixel vs Rate (bits/pixel) for MNIST, CIFAR-10, Faces, FSDD datasets. Figures in the bottom row are zoomed-in. } \label{fig:snr}
\end{figure}
\begin{figure}[hbt]
\centering
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_NQ_SNR_PBAPCA.png}
\endminipage\hfill
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_NQ_SNR_PBAPCA.png}
\endminipage\hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/FACES_NQ_SNR_PBAPCA.png}
\endminipage
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/SYN_PBA_PCA.png}
\endminipage
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_NQ_SNR_PBAPCA_UltraZoom.png}
\endminipage\hfill
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_NQ_SNR_PBAPCA_UltraZoom.png}
\endminipage\hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/FACES_NQ_SNR_PBAPCA_UltraZoom.png}
\endminipage
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/SYN_PBA_ULTRAZOOM.png}
\endminipage
\caption{SNR/pixel vs Rate (bits/pixel) for MNIST, CIFAR-10, Faces and Synthetic dataset. Reconstructions are not rounded to integers from $0$ to $255$. The bottom four plots are zoomed-in versions of the top four plots.}
\label{fig:nq}
\end{figure}
\begin{figure}[hbt]
\centering
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_EIG_LARGEST.png}
\endminipage\hfill
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_EIG_LARGEST.png}
\endminipage\hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/FACES_EIG_LARGEST.png}
\endminipage\hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/AUDIO_EIG_LARGEST.png}
\endminipage
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_EIG_SMALLEST.png}
\endminipage\hfill
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_EIG_SMALLEST.png}
\endminipage\hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/FACES_EIG_SMALLEST.png}
\endminipage\hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/AUDIO_EIG_SMALLEST.png}
\endminipage
\caption{Eigenvalue distribution of the datasets. The top three plots are the largest 25 eigenvalues for MNIST, CIFAR-10, Faces and FSDD dataset. The bottom four figures plot the remaining eigenvalues except the largest 500. }
\label{fig:eig}
\end{figure}
\begin{figure}[hbt]
\centering
\minipage{0.32\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_COMP.png}
\endminipage\hfill
\minipage{0.32\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_COMP.png}
\endminipage\hfill
\minipage{0.32\textwidth}%
\includegraphics[width=\linewidth]{images/FACES_COMP.png}
\endminipage
\caption{Plots of number of components sent vs rate (bits/pixel) for PBA and PCA. }\label{fig:num_comp}
\end{figure}
We begin by examining compression performance under
mean squared error, or equivalently, the SNR, defined as
$$
\text{SNR} = 10 \cdot \log_{10}\left(
\frac{P}{\text{MSE}}\right).
$$
where $P$ is the empirical second moment of the dataset.
This was the objective that PBA (and PCA) is designed to minimize.
In Figure~\ref{fig:reconstructions_pca_pba}, we display reconstructions
for a particular image in the Faces Dataset under PBA and PCA.
Figure~\ref{fig:snr} shows the tradeoff for PBA and PCA against
JPEG and JPEG2000 (for the image datasets) and
AAC (for the audio dataset). All of the image datasets
have integer pixel values between 0 and 255. Accordingly, we round
the reconstuctions of PBA and PCA to the nearest integer in this range. Figure~\ref{fig:nq} shows the same tradeoff for PBA and PCA when reconstructions are not rounded off to the nearest integer.
We see that PBA consistently
outperforms PCA and JPEG, and is competitive with JPEG2000, even
though the JPEG and JPEG2000 are variable-rate. \footnote{It should be
noted, however, that JPEG and JPEG2000 aim to minimize subjective
distortion, not MSE, and they
do not allow for training on sample images, as PBA and PCA do. A
similar caveat applies to AAC.} We estimate the size of the JPEG header by compressing an empty image and subtract this estimate from all the compression sizes produced by JPEG. We do not plot JPEG2000 performance for MNIST since it requires at least a 32x32 image. For audio data, we observe that PBA consistently outperforms PCA and AAC.
Since the image data all use $8$ bits per pixel, one can
obtain infinite SNR at this rate via the trivial encoding that
communicates the raw bits. PCA and PBA do not find this solution
because they quantize in the transform domain, where the
lattice-nature of the pixel distribution is not apparent.
Determining how to leverage lattice structure in the source
distribution for purposes of compression is an interesting
question that transcends the PBA and PCA algorithms and that
we will not pursue here.
The reason that PCA performs poorly is that it
favors sending the less significant bits of the most significant
components over the most significant bits of less significant
components, when the latter are more valuable for reconstructing
the source. Arguably, it does not identify the ``principal bits.''
Figure~\ref{fig:eig} shows the eigenvalue distribution of the
different datasets, and Figure~\ref{fig:num_comp} shows the number of distinct
components about which information is sent as a function of rate
for both PBA and PCA. We see that PBA sends information about many
more components for a given rate than does PCA. We discuss the
ramifications of this for downstream tasks, such as classification,
in Section~\ref{subsec:downstream}.
\subsection{SSIM Performance}
Structural similarity (SSIM) and Multi-Scale Structural similarity (MS-SSIM) are metrics that are tuned to perceptual similarity. Given two images, the SSIM metric outputs a real value between $0$ and $1$ where a higher value indicates more similarity between the images. We evaluate the performance of our algorithms on these metrics as well in Figure~\ref{fig:ssim}. We see that PBA
consistently dominates PCA, and although it was not optimized for
this metric, beats JPEG at low rates as well.
\begin{figure}[hbt]
\centering
\minipage{0.32\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_SSIM_PBAPCAJPEG.png}
\endminipage\hfill
\minipage{0.32\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_SSIM_PBAPCAJPEG.png}
\endminipage\hfill
\minipage{0.32\textwidth}%
\includegraphics[width=\linewidth]{images/FACES_SSIM_PBAPCAJPEG.png}
\endminipage
\caption{SSIM vs Rate (bits/pixel) for MNIST, CIFAR-10, Faces Dataset}\label{fig:ssim}
\end{figure}
\subsection{Performance on Downstream tasks}
\label{subsec:downstream}
Lastly, we compare the impact of using PBA and PCA on an important downstream task, namely classification. We evaluate the algorithms on MNIST and CIFAR-10 datasets and use neural networks for classification. Our hyperparameter and architecture choices are given in Table~\ref{tb:ARCHITECTURE}. We divide the dataset into three parts. From the first part, we obain the covariance matrix that we use for PCA and to obtain the PBA compressor. The second and third part are used as training and testing data for the purpose of classification. For a fixed rate, reconstructions are passed to the neural networks for training and testing respectively. Since our goal is to compare classification accuracy across the compressors, we fix both the architecture and hyperparameters, and do not perform any additional tuning for the separate algorithms.
Figure~\ref{fig:accuracy} shows that PBA outperforms PCA in terms of accuracy. The difference is especially significant for low rates; all algorithms attain roughly the same performance at higher rates.
\begin{table}[hbt]
\centering
\begin{tabular}[t]{lccc}
\hline
Hyperparameter&MNIST&CIFAR-10\\
\hline
Architecture & 2-layer fully connected NN & Convolutional Neural Network \\ & & with 2 convolutional layers, pooling and \\ & & three fully connected layers \\
\# Hidden Neurons&100& NA\\
Optimization Algorithm&Adam& SGD with momentum\\
Loss&Cross-entropy&Cross-entropy\\
Learning Rate&0.0005&0.01\\
\hline
\end{tabular}
\caption{Hyperparameter Choices and Architecture for Classification}\label{tb:ARCHITECTURE}
\end{table}%
\begin{figure}[hbt]
\centering
\minipage{0.45\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_ACCURACY_PBAPCAJPEG.png}
\endminipage\hfill
\minipage{0.45\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_ACCURACY_PBAPCAJPEG.png}
\endminipage
\caption{Accuracy vs Rate (bits/pixel) for MNIST, CIFAR-10}\label{fig:accuracy}
\end{figure}
\section{Introduction}
\emph{Autoencoders} are an effective method for representation learning
and dimensionality reduction. Given a centered dataset $\bm{x}_1, \bm{x}_2, \ldots, \bm{x}_n
\in \mathbb{R}^d$ (i.e., $\sum_i \bm{x}_i = 0$),
an autoencoder (with \emph{latent dimension} $k \le d$)
consists of an \emph{encoder} $f:
\mathbb{R}^d \mapsto \mathbb{R}^k$ and a \emph{decoder} $g: \mathbb{R}^k \mapsto
\mathbb{R}^d$. The goal is to select $f$ and $g$ from prespecified
classes $\mathcal{C}_f$ and $\mathcal{C}_g$ respectively such that if a random point $\bm{x}$ is picked from the data set then $g(f(\bm{x}))$ is close to $\bm{x}$ in some sense, for example in mean squared error.
If $\mathcal{C}_f$ and $\mathcal{C}_g$ consist of linear mappings then the autoencoder is called a \emph{linear autoencoder}.
Autoencoders have achieved striking successes when $f$ and $g$ are
selected through training from the class of functions realized
by multilayer perceptrons of a given architecture~\cite{HintonS06}.
Yet, the canonical autoencoder formulation described above has
a notable failing, namely that for linear autoencoders,
optimal choices of $f$ and $g$ do not necessarily identify the
principal components of the dataset; they merely identify the principal
subspace~\cite{BourlardK88, BaldiH89}. That is, the components of $f(\bm{x})$
are not necessarily proportional to projections of $\bm{x}$ against the eigenvectors of the covariance matrix
\begin{equation}\label{eq:cov}
\bm{K}\stackrel{\mathrm{def}}{=} \frac{1}{n} \sum_{i = 1}^n \bm{x}_i \cdot \bm{x}_i^\top,
\end{equation}
which we assume without loss of generality is full rank.
Thus, linear autoencoders do not recover Principal Component
Analysis (PCA). The reason {for this} is that both the objective
(the distortion) and the constraint (the dimensionality
of the latents) are invariant to an invertible transformation
applied after the encoder with its inverse applied
before the decoder. It is desirable for linear autoencoders to recover PCA for two reasons. First, from a representation learning standpoint, it guarantees that the autoencoder recovers uncorrelated features. Second, since a conventional linear autoencoder has a large number of globally optimal solutions corresponding to different bases of the principal subspace, it is preferable to eliminate this indeterminism.
Autoencoders are sometimes described as ``compressing'' the
data~\cite{Santo12,BourlardK88, LiaoZWLL21, Bishop06}, even though $f$
can be invertible even when $k < d$.
We show that by embracing this {compression-}view, one can obtain
autoencoders that are {able} to recover PCA.
Specifically, we consider linear autoencoders with quantized (or,
equivalently, noisy) latent variables with a constraint on the estimated
number of bits required to transmit the quantized latents
under fixed-rate coding. We call this problem
\emph{Principal Bit Analysis (PBA).}
The constraint turns out to be a strictly Schur-concave
function of the set of variances of the latent variables (see
the supplementary for a review of Schur-concavity). Although
finding the optimal $f$ and $g$ for this loss function is
a nonconvex optimization problem, we
show that for any strictly Schur-concave loss function,
an optimal $f$ must send projections of the data along the
principal components, assuming that the empirical covariance
matrix of the data has only simple eigenvalues. That is,
imposing a strictly Schur-concave loss in place of a
simple dimensionality constraint suffices to ensure recovery of PCA.
The idea is that the strict concavity of the loss function
eliminates the rotational invariance described above.
As we show, even a slight amount of ``curvature'' in the
constraint forces the autoencoder to spread the variances
of the latents out as much as possible, resulting in recovery
of PCA. If the loss function is merely Schur-concave, then
projecting along the principal components is optimal, but
not necessarily uniquely so.
Using this theorem, we can efficiently solve PBA.
We validate the solution experimentally by using it to construct
a fixed-rate compression algorithm for arbitrary vector-valued
data sources. We find that the PBA-derived compressor beats
existing linear, fixed-rate compressors both in terms of
mean squared error, for which it is optimized, and in terms
of the structural similarity index measure (SSIM) and downstream
classification accuracy, for which it is not.
A number of variable-rate multimedia compressors have recently
been proposed that are either related to, or directly
inspired by, autoencoders \cite{TschannenAL18, TodericiVJHMSC17,
BalleLS16, TodericiOHVMBCS16, TheisSCH17, RippelB17, HabibianRTC19,
AgustssonMTCTBG17, BalleMSHJ18, ZhouCGSW18, AgustssonTMTG19, BalleCMSJAHT20}.
As a second application of our result,
we show that for Gaussian sources, a linear form of such a
compressor is guaranteed to recover PCA. Thus we show that
ideas from compression can be fruitfully fed back into the
original autoencoder problem.
The contributions of the paper are
\begin{itemize}
\item We propose a novel linear autoencoder formulation in which
the constraint is Schur-concave. We show that this generalizes
conventional linear autoencoding.
\item If the constraint is strictly Schur-concave and the
covariance matrix of the data has only simple
eigenvalues, then we show that the autoencoder
provably recovers PCA, providing a new remedy
for a known limitation of linear autoencoders.
\item We use the new linear autoencoder formulation to efficiently
solve a fixed-rate compression problem that we
call \emph{Principal Bit Analysis (PBA).}
\item We demonstate experimentally that PBA outperforms
existing fixed-rate compressors on a variety
of data sets and metrics.
\item We show that a linear, variable-rate compressor that
is representative of many autoencoder-based
compressors in the literature effectively has
a strictly Schur-concave loss, and therefore it
recovers PCA.
\end{itemize}
\section{Notation}
\section{Linear Autoencoding with a Schur-Concave Constraint}\label{sec:FRAMEWORK}
Throughout this paper we consider $\mathcal{C}_f$ and
$\mathcal{C}_g$ to be the class of linear functions. The
functions $f\in\mathcal{C}_f$ and $g \in\mathcal{C}_g$ can then be represented by $d$-by-$d$ matrices,
respectively, which we denote by
$\bm{W}$ and $\bm{T}$, respectively. Thus we have
\begin{align}
f(\bm{x}) & = \bm{W}^\top \bm{x} \\
g(\bm{x}) & = \bm{T} \bm{x}.
\end{align}
We wish to design $\bm{W}$ and $\bm{T}$ to
minimize the mean squared error when the latent variables
$\bm{W}^\top \bm{x}$ are quantized, subject to a constraint
on the number of bits needed to represent the quantized latents.
We accomplish this via two modifications of the canonical autoencoder.
First, we perturb the $d$ latent variables with zero-mean
additive noise with covariance matrix $\sigma^2 \bm{I}$, which we denote by
$\bm{\eps}.$ Thus the input to the decoder is
\begin{equation}
\bm{W}^\top \bm{x} + \bm{\eps}
\end{equation}
and our objective is to minimize the mean squared error
\begin{equation}
\label{eq:1stmse}
\frac{1}{n} \sum\limits_{i=1}^{n} \mathbb{E}_{\bm{\eps}}\left[ \left \lVert \bm{x}_i - \bm{T}\left( \bm{W}^\top\bm{x}_i+ \bm{\eps} \right) \right \rVert_2^2 \right].
\end{equation}
This is equivalent to quantizing the latents, in the following sense~\cite{ZamirF92}.
Let $Q(\cdot)$ be the function that maps any real number to its nearest
integer and $\varepsilon$ be a random variable uniformly distributed over $[-1/2,1/2]$. Then for $X$ independent of $\varepsilon$, the quantities $Q(X + \varepsilon) - \varepsilon$
and $X + \varepsilon$ have the same joint distribution with $X$.
Thus (\ref{eq:1stmse}) is exactly the mean squared error if the latents
are quantized to the nearest integer and $\sigma^2 = \frac{1}{12}$,
assuming that the quantization is dithered. The overall system is depicted
in Fig.~\ref{fig:blockdiag}.
\tikzstyle{block} = [rectangle, draw, text width=5em, text centered, rounded corners, minimum height=4em]
\tikzstyle{line} = [draw, -latex']
\tikzstyle{input} = [coordinate]
\begin{figure}[!htb]
\begin{center}
\begin{tikzpicture}[node distance=1cm, auto]
\node [input, name=input] {};
\node [block,right = 1.5cm of input] (A) {Linear Encoder ($\bm{W}$)};
\node [block, right=1.5cm of A] (B) {Quantizer};
\node [block, right=2cm of B] (C) {Linear Decoder ($\bm{T}$)};
\node [text width = 1.5cm, right = 0.5cm of C] (D) {$\bm{T}\left(\bm{W}^\top \bm{x}_i + \bm{\varepsilon} \right)$};
\draw [draw,->] (input) -- node {$\bm{x}_i$} (A);
\path [line] (A) -- node [text width=2cm,midway,above,align=center ] {$\bm{W}^\top \bm{x}_i$} (B);
\path [line] (B) -- node [text width=2cm,midway,above,align=center ] {$\bm{W}^\top \bm{x}_i + \bm{\varepsilon}$} (C);
\path [line] (C) -- (D);
\end{tikzpicture}
\end{center}
\caption{Compression Block Diagram}\label{fig:blockdiag}
\end{figure}
We wish to constrain the number of bits needed to describe
the latent variables. We assume that the $j$th quantized latent
is clipped to the interval
$$\left(-\frac{\sqrt{ (2a)^2 \bm{w}_j^\top \bm{K} \bm{w}_j + 1}}{2},
\frac{\sqrt{(2a)^2 \bm{w}_j^\top \bm{K} \bm{w}_j + 1}}{2}\right],$$ where
$a > 0$ is a hyperparameter and the covariance matrix $\bm{K}$ is as defined in \eqref{eq:cov}. The idea is that for sufficiently
large $a$, the interval $$\left(-a\sqrt{ \bm{w}_j^\top \bm{K}
\bm{w}_j}, a\sqrt{ \bm{w}_j^\top \bm{K} \bm{w}_j}\right]$$
contains the latent with high probability, and adding $1$
accounts for the expansion due to the dither. The number of
bits needed for the $j$th latent is then
\begin{align}
\log \left(\sqrt{ 4a^2 \bm{w}_j^\top \bm{K} \bm{w}_j + 1}\right)
= \frac{1}{2} \log \left(4a^2 \bm{w}_j^\top \bm{K} \bm{w}_j + 1\right).
\end{align}
We arrive at our optimization problem:
\begin{equation}\label{eq:MASTEROPTIPBA}
\begin{aligned}
\inf\limits_{\bm{W},\bm{T}} \quad & \frac{1}{n} \sum\limits_{i=1}^{n} \mathbb{E}_{\bm{\eps}}\left[ \left \lVert \bm{x}_i - \bm{T}\left( \bm{W}^\top\bm{x}_i+ \bm{\eps} \right) \right \rVert_2^2 \right] \\
\text{subject to} \quad & R \geq \sum_{j = 1}^d \frac{1}{2} \log \left(4a^2 \bm{w}_i^\top \bm{K} \bm{w}_i + 1\right).
\end{aligned}
\end{equation}
Note that the function
$$\{\bm{w}_j^\top \bm{K} \bm{w}_j\}_{j = 1}^d \mapsto \sum_{j = 1}^d \frac{1}{2} \log \left(4a^2 \bm{w}_i^\top \bm{K} \bm{w}_i + 1\right)$$
is strictly Schur-concave (see Appendix~\ref{app:schur} for a brief review
of Schur-concavity). Our first result only requires that the constraint
is Schur-concave in the set of latent variances, so we will consider
the more general problem
\begin{equation}\label{eq:MASTEROPTI}
\begin{aligned}
\inf\limits_{\bm{W},\bm{T}} \quad & \frac{1}{n} \sum\limits_{i=1}^{n} \mathbb{E}_{\bm{\eps}}\left[ \left \lVert \bm{x}_i - \bm{T}\left( \bm{W}^\top\bm{x}_i+ \bm{\eps} \right) \right \rVert_2^2 \right] \\
\text{subject to} \quad & R \geq \rho\left(\left\{\bm{w}_j^\top \bm{K} \bm{w}_j\right\}_{j = 1}^d\right)
\end{aligned}
\end{equation}
where $\rho(\cdot)$ is any Schur-concave function.
Expressing the objective in~\eqref{eq:MASTEROPTI} in terms of $\bm{K}$, the optimization problem reduces to
\begin{equation}\label{eq:MASTEROPTI_LIN}
\begin{aligned}
\inf \limits_{\bm{W}, \bm{T}} \quad & \text{tr}\left( \bm{K} \right) - 2\text{tr}\left(\bm{K} \bm{W} \bm{T}^\top \right) + \text{tr}\left( \bm{T}\left( \bm{W}^\top \bm{K} \bm{W} + \sigma^2\bm{I} \right) \bm{T}^\top \right) \\
\text{subject to} \quad & R \geq \rho\left(\left\{\bm{w}_j^\top \bm{K} \bm{w}_j\right\}_{j = 1}^d\right).
\end{aligned}
\end{equation}
Since $\bm{T}$ does not appear in the rate constraint, the optimal $\bm{T}$ can be viewed as the Linear Least Squares Estimate (LLSE) of a random $\bm{x}$ given $\bm{W}^\top \bm{x} + \bm{\eps}$. Therefore, the optimal decoder, $\bm{T}^*$ for a given encoder $\bm{W}$ is (e.g. \cite{Kay98}):
\begin{equation}
\bm{T}^* = \bm{K} \bm{W} (\bm{W}^{\top} \bm{K} \bm{W} + \sigma^2\bm{I})^{-1}.
\end{equation}
Substituting for $\bm{T}$ in \eqref{eq:MASTEROPTI_LIN} yields an optimization problem over only $\bm{W}$
\begin{equation} \label{eq:MASTEROPTI_MMSE}
\begin{aligned}
\inf\limits_{\bm{W}} \quad & \text{tr}(\bm{K}) - \text{tr}(\bm{K} \bm{W} (\bm{W}^{\top} \bm{K} \bm{W} + \sigma^2\bm{I} )^{-1}\bm{W}^{\top} \bm{K} ) \\
\text{subject to} \quad & R \geq \rho\left(\left\{\bm{w}_j^\top \bm{K} \bm{w}_j\right\}_{j = 1}^d\right).
\end{aligned}
\end{equation}
This problem is nonconvex in general.
In the following subsection, we prove a structural result about
the problem for a Schur-concave $\rho$. Namely, we show that
the nonzero rows of $\bm{W}$ must be eigenvectors of $\bm{K}$.
In Section~\ref{sec:PBAPCA}, we solve the problem for the specific
choice of $\rho$ in (\ref{eq:MASTEROPTIPBA}). We also show how this
generalizes conventional linear autoencoders.
\subsection{Optimal Autoencoding with a Schur-Concave Constraint}\label{subsec:PROOF}
The following is the main theoretical result of the paper.
\begin{theorem}\label{thm:diagisoptimal}
For Schur-concave $\rho:\mathbb{R}_{\geq 0}^{d} \rightarrow \mathbb{R}_{\geq 0}$ and $R>0$, the set of matrices whose nonzero columns are eigenvectors of the covariance matrix $\bm{K}$ is optimal for \eqref{eq:MASTEROPTI_MMSE}. If $\rho$ is strictly Schur-concave and $\bm{K}$ contains distinct eigenvalues, this set contains all optimal solutions of \eqref{eq:MASTEROPTI_MMSE}.
\end{theorem}
\begin{proof}
Let the eigenvalues of $\bm{K}$ be $\{\sigma_i^2\}_{i=1}^{d}$ with $\sigma_1^2 \ge \sigma_2^2 \ge \ldots \ge \sigma^2_d$. Let the eigendecomposition of $\bm{K}$ be given by $\bm{K} = \bm{U} \bm{\Sigma} \bm{U}^{\top}$ where $\bm{U}$ is an orthogonal matrix whose columns are the eigenvectors of $\bm{K}$ and $\bm{\Sigma}$ is a diagonal matrix with entries $\left \lbrace \sigma_i^2 \right \rbrace_{i=1}^{d}$.
We first prove that the optimal value of \eqref{eq:MASTEROPTI_MMSE} can be achieved by \newest{a} $\bm{W}$ such that $\bm{W}^{\top} \bm{K} \bm{W}$ is a diagonal matrix. Let $\widetilde{\bm{W}} = \bm{W} \bm{Q}$ where $\bm{Q}$ is the orthogonal matrix obtained from the eigendecomposition of $\bm{W}^{\top} \bm{K} \bm{W}$ i.e., $$\bm{W}^{\top}\bm{K} \bm{W} = \bm{Q}\bm{\Lambda} \bm{Q}^{\top},$$ where $\bm{\Lambda}$ is a diagonal matrix formed from the eigenvalues of $ \bm{W}^{\top}\bm{K} \bm{W}$. Note that
\begin{align*}
\text{tr} \left( \bm{K} \widetilde{\bm{W}} \left( \widetilde{\bm{W}}^{\top} \bm{K} \widetilde{\bm{W}} + \sigma^2\bm{I} \right)^{-1} \widetilde{\bm{W}}^{\top} \bm{K}\right) &= \text{tr} \left( \bm{K} \bm{W} \bm{Q} \left( \bm{\Lambda} + \sigma^2\bm{I} \right)^{-1} \bm{Q}^{\top} \bm{W}^{\top} \bm{K} \right) \\
&= \text{tr} \left( \bm{K} \bm{W} \left( \bm{Q} \bm{\Lambda} \bm{Q}^{\top} + \sigma^2\bm{Q} \bm{Q}^{\top} \right)^{-1} \bm{W}^{\top} \bm{K} \right).
\end{align*}
Since $\bm{Q} \bm{\Lambda} \bm{Q}^{\top} = \bm{W}^{\top} \bm{K} \bm{W}$ and $\bm{Q} \bm{Q}^{\top}= \bm{I}$, the objective remains the same.
We now show that the constraint is only improved. Denoting the eigenvalues of $\bm{W}^{\top} \bm{K} \bm{W}$ by $\left \lbrace \nu_j \right \rbrace_{j=1}^d$, we have
\begin{equation*}
\rho\left( \left \lbrace \widetilde{\bm{w}}_j^{\top} \bm{K} \widetilde{\bm{w}}_j \right \rbrace_{j=1}^{d} \right)
= \rho\left( \left \lbrace \bm{q}_j^{\top} \bm{W}^{\top} \bm{K} \bm{W} \bm{q}_j \right \rbrace_{j=1}^{d} \right) = \rho \left( \left \lbrace \nu_j \right \rbrace_{j=1}^{d} \right).
\end{equation*}
Now since the eigenvalues of a Hermitian matrix majorize its diagonal elements by the Schur-Horn theorem \cite[Theorem~4.3.45]{HornJ85},
$$ \left \lbrace \bm{w}_j^{\top} \bm{K} \bm{w}_j \right \rbrace_{j=1}^{d} \prec \left \lbrace \nu_j \right \rbrace_{j=1}^{d}.$$
Since $\rho$ is Schur-concave, this implies
$$ \rho\left( \left \lbrace \bm{w}_j^{\top} \bm{K} \bm{w}_j \right \rbrace_{j=1}^{d} \right) \geq \rho \left( \left \lbrace \nu_j \right \rbrace_{j=1}^{d} \right) = \rho \left( \left \lbrace \widetilde{\bm{w}}_j^{\top} \bm{K} \widetilde{\bm{w}}_i \right \rbrace_{j=1}^{d} \right). $$
Therefore, if $\rho$ is Schur-concave, the rate constraint can only improve. This implies an optimal solution can be attained when
$\bm{W}$ is such that $\bm{W}^{\top} \bm{K} \bm{W}$ is diagonal. If $\rho$ is strictly Schur-concave, the rate constraint strictly improves implying that the optimal $\bm{W}$ must be such that $\bm{W}^{\top} \bm{K} \bm{W}$ is diagonal.
This implies that
\begin{align*}
\text{tr}\left( \bm{K} \bm{W} \left( \bm{W}^{\top} \bm{K} \bm{W} + \sigma^2\bm{I} \right)^{-1} \bm{W}^{\top} \bm{K} \right) &= \text{tr}\left( \bm{W}^{\top} \bm{K}^2 \bm{W} \left( \bm{W}^{\top} \bm{K} \bm{W} + \sigma^2\bm{I} \right)^{-1}\right) \\
&= \sum\limits_{i=1}^{d} \frac{\bm{w}_i^{\top} \bm{K}^2 \bm{w}_i}{\sigma^2 + \bm{w}_i^{\top} \bm{K} \bm{w}_i}.
\end{align*}
Note that minimizing the objective in \eqref{eq:MASTEROPTI_MMSE} is equivalent to maximizing the above expression. Perform the change of variable
\begin{align*}
\bm{w}_j & \mapsto \begin{cases}
\left(\frac{\bm{K}^{1/2} \bm{w}_j}{||\bm{K}^{1/2} \bm{w}_j||}, ||\bm{K}^{1/2} \bm{w}_j||^2\right) & \text{if $\bm{K}^{1/2} \bm{w}_j \ne \bm{0}$} \\
(\bm{0},0) & \text{if $\bm{K}^{1/2} \bm{w}_j = \bm{0}$}
\end{cases} \\
& = (\bm{y}_j,y_j).
\end{align*}
The assumption that $\bm{W}^\top \bm{K} \bm{W}$ is diagonal and
the normalization in the definition of $\bm{y}_j$ implies that
\[
\bm{Y} = [ \bm{y}_1 \bm{y}_2, \cdots, \bm{y}_d]
\]
is a matrix whose nonzero columns form an orthonormal set. Rewriting the objective in terms of
the $(\bm{y}_j,y_j)$, we have
\begin{equation}\label{eq:objinY}
\sum\limits_{i=1}^{d} \frac{\bm{w}_i^{\top} \bm{K}^2 \bm{w}_i}{\sigma^2 + \bm{w}_i^{\top} \bm{K} \bm{w}_i} = \sum\limits_{i=1}^{d} \bm{y}_i^{\top} \bm{K} \bm{y}_i \frac{y_i}{\sigma^2 + y_i} = \sum\limits_{i=1}^{d} \bm{y}_i^{\top} \bm{K} \bm{y}_i m_i,
\end{equation}
where $m_i = \frac{y_i}{\sigma^2+y_i}$.
Observe that under this new parametrization, the constraint only depends on $\{y_i \}_{i=1}^{d}$. Without loss of generality, we assume that $y_1 \geq y_2 \geq \cdots \ge y_d$, implying that $m_1 \geq m_2 \geq \cdots \geq m_d$.
We now prove that for given $\{y_i \}_{i=1}^{d}$, choosing the $\bm{y}_i$ along the eigenvectors of $\bm{K}$ is optimal.
Denote the diagonal elements of $\bm{Y}^\top \bm{K} \bm{Y}$ by $\{\lambda_i^2\}_{i = 1}^d$ and let $\{ \lambda_{i, \downarrow}^2 \}_{i=1}^{d}$ denote the same diagonal elements arranged in descending order. Denote the eigenvalues of $\bm{Y}^{\top} \bm{K} \bm{Y}$ by $\{ \mu_i^2 \}_{i=1}^{d}$ where $\mu_1 \geq \mu_2 \geq \cdots \geq \mu_d$.
Again invoking the Schur-Horn theorem, the eigenvalues of $\bm{Y}^{\top} \bm{K} \bm{Y}$
majorize its diagonal entries
\begin{equation}\label{eq:majorization}
\left \lbrace \lambda_i^2 \right \rbrace_{i=1}^{d} \prec \left \lbrace \mu_i^2 \right \rbrace_{i=1}^{d}.
\end{equation}
Substituting $\lambda_i^2 = \bm{y}_i^{\top} \bm{K} \bm{y}_i$ in \eqref{eq:objinY}, we have
\begin{align*}
\sum\limits_{i=1}^{d} \lambda_i^2 m_i \stackrel{(a)}{\leq} \sum\limits_{i=1}^{d} \lambda_{i,\downarrow}^2 m_i &= \lambda_{1,\downarrow}^2 m_1 + \sum\limits_{i=2}^{d} \left( \sum\limits_{j=1}^{i} \lambda_{j,\downarrow}^2 -\sum\limits_{j=1}^{i-1} \lambda_{j,\downarrow}^2 \right) m_i \\
&= \lambda_{1,\downarrow}^2 m_1 + \sum\limits_{i=2}^{d} m_i \sum\limits_{j=1}^{i} \lambda_{j,\downarrow}^2 - \sum\limits_{i=2}^{d} m_i \sum\limits_{j=1}^{i-1} \lambda_{j,\downarrow}^2 \\
&= \lambda_{1,\downarrow}^2 (m_1 - m_2) + m_d \left( \sum\limits_{j=1}^{d} \lambda_{j,\downarrow}^2 \right) + \sum\limits_{i=2}^{d-1} \left( m_i - m_{i+1}\right) \sum\limits_{j=1}^{i} \lambda_{j,\downarrow}^2 \\
&\stackrel{(b)}{\leq} \mu_1^2 (m_1 - m_2) + m_d \left( \sum\limits_{j=1}^{d} \mu_j^2 \right) + \sum\limits_{i=2}^{d-1} (m_i - m_{i+1}) \sum\limits_{j=1}^{i} \mu_j^2\\
&\stackrel{(c)}{\leq} \sigma_1^2 (m_1 - m_2) + m_d \left( \sum\limits_{j=1}^{d} \sigma_j^2 \right) + \sum\limits_{i=2}^{d-1} (m_i - m_{i+1}) \sum\limits_{j=1}^{i} \sigma_j^2\\
&= \sum\limits_{i=1}^{d} \sigma_i^2 m_i,
\end{align*}
where inequality $(a)$ follows from the assumption that $m_1 \geq m_2 \geq \cdots \geq m_d$, and $(b)$ from the definition in \eqref{eq:majorization}. Since $\bm{Y} $'s nonzero columns form an orthonormal set, the eigenvalues of $\bm{Y}^{\top} \bm{K} \bm{Y}$, when arranged in descending order, are at most the eigenvalues of $\bm{K}$ from Corollary 4.3.37 in \cite{HornJ85}, and therefore $(c)$ follows.
This upper bound is attained when $\bm{y}_i = \bm{u}_i$ for nonzero $y_i$, where $\bm{u}_i$ is the normalized eigenvector of $\bm{K}$ corresponding to eigenvalue $\sigma_i^2$. To see this, note that when $\bm{y}_i = \bm{u}_i$, $\lambda_i^2 = \mu_i^2 = \sigma_i^2$. From the definition of $\bm{y}_i, \bm{w}_i = \bm{K}^{-1/2}\bm{u}_i \sqrt{y_i} = \bm{u}_i \frac{\sqrt{y_i}}{\sigma_i}$. Therefore, for a Schur-concave $\rho$, the set of matrices whose nonzero columns are eigenvectors of $\bm{K}$ is optimal. We now prove that for a strictly Schur-concave $\rho$, if $\bm{K}$ has distinct eigenvalues, this set contains all of the optimal solutions $\bm{W}$.
We know that for a fixed $y_1 \geq y_2 \geq \cdots \geq y_d$, (implying a fixed $m_1 \geq m_2 \geq \cdots \geq m_d$) the upper bound $\sum\limits_{i=1}^{d} \sigma_i^2 m_i$ is attained by the previous choice of $\bm{y}_i$. Note that if all nonzero $m_i$ are distinct, equality in $(b)$ and $(c)$ is attained if and only if the nonzero diagonal elements of $\bm{Y}^{\top} \bm{K} \bm{Y}$ equal the corresponding eigenvalues of $\bm{K}$. This implies that, if all nonzero $m_i$ are distinct, the upper bound is attained if and only if $\bm{y}_i = \bm{u}_i$ for nonzero $y_i$. Therefore, it is sufficient to prove that for the following optimization problem
\begin{equation} \label{eq:ubschur}
\begin{aligned}
\sup\limits_{\{y_i \geq 0\}} \quad & \sum\limits_{i=1}^{d} \sigma_i^2 \frac{y_i}{\sigma^2 + y_i} \\
\text{subject to} \quad & R \geq \rho\left(\left\{y_i \right\}_{i= 1}^d\right),
\end{aligned}
\end{equation}
any optimal $\{y_i\}$ must be such that the nonzero $y_i$ are distinct. Firstly, note that since $\sigma_1^2 > \sigma_2^2 > \cdots > \sigma_d^2$, we must have $y_1 \geq y_2 \geq \cdots \geq y_d$. Assume to the contrary that for an optimal $\{ y_i \}_{i=1}^{d}$ there exists $1 \leq j, \ell < d$ such that $y_{j-1} > y_j = y_{j+1} = y_{j+2} = \cdots = y_{j+\ell} > y_{j+\ell+1} \geq 0$, where $y_0$ is chosen to be any real number strictly greater than $y_1$ and $y_{d+1}=0$.
Take $\delta > 0$ small.
Denote a new sequence $ \{ y'_i \}_{i=1}^{d}$ where $y'_{j} = y_j+\delta, y'_{j+\ell} = y_{j+\ell} - \delta$ and $y'_i = y_i$ for $1\leq i \leq d$ with $i \neq j$ and $j+\ell$.
Since $\rho$ is strictly Schur-concave, the constraint is strictly
improved,
\[ \rho\left( \{y'_i \}_{i=1}^{d} \right) < \rho\left( \{y_i \}_{i=1}^{d} \right). \]
Since $\sigma^2_j > \sigma^2_{j + \ell}$, the objective is strictly
improved for sufficiently small $\delta$,
$$
\sum\limits_{i=1}^{d} \sigma_i^2 \frac{y_i}{\sigma^2 + y_i} <
\sum\limits_{i=1}^{d} \sigma_i^2 \frac{y'_i}{\sigma^2 + y'_i},
$$
as desired.
\begin{comment}
the optimal nonzero $m_i$ are distinct. Firstly, note that since $\sigma_1^2 \geq \sigma_2^2 \geq \cdots \sigma_d^2$, $m_1 \geq m_2 \geq \cdots m_d$. Assume to the contrary that for an optimal $\{ m_i \}_{i=1}^{d}$ there exists $1 \leq j, \ell < d$ such that $m_{j-1} > m_j = m_{j+1} = m_{j+2} \cdots m_{j+\ell} > m_{j+\ell+1} \geq 0$, where $m_0$ is chosen to be any real number strictly greater than $m_1$ and $m_{d+1}=0$. Define $\delta \triangleq \min\left( \frac{m_{j-1} - m_j}{3}, \frac{m_{j+\ell}-m_{j+\ell+1}}{3} \right) > 0$. Denote a new sequence $ \{ m'_i \}_{i=1}^{d}$ where $m'_{j} = m_j+\delta, m'_{j+\ell} = m_{j+\ell} - \delta$ and $m'_i = m_i$ for $1\leq i \leq d$, $i \neq j,j+\ell$. Note that
\[ \{m'_i \}_{i=1}^{d} \succ \{m_i \}_{i=1}^{d}. \]
Since $\rho$ is strictly Schur-concave,
\[ \rho\left( \{m'_i \}_{i=1}^{d} \right) < \rho\left( \{m_i \}_{i=1}^{d} \right). \]
Also $\sigma_j^2 m'_j + \sigma_{j+\ell}^2 m'_{j+\ell} = \sigma_j^2 m_j + \sigma_{j+\ell}^2 m_{j+\ell} + \delta\left(\sigma_j^2 - \sigma^2_{j+\ell}\right) \geq \sigma_j^2 m_j + \sigma_{j+\ell}^2 m_{j+\ell}. $ Therefore, the objective only increases while the rate constraint strictly improves, contradicting the optimality assumption. This implies that the optimal nonzero $m_i$ are distinct, further implying that for a strictly Schur-concave $\rho$, the optimal value is attained only when a nonzero $\bm{y}_i = \frac{\bm{K}^{1/2} \bm{w}_i}{\left \lVert \bm{K}^{1/2} \bm{w}_i \right \rVert} = \bm{u}_i$. Therefore, $\bm{w}_i = \bm{K}^{-1/2}\bm{u}_i \sqrt{y_i} = \bm{u}_i \frac{\sqrt{y_i}}{\sigma_i}$.
\end{comment}
\end{proof}
As a consequence of Theorem~\ref{thm:diagisoptimal}, encoding via an optimal $\bm{W}$ can be viewed as a projection along the eigenvectors of $\bm{K}$, followed by different scalings applied to each component, i.e. $\bm{W} = \bm{U} \bm{S}$ where $\bm{S}$ is a diagonal matrix with entries $s_i \geq 0$ and
$\bm{U}$ is the normalized eigenvector matrix. Only $\bm{S}$ remains to be determined, and to this end, we may assume that
$\bm{K}$ is diagonal with nonincreasing diagonal entries, implying $\bm{U} = \bm{I}$. In subsequent sections, our choice of $\rho$ will be of the form $\sum\limits_{i=1}^{d} \rho_{sl}$, where $\rho_{sl}: \mathbb{R}_{\geq 0} \rightarrow \mathbb{R}_{\geq 0}$\footnote{``sl" stands for single-letter} is (strictly) concave, making $\rho$ (strictly) Schur-concave (see Proposition~\ref{prop:single-letter} in Appendix~\ref{app:schur}). Therefore, \eqref{eq:MASTEROPTI_MMSE} reduces to
\begin{equation} \label{eq:MASTEROPTI_S}
\begin{aligned}
\inf\limits_{\bm{S}} \quad & \text{tr}(\bm{K}) - \text{tr}(\bm{K} \bm{S} (\bm{S}^{\top} \bm{K} \bm{S} + \sigma^2\bm{I} )^{-1}\bm{S}^{\top} \bm{K} ) \\
\text{subject to} \quad & R \geq \rho_{sl}\left( \{s_i^2 \sigma^2_i \} \right),
\end{aligned}
\end{equation}
where the infimum is over diagonal matrices $\bm{S}$. To handle situations for which
\begin{equation}
\lim_{s \rightarrow \infty} \rho_{sl}(s) < \infty,
\end{equation}
we allow the diagonal entries of $\bm{S}$ to be $\infty$, with the objective for such cases defined via its continuous extension.
In the next section, we will solve \eqref{eq:MASTEROPTI_S} for several specific choices of $\rho_{sl}$.
\subsection{Organization}
\textbf{Notation.} We denote matrices by bold capital letters e.g. $\bm{M}$, and vectors by bold small, e.g. $\bm{v}$. The $j^{\text{th}}$ column of a matrix $\bm{M}$ is denoted by $\bm{m}_j$ and the $j^{\text{th}}$ entry of a vector $\bm{v}$ by $\left[\bm{v}\right]_j$. We denote the set $\left \lbrace 1,2, \cdots d \right \rbrace$ by $\left[ d \right]$. A sequence $a_1, a_2, \cdots a_n$ is denoted by $\{a_i\}_{i=1}^{n}$. We denote the zero column by $\bm{0}$. Logarithms without specified bases denote natural logarithms.
\textbf{Organization.} The balance of the paper is organized as follows. We describe our constrained linear autoencoder framework in Section~\ref{sec:FRAMEWORK}. This results in an optimization problem that we solve for any Schur-concave constraint in Section~\ref{subsec:PROOF}. In Section~\ref{sec:PBAPCA}, we recover linear autoencoders and PBA under our framework. We apply the PBA solution to a problem in variable-rate compression of Gaussian sources in Section~\ref{sec:VARIABLE}. Section~\ref{sec:EXPERIMENTS} contains experiments comparing the performance of the PBA-based fixed-rate compressor against existing fixed-rate linear compressors on image and audio datasets.
\section{Conclusion}
\section{Introduction}
\emph{Autoencoders} are an effective method for representation learning
and dimensionality reduction. Given a centered dataset $\bm{x}_1, \bm{x}_2, \ldots, \bm{x}_n
\in \mathbb{R}^d$ (i.e., $\sum_i \bm{x}_i = 0$),
an autoencoder (with \emph{latent dimension} $k \le d$)
consists of an \emph{encoder} $f:
\mathbb{R}^d \mapsto \mathbb{R}^k$ and a \emph{decoder} $g: \mathbb{R}^k \mapsto
\mathbb{R}^d$. The goal is to select $f$ and $g$ from prespecified
classes $\mathcal{C}_f$ and $\mathcal{C}_g$ respectively such that if a random point $\bm{x}$ is picked from the data set then $g(f(\bm{x}))$ is close to $\bm{x}$ in some sense, for example in mean squared error.
If $\mathcal{C}_f$ and $\mathcal{C}_g$ consist of linear mappings then the autoencoder is called a \emph{linear autoencoder}.
Autoencoders have achieved striking successes when $f$ and $g$ are
selected through training from the class of functions realized
by multilayer perceptrons of a given architecture~\cite{HintonS06}.
Yet, the canonical autoencoder formulation described above has
a notable failing, namely that for linear autoencoders,
optimal choices of $f$ and $g$ do not necessarily identify the
principal components of the dataset; they merely identify the principal
subspace~\cite{BourlardK88, BaldiH89}. That is, the components of $f(\bm{x})$
are not necessarily proportional to projections of $\bm{x}$ against the eigenvectors of the covariance matrix
\begin{equation}\label{eq:cov}
\bm{K}\stackrel{\mathrm{def}}{=} \frac{1}{n} \sum_{i = 1}^n \bm{x}_i \cdot \bm{x}_i^\top,
\end{equation}
which we assume without loss of generality is full rank.
Thus, linear autoencoders do not recover Principal Component
Analysis (PCA). The reason {for this} is that both the objective
(the distortion) and the constraint (the dimensionality
of the latents) are invariant to an invertible transformation
applied after the encoder with its inverse applied
before the decoder. It is desirable for linear autoencoders to recover PCA for two reasons. First, from a representation learning standpoint, it guarantees that the autoencoder recovers uncorrelated features. Second, since a conventional linear autoencoder has a large number of globally optimal solutions corresponding to different bases of the principal subspace, it is preferable to eliminate this indeterminism.
Autoencoders are sometimes described as ``compressing'' the
data~\cite{Santo12,BourlardK88, LiaoZWLL21, Bishop06}, even though $f$
can be invertible even when $k < d$.
We show that by embracing this {compression-}view, one can obtain
autoencoders that are {able} to recover PCA.
Specifically, we consider linear autoencoders with quantized (or,
equivalently, noisy) latent variables with a constraint on the estimated
number of bits required to transmit the quantized latents
under fixed-rate coding. We call this problem
\emph{Principal Bit Analysis (PBA).}
The constraint turns out to be a strictly Schur-concave
function of the set of variances of the latent variables (see
the supplementary for a review of Schur-concavity). Although
finding the optimal $f$ and $g$ for this loss function is
a nonconvex optimization problem, we
show that for any strictly Schur-concave loss function,
an optimal $f$ must send projections of the data along the
principal components, assuming that the empirical covariance
matrix of the data has only simple eigenvalues. That is,
imposing a strictly Schur-concave loss in place of a
simple dimensionality constraint suffices to ensure recovery of PCA.
The idea is that the strict concavity of the loss function
eliminates the rotational invariance described above.
As we show, even a slight amount of ``curvature'' in the
constraint forces the autoencoder to spread the variances
of the latents out as much as possible, resulting in recovery
of PCA. If the loss function is merely Schur-concave, then
projecting along the principal components is optimal, but
not necessarily uniquely so.
Using this theorem, we can efficiently solve PBA.
We validate the solution experimentally by using it to construct
a fixed-rate compression algorithm for arbitrary vector-valued
data sources. We find that the PBA-derived compressor beats
existing linear, fixed-rate compressors both in terms of
mean squared error, for which it is optimized, and in terms
of the structural similarity index measure (SSIM) and downstream
classification accuracy, for which it is not.
A number of variable-rate multimedia compressors have recently
been proposed that are either related to, or directly
inspired by, autoencoders \cite{TschannenAL18, TodericiVJHMSC17,
BalleLS16, TodericiOHVMBCS16, TheisSCH17, RippelB17, HabibianRTC19,
AgustssonMTCTBG17, BalleMSHJ18, ZhouCGSW18, AgustssonTMTG19, BalleCMSJAHT20}.
As a second application of our result,
we show that for Gaussian sources, a linear form of such a
compressor is guaranteed to recover PCA. Thus we show that
ideas from compression can be fruitfully fed back into the
original autoencoder problem.
The contributions of the paper are
\begin{itemize}
\item We propose a novel linear autoencoder formulation in which
the constraint is Schur-concave. We show that this generalizes
conventional linear autoencoding.
\item If the constraint is strictly Schur-concave and the
covariance matrix of the data has only simple
eigenvalues, then we show that the autoencoder
provably recovers PCA, providing a new remedy
for a known limitation of linear autoencoders.
\item We use the new linear autoencoder formulation to efficiently
solve a fixed-rate compression problem that we
call \emph{Principal Bit Analysis (PBA).}
\item We demonstate experimentally that PBA outperforms
existing fixed-rate compressors on a variety
of data sets and metrics.
\item We show that a linear, variable-rate compressor that
is representative of many autoencoder-based
compressors in the literature effectively has
a strictly Schur-concave loss, and therefore it
recovers PCA.
\end{itemize}
\section{Notation}
\section{Linear Autoencoding with a Schur-Concave Constraint}\label{sec:FRAMEWORK}
Throughout this paper we consider $\mathcal{C}_f$ and
$\mathcal{C}_g$ to be the class of linear functions. The
functions $f\in\mathcal{C}_f$ and $g \in\mathcal{C}_g$ can then be represented by $d$-by-$d$ matrices,
respectively, which we denote by
$\bm{W}$ and $\bm{T}$, respectively. Thus we have
\begin{align}
f(\bm{x}) & = \bm{W}^\top \bm{x} \\
g(\bm{x}) & = \bm{T} \bm{x}.
\end{align}
We wish to design $\bm{W}$ and $\bm{T}$ to
minimize the mean squared error when the latent variables
$\bm{W}^\top \bm{x}$ are quantized, subject to a constraint
on the number of bits needed to represent the quantized latents.
We accomplish this via two modifications of the canonical autoencoder.
First, we perturb the $d$ latent variables with zero-mean
additive noise with covariance matrix $\sigma^2 \bm{I}$, which we denote by
$\bm{\eps}.$ Thus the input to the decoder is
\begin{equation}
\bm{W}^\top \bm{x} + \bm{\eps}
\end{equation}
and our objective is to minimize the mean squared error
\begin{equation}
\label{eq:1stmse}
\frac{1}{n} \sum\limits_{i=1}^{n} \mathbb{E}_{\bm{\eps}}\left[ \left \lVert \bm{x}_i - \bm{T}\left( \bm{W}^\top\bm{x}_i+ \bm{\eps} \right) \right \rVert_2^2 \right].
\end{equation}
This is equivalent to quantizing the latents, in the following sense~\cite{ZamirF92}.
Let $Q(\cdot)$ be the function that maps any real number to its nearest
integer and $\varepsilon$ be a random variable uniformly distributed over $[-1/2,1/2]$. Then for $X$ independent of $\varepsilon$, the quantities $Q(X + \varepsilon) - \varepsilon$
and $X + \varepsilon$ have the same joint distribution with $X$.
Thus (\ref{eq:1stmse}) is exactly the mean squared error if the latents
are quantized to the nearest integer and $\sigma^2 = \frac{1}{12}$,
assuming that the quantization is dithered. The overall system is depicted
in Fig.~\ref{fig:blockdiag}.
\tikzstyle{block} = [rectangle, draw, text width=5em, text centered, rounded corners, minimum height=4em]
\tikzstyle{line} = [draw, -latex']
\tikzstyle{input} = [coordinate]
\begin{figure}[!htb]
\begin{center}
\begin{tikzpicture}[node distance=1cm, auto]
\node [input, name=input] {};
\node [block,right = 1.5cm of input] (A) {Linear Encoder ($\bm{W}$)};
\node [block, right=1.5cm of A] (B) {Quantizer};
\node [block, right=2cm of B] (C) {Linear Decoder ($\bm{T}$)};
\node [text width = 1.5cm, right = 0.5cm of C] (D) {$\bm{T}\left(\bm{W}^\top \bm{x}_i + \bm{\varepsilon} \right)$};
\draw [draw,->] (input) -- node {$\bm{x}_i$} (A);
\path [line] (A) -- node [text width=2cm,midway,above,align=center ] {$\bm{W}^\top \bm{x}_i$} (B);
\path [line] (B) -- node [text width=2cm,midway,above,align=center ] {$\bm{W}^\top \bm{x}_i + \bm{\varepsilon}$} (C);
\path [line] (C) -- (D);
\end{tikzpicture}
\end{center}
\caption{Compression Block Diagram}\label{fig:blockdiag}
\end{figure}
We wish to constrain the number of bits needed to describe
the latent variables. We assume that the $j$th quantized latent
is clipped to the interval
$$\left(-\frac{\sqrt{ (2a)^2 \bm{w}_j^\top \bm{K} \bm{w}_j + 1}}{2},
\frac{\sqrt{(2a)^2 \bm{w}_j^\top \bm{K} \bm{w}_j + 1}}{2}\right],$$ where
$a > 0$ is a hyperparameter and the covariance matrix $\bm{K}$ is as defined in \eqref{eq:cov}. The idea is that for sufficiently
large $a$, the interval $$\left(-a\sqrt{ \bm{w}_j^\top \bm{K}
\bm{w}_j}, a\sqrt{ \bm{w}_j^\top \bm{K} \bm{w}_j}\right]$$
contains the latent with high probability, and adding $1$
accounts for the expansion due to the dither. The number of
bits needed for the $j$th latent is then
\begin{align}
\log \left(\sqrt{ 4a^2 \bm{w}_j^\top \bm{K} \bm{w}_j + 1}\right)
= \frac{1}{2} \log \left(4a^2 \bm{w}_j^\top \bm{K} \bm{w}_j + 1\right).
\end{align}
We arrive at our optimization problem:
\begin{equation}\label{eq:MASTEROPTIPBA}
\begin{aligned}
\inf\limits_{\bm{W},\bm{T}} \quad & \frac{1}{n} \sum\limits_{i=1}^{n} \mathbb{E}_{\bm{\eps}}\left[ \left \lVert \bm{x}_i - \bm{T}\left( \bm{W}^\top\bm{x}_i+ \bm{\eps} \right) \right \rVert_2^2 \right] \\
\text{subject to} \quad & R \geq \sum_{j = 1}^d \frac{1}{2} \log \left(4a^2 \bm{w}_i^\top \bm{K} \bm{w}_i + 1\right).
\end{aligned}
\end{equation}
Note that the function
$$\{\bm{w}_j^\top \bm{K} \bm{w}_j\}_{j = 1}^d \mapsto \sum_{j = 1}^d \frac{1}{2} \log \left(4a^2 \bm{w}_i^\top \bm{K} \bm{w}_i + 1\right)$$
is strictly Schur-concave (see Appendix~\ref{app:schur} for a brief review
of Schur-concavity). Our first result only requires that the constraint
is Schur-concave in the set of latent variances, so we will consider
the more general problem
\begin{equation}\label{eq:MASTEROPTI}
\begin{aligned}
\inf\limits_{\bm{W},\bm{T}} \quad & \frac{1}{n} \sum\limits_{i=1}^{n} \mathbb{E}_{\bm{\eps}}\left[ \left \lVert \bm{x}_i - \bm{T}\left( \bm{W}^\top\bm{x}_i+ \bm{\eps} \right) \right \rVert_2^2 \right] \\
\text{subject to} \quad & R \geq \rho\left(\left\{\bm{w}_j^\top \bm{K} \bm{w}_j\right\}_{j = 1}^d\right)
\end{aligned}
\end{equation}
where $\rho(\cdot)$ is any Schur-concave function.
Expressing the objective in~\eqref{eq:MASTEROPTI} in terms of $\bm{K}$, the optimization problem reduces to
\begin{equation}\label{eq:MASTEROPTI_LIN}
\begin{aligned}
\inf \limits_{\bm{W}, \bm{T}} \quad & \text{tr}\left( \bm{K} \right) - 2\text{tr}\left(\bm{K} \bm{W} \bm{T}^\top \right) + \text{tr}\left( \bm{T}\left( \bm{W}^\top \bm{K} \bm{W} + \sigma^2\bm{I} \right) \bm{T}^\top \right) \\
\text{subject to} \quad & R \geq \rho\left(\left\{\bm{w}_j^\top \bm{K} \bm{w}_j\right\}_{j = 1}^d\right).
\end{aligned}
\end{equation}
Since $\bm{T}$ does not appear in the rate constraint, the optimal $\bm{T}$ can be viewed as the Linear Least Squares Estimate (LLSE) of a random $\bm{x}$ given $\bm{W}^\top \bm{x} + \bm{\eps}$. Therefore, the optimal decoder, $\bm{T}^*$ for a given encoder $\bm{W}$ is (e.g. \cite{Kay98}):
\begin{equation}
\bm{T}^* = \bm{K} \bm{W} (\bm{W}^{\top} \bm{K} \bm{W} + \sigma^2\bm{I})^{-1}.
\end{equation}
Substituting for $\bm{T}$ in \eqref{eq:MASTEROPTI_LIN} yields an optimization problem over only $\bm{W}$
\begin{equation} \label{eq:MASTEROPTI_MMSE}
\begin{aligned}
\inf\limits_{\bm{W}} \quad & \text{tr}(\bm{K}) - \text{tr}(\bm{K} \bm{W} (\bm{W}^{\top} \bm{K} \bm{W} + \sigma^2\bm{I} )^{-1}\bm{W}^{\top} \bm{K} ) \\
\text{subject to} \quad & R \geq \rho\left(\left\{\bm{w}_j^\top \bm{K} \bm{w}_j\right\}_{j = 1}^d\right).
\end{aligned}
\end{equation}
This problem is nonconvex in general.
In the following subsection, we prove a structural result about
the problem for a Schur-concave $\rho$. Namely, we show that
the nonzero rows of $\bm{W}$ must be eigenvectors of $\bm{K}$.
In Section~\ref{sec:PBAPCA}, we solve the problem for the specific
choice of $\rho$ in (\ref{eq:MASTEROPTIPBA}). We also show how this
generalizes conventional linear autoencoders.
\subsection{Optimal Autoencoding with a Schur-Concave Constraint}\label{subsec:PROOF}
The following is the main theoretical result of the paper.
\begin{theorem}\label{thm:diagisoptimal}
For Schur-concave $\rho:\mathbb{R}_{\geq 0}^{d} \rightarrow \mathbb{R}_{\geq 0}$ and $R>0$, the set of matrices whose nonzero columns are eigenvectors of the covariance matrix $\bm{K}$ is optimal for \eqref{eq:MASTEROPTI_MMSE}. If $\rho$ is strictly Schur-concave and $\bm{K}$ contains distinct eigenvalues, this set contains all optimal solutions of \eqref{eq:MASTEROPTI_MMSE}.
\end{theorem}
\begin{proof}
Let the eigenvalues of $\bm{K}$ be $\{\sigma_i^2\}_{i=1}^{d}$ with $\sigma_1^2 \ge \sigma_2^2 \ge \ldots \ge \sigma^2_d$. Let the eigendecomposition of $\bm{K}$ be given by $\bm{K} = \bm{U} \bm{\Sigma} \bm{U}^{\top}$ where $\bm{U}$ is an orthogonal matrix whose columns are the eigenvectors of $\bm{K}$ and $\bm{\Sigma}$ is a diagonal matrix with entries $\left \lbrace \sigma_i^2 \right \rbrace_{i=1}^{d}$.
We first prove that the optimal value of \eqref{eq:MASTEROPTI_MMSE} can be achieved by \newest{a} $\bm{W}$ such that $\bm{W}^{\top} \bm{K} \bm{W}$ is a diagonal matrix. Let $\widetilde{\bm{W}} = \bm{W} \bm{Q}$ where $\bm{Q}$ is the orthogonal matrix obtained from the eigendecomposition of $\bm{W}^{\top} \bm{K} \bm{W}$ i.e., $$\bm{W}^{\top}\bm{K} \bm{W} = \bm{Q}\bm{\Lambda} \bm{Q}^{\top},$$ where $\bm{\Lambda}$ is a diagonal matrix formed from the eigenvalues of $ \bm{W}^{\top}\bm{K} \bm{W}$. Note that
\begin{align*}
\text{tr} \left( \bm{K} \widetilde{\bm{W}} \left( \widetilde{\bm{W}}^{\top} \bm{K} \widetilde{\bm{W}} + \sigma^2\bm{I} \right)^{-1} \widetilde{\bm{W}}^{\top} \bm{K}\right) &= \text{tr} \left( \bm{K} \bm{W} \bm{Q} \left( \bm{\Lambda} + \sigma^2\bm{I} \right)^{-1} \bm{Q}^{\top} \bm{W}^{\top} \bm{K} \right) \\
&= \text{tr} \left( \bm{K} \bm{W} \left( \bm{Q} \bm{\Lambda} \bm{Q}^{\top} + \sigma^2\bm{Q} \bm{Q}^{\top} \right)^{-1} \bm{W}^{\top} \bm{K} \right).
\end{align*}
Since $\bm{Q} \bm{\Lambda} \bm{Q}^{\top} = \bm{W}^{\top} \bm{K} \bm{W}$ and $\bm{Q} \bm{Q}^{\top}= \bm{I}$, the objective remains the same.
We now show that the constraint is only improved. Denoting the eigenvalues of $\bm{W}^{\top} \bm{K} \bm{W}$ by $\left \lbrace \nu_j \right \rbrace_{j=1}^d$, we have
\begin{equation*}
\rho\left( \left \lbrace \widetilde{\bm{w}}_j^{\top} \bm{K} \widetilde{\bm{w}}_j \right \rbrace_{j=1}^{d} \right)
= \rho\left( \left \lbrace \bm{q}_j^{\top} \bm{W}^{\top} \bm{K} \bm{W} \bm{q}_j \right \rbrace_{j=1}^{d} \right) = \rho \left( \left \lbrace \nu_j \right \rbrace_{j=1}^{d} \right).
\end{equation*}
Now since the eigenvalues of a Hermitian matrix majorize its diagonal elements by the Schur-Horn theorem \cite[Theorem~4.3.45]{HornJ85},
$$ \left \lbrace \bm{w}_j^{\top} \bm{K} \bm{w}_j \right \rbrace_{j=1}^{d} \prec \left \lbrace \nu_j \right \rbrace_{j=1}^{d}.$$
Since $\rho$ is Schur-concave, this implies
$$ \rho\left( \left \lbrace \bm{w}_j^{\top} \bm{K} \bm{w}_j \right \rbrace_{j=1}^{d} \right) \geq \rho \left( \left \lbrace \nu_j \right \rbrace_{j=1}^{d} \right) = \rho \left( \left \lbrace \widetilde{\bm{w}}_j^{\top} \bm{K} \widetilde{\bm{w}}_i \right \rbrace_{j=1}^{d} \right). $$
Therefore, if $\rho$ is Schur-concave, the rate constraint can only improve. This implies an optimal solution can be attained when
$\bm{W}$ is such that $\bm{W}^{\top} \bm{K} \bm{W}$ is diagonal. If $\rho$ is strictly Schur-concave, the rate constraint strictly improves implying that the optimal $\bm{W}$ must be such that $\bm{W}^{\top} \bm{K} \bm{W}$ is diagonal.
This implies that
\begin{align*}
\text{tr}\left( \bm{K} \bm{W} \left( \bm{W}^{\top} \bm{K} \bm{W} + \sigma^2\bm{I} \right)^{-1} \bm{W}^{\top} \bm{K} \right) &= \text{tr}\left( \bm{W}^{\top} \bm{K}^2 \bm{W} \left( \bm{W}^{\top} \bm{K} \bm{W} + \sigma^2\bm{I} \right)^{-1}\right) \\
&= \sum\limits_{i=1}^{d} \frac{\bm{w}_i^{\top} \bm{K}^2 \bm{w}_i}{\sigma^2 + \bm{w}_i^{\top} \bm{K} \bm{w}_i}.
\end{align*}
Note that minimizing the objective in \eqref{eq:MASTEROPTI_MMSE} is equivalent to maximizing the above expression. Perform the change of variable
\begin{align*}
\bm{w}_j & \mapsto \begin{cases}
\left(\frac{\bm{K}^{1/2} \bm{w}_j}{||\bm{K}^{1/2} \bm{w}_j||}, ||\bm{K}^{1/2} \bm{w}_j||^2\right) & \text{if $\bm{K}^{1/2} \bm{w}_j \ne \bm{0}$} \\
(\bm{0},0) & \text{if $\bm{K}^{1/2} \bm{w}_j = \bm{0}$}
\end{cases} \\
& = (\bm{y}_j,y_j).
\end{align*}
The assumption that $\bm{W}^\top \bm{K} \bm{W}$ is diagonal and
the normalization in the definition of $\bm{y}_j$ implies that
\[
\bm{Y} = [ \bm{y}_1 \bm{y}_2, \cdots, \bm{y}_d]
\]
is a matrix whose nonzero columns form an orthonormal set. Rewriting the objective in terms of
the $(\bm{y}_j,y_j)$, we have
\begin{equation}\label{eq:objinY}
\sum\limits_{i=1}^{d} \frac{\bm{w}_i^{\top} \bm{K}^2 \bm{w}_i}{\sigma^2 + \bm{w}_i^{\top} \bm{K} \bm{w}_i} = \sum\limits_{i=1}^{d} \bm{y}_i^{\top} \bm{K} \bm{y}_i \frac{y_i}{\sigma^2 + y_i} = \sum\limits_{i=1}^{d} \bm{y}_i^{\top} \bm{K} \bm{y}_i m_i,
\end{equation}
where $m_i = \frac{y_i}{\sigma^2+y_i}$.
Observe that under this new parametrization, the constraint only depends on $\{y_i \}_{i=1}^{d}$. Without loss of generality, we assume that $y_1 \geq y_2 \geq \cdots \ge y_d$, implying that $m_1 \geq m_2 \geq \cdots \geq m_d$.
We now prove that for given $\{y_i \}_{i=1}^{d}$, choosing the $\bm{y}_i$ along the eigenvectors of $\bm{K}$ is optimal.
Denote the diagonal elements of $\bm{Y}^\top \bm{K} \bm{Y}$ by $\{\lambda_i^2\}_{i = 1}^d$ and let $\{ \lambda_{i, \downarrow}^2 \}_{i=1}^{d}$ denote the same diagonal elements arranged in descending order. Denote the eigenvalues of $\bm{Y}^{\top} \bm{K} \bm{Y}$ by $\{ \mu_i^2 \}_{i=1}^{d}$ where $\mu_1 \geq \mu_2 \geq \cdots \geq \mu_d$.
Again invoking the Schur-Horn theorem, the eigenvalues of $\bm{Y}^{\top} \bm{K} \bm{Y}$
majorize its diagonal entries
\begin{equation}\label{eq:majorization}
\left \lbrace \lambda_i^2 \right \rbrace_{i=1}^{d} \prec \left \lbrace \mu_i^2 \right \rbrace_{i=1}^{d}.
\end{equation}
Substituting $\lambda_i^2 = \bm{y}_i^{\top} \bm{K} \bm{y}_i$ in \eqref{eq:objinY}, we have
\begin{align*}
\sum\limits_{i=1}^{d} \lambda_i^2 m_i \stackrel{(a)}{\leq} \sum\limits_{i=1}^{d} \lambda_{i,\downarrow}^2 m_i &= \lambda_{1,\downarrow}^2 m_1 + \sum\limits_{i=2}^{d} \left( \sum\limits_{j=1}^{i} \lambda_{j,\downarrow}^2 -\sum\limits_{j=1}^{i-1} \lambda_{j,\downarrow}^2 \right) m_i \\
&= \lambda_{1,\downarrow}^2 m_1 + \sum\limits_{i=2}^{d} m_i \sum\limits_{j=1}^{i} \lambda_{j,\downarrow}^2 - \sum\limits_{i=2}^{d} m_i \sum\limits_{j=1}^{i-1} \lambda_{j,\downarrow}^2 \\
&= \lambda_{1,\downarrow}^2 (m_1 - m_2) + m_d \left( \sum\limits_{j=1}^{d} \lambda_{j,\downarrow}^2 \right) + \sum\limits_{i=2}^{d-1} \left( m_i - m_{i+1}\right) \sum\limits_{j=1}^{i} \lambda_{j,\downarrow}^2 \\
&\stackrel{(b)}{\leq} \mu_1^2 (m_1 - m_2) + m_d \left( \sum\limits_{j=1}^{d} \mu_j^2 \right) + \sum\limits_{i=2}^{d-1} (m_i - m_{i+1}) \sum\limits_{j=1}^{i} \mu_j^2\\
&\stackrel{(c)}{\leq} \sigma_1^2 (m_1 - m_2) + m_d \left( \sum\limits_{j=1}^{d} \sigma_j^2 \right) + \sum\limits_{i=2}^{d-1} (m_i - m_{i+1}) \sum\limits_{j=1}^{i} \sigma_j^2\\
&= \sum\limits_{i=1}^{d} \sigma_i^2 m_i,
\end{align*}
where inequality $(a)$ follows from the assumption that $m_1 \geq m_2 \geq \cdots \geq m_d$, and $(b)$ from the definition in \eqref{eq:majorization}. Since $\bm{Y} $'s nonzero columns form an orthonormal set, the eigenvalues of $\bm{Y}^{\top} \bm{K} \bm{Y}$, when arranged in descending order, are at most the eigenvalues of $\bm{K}$ from Corollary 4.3.37 in \cite{HornJ85}, and therefore $(c)$ follows.
This upper bound is attained when $\bm{y}_i = \bm{u}_i$ for nonzero $y_i$, where $\bm{u}_i$ is the normalized eigenvector of $\bm{K}$ corresponding to eigenvalue $\sigma_i^2$. To see this, note that when $\bm{y}_i = \bm{u}_i$, $\lambda_i^2 = \mu_i^2 = \sigma_i^2$. From the definition of $\bm{y}_i, \bm{w}_i = \bm{K}^{-1/2}\bm{u}_i \sqrt{y_i} = \bm{u}_i \frac{\sqrt{y_i}}{\sigma_i}$. Therefore, for a Schur-concave $\rho$, the set of matrices whose nonzero columns are eigenvectors of $\bm{K}$ is optimal. We now prove that for a strictly Schur-concave $\rho$, if $\bm{K}$ has distinct eigenvalues, this set contains all of the optimal solutions $\bm{W}$.
We know that for a fixed $y_1 \geq y_2 \geq \cdots \geq y_d$, (implying a fixed $m_1 \geq m_2 \geq \cdots \geq m_d$) the upper bound $\sum\limits_{i=1}^{d} \sigma_i^2 m_i$ is attained by the previous choice of $\bm{y}_i$. Note that if all nonzero $m_i$ are distinct, equality in $(b)$ and $(c)$ is attained if and only if the nonzero diagonal elements of $\bm{Y}^{\top} \bm{K} \bm{Y}$ equal the corresponding eigenvalues of $\bm{K}$. This implies that, if all nonzero $m_i$ are distinct, the upper bound is attained if and only if $\bm{y}_i = \bm{u}_i$ for nonzero $y_i$. Therefore, it is sufficient to prove that for the following optimization problem
\begin{equation} \label{eq:ubschur}
\begin{aligned}
\sup\limits_{\{y_i \geq 0\}} \quad & \sum\limits_{i=1}^{d} \sigma_i^2 \frac{y_i}{\sigma^2 + y_i} \\
\text{subject to} \quad & R \geq \rho\left(\left\{y_i \right\}_{i= 1}^d\right),
\end{aligned}
\end{equation}
any optimal $\{y_i\}$ must be such that the nonzero $y_i$ are distinct. Firstly, note that since $\sigma_1^2 > \sigma_2^2 > \cdots > \sigma_d^2$, we must have $y_1 \geq y_2 \geq \cdots \geq y_d$. Assume to the contrary that for an optimal $\{ y_i \}_{i=1}^{d}$ there exists $1 \leq j, \ell < d$ such that $y_{j-1} > y_j = y_{j+1} = y_{j+2} = \cdots = y_{j+\ell} > y_{j+\ell+1} \geq 0$, where $y_0$ is chosen to be any real number strictly greater than $y_1$ and $y_{d+1}=0$.
Take $\delta > 0$ small.
Denote a new sequence $ \{ y'_i \}_{i=1}^{d}$ where $y'_{j} = y_j+\delta, y'_{j+\ell} = y_{j+\ell} - \delta$ and $y'_i = y_i$ for $1\leq i \leq d$ with $i \neq j$ and $j+\ell$.
Since $\rho$ is strictly Schur-concave, the constraint is strictly
improved,
\[ \rho\left( \{y'_i \}_{i=1}^{d} \right) < \rho\left( \{y_i \}_{i=1}^{d} \right). \]
Since $\sigma^2_j > \sigma^2_{j + \ell}$, the objective is strictly
improved for sufficiently small $\delta$,
$$
\sum\limits_{i=1}^{d} \sigma_i^2 \frac{y_i}{\sigma^2 + y_i} <
\sum\limits_{i=1}^{d} \sigma_i^2 \frac{y'_i}{\sigma^2 + y'_i},
$$
as desired.
\begin{comment}
the optimal nonzero $m_i$ are distinct. Firstly, note that since $\sigma_1^2 \geq \sigma_2^2 \geq \cdots \sigma_d^2$, $m_1 \geq m_2 \geq \cdots m_d$. Assume to the contrary that for an optimal $\{ m_i \}_{i=1}^{d}$ there exists $1 \leq j, \ell < d$ such that $m_{j-1} > m_j = m_{j+1} = m_{j+2} \cdots m_{j+\ell} > m_{j+\ell+1} \geq 0$, where $m_0$ is chosen to be any real number strictly greater than $m_1$ and $m_{d+1}=0$. Define $\delta \triangleq \min\left( \frac{m_{j-1} - m_j}{3}, \frac{m_{j+\ell}-m_{j+\ell+1}}{3} \right) > 0$. Denote a new sequence $ \{ m'_i \}_{i=1}^{d}$ where $m'_{j} = m_j+\delta, m'_{j+\ell} = m_{j+\ell} - \delta$ and $m'_i = m_i$ for $1\leq i \leq d$, $i \neq j,j+\ell$. Note that
\[ \{m'_i \}_{i=1}^{d} \succ \{m_i \}_{i=1}^{d}. \]
Since $\rho$ is strictly Schur-concave,
\[ \rho\left( \{m'_i \}_{i=1}^{d} \right) < \rho\left( \{m_i \}_{i=1}^{d} \right). \]
Also $\sigma_j^2 m'_j + \sigma_{j+\ell}^2 m'_{j+\ell} = \sigma_j^2 m_j + \sigma_{j+\ell}^2 m_{j+\ell} + \delta\left(\sigma_j^2 - \sigma^2_{j+\ell}\right) \geq \sigma_j^2 m_j + \sigma_{j+\ell}^2 m_{j+\ell}. $ Therefore, the objective only increases while the rate constraint strictly improves, contradicting the optimality assumption. This implies that the optimal nonzero $m_i$ are distinct, further implying that for a strictly Schur-concave $\rho$, the optimal value is attained only when a nonzero $\bm{y}_i = \frac{\bm{K}^{1/2} \bm{w}_i}{\left \lVert \bm{K}^{1/2} \bm{w}_i \right \rVert} = \bm{u}_i$. Therefore, $\bm{w}_i = \bm{K}^{-1/2}\bm{u}_i \sqrt{y_i} = \bm{u}_i \frac{\sqrt{y_i}}{\sigma_i}$.
\end{comment}
\end{proof}
As a consequence of Theorem~\ref{thm:diagisoptimal}, encoding via an optimal $\bm{W}$ can be viewed as a projection along the eigenvectors of $\bm{K}$, followed by different scalings applied to each component, i.e. $\bm{W} = \bm{U} \bm{S}$ where $\bm{S}$ is a diagonal matrix with entries $s_i \geq 0$ and
$\bm{U}$ is the normalized eigenvector matrix. Only $\bm{S}$ remains to be determined, and to this end, we may assume that
$\bm{K}$ is diagonal with nonincreasing diagonal entries, implying $\bm{U} = \bm{I}$. In subsequent sections, our choice of $\rho$ will be of the form $\sum\limits_{i=1}^{d} \rho_{sl}$, where $\rho_{sl}: \mathbb{R}_{\geq 0} \rightarrow \mathbb{R}_{\geq 0}$\footnote{``sl" stands for single-letter} is (strictly) concave, making $\rho$ (strictly) Schur-concave (see Proposition~\ref{prop:single-letter} in Appendix~\ref{app:schur}). Therefore, \eqref{eq:MASTEROPTI_MMSE} reduces to
\begin{equation} \label{eq:MASTEROPTI_S}
\begin{aligned}
\inf\limits_{\bm{S}} \quad & \text{tr}(\bm{K}) - \text{tr}(\bm{K} \bm{S} (\bm{S}^{\top} \bm{K} \bm{S} + \sigma^2\bm{I} )^{-1}\bm{S}^{\top} \bm{K} ) \\
\text{subject to} \quad & R \geq \rho_{sl}\left( \{s_i^2 \sigma^2_i \} \right),
\end{aligned}
\end{equation}
where the infimum is over diagonal matrices $\bm{S}$. To handle situations for which
\begin{equation}
\lim_{s \rightarrow \infty} \rho_{sl}(s) < \infty,
\end{equation}
we allow the diagonal entries of $\bm{S}$ to be $\infty$, with the objective for such cases defined via its continuous extension.
In the next section, we will solve \eqref{eq:MASTEROPTI_S} for several specific choices of $\rho_{sl}$.
\section{Explicit Solutions: Conventional Linear Autoencoders and PBA} \label{sec:PBAPCA}
\subsection{Conventional Linear Autoencoders} \label{subsec:PCA}
Given a centered dataset $\bm{x}_1, \bm{x}_2, \cdots, \bm{x}_n \in \mathbb{R}^{d}$, consider a linear autoencoder optimization problem where the encoder and decoder, $\bm{W}$ and $ \bm{T}$, respectively, are $d$-by-$k$ matrices where $k \leq d$ is a parameter. The goal is to minimize the mean squared error as given by \eqref{eq:1stmse}. PCA corresponds to the global optimal solution of this optimization problem, where $\bm{W} = \bm{T} = \bm{U}_k$, where $\bm{U}_k \in \mathbb{R}^{d \times k}$ is a matrix whose columns are the $k$ eigenvectors corresponding to the $k$ largest eigenvalues of $\bm{K}$. However, there are multiple global optimal solutions, given by any encoder-decoder pair of the form $\left( \bm{U}_k \bm{V}, \bm{U}_k \bm{V} \right)$, where $\bm{V}$ is an orthogonal matrix \cite{BaldiH89}.
We now recover linear autoencoders through our framework in Section~\ref{sec:FRAMEWORK}. Consider the optimization problem in \eqref{eq:MASTEROPTI_S} where $\rho_{sl}: \mathbb{R}_{\geq 0} \rightarrow \{0,1 \}$ is a concave function defined as
\begin{equation} \label{eq:PCA_F}
\rho_{sl}(x) = \bm{1} \left[ x > 0 \right].
\end{equation}
Note that this penalizes the dimension of the latents, as desired.
Note also that this cost is Schur-concave but not strictly so.
The fact that PCA solves conventional linear autoencoding,
but is not necessarily the unique solution, follows immediately
from Theorem~\ref{thm:diagisoptimal}.
\begin{theorem}\label{thm:PCA}
If $\rho_{sl}(\cdot)$ is given by \eqref{eq:PCA_F}, then an optimal solution for \eqref{eq:MASTEROPTI_S} is given by a diagonal matrix $\bm{S}$ whose top $\min(\lfloor R \rfloor, d)$ diagonal entries are equal to $\infty$ and the remaining entries are $0$.
\end{theorem}
\begin{proof}
Let ${\cal F} \stackrel{\mathrm{def}}{=} \left \lbrace i \in \left[ d \right] : s_i > 0 \right \rbrace$, implying $\left \lvert {\cal F} \right \rvert \leq R$. Since $\bm{K}$ and $\bm{S}$ are diagonal, the optimization problem in \eqref{eq:MASTEROPTI_S} can be written as
\begin{equation} \label{eq:PCA_OPT}
\begin{aligned}
\inf\limits_{\left \lbrace s_{\ell} \right \rbrace} & \sum\limits_{j \in \left[d\right] \backslash {\cal F}} \sigma_j^2 + \sum\limits_{ \ell \in {\cal F}} \frac{\sigma^2 \sigma_{\ell}^2}{\sigma^2+ \sigma_{\ell}^2 s_{\ell}^2} \\
\text{subject to} \quad & R \geq \sum\limits_{i=1}^{d} \bm{1}\left[ s_i > 0 \right].
\end{aligned}
\end{equation}
Since the value of $s_{\ell}, \ell \in {\cal F}$ does not affect the rate constraint, each of the $s_{\ell}$ can be made as large as possible without changing the rate constraint. Therefore, the infimum value of the objective is $\sum\limits_{j \in \left[ d \right] \backslash {\cal F}} \sigma^2_j$. Since we seek to minimize the distortion, the optimal ${\cal F}$ is the set of indices with the largest $\left \lvert {\cal F} \right \rvert$ eigenvalues. Since the number of these eigenvalues cannot exceed $R$, we choose $\left \lvert {\cal F} \right \rvert = \min(\lfloor R \rfloor,d)$.
\end{proof}
Unlike the conventional linear autoencoder framework, in Section~\ref{sec:FRAMEWORK}, the latent variables $\bm{W}^{\top} \bm{x}$ are quantized, which we model with additive white noise of fixed variance. Therefore, an infinite value of $s_i$ indicates sending $\bm{u}_i^{\top} \bm{x}$ with full precision where $\bm{u}_i$ is the eigenvector corresponding to the $i^{th}$ largest eigenvalue. This implies that PCA with parameter $k$ corresponds to $\bm{W} = \bm{U} \bm{S}$, where $\bm{S}$ is a diagonal matrix whose top $k$ diagonal entries are equal to $\infty$ and the $d-k$ remaining diagonal entries are $0$. Therefore, for any $R$ such that $\lfloor R \rfloor = k$, an optimal solution to \eqref{eq:MASTEROPTI_S} corresponds to linearly projecting the data along the top $k$ eigenvectors, which is the same as PCA. Note that, like \cite{BaldiH89}, we only prove that projecting along the eigenvectors is one of possibly other optimal solutions. However, even a slight amount of curvature in $\rho$ would make it strictly Schur-concave, thus recovering the principal directions.
We next turn to a specific cost function with curvature, namely
the PBA cost function that was our original motivation.
\subsection{Principal Bit Analysis (PBA)}
Consider the choice of $\rho_{sl}:\mathbb{R}_{\geq 0} \rightarrow \mathbb{R}_{\geq 0}$ that provided the
original impetus for Theorem~\ref{thm:diagisoptimal}. For $\gamma>\frac{2}{\sigma^2}$,
\begin{equation}\label{eq:LOG_F}
\rho_{sl}(x) = \frac{1}{2} \log(\gamma x+1).
\end{equation}
The nature of the optimization problem depends on the value
of $\gamma$. For $1 \le \gamma \sigma^2 \le 2$, the problem can be
made convex with a simple change of variable. For $\gamma \sigma^2= 1$,
the problem coincides with the classical waterfilling procedure
in rate-distortion theory, in fact. For $\gamma \sigma^2 > 2$, the problem
is significantly more challenging. Since we are interested in
relatively large values of $\gamma$ for our compression application
(see Section~\ref{sec:EXPERIMENTS} to follow), we focus on the case $\gamma > 2/\sigma^2$.
\begin{algorithm}[h]
\caption{Principal Bit Analysis (PBA)} \label{alg:PBA}
\begin{algorithmic}[1]
\Require $\lambda > 0$, $\alpha > 2$,
\begin{equation}
\bm{K} = \left[ \begin{array}{cccc}
\sigma^2_1 & 0 & \cdots & 0 \\
0 & \sigma^2_2 & \cdots & 0 \\
\vdots & \vdots & \ddots & \vdots \\
0 & 0 & \cdots & \sigma^2_d \\
\end{array} \right] \succ 0,
\end{equation}
such that $\sigma^2_1 \ge \sigma^2_2 \ge \cdots \ge \sigma^2_d$.
\State If $\lambda \ge \sigma^2_1/(4(\alpha - 1))$, Output
$\bar{R}_{\text{opt}} = 0, \bar{D}_{\text{opt}} = \sum_{i = 1}^d \sigma^2_i$.
\State Set $\bar{d} = \max\left\{i : \lambda < \sigma_i^2/4(\alpha - 1)\right\}$.
\State Set $\bar{R}$, $\bar{D}$ to zero arrays of size $2\bar{d}$.
\For{$r \in \left \lbrace 1,2, \cdots \bar{d} \right \rbrace $}
\State $\bar{D}(2r-1) = \sum\limits_{i=1}^{r} \frac{\sigma_i^2}{2(\alpha -1)} \left( 1 - \sqrt{1 - \frac{4 \lambda (\alpha -1)}{\sigma_i^2}}\right) + \sum\limits_{i=r+1}^{d} \frac{\sigma_i^2}{\alpha } $,
\State $\bar{R}(2r-1)= \sum\limits_{i=1}^{r} \frac{1}{2} \log \left( \frac{\sigma_i^2}{4\lambda} \right) + \log\left( 1 + \sqrt{1 - \frac{4 \lambda (\alpha -1)}{\sigma_i^2}}\right)$.
\State $\bar{D}(2r) =\left( \sum\limits_{i=1}^{r-1} \frac{\sigma_i^2}{2(\alpha -1)} \left( 1 - \sqrt{1 - \frac{4 \lambda (\alpha -1)}{\sigma_i^2}}\right) + \frac{\sigma_{r}^2}{2(\alpha -1)} \left( 1 + \sqrt{1 - \frac{4 \lambda (\alpha -1)}{\sigma_{r}^2}}\right) + \sum\limits_{i=r+1}^{d} \frac{\sigma_i^2}{\alpha } \right)$.
\State $\bar{R}(2r)=\sum\limits_{i=1}^{r} \frac{1}{2} \log \left( \frac{\sigma_i^2}{4\lambda} \right) + \sum\limits_{i=1}^{r-1} \log\left( 1 + \sqrt{1 - \frac{4 \lambda (\alpha -1)}{\sigma_i^2}}\right) + \log\left( 1 - \sqrt{1 - \frac{4 \lambda (\alpha -1)}{\sigma_{r}^2}}\right)$.
\EndFor
\State $r^* \leftarrow \text{argmin}_{j \in \left[ 2\bar{d} \right]} \bar{D}(j) + \lambda \bar{R}(j). $
\State Output $\bar{R}_{\text{opt}} = \bar{R}(r^*), \bar{D}_{\text{opt}} = \bar{D}(r^*)$.
\end{algorithmic}
\end{algorithm}
\begin{theorem}
If $\rho_{sl}(\cdot)$ is given by \eqref{eq:LOG_F} for $\gamma > \frac{2}{\sigma^2}$, then for any $\lambda > 0$,
the pair $\bar{R}_{\text{opt}}, \bar{D}_{\text{opt}}$ obtained from the output of
Algorithm~\ref{alg:PBA} satisfies
\begin{equation}
\label{eq:PBA_OPT_LAGRANGE}
\bar{D}_{\text{opt}} + \lambda \bar{R}_{\text{opt}} =
\inf\limits_{\bm{S}} \quad \text{tr}(\bm{K}) - \text{tr}(\bm{K} \bm{S} (\bm{S}^{\top} \bm{K} \bm{S} + \sigma^2\bm{I} )^{-1}\bm{S}^{\top} \bm{K} )
+ \lambda \sum_{i = 1}^d \rho_{sl}\left( \{s_i^2 \sigma^2_i \} \right),
\end{equation}
\end{theorem}
\begin{proof}
Since $\bm{K}$ and $\bm{S}$ are diagonal, the optimization problem in
\eqref{eq:PBA_OPT_LAGRANGE} can be written as
\begin{equation} \label{eq:PBA_OPT}
\begin{aligned}
\inf\limits_{\left \lbrace s_i \right \rbrace} \quad & \sum\limits_{i=1}^{d} \frac{\sigma^2 \sigma_{i}^{2}}{\sigma^2+s_{i}^{2} \sigma_{i}^{2}}
+ \lambda \cdot \frac{1}{2}\sum\limits_{i=1}^{d} \log\left( 1 + \gamma s_{i}^2 \sigma_{i}^{2} \right).
\end{aligned}
\end{equation}
With the following change of variables $ \alpha = \gamma \sigma^2 $, $s_{i} \mapsto s_{i}'^2 = \alpha \frac{s_{i}'^2}{\sigma^2}$, we obtain
\begin{equation} \label{eq:PBA_OPT_p1}
\begin{aligned}
\inf\limits_{\left \lbrace s_i' \right \rbrace} \quad & \sum\limits_{i=1}^{d} \alpha \frac{ \sigma_{i}^{2}}{\alpha+s_{i}^{'2} \sigma_{i}^{2}} + \lambda \cdot \frac{1}{2}\sum\limits_{i=1}^{d} \log\left( 1 + s_i'^2 \sigma_{i}^{2} \right).
\end{aligned}
\end{equation}
Ignoring the constant factor in the objective, perform the change of variable $s'_{i} \mapsto D_i = \frac{\sigma_i^2}{\alpha +s_{i}'^2\sigma_{i}^{2}}$
to obtain
\begin{equation} \label{eq:PBA_REDUCED_OPT}
\begin{aligned}
\inf\limits_{\left \lbrace D_i \right \rbrace} \quad & \sum\limits_{i=1}^{d} D_i + \frac{\lambda}{2}\sum\limits_{i=1}^{d} \log\left( \frac{\sigma_i^2}{D_i} - (\alpha -1) \right), \\
\text{subject to} \quad &D_i \leq \frac{\sigma_i^2}{\alpha } \hspace{5mm} \text{ for all } i \in \left[ d \right].
\end{aligned}
\end{equation}
This optimization problem is nonconvex since the function $\log\left( \frac{\sigma_i^2}{D_i} - (\alpha -1) \right)$ is convex for $0 \leq D_i \leq \frac{\sigma_i^2}{2(\alpha -1)}$ but concave for $\frac{\sigma_i^2}{2(\alpha -1)} < D_i \leq \frac{\sigma_i^2}{\alpha}$ and the
latter interval is nonempty since $\alpha>2$.
Any optimizing $\{D_i\}$ must be a stationary point of
\begin{equation}\label{eq:LAG_PBA}
\cL\left( \left \lbrace D_i \right \rbrace_{i=1}^{d}, \lambda, \left \lbrace \mu_i \right \rbrace_{i=1}^{d} \right) = \sum\limits_{i=1}^{d} D_i + \lambda\left( \sum\limits_{i=1}^{d} \log\left( \frac{\sigma_i^2}{D_i} - (\alpha -1) \right) \right) + \sum\limits_{i=1}^{d} \mu_i \left( D_i - \frac{\sigma_i^2}{\alpha } \right).
\end{equation}
for some $\{\mu_i\}_{i = 1}^d$ with $\mu_i \geq 0$ for all $i \in \left[ d \right]$ and satisfying the complementary slackness condition~\cite[Prop.~3.3.1]{Bertsekas:Nonlinear}. The stationary points satisfy, for each $i$
\begin{equation}\label{eq:LAG_STAT}
\frac{\partial \cL}{\partial D_i} = 1 - \lambda\left( \frac{\frac{\sigma_i^2}{D_i^2}}{\left( \frac{\sigma_i^2}{D_i} - (\alpha -1) \right)}\right) + \mu_i = 0.
\end{equation}
Let ${\cal F} = \left \lbrace i : D_i < \frac{\sigma_i^2}{\alpha } \right \rbrace$.
For $i \in {\cal F}$, $\mu_i = 0$ due to complementary slackness. Substituting in \eqref{eq:LAG_STAT} we obtain a quadratic equation in $D_i$
\[
(\alpha-1)D_i^2 - \sigma_i^2 D_i + \lambda \sigma_i^2 = 0.
\]
which gives
$$ D_i = \frac{\sigma_i^2}{2(\alpha -1)} \left( 1 \pm \sqrt{1 - \frac{4\lambda (\alpha -1)}{\sigma_i^2}} \right).$$
Let $c_i = \sqrt{1 - \frac{4\lambda (\alpha -1)}{\sigma_i^2}}$. Note that $\frac{\sigma_i^2}{2(\alpha -1)} \left( 1 + c_i \right)$ is always in the concave region and $\frac{\sigma_i^2}{2(\alpha -1)} \left( 1 - c_i \right)$ is always in the convex region for a $\lambda$ chosen such that $D_i$ is a real number strictly less than $\frac{\sigma_i^2}{\alpha }$.
Therefore the optimal set of distortions are contained in the following set of $3^{d}$ points
$$ \prod_{i = 1}^d \left\{ \frac{\sigma_i^2}{2(\alpha -1)} \left( 1 + \sqrt{1 - \frac{4\lambda (\alpha -1)}{\sigma_i^2}} \right), \frac{\sigma_i^2}{2(\alpha -1)} \left( 1 - \sqrt{1 - \frac{4\lambda (\alpha -1)}{\sigma_i^2}} \right), \frac{\sigma_i^2}{\alpha}\right\}.$$
We now reduce the size of the above set by making a two observations:
\noindent \textbf{(1). ${\cal F}$ is contiguous.}
\begin{lemma}\label{lem:NORM_INC}
There exists an optimal $\left \lbrace D_i^* \right \rbrace_{i=1}^{d}$ for \eqref{eq:PBA_REDUCED_OPT} such that (a) $\frac{\sigma_i^2}{D_i^*}$ is a nonincreasing sequence and (b) ${\cal F} = \{ 1,2,\cdots \left \lvert {\cal F} \right \rvert \}$.
\end{lemma}
\begin{proof}
Substitute $x_i = \frac{\sigma_i^2}{D_i}$ in \eqref{eq:PBA_REDUCED_OPT}. This gives us
\begin{equation} \label{eq:PBA_NORM_OPT}
\begin{aligned}
\inf\limits_{\left \lbrace x_i \right \rbrace} \quad & \sum\limits_{i=1}^{d} \frac{\sigma_i^2}{x_i} + \frac{\lambda}{2}\sum\limits_{i=1}^{d} \log\left( x_i - (\alpha -1) \right), \\
\text{subject to} \quad & x_i \geq \alpha \hspace{5mm} \text{ for all } i \in \left[ d \right].
\end{aligned}
\end{equation}
Let $\{ x_i^* \}_{i=1}^{d}$ be an optimal solution for \eqref{eq:PBA_NORM_OPT}. If, for $i>j$, $x_i^* > x_j^* \ge \alpha$, then exchanging the values provides a solution that has the same rate and lower distortion since $\frac{\sigma_i^2}{x_i^*} + \frac{\sigma_j^2}{x_j^*} \geq \frac{\sigma_i^2}{x_j^*} + \frac{\sigma_j^2}{x_i^*}$. This proves (a). Part (b) follows immediately.
\end{proof}
\noindent\textbf{(2). No two solutions are concave.}
\begin{lemma}\label{lem:SINGLE_CONCAVE}
For $R>0$, let $\left \lbrace D_i^* \right \rbrace_{i=1}^{d}$ be an optimal solution for \eqref{eq:PBA_REDUCED_OPT}. There exists at most one $D_i^*$ such that $\frac{\sigma_i^2}{2(\alpha -1)} < D_i^* < \frac{\sigma_i^2}{\alpha }$.
\end{lemma}
\begin{proof}
Let $D_i^*, D_j^*$ be such that $\frac{\sigma_i^2}{2(\alpha -1)} < D_i^* < \frac{\sigma_i^2}{\alpha }$ and $\frac{\sigma_j^2}{2(\alpha -1)} < D_j^* < \frac{\sigma_j^2}{\alpha }$ . Without loss of generality, assume $D_i^* < D_j^*$. Denote the individual rate constraint function by $r\left( D_i \right) \triangleq \log \left( \frac{\sigma_i^2}{D_i} - (\alpha -1) \right)$. Since $r$ is concave in $\left(\frac{\sigma_i^2}{2(a-1)}, \frac{\sigma_i^2}{\alpha } \right)$, there exist an $\varepsilon > 0$ such that
\begin{align}
r\left(D_i^* - \varepsilon\right) +r\left( D_j^* + \varepsilon \right) &= r\left( D_i^* \right) - \varepsilon r'\left( D_i^* \right) + O(\varepsilon^2) + r\left(D_j^*\right) + \varepsilon r'\left(D_j^* \right) + O(\varepsilon^2) \\
&< r\left( D_i^* \right) + r\left(D_j^*\right)
\end{align}
The last inequality follows from concavity of $r$. Therefore, replacing $\left(D_i^*, D_j^*\right)$ with $\left(D_i^* - \varepsilon, D_j^*+ \varepsilon \right)$, the rate constraint can be improved while keeping the objective in \eqref{eq:PBA_REDUCED_OPT} constant, contradicting the optimality assumption of $\left \lbrace D_i^* \right \rbrace$.
\end{proof}
There is at most one $D_i^*$ such that $D_i^* = \frac{\sigma_i^2}{2(\alpha -1)}\left( 1 + c_i\right)$. Assuming such an $i$ exists, $x_i = \frac{2(\alpha -1)}{1+c_i} < 2(\alpha -1)$. For the convex roots, $x_i = \frac{2(\alpha -1)}{1-c_i} > 2(a-1)$. Therefore from Lemma~\ref{lem:NORM_INC}, all the convex roots are contiguous. Therefore, the set of potentially optimal solutions reduces to cardinality $2d$, where each solution is characterized by the number of components that send non-zero rate and whether or not a concave root is sent. PBA, detailed in Algorithm~\ref{alg:PBA} finds the minimum value of the Lagrangian across these $2d$ solutions for a fixed $\lambda$.
\end{proof}
Note that by sweeping $\lambda > 0$, one can compute the lower convex envelope of the $(D,R)$ curve.
Since every Pareto optimal $(D,R)$ must be a stationary point of (\ref{eq:PBA_OPT_LAGRANGE}),
one can also use Algorithm~\ref{alg:PBA} to compute the $(D,R)$ curve itself by sweeping $\lambda$
and retaining all those stationary points that are not Pareto dominated.
\section{Application to Variable-Rate Compression}\label{sec:VARIABLE}
We have seen that an autoencoder formulation inspired by
data compression succeeds in providing guaranteed
recovery the principal source components. Conversely,
a number of successful multimedia compressors have recently
been proposed that are either related to, or directly
inspired by, autoencoders~\cite{TschannenAL18, TodericiVJHMSC17,
BalleLS16, TodericiOHVMBCS16, TheisSCH17, RippelB17, HabibianRTC19,
AgustssonMTCTBG17, BalleMSHJ18, ZhouCGSW18, AgustssonTMTG19, BalleCMSJAHT20}.
In particular, Ball\'{e} \emph{et al.}~\cite{BalleMSHJ18}
show that the objective minimized by their compressor coincides with
that of variational autoencoders. Following
\cite{BalleCMSJAHT20}, we refer to this objective as \emph{nonlinear
transform coding (NTC)}. We next use Theorem~\ref{thm:diagisoptimal}
to show that
any minimizer of the NTC objective is guaranteed to recover
the principal source components if (1) the source is Gaussian,
(2) the transforms are restricted to be linear,
and (3) the entropy model is \emph{factorized}, as explained below.
Let $\bm{x} \sim {\cal N}\left(0,\bm{K}\right)$, where $\bm{K}$ is a positive semidefinite covariance matrix. As before, we consider an autoencoder defined by its encoder-decoder pair $(f,g)$, where for $k\leq d$, $f: \mathbb{R}^{d} \rightarrow \mathbb{R}^{k}$ and $g:\mathbb{R}^{k} \rightarrow \mathbb{R}^{d}$ are chosen from prespecified classes ${\cal C}_f$ and ${\cal C}_g$. The NTC framework
assumes dithered quantization during training, as in Section~\ref{sec:FRAMEWORK} and \cite{AgustssonT20, ChoiEL19}, and seeks to minimize the Lagrangian
\begin{equation}
\inf_{f \in {\cal C}_f,g \in {\cal C}_g} \mathbb{E}_{\bm{x}, \bm{\eps}} \left[ \left \lVert \bm{x} - g\left( Q\left( f(\bm{x}) + \bm{\eps} \right) - \bm{\eps} \right) \right \rVert_2^2 \right] + \lambda H\left(Q \left( f( \bm{x}) + \bm{\eps}\right) - \bm{\eps} | \bm{\eps} \right).
\end{equation}
where $\lambda > 0$ and $\bm{\eps}$ has i.i.d.\ $\text{Unif}\left[-0.5,0.5\right]$ components.
NTC assumes variable-length compression, and the quantity
$$
H\left(Q \left( f( \bm{x}) + \bm{\eps}\right) - \bm{\eps} | \bm{\eps} \right)
$$
is an accurate estimate of minimum expected codelength length for the discrete
random vector $Q \left( f( \bm{x}) + \bm{\eps}\right)$.
As we noted in Section~\ref{sec:FRAMEWORK}, \cite{ZamirF92} showed that for any random variable $\bm{x}$, $Q\left( \bm{x}+ \bm{\eps} \right) - \bm{\eps}$ and $\bm{x}+\bm{\eps}$ have the same joint distribution with $\bm{x}$. They also showed that $H\left(Q \left( \bm{x} + \bm{\eps}\right) - \bm{\eps} | \bm{\eps} \right) = I\left( \bm{x} + \bm{\eps};\bm{x} \right) = h(\bm{x} + \bm{\eps})$, where
$h(\cdot)$ denotes differential entropy. Therefore, the objective can be written as
\begin{equation} \label{eq:general_lag}
\inf_{f \in {\cal C}_f,g \in {\cal C}_g} \mathbb{E}_{\bm{x}, \bm{\eps}} \left[ \left \lVert \bm{x} - g\left( f\left( \bm{x} \right) + \bm{\eps} \right) \right \rVert_2^2 \right] + \lambda h\left( f\left( \bm{x} \right) + \bm{\eps}\right).
\end{equation}
(Compare eq.(13) in \cite{BalleCMSJAHT20}).
We consider the case in which ${\cal C}_f, {\cal C}_g$ are the class of linear functions. Let $\bm{W}, \bm{T}$ be $d$-by-$d$ matrices. Define $f\left(\bm{x}\right) = \bm{W}^{\top} \bm{x} $, $g\left( \bm{x} \right) = \bm{T} \bm{x}$. Substituting this in the above equation, we obtain
\begin{equation}\label{eq:RD_LAG}
\inf\limits_{\bm{W}, \bm{T}} \mathbb{E}_{\bm{x}, \bm{\eps}} \left[ \left \lVert \bm{x} - \bm{T} \left( \bm{W}^\top \bm{x} + \bm{\eps}\right) \right \rVert_2^2 \right] \\ + \lambda h\left( \bm{W}^\top \bm{x} + \bm{\eps}\right).
\end{equation}
Since $\bm{T}$ does not appear in the rate constraint, the optimal $\bm{T}$ can be chosen to be
the minimum mean squared error estimator of $\bm{x} \sim {\cal N}\left( 0,\bm{K} \right)$ given
$\bm{W}^\top \bm{x} + \bm{\eps}$, as in Section~\ref{sec:FRAMEWORK}. This gives
\begin{equation}\label{eq:RD_LAG_MMSE}
\inf\limits_{\bm{W}} \text{tr}(\bm{K}) - \text{tr}(\bm{K} \bm{W} \left(\bm{W}^{\top} \bm{K} \bm{W} + \frac{\bm{I}}{12} \right)^{-1}\bm{W}^{\top} \bm{K} ) + \lambda h\left( \bm{W}^\top \bm{x} + \bm{\eps}\right).
\end{equation}
As noted earlier, the rate term $h\left( \bm{W}^\top \bm{x} + \bm{\eps}\right)$ is an accurate
estimate for the minimum expected length of the compressed representation of $Q \left( \bm{W}^\top \bm{x}+
\bm{\eps}\right)$.
This assumes that the different components of this vector are encoded jointly, however. In
practice, one often encodes them separately, relying on the transform $\bm{W}$ to
eliminate redundancy among the components. Accordingly, we replace the rate term with
$$
\sum_{i = 1}^d
h\left( \bm{w}_i^\top \bm{x} + [\bm{\eps}]_i\right),
$$
to arrive at the optimization problem
\begin{equation}\label{eq:RD_LAG_MMSE_FACTOR}
\inf\limits_{\bm{W}} \text{tr}(\bm{K}) - \text{tr}(\bm{K} \bm{W} \left(\bm{W}^{\top} \bm{K} \bm{W} + \frac{\bm{I}}{12} \right)^{-1}\bm{W}^{\top} \bm{K} ) +
\lambda \cdot \sum_{i = 1}^d
h\left( \bm{w}_i^\top \bm{x} + [\bm{\eps}]_i\right).
\end{equation}
\begin{theorem}
Suppose $\bm{K}$ has distinct eigenvalues.
Then any $\bm{W}$ that achieves the infimum in~(\ref{eq:RD_LAG_MMSE_FACTOR})
has the property that all of its nonzero rows are eigenvectors of $\bm{K}$.
\label{thm:variable_rate}
\end{theorem}
\begin{proof}
Since the distribution of $\bm{\eps}$ is fixed, by the Gaussian assumption on $\bm{x}$,
$ h\left( \bm{w}_j^\top \bm{x} + [\bm{\eps}]_j\right)$ only depends on
$\bm{w}_j$ through $\bm{w}^{\top}_j \bm{K} \bm{w}_j$. Thus we may write
\begin{equation}
h(\bm{w}_j^{\top} \bm{x} + \epsilon)
= \rho_{sl}(\bm{w}_j^{\top} \bm{K} \bm{w}_j).
\end{equation}
By Theorem~\ref{thm:diagisoptimal}, it suffices to show that $\rho_{sl}(\cdot)$
is strictly concave.
Let $Z$ be a standard Normal random variable and let $\epsilon$
be uniformly distributed over $[-1/2,1/2]$, independent of $Z$.
Then we have
\begin{align}
\rho_{sl}(s)
& = h(\sqrt{s} \cdot Z + \epsilon).
\end{align}
Thus by de Bruijn's identity~\cite{CoverT06},
\begin{equation}
\rho'_{sl}(s) = \frac{1}{2} J(\epsilon + \sqrt{s} \cdot Z),
\end{equation}
where $J(\cdot)$ is the Fisher information. To show that $\rho'_{sl}(\cdot)$
is strictly concave, it suffices to show that $J(\epsilon + \sqrt{s} \cdot Z)$ is strictly
decreasing in $s$.\footnote{If $g'(\cdot)$ is strictly decreasing then for all $t > s$,
$g(t) = g(s) + \int_s^t g'(u) du < g(s) + g'(s)(t-s)$ and likewise for $t < s$.
That $g(\cdot)$ is strictly
concave then follows from the standard first-order test for
concavity~\cite{BoydVandenberghe}.}
To this end, let $t > s > 0$ and let $Z_1$ and $Z_2$
be i.i.d.\ standard Normal random variables, independent of $\epsilon$. Then
\begin{equation}
J(\epsilon + \sqrt{t} \cdot Z) = J(\epsilon + \sqrt{s}\cdot Z_1 + \sqrt{t-s}\cdot Z_2)
\end{equation}
and by the convolution inequality for Fisher information~\cite{Blachman65},
\begin{align}
\frac{1}{J(\epsilon + \sqrt{s} \cdot Z_1 + \sqrt{t - s} \cdot Z_2)} >
\frac{1}{J(\epsilon + \sqrt{s} \cdot Z_1)} + \frac{1}{J(\sqrt{t-s} \cdot Z_2)}> \frac{1}{J(\epsilon + \sqrt{s} \cdot Z_1)},
\end{align}
where the first inequality is strict because $\epsilon + \sqrt{s} \cdot Z_1$
is not Gaussian distributed.
\end{proof}
\subsection{Organization}
\textbf{Notation.} We denote matrices by bold capital letters e.g. $\bm{M}$, and vectors by bold small, e.g. $\bm{v}$. The $j^{\text{th}}$ column of a matrix $\bm{M}$ is denoted by $\bm{m}_j$ and the $j^{\text{th}}$ entry of a vector $\bm{v}$ by $\left[\bm{v}\right]_j$. We denote the set $\left \lbrace 1,2, \cdots d \right \rbrace$ by $\left[ d \right]$. A sequence $a_1, a_2, \cdots a_n$ is denoted by $\{a_i\}_{i=1}^{n}$. We denote the zero column by $\bm{0}$. Logarithms without specified bases denote natural logarithms.
\textbf{Organization.} The balance of the paper is organized as follows. We describe our constrained linear autoencoder framework in Section~\ref{sec:FRAMEWORK}. This results in an optimization problem that we solve for any Schur-concave constraint in Section~\ref{subsec:PROOF}. In Section~\ref{sec:PBAPCA}, we recover linear autoencoders and PBA under our framework. We apply the PBA solution to a problem in variable-rate compression of Gaussian sources in Section~\ref{sec:VARIABLE}. Section~\ref{sec:EXPERIMENTS} contains experiments comparing the performance of the PBA-based fixed-rate compressor against existing fixed-rate linear compressors on image and audio datasets.
\subsection{Related Work}
\textbf{Related Work.} Several recent works have examined how linear autoencoders can be
modified to guarantee recovery of PCA. Most solutions involve eliminating the invariant global optimal solutions by introducing regularization of some kind. \cite{OftadehSWS20} propose a loss function which adds $k$ penalties to recover the $k$ principal directions, each corresponding to recovering up to the first $i \leq k$ principal directions. \cite{KuninBGS19} show that $\ell_2$ regularization helps reduce the symmetry group to the orthogonal group. \cite{BaoLSG20} further break the symmetry by considering non-uniform $\ell_2$ regularization and deterministic dropout. \cite{LadjalNP19} consider a nonlinear autoencoder with a covariance loss term to encourage finding orthogonal directions. Recovering PCA is an important problem even in the stochastic counterpart of autoencoders. \cite{LucasTGN19} analyze linear variational autoencoders (VAEs) and show that the global optimum of its objective is identical to the global optimum of log marginal likelihood of probabilistic PCA (pPCA). \cite{RolinekZM19} analyze an approximation to the VAE loss function and show that the linear approximation to the decoder is orthogonal.
Our result on variable-rate compressors is connected to the sizable recent
literature on compression using autoencoder-like architectures. Representative
contributions to the literature were noted above. Those works focus mostly
on the empirical performance of deep, nonlinear networks, with a particular
emphasis on finding a differentiable proxy for quantization so as to train
with stochastic gradient descent. In contrast, this work considers provable
properties of the compressors when trained perfectly.
\section{Proofs from Section~\ref{sec:PBA}} \label{sec:pfsPBA}
\section{Compression Experiments}\label{sec:EXPERIMENTS}
We validate the PBA algorithm experimentally by comparing the performance
of a PBA-derived fixed-rate compressor against the performance of baseline fixed-rate
compressors. The code of our implementation can be found at \url{https://github.com/SourbhBh/PBA}. As we noted in the previous section, although variable-rate codes are more commonplace in practice, fixed-rate codes do offer some advantages over their more
general counterparts:
\begin{enumerate}
\item In applications where a train of source
realizations are compressed sequentially, fixed-rated
coding allows for simple concatenation of the compressed
representations. Maintaining synchrony between the encoder
and decoder is simpler than with variable-rate codes.
\item In applications where a dataset of source realizations
are individually compressed, fixed-rate coding allows for random access
of data points from the compressed representation.
\item In streaming in which a sequence of realizations will
be streamed, bandwidth provisioning is
simplified when the bit-rate is constant over time.
\end{enumerate}
Fixed-rate compressors exist for specialized sources such as
speech~\cite{McCreeB95, SchroederA85} and audio more
generally~\cite{Vorbis}. We
consider a general-purpose, learned, fixed-rate compressor
derived from PBA and the following two quantization operations.
The first, $Q_{CD}(a,\sigma^2,U,x)$\footnote{``CD'' stands for
``clamped dithered.''}
accepts the hyperparameter $a$, a variance estimate $\sigma^2$, a
dither realization $U$, and the scalar source realization to be
compressed, $x$, and outputs (a binary representation of) the
nearest point to $x$ in the set
\begin{equation}
\label{eq:quantizeset}
\left\{i + U : i \in \mathbb{Z} \ \text{and} \ i + U
\in \left(-\frac{\Gamma}{2},\frac{\Gamma}{2}\right] \right\}\newest{,}
\end{equation}
where
\begin{equation}
\Gamma = 2^{\lfloor \frac{1}{2} \log_2 (4a^2 \sigma^2 + 1) \rfloor }.
\end{equation}
This evidently requires $\log_2 \Gamma$ bits. The second
function, $Q_{CD}'(a^2,\sigma^2,U,b)$, where $b$ is a binary
string of length $\log_2 \Gamma$, maps the binary
representation $b$ to the point in~(\ref{eq:quantizeset}).
These quantization routines are applied separately to each
latent component. The $\sigma^2$ parameters are
determined during training. The dither $U$ is chosen
uniformly over the set $[-1/2,1/2]$, independently for
each component. We assume that $U$ is chosen pseudorandomly
from a fixed seed that is known to both the encoder and
the decoder. As such, it does not need to be explicitly
communicated. For our experiments, we fix the $a$ parameter
at $15$ and hard code this both at the encoder and at
the decoder. We found that this choice balances the
dual goals of minimizing the excess distortion due to
the clamping quantized points to the interval $(\Gamma/2,\Gamma/2]$
and minimizing the rate.
PBA compression proceeds by applying Algorithm~\ref{alg:PBA}
to a training set to determine the matrices $\bm{W}$ and
$\bm{T}$. The variance estimates $\sigma_1^2,\ldots,\sigma_d^2$
for the $d$ latent variances are chosen as the empirical
variances on the training set and are hard-coded in the encoder
and decoder. Given a data point $\bm{x}$,
the encoded representation is the concatenation of the bit strings
$b_1, \ldots, b_d$, where
$$
b_i = Q_{CD}(a^2,\sigma^2_i,U_i,\bm{w}_i^\top x),
$$
The decoder parses the received bits into $b_1, \ldots, b_d$.
and computes the latent reconstruction $\hat\bm{y}$, where
$$
\hat\bm{y}_i = Q_{CD}'(a^2,\sigma^2_i,U_i,b_i),
$$
The reconstruction is then $\bm{T}\hat\bm{y}$.
We evaluate the PBA compressor on MNIST \cite{LecunBBH98}, CIFAR-10 \cite{Krizhevsky09}, MIT Faces Dataset, Free Spoken Digit Dataset (FSDD) \cite{Zohar} and a synthetic Gaussian dataset. The synthetic Gaussian dataset is generated from a diagonal covariance matrix obtained from the eigenvalues of the Faces Dataset.
We compare our algorithms primarily using mean-squared error since our theoretical analysis uses mean squared error as the distortion metric. Our plots display Signal-to-Noise ratios (SNRs) for ease of interpretation. For image datasets, we also compare our algorithms using the Structural Similarity (SSIM) or the Multi-scale Strctural Similarity (MS-SSIM) metrics when applicable \cite{WangBSS04}. We also consider errors on downstream tasks, specifically classification, as a distortion measure.
For all datasets, we compare the performance of the PBA compressor
against baseline scheme derived from PCA that uses $Q_{CD}$
and $Q_{CD}'$.
The PCA-based scheme sends some of the principal components
essentially losslessly, and no information about the others.
Specifically, in the context of our framework,
for any given $k$, we choose the first $k$ columns
of $\bm{W}$ to be aligned with the first $k$ principal components
of the dataset; the remaining columns are zero. Each nonzero column
is scaled such that its Euclidean length multiplied by the eigenvalue has all the significant digits. This is done so that at high rates, the quantization procedure sends the $k$ principal components losslessly. The quantization
and decoder operations are as in the PBA-based scheme; in particular
the $a^2$ parameter is as specified above. By varying $k$, we trade off rate and distortion.
\subsection{SNR Performance}
\begin{figure}[hbt]
\centering
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PCA_1.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PCA_2.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PCA_3.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PCA_4.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PCA_5.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PBA_1.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PBA_2.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PBA_3.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PBA_4.png}
\endminipage\hfill
\minipage{0.18\textwidth}
\includegraphics[width=\linewidth]{images/Reconstructions_PBA_5.png}
\endminipage\hfill
\caption{Reconstructions at different bits/pixel values for PCA (top) and PBA (bottom)}\label{fig:reconstructions_pca_pba}
\end{figure}
\begin{figure}[hbt]
\centering
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_SNR_PBAPCAJPEG.png}
\endminipage\hfill
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_SNR_PBAPCAJPEG.png}
\endminipage\hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/FACES_SNR_PBAPCAJPEG.png}
\endminipage \hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/AUDIO_SNR.png}
\endminipage \hfill
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_SNR_PBAPCAJPEG_UltraZoom.png}
\endminipage\hfill
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_SNR_PBAPCA_UltraZoom.png}
\endminipage\hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/FACES_SNR_PBAPCAJPEG_UltraZoom.png}
\endminipage \hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/AUDIO_SNR_Zoom.png}
\endminipage
\caption{SNR/pixel vs Rate (bits/pixel) for MNIST, CIFAR-10, Faces, FSDD datasets. Figures in the bottom row are zoomed-in. } \label{fig:snr}
\end{figure}
\begin{figure}[hbt]
\centering
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_NQ_SNR_PBAPCA.png}
\endminipage\hfill
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_NQ_SNR_PBAPCA.png}
\endminipage\hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/FACES_NQ_SNR_PBAPCA.png}
\endminipage
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/SYN_PBA_PCA.png}
\endminipage
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_NQ_SNR_PBAPCA_UltraZoom.png}
\endminipage\hfill
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_NQ_SNR_PBAPCA_UltraZoom.png}
\endminipage\hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/FACES_NQ_SNR_PBAPCA_UltraZoom.png}
\endminipage
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/SYN_PBA_ULTRAZOOM.png}
\endminipage
\caption{SNR/pixel vs Rate (bits/pixel) for MNIST, CIFAR-10, Faces and Synthetic dataset. Reconstructions are not rounded to integers from $0$ to $255$. The bottom four plots are zoomed-in versions of the top four plots.}
\label{fig:nq}
\end{figure}
\begin{figure}[hbt]
\centering
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_EIG_LARGEST.png}
\endminipage\hfill
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_EIG_LARGEST.png}
\endminipage\hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/FACES_EIG_LARGEST.png}
\endminipage\hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/AUDIO_EIG_LARGEST.png}
\endminipage
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_EIG_SMALLEST.png}
\endminipage\hfill
\minipage{0.24\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_EIG_SMALLEST.png}
\endminipage\hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/FACES_EIG_SMALLEST.png}
\endminipage\hfill
\minipage{0.24\textwidth}%
\includegraphics[width=\linewidth]{images/AUDIO_EIG_SMALLEST.png}
\endminipage
\caption{Eigenvalue distribution of the datasets. The top three plots are the largest 25 eigenvalues for MNIST, CIFAR-10, Faces and FSDD dataset. The bottom four figures plot the remaining eigenvalues except the largest 500. }
\label{fig:eig}
\end{figure}
\begin{figure}[hbt]
\centering
\minipage{0.32\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_COMP.png}
\endminipage\hfill
\minipage{0.32\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_COMP.png}
\endminipage\hfill
\minipage{0.32\textwidth}%
\includegraphics[width=\linewidth]{images/FACES_COMP.png}
\endminipage
\caption{Plots of number of components sent vs rate (bits/pixel) for PBA and PCA. }\label{fig:num_comp}
\end{figure}
We begin by examining compression performance under
mean squared error, or equivalently, the SNR, defined as
$$
\text{SNR} = 10 \cdot \log_{10}\left(
\frac{P}{\text{MSE}}\right).
$$
where $P$ is the empirical second moment of the dataset.
This was the objective that PBA (and PCA) is designed to minimize.
In Figure~\ref{fig:reconstructions_pca_pba}, we display reconstructions
for a particular image in the Faces Dataset under PBA and PCA.
Figure~\ref{fig:snr} shows the tradeoff for PBA and PCA against
JPEG and JPEG2000 (for the image datasets) and
AAC (for the audio dataset). All of the image datasets
have integer pixel values between 0 and 255. Accordingly, we round
the reconstuctions of PBA and PCA to the nearest integer in this range. Figure~\ref{fig:nq} shows the same tradeoff for PBA and PCA when reconstructions are not rounded off to the nearest integer.
We see that PBA consistently
outperforms PCA and JPEG, and is competitive with JPEG2000, even
though the JPEG and JPEG2000 are variable-rate. \footnote{It should be
noted, however, that JPEG and JPEG2000 aim to minimize subjective
distortion, not MSE, and they
do not allow for training on sample images, as PBA and PCA do. A
similar caveat applies to AAC.} We estimate the size of the JPEG header by compressing an empty image and subtract this estimate from all the compression sizes produced by JPEG. We do not plot JPEG2000 performance for MNIST since it requires at least a 32x32 image. For audio data, we observe that PBA consistently outperforms PCA and AAC.
Since the image data all use $8$ bits per pixel, one can
obtain infinite SNR at this rate via the trivial encoding that
communicates the raw bits. PCA and PBA do not find this solution
because they quantize in the transform domain, where the
lattice-nature of the pixel distribution is not apparent.
Determining how to leverage lattice structure in the source
distribution for purposes of compression is an interesting
question that transcends the PBA and PCA algorithms and that
we will not pursue here.
The reason that PCA performs poorly is that it
favors sending the less significant bits of the most significant
components over the most significant bits of less significant
components, when the latter are more valuable for reconstructing
the source. Arguably, it does not identify the ``principal bits.''
Figure~\ref{fig:eig} shows the eigenvalue distribution of the
different datasets, and Figure~\ref{fig:num_comp} shows the number of distinct
components about which information is sent as a function of rate
for both PBA and PCA. We see that PBA sends information about many
more components for a given rate than does PCA. We discuss the
ramifications of this for downstream tasks, such as classification,
in Section~\ref{subsec:downstream}.
\subsection{SSIM Performance}
Structural similarity (SSIM) and Multi-Scale Structural similarity (MS-SSIM) are metrics that are tuned to perceptual similarity. Given two images, the SSIM metric outputs a real value between $0$ and $1$ where a higher value indicates more similarity between the images. We evaluate the performance of our algorithms on these metrics as well in Figure~\ref{fig:ssim}. We see that PBA
consistently dominates PCA, and although it was not optimized for
this metric, beats JPEG at low rates as well.
\begin{figure}[hbt]
\centering
\minipage{0.32\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_SSIM_PBAPCAJPEG.png}
\endminipage\hfill
\minipage{0.32\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_SSIM_PBAPCAJPEG.png}
\endminipage\hfill
\minipage{0.32\textwidth}%
\includegraphics[width=\linewidth]{images/FACES_SSIM_PBAPCAJPEG.png}
\endminipage
\caption{SSIM vs Rate (bits/pixel) for MNIST, CIFAR-10, Faces Dataset}\label{fig:ssim}
\end{figure}
\subsection{Performance on Downstream tasks}
\label{subsec:downstream}
Lastly, we compare the impact of using PBA and PCA on an important downstream task, namely classification. We evaluate the algorithms on MNIST and CIFAR-10 datasets and use neural networks for classification. Our hyperparameter and architecture choices are given in Table~\ref{tb:ARCHITECTURE}. We divide the dataset into three parts. From the first part, we obain the covariance matrix that we use for PCA and to obtain the PBA compressor. The second and third part are used as training and testing data for the purpose of classification. For a fixed rate, reconstructions are passed to the neural networks for training and testing respectively. Since our goal is to compare classification accuracy across the compressors, we fix both the architecture and hyperparameters, and do not perform any additional tuning for the separate algorithms.
Figure~\ref{fig:accuracy} shows that PBA outperforms PCA in terms of accuracy. The difference is especially significant for low rates; all algorithms attain roughly the same performance at higher rates.
\begin{table}[hbt]
\centering
\begin{tabular}[t]{lccc}
\hline
Hyperparameter&MNIST&CIFAR-10\\
\hline
Architecture & 2-layer fully connected NN & Convolutional Neural Network \\ & & with 2 convolutional layers, pooling and \\ & & three fully connected layers \\
\# Hidden Neurons&100& NA\\
Optimization Algorithm&Adam& SGD with momentum\\
Loss&Cross-entropy&Cross-entropy\\
Learning Rate&0.0005&0.01\\
\hline
\end{tabular}
\caption{Hyperparameter Choices and Architecture for Classification}\label{tb:ARCHITECTURE}
\end{table}%
\begin{figure}[hbt]
\centering
\minipage{0.45\textwidth}
\includegraphics[width=\linewidth]{images/CIFAR_ACCURACY_PBAPCAJPEG.png}
\endminipage\hfill
\minipage{0.45\textwidth}
\includegraphics[width=\linewidth]{images/MNIST_ACCURACY_PBAPCAJPEG.png}
\endminipage
\caption{Accuracy vs Rate (bits/pixel) for MNIST, CIFAR-10}\label{fig:accuracy}
\end{figure}
\section{Review of Schur-Convexity}\label{app:schur}
In this section, we review the key definitions and theorems related to Schur-convexity that we use in the proof of Theorem~\ref{thm:diagisoptimal}.
\begin{definition}{(\textbf{Majorization})}\cite{HornJ85}
For a vector $\bm{v} \in \mathbb{R}^d$, let $\bm{v}^{\downarrow}$ denote the vector with the same components arranged in descending order. Given vectors $\bm{a}, \bm{b} \in \mathbb{R}^{d}$, we say $\bm{a}$ majorizes $\bm{b}$ and denote $\bm{a} \succ \bm{b}$, if
\[\sum\limits_{i=1}^{d} \left[\bm{a} \right]_i = \sum\limits_{i=1}^{d} \left[\bm{b} \right]_i, \]
and for all $k \in \left[ d-1 \right]$,
\[ \sum\limits_{i=1}^{k} \left[\bm{a}^{\downarrow} \right]_i \geq \sum\limits_{i=1}^{k} \left[\bm{b}^{\downarrow} \right]_i.\]
\end{definition}
\begin{definition}{(\textbf{Schur-convexity})}
A function $f: \mathbb{R}^{d} \rightarrow \mathbb{R}$ is Schur-convex if for any vectors $\bm{a}, \bm{b} \in \mathbb{R}^d$, such that $\bm{a} \succ \bm{b}$,
\[ f\left( \bm{a} \right) \geq f\left( \bm{b} \right). \]
$f$ is strictly Schur-convex if the above inequality is a strict inequality for any $\bm{a} \succ \bm{b}$ that are not permutations of each other. $f$ is Schur-concave if the direction of the inequality is reversed and is strictly Schur concave if the direction of the inequality is reversed and it is a strict inequality for any $\bm{a} \succ \bm{b}$ that are not permutations of each other.
\end{definition}
\begin{proposition}\cite{MarshallOA11}\label{prop:single-letter}
If $f:\mathbb{R} \rightarrow \mathbb{R}$ is convex, then $\phi:\mathbb{R}^d \rightarrow \mathbb{R}$ given by
\[ \phi\left( \bm{v} \right) = \sum\limits_{i=1}^{d} f\left(\left[v \right]_i \right) \]
is Schur-convex. If $f$ is concave, then $\phi$ is Schur-concave. Likewise if $f$ is strictly convex, $\phi$ is strictly Schur-convex and if $f$ is strictly concave, $\phi$ is strictly Schur-concave.
\end{proposition}
\section{Explicit Solutions: Conventional Linear Autoencoders and PBA} \label{sec:PBAPCA}
\subsection{Conventional Linear Autoencoders} \label{subsec:PCA}
Given a centered dataset $\bm{x}_1, \bm{x}_2, \cdots, \bm{x}_n \in \mathbb{R}^{d}$, consider a linear autoencoder optimization problem where the encoder and decoder, $\bm{W}$ and $ \bm{T}$, respectively, are $d$-by-$k$ matrices where $k \leq d$ is a parameter. The goal is to minimize the mean squared error as given by \eqref{eq:1stmse}. PCA corresponds to the global optimal solution of this optimization problem, where $\bm{W} = \bm{T} = \bm{U}_k$, where $\bm{U}_k \in \mathbb{R}^{d \times k}$ is a matrix whose columns are the $k$ eigenvectors corresponding to the $k$ largest eigenvalues of $\bm{K}$. However, there are multiple global optimal solutions, given by any encoder-decoder pair of the form $\left( \bm{U}_k \bm{V}, \bm{U}_k \bm{V} \right)$, where $\bm{V}$ is an orthogonal matrix \cite{BaldiH89}.
We now recover linear autoencoders through our framework in Section~\ref{sec:FRAMEWORK}. Consider the optimization problem in \eqref{eq:MASTEROPTI_S} where $\rho_{sl}: \mathbb{R}_{\geq 0} \rightarrow \{0,1 \}$ is a concave function defined as
\begin{equation} \label{eq:PCA_F}
\rho_{sl}(x) = \bm{1} \left[ x > 0 \right].
\end{equation}
Note that this penalizes the dimension of the latents, as desired.
Note also that this cost is Schur-concave but not strictly so.
The fact that PCA solves conventional linear autoencoding,
but is not necessarily the unique solution, follows immediately
from Theorem~\ref{thm:diagisoptimal}.
\begin{theorem}\label{thm:PCA}
If $\rho_{sl}(\cdot)$ is given by \eqref{eq:PCA_F}, then an optimal solution for \eqref{eq:MASTEROPTI_S} is given by a diagonal matrix $\bm{S}$ whose top $\min(\lfloor R \rfloor, d)$ diagonal entries are equal to $\infty$ and the remaining entries are $0$.
\end{theorem}
\begin{proof}
Let ${\cal F} \stackrel{\mathrm{def}}{=} \left \lbrace i \in \left[ d \right] : s_i > 0 \right \rbrace$, implying $\left \lvert {\cal F} \right \rvert \leq R$. Since $\bm{K}$ and $\bm{S}$ are diagonal, the optimization problem in \eqref{eq:MASTEROPTI_S} can be written as
\begin{equation} \label{eq:PCA_OPT}
\begin{aligned}
\inf\limits_{\left \lbrace s_{\ell} \right \rbrace} & \sum\limits_{j \in \left[d\right] \backslash {\cal F}} \sigma_j^2 + \sum\limits_{ \ell \in {\cal F}} \frac{\sigma^2 \sigma_{\ell}^2}{\sigma^2+ \sigma_{\ell}^2 s_{\ell}^2} \\
\text{subject to} \quad & R \geq \sum\limits_{i=1}^{d} \bm{1}\left[ s_i > 0 \right].
\end{aligned}
\end{equation}
Since the value of $s_{\ell}, \ell \in {\cal F}$ does not affect the rate constraint, each of the $s_{\ell}$ can be made as large as possible without changing the rate constraint. Therefore, the infimum value of the objective is $\sum\limits_{j \in \left[ d \right] \backslash {\cal F}} \sigma^2_j$. Since we seek to minimize the distortion, the optimal ${\cal F}$ is the set of indices with the largest $\left \lvert {\cal F} \right \rvert$ eigenvalues. Since the number of these eigenvalues cannot exceed $R$, we choose $\left \lvert {\cal F} \right \rvert = \min(\lfloor R \rfloor,d)$.
\end{proof}
Unlike the conventional linear autoencoder framework, in Section~\ref{sec:FRAMEWORK}, the latent variables $\bm{W}^{\top} \bm{x}$ are quantized, which we model with additive white noise of fixed variance. Therefore, an infinite value of $s_i$ indicates sending $\bm{u}_i^{\top} \bm{x}$ with full precision where $\bm{u}_i$ is the eigenvector corresponding to the $i^{th}$ largest eigenvalue. This implies that PCA with parameter $k$ corresponds to $\bm{W} = \bm{U} \bm{S}$, where $\bm{S}$ is a diagonal matrix whose top $k$ diagonal entries are equal to $\infty$ and the $d-k$ remaining diagonal entries are $0$. Therefore, for any $R$ such that $\lfloor R \rfloor = k$, an optimal solution to \eqref{eq:MASTEROPTI_S} corresponds to linearly projecting the data along the top $k$ eigenvectors, which is the same as PCA. Note that, like \cite{BaldiH89}, we only prove that projecting along the eigenvectors is one of possibly other optimal solutions. However, even a slight amount of curvature in $\rho$ would make it strictly Schur-concave, thus recovering the principal directions.
We next turn to a specific cost function with curvature, namely
the PBA cost function that was our original motivation.
\subsection{Principal Bit Analysis (PBA)}
Consider the choice of $\rho_{sl}:\mathbb{R}_{\geq 0} \rightarrow \mathbb{R}_{\geq 0}$ that provided the
original impetus for Theorem~\ref{thm:diagisoptimal}. For $\gamma>\frac{2}{\sigma^2}$,
\begin{equation}\label{eq:LOG_F}
\rho_{sl}(x) = \frac{1}{2} \log(\gamma x+1).
\end{equation}
The nature of the optimization problem depends on the value
of $\gamma$. For $1 \le \gamma \sigma^2 \le 2$, the problem can be
made convex with a simple change of variable. For $\gamma \sigma^2= 1$,
the problem coincides with the classical waterfilling procedure
in rate-distortion theory, in fact. For $\gamma \sigma^2 > 2$, the problem
is significantly more challenging. Since we are interested in
relatively large values of $\gamma$ for our compression application
(see Section~\ref{sec:EXPERIMENTS} to follow), we focus on the case $\gamma > 2/\sigma^2$.
\begin{algorithm}[h]
\caption{Principal Bit Analysis (PBA)} \label{alg:PBA}
\begin{algorithmic}[1]
\Require $\lambda > 0$, $\alpha > 2$,
\begin{equation}
\bm{K} = \left[ \begin{array}{cccc}
\sigma^2_1 & 0 & \cdots & 0 \\
0 & \sigma^2_2 & \cdots & 0 \\
\vdots & \vdots & \ddots & \vdots \\
0 & 0 & \cdots & \sigma^2_d \\
\end{array} \right] \succ 0,
\end{equation}
such that $\sigma^2_1 \ge \sigma^2_2 \ge \cdots \ge \sigma^2_d$.
\State If $\lambda \ge \sigma^2_1/(4(\alpha - 1))$, Output
$\bar{R}_{\text{opt}} = 0, \bar{D}_{\text{opt}} = \sum_{i = 1}^d \sigma^2_i$.
\State Set $\bar{d} = \max\left\{i : \lambda < \sigma_i^2/4(\alpha - 1)\right\}$.
\State Set $\bar{R}$, $\bar{D}$ to zero arrays of size $2\bar{d}$.
\For{$r \in \left \lbrace 1,2, \cdots \bar{d} \right \rbrace $}
\State $\bar{D}(2r-1) = \sum\limits_{i=1}^{r} \frac{\sigma_i^2}{2(\alpha -1)} \left( 1 - \sqrt{1 - \frac{4 \lambda (\alpha -1)}{\sigma_i^2}}\right) + \sum\limits_{i=r+1}^{d} \frac{\sigma_i^2}{\alpha } $,
\State $\bar{R}(2r-1)= \sum\limits_{i=1}^{r} \frac{1}{2} \log \left( \frac{\sigma_i^2}{4\lambda} \right) + \log\left( 1 + \sqrt{1 - \frac{4 \lambda (\alpha -1)}{\sigma_i^2}}\right)$.
\State $\bar{D}(2r) =\left( \sum\limits_{i=1}^{r-1} \frac{\sigma_i^2}{2(\alpha -1)} \left( 1 - \sqrt{1 - \frac{4 \lambda (\alpha -1)}{\sigma_i^2}}\right) + \frac{\sigma_{r}^2}{2(\alpha -1)} \left( 1 + \sqrt{1 - \frac{4 \lambda (\alpha -1)}{\sigma_{r}^2}}\right) + \sum\limits_{i=r+1}^{d} \frac{\sigma_i^2}{\alpha } \right)$.
\State $\bar{R}(2r)=\sum\limits_{i=1}^{r} \frac{1}{2} \log \left( \frac{\sigma_i^2}{4\lambda} \right) + \sum\limits_{i=1}^{r-1} \log\left( 1 + \sqrt{1 - \frac{4 \lambda (\alpha -1)}{\sigma_i^2}}\right) + \log\left( 1 - \sqrt{1 - \frac{4 \lambda (\alpha -1)}{\sigma_{r}^2}}\right)$.
\EndFor
\State $r^* \leftarrow \text{argmin}_{j \in \left[ 2\bar{d} \right]} \bar{D}(j) + \lambda \bar{R}(j). $
\State Output $\bar{R}_{\text{opt}} = \bar{R}(r^*), \bar{D}_{\text{opt}} = \bar{D}(r^*)$.
\end{algorithmic}
\end{algorithm}
\begin{theorem}
If $\rho_{sl}(\cdot)$ is given by \eqref{eq:LOG_F} for $\gamma > \frac{2}{\sigma^2}$, then for any $\lambda > 0$,
the pair $\bar{R}_{\text{opt}}, \bar{D}_{\text{opt}}$ obtained from the output of
Algorithm~\ref{alg:PBA} satisfies
\begin{equation}
\label{eq:PBA_OPT_LAGRANGE}
\bar{D}_{\text{opt}} + \lambda \bar{R}_{\text{opt}} =
\inf\limits_{\bm{S}} \quad \text{tr}(\bm{K}) - \text{tr}(\bm{K} \bm{S} (\bm{S}^{\top} \bm{K} \bm{S} + \sigma^2\bm{I} )^{-1}\bm{S}^{\top} \bm{K} )
+ \lambda \sum_{i = 1}^d \rho_{sl}\left( \{s_i^2 \sigma^2_i \} \right),
\end{equation}
\end{theorem}
\begin{proof}
Since $\bm{K}$ and $\bm{S}$ are diagonal, the optimization problem in
\eqref{eq:PBA_OPT_LAGRANGE} can be written as
\begin{equation} \label{eq:PBA_OPT}
\begin{aligned}
\inf\limits_{\left \lbrace s_i \right \rbrace} \quad & \sum\limits_{i=1}^{d} \frac{\sigma^2 \sigma_{i}^{2}}{\sigma^2+s_{i}^{2} \sigma_{i}^{2}}
+ \lambda \cdot \frac{1}{2}\sum\limits_{i=1}^{d} \log\left( 1 + \gamma s_{i}^2 \sigma_{i}^{2} \right).
\end{aligned}
\end{equation}
With the following change of variables $ \alpha = \gamma \sigma^2 $, $s_{i} \mapsto s_{i}'^2 = \alpha \frac{s_{i}'^2}{\sigma^2}$, we obtain
\begin{equation} \label{eq:PBA_OPT_p1}
\begin{aligned}
\inf\limits_{\left \lbrace s_i' \right \rbrace} \quad & \sum\limits_{i=1}^{d} \alpha \frac{ \sigma_{i}^{2}}{\alpha+s_{i}^{'2} \sigma_{i}^{2}} + \lambda \cdot \frac{1}{2}\sum\limits_{i=1}^{d} \log\left( 1 + s_i'^2 \sigma_{i}^{2} \right).
\end{aligned}
\end{equation}
Ignoring the constant factor in the objective, perform the change of variable $s'_{i} \mapsto D_i = \frac{\sigma_i^2}{\alpha +s_{i}'^2\sigma_{i}^{2}}$
to obtain
\begin{equation} \label{eq:PBA_REDUCED_OPT}
\begin{aligned}
\inf\limits_{\left \lbrace D_i \right \rbrace} \quad & \sum\limits_{i=1}^{d} D_i + \frac{\lambda}{2}\sum\limits_{i=1}^{d} \log\left( \frac{\sigma_i^2}{D_i} - (\alpha -1) \right), \\
\text{subject to} \quad &D_i \leq \frac{\sigma_i^2}{\alpha } \hspace{5mm} \text{ for all } i \in \left[ d \right].
\end{aligned}
\end{equation}
This optimization problem is nonconvex since the function $\log\left( \frac{\sigma_i^2}{D_i} - (\alpha -1) \right)$ is convex for $0 \leq D_i \leq \frac{\sigma_i^2}{2(\alpha -1)}$ but concave for $\frac{\sigma_i^2}{2(\alpha -1)} < D_i \leq \frac{\sigma_i^2}{\alpha}$ and the
latter interval is nonempty since $\alpha>2$.
Any optimizing $\{D_i\}$ must be a stationary point of
\begin{equation}\label{eq:LAG_PBA}
\cL\left( \left \lbrace D_i \right \rbrace_{i=1}^{d}, \lambda, \left \lbrace \mu_i \right \rbrace_{i=1}^{d} \right) = \sum\limits_{i=1}^{d} D_i + \lambda\left( \sum\limits_{i=1}^{d} \log\left( \frac{\sigma_i^2}{D_i} - (\alpha -1) \right) \right) + \sum\limits_{i=1}^{d} \mu_i \left( D_i - \frac{\sigma_i^2}{\alpha } \right).
\end{equation}
for some $\{\mu_i\}_{i = 1}^d$ with $\mu_i \geq 0$ for all $i \in \left[ d \right]$ and satisfying the complementary slackness condition~\cite[Prop.~3.3.1]{Bertsekas:Nonlinear}. The stationary points satisfy, for each $i$
\begin{equation}\label{eq:LAG_STAT}
\frac{\partial \cL}{\partial D_i} = 1 - \lambda\left( \frac{\frac{\sigma_i^2}{D_i^2}}{\left( \frac{\sigma_i^2}{D_i} - (\alpha -1) \right)}\right) + \mu_i = 0.
\end{equation}
Let ${\cal F} = \left \lbrace i : D_i < \frac{\sigma_i^2}{\alpha } \right \rbrace$.
For $i \in {\cal F}$, $\mu_i = 0$ due to complementary slackness. Substituting in \eqref{eq:LAG_STAT} we obtain a quadratic equation in $D_i$
\[
(\alpha-1)D_i^2 - \sigma_i^2 D_i + \lambda \sigma_i^2 = 0.
\]
which gives
$$ D_i = \frac{\sigma_i^2}{2(\alpha -1)} \left( 1 \pm \sqrt{1 - \frac{4\lambda (\alpha -1)}{\sigma_i^2}} \right).$$
Let $c_i = \sqrt{1 - \frac{4\lambda (\alpha -1)}{\sigma_i^2}}$. Note that $\frac{\sigma_i^2}{2(\alpha -1)} \left( 1 + c_i \right)$ is always in the concave region and $\frac{\sigma_i^2}{2(\alpha -1)} \left( 1 - c_i \right)$ is always in the convex region for a $\lambda$ chosen such that $D_i$ is a real number strictly less than $\frac{\sigma_i^2}{\alpha }$.
Therefore the optimal set of distortions are contained in the following set of $3^{d}$ points
$$ \prod_{i = 1}^d \left\{ \frac{\sigma_i^2}{2(\alpha -1)} \left( 1 + \sqrt{1 - \frac{4\lambda (\alpha -1)}{\sigma_i^2}} \right), \frac{\sigma_i^2}{2(\alpha -1)} \left( 1 - \sqrt{1 - \frac{4\lambda (\alpha -1)}{\sigma_i^2}} \right), \frac{\sigma_i^2}{\alpha}\right\}.$$
We now reduce the size of the above set by making a two observations:
\noindent \textbf{(1). ${\cal F}$ is contiguous.}
\begin{lemma}\label{lem:NORM_INC}
There exists an optimal $\left \lbrace D_i^* \right \rbrace_{i=1}^{d}$ for \eqref{eq:PBA_REDUCED_OPT} such that (a) $\frac{\sigma_i^2}{D_i^*}$ is a nonincreasing sequence and (b) ${\cal F} = \{ 1,2,\cdots \left \lvert {\cal F} \right \rvert \}$.
\end{lemma}
\begin{proof}
Substitute $x_i = \frac{\sigma_i^2}{D_i}$ in \eqref{eq:PBA_REDUCED_OPT}. This gives us
\begin{equation} \label{eq:PBA_NORM_OPT}
\begin{aligned}
\inf\limits_{\left \lbrace x_i \right \rbrace} \quad & \sum\limits_{i=1}^{d} \frac{\sigma_i^2}{x_i} + \frac{\lambda}{2}\sum\limits_{i=1}^{d} \log\left( x_i - (\alpha -1) \right), \\
\text{subject to} \quad & x_i \geq \alpha \hspace{5mm} \text{ for all } i \in \left[ d \right].
\end{aligned}
\end{equation}
Let $\{ x_i^* \}_{i=1}^{d}$ be an optimal solution for \eqref{eq:PBA_NORM_OPT}. If, for $i>j$, $x_i^* > x_j^* \ge \alpha$, then exchanging the values provides a solution that has the same rate and lower distortion since $\frac{\sigma_i^2}{x_i^*} + \frac{\sigma_j^2}{x_j^*} \geq \frac{\sigma_i^2}{x_j^*} + \frac{\sigma_j^2}{x_i^*}$. This proves (a). Part (b) follows immediately.
\end{proof}
\noindent\textbf{(2). No two solutions are concave.}
\begin{lemma}\label{lem:SINGLE_CONCAVE}
For $R>0$, let $\left \lbrace D_i^* \right \rbrace_{i=1}^{d}$ be an optimal solution for \eqref{eq:PBA_REDUCED_OPT}. There exists at most one $D_i^*$ such that $\frac{\sigma_i^2}{2(\alpha -1)} < D_i^* < \frac{\sigma_i^2}{\alpha }$.
\end{lemma}
\begin{proof}
Let $D_i^*, D_j^*$ be such that $\frac{\sigma_i^2}{2(\alpha -1)} < D_i^* < \frac{\sigma_i^2}{\alpha }$ and $\frac{\sigma_j^2}{2(\alpha -1)} < D_j^* < \frac{\sigma_j^2}{\alpha }$ . Without loss of generality, assume $D_i^* < D_j^*$. Denote the individual rate constraint function by $r\left( D_i \right) \triangleq \log \left( \frac{\sigma_i^2}{D_i} - (\alpha -1) \right)$. Since $r$ is concave in $\left(\frac{\sigma_i^2}{2(a-1)}, \frac{\sigma_i^2}{\alpha } \right)$, there exist an $\varepsilon > 0$ such that
\begin{align}
r\left(D_i^* - \varepsilon\right) +r\left( D_j^* + \varepsilon \right) &= r\left( D_i^* \right) - \varepsilon r'\left( D_i^* \right) + O(\varepsilon^2) + r\left(D_j^*\right) + \varepsilon r'\left(D_j^* \right) + O(\varepsilon^2) \\
&< r\left( D_i^* \right) + r\left(D_j^*\right)
\end{align}
The last inequality follows from concavity of $r$. Therefore, replacing $\left(D_i^*, D_j^*\right)$ with $\left(D_i^* - \varepsilon, D_j^*+ \varepsilon \right)$, the rate constraint can be improved while keeping the objective in \eqref{eq:PBA_REDUCED_OPT} constant, contradicting the optimality assumption of $\left \lbrace D_i^* \right \rbrace$.
\end{proof}
There is at most one $D_i^*$ such that $D_i^* = \frac{\sigma_i^2}{2(\alpha -1)}\left( 1 + c_i\right)$. Assuming such an $i$ exists, $x_i = \frac{2(\alpha -1)}{1+c_i} < 2(\alpha -1)$. For the convex roots, $x_i = \frac{2(\alpha -1)}{1-c_i} > 2(a-1)$. Therefore from Lemma~\ref{lem:NORM_INC}, all the convex roots are contiguous. Therefore, the set of potentially optimal solutions reduces to cardinality $2d$, where each solution is characterized by the number of components that send non-zero rate and whether or not a concave root is sent. PBA, detailed in Algorithm~\ref{alg:PBA} finds the minimum value of the Lagrangian across these $2d$ solutions for a fixed $\lambda$.
\end{proof}
Note that by sweeping $\lambda > 0$, one can compute the lower convex envelope of the $(D,R)$ curve.
Since every Pareto optimal $(D,R)$ must be a stationary point of (\ref{eq:PBA_OPT_LAGRANGE}),
one can also use Algorithm~\ref{alg:PBA} to compute the $(D,R)$ curve itself by sweeping $\lambda$
and retaining all those stationary points that are not Pareto dominated.
\section{Proofs from Section~\ref{sec:PBA}} \label{sec:pfsPBA}
\section{Conclusion}
\subsection{Related Work}
\textbf{Related Work.} Several recent works have examined how linear autoencoders can be
modified to guarantee recovery of PCA. Most solutions involve eliminating the invariant global optimal solutions by introducing regularization of some kind. \cite{OftadehSWS20} propose a loss function which adds $k$ penalties to recover the $k$ principal directions, each corresponding to recovering up to the first $i \leq k$ principal directions. \cite{KuninBGS19} show that $\ell_2$ regularization helps reduce the symmetry group to the orthogonal group. \cite{BaoLSG20} further break the symmetry by considering non-uniform $\ell_2$ regularization and deterministic dropout. \cite{LadjalNP19} consider a nonlinear autoencoder with a covariance loss term to encourage finding orthogonal directions. Recovering PCA is an important problem even in the stochastic counterpart of autoencoders. \cite{LucasTGN19} analyze linear variational autoencoders (VAEs) and show that the global optimum of its objective is identical to the global optimum of log marginal likelihood of probabilistic PCA (pPCA). \cite{RolinekZM19} analyze an approximation to the VAE loss function and show that the linear approximation to the decoder is orthogonal.
Our result on variable-rate compressors is connected to the sizable recent
literature on compression using autoencoder-like architectures. Representative
contributions to the literature were noted above. Those works focus mostly
on the empirical performance of deep, nonlinear networks, with a particular
emphasis on finding a differentiable proxy for quantization so as to train
with stochastic gradient descent. In contrast, this work considers provable
properties of the compressors when trained perfectly.
\section{Review of Schur-Convexity}\label{app:schur}
In this section, we review the key definitions and theorems related to Schur-convexity that we use in the proof of Theorem~\ref{thm:diagisoptimal}.
\begin{definition}{(\textbf{Majorization})}\cite{HornJ85}
For a vector $\bm{v} \in \mathbb{R}^d$, let $\bm{v}^{\downarrow}$ denote the vector with the same components arranged in descending order. Given vectors $\bm{a}, \bm{b} \in \mathbb{R}^{d}$, we say $\bm{a}$ majorizes $\bm{b}$ and denote $\bm{a} \succ \bm{b}$, if
\[\sum\limits_{i=1}^{d} \left[\bm{a} \right]_i = \sum\limits_{i=1}^{d} \left[\bm{b} \right]_i, \]
and for all $k \in \left[ d-1 \right]$,
\[ \sum\limits_{i=1}^{k} \left[\bm{a}^{\downarrow} \right]_i \geq \sum\limits_{i=1}^{k} \left[\bm{b}^{\downarrow} \right]_i.\]
\end{definition}
\begin{definition}{(\textbf{Schur-convexity})}
A function $f: \mathbb{R}^{d} \rightarrow \mathbb{R}$ is Schur-convex if for any vectors $\bm{a}, \bm{b} \in \mathbb{R}^d$, such that $\bm{a} \succ \bm{b}$,
\[ f\left( \bm{a} \right) \geq f\left( \bm{b} \right). \]
$f$ is strictly Schur-convex if the above inequality is a strict inequality for any $\bm{a} \succ \bm{b}$ that are not permutations of each other. $f$ is Schur-concave if the direction of the inequality is reversed and is strictly Schur concave if the direction of the inequality is reversed and it is a strict inequality for any $\bm{a} \succ \bm{b}$ that are not permutations of each other.
\end{definition}
\begin{proposition}\cite{MarshallOA11}\label{prop:single-letter}
If $f:\mathbb{R} \rightarrow \mathbb{R}$ is convex, then $\phi:\mathbb{R}^d \rightarrow \mathbb{R}$ given by
\[ \phi\left( \bm{v} \right) = \sum\limits_{i=1}^{d} f\left(\left[v \right]_i \right) \]
is Schur-convex. If $f$ is concave, then $\phi$ is Schur-concave. Likewise if $f$ is strictly convex, $\phi$ is strictly Schur-convex and if $f$ is strictly concave, $\phi$ is strictly Schur-concave.
\end{proposition}
\section{Application to Variable-Rate Compression}\label{sec:VARIABLE}
We have seen that an autoencoder formulation inspired by
data compression succeeds in providing guaranteed
recovery the principal source components. Conversely,
a number of successful multimedia compressors have recently
been proposed that are either related to, or directly
inspired by, autoencoders~\cite{TschannenAL18, TodericiVJHMSC17,
BalleLS16, TodericiOHVMBCS16, TheisSCH17, RippelB17, HabibianRTC19,
AgustssonMTCTBG17, BalleMSHJ18, ZhouCGSW18, AgustssonTMTG19, BalleCMSJAHT20}.
In particular, Ball\'{e} \emph{et al.}~\cite{BalleMSHJ18}
show that the objective minimized by their compressor coincides with
that of variational autoencoders. Following
\cite{BalleCMSJAHT20}, we refer to this objective as \emph{nonlinear
transform coding (NTC)}. We next use Theorem~\ref{thm:diagisoptimal}
to show that
any minimizer of the NTC objective is guaranteed to recover
the principal source components if (1) the source is Gaussian,
(2) the transforms are restricted to be linear,
and (3) the entropy model is \emph{factorized}, as explained below.
Let $\bm{x} \sim {\cal N}\left(0,\bm{K}\right)$, where $\bm{K}$ is a positive semidefinite covariance matrix. As before, we consider an autoencoder defined by its encoder-decoder pair $(f,g)$, where for $k\leq d$, $f: \mathbb{R}^{d} \rightarrow \mathbb{R}^{k}$ and $g:\mathbb{R}^{k} \rightarrow \mathbb{R}^{d}$ are chosen from prespecified classes ${\cal C}_f$ and ${\cal C}_g$. The NTC framework
assumes dithered quantization during training, as in Section~\ref{sec:FRAMEWORK} and \cite{AgustssonT20, ChoiEL19}, and seeks to minimize the Lagrangian
\begin{equation}
\inf_{f \in {\cal C}_f,g \in {\cal C}_g} \mathbb{E}_{\bm{x}, \bm{\eps}} \left[ \left \lVert \bm{x} - g\left( Q\left( f(\bm{x}) + \bm{\eps} \right) - \bm{\eps} \right) \right \rVert_2^2 \right] + \lambda H\left(Q \left( f( \bm{x}) + \bm{\eps}\right) - \bm{\eps} | \bm{\eps} \right).
\end{equation}
where $\lambda > 0$ and $\bm{\eps}$ has i.i.d.\ $\text{Unif}\left[-0.5,0.5\right]$ components.
NTC assumes variable-length compression, and the quantity
$$
H\left(Q \left( f( \bm{x}) + \bm{\eps}\right) - \bm{\eps} | \bm{\eps} \right)
$$
is an accurate estimate of minimum expected codelength length for the discrete
random vector $Q \left( f( \bm{x}) + \bm{\eps}\right)$.
As we noted in Section~\ref{sec:FRAMEWORK}, \cite{ZamirF92} showed that for any random variable $\bm{x}$, $Q\left( \bm{x}+ \bm{\eps} \right) - \bm{\eps}$ and $\bm{x}+\bm{\eps}$ have the same joint distribution with $\bm{x}$. They also showed that $H\left(Q \left( \bm{x} + \bm{\eps}\right) - \bm{\eps} | \bm{\eps} \right) = I\left( \bm{x} + \bm{\eps};\bm{x} \right) = h(\bm{x} + \bm{\eps})$, where
$h(\cdot)$ denotes differential entropy. Therefore, the objective can be written as
\begin{equation} \label{eq:general_lag}
\inf_{f \in {\cal C}_f,g \in {\cal C}_g} \mathbb{E}_{\bm{x}, \bm{\eps}} \left[ \left \lVert \bm{x} - g\left( f\left( \bm{x} \right) + \bm{\eps} \right) \right \rVert_2^2 \right] + \lambda h\left( f\left( \bm{x} \right) + \bm{\eps}\right).
\end{equation}
(Compare eq.(13) in \cite{BalleCMSJAHT20}).
We consider the case in which ${\cal C}_f, {\cal C}_g$ are the class of linear functions. Let $\bm{W}, \bm{T}$ be $d$-by-$d$ matrices. Define $f\left(\bm{x}\right) = \bm{W}^{\top} \bm{x} $, $g\left( \bm{x} \right) = \bm{T} \bm{x}$. Substituting this in the above equation, we obtain
\begin{equation}\label{eq:RD_LAG}
\inf\limits_{\bm{W}, \bm{T}} \mathbb{E}_{\bm{x}, \bm{\eps}} \left[ \left \lVert \bm{x} - \bm{T} \left( \bm{W}^\top \bm{x} + \bm{\eps}\right) \right \rVert_2^2 \right] \\ + \lambda h\left( \bm{W}^\top \bm{x} + \bm{\eps}\right).
\end{equation}
Since $\bm{T}$ does not appear in the rate constraint, the optimal $\bm{T}$ can be chosen to be
the minimum mean squared error estimator of $\bm{x} \sim {\cal N}\left( 0,\bm{K} \right)$ given
$\bm{W}^\top \bm{x} + \bm{\eps}$, as in Section~\ref{sec:FRAMEWORK}. This gives
\begin{equation}\label{eq:RD_LAG_MMSE}
\inf\limits_{\bm{W}} \text{tr}(\bm{K}) - \text{tr}(\bm{K} \bm{W} \left(\bm{W}^{\top} \bm{K} \bm{W} + \frac{\bm{I}}{12} \right)^{-1}\bm{W}^{\top} \bm{K} ) + \lambda h\left( \bm{W}^\top \bm{x} + \bm{\eps}\right).
\end{equation}
As noted earlier, the rate term $h\left( \bm{W}^\top \bm{x} + \bm{\eps}\right)$ is an accurate
estimate for the minimum expected length of the compressed representation of $Q \left( \bm{W}^\top \bm{x}+
\bm{\eps}\right)$.
This assumes that the different components of this vector are encoded jointly, however. In
practice, one often encodes them separately, relying on the transform $\bm{W}$ to
eliminate redundancy among the components. Accordingly, we replace the rate term with
$$
\sum_{i = 1}^d
h\left( \bm{w}_i^\top \bm{x} + [\bm{\eps}]_i\right),
$$
to arrive at the optimization problem
\begin{equation}\label{eq:RD_LAG_MMSE_FACTOR}
\inf\limits_{\bm{W}} \text{tr}(\bm{K}) - \text{tr}(\bm{K} \bm{W} \left(\bm{W}^{\top} \bm{K} \bm{W} + \frac{\bm{I}}{12} \right)^{-1}\bm{W}^{\top} \bm{K} ) +
\lambda \cdot \sum_{i = 1}^d
h\left( \bm{w}_i^\top \bm{x} + [\bm{\eps}]_i\right).
\end{equation}
\begin{theorem}
Suppose $\bm{K}$ has distinct eigenvalues.
Then any $\bm{W}$ that achieves the infimum in~(\ref{eq:RD_LAG_MMSE_FACTOR})
has the property that all of its nonzero rows are eigenvectors of $\bm{K}$.
\label{thm:variable_rate}
\end{theorem}
\begin{proof}
Since the distribution of $\bm{\eps}$ is fixed, by the Gaussian assumption on $\bm{x}$,
$ h\left( \bm{w}_j^\top \bm{x} + [\bm{\eps}]_j\right)$ only depends on
$\bm{w}_j$ through $\bm{w}^{\top}_j \bm{K} \bm{w}_j$. Thus we may write
\begin{equation}
h(\bm{w}_j^{\top} \bm{x} + \epsilon)
= \rho_{sl}(\bm{w}_j^{\top} \bm{K} \bm{w}_j).
\end{equation}
By Theorem~\ref{thm:diagisoptimal}, it suffices to show that $\rho_{sl}(\cdot)$
is strictly concave.
Let $Z$ be a standard Normal random variable and let $\epsilon$
be uniformly distributed over $[-1/2,1/2]$, independent of $Z$.
Then we have
\begin{align}
\rho_{sl}(s)
& = h(\sqrt{s} \cdot Z + \epsilon).
\end{align}
Thus by de Bruijn's identity~\cite{CoverT06},
\begin{equation}
\rho'_{sl}(s) = \frac{1}{2} J(\epsilon + \sqrt{s} \cdot Z),
\end{equation}
where $J(\cdot)$ is the Fisher information. To show that $\rho'_{sl}(\cdot)$
is strictly concave, it suffices to show that $J(\epsilon + \sqrt{s} \cdot Z)$ is strictly
decreasing in $s$.\footnote{If $g'(\cdot)$ is strictly decreasing then for all $t > s$,
$g(t) = g(s) + \int_s^t g'(u) du < g(s) + g'(s)(t-s)$ and likewise for $t < s$.
That $g(\cdot)$ is strictly
concave then follows from the standard first-order test for
concavity~\cite{BoydVandenberghe}.}
To this end, let $t > s > 0$ and let $Z_1$ and $Z_2$
be i.i.d.\ standard Normal random variables, independent of $\epsilon$. Then
\begin{equation}
J(\epsilon + \sqrt{t} \cdot Z) = J(\epsilon + \sqrt{s}\cdot Z_1 + \sqrt{t-s}\cdot Z_2)
\end{equation}
and by the convolution inequality for Fisher information~\cite{Blachman65},
\begin{align}
\frac{1}{J(\epsilon + \sqrt{s} \cdot Z_1 + \sqrt{t - s} \cdot Z_2)} >
\frac{1}{J(\epsilon + \sqrt{s} \cdot Z_1)} + \frac{1}{J(\sqrt{t-s} \cdot Z_2)}> \frac{1}{J(\epsilon + \sqrt{s} \cdot Z_1)},
\end{align}
where the first inequality is strict because $\epsilon + \sqrt{s} \cdot Z_1$
is not Gaussian distributed.
\end{proof}
|
\section{Introduction}
ASR models are typically designed to assume that all input sources to the model are always available for each sample. The inputs could be acoustic data from a single audio channel, from multiple channels, or acoustic data combined with context vector embeddings from prior utterances in a conversational setting \cite{rnnt, minhua2019, Kim2019}. This design choice prevents the ASR model to accept additional input sources that contain useful information but are only \textit{occasionally} available. For instance, consider a classroom scenario where there is a central listener device with a microphone array and an additional lapel microphone that is occasionally used by the teacher. The input to the ASR could be a single audio channel that comes from an on-device beamformer in the central listener. If the upload bandwidth allows it, additional raw microphone channels (auxiliary inputs) could also be streamed. If the lapel is used, another auxiliary source is available. One would need to build separate ASR models for these scenarios with their own datasets.
In this work, we propose a unified model that can serve all these predefined scenarios. The unified model has separate frontends for each scenario (Sec. \ref{sec:architecture}) and employs a unique training methodology that combines datasets with different number of sources (Sec. \ref{sec:training}). In the rest of this paper, we present the results of such a unified model in a far-field ASR scenario: a wide variety of devices stream single-channel (SC) audio, but a subset of them might conditionally stream additional raw audio from microphones to create a multi-channel (MC) input. The unified model, coupled with the proposed training methodology, leads to lower WER than building an MC-only model that can only be trained with MC audio (Sec. \ref{sec:results}).
\begin{figure}[t!]
\centering
\includegraphics[width=0.85\linewidth]{./images/Architecturev2.png}
\caption{A Unified ASR model architecture. Separate frontends for each input type (primary and primary+auxiliary) share a backend, enabling a single model that serves both data types.}
\label{fig:architecture}
\end{figure}
\section{Related Work}
\label{sec:related-work}
Far-field ASR systems are designed to operate in more challenging acoustic conditions compared to a near-field system where the speaker is close to the microphone. Lower signal-to-noise ratio (SNR) in the received signal in the microphones reduces the word error rate (WER) of the following ASR system. Increased signal degradation with distance, room reverberation, noise, and background speech contribute to this reduction \cite{umbach2021}.
A complete distant speech recognition (DSR) system typically consists of distinct components such as a voice activity detector (VAD), speaker localizer (SL), dereverberator, beamformer and acoustic model \cite{Omolog2001,Wolfel2009,KumataniAYMRST12,KinoshitaDGHHKL16,VirtanenBook2012}. Beamforming techniques take advantage of multiple microphones to enhance the audio signal, and is a key component to improve noise robustness of the DSR. Beamforming can be categorized into fixed beamforming or adaptive beamforming. In comparison to fixed beamforming, adaptive techniques have shown that noise robustness of ASR system can be improved with a dereverberation approach or high-order statistics. However, adaptive techniques rely on accurate VAD or SL, and therefore they can underperform in comparison to a fixed beam former; especially when these dependent components are not performing reliably. According to previous studies, individually optimizing various DSR components is sub-optimal \cite{McDonough2008, Seltzer2008}.
More recently, multi-channel deep neural network (MC-DNN) approaches have been applied to ASR by training a unified MC-DNN model where the MC processing modules are part of the DNN structure \cite{xiao2016deep,ochiai2017multichannel,sainath2017multichannel,minhua2019frequency}. Aside from unified MC-DNN approaches, a DNN is also employed to construct a clean speech signal. A mask-based method was proposed to estimate the statistics of the target clean speech via an LSTM \cite{heymann2018performance, higuchi2018frame}. However, this method needs accumulated statistics from adequate amount of adaptation data to perrform well. Accumulating the statistics might cause additional latency and is less applicable to real-time applications.
\section{Model Architecture}
\label{sec:architecture}
\subsection{Overview}
\label{subsec:overview}
The unified architecture diagram is given in Fig. \ref{fig:architecture}. The model includes two separate frontends for SC and MC audio, with a shared backend. SC audio, or the \textit{primary channel}, is either obtained from the only existing microphone in the device or from an on-device beamformer that combines the outputs of multiple microphones. In our experiments, MC audio has three channels: the primary channel and two \textit{auxiliary channels} that are obtained from the raw audio outputs of two microphones.
The model is trained with a mix of SC and MC audio. MC samples propagate through the MC frontend (FE) and the shared backend, whereas SC samples propagate through the SC FE and the shared backend. During inference, audio is propagated through the corresponding FE for the incoming data type, followed by the shared backend (Sec. \ref{sec:training}).
\subsection{SC Frontend}
\label{subsec:sc-frontend}
In the SC FE, extracted audio features (Sec. \ref{subsubsec:fex}) are processed by a multi-view frequency LSTM (MV-FLSTM) \cite{maarten_flstm} (Sec. \ref{subsubsec:sc-mv-flstm}), followed by a shared backend (Sec. \ref{subsec:backend}).
\subsubsection{Feature Extraction}
\label{subsubsec:fex}
Global mean and variance (GMV) values computed from the received channel are used to normalize the input. Complex STFT features with 256 frequency bins are extracted from the normalized waveform with a window size of 25ms produced in 10ms steps. Three consecutive input frames are stacked into a single vector to reduce the number of time steps in the acoustic model, known as the Lower Frame Rate model \cite{lfr2016}. Log-power of the complex features is consumed by the MV-FLSTM (Sec. \ref{subsubsec:sc-mv-flstm}).
\begin{figure}[t!]
\centering
\renewcommand{\arraystretch}{0.2}
${
\begin{bmatrix}
FR_1, freq_1 \\
\vdots \\
FR_1, freq_{256}\\
FR_2, freq_1\\
\vdots \\
FR_3, freq_1\\
\vdots \\
FR_3, freq_{256}
\end{bmatrix}_{(256\times3,1)}
}$
$\Rightarrow$
\renewcommand{\arraystretch}{0.5}
${
\begin{bmatrix}
FR_1, freq_1 \\
FR_2, freq_1 \\
FR_3, freq_1\\
FR_1, freq_2 \\
FR_2, freq_2\\
\vdots \\
FR_1, freq_{256}\\
FR_2, freq_{256}\\
FR_3, freq_{256}
\end{bmatrix}_{(256\times3,1)}
}$
\caption{Input to the SC MV-FLSTM is reordered to put frequency bins closer. A similar procedure is done before the MC MV-FLSTM, incorporating the look directions in the reorder.}
\label{fig:mvflstm2}
\end{figure}
\subsubsection{SC Frontend MV-FLSTM}
\label{subsubsec:sc-mv-flstm}
A Frequency LSTM (FLSTM) operates over the frequencies contained in an individual input frame with a sliding, overlapping window \cite{flstmli}. Multi-view FLSTM extends this concept by having several window sizes (\textit{views}) to span different frequency ranges at each step \cite{maarten_flstm}. Outputs of the views are concatenated to generate a feature vector that is later processed by a conventional time-domain LSTM (TLSTM). Concatenated feature vector can optionally be projected to a lower dimensional space to reduce computational requirements of the following TLSTM. Prior work has shown 3\%-7\% relative gains in WER by using the MV-FLSTM approach over a single-view FLSTM \cite{maarten_flstm}.
Since FLSTMs operate over the frequency dimension, the output of the feature extraction stage consisting of three stacked frames is reordered. This reordering ensures the same frequency bins are clustered together before being processed by the MV-FLSTM (Fig. \ref{fig:mvflstm2}). In this paper, four FLSTM views are used. Each view consists of a three-layer bidirectional FLSTM with 32 cells in each layer. FLSTMs operate on data that is contained in a single frame, hence bidirectionality does not look ahead in time and disrupt causality. Window sizes of the views are [24, 48, 96, 192] and hop size is half of the window size.
\subsection {MC Frontend}
MC audio is normalized by GMV statistics for each channel. Neural Beamforming Layer (NBL) combines the audio channels and produces a number of look directions, which are processed by a separate MV-FLSTM and the shared backend.
\begin{figure}[t!]
\centering
\includegraphics[width=0.35\linewidth]{./images/NeuralBeamformer.png}
\caption{Neural Beamforming Layer.}
\label{fig:beamformer}
\end{figure}
\subsubsection{Neural Beamforming Layer}
\label{subsubsec:beamformer}
NBL implementation is adopted from the Elastic Spatial Filter described in \cite{minhua2019}. The Discrete Fourier Transform of the normalized input signal can be described as below:
\begin{equation}
\mathbf{X}\left(t, \omega_{k}\right)=\left[X_{1}\left(t, \omega_{k}\right), \cdots, X_{M}\left(t, \omega_{k}\right)\right]^{T}
\end{equation}
Using this notation, we can express the complex weight vector for source position $\mathbf{p}$ as follows:
\begin{equation}
\mathbf{w}^H\left(t, \omega_{k}, \mathbf{p}\right)=\left[w_{1}\left(t, \omega_{k}, \mathbf{p}\right), \cdots, w_{M}\left(t, \omega_{k}, \mathbf{p}\right)\right]
\end{equation}
Thus, the block affine transform (BAT) can be expressed as:
\begin{equation}\footnotesize
\left[\begin{array}{c}{Y_{1}\left(\omega_{1}\right)} \\ {\dots} \\ {Y_{D}\left(\omega_{1}\right)} \\ {\dots} \\ {Y_{1}\left(\omega_{K}\right)} \\ {\dots} \\ {Y_{D}\left(\omega_{K}\right)}\end{array}\right] = \left[\begin{array}{c}{\mathbf{w}_{\mathrm{SD}}^{H}\left(\omega_{1}, \mathbf{p}_{1}\right) \mathbf{X}\left(\omega_{1}\right)+\mathbf{b}_{1}} \\ {\dots} \\ {\mathbf{w}_{\mathrm{SD}}^{H}\left(\omega_{1}, \mathbf{p}_{D}\right) \mathbf{X}\left(\omega_{1}\right)+\mathbf{b}_{D}} \\ {\dots} \\ {\mathbf{w}_{\mathrm{SD}}^{H}\left(\omega_{K}, \mathbf{p}_{1}\right) \mathbf{X}\left(\omega_{K}\right)+\mathbf{b}_{D (K-1)+1}} \\ {\dots} \\ {\mathbf{w}_{\mathrm{SD}}^{H}\left(\omega_{K}, \mathbf{p}_{D}\right) \mathbf{X}\left(\omega_{K}\right)+\mathbf{b}_{DK}}\end{array}\right]
\normalsize
\end{equation} where $\mathbf{b}$ is bias term, $D$ is the number of look directions and $K$ is the number of frequency bins.
Complex STFT features from the two microphones are processed by the trainable BAT, generating 12 look directions (Fig. \ref{fig:beamformer}). BAT is initialized with super directive beamformer weights. Log-power features of the look directions and the primary channel are concatenated and further processed by the MC FE's MV-FLSTM. Early experiments showed that using all three channels in the MC FE had 4-7\% relative WERR compared to using only the two auxiliary (raw) channels, hence the primary input is concetanated to the BAT output.
\subsubsection{MC Frontend MV-FLSTM}
\label{subsubsec:mc-mv-flstm}
The input to the MC FE's MV-FLSTM is 13 times the size of the input to the SC FE's MV-FLSTM. This is due to the concatenation of the primary channel and the look directions from the NBL. In order to ensure the same range of frequencies are spanned, each view in MC FE's MV-FLSTM has 13 times the window and hop size compared to the SC FE implementation. The number of views, layers and units-per-layer are the same for both frontends. Similar to the SC FE implementation, the input to the MV-FLSTM is reordered and processed frequency bins from different look directions (in addition to the stacked frames) are arranged to be next to each other (Fig. \ref{fig:mvflstm2}).
\subsection{Shared Backend}
\label{subsec:backend}
The shared backend has a projection (linear) layer to reduce the input dimensionality to the five-layer, unidirectional TLSTM with 768 cells per layer. Unidirectionality is required to preserve causality for a streamable ASR model. TLSTM output is connected to a classification layer with softmax outputs to generate tied tri-phone states used in a hybrid ASR model \cite{denglibook}. Unified model has a total of 28M parameters.
\begin{figure}[t!]
\centering
\includegraphics[width=0.7\linewidth]{./images/Datasets.png}
\caption{Datasets in this work: (a) SC training, with [500, 10K, 65K] hours, (b) MC training, with [500, 10K, 20K] hours, (c) Test, 45 hours. SC data has one primary channel, MC data has one primary and two auxiliary channels. Large datasets include smaller ones. MC and SC datasets don't overlap.}
\label{fig:datasets}
\end{figure}
\section{Datasets}
\label{sec:datasets}
Datasets used in this work are shown in Fig. \ref{fig:datasets}. SC datasets D500\textsubscript{sc}, D10K\textsubscript{sc} and D65K\textsubscript{sc} are human transcribed with [500, 10K, 65K] hours of audio respectively and they contain only the primary channel. MC datasets D500\textsubscript{mc}, D10K\textsubscript{mc} and D20K\textsubscript{mc} have [500, 10K, 20K] hours of audio respectively and contain three channels: primary channel and two auxiliary channels obtained from raw audio of two microphones. MC datasets are transcribed by an accurate SC model that is free from production constraints such as latency, causality and size.
DTest\textsubscript{mc} is a human-transcribed MC test set with 45 hours of audio, containing three channels (one primary and two auxiliary). For evaluation of SC models, only the primary channel in this set is used. The test set contains a mix of single and multi-speaker utterances (i.e., background speech). SNR value for each utterance is also available in the test set to evaluate performance across different noise conditions. All audio data is de-identified for privacy reasons.
\section{Training Methodology}
\label{sec:training}
During training, each batch of data contains a mix of MC and SC audio. Since the primary channel is present in the MC dataset, the auxiliary channels can be removed and the resulting SC data is added to expand the SC dataset. Gradient updates obtained from the expanded SC dataset are used to train the SC FE and the shared backend. Gradients generated from the MC data are used to update the MC FE and the shared backend. Notably, the unified ASR architecture described in Sec. \ref{sec:architecture} allows the shared backend to learn from both SC and MC data.
There are other alternatives to incorporate SC data to MC model training. One can pad the missing two channels in the SC data with zeros to expand the MC dataset and train an MC model with this expanded dataset. However, experiments show this is inferior to unified model training with two separate frontends (Sec. \ref{subsec:unified-vs-zeropad}). Another alternative is to first train an SC-only model to obtain the backend, freeze its weights, append an MC FE and then train the MC FE with MC-only data. Freezing the weights is required to enable a single model for SC and MC audio. This approach is undesirable since it requires a two-step methodology and the backend will not be updated for MC data. In practice, training in this manner led to non-convergence of the MC FE in large-scale datasets.
During inference of the unified ASR model, either the SC or the MC FE is used based on the incoming data type.
\begin{table}[t!]
\caption{Normalized WER Comparison of Unified Model with Two Frontends to Zero-Padding Missing Channels.}
\label{tab:results0}
\begin{tabular}{|C{0.9cm} | C{1.1cm} | C{2.5cm}| C{2.0cm}|}
\hline
Experiment & Training Data & Model / Inference Path & nWER (Single/Multi-Speaker)\\ \Xhline{4\arrayrulewidth}
E1 & D500\textsubscript{sc} +D500\textsubscript{mc} & MC, Zero-pad missing channels & 100.0/100.0 \\ \hline
E2 & D500\textsubscript{sc} +D500\textsubscript{mc} & Unified (MC FE + Shared Backend) & \textbf{96.0}/\textbf{96.9} \\ \hline
\end{tabular}
\end{table}
\section{Experimental Results}
\label{sec:results}
Models are trained with cross-entropy (CE) loss, followed by CTC loss \cite{ctc} for the same number of training epochs as the corresponding baseline. WER results are obtained from the test set DTest\textsubscript{mc }(Sec. \ref{sec:datasets}) and normalized to the baseline. Absolute WER values are below 10\% for medium and large-scale datasets. Only the primary channel is used for SC-only models during training and test.
\subsection{Comparison of Proposed Approach to Zero Padding}
\label{subsec:unified-vs-zeropad}
Table \ref{tab:results0} shows the Normalized WER (nWER) of the zero-padding baseline (E1) to the proposed unified model architecture (E2). E1 is a combined MC/SC model that consists of the same MC FE and the shared backend as E2 (Sec. \ref{sec:architecture}), but does not include an SC FE. E1 always expects three input channels (one primary and two auxiliary); if only SC data is available during training or inference, the missing channels are filled with zeros. E2 is trained with two separate frontends (Sec. \ref{sec:training}), with each input type going through its respective frontend.
During test time, all three channels in DTest\textsubscript{mc} are used as input to E1 and E2. WER values of E1 for single and multi-speaker cases are arbitrarily set to 100.0, and E2's WER values are normalized to E1. The unified model architecture achieves 3.1\%-4.0\% relative WERR compared to the baseline, demonstrating its advantage over the alternative method.
\begin{table*}[t!]
\centering
\caption{Normalized WER Comparison of MC and SC models.}
\label{tab:results1}
\begin{tabular}{|C{1.5cm} | C{2.8cm} | C{5.5cm}| C{3.5cm}|}
\hline
Experiment & Training Data & Model / Inference Path & { nWER \newline (Single/Multi-Speaker)} \\ \Xhline{4\arrayrulewidth}
E3 & D10K\textsubscript{sc} + D10K\textsubscript{mc} & Standalone SC & 100.0/100.0 \\ \hline
E4 & D10K\textsubscript{mc} & Standalone MC & 89.7/94.8 \\ \hline
E5 & D10K\textsubscript{sc} + D10K\textsubscript{mc} & Unified (SC FE + Shared Backend) & 101.5/99.3 \\ \hline
E6 & D10K\textsubscript{sc} + D10K\textsubscript{mc} & Unified (MC FE + Shared Backend) & \textbf{87.5}/\textbf{92.3} \\ \hline \hline
E7 & D65K\textsubscript{sc} + D20K\textsubscript{mc} & Unified (SC FE + Shared Backend) & 85.9/88.7 \\ \hline
E8 & D65K\textsubscript{sc} + D20K\textsubscript{mc} & Unified (MC FE + Shared Backend) & \textbf{82.6}/\textbf{85.7} \\ \hline
\end{tabular}
\end{table*}
\subsection{Impact of Including Multi-Channel Audio}
\label{subsec:results-mc}
Table \ref{tab:results1} shows the nWER comparison of the SC and MC models in medium and large-scale datasets. E3 has the same SC FE and the shared backend architecture in Fig. \ref{fig:architecture}, but does not contain the MC FE. Datasets D10K\textsubscript{sc} and D10K\textsubscript{mc} are combined for training (Sec. \ref{sec:datasets}); however E3 only uses the primary channel in D10K\textsubscript{mc} since it is an SC-only model. WER values of E3 for single and multi-speaker cases are arbitrarily set to 100.0, and following experiments' WER values are normalized to E3.
E4 is an MC model that combines the MC FE and the shared backend, and no SC FE. It is trained with D10K\textsubscript{mc} and uses all three channels. E3 and E4 are trained separately, with no weight sharing. E4 obtains a 10.3\% relative WER reduction compared to the SC model (E3) in the single-speaker test set, even though E4 is trained with half the data of E3. This clearly demonstrates the improvements that can be obtained from using multiple audio channels. Multi-speaker test set has 5.2\% reduction.
\subsection{Impact of Unified Model Training}
\label{sec:umt}
Experiment E5 shows the results for the SC path (SC FE + shared backend) in the unified model and E6 shows the MC path results for the same model. E5 and E6 are trained together (Sec. \ref{sec:training}). Compared to E4, E6 achieves a further 2.2\%/2.5\% WERR in single/multi-speaker conditions. This demonstrates an advantage of the proposed architecture.
E5 has a 0.7\% improvement in multi-speaker over E3, but a 1.5\% degradation in single-speaker conditions is observed. It is not clear if this degradation is an artifact of the unified model methodology or due to the inherent variability in training dynamics. Multiple runs with different initialization seeds might be warranted for further study.
\subsection{Impact of Large-Scale Data}
\label{subsec:results-lsd}
Table \ref{tab:results1} also shows the nWER comparison of the unified model with SC (E7) and MC (E8) paths in a large-scale data setting. E7 obtains 14.1\%/11.3\% reduction in relative WER compared to the SC baseline model E3. E8 has an additional 3.3\%/3.0\% relative gain over E7. SC data is more widely available due to the prevalence of SC ASR models, and our dataset distribution reflects this fact. Additional experiments with a different data distribution (e.g., oversampling the MC data to reach a 50/50 distribution) could determine if MC results can be further improved.
\subsection{Impact of Utterance SNR}
\label{subsec:snr}
Fig. \ref{fig:snr} shows the nWER for the models with respect to SNR of utterances. The test set is split into three bins according to individual utterances' SNR, and WER is computed for each bin. WER for each bin is normalized to the corresponding SNR bin of E3, whose WER is arbitrarily set to 100.0.
Comparing the MC model E4 to the SC model E3, the advantage of the MC FE is more pronounced in low-SNR ($<10dB$) conditions. A significant 14.8\% relative WERR is observed in this regime. This is perhaps not surprising since additional information available in the auxiliary channels is even more valuable in these noisy utterances. Conversely, the advantage of incorporating SC data in E6 is more evident in medium and high-SNR conditions, with 2.8\% and 2.4\% relative WERR compared to E4. In low-SNR conditions, a smaller 1.2\% improvement is observed. E6 achieves 16.0\% relative WERR in low-SNR conditions compared to the SC baseline E3.
Adding large-scale data also helps significantly in low-SNR conditions. Comparing SC-models E5 and E7, a 17.3\% relative WERR is observed. In medium-SNR ($[10dB, 20dB]$) and high-SNR ($>20dB$) conditions, still significant reductions of 14.0\% and 12.2\% are seen. The combination of all three techniques (unified model, MC FE and additional data) leads to an impressive 21.7\% relative WERR in low-SNR conditions (E3 vs. E8), with 16.2\% and 13.0\% reductions in medium and high-SNR utterances.
\begin{figure}[t!]
\centering
\includegraphics[width=0.77\linewidth]{./images/SNR.png}
\caption{Relative WERR is more pronounced in low-SNR conditions when an MC model is used (e.g., E4 vs E3). Additional data helps more in low-SNR conditions (E7 vs. E5).}
\label{fig:snr}
\end{figure}
\section{Conclusion}
We propose a unified MC/SC model that can be trained with both types of input, allowing a single model to support a variety of scenarios. Proposed approach achieves up to 2.5\% relative WERR compared to the MC baseline and up to 16.0\% relative WERR compared to the SC baseline in low-SNR conditions.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Deep convolutional neural networks (DCNNs), initially inspired by the primate visual cortex architecture, have taken big strides in solving computer vision tasks in the last decade. State-of-the-art networks can learn to classify images with high accuracy from huge labeled datasets~\cite{krizhevsky2012imagenet,simonyan2014very,he2015deep,huang2016densely,al2021reconstructing,tan2019efficientnet}. This rapid progress and the resulting interest in these techniques have also highlighted their various shortcomings. Most widely studied is the sensitivity of neural networks, not only to perturbations specifically designed to fool them (so-called ``adversarial examples'') but also to regular noises typically observed in natural scenes~\cite{szegedy2013intriguing,hendrycks2019benchmarking,nguyen2014deep}. These shortcomings indicate that there is still room for improvement in current techniques.
One possible way to improve the robustness of artificial neural networks could be to take further inspiration from the brain. In particular, one major aspect of the cerebral cortex that is missing from standard feedforward DCNNs is the presence of feedback connections. Recent studies have stressed the importance of feedback connections in the brain~\cite{kietzmann2019recurrence,kar2019evidence}, and have shown how artificial neural networks can take advantage of such feedback for various tasks such as object recognition with occlusion~\cite{ernst2019recurrent}, or panoptic segmentation~\cite{linsley2020stable}. Feedback connections convey contextual information about the state of the higher layers down to the lower layers of the hierarchy; in this way, they can constrain lower layers to represent inputs in meaningful ways. In theory, this could make neural representations more robust to image degradation~\cite{wyatte2012limits}. Merely including feedback in the pattern of connections, however, may not always be sufficient; rather, it should be combined with proper mechanistic principles.
To that end, we explore the potential of recurrent dynamics for augmenting deep neural networks with brain-inspired predictive coding (supported by ample neuroscience evidence~\cite{bastos2012canonical,huang2011predictive,heilbron2018great,aitchison2017or,walsh2020evaluating}). We build large-scale hierarchical networks with both feedforward and feedback connections that can be trained using error backpropagation. Several prior studies have explored this interesting avenue of research~\cite{chalasani2013deep,lotter2016deep,wen2018deep,boutin2019sparse}, but with important differences with our approach (see Section~\ref{prior_work}). We demonstrate that our proposed method adds desirable properties to feedforward DCNNs, especially when viewed from the perspective of robustness.
Our contributions can be summarized as follows:
\begin{itemize}
\item We propose a novel strategy for effectively incorporating recurrent feedback connections based on the neuroscientific principle of predictive coding.
\item We implement this strategy in two pre-trained feedforward architectures with unsupervised training of the feedback weights, and show that this improves their robustness against different types of natural and adversarial noise.
\item We suggest and verify that an emergent property of the network is to iteratively shift noisy representations towards the corresponding clean representations---a form of ``projection towards the learned manifold'' as implemented in certain adversarial defense methods.
\item To facilitate research aimed at using such neuroscientific principles in machine learning, we provide a Python package called \textit{Predify} that can easily implement the proposed predictive coding dynamics in any convolutional neural network with a few lines of code.
\end{itemize}
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{pc_schematic_v7.pdf}
\caption{\textbf{General overview of our predictive coding strategy} as implemented in a feedforward hierarchical network with generative feedback connections. The architecture (roughly similar to stacked auto-encoders) consists of $N$ encoding layers $e_{n}$ and $N$ decoding layers $d_{n}$. $W_{m,n}$ denotes the connection weights from layer $m$ to layer $n$, with $W^{f}$ and $W^{b}$ for feedforward and feedback connections, respectively. The reconstruction errors at each layer are denoted by $\epsilon_{n}$. The feedforward connections (green arrows) are trained for image classification (in a supervised fashion), while the feedback weights (red arrows) are optimized for a prediction (i.e. reconstruction) objective (unsupervised). Predictive coding minimizes the reconstruction errors in each layer by updating activations in the next layer accordingly (black arrows). Self-connections (memory) are represented by blue arrows.}
\label{fig:pc_schematic}
\end{figure}
\section{Our Approach}
\label{our_approach}
\subsection{The proposed predictive coding dynamics}
Predictive coding, as introduced by~\cite{rao1999predictive}, is a neurocomputational theory positing that the brain maintains an internal model of the world, which it uses to actively predict the observed stimulus. Within a hierarchical architecture, each higher layer attempts to predict the activity of the layer immediately below, and the errors made in this prediction are then utilized to correct the higher-layer activity.
To establish our notation, let us consider a hierarchical feedforward network equipped with generative feedback connections, as represented in Figure~\ref{fig:pc_schematic}. The network contains $N$ encoding layers $e_{n}$ ($n \in \mathbb{N}$) and $N$ corresponding decoding layers $d_{n-1}$. The feedforward weights connecting layer $n-1$ to layer $n$ are denoted by $W^{f}_{n-1,n}$, and the feedback weights from layer $n+1$ to $n$ by $W^{b}_{n+1,n}$. For a given input image, we first initiate the activations of all encoding layers with a feedforward pass. Then, over successive recurrent iterations (referred to as timesteps $t$), both the decoding and encoding layer representations are updated using the following equations (also refer to Pseudocode~\ref{pseudocode}):
\begin{equation}
\boldsymbol{d}_{n}(t) = W^{b}_{n+1,n}\boldsymbol{e}_{n+1}(t)
\end{equation}
\begin{equation}
\label{eq:pc_equation}
\boldsymbol{e}_{n}(t+1) = \beta_{n} W^{f}_{n-1,n}\boldsymbol{e}_{n-1}(t+1) + \lambda_{n} \boldsymbol{d}_{n}(t) + (1 - \beta_{n} - \lambda_{n}) \boldsymbol{e}_{n}(t) - \alpha_{n} \nabla{\epsilon_{n-1}(t)} ,
\end{equation}
where $\beta_{n}$, $\lambda_{n}$ ($0\leq \beta_n + \lambda_n \leq 1$), and $\alpha_{n}$ act as layer-dependent balancing coefficients for the feedforward, feedback, and error-correction terms, respectively. $\epsilon_{n-1}(t)$ denotes the reconstruction error at layer $n-1$ and is defined as the mean squared error (MSE) between the representation $e_{n-1}(t)$ and the predicted reconstruction $d_{n-1}(t)$ at that particular timestep. Layer $e_0$ is defined as the input image and remains constant over timesteps. All the weights $W^{f}_{n-1,n}$ and $W^{b}_{n+1,n}$ are fixed during these iterations.
\floatname{algorithm}{\scriptsize Pseudocode}
\begin{algorithm}[b!]
\scriptsize
\algsetup{linenosize=\scriptsize}
\caption{\scriptsize Predictive Coding Iterations}
\label{pseudocode}
\begin{algorithmic}[1]
\STATE{Input image: $e_{0}$}
\FOR{$n = 1$ to $N$}
\STATE{$e_{n} \gets Conv(e_{n-1})$}
\STATE{$d_{n-1} \gets deConv(e_{n})$}
\STATE{$\epsilon_{n-1} \gets ||d_{n-1} - e_{n-1}||_{2}^{2}$ }
\ENDFOR
\FOR{$t=1$ to $T$}
\FOR{$n = 1$ to $N$}
\STATE{$ff \gets \beta_{n}\cdot Conv(e_{n-1})$}
\STATE{$fb \gets 0$}
\IF{$n < N$}
\STATE $fb \gets \lambda_{n}\cdot d_{n}$
\ENDIF
\STATE{$e_{n} \gets ff+fb+(1 - \beta_{n} - \lambda_{n})\cdot e_{n}-\alpha_{n} \cdot \nabla{\epsilon_{n-1}}$}
\STATE{$d_{n-1} \gets deConv(e_{n})$}
\STATE{$\epsilon_{n-1} \gets ||d_{n-1} - e_{n-1}||_{2}^{2}$ }
\ENDFOR
\ENDFOR
\end{algorithmic}
\end{algorithm}
Each of the four terms in Equation~\ref{eq:pc_equation} contributes different signals, reflected by different arrow colors in Figure \ref{fig:pc_schematic}: (i) the feedforward term (green arrows; controlled by parameter $\beta$) provides information about the (constant) input and changing representations in the lower layers, (ii) the feedback correction term (red arrows; parameter $\lambda$), as proposed in~\cite{rao1999predictive,heeger2017theory}, guides activations towards their representations from the higher levels, thereby reducing the reconstruction errors over time, (iii) the memory term (blue arrows) acts as a time constant to retain the current representation over successive timesteps, and (iv) the feedforward error correction term (black arrows; controlled by parameter $\alpha$) corrects representations in each layer such that their generative feedback can better match the preceding layer. For this error correction term, we directly use the error gradient $\nabla{\epsilon_{n-1}}=[\frac{\partial\epsilon_{n-1}}{\partial e_{n}^{0}}, ..., \frac{\partial\epsilon_{n-1}}{\partial e_{n}^{k}}]$ to take full advantage of modern machine learning capabilities (where $k$ is the number of elements in $e_n$). While the direct computation of this error gradient is biologically implausible, it has been noted before that it is mathematically equivalent to propagating error residuals up through the (transposed) feedback connection weights $(W^{b})^{T}$, as often done in other predictive coding implementations~\cite{wen2018deep,rao1999predictive}. Together, the feedforward and feedback error correction terms fulfill the objective of predictive coding as laid out by Rao and Ballard~\cite{rao1999predictive}. We discuss the similarities and differences between our equations and those proposed in the original Rao and Ballard implementation in the Appendix~\ref{apndx:rao_ballard}.
While it is certainly possible to train such an architecture in an end-to-end fashion, by combining a classification objective for the feedforward weights $W^{f}$ with an unsupervised predictive coding objective (see Section~\ref{methods}) for the feedback weights $W^{b}$, we believe that the benefits of our proposed scheme are best demonstrated by focusing on the added value of the feedback pathway onto a pre-existing state-of-the-art feedforward network. Consequently, we implement the proposed strategy with two existing feedforward DCNN architectures as backbones: VGG16 and EfficientNetB0, both trained on ImageNet. We show that predictive coding confers higher robustness to these networks.
\subsection{Model architectures and training}
\label{methods}
We select VGG16 and EfficientNetB0, two different pre-trained feedforward networks on ImageNet, and augment them with the proposed predictive coding dynamics. The resulting models are called PVGG16 and PEfficientNetB0, respectively. The networks' ``bodies'' (without the classification head) are split into a cascade of $N$ sub-modules, where each plays the role of an $e_{n}$ in equation~(\ref{eq:pc_equation}). We then add deconvolutions as feedback layers $d_{n-1}$ connecting each $e_{n}$ to $e_{n-1}$, with kernel sizes accounting for the increased receptive fields of the neurons in $e_{n}$ or upsampling layers to match the size of the predictions and their targets (see Appendix~\ref{apndx:netarc}). We then train the parameters of the feedback deconvolution layers with an unsupervised reconstruction objective (with all feedforward parameters frozen). We minimize the reconstruction errors just after the first forward pass, and after a single deconvolution step (i.e. no error correction or predictive coding recurrent dynamics are involved at this stage):
\begin{equation}\label{eq:rec_loss}
\mathcal{L} = \sum_{n=0}^{N-1}\parallel e_{n} - d_{n} \parallel_2^{2},
\end{equation}
where $e_n$ is the output of the $n^{th}$ encoder after the first forward pass and $d_n$ is the estimated reconstruction of $e_n$ via feedback/deconvolution (from $e_{n+1}$).
For both the networks, after training the feedback deconvolution layers, we freeze all of the weights, and set the values of hyperparameters to $\beta_{n}=0.8$, $\lambda_{n}=0.1$, and $\alpha_{n}=0.01$ for all the encoders/decoders in Equation~(\ref{eq:pc_equation}). We also explore various strategies for further tuning hyperparameters to improve the results (see Appendix~\ref{apndx:tuninghps} for the chosen hyperparameter values).
\subsection{\textit{Predify}}
\label{methods_predify}
To facilitate and automate the process of adding the proposed predictive coding dynamics to existing deep neural networks, we have developed an open-source Python package called \textit{Predify}. The package is developed based on PyTorch~\cite{NEURIPS2019_9015} and provides a flexible object oriented framework to convert any PyTorch-compatible network into a predictive network. While an advanced user may find it easy to integrate \textit{Predify} in their project manually, a simple text-based user interface (in TOML\footnote{\url{https://toml.io/en/}} format) is also provided to automate the steps. For the sake of improved performance and flexibility, \textit{Predify} generates the code of the predictive network rather than the Python object. Given the original network and a configuration file (e.g. \texttt{{\color{orange}\textquotesingle config.toml\textquotesingle}}) that indicates the intended source and target layers for the predictive feedback, three lines of code are enough to construct the corresponding predictive network:
\lstdefinestyle{mystyle}{
backgroundcolor=\color{white},
commentstyle=\color{TealBlue},
keywordstyle=\color{purple},
stringstyle=\color{orange},
basicstyle=\ttfamily\footnotesize,
breakatwhitespace=false,
breaklines=true,
captionpos=b,
keepspaces=true,
showspaces=false,
showstringspaces=false,
showtabs=false,
tabsize=2
}
\lstset{style=mystyle}
\begin{lstlisting}[language=Python]
from predify import predify
net = # load PyTorch network
predify(net,'./config.toml') # config file indicates the layers that
# will act as outputs of encoders.
\end{lstlisting}
The Appendix~\ref{apndx:predify} provides further details on the package, along with a sample config file and certain default behaviours. \textit{Predify} is an ongoing project available on GitHub\footnote{\url{https://github.com/miladmozafari/predify}} under GNU General Public License v3.0. Scripts for creating PVGG16 and PEfficientNetB0 from their feedforward instances and reproducing the results presented in this paper, as well as the pre-trained weights are also available on another GitHub repository\footnote{\url{https://github.com/bhavinc/predify2021}}.
\section{Related work}
\label{prior_work}
There is a long tradition of drawing inspiration from neuroscience knowledge to improve machine learning performance.~Some studies suggest using sparse coding, a concept closely related to predictive coding~\cite{huang2011predictive,olshausen1996emergence, oja1999image,chalk2018toward,paiton2020selectivity}, for image denoising~\cite{lu2013sparse} and robust deep learning~\cite{sulam2020adversarial,kim2020modeling}, while other studies focus on implementing feedback and horizontal recurrent pathways to tackle challenges beyond the core object recognition~\cite{linsley2020stable,wyatte2012limits,heeger2017theory,spoerer2017recurrent,linsley2018learning,Frosst2018DARCCCDA,krotov2018dense,kubilius2019brain,rajaei2019beyond}.
Here, we focus specifically on those studies that tried implementing predictive coding mechanisms in machine learning models~\cite{chalasani2013deep,lotter2016deep,wen2018deep,boutin2019sparse}. Out of these, our implementation is most similar to the Predictive Coding Networks (PCNs) of~\cite{wen2018deep}. These hierarchical networks were designed with a similar goal in mind: improving object recognition with predictive coding dynamics. However, their network (including the feedback connection weights) is solely optimized with a classification objective. As a result, their network does not learn to uniformly reduce reconstruction errors over timesteps, as the predictive coding theory would mandate. We also found that their network performs relatively poorly until the final timestep (see corresponding Figure~\ref{fig:prednet} in the Appendix~\ref{apndx:purdue_paper}), which does not seem biologically plausible: biological systems typically cannot afford to wait until the last iteration before detecting a prey or a predator. In the proposed method, we incorporate the feedforward drive into a similar PC dynamics and train the feedback weights in an unsupervised way using a reconstruction loss. We then show that these modifications help resolve PCNs' issues. We discuss these PCNs~\cite{wen2018deep} further in the Appendix~\ref{apndx:purdue_paper}, together with our own detailed exploration of their network's behavior.
Other approaches to predictive coding for object recognition include Boutin et al.~\cite{boutin2019sparse}, who used a PCN with an additional sparsity constraint. The authors showed that their framework can give rise to receptive fields which resemble those of neurons in areas V1 and V2 of the primate brain. They also demonstrated the robustness of the system to noisy inputs, but only in the context of reconstruction. Unlike ours, they did not show that their network can perform (robust) classification, and they did not extend their approach to deep neural networks.
Spratling~\cite{spratling2017hierarchical} also described PCNs designed for object recognition, and demonstrated that their network could effectively recognise digits and faces, and locate cars within images. Their update equations differed from ours in a number of ways: they used divisive/multiplicative error correction (rather than additive), and a form of biased competition to make the neurons ``compete'' in their explanatory power. The weights of the network were not trained by error backpropagation, making it difficult to scale it to address modern machine learning problems. Conversely, our proposed network architecture and PC dynamics are fully compatible with error backpropagation, making them a suitable option for large-scale problems. Indeed, the tasks on which they tested their network are simpler than ours, and the datasets are much smaller.
Huang et al.~\cite{huang2020neural} also aimed to extend the principle of predictive coding by incorporating feedback connections such that the network maximizes ``self consistency'' between the input image features, latent variables and label distribution. The iterative dynamics they proposed, though different from ours, improved the robustness of neural networks against gradient-based adversarial attacks on datasets such as Fashion-MNIST and CIFAR10.
\section{Results}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{fig2_newrecs_v1.png}
\caption{\textbf{Performance under Gaussian noise and projection towards the learned manifold.} (a) Improvement in recognition accuracy with reference to the feedforward baseline under various levels of Gaussian noise. Both networks demonstrate significant accuracy improvement across timesteps under noisy conditions, while maintaining a performance close to the feedforward level for clean images. (b) Normalized MSE distance between the image reconstruction ($d_0$) and the clean image ($e_0$). Irrespective of the noise level, image reconstruction consistently gets closer to the clean image across timesteps in both models. (c) Examples of clean and noisy input images together with their final reconstruction by the model (the row order from top to bottom is: original image, PVGG16 reconstruction, PEfficientNetB0 reconstruction; noisy image, PVGG16 reconstruction, PEfficientNetB0 reconstruction). For best viewing, we recommend zooming in on the electronic version. (d) Normalized correlation distance between representation of clean and noisy images for each encoder ($e_i$) across timesteps. The values are normalized with respect to the feedforward baseline (timestep 0). In both models and all encoders, the noisy representations tend to move toward the clean copies.}
\label{fig:manifold_projection}
\end{figure}
Here we contrast the behavior of feedforward networks with their predictive coding augmentations. When considered at timestep 0 (i.e., after a single feedforward and feedback pass through the model), the deep predictive coding networks (DPCNs) and their accuracy are---by construction---exactly identical to their standard pretrained feedforward versions. Over successive timesteps, however, the influence of feedback and predictive coding iterations becomes visible. Here, we investigate for both DPCNs (PVGG16 and PEfficientNetB0): (i) how the PC dynamics update the networks' representations across timesteps, and in which direction relative to the learned manifold; (ii) how the networks benefit from PC under noisy conditions, or against adversarial attacks.
\subsection{Performance under Gaussian noise}
To understand the evolution of representations and the behavior of the proposed DPCNs, we first investigate their performance under the influence of different levels of Gaussian noise. To this end, we inject additive Gaussian noise to the ImageNet validation set, and monitor the models' performance across timesteps.
In Figure~\ref{fig:manifold_projection}a we provide the classification accuracy on these noisy images and absolute values in the Table~\ref{tab:absolute_values}. We observed that both models progressively improve their recognition accuracy relative to their feedforward baseline (timestep 0) over successive iterations while imposing only a minor performance reduction on clean images. In other words, the networks are able to discard some of the noise by leveraging the predictive coding dynamics over timesteps.
\subsection{Projection towards the learned manifold}
In order to quantify DPCNs' denoising ability, we evaluate the quality of image reconstructions generated by each network using the mean squared error (MSE) between the clean image and its reconstruction generated by the first decoder. For each DPCN, we normalize these distances, by dividing them by the value obtained for the corresponding feedforward network (at t=0). We provide the absolute values in the Table~\ref{tab:mse_values_absolute}. As Figures~\ref{fig:manifold_projection}b-c illustrate, the reconstructions become progressively cleaner over timesteps. It should be noted that the feedback connections were trained only to reconstruct clean images; therefore, this denoising property is an emerging feature of the PC dynamics.
Next, we test whether the higher layers of the proposed DPCNs also manifest this denoising property. Hence, we pass clean and noisy versions of all images from the ImageNet validation set through the networks, and measure the average correlation distance between the clean and noisy representations of each encoder at each timestep. As done above, these correlation distances are then normalized with the distance measured at timestep 0 (i.e., relative to the standard feedforward network). For both the networks, the correlation distances decrease consistently over timesteps across all layers (see Figure~\ref{fig:manifold_projection}d). This implies that predictive coding iterations help the networks to steer the noisy representations closer to the representations elicited by the corresponding (unseen) clean image.
This is an important property for robustness. When compared to clean images, noisy images can result in different representations at higher layers~\cite{xie2019feature} and consequently, produce significant classification errors. Various defenses have aimed to protect neural networks from perturbations and adversarial attacks by constraining the images to the ``original data manifold''. Accordingly, studies have used generative models such as GANs~\cite{samangouei2018defense,shen2017ape,meng2017magnet,jalal2017robust} or PixelCNNs~\cite{song2018pixeldefend} to constrain the input to the data manifold. Similarly, multiple efforts have been made to clean the representations in higher layers and keep them closer to the learned latent space~\cite{xie2019feature,tao2018attacks,orhan2019improving,roth2019odds}.
Here, we demonstrate that feedback predictive coding iterations can achieve a similar goal by iteratively projecting noisy representations towards the manifolds learned during training, both in pixel (Figure~\ref{fig:manifold_projection}b-c) and representation spaces (Figure~\ref{fig:manifold_projection}d).
\begin{figure}[t!]
\centering
\includegraphics[width=0.7\textwidth]{fig3_new_sharedaxisv3.pdf}
\caption{\textbf{Benchmarking robustness to ImageNet-C.} (a) Normalized corruption errors (CE) of PVGG16 and PEfficientNetB0 under four types of additive noise corruptions. The values are normalized with respect to the feedforward baseline. Both networks show consistent reductions in the errors across timesteps. (b) Normalized mean Corruption Error (mCE) scores for PVGG16 and PEfficientNetB0 on all the 19 corruptions available in the ImageNet-C dataset, when optimized hyperparameters are used (as described in the Appendix~\ref{apndx:tuninghps}). The values are normalized with respect to the feedforward baseline. In both the panels, error bars represent the standard deviation of the bootstrapped estimate of the mean value.}
\label{fig:robustness_imagenet_c}
\end{figure}
\subsection{Benchmarking robustness to ImageNet-C}
Given the promising results with additive Gaussian noise (Figure~\ref{fig:manifold_projection}), we extend the noise variety and quantify the classification accuracy of the networks under different types of perturbations. We use ImageNet-C, a benchmarking dataset for noise robustness provided by~\cite{hendrycks2019benchmarking}, including 19 types of image corruptions across 5 severity levels each. To begin with, we evaluate DPCNs with pre-defined hyperparameter values (as provided in subsection~\ref{methods}). We observe that they improve the Corruption Error (CE) scores over timesteps for several of the additive-noise corruptions: Gaussian noise, shot noise, impulse noise or speckle noise (see Figure~\ref{fig:robustness_imagenet_c}), but fail to improve the overall mean Corruption Error, or mCE score (the recommended score for this benchmark~\cite{hendrycks2019benchmarking}).
Thus, instead of using pre-defined hyperparameter values, we fine-tune them using two different methods (see Appendix~\ref{apndx:tuninghps}), and repeat the above experiment. As shown in Figure~\ref{fig:robustness_imagenet_c}b, when the hyperparameters are more appropriately tuned for the task, the PC dynamics can increase noise robustness more generally across noise types, resulting in improvements of the mean Corruption Error (mCE) score. The CE plots for individual perturbations along with other recommended metrics (values normalized with AlexNet scores, Relative mCE scores) are provided in the Appendix~\ref{apndx:mce_alexnet}.
Furthermore, in the Appendix~\ref{apndx:mce_robust_net}, we demonstrate that we can replicate these observations with a version of PEfficientNetB0 provided by~\citep{xie2020adversarial} that is robust to corruptions in the ImageNet-C dataset. We show that the recurrent dynamics we propose still help in further improving the mCE score of this already robust network.
\subsection{Benchmarking robustness to adversarial attacks}
Finally, we evaluate the robustness of the networks across timesteps against adversarial attacks. The proposed DPCNs are recurrent models, meaning that their layer representations change on every timestep, and consequently, so do the classification boundaries in the last layer, leading to different accuracy and generalization errors across time (as seen above). To mitigate this effect and properly assess the changes in robustness due to the PC dynamics, for each network we start by selecting 1000 images from the ImageNet validation dataset such that they are correctly classified across all timesteps. Also, we only perform \emph{targeted} attacks so that for each image, the same attack target is given for all timesteps. Using the \textit{Foolbox} library~\cite{rauber2017foolbox}, we conduct targeted Basic\_Iterative\_Method attacks (BIM, with $L_{\infty}$ norm)~\cite{goodfellow2014explaining} for both networks; although it would prove computationally prohibitive to systematically explore all standard types of adversarial attacks, we also evaluated random Projected\ Gradient\ Descent\ attacks (RPGD, with $L_2$ norm)~\cite{madry2018towards}, and non-gradient-based HopSkipJump attacks~\citep{chen2020hopskipjumpattack} on a subset of 100 images, specifically for PEfficientNetB0. Across various levels of allowed image perturbations (denoted as $\epsilon$s), the predictive coding iterations tend to decrease the success rate of the attacks across timesteps, for both networks and attacks (see Figure~\ref{fig4}). That is, DPCNs are more robust against these adversarial attacks than their feedforward counterparts.
\begin{figure}
\centering
\includegraphics[width=0.9\textwidth]{advattacks.pdf}
\caption{\textbf{Benchmarking robustness to adversarial attacks.} Plots show the success rate of targeted adversarial attacks against DPCNs across timesteps. The values are baseline-corrected, relative to the success rate at timestep 0 (feedforward baseline). Though we see some signs of reversals for a few perturbations, on average, both networks demonstrate improving robustness across timesteps to different types and/or levels of perturbations.}
\label{fig:robustness_adversarial_attacks}
\label{fig4}
\end{figure}
\section{Discussion and Conclusion}
In this work, we explore the use of unsupervised recurrent predictive coding (PC) dynamics, based on neuroscientific principles, to augment modern deep neural networks. The resulting models have an initial feedforward sweep, compatible with visual processing in human and macaque brains~\cite{kar2019evidence,thorpe1996speed,hung2005fast,vanrullen2007power}. Following this feedforward sweep, consecutive layers iteratively exchange information regarding predictions and prediction errors, aiming to converge towards a stable explanation of the input. This dynamic system is inspired by, and reminiscent of, the ``canonical microcircuit'' (a central component of cortical structure~\cite{bastos2012canonical}) that relies on feedback signaling between hierarchically adjacent layers to update its activity. Overall, the augmented networks are closer to the architecture of biological visual systems, while gaining some desirable functional properties. For example, in~\cite{PANG2021164}, we also demonstrated that the proposed dynamics help the networks perceive illusory contours in a similar way to humans.
Here, we implemented these PC dynamics in two state-of-the-art DCNs, VGG16 and EfficientNetB0, and showed that they helped to improve the robustness of the networks against various corruptions (e.g. ImageNet-C). We demonstrated that this behavior, at least partly, stems from PC's ability to project both the corrupted image reconstructions and neural representations towards their clean counterparts.
We also tested the impact of our network augmentations against adversarial attacks; here again, we showed that PC helps to improve the robustness of the networks. So far, the most promising strategy for achieving robustness has been adversarial training, whereby adversarial datapoints are added to the training dataset. While efficient, this strategy was also shown to be strongly limited~\cite{schmidt2018adversarially,zhang2019defense}. Apart from factors like the choice of the norms used for training, or the high computation requirements, it is ultimately performed with a supervised loss function that can alter the decision boundaries in undesirable ways~\cite{zhang2019defense,tanay2016boundary}.
Most importantly, adversarial training shares very little, if any, resemblance to the way the brain achieves robustness. Instead, here we start from biological principles and show that they can lead to improved adversarial robustness. It is worth mentioning that both our networks achieved robustness totally via unsupervised training of the feedback connections (while of course, the backbone feed-forward networks that we used were pretrained in a supervised manner). We avoided using costly adversarial training, or tuning our hyperparameters specifically for classification under each attack. This likely explains why the models, while improving in robustness compared to their feedforward versions, remain far from state-of-the-art adversarial defenses. On the other hand, we believe that addition of these methods (adversarial training, hyperparameter tuning) to the training paradigm, in future work, could further improve the networks' adversarial robustness.
For the present experiments, we made a choice of using different objectives for training the feedforward and feedback weights: pre-trained feedforward weights optimized for classification, feedback weights trained with a reconstruction objective (computed after a single time-step). On the one hand, we note that it is perfectly feasible to train a similar predictive coding architecture with a single objective (classification, reconstruction, or otherwise) for both feedforward and feedback weights~\citep{PANG2021164,alamia2021role}. On the other hand, our choice has several advantages. First, using a feedforward backbone pretrained for classification allowed us to demonstrate the effect of our dynamics on pre-existing state-of-the-art neural networks. Some authors have tried training both feedforward and feedback connections together for classification~\citep{wen2018deep} at the final timestep for relatively smaller networks, but as we discussed in our explorations in the Appendix~\ref{apndx:purdue_paper}, we found that the resulting network ended up classifying correctly at the last timestep, with very poor performance during early timesteps. This problem could be addressed by training over time-averaged metrics, such as the average cross-entropy loss for $N$ timesteps. Nonetheless, training the feedback weights for reconstruction instead of classification has the additional advantage that it can be done entirely without supervision. We chose to train the feedback weights for a single time-step, because training with recurrence over multiple timesteps would have required unrolling the network over time. Hence, training a large network like PVGG16 for say 5 or 10 timesteps would incur significant computational challenges. Furthermore, our use of a one-step reconstruction objective allowed us to train the feedback weights independently of the various hyperparameters of our predictive coding dynamics ($\beta$, $\lambda$, and $\alpha$), which only influence the model behavior after the second timestep. Training these weights using recurrence would have required to (i) either fix the values of these hyperparameters beforehand, leading to constraints of expensive hyperparameter explorations; (ii) or directly train these hyperparameters as parameters of the model, probably with additional constraints to prevent the network from reaching trivial values (e.g., if all hyperparameters but the feedforward term $\beta$ converge to zero, the network performs identically to a feedforward one). Finally, from a neuroscience perspective, whether and how the brain combines discriminative and generative representations has been an open question addressed by many researchers, e.g.~\citep{al2021reconstructing,dicarlo2021does,huffman2014multivariate}. Our approach of a discriminative (classification-trained) feedforward coupled with generative (reconstruction-trained) feedback could be considered another attempt in this direction.
We speculate that the proposed PC dynamics could help improve robustness in most feedforward neural architectures. To facilitate further explorations in this direction, we provided a Python package, called \textit{Predify}, which allows users to implement recurrent PC dynamics in any feedforward DCN, with only a few lines of code. \textit{Predify} automates the network building, and thus simplifies experiments. On the other hand, there is as yet no established method or criteria to automate the process of identifying the appropriate number of encoding layers, their source and target layers in the DCN hierarchy, and the corresponding hyperparameter values. This remains an open research question, and a requirement for manual explorations and tuning from \textit{Predify} users. For instance, our own explorations with augmenting ResNets through \textit{Predify} proved difficult, and failed in some situations but succeeded in others. More specifically, as developed in~\citep{alamia2021role} using \textit{Predify}, ResNet augmentations always achieved noise robustness when the hyperparameter values (controlling the feedforward, feedback, and memory terms) could be tuned separately for each noise type; but we found it challenging to identify a single set of hyperparameters that could generalize to all noise types. Nonetheless, we are hopeful that the package will prove useful to the community. The code is structured such that users can readily adapt it to test their hypotheses. In particular, it should allow both proponents and opponents of the predictive coding theory to investigate its effects on any DCN.
Overall, this work contributes to the general case for continuing to draw inspiration from biological visual systems in computer vision, both at the level of model architecture and dynamics. We believe that our user-friendly Python package \textit{Predify} can open new opportunities, even for neuroscience researchers with little background in machine learning, to investigate bio-inspired hypotheses in deep computational models, and thus bridge the gap between the two communities.
\section{Broader Impacts}
\label{broader_impacts}
The research discussed above proposes novel ways of using brain-inspired dynamics in current machine learning models. Specifically, it demonstrates a neuro-inspired method for improving the robustness of machine learning models. Given that such models are employed by the general public, and are simultaneously shown to be heavily vulnerable, research efforts to increase (even marginally) or to understand their robustness against mal-intentioned adversaries has high societal relevance.
Importantly, the research also aims to bridge techniques between two different fields--neuroscience and machine learning, which can potentially open new avenues for studying the human brain. For example, it could help better understand the unexplained neural activities in patients, to improve their living conditions, and in the best case, in the treatment of their conditions. While this may also be associated with inherent risks (related to privacy or otherwise), there are clear potential benefits to society.
The likelihood of sentient AI arising from this line of research is estimated to be rather low.
\clearpage
\begin{ack}
This work was funded by an ANITI (Artificial and Natural Intelligence Toulouse Institute) Research Chair to RV (ANR grant ANR-19-PI3A-0004), as well as ANR grants AI-REPS (ANR-18-CE37-0007-01) and OSCI-DEEP (ANR-19-NEUC-0004).
\end{ack}
|
\section{Introduction}
Rapid progress in text-to-speech (TTS) was initiated by autoregressive models. These models replaced traditional methods \cite{griffin1984signal, yoshimura1999simultaneous, kawahara1999restructuring, suk2011subject, lee2018high} and are typically sequence-to-sequence in an encoder-decoder framework with attention mechanism \cite{arik2017deep, gibiansky2017deep, wang2017tacotron, shen2018natural, li2019neural, li2020robutrans, weiss2020wave}. The purpose of the encoder is to extract hidden representation feature vectors from phoneme sequence, and the decoder generates mel-spectrograms from produced vectors. Despite the advantages, end-to-end attention within autoregressive models have limitations such as slow inference speed, word skipping, and reading \cite{ren2019fastspeech, li2020robutrans}. As a means of remedy to this problem, non-autoregressive models are proposed for parallel generation of mel-spectrograms from text or phoneme \cite{ren2019fastspeech, lim2020jdi, vainer2020speedyspeech, lancucki2020fastpitch, lee2020multi, elias2020parallel, luo2021lightspeech}. Although the new architecture alleviates some of the drawbacks from autoregressive models, the duration aligner of non-autoregressive models still require guidance from external aligners. The most critical issue of having an external aligner is an increase in complexity of the training process. Properly aligned text and speech attention maps are required from an autoregressive model before training. This delays the training process and the non-autoregressive model becomes reliant on the quality of alignments generated by the external aligner. Therefore, recent synthesis pipelines are designed with the objective of robust end-to-end internal aligners \cite{kim2020glow, donahue2020end, miao2020efficienttts, shen2020non, liu2021vara}.
Most recent works on TTS related to our work are EATS \cite{donahue2020end} and HiFi-GAN \cite{kong2020hifi}. EATS is an end-to-end text to waveform network with an internal aligner that approximates phoneme to mel-spectrogram sequence alignments with Gaussian kernels. Although the model proposes robust text-to-wave synthesis, the alignments are not additionally trained to ensure improved duration alignment. Hifi-GAN generates high quality audio waveforms through multi-scale and multi-period discriminators, but the model is limited to considering only mel-spectrograms as input. Our proposed text-to-waveform network represents the best of both worlds and improvement on the drawbacks.
In this paper we propose an end-to-end text-to-wave network with reinforce-aligner, which is a reinforcement learning based alignment search method for robust speech synthesis. Our agent interacts with the environment over a sequence of steps to select the best action given the current state. Then, the environment applies an update to the action and returns the reward of the action for the agent to take into account for the next step. This training process is repeated until convergence, and enables the network to internally learn its own alignment. Our experimental results show the positive impact of the reinforce-aligner on the duration alignment and quality of generated audio waveforms. The synthesized audio samples are provided on our online demo webpage.\footnote{\url{https://prml-lab-speech-team.github.io/demo/Reinforce-Aligner}}
\begin{figure}[!t]
\begin{center}
\includegraphics[width=1\columnwidth]{figures/reinforce_aligner.png}
\end{center}
\caption{Overall architecture of reinforce-aligner. The agent (duration predictor) interacts with the environment (text-to-waveform network) to receive current state (phoneme sequence) and reward feedback.
}
\label{fig:1}
\end{figure}
\begin{figure*}[!t]
\centering
\subfigure[Generator training process]{
\centering
\includegraphics[width=1.50\columnwidth]{figures/generator_overall.png}
\label{generator}
}
\subfigure[Reward setups]{
\includegraphics[width=0.55\columnwidth]{figures/reward_setups.png}
\label{rewards}
}
\caption{Figure (a) is the overall architecture of the generator in training process. The aligner is shown in detail to describe the actions available for the agent. Figure (b) is the phoneme-wise and segment-wise rewards. Each arrow represents a shift that deviates towards the lower mel-spectrogram loss value.}
\label{fig:2}
\end{figure*}
\section{Model Architecture}
Our proposed fully convolutional generator utilizes text or phoneme as input and generates a raw audio waveform as output. The encoder contains a modified multi-receptive field fusion module (MRF) \cite{kong2020hifi}. The original implementation of MRF uses mel-spectrograms as input and upsamples the mel-spectrogram with transposed convolution to a raw waveform. Our implementation utilizes phoneme embeddings as input and outputs hidden representations without upsampling. The MRF based phoneme encoder contains multiple residual blocks with multiple kernel sizes and dilation rates, which is an essential component of our network because various receptive fields are able to extract distinct contextual features from phoneme embeddings. Then, the encoder output progresses through the reinforce-aligner to produce output frames. The output frames are randomly segmented by $\gamma$. Finally, the decoder upsamples $\gamma$ by 256 to produce raw audio waveforms.
Our model contains two discriminators for different objectives during training. The first discriminator is the Multi-Scale Discriminator proposed in \cite{kumar2019melgan}. This discriminator effectively learns different frequency components of audio through variation in scales. Each of the three sub-discriminators contain convolution on different scales: raw audio, downsampled by factor of 2, and downsampled by factor of 4. The second discriminator is Multi-Period Discriminator \cite{kong2020hifi}, which captures the distinct features of audio with convolutions in periodic variations. Each of the different periodic values consider different periodic segments of the input audio.
\begin{figure*}[!t]
\begin{center}
\includegraphics[width=180mm]{figures/alignments3.png}
\end{center}
\caption{Duration alignments between input phonemes and output frames in our proposed model. The three unsupervised models are compared to the last predicted alignment computed from supervision of target durations. (a) Without Soft-DTW, (b) With Soft-DTW, (c) Phoneme-wise ($\alpha=2$) (d) With duration target
}
\label{fig:3}
\end{figure*}
\section{Reinforce-Aligner}
\subsection{Reinforcement Learning Setup}
The general architecture of the reinforce-aligner is shown in Figure \ref{fig:1}. The duration predictor is the agent that interacts with the environment for each step of the training process. The environment is the overall text-to-waveform network. As shown in the generator training process of Figure \ref{generator}, duration predictions are each upsampled to waveforms, and reward is calculated by mel-spectrogram losses of the respective mel-spectrograms. The reward feedback is given back to the agent for the final action selection.
\subsection{Agent}
The duration predictor returns one scalar value for each phoneme duration prediction. Predictor consists of two 1D convolutional layers, each with layer normalization, ReLU activation, and dropout following \cite{ren2019fastspeech}. The last linear layer reshapes the convolutional output into a single scalar value. There are two actions available for the agent:
\begin{itemize}
\item \textit{KEEP}: Keep phoneme duration prediction without any alterations to the prediction output.
\item \textit{SHIFT}: Shift phoneme duration prediction with shift value applied in alternating signs.
\end{itemize}
Shift applied to each phoneme duration consists of alternating signs, which is important to maintain the total sum of the duration prediction outputs. Additionally, we designed two types of shifts: Segments-wise and phoneme-wise shifts. Segment-wise shift corresponds to shift applied to the entire phoneme sequence segments. Phoneme-wise shift is shift applied to each phoneme duration of a phoneme sequence. We examine the effect of shift type and value on the outcome of alignments and speech quality in our ablation study.
\subsection{Environment}
In this reinforcement learning based setup, the environment is the trained text-to-waveform network that outputs audio waves from phoneme inputs. There are two main objectives of the environment in the reinforce-aligner: (i) Provide input phoneme sequence information to the agent before it takes action, (ii) Give feedback to the agent after considering each of the two possible actions.
\subsubsection{State}
The environment produces phoneme embedding outputs for the agent in each training step. From phoneme sequence inputs, the multi-receptive field fusion based phoneme encoder generates phoneme embedding outputs through multiple residual blocks. These encoder outputs are the current state inputs for the agent to decide on an action.
\subsubsection{Rewards}
We have two different rewards depending on the type of shift. Both rewards consider the mel-spectrogram loss, which is the L1 loss between mel-spectrograms of ground-truth and generated waveforms. As shown in Figure \ref{generator}, each mel-spectrograms are produced by the waveforms synthesized from predicted (\textit{KEEP}) and shifted (\textit{SHIFT}) durations. The segment-wise reward compares loss values of the entire wave segments used for training. Lower loss value implies higher similarity of the generated waveform to the ground-truth. The phoneme-wise reward considers phoneme-wise mel-spectrogram loss values. Specifically, the mel-spectrogram loss values are interpolated by downsampling into the shape of the phoneme duration sequences. Denote the phoneme duration sequences of predicted duration as $D_k = [d_1, d_2, d_3, ..., d_j]$ and shifted duration as $D_s = [d_1+\alpha, d_2-\alpha, d_3+\alpha ..., d_j\pm\alpha]$, where $\alpha$ represents the shift value. Then, our reward is formulated as:
\begin{equation}
r_{j}=
\begin{cases}
r_{kj}=1 \text{ and } r_{sj}=0, & \mbox{if } \mathcal{L}_{kj} \leq \mathcal{L}_{sj} \\
r_{kj}=0 \text{ and } r_{sj}=1, & \mbox{if } \mathcal{L}_{kj} > \mathcal{L}_{sj}
\end{cases}
\end{equation}
Here, $\mathcal{L}_k$ represents the L1 loss between predicted and ground-truth mel-spectrograms, and $\mathcal{L}_s$ is the L1 loss between shifted and ground-truth mel-spectrograms. $r_k$ and $r_s$ are the keep reward and shift reward values, respectively. Each $j$ is a phoneme duration index value in the phoneme duration sequence for a total of $N$ indices. For the segment-wise reward, keep reward values and shift reward values each have equal values $\forall{j}$. For the phoneme-wise reward, there are unique keep and shift reward values for each $j$.
\iffalse
\subsection{Policy}
Policy($\pi$) determines the agent's optimal action at a certain time. Given reward feedback in training step $t$, the policy for segment-wise reward is as follows:
\begin{equation}
\pi_{t} = {a}_{t}\cdot{r}_{t}
\end{equation}
where $a=\begin{cases}A_{c},&\mbox{if\enskip}r_{p}=1 \\A_{s},&\mbox{if\enskip}r_{s}=1\end{cases}$.
$a_{t}$ and $r_{t}$ are fixed actions and rewards for each of our duration prediction segments in training step $t$. The policy for phoneme-wise reward is:
\begin{equation}
\pi_{nt} = {a}_{pnt}\cdot{r}_{pnt} + {a}_{snt}\cdot{r}_{snt}
\end{equation}
here, $a_{pnt}$, $a_{snt}$ represent each phoneme-wise and segment-wise actions, and $r_{pnt}$, $r_{snt}$ represents each phoneme-wise and segment-wise rewards, respectively.
\fi
\subsection{Gaussian Upsampling}
Predicted durations are scaled to the length of frame sequence outputs. Each scaled predictions are used to find cumulative sum of scaled token lengths and their center positions, as introduced in \cite{donahue2020end}. We first compute the weights:
\begin{equation}
w_{t}^{i} = \cfrac{\exp({\sigma^{-2}(t-c_{i})^{2})}}
{ \sum_{j=1}^{i}\exp({\sigma^{-2}(t-c_{j})^{2})} }
\end{equation}
given fixed temperature parameter $\sigma^{2}$, scaled
token center position $c$, and time step $t$. We finalize upsampling by producing weighted-sum between the encoder outputs and weights. The output features are used as decoder inputs and transposed convolutions upsample the features to a raw waveform.
\subsection{Reinforced Duration Loss}
We provide the agent with appropriate feedback by a loss that incorporates the rewards and duration prediction actions. For each phoneme sequence index $j$, the duration values are compared between the original duration prediction and the duration of the selected action. The loss is defined as:
\begin{equation}
\mathcal{L}_{re} = \sum_{j=1}^{N}(||{D}_{predj}-({D}_{kj} \times {r}_{kj} + {D}_{sj} \times {r}_{sj})||_{1})
\end{equation}
Given $N$ total tokens, each $D_k$, $r_k$ pairs and $D_s$, $r_s$ pairs represent duration and reward values for keep, shift actions, respectively. $D_{pred}$ is equal to $D_k$ because the keep action does not shift the predicted duration values. Therefore, the loss returns positive loss for shift action, and zero loss for keep action.
\section{Auxiliary Loss}
We utilize GAN loss \cite{mao2017least} and reinforced duration loss. Additionally, auxiliary losses were used to support training of our text-to-wave network.
\subsection{Total Duration Loss}
The main purpose of the aligner is to produce accurate alignments of phoneme-to-frame sequences. However, the aligner does not have "correct" duration alignments to refer to during training, and therefore is not certain the duration outputs are accurate. Therefore, total duration loss is utilized as guidance for the model. We refer to the aligner length loss in \cite{donahue2020end}. Let $m_{length}$ be the length of ground-truth mel-spectrogram, and $l_{j}$ be the predicted length of $j$th token. The total duration loss is:
\begin{equation}
\mathcal{L}_{total} = ({m}_{length}-\sum_{j=1}^{i}{l}_{j})^{2}
\end{equation}
\subsection{Mel-Spectrogram Loss}
In \cite{kong2020hifi}, mel-spectrogram loss is mentioned to be able to optimize the generator and improve quality of generated waveforms. Additionally, reward designs for the reinforce-aligner depends on the mel-spectrogram loss values to produce quality feedback for the agent in our model. The loss is formulated as:
\begin{equation}
\mathcal{L}_{mel} = \sum_{t=1}^{T}(||{mel}_{gt}[t]-{mel}_{pred}[t]||_{1})
\end{equation}
where ${mel}_{gt}$, ${mel}_{pred}$ are mel-spectrograms of ground-truth and synthesized waveforms for $T$ time steps.
\subsection{Soft Dynamic Time Warping}
We enable the mel-spectrograms to have room for error by iteratively finding an alignment path between ground-truth and synthesized spectrograms with dynamic time warping (DTW) \cite{sakoe1971dynamic, donahue2020end}. The main objective of this method is to alleviate the requirement that both spectrograms must be exactly aligned.
The total cost is defined as:
\begin{equation}
{c}_{total} = \sum_{l=1}^{L}(\delta_{l} \cdot\omega + \sum_{t=1}^{T}(||{mel}_{gt}[t]-{mel}_{pred}[t]||_{1}))
\end{equation}
where $L$ is the mel-spectrogram length, and $\omega$ is the warp penalty that occurs for actions 2 and 3. $\delta$ is a binary indicator that is 1 when warp penalty is greater than zero. For our loss, we use the soft minimum from \cite{cuturi2017soft} to produce Soft-DTW.
\begin{table}[!t]
\caption{MOS with $95\%$ confidence intervals and duration prediction error results of the ablation study.}
\centering
\begin{tabular}{l|c|c} \Xhline{3\arrayrulewidth}
\textbf{Setting} &\textbf{MOS} &\textbf{Dur. Error} \\ \hline
Ground truth &${4.52}\pm{{0.06}}$ &- \\ \hline
w/ duration target &${4.42}\pm{{0.04}}$ &${1.03}$ \\ \hline
w/o Soft-DTW &$1.49\pm{0.08}$ &$2.85$ \\
w/ Soft-DTW &$2.91\pm{0.06}$ &$2.53$ \\
Segment-wise ($\alpha=1$) &$3.48\pm{0.08}$ &$2.24$ \\
Segment-wise ($\alpha=2$) &$3.93\pm{0.07}$ &$1.84$ \\
Phoneme-wise ($\alpha=1$) &$3.78\pm{0.05}$ &$2.14$ \\
Phoneme-wise ($\alpha=2$) &$\textbf{4.06}\pm{\textbf{0.06}}$ &$\textbf{1.78}$ \\
\Xhline{3\arrayrulewidth}
\end{tabular}
\label{ablation_table}
\end{table}
\section{Experimental Results and Analysis}
\subsection{Experimental Settings}
For our experiments, we use the National Institute of Korean Language (NIKL) corpus \cite{niklcorpus} and the LJSpeech dataset \cite{ljspeech17}. The NIKL corpus contains about 45K audio samples of 50 native Korean speakers reading Korean text. The LJSpeech dataset contains 13,100 samples recorded by a single English female speaker. For both datasets we unified the sampling rates to 22,050 Hz. In NIKL corpus, we randomly split 10 audio samples for validation, 10 audio samples for testing, and the rest for training in all of the speakers. In LJSpeech dataset, we randomly divided 300 samples for validation, 300 samples for testing, and the rest for training. We conducted our experiments on 4 NVIDIA A100 GPUs.
\subsection{Training Setup}
All experiments are conducted with batch size of 64, using AdamW optimizer \cite{loshchilov2017decoupled} with $\beta_{1}$=0.8, $\beta_{2}$=0.99, and weight decay of $\lambda$=0.01. Learning rate decay of 0.999 was used for each epoch, and the initial learning rate started with 0.0002. We used hidden representation of 256 , frame segment $\gamma$=128, fixed temperature parameter $\sigma^{2}$=10, and other hyper-parameters identical to HiFi-GAN V1, V2 \cite{kong2020hifi} for all models to ensure fairness of experiments. The FFT, window size, and hop size were set to 1024, 1024, and 256. The english texts were converted to phoneme using the method of \cite{g2pE2019}. All models were trained for 350k steps.
\subsection{Ablation Study}
We conduct ablation study with NIKL corpus to determine the optimal reward for our model. We feed in speaker embedding alongside the encoder output to the reinforce-aligner for training in the multi-speaker dataset. All models are trained with the text-to-waveform network proposed in this paper, and each settings are variations on the alignment method, reward, and shift values. "with duration target" setting represents experiments conducted with attention alignments extracted from Tacotron 2 \cite{shen2018natural}. We synthesize 100 utterances randomly sampled from NIKL corpus test dataset. Afterwards, 12 subjects rated the quality of synthesized speech with scores in the range of 1 (worst) to 5 (best) in increments of 1. MOS score and duration error results are displayed in Table \ref{ablation_table}. "Phoneme-wise ($\alpha=1$)" represents a phoneme-wise reward shifted by scalar value of 1 in alternating signs. In table \ref{ablation_table}, "Phoneme-wise ($\alpha=2$)" showed the highest MOS score and lowest duration error. The duration error is the L1 loss between predicted duration and target duration extracted from attention alignments of Tacotron 2 . Additionally, Figure \ref{fig:3} visualizes the effectiveness of each settings with different alignment methods.
\subsection{Comparison with Other Methods}
We compare our model with other state-of-the-art methods developed for the task of TTS. Both Glow-TTS and BVAE-TTS use internal aligners. In this comparison, we utilize the "Phoneme-wise ($\alpha=2$)" setting, which represents the best MOS from our ablation study. Our model produces raw waveforms directly from text, and other methods require a vocoder to synthesize waveform from produced spectrograms. For the vocoder, HiFi-GAN \cite{kong2020hifi} is used. We conducted a subjective 5 scale MOS test on Amazon Mechanical Turk \cite{mturk}. At least 20 subjects rated naturalness of audio on a scale of 1 to 5 with 1 point increments. In Table \ref{main_table}, we display the MOS score followed by computed MCD$_{13}$ \cite{kubichek1993mel} and RMSE$_{f0}$ \cite{hayashi2017investigation} results. We synthesize 100, 200 utterances for subjective, objective evaluations, respectively. Our model has the highest MOS score, and lowest MCD$_{13}$, RMSE$_{f0}$ values.
\begin{table}[!t]
\caption{MOS with $95\%$ confidence intervals, MCD$_{13}$, and RMSE$_{f0}$ results of proposed and state-of-the-art methods.}
\centering
\resizebox{1\columnwidth}{!}{
\begin{tabular}{l|c|c|c} \Xhline{3\arrayrulewidth}
\textbf{Method} &\textbf{MOS} &\textbf{{MCD}$_{13}$} &\textbf{RMSE$_{f0}$} \\ \hline
Ground Truth &${4.32}\pm{{0.03}}$ &- &-\\
Vocoded \cite{kong2020hifi} &$4.10\pm{0.04}$ &$1.464$ &$38.013$\\ \hline
Tacotron2 \cite{shen2018natural} &$4.02\pm{0.05}$ &$4.455$ &$39.260$\\
Glow-TTS \cite{kim2020glow} &$3.93\pm{0.05}$ &$3.944$ &$40.438$\\
BVAE-TTS \cite{lee2020bidirectional} &$3.70\pm{0.06}$ &$4.002$ &$41.686$\\
Proposed (Ours) &$\textbf{4.07}\pm{\textbf{0.04}}$ &$\textbf{2.887}$ &$\textbf{38.988}$\\
\Xhline{3\arrayrulewidth}
\end{tabular}
}
\label{main_table}
\end{table}
\section{Conclusion}
We propose an end-to-end text-to-waveform network with a novel reinforcement learning based duration alignment search method. The advantage of this model is in the agent's ability to actively search for the optimal duration alignment through action based on reward feedback. We conducted a series of experiments to select the optimal reward for our reinforce-aligner. Our proposed model was able to outperform other state-of-the-art methods with more accurate duration alignments and enhanced naturalness of synthesized audio.
\clearpage
\bibliographystyle{IEEEtran}
|
\section{Introduction}\label{S1}
\IEEEPARstart{}{}Millimeter-Wave (MmWave) multiple-input multiple-output (MIMO) techniques have a high promise for next-generation wireless communication systems due to the abundance of bandwidth\cite{1,2}. The short wave-lengths of mmWave frequencies allow a large number of antennas to be packed in a compact physical size for providing a high beamforming gain for compensating the heavy path loss of mmWave signals. Unfortunately, these densely packed antennas are subject to blockages by humans and dust, snow or water drops\cite{4,5,6}. These blockages will inevitably change the array's geometry, causing a change of the composite channels \cite{16}. Therefore, it is necessary to diagnose the blocked/faulty antennas.
Hence several antenna array diagnosis (AAD) techniques have been proposed in\cite{7,9,10,11}. Explicitly, Bucci et al. \cite{7} studied both partial blockage as well as complete blockage, and then proposed a modified genetic algorithm for calculating the characteristic parameters of blocked antennas. Bucci et al. \cite{9} also proposed a method of analyzing the near-field data to detect the location of blocked antennas. However, the AAD methods proposed in \cite{7} and \cite{9} require a large number of measurements, which will substantially increase the time required for concluding the diagnosis. To solve this problem, sophisticated AAD techniques based on compressive sensing (CS) were proposed in \cite{10,11}. Specifically, the AAD problem was formulated as a sparse signal recovery problem by Migliore \cite{10}, which can be solved by CS algorithms. By contrast, upon exploiting the correlations between blocked antennas, a group-blockage diagnosis technique was proposed by Eltayeb et al.\cite{11}. However, this method cannot be used, when both the receive and transmit antennas are blocked at the same time.\\
\hspace*{0.36cm}Against this background, we propose a cross-entropy-inspired AAD method (CE-AAD) for detecting the location of blocked antennas and for calculating the corresponding characteristic parameters of blocked antennas. Specifically, our proposed CE-AAD exploits the correlations among the blocked antennas, where the required number of measurements is significantly reduced. Then, we extend the proposed CE-AAD method to the scenario, where blockages occur simultaneously both at the transmitter and receiver. Finally, our simulations show that the proposed CE-AAD method outperforms its traditional counterparts. \\
\emph{Notation}: We use the following notation throughout the paper. We let a, $\mathbf{a}$, $\mathbf{A}$ represent the scalar, vector, and matrix respectively. $\rm vec\{\cdot \}$ denotes the vectorization of a matrix. $\rm ivec\{\cdot \}$ denotes the invectorization of a vector.
$(\cdot)^{\rm{T}}$,
$(\cdot)^{\text{H}}$, and $(\cdot)^{-1}$ denote the transpose,
conjugate transpose, and inverse of a matrix, respectively. The operator $ \circ $ and $\otimes$ represent the Hadamard-product and Kronecker product, respectively. $\mathbf {1}_{{N} \times {M}}$ denotes all $ {1}$ matrix of size $N \times M$.
\section{System Model}\label{S2}
In this section, we consider a uniform planar array (UPA) of antennas at the transmitter with $N_x$ equally spaced elements along the x-axis, and $N_y$ equally spaced elements along the y-axis. Thus the total number of antennas at the transmitter is $N_{\rm{T}}=N_x\times N_y$ . The receiver is equipped with a single antenna. At the $k$-th measurement $(k =1,2 \cdots K)$, the received signal $g_k$ can be described as
\begin{align}\label{eq_1}
g_k=\mathbf{f}^{\rm{T}}_k\cdot {\rm{vec}}(\mathbf{H}){s} + {{n_k}},
\end{align}
where $\mathbf{H}\in \mathbb{C}^{N_x\times N_y}$ is the channel matrix, $\mathbf{f}_k\in \mathbb{C}^{N_T\times 1}$ is the transmit precoding (TPC) vector at the $k$-th measurement, and $s$ is the transmitted symbol, while ${n_k} \sim {CN}(0,{\delta}^2)$ is the additive noise and ${\delta}^2$ is the noise power. Considering the classical multi-path channel model, $\mathbf{H}$ can be written as \cite{17}
\begin{align}\label{eq_1}
\mathbf{H} = \sum\limits_{{\ell} = 1}^L {{\beta _{\ell}}\cdot \mathbf{A}({\theta _{\ell}},{\varphi _{\ell}})},
\end{align}
where $L$ is the number of propagation paths and $\beta _{\ell}$ is the complex gain of the $\ell$-th path. Furthermore, $\mathbf{A}({\theta _{\ell}},{\varphi _{\ell}}) \in \mathbb{C}^{N_x \times N_y}$ is the antenna array response, where $\theta _{\ell}$ is the elevation angle-of-departure (AoD) and $\varphi _{\ell}$ is the azimuth AoD. The $(m,n)$-th element of $\mathbf{A}({\theta _{\ell}},{\varphi _{\ell}})$ is given by ${\frac{1}{{\sqrt {{N_x}{N_y}} }}e^{j(m-1)\frac{{2\pi d_x}}{\lambda }\sin \theta _{\ell} \cos \varphi _{\ell}}}{e^{j(n-1)\frac{{2\pi d_y}}{\lambda }\sin \theta _{\ell} \sin \varphi _{\ell} }}$ \cite{17}, where ${d_x}$ and ${d_y}$ are the antenna spacing along the x-axis and y-axis, $\lambda$ is the wavelength.
When the antenna array is blocked, the received signal at the $k$-th measurement is given by
\begin{align}\label{eq_3}
r_k = \mathbf{f}^{\rm{T}}_k\cdot(\mathbf{b} \circ {\rm{vec}}(\mathbf{H}))s + {n_k} {,}
\end{align}
where $\mathbf{b}\in {\mathbb{C}^{N_{\rm{T}}{\rm{ \times }}1}}$ and the $n$-th element of vector ${\mathbf{b}}$ can be defined as
\begin{align}\label{eq_4}
{b_n} = \left\{ {\begin{array}{*{20}{c}}
{{a_n}}\\
1
\end{array}\begin{array}{*{20}{c}}
{{\rm{if}}{\mkern 1mu} {\kern 1pt} {\rm{the}}{\mkern 1mu} {\kern 2pt} n{{ {\tiny{-}} {\rm{th}}}}{\mkern 1mu} {\kern 2pt} {\rm{antenna}}{\mkern 2mu} {\kern 2pt} {\rm{is}}{\mkern 1mu} {\kern 2pt} {\rm{blocked}}}\\
{{\rm{otherwise}}}
\end{array}} \right. ,
\end{align}
where ${a_n} = {\tau _n}{e^{i{\Psi _n}}}$ is the characteristic parameter of the blocked antenna $n$. Here ${\tau _n}$ is the absorption coefficient following the uniform distribution between 0 and 1. If the $n$-th antenna is completely blocked, ${\tau _n}$ is equal to 0. Otherwise, ${\tau _n}$ is a positive value less than 1. The variable ${\Psi _n}$ denotes the blockage scattering coefficient following the uniform distribution between 0 and $2\pi$ \cite{20}.
Without loss of generality, we let the training symbol $s$ be equal to 1. Using (1) and (3), we have
\begin{align}\label{eq_5}
{y_k} = {r_k} - {g_k} = {\mathbf{f}_k^{\rm{T}}}\underbrace {(\rm{vec}(\mathbf{H}) \circ \mathbf{b} - vec(\mathbf{H}))}_\mathbf{q} + {\widetilde n_k},
\end{align}
where $\mathbf{q}=[q_1,q_2,\cdots, q_{N_{\rm{T}}}]^{\rm{T}} \in \mathbb{C}^{{N_{\rm{T}}}\times1}$ and ${\widetilde n_k}\sim CN(0,{\delta}^2)$. After \emph{K} measurements, we obtain
\begin{align}
\mathbf{y}=\mathbf{Fq}+\widetilde {\mathbf{n}},
\end{align}
where $\mathbf{y}=[y_1,y_2,\cdots y_K]^{\rm{T}} \in \mathbb{C}^{K \times1}$ and $\mathbf{F}=[\mathbf{f}_1,\mathbf{f}_2,\cdots \mathbf{f}_K]^{\rm{T}} \in \mathbb{C}^{K\times N_{\rm{T}}}$. The TPC matrix $\mathbf{F}$ can be designed as a random matrix, which can be realized by 2-bit phase shifters, whose elements are randomly selected from $\{1+i, 1-i, -1-i, -1+i\}$ \cite{11}. Since the number of blocked antennas is usually small, the vector $\mathbf{q}$ is a sparse vector. Thus, the AAD problem becomes a sparse signal recovery problem.
Through the estimation of $\mathbf{q}$, which is denoted by $\hat{\mathbf{q}}$, the locations of blocked antennas can be recognized by finding the positions of non-zero elements of $\widehat {\mathbf{q}}$. Then, the characteristic parameters of blocked antennas ${\tau _n}$ and ${\Psi _n}$ can be readily obtained as
${\tau _n} = \left| {\frac{{{{\widehat q}_{n}}}}{{{{{{h}}}_{n}}}} + 1} \right|$ ,${\Psi _n} = \angle (\frac{{{{\widehat q}_{n}}}}{{{{ {{h}}}_{n}}}} + 1)$, where ${n \in \mathcal{S}_{q}}$, which is the index set of non-zero elements in the vector $\mathbf{q}$. Furthermore, $\mathbf{ {\mathbf{h}}}=\rm{vec}(\mathbf{H})$ and ${ {h}_n}$ is the $n$-th element of $\mathbf{ {\mathbf{h}}}$. While $\widehat q_n$ denotes the $n$-th element of $\widehat {\mathbf{q}}$.
\section{Proposed CE-AAD Algorithm}\label{S3}
Let us now outline the CE-AAD algorithm by exploiting the correlations between the adjacent blocked antennas to detect the location of blocked antennas and calculate the corresponding characteristic parameters.
Since numerous antennas are packed in a compact physical size in mmWave MIMO system, several neighboring antenna elements might be blocked. This results in $\mathbf{Q}$ = {\rm ivec}($\mathbf{q}$) $\in \mathbb{C}^{N_x\times N_y}$ of (6) becoming a block-sparse matrix\cite{11}. An example of the block-sparse matrix $\mathbf{Q}$ is shown in Fig. 1.
\begin{figure}[t]
\center{\includegraphics[height=1.7in,width=2.65in]{upa_0810.eps}}
\caption{An example of the block-sparse matrix $\mathbf{Q}$. }
\label{ula}
\end{figure}
By exploiting the block-sparsity of matrix $\mathbf{Q}$, we propose a CE-AAD algorithm for recovering Q using a low number of measurements. Specifically, we use the vector $\mathbf{d}$ of size $N_{\rm{T}} \times 1$ to denote the positions of non-zero elements of vector $\mathbf{q}$, where the elements of $\mathbf{d}$ are selected from \{0,1\}. That is to say, ${q_n}=0$ means $d_n=0$ and ${q_n}\ne 0$ means ${d_n}=1$, where $d_n$ is the $n$-th element of $\mathbf{d}$. Thus, we can rewrite (6) as
\begin{align}
\mathbf{y}=\mathbf{F} ({\mathbf{q}} \circ {\mathbf d})+\widetilde {\mathbf{n}}.
\end{align}
Next, we propose the CE-AAD algorithm for estimating $\mathbf{d}$ and $\mathbf{q}$, as shown in Algorithm 1. Firstly, we initialize the algorithm by setting the probability matrix to ${\mathbf{P}}_{\rm B}^0 = {\frac{1}{2}} \times \mathbf{1}_{({N_x}/{N_{b_x}}) \times ({N_y}/{N_{b_y}})}$, where we assume that each block is of size ${N_{b_x}} \times {N_{b_y}}$.
\begin{algorithm}
\caption{The proposed CE-AAD algorithm}
\KwIn{Precoding matrix $\mathbf{F}$; Measurement vector $\mathbf{y}$; Number of candidates $N_c$; Number of elites $N_e$; Number of iterations ${N_{iter}}$.}
$\mathbf{Initialization}$:
${\mathbf{P}}_{\rm B}^0= {1/2} \times {\mathbf{1}_{{({N_x}/{N_{b_x}})} \times ({N_y}/{N_{b_y}})}}$; $i=0$.
\For{$0\le i \le N_{iter}-1$}
{
1. Randomly generate $N_c$ candidate vectors $\{ {\mathbf{d}_{n_c}}\} _{n_c = 1}^{N_c}$ according to ${\mathbf{P}_{\rm{B}}^{i}}$, where $\mathbf{d}_{n_c} \! \in\! \{0, 1\}^{N_{\rm{T}} \times 1}$;
2. ${\widehat {\mathbf{q}}^i}_{n_c} \!=\! {({\mathbf{F}_{n_c}^{\rm{H}}}\mathbf{F}_{n_c})^{ - 1}}{\mathbf{F}_{n_c}^{\rm{H}}}\mathbf{y}$;
3. ${\zeta }^i_{n_c} = ||\mathbf{y} - \mathbf{F}_{n_c}{\widehat {\mathbf{q}}^i}_{n_c}|{|_2} + \epsilon ||{\mathbf{d}_{n_c}}|{|_0}$;
4. Sort the objective function ${\zeta }^i_{n_c}$ in ascending order: $ {\zeta }^i_{d_{e,1}} \le {\zeta }^i_{d_{e,2}} \le\cdots \le{\zeta }^i_{d_{e,{N_{\rm{c}}}}}$;
5. ${{\mathbf{P}_{\rm{B}}^{i + 1}}\! =\! \frac{1}{{N_e}}({\mathbf{C}_{\rm{B}}^{d_{e,1}}}\! + \!{\mathbf{C}_{\rm{B}}^{d_{e,2}}}\! + \! \cdots \!+ \!{\mathbf{C}_{\rm{B}}^{d_{e,N_e}}})}$;
6. $i=i+1$;
}
\KwOut{Estimated value ${\widehat {\mathbf{q}}}$}
\end{algorithm}
\hspace*{-0.36cm}In Step 1, we divide the matrix $\mathbf{Q}$ into $({N_x}/{N_{b_x}}) \times ({N_y}/{N_{b_y}})$ blocks by exploiting the block-sparsity of $\mathbf{Q}$. We then define the probability matrix in the $i$-th iteration $\mathbf{P}_{\rm{B}}^i\in {\mathbb{C}^{N_x/{N_{b_x}} \times {N_y}/{N_{b_y}}}}$ , where the $(m,n)$-th element of $\mathbf{P}_{\rm{B}}^i$ denotes the probability that $\mathbf{C}_{\rm{B}}^{n_c}(m,n)=1$. Here, $\mathbf{C}_{\rm{B}}^{n_c}(m,n)$ is the $(m,n)$-th element of $\mathbf{C}_{\rm{B}}^{n_c} \in {\{0,1\}}^{({N_x}/{N_{b_x}})\times({N_y}/{N_{b_y}})}$, representing whether the $(m,n)$-th block of $\mathbf {Q}$ is a non-zero block. We generate $N_c$ candidates $\{ {\mathbf{C}_{\rm{B}}^{n_c}}\} _{n_c = 1}^{N_c }$ based on the probability matrix $\mathbf{P}_{\rm{B}}^i$. Then, we vectorize $\mathbf{C}_{n_c}$ as $\mathbf{d}_{n_c}= {\rm{vec}}(\mathbf{C}_{n_c}) \in \mathbb{R}^{N_{\rm{T}} \times 1}$, where $\mathbf{C}_{n_c}= \mathbf{C}_{\rm{B}}^{n_c} \otimes \mathbf {1}_{{N_{b_x}} \times {N_{b_y}}} \in \mathbb{R}^{N_x \times N_y} $. We also define the index set $\mathcal{D}_{n_c}$ for recording the indices of non-zero elements in vector $\mathbf{d}_{n_c}$.
In Step 2, we use the least squares (LS)\cite{21} algorithm to compute ${\widehat {\mathbf{q}}^i}_{n_c}\in {\mathbb{C}}^{N_{n_c}\times 1}$, which is the estimate of the non-zero elements of $\mathbf{q}$ associated with the $n_c$-th candidate ${\mathbf d}_{n_c}$ in the $i$-th iteration, where $N_{n_c}$ is the number of none-zero elements in candidate ${\mathbf{ d}_{n_c}}$. We then define a submatrix $\mathbf{F}_{n_c} \in {\mathbb{C}}^{K \times N_{n_c}}$, which is composed of the column vectors of matrix $\mathbf F$ with indices belonging to $\mathcal{D}_{n_c}$.
In Step 3, we define an objective function as ${\zeta }^i_{n_c} = ||\mathbf{y} - \mathbf{F}_{n_c}{\widehat {\mathbf{q}}^i}_{n_c}|{|_2} + \epsilon ||{\mathbf{d}_{n_c}}|{|_0}$, where $||{\mathbf{d}_{n_c}}|{|_0}$ is the ``$\mathcal{L}0$" norm of $\mathbf{d}_{n_c}$, i.e. the number of none-zero elements in $\mathbf{d}_{n_c}$, while $\epsilon$ is a super-parameter \cite{18} \cite{19}, which is set to 0.6 for our simulations.
In Step 4, we sort ${\zeta }^i_{n_c}$ in an ascending order and retain the first $N_e$ elements. Then, we select the corresponding $N_e$ candidates $\mathbf{d}_{n_c}$ as elites and record their indices as $\{d_{e,1} ,d_{e,2}, \cdots ,d_{e,N_e}\}$.
In step 5, we update the probability matrix as ${{\mathbf{P}_{\rm{B}}^{i + 1}}\! =\! \frac{1}{{N_e}}({\mathbf{C}_{\rm{B}}^{d_{e,1}}}\! + \!{\mathbf{C}_{\rm{B}}^{d_{e,2}}}\! + \! \cdots \!+ \!{\mathbf{C}_{\rm{B}}^{d_{e,N_e}}})}$.
We then repeat this procedure until the maximum number of iterations $N_{iter}$ is reached.
The traditional cross-entropy (CE) algorithm \cite{14} ignores the block-sparsity of matrix $\mathbf{Q}$. In contrast to the CE algorithm, the proposed CE-AAD algorithm generates ${\mathbf{d}}_{n_c}$ by exploiting the block-sparsity of $\mathbf{Q}$ in Step 1. The corresponding candidate matrix $\mathbf{C}_{n_c}$ exhibits a block-structure. In this way, the CE-AAD algorithm achieves better performance of recovery and get faster convergence.
Furthermore, when complete blockages occur, we have $a_n=0$. Then, the proposed algorithm will be further simplified, since Step 2 and Step 3 can be combined. The objective function becomes ${\zeta }^i_{n_c} = ||\mathbf{y} - \mathbf{F}{\mathbf{d}_{n_c}}|{|_2} + \epsilon ||{\mathbf{d}_{n_c}}|{|_0}$. Hence it is not necessary to estimate the value of ${\widehat {\mathbf{q}}^i}_{n_c}$ by exploiting the LS algorithm. After ${N_{iter}}$ iterations, $\mathbf{d}_{d_{e,1}}$ is the estimated value of $\mathbf {q}$.
\section{Joint CE-AAD}\label{S4}
In the previous section, we proposed the CE-AAD for the case where blockages only occur at the transmitter. In this section, we extend the proposed CE-AAD algorithm to the case where blockages simultaneously occur both at the transmitter and receiver.
We consider the uniform linear arrays (ULAs) of antennas at both the transmitter and receiver with $N_{\rm t}$ transmit and $N_{\rm r}$ receive antennas\footnote{The proposed technique can be easily extended to the UPA scenario.}, respectively. Similar to (5), at the $k$-th measurement, we have
\begin{figure}[t]
\center{\includegraphics[height=1.9in,width=2.7in]{ula_0810.eps}}
\caption{An example of sparse structure of matrix $\mathbf{Q}$.}
\label{ula}
\end{figure}
\begin{align}\label{eq_13}
{y_k} = {r_k} - {g_k} = \mathbf{w}_k^{\rm{H}}\underbrace {(\mathbf{H} \circ \mathbf{B} - \mathbf{H})}_{\mathbf{Q}}{\mathbf{f}_k} + {\tilde n_k},
\end{align}
where ${\mathbf{w}_k}\in {\mathbb{C}^{{N_{\rm r}} \times 1}}$ denotes the combining vector at the receiver and ${\mathbf{f}_k}\in {\mathbb{C}^{{N_{\rm t}} \times 1}}$ represents the TPC vector at the transmitter. Furthermore, $\mathbf{B}=\mathbf{b}_{\rm r}\mathbf{b}_{\rm t}^{\rm{T}}$, where $\mathbf{b}_{\rm r} \in {\mathbb{C}^{{N_{\rm r}} \times 1}} $ and $\mathbf{b}_{\rm t} \in {\mathbb{C}^{{N_{\rm t}} \times 1}}$ denote the blockage coefficients at the receiver and transmitter, which have similar definitions as $\mathbf{b}$ in (3), while $\mathbf{H} \in {\mathbb{C}^{N_{\rm r} \times N_{\rm t}}}$ is the channel matrix of
\begin{align}\label{eq_13}
\mathbf{H} = \sum\limits_{\ell = 1}^L {{\beta _\ell}{\mathbf{a}_{\rm r}}({\theta _\ell^r})} {\mathbf{a}_{\rm t}^{\rm H}}({\theta _\ell^t}),
\end{align}
where $L$ is the number of propagation paths and $\beta _l$ is the complex gain of the $\ell$-th path. Furthermore, $\mathbf{a}_{\rm r}({\theta _\ell^r}) \in {\mathbb C}^{N_{\rm r} \times 1}$ and $\mathbf{a}_{\rm t}({\theta _\ell^t})\in {\mathbb C}^{N_{\rm t} \times 1}$ denote the antenna array responses at the receiver and transmitter, respectively. The steering vector $\mathbf{a}_{\rm x}({\theta _\ell^{\rm x}})$ can be expressed as $[1,{e^{ - j2\pi \frac{d}{\lambda }\sin (\theta _\ell^{\rm x} )}},\cdots,$ ${e^{ - j2\pi \frac{{({N_{\rm x}} - 1)d}}{\lambda }\sin (\theta _\ell^{\rm x})}}]^{\rm T}$, where ${\rm x} \in \{\rm t,r\}$ and $d$ is the antenna spacing. Vectorizing the matrix $\mathbf{Q} \in {\mathbb{C}^{N_{\rm r} \times N_{\rm t}}}$, (8) can be rewritten as
\begin{align}\label{eq_14}
{y_k} = \underbrace {({{\mathbf{f}}}_k^{\rm{T}} \otimes {\mathbf{w}}_k^{\rm{H}})}_{{\mathbf{u}}_k}\underbrace {\rm{vec}({\mathbf{Q}})}_\mathbf{q} + {\widetilde n_k},
\end{align}
where the matrix $\mathbf{Q}$ has a sparse structure, as shown in Fig. 2. Therefore, after $K$ measurements, we have
\begin{align}
\mathbf{y}=\mathbf{Uq}+{\widetilde {\mathbf n}},
\end{align}
where $\mathbf y = [y_1,y_2,\cdots, y_K]^{\rm T} \in \mathbb C^{K \times 1}$ and $\mathbf U=[{{\mathbf u}_1^{\rm T},{\mathbf u}_2^{\rm T},\cdots,{\mathbf u}_K^{\rm T}}]^{\rm T} \in \mathbb C^{K \times N_rN_t}$. The joint antenna array diagnosis becomes a sparse signal recovery problem.
\begin{figure*}[t]\centering\subfigure[] {\includegraphics[height=2.1in,width=2.25in,angle=0]{t1.eps}}\subfigure[] {\includegraphics[height=2.1in,width=2.25in,angle=0]{t2.eps}}\subfigure[] {\includegraphics[height=2.1in,width=2.25in,angle=0]{t3.eps}}\caption{NMSE performance comparison: (a) and (b) show the NMSE comparison when blockages occur at the transmitter. (c) shows the NMSE comparison when blockages occur at both the transmitter and receiver.}\label{fig5}\end{figure*}
Since $\mathbf Q$ has different sparse structures, the CE-AAD algorithm of this section is different from that in Section III. The main difference is reflected in Step 1 and Step 5.
In Step 1, we randomly generate $N_c$ candidate vectors $\mathbf{d}^{n_c}_{\rm t} \in {\{0,1\}^{{N_{\rm{t}}} \times 1}}$ according to the probability vector $\mathbf{p}_{\rm t}^i \in {\mathbb{R}}^{N_t \times 1}$ as well as $N_c$ candidate vectors ${\mathbf{d}}^{n_c}_{\rm r} \in {\{0,1\}^{{N_{\rm{r}}} \times 1}}$ according to the probability vector $\mathbf{p}_{\rm r}^i \in {\mathbb{R}}^{N_r \times 1}$. Then we have $N_c$ candidates $\mathbf{d}_{n_c} = \rm{vec}(\mathbf{C}_{n_c})$, where $\mathbf{C}_{n_c}(m,n) = \mathbf{d}^{n_c}_{\rm r}(m) \oplus \mathbf{d}^{n_c}_{\rm t}(n) $ denotes the $(m,n)$-th elements of $\mathbf{C}_{n_c}$. The operator $\oplus$ denotes the `or' operation. In Step 5, we use ${N_e}$ elites ${\mathbf{d}}^{d_{e,k}}_{\rm r}$ and ${\mathbf{d}}^{d_{e,k}}_{\rm t}$ $(k=1,2,\cdots N_e)$ to update the probability $\mathbf{p}^{i+1}_{\rm r}$ and $\mathbf{p}_{\rm t}^{i+1}$, where $d_{e,k}$ has the same definition as in Section III. After obtaining the estimation of $\mathbf{q}$, we can identify the locations of blocked antennas and calculate the corresponding characteristic parameters.
\section{Simulation Results}\label{S5}
In this section, we characterize the performance of the proposed CE-AAD algorithm. We adopt the ray-based mmWave channel model of \cite{15}. The system parameters used are as follows: $d=\frac{d_x}{\lambda}=\frac{d_y}{\lambda}=\frac{1}{2}$, $N_c=400$, $N_e=50$, $N_{iter}=20$, ${\epsilon = 0.6}$, ${\tau}_n \sim U(0,1)$, and ${{\Psi}_n \sim U(0,2\pi)}$. The number of dominant paths $L$ is 10, $\beta _\ell$ obeys the Gaussian distribution with zero mean and unit variance, while $\theta_\ell, \varphi_\ell, \theta_\ell^r, \theta_\ell^t$ are randomly chosen from $[-\pi/2,\pi/2]$. We assume that the blockage probability of antennas $p_b$ is 0.1. We simulate the normalized mean square error (NMSE) of $\mathbf b$ and $\mathbf{B}$ estimated both by the proposed CE-AAD method and by a pair of traditional AAD methods [10], \cite{13}.
In Fig. 3 (a) and (b), we consider a single free antenna at the receiver and a UPA having $10\times 10$ antennas at the transmitter. Fig. 3 (a) shows the NMSE comparison of different algorithms vs. the number of measurements when the signal-to-noise ratio (SNR) $\frac{1}{{\delta}^2}$ is 5dB. NMSE-ideal represents the performance upper-bound, which is achived when the exact locations of blocked antennas are known. We observe that the proposed algorithm outperforms other algorithms.In Fig. 3 (b), we compare the NMSE performance of the proposed algorithm and NMSE-ideal at different SNRs. We observe that the proposed algorithm achieves near-ideal performance using only a few measurements.
In Fig. 3 (c), we consider the case where blockages occur at both the transmitter and receiver, which rely on ULAs having 10 elements. Fig. 3 (c) shows our NMSE performance comparison vs. the SNRs, when the number of measurements is 50. We observe that the proposed joint CE-AAD algorithm has a good performance.
\section{Conclusions}\label{S5}
In this paper, we studied the AAD problem of mmWave MIMO systems. By exploiting the correlations between adjacent blocked antennas, we proposed the CE-AAD algorithm for identifying the locations and characteristic parameters of blocked antennas when blockages occurs at the transmitter. Then, we extended the proposed CE-AAD algorithm to the case, where blockages occur both at the transmitter and receiver simultaneously. Our simulation results verify that the proposed method achieves near-optimal performance, approaching that of the genie solution having explicit knowledge of the blocked antenna indices.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
The plasma, which is termed as the fourth state of the matter, is a collection of negatively charged electrons and positively charged ions. These charged species (electrons and ions) are not free but strongly affected by the electromagnetic field of surrounding charged species and capable of exhibiting the collective response. In laboratory plasma, a fraction of neutrals is present along with the charged species and the ionization fraction depends on the density of neutrals. However, the plasma is resulting from the ionization of neutral atoms; therefore, it generally contains nearly equal numbers of positive and negative charge carriers and is termed as quasi-neutral plasma. In hot-filament discharges, the emitted energetic electrons from heated filaments are the source of ionization, and the chamber wall is a sink for the plasma losses. In steady-state, the ionization rate (plasma production rate) and loss rate are same to hold the quasi-neutrality of plasma \cite{chenplasmaphysicsbook,bittencourtplasmaphysicsbook,particlebalancemodel1}.\par
The response of weakly ionized hot-filament discharge to a biased auxiliary metal electrode above the plasma potential gives some very interesting features such as formation of fire ball\cite{fireballstenzel1,fireballstenzel2}, excitation of solitary electron hole \cite{satyasolitaryelectronhole,satyatransientrespons}, electrostatic confinement\cite{electrostaticconfinment1,globalambipolardiffusion}, plasma potential locking\cite{plasmapotentiallocking,anodesizeplasmareponse} etc. At a given discharge condition, the size of an auxiliary metal electrode (diameter in case of disk electrode) and bias voltage above the plasma potential mainly determine the characteristics of plasma. The ratio of area, $A_d/A_w$, determines whether an electron sheath ($A_d/A_w$ $< \mu$) or a double sheath ($A_d/A_w$ $\simeq$ 1.7 $\mu$) or ion sheath ($A_d/A_w$ $>$ 1.7 $\mu$) will be formed around a positively biased ($V_b > V_p$) auxiliary electrode (disk electrode) in plasma volume \cite{plasmapotentiallocking,globalambipolardiffusion,anodesizeplasmareponse}. Here $A_d$ is area of disk electrode, $A_w$ is the plasma facing area of the vacuum chamber, $\mu$ = $\sqrt{2.3 m_e/M_i}$ and $V_b$ is the disk bias voltage. Hence characteristics of bulk plasma strongly depend on the ratio $A_d/A_w$ to $\mu$. The earlier work performed in double plasma device (HOT--filament discharge) suggests lowering the plasma density if the biased voltage of the disk is increased above the plasma potential\cite{globalambipolardiffusion,anodesizeplasmareponse}. \par
In our recent experiment \cite{mangilalcpp}, we observed the plasma density enhancement after application of transient high voltage positive pulses to the axillary disk electrode in low pressure helium plasma. It is expected to observe the plasma density enhancement if a positively biased ($V_b > V_p$) disk electrode is immersed in the hot-filament helium discharge. Furthermore, the experimental device used in our study \cite{mangilalcpp} was different than the double plasma device \cite{anodesizeplasmareponse,plasmapotentiallocking} therefore plasma response to a positively biased auxiliary electrode (disk) should be different than that reported in references\cite{plasmapotentiallocking,globalambipolardiffusion,anodesizeplasmareponse}. Some of these open questions motivate us to perform experiments in the hot-filament helium discharge to see the effect of the positively biased auxiliary disk electrode on the plasma characteristics. \par
In the present investigation, a metallic electrode in the form of disk immersed in unmagnetized helium plasma was biased with a DC voltage source and the response of plasma was studied by measuring plasma parameters with the help of a planar Langmuir probe and emissive probe. The increase in plasma potential, electron temperature and plasma density is understood by loss of bulk plasma electrons to biased disk electrode and the role played by primary electrons in the ionization process. A detailed description of the experimental set-up and plasma production is given in Sec.\ref{sec:exp_setup}. The variation of plasma parameters with bias voltage on the disk electrode and filament heating currents (primary electron population) are discussed in Sec.\ref{sec:exp_results}. The observed results are discussed in Sec.\ref{sec:results_discussion}. A brief summary of the work along with concluding remarks is provided in Section~\ref{sec:summary}.
\section{Experimental Setup} \label{sec:exp_setup}
The experiments were conducted in a grounded cylindrical chamber of stainless steel (SS) having an inner diameter of 29 cm and length of 50 cm. The vacuum chamber was evacuated with the help of a pumping assembly consisting of a combination of rotary pump and diffusion pump. The chamber was evacuated to base pressure of $\sim$ $1\times 10^{-5}$ mbar. The pressure inside the vacuum chamber was measured by a Pirani gauge. A Needle valve attached to the chamber was used to regulate the gas pressure inside the vacuum chamber. The plasma is generated by electron impact ionization of helium atoms at working pressure of $1\times 10^{-3}$ mbar by primary energetic electrons (60 to 80 eV) emitted from four DC biased hot thoriated tungsten filaments of radius 0.125 mm. These filaments were mounted on two SS rings of 12 cm radius. The filaments were heated by using a direct current power supply (32 V, 30 A) and then biased to a potential of -65 V with respect to the grounded chamber by using a discharge power supply (0--300 V, 5 A). Since electrons emitted from the heated filaments (thermionic electrons) have very low energy (0.2 to 0.3 eV), an external negative biasing is required to accelerate the thermionic emitted electrons above the ionization potential of helium gas ($\sim$ 24.58 eV). A one-sided stainless steel disk (or auxiliary electrode) of diameter 3 cm or 8 cm was used to modify the steady-state low-pressure plasma. A DC voltage supply was used for biasing the metal disk from -50 V to +50 V. A one sided-planar Langmuir probe \cite{probereview1,chenprobe1,mangilalthesis} of radius 4 mm was used to measure the plasma density ($n$) and electron temperature ($T_e$). The plasma potential at a given experimental condition was measured using the emissive probe \cite{emissivesheehan}. Schematic diagram of the experimental setup is given in Fig.\ref{fig:fig1}. \par
\begin{figure*}
\centering
\includegraphics[scale= 0.8000]{Fig1m1}
\caption{\label{fig:fig1}Schematic of the experimental set-up: (1) Experimental vacuum chamber, (2) pumping assembly, (3) stainless steel ring, (4) and (5) thoriated tungsten filaments, (6) filament heating supply, (7) discharge power supply, (8) disk bias supply, (9) ceramic tube, (10) metallic SS disk, (11) Planar Langmuir probe and (12) Langmuir probe supply.}
\end{figure*}
\begin{figure}
\includegraphics[scale= 0.3505000]{Fig2m}
\caption{\label{fig:fig2}I--V characteristics of the planar probe (smoothed) at three different bias voltages to the disk electrode. The experiment was performed in helium plasma at pressure $1\times 10^{-3}$ mbar and discharge current, 50 mA. Diameter of the disk electrode was 3 cm and probe was kept at 10 cm away from the biased metal disk on the discharge axis.}
\end{figure}
\section{Experimental Results} \label{sec:exp_results}
To see the effect of either positively or negatively biased auxiliary (additional) metal electrode (disk) on the low-pressure filament discharge, plasma parameters ($n_e, T_e$ and $V_p$) were measured using the electrostatic probes. A stainless steel metal disk of diameter, D = 3 cm or 8 cm and thickness 0.5 mm was placed at a fixed position along the discharge axis (see Fig.\ref{fig:fig1}). The backside of the disk was covered by a dielectric (Nylon) material and only the front side was exposed to helium plasma. A planar probe at 10 cm away from the metal disk on the same axis was used to get the current-voltage characteristics (I--V curve) by applying a voltage ramp of range -50 to 50 V. The plasma potential at same position was measured by a radial moving emissive probe. We also measured the plasma potential using the cold probe technique (first derivative of I--V curve of planar probe) for the given experimental conditions. Difference in the measured plasma potential by both hot and cold probe techniques was $<$ 5\%. For getting the aimed data, metal disk was biased in the potential range of -30 V to +50 V and the corresponding probe current was measured to construct the I--V characteristics for a given disk bias voltage. \par
For the first set of experiments, helium gas pressure and discharge current were kept fixed at $1 \times 10^{-3}$ mbar and 0.05 A respectively. The discharge current depends on the filament heating current (or thermionic emitted current flux) and negative DC bias voltage applied to heated filaments. In this case, filaments were heated by passing the current of 24 A, and thermionic emitted electrons were accelerated by applying the negative bias of 65 V to filaments. The I--V characteristics of the planar probe while the metal disk (D = 3 cm) was kept at -3.5 V, 30 V, and 45 V are depicted in Fig.\ref{fig:fig2}. These I--V characteristics at fixed discharge conditions indicate change in the plasma characteristics after immersing or placing a positively biased auxiliary metal disk in the plasma column. For getting the variation of plasma parameters against metal disk biasing at a given discharge condition, the I--V curves taken at various disk bias voltages were analyzed.\\
\begin{figure*}
\centering
\subfloat{{\includegraphics[scale=0.36050]{Fig3a}}}%
\hspace*{-0.5in}
\subfloat{{\includegraphics[scale=0.36050]{Fig3b}}
\qquad
\subfloat{{\includegraphics[scale=0.36050]{Fig3c}}
\caption{\label{fig:fig3}(a)Plasma potential variation, (b) Electron temperature variation and (c) plasma density variation against metal disk bias voltage. The helium pressure was set at $1\times 10^{-3}$ mbar and filament heating current was 24 A. The diameter (D) of the metal disk was 3 cm. Error over the averaged value of the measure plasma parameters are $< \pm$ 7\%.}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[scale= 0.36000]{Fig4}
\caption{\label{fig:fig4}The plasma potential ($V_p$) variation with the disk bias voltage at three different locations from the biased disk. The helium gas pressure was set at $1\times 10^{-3}$ mbar and filament heating current was 24 A. The diameter (D) of the metal disk was 8 cm. Error over the averaged value of the measure plasma potential is $< \pm$ 5\%. }
\end{figure*}
\begin{figure*}
\centering
\subfloat{{\includegraphics[scale=0.3550]{Fig5a}}}%
\hspace*{-0.5in}
\subfloat{{\includegraphics[scale=0.3550]{Fig5b}}
\qquad
\caption{\label{fig:fig5}(a) The radial plasma potential and (b) plasma density variation against the disk bias voltage. The measurements were taken at d = 10 cm. The helium gas pressure was set at $1\times 10^{-3}$ mbar and filament heating current was 24 A. The diameter (D) of the metal disk was 8 cm. Error over the averaged value of the measure plasma potential is $< \pm$ 5\%.}
\end{figure*}
In Fig.\ref{fig:fig3}(a), we have plotted the plasma potential variation against the disk bias voltages. The plasma potential varies slowly at low positive bias voltage ($<$ 20 V) and has an approximate linear growth above +20 V. The measured plasma potential is always found to be lower than the disk bias voltage. The negative value of plasma potential is expected due to the presence of primary energetic electrons (beam electrons) in the plasma volume. At low disk bias voltage, it is difficult to determine the population of these beam electrons but can be possible at large positive bias voltage on the disk electrode \cite{mangilalcpp}. The increase in plasma potential (towards more positive) is expected by removing the confined electrons of the plasma medium \cite{particlebalancemodel1,particlebalancemodel3exp,anodesizeplasmareponse}.
The rate of change of $V_p$ depends on the background plasma density, population of energetic electrons, and size of the auxiliary disk electrode. Similar to the plasma potential, the plasma density ($n_e$) and electron temperature ($T_e$) were calculated using the recorded I--V data at various metal disk bias voltages. The $T_e$ variation with the disk bias voltages is given in Fig.\ref{fig:fig3}(b). We observe an increment in $T_e$ with increasing the positive bias voltage on the disk which is expected in case of higher positive potential \cite{particlebalancemodel1,particlebalancemodel2,particlebalancemodel3exp,
anodesizeplasmareponse}. The plasma density variation against disk bias voltages is shown in Fig.\ref{fig:fig3}(c). In Fig.\ref{fig:fig3}(c), we can see that $n_e$ increases with increase in the disk bias voltage above the plasma potential. At large positive bias voltage (+50 V), plasma density is observed to be maximum. Thus, we observe the plasma density enhancement from its equilibrium (initial) value in the presence of steady state perturbation applied by a positively biased metal disk electrode. \par
In the next set of experiments, we carried out the experiments with a large diameter disk electrode (D = 8 cm) at same discharge conditions. The plasma potential with disk bias voltages at three different locations from the biased disk (d = 0 cm). For this biased disk electrode (D = 8 cm), the ratio $A_d/A_w < \mu$ which suggests a slow variation of plasma potential with biasing voltage to the disk \cite{globalambipolardiffusion,plasmapotentiallocking}. However, the plasma potential increases at a higher rate than that expected from the theoretical model as we increase the bias voltage to metal disk (see Fig.\ref{fig:fig4}). The plasma potential profiles are nearly same at different locations from biased disk which confirms the effect of the positively biased electrode on the entire bulk plasma (or plasma column). The plasma density and $T_e$ increase with increasing the bias voltage to disk electrode and found a nearly similar trend that of observed for small-sized biased metal disk (see Fig.\ref{fig:fig3}). With the same experimental configuration, we also measured plasma potential and plasma density radially from the discharge axis ($r$ = 0 cm) at d = 10 cm. In fig.\ref{fig:fig5}, we see a slight change in $V_p$ and $n$ radially. It shows that the effect of biased disk electrode is seen throughout the bulk plasma.\par
The last set of experiments was conducted to see the effect of primary energetic electron flux on the plasma parameters when an additional positively biased metal disk is immersed in the plasma column (see Fig.\ref{fig:fig1}). The I--V characteristics of probe were recorded for two disk bias voltages (-3.5 and +30 V) at different filament currents (or thermionic electron currents). The primary energetic electrons flux increases with increasing the filament current at a fixed DC bias voltage (-65 V). Variation of plasma density with filament heating currents for two disk bias voltages is shown in Fig.\ref{fig:fig6}(a). The plasma density increases if flux of primary energetic electrons is increased by passing more current through the filaments. The plasma density increment rate depends on the bias voltage of the metal disk, as can be seen in Fig.\ref{fig:fig6}(a). Apart from density variation plots, we have also plotted $V_p$ and $T_e$ variation against the filament currents in Fig.\ref{fig:fig6}(b). It is clear from fig.\ref{fig:fig6}(b) that $T_e$ has different values for bias voltage -3.5 V and +30 V for a given filament current but a slight increment in $T_e$ was observed with increasing the population of primary energetic electrons or filaments heating. We observe slight lower plasma potential at $V_b$ = 30 V as filament heating current is increased (see Fig.\ref{fig:fig6}(b)). It shows that plasma potential is going to decrease with increase in the population of primary energetic electrons. We expect similar behavior of $V_p$ against filament heating current if the metal disk is positively biased above the plasma potential. This behavior suggests the role of primary energetic electrons (or electrons beam) on plasma characteristics in the presence of an auxiliary positively biased metal electrode (disk). A detailed discussion on the observed experimental results is provided in Sec.\ref{sec:results_discussion}.
\begin{figure*}
\centering
\subfloat{{\includegraphics[scale=0.3550]{Fig6a}}}%
\hspace*{-0.3in}
\subfloat{{\includegraphics[scale=0.3550]{Fig6b}}
\caption{\label{fig:fig6}(a) Plasma density variation with filament heating current at two disk bias voltages. (b) Electron temperature and plasma potential variation with filament heating current at two disk bias voltages. The helium pressure was set at $1\times 10^{-3}$ mbar. The diameter (D) of the metal disk was 3 cm. Error over the averaged value of the measure plasma parameters are $< \pm$ 5\%.}
\end{figure*}
\section{Discussion} \label{sec:results_discussion}
In a steady-state low-pressure filament discharge, the plasma production rate is balanced by the loss rate to the chamber wall. The plasma production rate depends on the energetic emitted electron flux from the negatively biased heated filaments\cite{particlebalancemodel1,particlebalancemodel2,particlebalancemodel3exp}. It was observed that the plasma characteristics get modified in the presence of a positively biased auxiliary metal electrode (disk) in the plasma column. For a negatively biased disk electrode in plasma, an ion sheath is formed in front of it which reduces the electron flux and increases the ion flux towards the disk. Since this field is localizing up to a few mm distance and after that its effects are not dominating inside the plasma, the plasma parameters are not changing with the negatively biased disk at a given discharge condition. If we give positive bias to the metal disk then an electron or ion sheath is formed around the positively biased metal disk in the hot-filament discharge \cite{globalambipolardiffusion,plasmapotentiallocking}. In such a case, positively biased metal electrode perturbs the entire plasma and influence the bulk plasma properties. As disk bias voltage is kept above the initial plasma potential, the confined electrons (plasma electrons) are lost to the surface of the disk. Due to the loss of plasma electrons, plasma potential becomes more positive to maintain the quasi-neutrality condition of plasma. However, the plasma potential variation of bulk plasma with the disk bias voltage depends on the area of the biased metal disk ($A_d$) in order to maintain current balance \cite{globalambipolardiffusion,plasmapotentiallocking}. The response of the bulk plasma to a biased metal electrode above the plasma potential is mainly determined by the ratio of area $A_d/A_w$ and the ratio of ion-electron mass $(M_i/m_e)$. The ratio of area $A_d/A_w$ determines whether an electron sheath ($A_d/A_w$ $< \mu$) or a double sheath ($A_d/A_w$ $\simeq$ 1.7 $\mu$) or ion sheath ($A_d/A_w$ $>$ 1.7 $\mu$) will be formed around the positively biased ($V_b > V_p$) auxiliary electrode (disk) in the helium plasma \cite{plasmapotentiallocking,globalambipolardiffusion,anodesizeplasmareponse}. Here $\mu$ = $\sqrt{2.3 m_e/M_i}$ and $V_b$ is the disk bias voltage. For helium gas, $\mu$ is $\sim$ 1.76 $\times 10^{-2}$. In the present experiment, the area of the plasma-facing chamber wall $A_w$ $\sim$ 6350 $cm^2$ and area of disk electrode, D = 3 cm and D = 8 cm are $A_d$ $\sim$ 7 $cm^2$ and $A_d$ $\sim$ 50 $cm^2$ respectively. So the ratio $A_d/A_w$ $< \mu$ for both biased disks which predicts the electron sheath near the positively biased disk \cite{globalambipolardiffusion}. Since the characteristics of bulk plasma depends on the type of sheath region around a positively biased electrode, plasma parameters ($n_e$, $T_e$ and $V_p$) should also vary according to that of predicted in the theoretical model as well as in experiments \cite{globalambipolardiffusion,anodesizeplasmareponse,plasmapotentiallocking}.\par
The plasma potential variation with the bias voltages to the metal disk, Fig.\ref{fig:fig3}(a) and Fig.\ref{fig:fig4}, does not follow the trend (sub-linear) that were predicted by theoretical model if electron sheath is formed ($A_d/A_w$ $< \mu$) in front of metal disk. The plasma potential increases with a higher rate (super-linear) when we give more positive bias voltage ($>$ +20 V) to the metal disk. Such super-linear variation of plasma potential is expected for a large sized biased disk electrode ($A_d/A_w$ $\geq$ 1.7 $\mu$).
The electron temperature of bulk plasma ($T_e$) also increases in the presence of a positively biased auxiliary metal disk. The variation of $T_e$ is somewhat similar to that predicted in theory and experiment \cite{globalambipolardiffusion,anodesizeplasmareponse}. There is a major difference in the plasma density variation with bias voltages above the plasma potential in the case of $A_d/A_w$ $< \mu$. In our experiments, we have observed the enhancement of plasma density (see Fig.\ref{fig:fig3}(c)) instead of electron density depletion \cite{anodesizeplasmareponse,globalambipolardiffusion} in the presence of a smaller sized positively biased auxiliary metal electrode (disk). The plasma density increases linearly above the disk bias voltage of 25 V. This opposite behavior of plasma density variation could be due to the presence of primary energetic electrons (electron beam) in the bulk plasma.\\
Plasma response to a fixed biased disk (+30 V) at different fluxes of primary energetic electrons (heating currents) may help in understanding the observed results. We see an effective change in plasma density (see Fig.\ref{fig:fig6}(a)) for the negatively biased (-3.5 V) and positively biased (+30 V) disk at a large filament heating current (24.7 A). It indirectly shows the role of the population of primary energetic electrons (or thermionic electrons) on the plasma properties in the presence of a positively biased disk. In Fig.\ref{fig:fig6}(b), the gap between plasma potential measured at $V_b$ = -3.5 V and +30 V is decreased by $\sim$ 2 V if filament current is increased by 1.2 A (see Fig.\ref{fig:fig6}(b)). It indicates that the beam electrons (primary energetic electrons) lower the plasma potential that was also observed in other hot-filament discharge\cite{negativeplasmapotentialcpp}. \par
The observed results can also be understood qualitatively based on the past theoretical and experimental work in the low-pressure hot-filament discharge \cite{plasmapotentiallocking,particlebalancemodel1,particlebalancemodel2,particlebalancemodel3exp,mangilalcpp}. Once the disk bias voltage is higher than the plasma potential ($V_b > V_p$), bulk plasma electrons loss rate to the metal disk increases, resulting in increasing the plasma potential. With increasing the plasma potential by $\sim$ 30 V (Fig.\ref{fig:fig3}(a)), the energy of thermionic emitted electrons is increased by $\sim$ 30 eV. These energetic primary electrons from heated filaments (90 to 100 eV) create the plasma by the impact ionization of helium gas. Since the chamber wall is grounded ($V_w$ = 0), the bulk plasma electrons are confined in a potential well which is determined by the plasma potential. As the disk bias voltage is above 25 V, plasma density increases at a higher rate. It is only possible with a higher ionization rate in the bulk plasma volume. According to the theoretical model\cite{particlebalancemodel1,particlebalancemodel2,particlebalancemodel3exp,secondaryelectronsinplasma}, secondary electrons (hot confined electrons) which are released from the chamber wall and produced from the primary energetic electrons are also confined in the potential well. The secondary electrons of kinetic energy (E $>$ 25 eV) above the ionization potential of helium gas ($\sim$ 24.5 eV) may also take part in the ionization process and increase the plasma density. If we increase the density of the primary energetic electrons, density growth rates are different for negatively (-3.5 V) and positively (+30 V) biased disk (see Fig.\ref{fig:fig6}(a)). It confirms that the density of primary energetic electrons of the same energy plays a dominant role in determining the plasma formation rate (or ionization rate) in presence of a positively biased auxiliary metal disk.\par
The positively biased disk ($V_b$ $>$ $V_p$) attracts the low-energy plasma electrons therefore average energy of the bulk plasma electrons is increased\cite{anodesizeplasmareponse,globalambipolardiffusion}. It is also expected to increase the average energy bulk plasma electrons by collisions of low energy electrons and secondary high energy electrons in the potential well \cite{robertsonincreasete,particlebalancemodel3exp,secondaryelectronsinplasma}.
Hence we observe increment in electron temperature ($T_e$) of the bulk plasma with increasing the disk bias voltage.
\section{Summary} \label{sec:summary}
In this work, we have seen the response of plasma with an auxiliary (additional) biased disk electrode immersed in the hot-filament low-pressure helium plasma. The plasma parameters ($V_p$, $T_e$ and $n$) were measured using the planar and emissive probe at various locations of the plasma column. For both sizes of the metal disk (D = 3 cm and D = 8 cm), plasma potential, plasma density, and electron temperature were observed to increase with increasing the bias voltage to the metal disk above the plasma potential at given discharge conditions. However, the rate of change of plasma parameters with disk bias voltage depends on the size of the disk electrode and equilibrium plasma density (or population of primary energetic electrons). Such response of helium plasma is understood with the help of available theoretical models. The ionization rate (plasma production rate) and plasma loss rate to available boundaries mainly determine the characteristics of bulk plasma. Discrepancy in the observed results (plasma parameters variation) from that of predicated or observed in hot-filament discharge could be possible due to the presence of energetic primary electrons in the bulk plasma. The experimental results suggest to incorporate the role of primary energetic electrons emitted from heated filaments in the available theoretical model. These experimental results also provides an insight into electrostatic confinement by using an positively biased additional electrode. In future, our focus will be to study the transient response of hot-filament plasma at different electrostatics confinement strengths which depends on the height of potential well. The height of potential well can be modified by changing the plasma potential with respect to grounded chamber.\\
\section{Acknowledgement}
The authors are very grateful to Prof. S. Mukherjee for guiding during the experiments at Institute for Plasma Research, Gandhinagar.
|
\section{Introduction}
Recommender systems have become increasingly important for online marketplaces to achieve their business success. Among common recommendation tasks, Click-Through-Rate (CTR) prediction has generated significant research interest because it is a practically important business performance measure influencing user behavior \cite{zhou2018deep,lian2018xdeepfm} that is also closely related to the revenue generating performance metrics used in online platforms. In particular, researchers have proposed to use sequential recommendation models \cite{sutskever2014sequence,wang2019sequential} to automatically extract latent user and item representations and their high-order interactions to achieve better CTR prediction performance, including GRU \cite{hidasi2015session}, LSTM \cite{zhu2017next}, and BERT \cite{sun2019bert4rec}. However, these methods also suffer from the cold start and data sparsity problems \cite{schein2002methods,adomavicius2005toward} commonly encountered in recommendation tasks, which can severely affect the performance metrics since we only have access to a small proportion of past consumption records.
To address these problems, researchers have proposed to use cross domain recommender systems \cite{cantador2015cross} to transfer and aggregate user preference across different domains \cite{pan2010survey}. For example, cross domain sequential recommendation models \cite{ma2019pi,ouyang2020minet} leverages auxiliary user information from a source domain to improve the CTR prediction performance of a target domain. However, existing cross domain sequential recommendation models primarily focus on the unidirectional transfer of user preferences from the source domain to the target domain without taking into account the dual nature of providing recommendations in source-target domain pairs. While achieving satisfying CTR prediction performance in the target domain, these models might not achieve the optimal CTR prediction performance in the original source domain, which is unfortunate as industrial platforms typically have many category domains, and it is important to improve the recommendation performance in all domains simultaneously.
Meanwhile, it is more desirable to also transfer user preferences in the other direction, i.e., from the target domain back to the source domain, since improving recommendation performance in one domain could also lead to the improvements of recommendation performance in the other domain \cite{li2020ddtcdr}, especially for CTR predictions as users may simultaneously seek purchases of items from multiple categories such as buying electronics for Christmas gifts and candles for Christmas decorations. If we can provide better CTR predictions for the electronics domain, we can use those predictions to improve our recommendations for the decorations domain, and vice versa.
Therefore in this paper, we propose to apply the dual learning mechanism to sequantial learning models to bidirectionally transfers user preferences across different domains simultaneously for cross-domain click-through rate prediction task. Our proposed dual learning-based model consists of the following two novel components:
\begin{itemize}
\item \textbf{Dual Embedding}. As user purchasing actions in a certain domain could be affected by their exposures to previous recommendations in other domains, we propose a novel \emph{dual embedding} component that unifies the learning process of user representations. In particular, we adopt the idea of metric learning \cite{kulis2012metric} and map the latent user and item representations in different domains into the shared latent space, where we aim to minimize the metric distances between different representations. By doing so, we could not only bring two item embeddings closer to each other if they are purchases by the same user, but also bring two user embeddings closer to each other through the similar items consumed across the two domains. Therefore, we progressively improve the modeling of user preferences for all domains through the iterative learning process in the Dual Embedding component.
\item \textbf{Dual Attention}. As the transaction records for sequential recommendations often contain many items irrelevant to the next choice, attention mechanism \cite{shaw2018self,zhou2018deep} is often utilized to weight the observed items with different relevance to build an attentive context that outputs the proper next item with a high probability. In this paper, we propose a novel \emph{dual attention} component to estimate relative importance of historical items in \emph{both} domains simultaneously. Unlike classical attention-based recommendation models that only takes into account user behavior in one single domain, the Dual Attention component incorporates user behaviors in multiple different domains to obtain the associated attention values.
\end{itemize}
\noindent
By combining these two novel components - dual embeddings and dual attention mechanism - we develop a new \emph{Dual Attentive Sequential Learning (DASL)} model that bidirectionally transfers user preferences between domain pairs and thus provides more dynamic and effective cross domain recommendations.
As a part of this work, we conduct extensive offline experiments on three real-world datasets and show that the proposed DASL model significantly and consistently outperforms all the state-of-the-art baseline models in terms of the CTR-based cross domain recommendation performance. We also conduct online A/B test at a major video streaming platform Alibaba-Youku and show that the DASL model outperforms the current production system by 7.07\% in the total number of video views.
In this paper, we make the following research contributions:
(1) We propose novel dual embedding and dual attention techniques and a method of incorporating them into the cross-domain recommendation model. We show in the paper that these two new methods are effective for providing cross domain sequential recommendations that progressively improve CTRs through the iterative training process.
(2) We develop the Dual Attentive Sequential Learning (DASL) model that bidirectionally transfers user preferences and user-item interactions between domain pairs by deploying the proposed dual embeddings and dual attention mechanisms.
(3) We conduct extensive offline and online experiments to demonstrate superiority of the proposed DASL model, which consistently and significantly outperforms all the baseline models, as well as the latest production system at a major video streaming platform Alibaba-Youku.
The rest of the paper is organized as follows. We discuss the related work in Section 2 and present our proposed DASL model for providing sequential cross domain recommendations in Section 3. Experimental design on three real-world datasets is described in Section 4. The results and discussions are presented in Section 5. We also present the online A/B experiments in Section 6. Finally, Section 7 summarizes our contributions and concludes the paper.
\section{Related Work}
In this section, we summarize the related work of our proposed model as three categories: sequential recommendations, cross domain recommendations and dual learning methods.
\subsection{Sequential Recommendations}
Our work is related to deep-learning based sequential recommender system \cite{wang2019sequential} which suggests items that may be of interest to a user by modelling the sequential dependencies over the user-item interactions and capturing the current and recent preferences of the user. Popular sequential recommendation models utilize sequential neural network structures, including GRU \cite{hidasi2015session}, LSTM \cite{zhu2017next}, and BERT \cite{sun2019bert4rec} for extracting latent user and item representations as well as their high-order interactions and achieving satisfying recommendation performance. Some sequential recommendation models also take into account personalized and session-based information during the recommendation process, including DIN \cite{zhou2018deep}, DSIN \cite{feng2019deep}, DIEN \cite{zhou2019deep}, DeepFM \cite{guo2017deepfm}, Wide \& Deep \cite{cheng2016wide}, PNN \cite{qu2016product} and PURS \cite{li2020purs}, as different people would have different preferences towards the recommended items under different circumstances. In this paper, we inherit the fundamental idea of sequential recommendations to construct our novel cross-domain CTR prediction model.
\subsection{Cross Domain Recommendations}
Cross domain recommender system \cite{fernandez2012cross} is a powerful method to handle the cold-start and data sparsity problem commonly encountered in recommendation tasks. By leveraging auxiliary information from other domains, cross domain recommendation models are extended from single-domain recommendation settings based on the assumption that different behavioral patterns jointly characterize the way users interact with items of a certain domain \cite{sahebi2017cross,singh2008relational,hu2013personalized,loni2014cross}. Researchers have also proposed to use transfer learning techniques \cite{pan2010survey} to transfer user preferences from the target domain to the source domain for providing enhanced recommendations, including CoNet \cite{hu2018conet}, CCCFNet \cite{lian2017cccfnet} and so on. As CTR prediction task become increasingly important for online commerce platforms, some recent studies aim for providing cross domain sequential recommendations \cite{ma2019pi,ouyang2020minet} and have achieved great success in addressing the cold-start and data sparsity problem. While most of the existing studies focus on unidirectional user preference transfer from the source domain to the target domain, in this paper we propose a novel dual learning-based cross domain sequential recommender system that greatly improves its business performance.
\subsection{Dual Learning}
As we previously discussed, transfer learning \cite{pan2010survey} constitute an important method for constructing cross domain recommendation models, which identifies and incorporates the common knowledge structure that defines the domain relatedness during the learning process. Several cross domain recommendation models \cite{zhuang2010collaborative,liu2013multi,long2013transfer} have successfully applied the idea of transfer learning to achieve satisfying recommendation performance. As the dual learning mechanism \cite{long2012dual,zhong2009cross,wang2011cross,he2016dual} is capable of exploiting the duality between two recommendation tasks and enhancing the transfer learning capability, researchers has purposed to utilize dual learning for providing cross domain recommendations \cite{li2020ddtcdr,zhu2019dtcdr,tkde2021} In particular, these models assume that if two users have similar preferences in a certain domain, their preferences should also be similar across other domains as well. While these dual learning-based cross domain recommendation models achieves significantly better performance over the classical models, they focus primarily on the static rating predict task, without taking into account user sequential behavior and the next-item recommendation task, which is the main focus of this paper.
\section{Model}
In this section, we present the proposed Dual Attentive Sequential Learning (DASL) model for providing cross domain sequential recommendations, which is shown in Figure \ref{model}. First, we construct latent representations of users and items in both domains using the autoencoding technique. Then we utilize the GRU network that takes sequences of user consumptions as inputs and extracts user preferences accordingly. During this process, we design a novel Dual Embedding component that maps user representations into a shared latent space through latent orthogonal mapping functions. By doing so, we bidirectionally transfer latent user preferences across different domains through minimization of metric distances in the shared space. The final recommendation is produced through the matching process between user preferences and the candidate items, where we utilize a novel Dual Attention component that takes into account relative importance of each consumption record in both domains. We will explain the details of our design in the following sections.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{picture/framework.jpg}
\caption{Overview of the proposed DASL framework for providing cross domain sequential recommendations.}
\label{model}
\end{figure*}
\subsection{User and Item Representations}
To effectively extract latent user preferences and efficiently model the interrelations of explicit features of users and items, we utilize the common practices of autoencoding techniques that automatically transforms the heterogeneous and discrete feature vectors into continuous feature embeddings. We denote the explicit features of user $a$ as $u_{a} = \{u_{a_{1}},eu_{a_{2}},\cdots,u_{a_{m}}\}$ and the explicit features of item $b$ as $i_{b} = \{i_{b_{1}},i_{b_{2}},\cdots,i_{b_{n}}\}$. The autoencoder framework trains two separate neural networks simultaneously: the encoder network, which maps explicit features into latent embeddings; and the decoder network, which reconstructs feature vectors from latent embeddings. Due to the effectiveness and efficiency of the Multi-Layer Percetron (MLP) network, we formulate both the encoder and the decoder as MLP, which learns the hidden representations by optimization the reconstruction loss $L$ for users and items as $L_{u} = ||u_{a}-MLP_{dec}(MLP_{enc}(u_{a}))||$ and $L_{i} = ||i_{b}-MLP_{dec}(MLP_{enc}(i_{b}))||$,where $MLP_{enc}$ and $MLP_{dec}$ represent the MLP networks for the encoder and the decoder respectively. Note that, during the learning process we construct different autoencoder networks for users and items in different domains to avoid information leakage between the domains.
\subsection{Dual Embedding}
While the user representations in different domains are obtained through the same autoencoding framework described in the previous section, the distribution of those latent embeddings would still be vastly different, as the user behaviors and preferences are highly heterogeneous across different domains. To utilize heterogeneous user behavioral information and address such phenomenon, we identify and extract overlap users that have interacted with items in both domains and use them as `pivots' to learn the relations of user preferences and behaviors in different domains.
In this section, we present the novel Dual Embedding component, which is constructed by identifying users with similar preferences in one domain and bidirectionally transfer this similarity information to the other domain. In particular, we adopt the idea of metric learning \cite{kulis2012metric} and map the latent embeddings in different domains into a shared latent space through the transitional mapping. By minimizing the metric distances in the shared latent space, this transitional mapping would pull the mapped embeddings of the same user from different domains closer, while pushing embeddings from different users further apart. According to the triangle inequality, this metric learning process will also pull those users who purchase similar items across different domains together. Eventually, the nearest neighbor points for any given user in the shared latent space will become its representation in the other domains respectively. Through the metric learning mapping that effectively captures the relations between overlap users, we propagate the learning process that not only models the information of overlap users, but also non-overlap users and user-item interactions, which is typically difficult to capture in cross-domain recommendation tasks.
Different from the classical metric learning settings, we restrict this transitional mapping to be an orthogonal mapping to obtain certain benefits for the model optimization process, following the discussion in \cite{li2020ddtcdr}:
\begin{itemize}
\item Orthogonality preserves similarity between embeddings of different users during the metric learning process since orthogonal transformation preserves inner product of latent vectors.
\item Orthogonal mapping $X$ has transpose mapping $X^T$ as its inverse \cite{greub2012linear}, which simplifies the learning procedure and enables efficient dual learning process.
\item Orthogonality restricts the parameter space of the transitional mapping and thus reduces the computational complexity of the learning process.
\end{itemize}
Specifically, we denote the user embeddings for domain $A$ and $B$ as $W_{u_{A}}$, $W_{u_{B}}$ that we obtained in the previous section. The optimization goal is to search for the mapping matrix $X$ that minimizes the sum of squared Euclidean distances between the mapped user embeddings $XW_{u_{A}}$ and the target user embeddings $W_{u_{B}}$ for the same overlap users:
\begin{equation}
L_{X} = argmin_{X} \sum_{W_{u_{A}},W_{u_{B}}\in u_{A},u_{B}} ||XW_{u_{A}}-W_{u_{B}}||^{2}
\end{equation}
This optimization is equivalent to its dual form:
\begin{equation}
L_{X^{T}} = argmin_{X} \sum_{W_{u_{A}},W_{u_{B}}\in u_{A},u_{B}} ||W_{u_{A}}-X^{T}W_{u_{B}}||^{2}
\end{equation}
We constrain the metric learning mapping $X$ to be an orthogonal mapping (i.e. $XX^{T}=X^{T}X=I$), which serves to enforce structural invariance of user preferences in each domain, while preventing a degradation in mono-domain recommendation performance for learning better transitional mappings.
As we simultaneously map user embeddings from different domains into the shared latent space, we effectively utilize the dual learning mechanism that updates user embeddings for both domains using equations (1) and (2) simultaneously. The learning process can then be repeated iteratively to obtain progressively better cross-domain recommendation performance until the convergence criterion is met. As a result, it pushes the metric learning framework to better capture user preferences and thus provide even better recommendation performance.
\subsection{Sequential Learning}
Next, we construct latent representations of user preferences from their past consumptions that account for the personalized, dynamic and contextual factors during the recommendation process. Since user behavior streams can be modeled as temporal sequences, we adopt the state-of-the-art deep sequential learning models to condense user behavior streams into latent representations of user preferences to facilitate the learning process. In particular, we select the recurrent neural network to model user interests, for it is capable of capturing the time information and the sequence of user transactions, as more recent interactions would naturally have a higher impact on the current recommendation than previous interactions. As the GRU \cite{cho2014learning} neural network is shown to be computationally more efficient and is capable of providing better performance \cite{chung2014empirical}, we select GRU as our choice of the RNN network.
During the learning process, we first map the behavior sequence to the corresponding item embeddings obtained in the previous stage. To illustrate the GRU learning procedure, we denote $W_{z}$,$W_{r}$,$U_{z}$ and $U_{r}$ as the weight matrices of current information and the past information for the update gate and the reset gate respectively. $x_{t}$ is the user state input at timestep $t$, while $h_{t}$ stands for the user state output. $z_{t}$ denotes the update gate status and $r_{t}$ represents the status of reset gate. Therefore, the hidden state at timestep $t$ could be obtained following these equations:
\begin{equation}
z_t = \sigma_g(W_{z} x_t + U_{z} h_{t-1} + b_z)
\end{equation}
\begin{equation}
r_t = \sigma_g(W_{r} x_t + U_{r} h_{t-1} + b_r)
\end{equation}
\begin{equation}
h_t = (1-z_t) \circ h_{t-1} + z_t \circ \sigma_h(W_{h} x_t + U_{h} (r_t \circ h_{t-1}) + b_h)
\end{equation}
By iteratively calculating hidden states throughout user behavior sequences, we would obtain the final hidden state at the end of user behavior sequences, which constitutes the representation of user preferences for the current recommendation. The latent representations of user preferences would be subsequently used for providing cross-domain recommendations.
\subsection{Dual Attention}
Note that each user past transaction might have different effect on the current recommendation depending on its time of occurrence and other contexts. As historical records of user transactions may contain items that are not directly relevant to the current recommendation choices, attention mechanism \citep{bahdanau2014neural} is often utilized to assign the transaction records with different weights to build an attentive context that outputs the proper next item with a high probability of a positive outcome. While previously proposed models have incorporated the self-attention mechanism \cite{shaw2018self,zhou2018deep} during the sequence learning process, they only consider the case of providing single-domain recommendations. To provide effective cross-domain recommendations, in this paper we propose a novel Dual Attention component, which aggregates user behaviors in \emph{both} domains that jointly determine the relevance of each consumption record towards the current item recommendation.
The Dual Attention component is constructed based on the framework of scaled dot-product attention \cite{vaswani2017attention}, which is much faster and more space-efficient in practice, compared to the additive attention framework \citep{bahdanau2014neural}. Unlike classical attention-based recommendation models that only takes into account user behavior in one single domain, we propose to incorporate user behaviors in multiple different domains simultaneously to obtain the associated attention values. By doing so, we could provide a more comprehensive and effective modeling of the matching process between user preferences and candidate items, as the relative importance of each historical interaction towards current recommendations would be affected by user interactions in the other domains as well. For example, if we know that a user has watched videos related to Harry Potter, we might be willing to recommend another Harry Potter video to that user, although we are not entirely certain about the user's preferences as the user has also watched other types of videos. However, if we also know that this user has finished reading the entire series of Harry Potter novels from her/his activities in the Book domain, we would assign a much higher weight to that user's Harry Potter video watching experiences in the Video domain, thus making it more likely and with higher confidence to recommend similar content to that user. Similarly, user information in the Video domain would also help us to estimate better the attention values in the Book domain.
To illustrate the learning process of the Dual Attention component, we denote the input of queries in both domains as $Q_{1}$, $Q_{2}$ consisting past consumption records of the user in both domains. We also denote keys as $K$ of dimension $d_k$, and values as $V$ of dimension $d_v$, which will be updated during the learning process. We compute the dot products of both queries with all keys, divide each by $\sqrt{d_k}$ to counteract the effect that the dot products grow large in magnitude for large values of $d_k$. We subsequently apply a softmax function to obtain the attention weights on these values for both domains simultaneously as:
\begin{equation}
\mathrm{Attention}(Q_{1}, K, V) = \mathrm{softmax}(\frac{(Q_{1};Q_{2})K^T}{\sqrt{d_k}})V
\end{equation}
\begin{equation}
\mathrm{Attention}(Q_{2}, K, V) = \mathrm{softmax}(\frac{(Q_{2};Q_{1})K^T}{\sqrt{d_k}})V
\end{equation}
The final recommendations of our proposed Dual Attentive Sequential Learning model are obtained by aggregating the user embeddings generated through the Dual Embedding component described in Section 3.2, the latent representations of user preferences described in Section 3.3, and the attention values representing the relative importance of each consumption record obtained in Section 3.4. We feed the aggregated vectors into the Multi-Layer Perceprton (MLP) network to obtain the predicted values of the utility function. We produce the final recommendations by selecting the top-n items with the highest utility values. To demonstrate the superiority of our proposed model, we implement it on several industrial recommendation applications, which will be described in the next section.
\section{Experiments}
\subsection{Dataset}
We conduct extensive offline evaluations on three industrial cross-domain recommendation datasets to evaluate the performance of our proposed model. The Imhonet dataset \cite{DBLP:conf/cla/BobrikovNI16} contains user ratings and feedback information over the period of 2011 to 2014 across multiple domains. We select the two largest domains in the Imhonet, namely Books and Movies, to conduct our experiments. The Amazon dataset \cite{ni2019justifying} consists of user purchase actions and rating information collected from the Amazon platform, and we select two domains with sufficient amount of overlap users for conducting our experiments: Toys and Video Games. Finally, we test our model on the Youku dataset, which contains user logs for the domains of TV Shows and Short Videos of a major video-streaming platform Alibaba-Youku in China during the month of December 2020.
We list the basic statistics and the number of overlap users of these three datasets in Tables \ref{imhonet}, \ref{amazon} and \ref{imhonet} respectively. During the recommendation process, we identify the overlap users across different domains using a common user ID. We binarize the rating information as labels of click and non-click for the click-through-rate prediction task. Note that while each user might purchase items from different domains, each item belongs only to a single domain in all the three datasets. For providing sequential recommendations, we construct the purchasing history for each user as the last 10 purchased items based on the recorded timestamp for all experimental settings.
\begin{table}
\centering
\begin{tabular}{|l l l|}
\hline
Domain & Book & Movie \\ \hline
\#Users & 804,285 & 959,502 \\ \hline
\#Items & 182,653 & 79,866 \\ \hline
\#Records & 223,007,805 & 51,269,130 \\ \hline
Sparsity & 0.0157\% & 0.0669\% \\ \hline
Overlap Users & \multicolumn{2}{c|}{318,225}\\ \hline
\end{tabular}
\caption{Descriptive Statistics of the Imhonet Dataset}
\label{imhonet}
\end{table}
\begin{table}
\centering
\begin{tabular}{|l l l|}
\hline
Domain & Toys & Video Games \\ \hline
\#Users & 1,342,911 & 826,767 \\ \hline
\#Items & 327,698 & 50,210 \\ \hline
\#Records & 2,252,771 & 1,324,753 \\ \hline
Sparsity & 0.0005\% & 0.0032\% \\ \hline
Overlap Users & \multicolumn{2}{c|}{131,684}\\ \hline
\end{tabular}
\caption{Descriptive Statistics of the Amazon Dataset}
\label{amazon}
\end{table}
\begin{table}
\centering
\begin{tabular}{|l l l|}
\hline
Domain & TV Shows & Short Videos \\ \hline
\#Users & 63,360 & 63,360 \\ \hline
\#Items & 17,966 & 941,507 \\ \hline
\#Records & 11,558,124 & 19,162,111 \\ \hline
Sparsity & 1.0150\% & 0.0321\% \\ \hline
Overlap Users & \multicolumn{2}{c|}{63,360}\\ \hline
\end{tabular}
\caption{Descriptive Statistics of the Youku Dataset}
\label{industrial}
\end{table}
\subsection{Baseline Models and Evaluation Metrics}
To demonstrate the superiority of our proposed Dual Attentive Sequential Learning (DASL) model, we compare it with the following three categories of state-of-the-art baseline recommendation models. We also select two popular accuracy metrics for the evaluation process, namely \textbf{AUC} and \textbf{HR@10} \cite{gunawardana2015evaluating}. For all the experiments, we conduct 5-fold cross validation and report the average recommendation performance.
\subsubsection{Cross-Domain Sequential Recommendations}
\begin{itemize}
\item \textbf{Pi-Net \cite{ma2019pi}} The Parallel Information-sharing Network (Pi-Net) simultaneously generate recommendations for two domains through shared user accounts. It consists of two core units: a shared account filter unit to learn user-specific representation and a cross-domain transfer unit to transfer user information.
\item \textbf{MiNet \cite{ouyang2020minet}} The Mixed Interest Network (MiNet) jointly models three types of user interest: 1) long-term interest across domains, 2) short-term interest from the source domain and 3) short-term interest in the target domain to provide cross-domain recommendations.
\end{itemize}
\subsubsection{Cross-Domain Recommendations}
\begin{itemize}
\item \textbf{DDTCDR \cite{li2020ddtcdr}} Deep Dual Transfer Cross Domain Recommendation (DDTCDR) efficiently transfers user preferences across domain pairs through dual learning mechanism.
\item \textbf{CoNet \cite{hu2018conet}} Collaborative Cross Networks (CoNet) enables knowledge transfer process across domains through cross connections between base networks.
\end{itemize}
\subsubsection{Sequential Recommendations}
\begin{itemize}
\item \textbf{DIN \cite{zhou2018deep}} Deep Interest Network (DIN) designs a local activation unit to adaptively learn the representation of user interests from historical behaviors with respect to a certain item.
\item \textbf{Wide \& Deep \cite{cheng2016wide}} Wide \& Deep utilizes the wide model to handle the manually designed cross product features, and the deep model to extract nonlinear relations among features.
\end{itemize}
\section{Results}
\begin{figure}[h]
\centering
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{picture/scores_par_imhonet.png}
\caption{Imhonet Dataset}
\end{subfigure}%
\vskip\baselineskip
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{picture/scores_par_amazon.png}
\caption{Amazon Dataset}
\end{subfigure}%
\vskip\baselineskip
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{picture/scores_par_youku.png}
\caption{Youku Dataset}
\end{subfigure}%
\caption{Comparison of cross domain recommendation performance in terms of AUC and HR@10 measures in three datasets.}
\label{comparison}
\end{figure}
\subsection{Cross-Domain Recommendation Performance}
The results of applying our proposed DASL model and the corresponding baseline models to the three industrial datasets under the experimental settings described in Section 4 are presented in Table \ref{result} and Figure \ref{comparison}. As Table \ref{result} and Figure \ref{comparison} illustrate, our DASL model significantly and consistently outperforms all other baselines in terms of the AUC and HR@10 metrics across all the three datasets. Most importantly, we improve the recommendation performance for both domains in every recommendation task, instead of only improving the performance of the target domain, as was typically done by the classical cross-domain recommendation models. In particular, we observe an average increase of 1.36\% for the AUC metric and 1.72\% for the HR@10 metric, compared to the second-best baseline approach. This clearly demonstrates the effectiveness of our dual learning design.
We also observe in our experimental settings that cross domain recommendation models (i.e., DASL, Pi-Net, MiNet, DDTCDR and CoNet) significantly outperform the single-domain recommendation models (i.e., DIN and Wide \& Deep). This is the case, as cross-domain recommendation models utilize user behavioral information from multiple domains, therefore obtaining more comprehensive understanding of user preferences for recommendation purposes, compared to single-domain recommendation models. In addition, we observe that cross domain sequential recommendation models (DASL, Pi-Net and MiNet) achieve significantly better recommendation results than static cross domain recommendation methods (DDTCDR and CoNet), which illustrates the importance of incorporating sequential user behavior during the recommendation process.
To summarize, all these results show that the dual learning mechanism is powerful for the cross domain sequential recommendation tasks and works well in practice, as the performance of our proposed DASL model attests to.
\begin{table*}
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{|c|cccc|cccc|cccc|} \hline
Algorithm & \multicolumn{4}{c|}{Imhonet} & \multicolumn{4}{c|}{Amazon} & \multicolumn{4}{c|}{Youku} \\ \cline{2-13}
& \multicolumn{2}{c|}{Books} & \multicolumn{2}{c|}{Movies} & \multicolumn{2}{c|}{Toys} & \multicolumn{2}{c|}{Video Games} & \multicolumn{2}{c|}{TV Shows} & \multicolumn{2}{c|}{Short Videos} \\ \cline{2-13}
& AUC & HR@10 & AUC & HR@10 & AUC & HR@10 & AUC & HR@10 & AUC & HR@10 & AUC & HR@10 \\ \hline
\textbf{DASL} & \textbf{0.8375*} & \textbf{0.8752*} & \textbf{0.8380*} & \textbf{0.8752*} & \textbf{0.8520*} & \textbf{0.7844*} & \textbf{0.8511*} & \textbf{0.7844*} & \textbf{0.8825*} & \textbf{0.8565*} & \textbf{0.8635*} & \textbf{0.8539*} \\
Increased \% & (+0.98\%) & (+1.25\%) & (+1.64\%) & (+1.25\%) & (+1.30\%) & (+1.83\%) & (+1.43\%) & (+1.96\%) & (+1.08\%) & (+1.63\%) & (+1.73\%) & (+2.37\%)\\ \hline
Pi-Net & 0.8278 & 0.8644 & 0.8237 & 0.8630 & 0.8399 & 0.7697 & 0.8382 & 0.7699 & 0.8723 & 0.8422 & 0.8488 & 0.8324 \\
MiNet & 0.8294 & 0.8630 & 0.8245 & 0.8644 & 0.8411 & 0.7703 & 0.8391 & 0.7693 & 0.8731 & 0.8428 & 0.8486 & 0.8341 \\ \hline
DDTCDR & 0.8235 & 0.8553 & 0.8203 & 0.8577 & 0.8403 & 0.7685 & 0.8366 & 0.7661 & 0.8705 & 0.8395 & 0.8459 & 0.8308 \\
CoNet & 0.8186 & 0.8562 & 0.8198 & 0.8590 & 0.8382 & 0.7663 & 0.8354 & 0.7659 & 0.8698 & 0.8401 & 0.8444 & 0.8310 \\ \hline
DIN & 0.8233 & 0.8579 & 0.8198 & 0.8571 & 0.8317 & 0.7629 & 0.8320 & 0.7622 & 0.8642 & 0.8366 & 0.8410 & 0.8295 \\
Wide \& Deep & 0.8198 & 0.8608 & 0.8172 & 0.8568 & 0.8300 & 0.7618 & 0.8308 & 0.7614 & 0.8633 & 0.8357 & 0.8395 & 0.8295 \\ \hline
\end{tabular}
}
\newline
\caption{Comparison of Cross-Domain Sequential Recommendation Performance}
\label{result}
\end{table*}
\subsection{Ablation Study}
\begin{table*}
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{|c|cccc|cccc|cccc|} \hline
Algorithm & \multicolumn{4}{c|}{Imhonet} & \multicolumn{4}{c|}{Amazon} & \multicolumn{4}{c|}{Youku} \\ \cline{2-13}
& \multicolumn{2}{c|}{Books} & \multicolumn{2}{c|}{Movies} & \multicolumn{2}{c|}{Toys} & \multicolumn{2}{c|}{Video Games} & \multicolumn{2}{c|}{TV Shows} & \multicolumn{2}{c|}{Short Videos} \\ \cline{2-13}
& AUC & HR@10 & AUC & HR@10 & AUC & HR@10 & AUC & HR@10 & AUC & HR@10 & AUC & HR@10 \\ \hline
\textbf{DASL} & \textbf{0.8375*} & \textbf{0.8752*} & \textbf{0.8380*} & \textbf{0.8752*} & \textbf{0.8520*} & \textbf{0.7844*} & \textbf{0.8511*} & \textbf{0.7844*} & \textbf{0.8825*} & \textbf{0.8565*} & \textbf{0.8635*} & \textbf{0.8539*} \\ \hline
DASL-DE & 0.8333 & 0.8704 & 0.8355 & 0.8714 & 0.8468 & 0.7803 & 0.8471 & 0.7799 & 0.8771 & 0.8523 & 0.8600 & 0.8498 \\
DASL-DA & 0.8341 & 0.8712 & 0.8358 & 0.8710 & 0.8482 & 0.7795 & 0.8474 & 0.7797 & 0.8777 & 0.8529 & 0.8600 & 0.8502 \\
Single Domain & 0.8234 & 0.8580 & 0.8197 & 0.8569 & 0.8319 & 0.7629 & 0.8320 & 0.7621 & 0.8642 & 0.8366 & 0.8412 & 0.8293 \\ \hline
\end{tabular}
}
\newline
\caption{The Ablation Study}
\label{ablation}
\end{table*}
As discussed in the previous section, our proposed DASL model achieves significant improvements over other baselines. These improvements indeed come from incorporating the two novel components into the design of recommendation model: \textbf{Dual Embedding}, which unifies the learning process of user representations in both domains, thus enabling progressive improvements of user preference modeling through an iterative learning process; and \textbf{Dual Attention}, which estimates the relative importance of historical items in both domains simultaneously for providing next-item recommendations.
In this section, we conduct the ablation study to justify importance of each factor. Specifically, we compare the proposed model with the following variations:
\begin{itemize}
\item \textbf{DASL-DE} This is a variant of our proposed model, where we drop the Dual Embedding (DE) component and generate latent representation of explicit user features separately for each domain.
\item \textbf{DASL-DA} This is a variant of our proposed model, where we drop the Dual Attention (DA) component and obtain the attention values of user records separately for each domain.
\item \textbf{Single-Domain} This is a variant of our proposed model, where we drop both the Dual Embedding and the Dual Attention components during the recommendation process. That is to say, we train the recommendation model separately for different domains without transferring user preferences.
\end{itemize}
As shown in Table \ref{ablation}, if we remove any of these two novel components out of the recommendation model, we observe significant loss in both the AUC and HR@10 metrics. Therefore, the ablation study demonstrates that the superiority of our proposed model really comes from the combination of these novel features that all play significant role in the dual learning-based model and all contribute to superior cross-domain recommendation performance.
\section{Online A/B Test}
We have also conducted an online A/B test at a major video streaming platform Alibaba-Youku providing cross-domain recommendations of TV Shows and Short Videos in January 2021. During the testing period, we compared the proposed DASL model with the latest production model deployed in the company. We measured model performance using the standard business metric \textbf{VV} (Video View) extensively used in the company that specifies the average amount of videos or TV shows each user watches in one session. It is closely related to the Click-Through-Rate (CTR) metric that we optimize in our recommendation model. We find that our DASL model increased the VV metric in both the TV Shows and the Short Videos domains by 7.07\%, compared to the latest production system in the company, and this improvement is significant and consistent through our experimental process. Therefore, we demonstrate the superior performance of our DASL model and the power of incorporating the proposed Dual Embedding and the Dual Attention components into the cross domain recommendation toolbox. Since the proposed DASL model has achieved such strong performance results, it is currently in the process of being deployed at the company.
\section{Conclusions}
Previously proposed cross domain sequential recommendation models primarily leverage information from the source domain to improve CTR predictions in the target domain, without taking into account the duality nature of providing cross domain recommendations simultaneously for source-target domain pairs. In this paper, we propose to apply the dual learning mechanism to sequantial learning models to bidirectionally transfers user preferences across different domains simultaneously for cross-domain click-through rate prediction task in an iterative manner until the learning process stabilizes. In particular, the proposed Dual Attentive Sequential Learning (DASL) model consists of two novel components: the Dual Embedding component that simultaneously extracts user preferences in both domains, and the Dual Attention component that jointly determines the relative importance of each consumption record in both domains to produce final recommendations.
To demonstrate the superiority of our proposed model, we conduct extensive offline experiments on three real-world datasets and demonstrate that our proposed DASL model significantly and consistently outperformed several state-of-the-art baselines across all the experimental settings. We also conduct an online A/B test at a major video streaming platform, where our DASL model also significantly improves business performance results vis-a-vis the latest production system. Since the proposed model has achieved such strong performance results, it is currently in the process of being deployed at the company.
As the future work, we plan to extend the proposed components of Dual Embedding and Dual Attention to provide recommendations across multiple domains rather than only in domain pairs. We also plan to study theoretical properties, especially the convergence behavior, of our proposed framework.
\bibliographystyle{ACM-Reference-Format}
\balance
|
\section{Introduction}
It is widely known that the solar atmosphere is dominated by magnetic fields and almost all the activities taking place in the atmosphere are associated with the evolution of magnetic fields (e.g.,\citealp{tho11,wie14,van15,cai19}). Therefore, the question of the origin of magnetic fields is the most significant aspect in understanding the nature of solar activities. The observational sunspots and active regions (ARs) are thought to be born of the emerging magnetic flux, which is the concentration of the magnetic fields \citep{van15}. On the other hand, many solar activities are triggered by newly emerging magnetic flux, such as coronal mass ejections (CMEs), flares, formations or eruptions of the filaments, jets, and small-scale bursts (e.g., \citealp{fey95,che00,iso17,yan17,che18,wan18,wan19,you18,tia18}). Such an emerging magnetic field not only plays a vital role in understanding active regions and sunspots, but it also has a great influence on many of the activities in taking place in the solar atmosphere.
Flux-emerging regions (FERs), or general ARs, are thought to be formed by the rise of flux tubes from the convection zone to the solar surface \citep{van15}. \cite{fu16} found that about half of the investigated emerging active regions show a two-step emergence pattern that indicates that the flux emergence rate was relatively low at the early phase, undergoing a significant enhancement in the next phase. According to the numerical simulations, a twisted flux tube can emerge from the convection zone into the corona, passing through the photosphere and chromosphere as a result of the magnetic buoyancy or Parker instability, which leads to the formation of an $\Omega$-shaped flux tube \citep{lek96,fan01,mag03}. However, some studies have suggested that not all the twisted flux tubes embedded in the subsurface could emerge into the upper atmosphere. \cite{mur06} simulated different emerging flux tubes with different twist values. They found that the highly twisted flux tube could easily emerge into the upper atmosphere, while the weakly twisted flux tube fails to do so. \cite{tor11} also obtained a similar result that only the flux tube with high twist (q$H_0$ $>$ 0.1, $q=B_\theta/rB_l$, $H_0$=170 km) can be launched into the corona and the low twisted flux tube could hardly emerge into the photosphere due to the influence of the horizontal expansion. However, it is rare to observe a high twisted flux tube emerging from the subsurface. Several researchers have provided some observational evidence to ascertain the emergence of a small twisted flux rope \citep{oka09,yan17}. \cite{poi15} analyzed the ``magnetic tongue '' of 41 emerging bipolar active regions with a new systematic and user-independent method, deducing that flux ropes embedded in the subsurface before the emergence have a low amount of twist and that highly twisted flux tubes are relatively rare. The discrepancy between the numerical simulations and the observations raises a significant question regarding how much twist the flux tube needs for its emergence.
Magnetic fields emerging from the Sun's interior would carry intrinsic information about the magnetic field in the subsurface to the upper atmosphere. Emerging magnetic fields play a key role in the transportation of the magnetic energy and helicity. When a flux tube emerges into the solar atmosphere from the convective zone (CZ), it not only brings magnetic flux that contributes to the formations of active regions or sunspots but that also injects magnetic energy and magnetic helicity into the solar atmosphere \citep{dem09}. Using a three-dimensional MHD numerical simulation, \cite{mag03} investigated the injection process of magnetic energy and relative magnetic helicity into the upper atmosphere. They found that both the magnetic energy and relative magnetic helicity are contributed by the emergence term during the early phase and the shear term during the latter phase. \cite{liu12} studied the injections of magnetic energy and helicity in two emerging active regions, and they found that rates of energy injection from both the emergence term and the shear term have a consistent evolution in phase during the entire process of flux emergence. They also derived that the injection of magnetic helicity and magnetic energy in the corona are mainly contributed by the shear term and the emergence term, respectively. A similar result was also obtained by \cite{vem15}. However, with the analysis of 28 emerging active regions, \cite{liu14} certified again that the shear term is dominant in the helicity injection, however, both terms, namely, the shear and emergence terms contribute approximately equivalent energy to the solar atmosphere. On the other hand, \cite{vem17} even found that opposite magnetic helicity was injected into the corona during the emergence of the active region and suggested that it may be caused by the emergence of a flux rope with different sign of twist between its apex and its legs. Investigations of the magnetic helicity or energy injection during the emergence of magnetic flux tube provide a good window for us to improve the understanding of the non-potential origination of emerging magnetic flux and a potential way to explore the nature of the magnetic flux embedded in the subsurface.
In general, the leading sunspots of a bipolar AR in the northern hemisphere have the opposite polarity with regard to leading spots in the southern hemisphere and the leading magnetic polarities alternate in successive sunspot cycle, known as Hale's law \citep{hal25}. This feature can be interpreted well in the models of ``dynamo theory'' (e.g., \citealp{par55,bab61,lei64,wan91,cho95}). Shearing motion by differential rotation, called the $\Omega$-effect, plays a key role in these models \citep{kra80,dik01}. According to the Hale's law, the leading spots of bipolar ARs should be negative and the following spots should be positive in the northern hemisphere in the solar cycle 24. The ARs with the reverse polarities in leading spots and following spots are considered as anti-Hale ARs. People found that a few ARs could violate Hale's law and the percentage of anti-Hale ranges from a few to 10\% \citep{wan89,li12,mcc14,li18,zhu20}. \cite{li18} found that the distribution of tilt angles and magnetic fluxes between Hale and anti-Hale is different while the pole separation of anti-Hale ARs is smaller than Hale ones. The anti-Hale and Hale ARs exhibit similar latitudinal distributions \citep{mcc14,li18}. On the other hand, anti-Hale ARs may hide the imprint of fundamental processes in solar interior. With the clue of the increase of the percentage of anti-Hale regions during the solar minimum, \cite{sok15} concluded that small-scale dynamo is active in the solar interior. Based on a surface flux transport simulation and observational data, \cite{jia15} proposed that bigger anti-Hale ARs emerging around low latitudes are responsible for the weak polar fields that affect the activity of the subsequent cycle. Thus, studies of anti-Hale ARs may shed light on our understanding of solar magnetic field evolution and supply a clue to probe the performing mechanism of magnetic field.
In this paper, we focus on the evolution and the magnetic characteristics in a strictly naked emerging anti-Hale active region NOAA 12720. The evolution of the active region and variations of different related-magnetic parameters during its emergence are studied in detail. The sections of this paper are organized as follows. The observations and methods are described in Section \ref{sec:obser methods}, the results are given in Section \ref{sec:results}, and the summary and discussions are presented in Section 4.
\section{Observations and methods} \label{sec:obser methods}
\subsection{Observations}
In this study, data from Helioseismic and Magnetic Imager (HMI: \citealp{sch12}) on board the Solar Dynamic Observatory\footnote{\url{https://sdo.gsfc.nasa.gov}} (SDO: \citealp{pes12}) are used to analyze the magnetic properties of an emerging Active Region NOAA 12720 during the period from 16:12 UT on August 24 to 12:00 UT on August 25, 2018. The SDO/HMI provides continuous full-disk measurement of the Stokes vector of the photospheric Fe $\rm{I}$ 6173 $\rm{\AA}$ line about every 135 s. The HMI science team had pipelined the process of retrieving vector field information from filtergrams and derived the available data (hmi.sharp.cea.720s) for solar researchers \citep{hoe14}. The pipeline procedure involves inversion of stokes vectors using the Very Fast Inversion of the Stokes Vector algorithm (VFSIV: \citealp{bor11,cen14}) based on the Milne-Eddington atmospheric model and removing the 180$\degr$ azimuthal ambiguity using the minimum energy method \citep{met94,lek09}. Space Weather HMI Active Region Patch (SHARP) products provide the data of AR's line-of-sight and vector magnetic field, continuum intensity, Doppler velocity, with a pixel scale of about 0.5 $\arcsec$ and a cadence of 12 minutes \citep{hoe14}. These data are remapped using Lambert (cylindrical equal area method \citep{cal02}) projection centered on the midpoint of the active region, which is tracked at the Carrington rotation rate. The continuum intensity images from SHARP permit us to investigate the evolution of the active region, whereas the vector magnetograms (VMs) from SHARP allow us to analyze the magnetic properties of the active region. Moreover, the high-resolution TiO and H$\alpha$ images from the New Vacuum Solar Telescope\footnote{\url{http://fso.ynao.ac.cn}} (NVST:\citealp{liu14b,yan20}) are also utilized to exhibit the photospheric {and chromospheric} details. The TiO images have a pixel size of 0.\arcsec052 and a cadence of 30 s. The field of view (FOV) of TiO images is 120\arcsec $\times$ 100\arcsec. The field of view (FOV) of H$\alpha$ images is 150\arcsec $\times$ 150\arcsec, with a 45s cadence and a spatial resolution of 0\arcsec165 per pixel. These data are calibrated from Level 0 to Level 1 with dark current subtracted and flat field corrected, and then speckle masking method was used to reconstruct the calibrated images from Level 1 to Level 1+ \citep{xia16}.
\subsection{Methods}\label{methods}
\subsubsection{Pole separation and tilt angle}
In order to determine the pole separation of the active region, we first need to determine the locations of each polarity. For each moment, the centroids of positive and negative polarities is determined by using the flux-weighted method as follows:
\begin{equation}
(\tilde{x}_{\pm},\tilde{y}_{\pm})=(\frac{\Sigma xB_{\pm}}{\Sigma B_{\pm}}, \frac{\Sigma yB_{\pm}}{\Sigma B_{\pm}}),\label{eq1}
\end{equation}
in which $B_{\pm}$ is the magnetic field strength in each pixel and ($x$, $y$) is the position coordinate of each pixel. The subscript signs ``+'' and ``-'' denote positive and negative polarities, respectively. As the centroids $(\tilde{x}_{\pm},\tilde{y}_{\pm})$ of the positive and negative polarities have been derived, the pole separation is calculated by using the following equation:
\begin{equation}
d=\sqrt{(\tilde{x}_+-\tilde{x}_-)^2+(\tilde{y}_+-\tilde{y}_-)^2}.\label{eq2}
\end{equation}
On the other hand, we define the tilt angle as the angle between the pole separation vector and the solar east. Thus, the tilt angle can be calculated by using the following equation:
\begin{equation}
\vartheta = -\arctan (\frac{\tilde{y}_+-\tilde{y}_-}{\tilde{x}_+-\tilde{x}_-}).\label{eq3}
\end{equation}
\subsubsection{Magnetic flux, vertical electric current, and force-free parameter $\alpha$}
Magnetic fluxes are integrated the positive and negative polarity magnetic fields by their area, respectively. Positive and negative magnetic fluxes are calculated by using the following equations:
\begin{eqnarray}
\varphi_{zp} = \int_{S_{ph}} B_{z}(x,y)_+dS, \nonumber\\
\varphi_{zn} = \int_{S_{ph}} |B_{z}(x,y)_-|dS, \label{eq4}
\end{eqnarray}
where $B_{z}(x,y)_+$/$B_{z}(x,y)_-$ is the vertical positive/negative magnetic field and the S denotes the integrated area with positive or negative magnetic field.
According to Ampere's law, the electric current density in the corona can be derived by the following equation:
\begin{equation}
\textbf{\emph J}=\frac{c}{4\pi}(\triangledown \times \textbf{\emph B}),\label{eq5}
\end{equation}
in which $c$ is the speed of the light and $\emph{\textbf{B}}$ denotes the magnetic field vector in the corona. According to Eq. \ref{eq5}, by neglecting the effect of the electric displacement current, the electric current density component perpendicular to the solar surface ($J_z$) can be calculated by the following equation:
\begin{equation}
J_z(x,y)=\frac{c}{4\pi}(\triangledown \times \emph{\textbf{B}})_z=\frac{c}{4\pi}(\frac{\partial B_y(x,y)}{\partial x}-\frac{\partial B_x(x,y)}{\partial y}),\label{eq6}
\end{equation}
where $B_x$ and $B_y$ are the two perpendicular components of the transverse magnetic fields. With the vector magnetic fields from SDO/HMI and Eq. \ref{eq6}, the distribution of vertical electric current density in the photosphere can be approximated by using a five-point stencil method. Therefore, the vertical current can be integrated by the vertical current density in the active region. The positive and negative currents calculated by using the following equations:
\begin{eqnarray}
I_{zp}&=&\int _{S_{ph}} J_{z}(x,y)_+ dS, \nonumber \\
I_{zn}&=&\int_{S_{ph}} |J_{z}(x,y)_-|dS,\label{eq7}
\end{eqnarray}
where $J_{z}(x,y)_+$/$J_{z}(x,y)_-$ is the vertical positive or negative current density and S denotes the integrated area. On the other hand, we also calculate the current neutralization ratio ($\Re$)\citep{tor14,liu17}:
\begin{equation}
\Re = \frac{I_{ret}}{I_{dir}}, \label{eq8}
\end{equation}
where $I_{ret}$ and $I_{dir}$ are the return and direct currents, respectively. They are express as
\begin{eqnarray}
I_{ret}=\int _{S_{ph}} \kappa_{ret}(x,y) J_{z}(x,y)dS, \nonumber\\
I_{dir}=\int_{S_{ph}}\kappa_{dir}(x,y)J_{z}(x,y)dS,
\end{eqnarray}
with
\begin{eqnarray}
\kappa_{ret}=\left\{
\begin{array}{cl}
0& $ $\text{if}$ $J_z(x,y)B_z(x,y)<0 \\
1& $ $\text{otherwise,}
\end{array}
\right.\nonumber
\end{eqnarray}
\begin{eqnarray}
\kappa_{dir}=\left\{
\begin{array}{cl}
0& $ $\text {if}$ $J_z(x,y)B_z(x,y)>0 \\
1& $ $\text{otherwise.}
\end{array}
\right.\nonumber
\end{eqnarray}
The current neutralization ratio ($\Re$) is an important parameter in evaluating the equilibrium of the AR system, which is related to flares and CMEs \citep{mel91,liu17,ava20}.
Under the assumption of force-free field, electric currents are parallel or anti-parallel to the magnetic field lines. The electric current could be expressed as:
\begin{equation}
J=\alpha B,\label{10}
\end{equation}
in which $\alpha$ is called force-free parameter, and it is constant along each field line. The parameter $\alpha$ could usually be used to represent a measure of magnetic twist in an AR \citep{see90,pev95,hag04}. A calculable quantity ($\alpha_z=\frac{J_z(x,y)}{B_z(x,y)}$) is a commonly used proxy of the twist at the photospheric level \citep{pev94}. Here, we calculate an averaged flux-weighted force-free parameter $\tilde{\alpha}$ over the whole AR in the photosphere, based on the following equation \citep{hag04,kut19}:
\begin{equation}
\tilde{\alpha}=\frac{\int j_z(x,y)B_z(x,y)dS}{\int B_z(x,y)^2dS},\label{eq11}
\end{equation}
in which $B_z$ is the vertical magnetic field. S denotes the area of the whole active region.
\subsubsection{Magnetic helicity injection flux}
Relative magnetic helicity (for simplicity, we use magnetic helicity to refer to relative magnetic helicity in this paper) can be transported from the solar interior to the corona by the new emerging fluxes and the various motions of magnetic flux in the photosphere. The magnetic helicity injection flux across a surface S is expressed as \citep{ber84,dem03}:
\begin{equation}
\dot H = 2\int_S(\textbf{\emph A}_p\cdot \textbf{\emph B}_t)V_{\bot n}dS - 2\int_S(\textbf{\emph A}_p\cdot \textbf{\emph V}_{\bot t})B_n dS,\label{eq12}
\end{equation}
where $\textbf{\emph A}_p$ is the vector potential of the potential field $\text{\emph B}_p$, $\textbf{\emph B}_t$, and $B_n$ are the tangential and normal components of the magnetic field, respectively. $\textbf{\emph V}_{\bot t}$ and $\textbf{\emph V}_{\bot n}$ are the tangential and normal components of the velocity perpendicular to the magnetic field lines, respectively. The velocity ($\textbf{\emph V}_\bot$) is further corrected by removing the irrelevant field aligned plasma flow as $\textbf{\emph V}_\bot=\textbf{\emph V}-\frac{\textbf{\emph V}\cdot \textbf{\emph B}}{\textbf{\emph B}^2}\textbf{\emph B}$. The velocity vector field ($\textbf{\emph V}$) is derived by using the Differential Affine Velocity Estimator for Vector Magnetograms (DAVE4VM) method \citep{sch08} in our study. The DAVE4VM method implements a variational principle with the magnetic induction equation to track the velocity of magnetic footpoints. We adopt 19 $\times$ 19 pixels as the window size for DAVE4VM, which is determined by examining the slope, Pearson linear correlation coefficient, and Spearman rank order between $\bigtriangledown_n \cdot (\textbf{V}B_n) $ and $\Delta B_n/ \Delta t$ \citep{sch06,sch08}. The first term in the right of the equation is named emergence term ($\dot H_e$), which is associated with the emergence of magnetic flux tube from the solar subsurface. The second term is named shear term ($\dot H_s$), which is generated by shearing magnetic field lines due to tangential motions on the surface \citep{ber84,kus02,par05,liu14}. As the target active region is small enough, the curvature effect could be ignored. The solar photosphere S can be assumed to be a planar. Therefore, the helicity ejection flux can be expressed as follows \citep{par05,liu12}:
\begin{eqnarray}
\dot H &=& \dot H_e +\dot H_s \nonumber \\
&=&\frac{1}{2\pi}\int_s \int_{s\arcmin}\textbf{\emph n}\cdot \frac{\textbf{\emph{x}}-\textbf{\emph{x}}\arcmin}{|\textbf{\emph{x}}-\textbf{\emph{x}}{\arcmin}|^2} \times \{\textbf{\emph B}_t(\textbf{\emph{x}})V_{\bot n}(\textbf{\emph{x}})B_n(\textbf{\emph{x}}\arcmin) \nonumber \\
& & -\textbf{\emph B}_t (\textbf{\emph{x}} \arcmin)V_{\bot n}(\textbf{\emph{x}} \arcmin)B_n(\textbf{\emph{x}})\} dS\arcmin dS \nonumber \\
& &-\frac{1}{2\pi}\int _s\int _{s\arcmin}\textbf{\emph n}\cdot \frac{\textbf{\emph{x}}-\textbf{\emph{x}}\arcmin}{|\textbf{\emph{x}}-\textbf{\emph{x}}{\arcmin}|^2} \times \{[\textbf{\emph V}_{\bot t}(\textbf{\emph{x}}) \nonumber \\
& &-\textbf{\emph V}_{\bot t}(\textbf{\emph{x}}\arcmin)]B_n(\textbf{\emph{x}})B_n(\textbf{\emph{x}}\arcmin)\}dS\arcmin dS \nonumber \\
&=& \int_s G_\theta(\textbf{\emph{x}}) dS,\label{eq13}
\end{eqnarray}
in which $\textbf{\emph{x}}$ and $\textbf{\emph{x}}\arcmin$ represent two photospheric positions and $\textbf{\emph n}$ is the surface normal vector pointing into the corona. $G_\theta (\textbf{\emph{x}})$ could be used to present the approximate proxy of the magnetic helicity injection flux density \citep{par05}. We use Eq.\ref{eq13} to calculate the total helicity flux in the whole active region. Accumulated helicity can be derived by the integration of the helicity injection rate with time.
To minimize the influence of measurement error in the magnetic field of HMI, we only consider pixels with a magnetic field strength exceeding 300 G (roughly 3$\sigma$ noise level of the vector magnetic field) to calculated above magnetic parameters \citep{hoe14}. The errors of positive and negative fluxes are estimated by the uncertainties from HMI observations of SHARP data. On the other hand, a Monte-Carlo experiment was used to estimate the errors in Eqs.\ref{eq1}, \ref{eq2}, \ref{eq3}, \ref{eq7}, \ref{eq8}, \ref{eq11}, and \ref{eq13}. In the Monte Carlo experiment, we randomly added noises to three components of the vector magnetic field, and repeated the computations of above equations. The added noises have a Gaussian distribution, and the widths ($\sigma$) of Gaussian distributions are set to be the uncertainties from HMI observations in each components of the vector magnetic field. This test was repeated 200 times. The two-hour running average of the root mean square (rms; $\sigma$) of these 200 experiments is used to be the representative error as the method estimated by \cite{liu12}.
\section{Results} \label{sec:results}
\subsection{Emerging process of the active region}
During the period from August 24 to 25 2018, an active region NOAA 12720 gradually emerged into the solar surface in the northern hemisphere, and appeared near the center of the solar disk (N07, W16). Figure \ref{fig1} shows the overview of this emerging active region at around 09:24 UT on August 24, 2018. Panel (a) displays the full disk line-of-sight magnetic field observed by SDO/HMI. This active region occurred in a purely clean environment, which is a perfect isolated AR. Panel (b) shows the TiO image observed by NVST. A lot of elongated granules and stretched dark lanes marked by black arrows in panels (b) could be identified between two main pores or sunspots, which are considered to be the signatures of the emerging flux through the photosphere \citep{str96,sch10}. The panels of the right column exhibit the images from SHARP data at the corresponding moment. Panel (c) is the continuum intensity image. The elongated granules and stretched dark lanes marked by the black arrows also could be found in the flux-emerging region. Panel (d) shows the vector magnetogram at 09:24 UT on August 24. The horizontal magnetic field ($\textbf{B}_t$) is depicted by blue and red arrows indicating the direction and strength, while the background is the map of the vertical magnetic field component ($B_z$). The relatively concentrated leading positive polarity and the dispersed following negative polarity indicate the fact that this active region does not obey with the hemisphere pattern of Hale's law \citep{hal25} for the active region in northern hemisphere in the solar cycle 24. Thus, the emerging active region is an anti-Hale one. Panel (e) shows the Doppler velocity map in the photosphere. Blue and red colors indicate upflows and downflows, respectively. The fascinating feature is that the regions with strong magnetic flux show strong red shift in the photosphere. This may be associated with the falling plasma. As the magnetic fields ascend into higher altitude, the plasma residing in the emerging magnetic field would fall down through the footpoints along the magnetic field lines due to the solar gravity, and returns to the solar surface. These falling plasma shows the red shifted signatures at the footpoints of the magnetic field lines.
\begin{figure*}
\centering
\includegraphics{fig1.eps}
\caption{Overview of the emerging active region. (a) The global line-of-sight magnetic field observed by SDO/HMI. The yellow dashed line marks the solar equator. The yellow and black boxes outline the fields of view of panel (b) and panels (c)-(e), respectively; (b) TiO image observed by NVST; (c) continuum intensity image from SHARP data; (d) vector magnetograms from SHARP data; blue and red arrows indicate the transverse field with positive and negative flux, respectively; (e) Doppler velocity from SHARP data.}\label{fig1}
\end{figure*}
Figure \ref{fig2} exhibits the emerging process of the active region NOAA 12720. The panels of left column show vector magnetograms at different moments, while the ones in the mid and right columns are corresponding continuum intensity and horizontal velocity, respectively. The active region initially appeared as a simple small bipolar pore (see panels (a1) \& (b1)), which eventually developed into a large active region including a leading positive sunspot, a following negative sunspot and some mixed polarities between two main sunspots (see panels (a4) \& (b4)). The detailed evolution of the active region is displayed by the animation of the Fig. \ref{fig2}. Most of the newly emerging flux initially appeared between two main polarities and gradually accumulated toward each polarity. Some moving dipolar features \citep{ber02} could be found between two main polarities, which is marked by the yellow circle in the panel (a3). These types of structures are important to form the long coronal loop connecting two main polarities and release the mass from rising magnetic field at the later stage of emergence \citep{cen12}. Panels (b1)-(b4) show the evolution of two main sunspots during the emergence of the active region in continuum intensity images. The leading sunspot tends to be more concentrated than the following one. On the other hand, the leading sunspot drifted westward while the following sunspot drifted eastward. The leading sunspot moved faster than the following one, which is demonstrated by the inclinations of two dashed blue lines in the mid column. This feature is reasonable for the leading (or following) sunspot during emergence. In general, the leading sunspot is close to the solar equator and drift westward while the following sunspot is far away from the solar equator and drift opposite direction. With the influence of solar rotation and differential rotation in latitude, the leading sunspot is expected to drift faster than the following one during their emergence. Based on the velocity map derived by DAVE4VM in the panels of the right column, the positive polarity exhibited a strong motion toward the southwest at the early phase of emergence, while the negative polarity moved toward the northeast (panel (c1)-(c3)). At the later phase of emergence, these motions became weak and inconsistent (panel (c4)).
\begin{figure*}
\centering
\includegraphics{fig2.eps}
\caption{Evolution of the emerging active region NOAA 12740. (a1)-(a4) The vector magnetograms. (b1)-(b4) The continuum intensity images. The white and black contours indicate the magnetic field strength (800G) with positive and negative polarities, respectively. The white asterisks in panel (b2) denote the centroids of positive and negative polarities; (c1)-(c2) The transverse velocity maps derived with DAVE4VM. The blue and green arrows indicate the transverse velocity with positive and negative polarities, respectively.}\label{fig2}
\end{figure*}
Figure \ref{fig31} shows the changes in photosphere and chromosphere between the early and latter stages of emergence. The images in top row are TiO observations from NVST, while the ones in bottom row are corresponding H$\alpha$ observations from NVST. At early stage of emergence, a lot of elongated granules and stretched dark lanes can be found in the middle of active region in the photosphere, which are marked by black circle in panel (a) of Fig. \ref{fig31}. Meanwhile, the sheared-arcade structures appeared in the chromosphere (see in panel (c)). At the latter stage of emergence, the active region was approaching mature, which consists of leading, following sunspots and some fragments. It is interesting to find that the leading sunspot has a complete penumbra while the following sunspot does not (see panel (b)). Meanwhile, more complicated structures (such as filaments or small-scale filamentary structures) could be seen in the chromosphere (see panel (d)).
\begin{figure*}
\centering
\includegraphics{fig31.eps}
\caption{Photospheric and chromospheric observations during the emergence. (a) \& (b) TiO images observed by NVST; (c) \& (d) H$\alpha$ images observed by NVST. Black circle in panel (a) points out some elongated granules and stretched dark lanes.}\label{fig31}
\end{figure*}
\subsection{Magnetic flux, pole separation and tilt angle}
Based on the equations described in Sect. \ref{methods}, we calculate the magnetic flux, pole separation, and tilt angle of the emerging active region during its emergence. Figure \ref{fig3} (a) exhibits the temporal variations of positive and negative magnetic fluxes. The error bars are derived by integrating the uncertainties from HMI observation with integrated area. Both positive and negative magnetic fluxes show a equivalent increase from about 0.2 $\times 10^{21}$ Mx at 16:12 UT on August 23 to 3.6 $\times 10^{21}$ Mx at 12:00 UT on August 25. The mean increasing rate of magnetic flux is about 0.78 $\times 10^{20}$ Mx/hr in each polarity. We can find that both positive and negative magnetic fluxes maintained a slow increase during the period before 05:00 UT on August 24. The mean increasing rate of each polarity was only about 0.17 $\times 10^{20}$ Mx/hr during this period. We define this period as the early phase of emergence. Both positive and negative magnetic fluxes experienced a rapid increase during the period from 05:00 UT on August 24 to 9:00 UT on August 25. The mean increasing rate of each polarity was 1.16 $\times 10^{20}$ Mx/hr. We define this period as the main phase of emergence. Therefore, the emergence of this naked anti-Hale AR is also consistent with the two-step emergence pattern \citep{fu16}. Unfortunately, the data during the periods from 06:24 to 08:48 UT on August 24 and 25 were unavailable because of the eclipse in the SDO orbit \citep{pes12}. Both positive and negative fluxes exhibit little change during the period from 09:00 UT to 12:00 UT on August 25. This indicates that the active region became mature at around 12:00 on August 25. Panel (b) shows the time variations of tilt angle and pole separation. The blue line denotes the tilt angle of the active region, while the black line denotes the pole separation. The uncertainties are depicted by the black and blue representative error bars. They are estimated by conducting a Monte Carlo experiment, as described in Sect. \ref{methods}. There are several features on the profile of tilt angle. Firstly, at the early phase of the flux emergence, the active region maintained a high value for the tilt angle (about 23$\degr$) and began to decrease. Secondly, the main decrease of the tilt angle took place during the phase from 03:36 UT to 18:00 UT on August 24. Thirdly, the active region kept at low value of tilt angle (about 12$\degr$) after 18:00 UT, when the magnetic flux continued to increase. The pole separation shows a continuously increasing pattern with a constant rate during the entire time, which was from about 18 Mm at the beginning to about 70 Mm in the end. The mean increasing rate of pole separation was about 2.92 Mm/hr. We divide the pole separation into two components: longitude component, $d_x$, and latitude component, $d_y$. Panel (c) shows the evolutions of $d_x$ and $d_y$. The longitude component $d_x$ increased significantly all the time from about 16 Mm to 68 Mm. The enhancement of pole separation in latitude direction ($d_y$) is much smaller than that in the longitude direction. Therefore, the increase of the polar separation is mainly contributed from the enhancement in the longitude direction ($d_x$). This means that the two poles are separated from each other in longitude as the active region emerges.
\begin{figure*}
\centering
\includegraphics{fig3.eps}
\caption{Variations of magnetic flux, pole separation, and tilt angle of the emerging active region during the period from 16:12 UT on August 23 to 12:00 UT on August 25, 2018. (a) The evolution of positive and negative magnetic flux. The red and blue lines indicate the positive and negative magnetic flux. The errors in the positive and negative magnetic flux are estimated by the uncertainties from HMI observations of SHARP data; (b) The evolutions of pole separation and tilt angle. The black line denotes pole separation and the blue one denotes tilt angle; (c) The evolution of pole separation in longitude direction ($d_x$) and latitude direction ($d_y$). The black and blue lines denote the pole separation in longitude direction and latitude direction, respectively. Error bars, reported at several representative times in panels (b) \& (c), are estimated by using the Monte-Carlo method.}\label{fig3}
\end{figure*}
In order to investigate the relationships between the magnetic flux and pole separation or tilt angle during the emergence of the active region, we plot the total magnetic flux as a function of pole separation and tilt angle in Fig. \ref{fig4}. Figure \ref{fig4} (a) shows the total magnetic flux ($\varphi$) as a function of the pole separation ($d$). A positive correlation between these two parameters is found. By using a linear fitting for the logarithm of these variables, a correlation of $log_{10}(\varphi)=2.72*log_{10}(d)+16.94$ plotted by blue dotted line can be derived. This correlation is equivalent to $d$ $\sim$ $\varphi^{0.37}$. In detail, when the pole separation is smaller than $\sim$ 36.5 Mm ($log_{10} d$ $<$ $10^{1.56}$ Mm), we get that $ log_{10}(\varphi)=1.17*log_{10}(d)+19.11$ shown by the yellow line and is equivalent to $d$ $\sim$ $\varphi^{0.85}$. When the pole separation is between 36.5 Mm and 58.4 Mm ($10^{1.56}<log_{10}d$ $\leqslant10^{1.77}$ Mm), the logarithm of magnetic flux equals to 3.00$*log_{10}(d)+16.53$ shown by the red line, which is equivalent to $d$ $\sim$ $\varphi^{0.33}$. When the pole separation is larger than 58.4 Mm ($log_{10}d$ $>$ $10^{1.77}$ Mm), the logarithm of magnetic flux equals to 0.55$*log_{10}(d)+20.85$ plotted by the pink line, which is equivalent to $d$ $\sim$ $\varphi^{1.82}$. In other words, during the emergence of active region, the magnetic flux has a relatively small increasing rate when the pole separation is shorter than 36.5 Mm and longer than 58.4 Mm. The magnetic flux experiences a significant increase at the moderate pole separation of the range from 36.5 Mm to 58.4 Mm. This fact manifests that the magnetic flux of naked AR main increases in the moderate pole separation. Fig. \ref{fig4} (b) exhibits the total magnetic flux as a function of the tilt angle. We also use a linear function to fit the scatter plot. A correlation of $\vartheta$ $=$ $-10.66*log_{10}(\varphi)+224.10$ can be derived. This means that there is a negative correlation between the tilt angle and magnetic flux during the dipolar AR emergence. In detail, the active region with small magnetic flux ($<$ 10$^{20.8}$ Mx) kept a high and almost constant tilt angle of $\sim$ 23$\degr$-24$\degr$. The active region with large magnetic flux ($>$ 10$^{21.6}$ Mx) kept in a small and almost constant tilt of $\sim$ 12$\degr$-13$\degr$. The tilt angle decreased gradually from about 23$\degr$ to 13$\degr$, when the total magnetic flux increased from 10$^{20.8}$ to 10$^{21.6}$ Mx. This also manifests that the newly emerging magnetic flux with a high tilt angle emerges from subsurface at first. And then two polarities of emerging magnetic flux separate mainly from each other in longitude direction, which results in the decrease in tilt angle.
\begin{figure*}
\centering
\includegraphics{fig4.eps}
\caption{Correlation between the magnetic flux and pole separation or tilt angle; (a) profile of total magnetic flux as a function of the pole separation. The different color lines denote the fitting lines with different slopes. The different color functions depict the results of different fitting lines by using the linear fit; (b) profile of total magnetic flux as a function of the tilt angle. The fitting line by linear function is indicated by the blue dotted line.}\label{fig4}
\end{figure*}
\subsection{Magnetic properties}
Figure \ref{fig5} shows the distributions of the force-free parameter $\alpha_z$, vertical electric current ($J_z$), and proxy of heilicty flux density ($G_\theta$) in the active region. Panels (a1)-(a4) exhibit the distributions of the force-free parameter $\alpha_z$ in the photosphere at different moments. The force-free parameter shows a disordered pattern in each polarity. It is interesting to find that the strong force-free parameter mainly located in the periphery of main polarities. Panels (b1)-(b4) show the distributions of the vertical electric current density ($J_z$) in the photosphere at the corresponding moments. We also find that the vertical electric current shows a disordered in each polarity and the strong vertical electric current also located in the periphery of main polarities. Panels (c1)-(c4) display the distributions of magnetic helicity ejection flux density ($G_\theta$) in the photosphere at different moments. The density of the helicity ejection flux shows a relative uniform pattern unlike the force-free parameter and vertical electric current (positive and negative helicity ejection fluxes are dominant in each polarity). At around 04:54 UT and 14:42 UT on August 24, the helicity ejection flux was small and dominated by positive value (see panels (c1)-(c2)). At around 00:30 UT on August 25, it was also dominated by positive helicity ejection flux for the whole active region. However, the leading sunspot (positive polarity) was dominated by negative helicity ejection flux (see panel (c3)). At around 10:18 UT on August 25, the negative helicity ejection flux was dominant in the active region, especially in the leading sunspot (positive polarity).
\begin{figure*}
\centering
\includegraphics{fig5.eps}
\caption{Distributions of force-free parameter ($\alpha_z$), vertical electric current density, proxy of heilicty flux density ($G_\theta$); (a1)-(a4) the force-free parameter ($\alpha_z$); (b1)-(b4) the vertical electric current density ($J_z$). The pink and black contours denote the positive and negative magnetic field with the levels of 500 G, 1500 G, respectively; (c1)-(c4) the proxy of helicity flux density (G$_\theta$). It only plots the pixels with the total magnetic field strength bigger than 300 G.}\label{fig5}
\end{figure*}
Figure \ref{fig6} shows the evolution of different magnetic parameters in the whole active region during the emergence of the active region. Panel (a) exhibits the variation of the average flux-weighted force-free parameter $\tilde{\alpha}$ calculated by the Eq.\ref{eq11}. The average force-free parameter is almost positive during the emergence of the active region, which does not follow the hemisphere rule. The magnitude of the average force-free parameter is the order of $\sim$ $10^{-8}$ m$^{-1}$, which is consistent with results from previous studies \citep{wan00,liu14,kut19}. It rises up to 6 $\times$ $10^{-8}$ m$^{-1}$ at 10:00 UT on Aug 24. At the end of the emergence, the average force-free parameter is only $\sim$ 0.5 $\times$ $10^{-8}$ m$^{-1}$. In contrast to the result from \cite{wan00}, it is variant instead of constant during the emergence of the active region. Panel (b) exhibits the evolutions of the vertical electric current and the current neutralization ratio ($\Re$). The red line indicates variation of the positive electric current while the blue line indicates variation of the negative electric current. The positive and negative electric currents are almost equivalent during the entire emergence of the AR. This fact manifests that no any net current is carried into the atmosphere during the emergence of this isolated AR. It is reasonable to assume that the total current in the isolated AR stays neutralized during its emergence; this is because the total current $I_{total}$ over the whole photospheric AR is equal to $\frac{c}{4\pi}\oint \bf{B_t} \cdot \it d \bf{l,}$ and $\bf{B_t}$ would be balanced along the path outside the boundary of the isolated AR \citep{par96}. At the early phase of emergence, they remained in low magnitude of less than 1 $\times$ $10^{12}$ A. At the main phase of emergence, they experienced a rapid increase at first and then decreased at the latter phase. During the period from around 03:48 UT on Aug 24 to 03:24 UT on Aug 25, both positive and negative electric currents increased mainly from about 0.5 $\times$ $10^{12}$ A to about 7.5 $\times$ $10^{12}$ A. After around 04:00 UT on Aug 25, they experienced a decrease until the end time of investigation. The peak time of electric current (about 03:24 UT on August 25) was earlier than that of the magnetic flux (about 10:00 UT on August 25) (see Figs. \ref{fig3} \& \ref{fig6}). The black line shows the variation of the current neutralization ratio ($\Re$). The $\Re$ values are in the range from about 0.8 to 1.1, which suggests that the AR kept neutral current during its emergence. This result is consistent with the previous observational studies, which also supports the premise that currents become more neutralized during magnetic flux emergence in flare-quiet ARs \citep{liu17,ava20}.
\begin{figure*}
\centering
\includegraphics{fig6.eps}
\caption{Evolution of different magnetic parameters in the active region during the period from 16:12 UT on August 23 to 12:00 UT on August 25, 2018: (a) The flux-weighted force-free parameter ($\tilde{\alpha}$); (b) Vertical electric currents and the current neutralization ratio. Red and blue lines denote the positive and negative vertical current, respectively. The black line plots the profile of current neutralization ratio; (c) The magnetic helicity injection flux. Red and blue lines denote the emergence term and shear term, respectively; (d) Accumulative helicity. Red and blue lines indicate the accumulative helicity contributed from emergence term and shear term. The total accumulative helicity is shown by the black line. The green line denotes the flux of GOES soft X-ray 1-8 $\rm\AA$. The two vertical dashed lines denote the times of 05:00 UT on August 24 and 09:00 UT on August 25. Some B-class flares are marked by blue arrows. Error bars estimated by the Monte-Carlo experiment are plotted at representative times.}\label{fig6}
\end{figure*}
With the methods described in Sect. \ref{methods}, we also calculated the magnetic helicity injection flux of the active region during the emergence of the active region. According to Eq.\ref{eq13}, we could obtain the magnetic helicity injection flux from the emergence term and the shear term. Figure \ref{fig6} (c) exhibits the evolution of magnetic helicity injection flux. The magnetic helicity flux contributed by the emergence term is plotted by the red line while the magnetic helicity flux contributed by the shear term is plotted by the blue line. There are several features in the magnetic helicity flux. Firstly, both terms of the helicity injection flux kept a low value at the early phase of the emergence. Secondly, the shear term changed its sign from positive to negative at about 04:00 UT on Aug 25. Thirdly, the emergence term stayed in positive during the whole period. On the other hand, we calculate the accumulated helicity by the time integral of the helicity injection flux with setting the accumulated helicity to be zero at the beginning of investigation. Figure \ref{fig6} (d) shows the time variation of accumulative helicity. At the early phase, the accumulative helicity was very low. At the main phase of the emergence, the total accumulative helicity also showed the same pattern as the vertical current, which was increasing at first and then began to decrease. Overall, this active region is mainly dominated by the positive helicity. This feature does not obey the hemisphere rule (being positive in the northern hemisphere and negative in the southern hemisphere at the 24 solar cycle). Furthermore, during the most of time, the total magnetic helicity in the active region was contributed mainly by the shear term, which is also consistent with previous findings \citep{liu12,liu14,wan19}.
The green line in panel (d) denotes the variation of soft X-ray 1-8 $\rm\AA$ flux during the emergence of the active region. We can find that several B-class flares took place in this active region. These flares marked by blue arrows in panel (d) mainly appeared at the latter phase of the emergence. As is well known, both the electric current and magnetic helicity are the proxies of non-potential magnetic field. In general, the flares would transport the magnetic energy to other forms of energy (such as thermal energy and kinetic energy of plasma) by magnetic reconnection. It implies that these flares may be associated with the decrease of electric current and the magnetic helicity at the latter phase.
\subsection{Magnetic twist}
\begin{figure*}
\centering
\includegraphics{fig7.eps}
\caption{Evolution of twist numbers ($\bar{n}$ and $T_{n\alpha}$) in the entire active region during the emergence of the active region. The asterisk denotes the profile of $\bar{n}$ while the solid line denotes the one of $T_{n\alpha}$. The horizontal dotted line denotes zero of twist number. The two vertical dashed lines denote the times of 05:00 UT on August 24 and 09:00 UT on August 25. The errors are transferred from the parameter's errors estimated by the Monte-Carlo experiment. }\label{fig7}
\end{figure*}
On one hand, the magnetic helicity inside a uniformly twisted flux tube with n turns, can be estimated by $\rm H = n \varphi^2$ \citep{ber84}, where $\varphi$ is axial magnetic flux. Therefore, if we assume that the magnetic field of active region is a uniformly twisted flux tube, the twist number of active region can be derived by the magnetic helicity and magnetic flux. The black line of asteriskes in Figure \ref{fig7} plots the twist number, $\bar{n} \approx \rm H(t)/\bar{\varphi}^2$, where $\bar{\varphi}$ is the average magnetic flux between two polarities ($\bar{\varphi}=(\varphi_{zp}+\varphi_{zn})/2$) and H(t) is total accumulated helicity. This quantity ($\bar{n}$) indicates how much twist is in the the magnetic configuration of the entire active region. At the early phase, the twist number had changed its sign several times. This feature may be related to the small and mixed helicity injection flux from the subsurface at the early phase. At the main phase, the twist number increased to about 0.05 at first and then it lowered slightly and remained at about 0.04. After that, it experienced a sharp decrease until the end of our investigation time frame. Except for the early phase, the twist number almost kept in positive during the emergence of the active region. Due to the small magnetic helicity and magnetic flux at the early phase, we only focus on the twist number in the main phase of emergence. We were able to find that the twist number was on the order of 10$^{-1}$ during the emergence. The positive sign of the twist number is consistent with the positive sign of force-free parameter, $\tilde{\alpha}$, which does not follow the hemisphere rule either. On the other hand, we assume that the emerged magnetic field is a linear force-free field. The force-free parameter is set to be $\tilde{\alpha}$. Therefore, the twist number (T$_{n\alpha}$) of magnetic field could be estimated by the function of T$_{n\alpha}=\frac{1}{4\pi}\int \tilde{\alpha}dL$, in which the L is the length of the magnetic field and simply assume to be $\frac{\pi d}{2}$. Based on the above equation and reasonable assumptions, the T$_{n\alpha}$ could be calculated by using the evolutions of $\tilde{\alpha}$ and pole separation (d). The profile of T$_{n\alpha}$ shows by the solid blue line as the function of time in Fig. \ref{fig7}. Interestingly, T$_{n\alpha}$ was also on the order of 10$^{-1}$ turns, which is consistent with the twist number ($\bar{n}$) derived by the helicity calculation. In detail, the T$_{n\alpha}$ is slightly bigger than $\bar{n}$ at the early stage of main emergence. At the end of emergence, the T$_{n\alpha}$ is also about 0.04 turns, which is consistent with $\bar{n}$. This might be related to the continuous increasing twist in the emerged magnetic field at the early stage of main emergence. Nevertheless, both T$_{n\alpha}$ and $\bar{n}$ show almost the same value at the end of emergence. These results manifest that the twist of this naked anti-Hale active region is very low and in the order of 10$^{-1}$ turns. The errors of $\bar{n}$ mainly come from the uncertainty of accumulated helicity related to photospheric velocity and magnetic field, while the errors of T$_{n\alpha}$ mainly come from the uncertainty of force-free parameters related to magnetic field. We use a Monte Carlo experiment method to estimate the uncertainty of these parameters. Therefore, the errors of T$_{n\alpha}$ and $\bar{n}$ can be estimated by the propagation from these parameters' errors, which are shown by using error bars in Fig. \ref{fig7}. On the other hand, the lack of data during periods from 06:24 to 08:48 UT on August 24 and 25 might lead to an underestimation of the accumulated helicity, which results in the underestimation of $\bar{n}$. Furthermore, we use a photospheric averaged flux-weighted force-free parameter $\tilde{\alpha}$ to represent the force-free parameter of magnetic field, which might bring on some uncertainties in estimating T$_{n\alpha}$. Therefore, we take a conservative estimate that the twist number of this emerging active region is on the order of 10$^{-1}$ turns. Nevertheless, we use two different methods to estimate the twist number of the active region and obtain the consistent result, which demonstrates that the result is reliable.
\section{Summary and discussion}\label{sec:conclusion}
This paper presents a detailed study on the evolution and magnetic properties of an anti-Hale emerging active region NOAA 12720. Our main results are as follows.
(1) During the emergence of the anti-Hale active region, the leading sunspot with positive magnetic field drifted faster than the following one with negative magnetic field. Most of the newly emerging magnetic fluxes initially occurred between the two main polarities.
(2) A positive correlation between magnetic flux and pole separation is found during the emergence. The increase of the pole separation is mainly contributed from the enhancement in the longitudinal direction. The tilt angle of the active region maintains a high value (about 23$\degr$) at the early phase and a low value (about 12$\degr$) at the latter phase. In contrast to the pole separation, the tilt angle shows a negative correlation with total magnetic flux. On the other hand, a power-law relationship between pole separation and total flux is found during the emergence of the active region.
(3) The force-free parameter ($\alpha_z$) and vertical electric current density ($J_z$) exhibit a disorderly distribution in each polarity. The average flux-weighted force-free parameter and accumulative helicity in the entire active region was positive, a result that does not obey the hemispheric rule. The electric current and magnetic helicity in the active region are mainly built up at the main phase of the emergence. In addition, the magnetic helicity injection in the active region is mainly contributed by the shear term.
(4) Through the calculation of the helicity and the force-free parameter to estimate the twist number, we infer that the twist number of the whole active region is on the order of 10$^{-1}$ turns.
According to the statistical result of 4385 sunspot groups determined by \cite{li18}, only 8.1 $\pm$ 0.4 \% of these active regions are the anti-Hale sunspot groups. Therefore, anti-Hale active region is relatively rare. This anti-Hale active region NOAA 12720 emerged in a clean environment, which provides a unique opportunity for us to investigate the nature of emerging magnetic flux and anti-Hale active region. We find that the active region emerged with a high tilt angle at the early phase. The tilt angle of the active region decreased gradually when the active region continued to emerge. The tilt angle of the active region came up to a low value at the mid-phase of emergence and kept invariant until the end of the emergence. If we hypothesize that the anti-Hale active action before emergence is a flux tube embedded in the subsurface, the evoution of the tilt angle in this anti-Hale active region is consistent with the emergence of the flux tube with low twist. As a flux tube emerges with low twist from the convective zone into the solar atmosphere, the tilt angle of two main polarities exhibits high value at the initial stage and then decreases to a lower value. As the bottom of the lifting flux tube escapes, the tilt angle remains invariant at a low value \citep{kni18}. A negative correlation between tilt angle and magnetic flux could be found during this bipolar AR emergence. However, \cite{kos08} studied 715 bipolar ARs and deduced that the tilt angle does not dependent on the amount of magnetic flux. These facts manifest the finding that the evolution of tilt angle depends on the magnetic flux during the emergence of one active region but the tilt angle does not depend on the amount of magnetic flux in various active regions.
During the emergence of this naked active region, the pole separation has a power-law dependence on the magnetic flux. The power-law index in our study is 0.37. Interestingly, \cite{kut19} also found a similar relationship between pole separation and maximum of magnetic flux using 423 ARs. Their power-law index is about 0.36 which is a similar value to ours. It should be noted that our study just focuses on the evolution of one active region, while the statistical study of \cite{kut19} includes the properties of many different emerging active regions. Therefore, the dependence of polarities separation on magnetic flux is a fundamental relationship -- either in the evolution of one emerging active region or between different emerging active regions. On the other hand, there are three states on the relation between pole separation and magnetic flux (see Fig. \ref{fig4}(a)). When the pole separation is shorter than 36.5 Mm and longer than 58.4 Mm, the magnetic flux increases with a relatively small increasing rate. When the pole separation is moderate in the range from 36.5 Mm to 58.4 Mm, the magnetic flux increases with a large increasing rate. The similar relationship between pole separation and magnetic flux is also seen in the emergence phase of NOAA active region 8167 (see Fig. 1 (f) in \cite{kos08}). Combining (a) \& (b) in Fig. 3, we see that these three states are mainly associated with the evolution of magnetic field while the pole separation increase with an almost constant rate. The first two states may be related to a two-step emergence pattern in the active region \citep{fu16}, with a relatively gradual emergence followed by a relatively sudden change to more rapid emergence. This feature has been seen in simulations of ``two-step'' model \citep{mat93,mag01,tor10}. In models of two-step emergence, the rate of emergence in the emerging flux tube becomes rapidly increasing from a relatively small as the emerging flux tube reaches around the photosphere or chromosphere. In addition, \cite{chi13} suggested that the emerging magnetic structure with a weaker-field structure preceding a stronger-field structure is responsible for the two-step emergence of AR 11158. The third presumes that the magnetic flux has a low increasing rate with pole separation at the later stage, which might be related to emergence of the serpentine magnetic field at the latter stage of emergence. These serpentine magnetic fields would release mass through magnetic reconnection \citep{che10,cen12}, resulting in the cancellation of magnetic flux on the photosphere. The cancellation of magnetic flux would slow down the growth of magnetic flux in the whole active region. Otherwise, these cancelled magnetic fluxes were often seen as the signatures of moving dipolar features (see Fig. \ref{fig2} (a3)) on the photosphere before cancellation.
The injection of magnetic helicity in this anti-Hale active region was contributed mainly from the shear term, which is consistent with previous results \citep{liu12,liu14}. According to the 3D numerical simulation of a twisted flux tube, \cite{mag03} found that relative magnetic helicity is dominated by the emergence term at the early stage of the emergence. However, we does not find this feature in our study. This may be related to the low twist in this emerging active region. We obtain that the twist number of the emerging active region is in the order of 10$^{-1}$ turns. This means that the emerging active region carried the low twist into the atmosphere. The low twist in the magnetic flux leads to a little injection of magnetic helicity contributed by the emergence term at the early stage. In other words, the low magnetic helicity injection flux contributed by the emergence term at the early stage manifests that the twist of the emerging magnetic flux is low.
According to the definition of the twist number in the axial symmetric cylindrical flux tube, we can get $B_\theta / B_l$ = $n*\frac{2\pi r}{L}$, where $B_\theta$ and $B_l$ are, respectively, the magnetic field strength of azimuthal and longitudinal components, while r is the radius of flux tube and L is the length of the flux tube. According to the definition of the twist parameter ($q=\frac{B_\theta}{rB_l}$), we can obtain that q$H_0$ = $n*\frac{2\pi H_0}{L}$, where $H_0$ is the pressure scale height ($H_0$=170 km) \citep{tor11}. With the assumption that the whole emerging active region is a flux tube, the length L equals $\frac{\pi d}{2}$. At 04:00 UT on August 25, the pole separation (d) of the active region equals to about 60 Mm. Combinated with the above values of the parameters, the twist parameter (q$H_0$) in this active region is estimated to be about 1 $\times$ 10$^{-3}$. Comparing with the critical value (q$H_0$ = 0.1) for the emergence of flux tube derived by \cite{tor11}, the twist of this emerging active region is relatively weak. Therefore, this result is opposed to the conclusion in which the flux tube with weak twist would fails to emerge into the upper atmosphere \citep{mur06,tor11}.
While the magnetic twist plays an important role in the activity and evolution of the active region, the studies of twist in the active region are needed to reveal the intrinsic properties of magnetic field \citep{nan06,luo11,poi15}. Our result indicates that the low twist in the sub-photospheric emerging part of a flux rope is coherent with previous studies \citep{liu06,poi15}. \cite{yan09} studied 58 emerging Hale active regions and obtained that the average twist number in these Hale active region is about 0.039 turns, a value that is comparable to our result. This implies that either Hale or anti-Hale active regions operate under the same rising mechanism in the solar interior. Otherwise, this can also demonstrate that this anti-Hale active region is likely to be generated by rising of toroidal flux tube with opposite orientation in the interior \citep{ste12}, instead of causing by the kink instability mechanism of highly twisted flux tube \citep{nan06,kni18}. However, the positive sign of twist including positive twist number and average force-free parameter ($\tilde{\alpha}$) is opposite the sign of structural chirality of solar active region in northern hemisphere \citep{see90}. The active region violating Hale rule might be responsible for this phenomenon.
Here, we consider the role played by the twist in the emergence of the magnetic flux tube. In our observational study, we use two different methods which are based on magnetic helicity calculation and force-free parameter to estimate the twist in this anti-Hale active region. We find that the emerged magnetic field covering the whole active region exhibits a lowly twisted structure. The twist number is in the order of 10$^{-1}$ turns. More recently, \cite{kni18} simulated the emergence of different twisted flux tubes from the convection zone into the corona. They found that the behaviours of low twisted flux tube emergence matched the observed properties of the single active region (such as $\beta$ style), while the ones of high twisted flux tube matched the complex active region ($\delta$ style). Moreover, some authors found that an untwisted flux tube could also be able to rise through the upper convection zone and emerge into the photosphere to form spots \citep{ste11,rem14,kni20}. \cite{kni20} explained that the rise of untwisted toroidal flux rope is triggered by the undular instability instead of buoyant instability. On the other hand, most studies of numerical simulation often set a mature magnetic flux rope embedding in the convection zone bottom as the initial condition. The reality is that the origin of the magnetic field of the active region is still unclear at present. One possibility is that the magnetic field of the flux rope or flux tube is forming while they are rising under the subsurface. Thus, a greater number of episodes ought to be considered in the numerical simulation. After all, we conclude that the simple active region does not need too much initial twist in the intrinsic magnetic field for its emergence.
\begin{acknowledgements}
We appreciate the referee's careful reading of the manuscript and many constructive comments, which helped greatly in improving the paper. SDO is a mission of NASA's Living With a Star Program. The authors are indebted to the SDO, NVST teams for providing the data. This work is supported by the National Key R\&D Program of China (2019YFA0405000), the National Science Foundation of China (NSFC) under grant numbers 12003064, 11873087, 11803085, Yunnan Science Foundation of China (2019FD085), the CAS ``Light of West China" Program under number Y9XB018001, the Open Research Program of the Key Laboratory of Solar Activity of Chinese Academy of Sciences (grant No.KLSA202014), the Yunnan Talent Science Foundation of China (2018FA001), the Yunnan Science Foundation for Distinguished Young Scholars under No. 202001AV070004, and the Key Research and Development Project of Yunnan Province under number 202003AD150019..
\end{acknowledgements}
|
\section{\label{sec:Intro} Introduction}
Neutrinos were originally formulated, within the weak interactions of the Standard Model (SM), to be massless fermions.
However, hints of massive neutrinos appeared in the second half of the 20th century with the solar neutrino problem\cite{Bahcall:2002ng}.
Eventually, updated solar models and experiments from the Kamiokande laboratory and the Sudbury Neutrino Observatory (SNO) proved the disappearance is caused by flavor oscillations\cite{Abe:2010hy,Aharmim:2009gd}.
For neutrinos the existence of flavor oscillations mean there is a mixing between mass and flavor eigenstates.
The mixing between mass and flavor is governed by the unitary PMNS matrix, which has six (four) free parameters in the $3\times3$ instance\cite{Pontecorvo:1957qd,Maki:1962mu}.
Thus, the mixing caused by flavor oscillations is direct evidence that neutrinos are massive and, consequently, require physics beyond the SM.
In this decade, the six parameters related to the PMNS matrix; the oscillation angles $\theta_{12},\theta_{23}, \text{and }\theta_{13}$, the mass squared differences $\Delta m_{21}^2\text{ and }\Delta m^2_{31}$, and the Dirac CP violating phase $\delta_{cp}$, are expected to be measured within a few percentages by neutrino oscillation experiments\cite{Abi:2020evt,Abe:2015zbg,Abe:2016ero,Baussan:2013zcy,Kumar:2017sdq,An:2015jdp}. In addition, the absolute mass of the lightest neutrino, $m_1$ or $m_3$, is expected to be directly limited down to the sub-eV range by the Karlsruhe Tritium Neutrino Experiment (KATRIN)\cite{Aker:2019qfn}. However, even with these precision measurements, questions remain in neutrino phenomenology. Specifically, the question of what is the absolute mass hierarchy for neutrinos, normal $m_1<m_2\ll m_3$ or inverted $m_3\ll m_1<m_2$,
and the question of what type of mass neutrinos possess, Dirac or Majorana. Those questions could be answered by future neutrinoless double-$\beta$ $(0\nu\beta\beta)$ decay experiments\cite{Dolinski:2019nrj,Furry:1939qr,Bahcall:1978jn,Schechter:1980gk,Xing:2013ty}. Arguably, additional approaches to answer those questions would be ideal.
For this work, we present a unified description of neutrino phenomenology that leads to an additional approach for investigations of the previously mentioned questions. We define unified to mean a description that includes neutrino flavor oscillations and, particle-anti-particle or chiral oscillations.
In addition, our description is different from the usual neutrino oscillation theory that assumes neutrinos are relativistic.
The relativistic assumption is often taken, because cosmological limits on the neutrinos absolute masses place them in the sub-eV range\cite{Aghanim:2018eyx}; whereas, experiments are performed on neutrinos with energies in the KeV, MeV, GeV, and recently the PeV ranges\cite{Formaggio:2013kya}.
Nonrelativistic neutrinos are predicted to exist in nature as the cosmic neutrino background (C$\nu$B).
Experiments to detect the C$\nu$B are under preparation (see for example, \cite{Betti:2019ouf}), and we plan future studies of the C$\nu$B as an application of our description.
For studies of the C$\nu$B we need to build a framework that uses a momentum distribution \cite{Dolgov:2002wy}.
For oscillation experiments, if the neutrino is produced with a localized space a momentum distribution description is necessary.
Descriptions of that type have been done using wave packet formulations \cite{Giunti:1991ca,Giunti:1993se,Giunti:2002xg,Beuthe:2001rc,Kayser:1981ye}.
To incorporate a momentum distribution in our framework, we consider densities of the lepton family numbers that are localized in space.
We have previously discussed details of the formulation for the Majorana mass case using plane waves \cite{Adam:2021qiq}.
\section{Lepton Number Density}
\subsection{Majorana neutrino formulation}
We consider the following Lagrangian $\mathcal{L}^{M}$ for Majorana neutrinos,
\begin{equation}
\mathcal{L}^{M} = \overline{\nu_{L\alpha}} i\gamma^\mu\partial_\mu \nu_{L\alpha}
- \theta(t) \left( \frac{m_{\alpha \beta}}{2}\overline{(\nu_{L\alpha})^{C}}\, \nu_{L\beta}^{} + \text{h.c} \right),
\label{Eq:Lagrangian}
\end{equation}
in which the first term is kinetic one and the second term is the Majorana mass one.
A step-function controls the second term, and initially guarantees a pure lepton family number.
A weak interaction produces the neutrinos at $t =-\epsilon$ for ${\epsilon \rightarrow 0+}$, and we treat them as part of a $\text{SU}(2)_L$ doublet.
The $\text{SU}(2)_L$ doublet is used to assign a lepton family number of $L_e , L_\mu, \text{ or } L_\tau$ to the produced neutrino \cite{Bilenky:2001yh}.
We then formulate a Heisenberg operator for $L_{\alpha}$, where $\alpha=(e,\mu,\tau)$, that is distributed over space.
\begin{gather}
L^M_{\alpha}(t)=\int d^3x \, l^M_{\alpha}(t,\mathbf{x}) ,
\label{Eq:LeptonNumber} \\
l^M_{\alpha}(t,\mathbf{x}) = \, :\overline{\nu_{L\alpha}}(t,\mathbf{x})\gamma^0 \nu_{L\alpha}(t,\mathbf{x}):\,,
\label{Eq:LeptonNumberDensity}
\end{gather}
where the colon : denotes normal ordering.
For $t=-\epsilon$ with ${\epsilon \rightarrow 0+}$ the neutrinos are massless spinors that follow the usual Fourier expansion,
\begin{equation}
\lim_{\epsilon\rightarrow0+}\nu_{L\alpha}(-\epsilon,\mathbf{x})=\int \frac{d^3\mathbf{p}}{(2\pi)^32|\mathbf{p}|}
\left(a_{\alpha}^{}(\mathbf{p})u_L(\mathbf{p})e^{i\mathbf{p}\cdot\mathbf{x}}+b^\dagger_{\alpha}(\mathbf{p})v_L(\mathbf{p})e^{-i\mathbf{p}\cdot\mathbf{x}}\right).
\label{Eq:WeylExpasion}
\end{equation}
At $t=0$ we assign a Majorana mass to the neutrino by way of the step-function in Eq.(\ref{Eq:Lagrangian}).
Then, the time evolution of the lepton family number in Eq.(\ref{Eq:LeptonNumber}) becomes dependent on the mass.
We rewrite Eq.(\ref{Eq:LeptonNumberDensity}) using Eq.(\ref{Eq:WeylExpasion});
\begin{multline}
l^M_{\alpha}(t,\mathbf{x}) = \int'\frac{d^3 \mathbf{k}}{(2\pi)^3 2|\mathbf{k}|}\int'\frac{d^3 \mathbf{p}}{(2\pi)^3 2|\mathbf{p}|} \\
\times\left[a^\dagger_\alpha(\mathbf{k},t)a_\alpha(\mathbf{p},t)\overline{u_L}(\mathbf{k})\gamma^0u_L(\mathbf{p})e^{-i(\mathbf{k}-\mathbf{p})\cdot\mathbf{x}}
+b_\alpha(\mathbf{k},t)a_\alpha(\mathbf{p},t)\overline{v_L}(\mathbf{k})\gamma^0u_L(\mathbf{p})e^{i(\mathbf{k}+\mathbf{p})\cdot\mathbf{x}} \right.\\
+\left.a^\dagger_\alpha(\mathbf{k},t)b^\dagger_\alpha(\mathbf{p},t)\overline{u_L}(\mathbf{k})\gamma^0v_L(\mathbf{p})e^{-i(\mathbf{k}+\mathbf{p})\cdot\mathbf{x}}
-b^\dagger_\alpha(\mathbf{p},t)b_\alpha(\mathbf{k},t)\overline{v_L}(\mathbf{k})\gamma^0v_L(\mathbf{p})e^{i(\mathbf{k}-\mathbf{p})\cdot\mathbf{x}}\right],
\label{Eq:MajoranaDensity}
\end{multline}
where the integration region $\int'$ does not include the zero momentum mode.
We take the expectation value of the Majorana density operator to study the spacetime evolution $\langle \psi_\sigma(q^0;\sigma_q)|l^M_{\alpha}(t,\mathbf{x})|\psi_\sigma(q^0;\sigma_q)\rangle$.
The initial momentum state of $\lvert \psi_\sigma(q^0;\sigma_q)\rangle$ is equally distributed in the first and third dimensions, but has a Gaussian distribution in the second, i.e., $\mathbf{q}=(0,q,0)$.
This leads to a shape similar to a 1-D Gaussian wave packet,
\begin{equation}
\lvert \psi_\sigma(q^0;\sigma_q) \rangle = \frac{1}{\sqrt{\sigma_q}(2\pi)^{3/4}\delta(0)}
\int'\frac{dq}{2\pi\sqrt{2|q|}}e^{-\frac{(q-q^0)^2}{4\sigma_q^2}}
a_\sigma^\dagger(0,q,0)\lvert 0\rangle ;
\label{Eq:Wavepacket}
\end{equation}
where $\sigma_q$ is the width of the Gaussian distribution in the second component of the momentum.
The second component of the mean momentum $q^0$ is positive for the Gaussian distribution.
Sandwiching the Majorana density operator of Eq.(\ref{Eq:MajoranaDensity}) with Eq.(\ref{Eq:Wavepacket}) and taking the integration over $\mathbf{k}$ and $\mathbf{p}$ results in,
\begin{multline}
\langle \psi_\sigma(q^0;\sigma_q)|l^M_{\alpha}(t,\mathbf{x})|\psi_\sigma(q^0;\sigma_q)\rangle =
\frac{1}{\sigma_q(2\pi)^{3/2}\delta(0)^2} \iint' \frac{dq'dq}{(2\pi)^2}
e^{-\frac{(q'-q^0)^2+(q-q^0)^2}{4\sigma_q^2}-i(q'-q)\mathbf{e}_2\cdot\mathbf{x}} \\
\times\left[\sum_{i,j} V^{\ast}_{\alpha i}V^{}_{\sigma i}V^{}_{\alpha j}V^{\ast}_{\sigma j}
\left(\cos{E_i(q')t}+i\frac{|q'|}{E_i(q')}\sin{E_i(q')t}\right)
\left(\cos{E_j(q)t}-i\frac{|q|}{E_j(q)}\sin{E_j(q)t}\right) \right. \\
\left. -\sum_{i,j} V^{\ast}_{\alpha i}V^{\ast}_{\sigma i}V^{}_{\alpha j}V^{}_{\sigma j}
\frac{m_j}{E_j(q')}\sin{E_j(q')t}\frac{m_i}{E_i(q)}\sin{E_i(q)t}\right];
\label{Eq:kpExpectationValue}
\end{multline}
where $\mathbf{e}^T_2=(0,1,0)$ is the unit vector.
To perform the integration over $q'$ and $q$ we must assume two properties about the Gaussian distributions.
First, the distributions are sharply peaked around the mean momentum value $q^0$, i.e., $\sigma_q \ll q^0$.
Second, the width (variance) of the distribution $\sigma_q$ does not change in spacetime.
Those two assumptions allow us to approximate the $q'$ and $q$ integration as Gaussian; because,
\begin{equation}
E_{i,j}(q^{(\prime)}) \simeq E_{i,j}(q^0)+\frac{q^0}{E_{i,j}(q^0)}(q^{(\prime)}-q^0).
\label{Eq:IntegralApprox}
\end{equation}
The Gaussian integration of Eq.(\ref{Eq:kpExpectationValue}), and integration over $x_1$ and $x_2$, results in a linear density,
\begin{multline}
\begin{aligned}
\lambda^M_{\sigma\rightarrow\alpha}(t,x_2) \simeq & \frac{\sigma_q}{(2\pi)^{1/2}}\sum_{i,j} V^{\ast}_{\alpha i}V^{}_{\sigma i}V^{}_{\alpha j}V^{\ast}_{\sigma j} \\
&\times\frac{1}{2}\left[(v_{i0}+v_{j0}+1+v_{i0}v_{j0})e^{i(E_i(q^0)-E_j(q^0))t}e^{-\sigma^2_q[(x_2-v_{i0}t)^2+(x_2-v_{j0}t)^2]}\right.\\
&\quad\quad-(v_{i0}+v_{j0}-1-v_{i0}v_{j0})e^{-i(E_i(q^0)-E_j(q^0))t}e^{-\sigma^2_q[(x_2+v_{i0}t)^2+(x_2+v_{j0}t)^2]}\\
&\quad\quad+(v_{i0}-v_{j0}+1-v_{i0}v_{j0})e^{i(E_i(q^0)+E_j(q^0))t}e^{-\sigma^2_q[(x_2-v_{i0}t)^2+(x_2+v_{j0}t)^2]} \\
&\left.\quad\quad-(v_{i0}-v_{j0}-1+v_{i0}v_{j0})e^{-i(E_i(q^0)+E_j(q^0))t}e^{-\sigma^2_q[(x_2+v_{i0}t)^2+(x_2-v_{j0}t)^2]} \right] \\
&-\frac{\sigma_q}{(2\pi)^{1/2}}\sum_{i,j} V^{\ast}_{\alpha i}V^{\ast}_{\sigma i}V^{}_{\alpha j}V^{}_{\sigma j} \sqrt{1-v^2_{i0}}\sqrt{1-v^2_{j0}}
\end{aligned} \\
\begin{aligned}
\times&\frac{1}{2}\left[e^{i(E_i(q^0)-E_j(q^0))t}e^{-\sigma^2_q[(x_2-v_{i0}t)^2+(x_2-v_{j0}t)^2]} +e^{-i(E_i(q^0)-E_j(q^0))t}e^{-\sigma^2_q[(x_2+v_{i0}t)^2+(x_2+v_{j0}t)^2]}\right. \\
&\quad-e^{i(E_i(q^0)+E_j(q^0))t}e^{-\sigma^2_q[(x_2-v_{i0}t)^2+(x_2+v_{j0}t)^2]}
\left.-e^{-i(E_i(q^0)+E_j(q^0))t}e^{-\sigma^2_q[(x_2+v_{i0}t)^2+(x_2-v_{j0}t)^2]} \right]
\end{aligned},
\label{Eq:ExpectationValue}
\end{multline}
where $v_{i,j0}=q^0/E_{i,j}(q^0)$ are the group velocities of the distributions, $x_2=\mathbf{e}_2\cdot\mathbf{x}$, and $ \lambda^M_{\sigma\rightarrow\alpha}(t,x_2) = \iint dx_1 dx_3 \langle \psi_\sigma(q^0;\sigma_q)|l^M_{\alpha}(t,\mathbf{x})|\psi_\sigma(q^0;\sigma_q)\rangle$.
The terms with the PMNS matrix combination $V^{\ast}_{\alpha i}V^{\ast}_{\sigma i}V^{}_{\alpha j}V^{}_{\sigma j}$ are dependent on the Majorana phases and are suppressed by the small masses of neutrinos $\sqrt{1-v^2_{i,j0}}=m_{i,j}/E_{i,j}(q^0)$.
Lastly, to compare with our previous work's result that considered only the time evolution of plane waves, we integrate the linear density of Eq.(\ref{Eq:ExpectationValue}) over all space,
\begin{multline}
\begin{aligned}
\int \lambda^M_{\sigma\rightarrow\alpha}(t,x_2) dx_2 \simeq & \frac{1}{4}\sum_{i,j}V^{\ast}_{\alpha i}V^{}_{\sigma i}V^{}_{\alpha j}V^{\ast}_{\sigma j} \\
&\times\left[(v_{i0}+v_{j0}+1+v_{i0}v_{j0})e^{i(E_i(q^0)-E_j(q^0))t}e^{-\sigma^2_q\frac{(v_{i0}+v_{j0})^2t^2}{2}}\right. \\
&\quad\quad-(v_{i0}+v_{j0}-1-v_{i0}v_{j0})e^{-i(E_i(q^0)-E_j(q^0))t}e^{-\sigma^2_q\frac{(v_{i0}+v_{j0})^2t^2}{2}} \\
&\quad\quad+(v_{i0}-v_{j0}+1-v_{i0}v_{j0})e^{i(E_i(q^0)+E_j(q^0))t}e^{-\sigma^2_q\frac{(v_{i0}-v_{j0})^2t^2}{2}} \\
&\left.\quad\quad-(v_{i0}-v_{j0}-1+v_{i0}v_{j0})e^{-i(E_i(q^0)+E_j(q^0))t}e^{-\sigma^2_q\frac{(v_{i0}-v_{j0})^2t^2}{2}} \right] \\
&-\frac{1}{4}\sum_{i,j}V^{\ast}_{\alpha i}V^{\ast}_{\sigma i}V^{}_{\alpha j}V^{}_{\sigma j}\sqrt{1-v^2_{i0}}\sqrt{1-v^2_{j0}}
\end{aligned} \\
\times\left[ (e^{i(E_i(q^0)-E_j(q^0))t}+e^{-i(E_i(q^0)-E_j(q^0))t}) e^{-\sigma^2_q\frac{(v_{i0}+v_{j0})^2t^2}{2}}\right. \\
\left.-(e^{i(E_i(q^0)+E_j(q^0))t}+e^{-i(E_i(q^0)+E_j(q^0))t})e^{-\sigma^2_q\frac{(v_{i0}-v_{j0})^2t^2}{2}} \right] .
\label{Eq:MajoranaSpaceTimeInt}
\end{multline}
The result of our previous work \cite{Adam:2021qiq} is recovered in the plane wave limit, for which we take the width $\sigma_q$ of the density in momentum space to zero.
\subsection{Dirac neutrino formulation}
We consider the following Lagrangian $\mathcal{L}^D$ for Dirac neutrinos,
\begin{equation}
\mathcal{L}^D=
\overline{\nu_{L\alpha}} i \gamma^\mu \partial_\mu \nu_{L\alpha}
+\overline{\nu_{R\alpha}} i \gamma^\mu \partial_\mu \nu_{R\alpha}
-\theta(t)\left(\overline{\nu_{R\alpha}}m_{\alpha\beta}\nu_{L\beta} + \text{h.c.}
\right),
\label{eq:Lagmulti1}
\end{equation}
The first and second terms are kinetic ones and third term is Dirac mass one.
Similar to the Majorana case, a time dependent step function controls the third term.
In contrast to the Majorana case, the Dirac mass matrix is diagonalized by two mixing matrices,
\begin{gather}
\nu_{L\beta}=V_{\beta j}\nu_{Lj},
\\
\nu_{R\alpha}=U_{\alpha i}\nu_{Ri},
\\
(U^\dagger)_{i\alpha}m_{\alpha\beta}V_{\beta j}=m_i\delta_{ij} .
\end{gather}
We relate the two regions of the step-function in Eq.(\ref{eq:Lagmulti1}) with the continuity condition at $t=0$:
\begin{gather}
\lim_{\epsilon\rightarrow0+}\nu_{Li}(t=+\epsilon) =\lim_{\epsilon\rightarrow0+} \sum_{\beta=e}^\tau V_{\beta i}^\ast \nu_{L\beta}(t=-\epsilon)
\label{Eq:LeftHandedContinuity}
\\
\lim_{\epsilon\rightarrow0+}\nu_{Ri}(t=+\epsilon) =\lim_{\epsilon\rightarrow0+} \sum_{\alpha=e}^\tau U_{\alpha i}^\ast \nu_{R\alpha}(t=-\epsilon).
\label{Eq:RightHandedContinuity}
\end{gather}
The Weyl fields of $\nu_{R\alpha}$ and $\nu_{L\beta}$ are Fourier expanded,
\begin{gather}
\lim_{\epsilon\rightarrow 0+}\nu_{L\alpha}(-\epsilon,\mathbf{x}) = \int\frac{d^3\mathbf{p}}{(2\pi)^3 2|\mathbf{p}|} \left( u_L(\mathbf{p})a_{L\alpha}(\mathbf{p})e^{i \mathbf{p}\cdot\mathbf{x}} + v_L(\mathbf{p}) b_{L\alpha}^\dagger(\mathbf{p})e^{-i\mathbf{p}\cdot\mathbf{x}} \right), \label{eq:leftmasslessexpand}
\\
\lim_{\epsilon\rightarrow0+}\nu_{R\alpha}(-\epsilon,\mathbf{x}) = \int\frac{d^3\mathbf{p}}{(2\pi)^3 2|\mathbf{p}|} \left( u_R(\mathbf{p})a_{R\alpha}(\mathbf{p})e^{i\mathbf{p}\cdot\mathbf{x}} + v_R(\mathbf{p}) b_{R\alpha}^\dagger(\mathbf{p})e^{-i\mathbf{p}\cdot\mathbf{x}} \right).
\end{gather}
Then the operators obey the usual anti-commutation relations of,
\begin{align}
\{a_{L\alpha}(\pm \mathbf{p}), a_{L\beta}^\dagger(\pm \mathbf{q})\} &= 2|\mathbf{p}|(2\pi)^3\delta^{(3)}(\mathbf{p}-\mathbf{q})\delta_{\alpha\beta}, \label{eq:flavorOperatorsa}
\\
\{b_{L\alpha}(\pm \mathbf{p}), b_{L\beta}^\dagger(\pm \mathbf{q})\} &= 2|\mathbf{p}|(2\pi)^3\delta^{(3)}(\mathbf{p}-\mathbf{q})\delta_{\alpha\beta}, \label{eq:flavorOperatorsb}
\end{align}
where we denote the three-dimensional momenta $+\mathbf{p}$ and $-\mathbf{p}$. $+\mathbf{p}$ belongs to a hemisphere region where the azimuthal angle $\phi$ in the momentum direction is
$ \phi \in [0, \pi)$ and $-\mathbf{p}$ belongs to the other hemisphere region that corresponds to $ \phi \in [\pi, 2\pi)$
\cite{Adam:2021qiq}.
We take the usual definition of a Dirac field $\psi_i$ as $\psi_i=\nu_{R i}+\nu_{L i}$ with the on-shell Fourier expansion,
\begin{equation}
\lim_{\epsilon\rightarrow0+}\psi_i(+\epsilon,\mathbf{x}) = \int\frac{d^3\mathbf{p}}{(2\pi)^3 2E_i(\mathbf{p})} \sum_{h} \left(
u_i(\mathbf{p},h)a_i(\mathbf{p},h)e^{i\mathbf{p} \cdot \mathbf{x}}
+ v_i(\mathbf{p},h)b_i^\dagger(\mathbf{p},h)e^{-i\mathbf{p}\cdot \mathbf{x}}
\right),
\label{eq:onshellDirac}
\end{equation}
and $h$ is the helicity.
The mass operators $a_i(\mathbf{p},h)$ and $b_i(\mathbf{p},h)$ obey the usual anti-commutation relations,
\begin{align}
\{a_i(\mathbf{p},h), a_j^\dagger(\mathbf{q},h)\} & = (2\pi)^3 2E_i(\mathbf{p}) \delta_{ij}\delta^{(3)}(\mathbf{p}-\mathbf{q}), \label{eq:massOperatorsa}
\\
\{b_i(\mathbf{p},h), b_j^\dagger(\mathbf{q},h)\} & = (2\pi)^3 2E_i(\mathbf{p}) \delta_{ij}\delta^{(3)}(\mathbf{p}-\mathbf{q}). \label{eq:massOperatorsb}
\end{align}
All other anti-commutation relations are zero.
The continuity conditions of Eqs.(\ref{Eq:LeftHandedContinuity}-\ref{Eq:RightHandedContinuity}) are used to derive relations between the operators of Eqs.(\ref{eq:flavorOperatorsa}, \ref{eq:flavorOperatorsb}) and Eqs.(\ref{eq:massOperatorsa}, \ref{eq:massOperatorsb});
\begin{align}
\frac{1}{\sqrt{2|\mathbf{p}|}}a_{L\alpha}(\pm \mathbf{p}) & = \sum_{i}^3
V_{\alpha i}\frac{\sqrt{E_i(\mathbf{p})+|\mathbf{p}|}}{2E_i(\mathbf{p})}
\left(a_i(\pm \mathbf{p},-)\pm i\frac{m_i}{E_i(\mathbf{p})+|\mathbf{p}|}b_i^\dagger(\mp \mathbf{p},-)\right), \label{eq:operatorarelations}
\\
\frac{1}{\sqrt{2|\mathbf{p}|}}b_{L\alpha}^\dagger(\pm \mathbf{p}) & = \sum_{i}^3
V_{\alpha i}\frac{\sqrt{E_i(\mathbf{p})+|\mathbf{p}|}}{2E_i(\mathbf{p})}
\left(b_i^\dagger(\pm \mathbf{p},+)\mp i\frac{m_i}{E_i(\mathbf{p})+|\mathbf{p}|} a_i(\mp \mathbf{p},+)\right). \label{eq:operatorbrelations}
\end{align}
Where the relations for the right-handed operators $a_{R\alpha},\,b_{R\alpha}$ are found by replacing the PMNS matrices $V\rightarrow U$ and flipping the operator helicity $a_i(\pm \mathbf{p},\pm)\rightarrow a_i(\pm \mathbf{p},\mp)$, $b_i(\pm \mathbf{p},\pm)\rightarrow b_i(\pm \mathbf{p},\mp)$.
The time evolution of the operators from Eqs.(\ref{eq:operatorarelations}, \ref{eq:operatorbrelations}) are given as,
\begin{gather}
\begin{split}
a_{L\alpha}(\pm\mathbf{p},t) = \sum_{i=1}^3\sum_{\beta=e}^\tau
& \left[V^{}_{\alpha i}V^\ast_{\beta i}\left(\cos E_i(\mathbf{p})t - i\frac{|\mathbf{p}|}{E_i(\mathbf{p})}\sin E_i(\mathbf{p})t \right)a_{L\beta}(\pm\mathbf{p}) \right. \\
& \left. \quad \mp V^{}_{\alpha i}U^\ast_{\beta i}\frac{m_i}{E_i(\mathbf{p})}\sin E_i(\mathbf{p})t \, b^\dagger_{R\beta}(\mp\mathbf{p}) \right],
\end{split} \\
\begin{split}
a^{\dagger}_{L\alpha}(\pm\mathbf{p},t) = \sum_{i=1}^3\sum_{\gamma=e}^\tau
& \left[V^\ast_{\alpha i}V^{}_{\gamma i} \left(\cos E_i(\mathbf{p})t + i\frac{|\mathbf{p}|}{E_i(\mathbf{p})}\sin E_i(\mathbf{p})t \right)a^\dagger_{L\gamma}(\pm\mathbf{p}) \right. \\
& \left. \quad\mp V^\ast_{\alpha i}U^{}_{\gamma i}\frac{m_i}{E_i(\mathbf{p})}\sin E_i(\mathbf{p})t \, b_{R\gamma}(\mp\mathbf{p}) \right],
\end{split}\\
\begin{split}
b_{L\alpha}(\pm\mathbf{p},t) = \sum_{i=1}^3\sum_{\beta=e}^\tau
& \left[V^\ast_{\alpha i}V^{}_{\beta i} \left(\cos E_i(\mathbf{p})t - i\frac{|\mathbf{p}|}{E_i(\mathbf{p})}\sin E_i(\mathbf{p})t \right)b_{L\beta}(\pm\mathbf{p}) \right. \\
& \left. \quad\mp V^\ast_{\alpha i}U^{}_{\beta i}\frac{m_i}{E_i(\mathbf{p})}\sin E_i(\mathbf{p})t \, a^\dagger_{R\beta}(\mp\mathbf{p}) \right],
\end{split}\\
\begin{split}
b^{\dagger}_{L\alpha}(\pm\mathbf{p},t) = \sum_{i=1}^3\sum_{\gamma=e}^\tau
& \left[V^{}_{\alpha i}V^\ast_{\gamma i} \left(\cos E_i(\mathbf{p})t + i\frac{|\mathbf{p}|}{E_i(\mathbf{p})}\sin E_i(\mathbf{p})t \right)b^\dagger_{L\gamma}(\pm\mathbf{p}) \right. \\
& \left. \quad\mp V^{}_{\alpha i}U^\ast_{\gamma i}\frac{m_i}{E_i(\mathbf{p})}\sin E_i(\mathbf{p})t \, a_{R\gamma}(\mp\mathbf{p}) \right],
\end{split}
\end{gather}
where the right-handed operators are just replacements of the PMNS matrices $U\rightarrow V,\quad V\rightarrow U$ and the handedness $a_{(L,R)} \rightarrow a_{(R,L)},\quad b_{(L,R)} \rightarrow b_{(R,L)}$.
We define the lepton family number density, for both the left- and right-handed Dirac neutrinos,
\begin{gather}
l^L_\alpha (t,\mathbf{x}) = \,:\overline{\nu_{L\alpha}}(t,\mathbf{x})\gamma^0 \nu_{L\alpha}(t,\mathbf{x}):\,,\label{Eq:left-hand_Lep}\\
l^R_\alpha (t,\mathbf{x}) = \,:\overline{\nu_{R\alpha}}(t,\mathbf{x})\gamma^0 \nu_{R\alpha}(t,\mathbf{x}): \,. \label{Eq:right-hand_Lep}
\end{gather}
The Fourier expansion of $\nu_{L\alpha}(t,\mathbf{x})$ follows the massless form of Eq.(\ref{eq:leftmasslessexpand});
\begin{equation}
\nu_{L\alpha}(t,\mathbf{x}) = \int\frac{d^3\mathbf{p}}{(2\pi)^3 2|\mathbf{p}|} \left( u_L(\mathbf{p})a_{L\alpha}(\mathbf{p},t)e^{i \mathbf{p}\cdot\mathbf{x}} + v_L(\mathbf{p}) b_{L\alpha}^\dagger(\mathbf{p},t)e^{-i\mathbf{p}\cdot\mathbf{x}} \right).
\label{eq:timeleftmassexpand}
\end{equation}
Then, the evolution of the lepton family number density for the left-handed case Eq.(\ref{Eq:left-hand_Lep}) is obtained by substituting Eq.(\ref{eq:timeleftmassexpand}),
\begin{multline}
l^L_{\alpha}(t,\mathbf{x}) = \int'\frac{d^3 \mathbf{k}}{(2\pi)^3 2|\mathbf{k}|}\int'\frac{d^3 \mathbf{p}}{(2\pi)^3 2|\mathbf{p}|} \\
\times\left[a^\dagger_{L\alpha}(\mathbf{k},t)a_{L\alpha}(\mathbf{p},t)\overline{u_L}(\mathbf{k})\gamma^0u_L(\mathbf{p})e^{-i(\mathbf{k}-\mathbf{p})\cdot\mathbf{x}}
+b_{L\alpha}(\mathbf{k},t)a_{L\alpha}(\mathbf{p},t)\overline{v_L}(\mathbf{k})\gamma^0u_L(\mathbf{p})e^{i(\mathbf{k}+\mathbf{p})\cdot\mathbf{x}} \right.\\
+\left.a^\dagger_{L\alpha}(\mathbf{k},t)b^\dagger_{L\alpha}(\mathbf{p},t)\overline{u_L}(\mathbf{k})\gamma^0v_L(\mathbf{p})e^{-i(\mathbf{k}+\mathbf{p})\cdot\mathbf{x}}
-b^\dagger_{L\alpha}(\mathbf{p},t)b_{L\alpha}(\mathbf{k},t)\overline{v_L}(\mathbf{k})\gamma^0v_L(\mathbf{p})e^{i(\mathbf{k}-\mathbf{p})\cdot\mathbf{x}}\right],
\label{Eq:DiracDensity}
\end{multline}
whereas the right-handed form is a replacement of the spinors and operators from $L$ to $R$.
Next, we write the expectation value using a similar state $|\psi_\sigma(q^0;\sigma_q)\rangle$ as Eq.(\ref{Eq:Wavepacket}), but with the left-handed operator replaced as $a^\dagger_{\sigma}(0,q,0)\rightarrow a^\dagger_{L\sigma}(0,q,0)$.
We sandwich the left-handed density for the lepton family number of Eq.(\ref{Eq:DiracDensity}) with the left-handed state $\vert\psi^L_\sigma(q^0;\sigma_q)\rangle$ and integrate over $\mathbf{k},\mathbf{p}$ to result in,
\begin{multline}
\langle \psi^L_\sigma(q^0;\sigma_q)\lvert l^L_{\alpha}(t,\mathbf{x})\rvert \psi^L_\sigma(q^0;\sigma_q)\rangle =
\frac{1}{\sigma_q(2\pi)^{3/2}\delta(0)^2} \iint' \frac{dq'dq}{(2\pi)^2}
e^{-\frac{(q'-q^0)^2+(q-q^0)^2}{4\sigma_q^2}-i(q'-q)\mathbf{e}_2\cdot\mathbf{x}} \\
\times\left[\sum_{i,j} V^{\ast}_{\alpha i}V^{}_{\sigma i}V^{}_{\alpha j}V^{\ast}_{\sigma j}
\left(\cos{E_i(q')t}+i\frac{|q'|}{E_i(q')}\sin{E_i(q')t}\right)
\left(\cos{E_j(q)t}-i\frac{|q|}{E_j(q)}\sin{E_j(q)t}\right)\right] .
\label{Eq:DiracExpectationValue}
\end{multline}
An initial right-handed state of Eq.(\ref{Eq:DiracExpectationValue}), $\langle \psi^R_\sigma(q^0;\sigma_q)\lvert l^R_{\alpha}(t,\mathbf{x})\rvert\psi^R_\sigma(q^0;\sigma_q)\rangle$, is obtained by interchanging the matrices as $U\rightarrow V,\quad V\rightarrow U$.
Lastly, we preform the integration over $q$ and $q'$ using the approximation of Eq.(\ref{Eq:IntegralApprox}) to write the integrand in Gaussian form.
The result is the linear density expectation value of the left-handed lepton family number,
\begin{equation}
\begin{split}
\lambda^L_{\sigma\rightarrow\alpha}(t,x_2)\simeq & \frac{\sigma_q}{(2\pi)^{1/2}}\sum_{i,j} V^{\ast}_{\alpha i}V^{}_{\sigma i}V^{}_{\alpha j}V^{\ast}_{\sigma j} \\
&\times\frac{1}{2}\left[ (v_{i0}+v_{j0}+1+v_{i0}v_{j0})e^{i(E_i(q^0)-E_j(q^0))t}e^{-\sigma^2_q[(x_2-v_{i0}t)^2+(x_2-v_{j0}t)^2]}\right. \\
& \quad\quad -(v_{i0}+v_{j0}-1-v_{i0}v_{j0})e^{-i(E_i(q^0)-E_j(q^0))t}e^{-\sigma^2_q[(x_2+v_{i0}t)^2+(x_2+v_{j0}t)^2]} \\
& \quad\quad +(v_{i0}-v_{j0}+1-v_{i0}v_{j0})e^{i(E_i(q^0)+E_j(q^0))t}e^{-\sigma^2_q[(x_2-v_{i0}t)^2+(x_2+v_{j0}t)^2]} \\
& \left.\quad\quad -(v_{i0}-v_{j0}-1+v_{i0}v_{j0})e^{-i(E_i(q^0)+E_j(q^0))t}e^{-\sigma^2_q[(x_2+v_{i0}t)^2+(x_2-v_{j0}t)^2]} \right].
\end{split}
\label{Eq:LeftHandedExpectationValue}
\end{equation}
Where we have defined $\lambda^L_{\sigma\rightarrow\alpha}(t,x_2)=\iint dx_1 dx_3 \langle \psi^L_\sigma(q^0;\sigma_q)|l^L_{\alpha}(t,\mathbf{x})|\psi^L_\sigma(q^0;\sigma_q)\rangle$.
The process for solving for the linear density of the right-handed lepton family number of Eq.(\ref{Eq:right-hand_Lep}) is the same.
Thus, we only write the resulting linear density expectation value of the right-handed lepton family number,
\begin{multline}
\lambda^R_{\sigma\rightarrow\alpha}(t,x_2) \simeq \frac{\sigma_q}{\sqrt{2\pi}}\sum_{i,j}\sqrt{1-v_{i0}^2} \sqrt{1-v_{j0}^2} e^{-\sigma_q^2\left((v_{i0}^2+v_{j0}^2)t^2+2 x_2^2\right)} \\
\times \left\{\text{Re}\left(U^{\ast}_{\alpha i}V^{}_{\sigma i}U^{}_{\alpha j}V^{\ast}_{\sigma j} \right)
\left[\cosh\left(2\sigma_q^2(v_{i0}+v_{j0})tx_2\right)\cos\left((E_i(q^0)-E_j(q^0))t\right)\right.\right. \\
\left.-\cosh\left(2\sigma_q^2(v_{i0}-v_{j0}\right)tx_2)\cos\left((E_i(q^0)+E_j(q^0))t\right) \right] \\
-\text{Im}\left(U^{\ast}_{\alpha i}V^{}_{\sigma i}U^{}_{\alpha j}V^{\ast}_{\sigma j} \right)\left[\sinh\left(2\sigma_q^2(v_{i0}+v_{j0})tx_2\right)\sin\left((E_i(q^0)-E_j(q^0))t\right)\right.\\
\left.\left.-\sinh\left(2\sigma_q^2(v_{i0}-v_{j0})tx_2\right)\sin\left((E_i(q^0)+E_j(q^0))t\right) \right] \right\}.
\label{Eq:RightHandedExpectationValue}
\end{multline}
\subsection{Comparison to Wave packet formulations}\label{sec:Decoherence}
The results we present in Eqs.(\ref{Eq:ExpectationValue}, \ref{Eq:LeftHandedExpectationValue}, \ref{Eq:RightHandedExpectationValue}) are based on the evolution of a density operator.
We will now discuss an interpretation of those results based on wave packet formulations.
Specifically, wave packets have been used in connection with neutrino oscillation phenomenology (see Ref. \cite{Beuthe:2001rc} for a review) as several forms; the quantum mechanical (QM) formulation \cite{Giunti:1991ca}, and the external wave packet QFT formulation \cite{Giunti:1993se,Giunti:2002xg}.
We will focus on comparisons to interpretations in the quantum mechanical formulation, because it is the simplest in literature.
Often discussed in literature see \cite{Akhmedov:2019iyt,Kayser:1981ye,Akhmedov:2017xxm,Akhmedov:2009rb}, neutrino oscillations occur when three types of coherence are satisfied.
\begin{enumerate}
\item The different massive neutrino components are coherently produced.
\item Coherent propagation of massive neutrino components.
\item Coherent detection of the different massive components.
\end{enumerate}
We do not consider any production or detection processes, so we can not discuss the possibility of coherent detection in our formulation.
The initial density state of Eq.(\ref{Eq:Wavepacket}) captures the properties of coherent propagation, similar to the QM wave packet.
A prediction of the QM wave packet formulation is the coherence length $L^{\text{coh}}_{i,j}$ for oscillations \cite{Nussinov:1976uw}.
The coherence length is a measure of the group velocities for the mass eigenstates that appears after integration over time.
It acts as a damping factor on the oscillations in space.
We find a similar damping in our density formulation.
The damping appears as a real exponential component that depends quadratically on spacetime,
\begin{gather}
e^{-\sigma^2_q[(x_2\pm v_{i0}t)^2+(x_2\pm v_{j0}t)^2]} \label{Eq:DampingLinearSquared1} \\
e^{-\sigma^2_q[(x_2\pm v_{i0}t)^2+(x_2\mp v_{j0}t)^2]}. \label{Eq:DampingLinearSquared2}
\end{gather}
These real components are a non-linear damping that is applied to the density oscillation.
An important distinction between our density formulation and the QM formulation are the types of damping.
The QM formulation of Eq.(3) in Ref. \cite{Giunti:1991ca} has a single type of damping to coincide with the single type of oscillation
\begin{equation}
e^{-\sigma^2_q[(x_2-v_{i0}t)^2+(x_2-v_{j0}t)^2]},
\end{equation}
which we denote as $(-,-)$.
Whereas, our density formulation has four different types damping, $(-,-),(-,+),(+,-),(+,+)$.
Each damping is applied to a different oscillation in Eqs.(\ref{Eq:ExpectationValue}, \ref{Eq:LeftHandedExpectationValue}, \ref{Eq:RightHandedExpectationValue}).
The $(+,+)$ is the strongest damping factor and $(-,-)$ is the weakest damping, which we illustrate by minimizing the polynomials inside the exponentials assuming $x_2>0$;
\begin{align}
P_1\left(x_2,t=\frac{2x_2}{v_{i0}+v_{j0}}\right) & = 2x_2^2\frac{(v_{i0}-v_{j0})^2}{(v_{i0}+v_{j0})^2} &&\text{from }(-,-),
\label{eq:p1min} \\
P_2\left(x_2,t=\frac{2x_2}{v_{i0}+v_{j0}}\right) & = 2x_2^2\frac{3(v_{i0}+v_{j0})^2+2(v_{i0}^2+v_{j0}^2)}{(v_{i0}+v_{j0})^2} && \text{from }(+,+), \\
P_3\left(x_2,t=\frac{2x_2}{v_{i0}+v_{j0}}\right) & = 2x_2^2+8x_2^2\frac{v_{j0}^2}{(v_{i0}+v_{j0})^2} && \text{from }(-,+), \\
P_4\left(x_2,t=\frac{2x_2}{v_{i0}+v_{j0}}\right) & = 2x_2^2+8x_2^2\frac{v_{i0}^2}{(v_{i0}+v_{j0})^2} && \text{from }(+,-).
\end{align}
Notice the peak value of $(-,-)$ is damped at the peak by $P_1<2x_2^2$ causing it to be weak.
Whereas the peak values of $(+,+)$, $(-,+)$, and $(+,-)$ are damped with $P_{2,3,4}>2x_2^2$ strengthening their suppression of the oscillations.
If we perturb slightly away from the peak time, the strength of the damping factors depend directly on the velocities. For example,
\begin{equation}
P_1\left(x_2,t=\frac{2x_2}{v_{i0}+v_{j0}}+\Delta t\right)-P_1\left(x_2,t=\frac{2x_2}{v_{i0}+v_{j0}}\right) = (v_{i0}^2+v_{j0}^2)(\Delta t)^2+\frac{2x_2(v_{i0}-v_{j0})^2}{v_{i0}+v_{j0}}\Delta t .
\label{eq:p1perturbation}
\end{equation}
So, a decrease in the velocities lead to longer damping times.
After we integrate over all space to arrive at Eq.(\ref{Eq:MajoranaSpaceTimeInt}), the damping factors appear as two types $\pm$.
\begin{equation}
t^{coh}_{i,j}\simeq \frac{\sqrt{2}}{\sigma_q(v_{i0}\pm v_{j0})},
\end{equation}
which we write as a coherence time $t_{i,j}^{coh}$.
In the usual QM case only a single damping factor appears for the oscillations $(v_{i0} + v_{j0})$.
However, new oscillation terms appear in our formulation corresponding to $e^{\pm i[E_i(q^0)+E_j(q^0)]t}$, to which a new damping factor $(v_{i0} - v_{j0})$ is applied.
\section{Comparison of the Dirac and Majorana formulations}\label{sec:DiracVsMajorana}
From the expectation value equations of Eq.(\ref{Eq:ExpectationValue}), Eq.(\ref{Eq:LeftHandedExpectationValue}), and Eq.(\ref{Eq:RightHandedExpectationValue}) we find signatures of the Dirac and Majorana formulations.
Firstly, the terms proportional to the PMNS matrix combination of $V_{\alpha i}^\ast V^{}_{\sigma i}V^{}_{\alpha j} V_{\sigma j}^\ast$ are common between the formulations.
Those terms are the first summation from Eq.(\ref{Eq:ExpectationValue}) of the Majorana formulation and the left-handed Dirac formulation Eq.(\ref{Eq:LeftHandedExpectationValue}).
The difference between the formulations is the secondary term of the Majorana formulation and the right-handed Dirac formulation Eq.(\ref{Eq:RightHandedExpectationValue}).
The secondary term of the Majorana formulation has the PMNS combination $V^{\ast}_{\alpha i}V^{\ast}_{\sigma i}V_{\alpha j}V_{\sigma j}$ that is related to the Majorana phases and is subtracted from the common oscillation terms.
Whereas, the second unitary matrix $U_{\alpha i}$ from the right-handed Dirac formulation is present in Eq.(\ref{Eq:RightHandedExpectationValue}) and is added to Eq.(\ref{Eq:LeftHandedExpectationValue}) as $\lambda^D_{\sigma\rightarrow\alpha}(t,x_2)=\iint dx_1 dx_3 \langle \psi^L_\sigma(q^0;\sigma_q)|l^L_{\alpha}(t,\mathbf{x})+l^R_{\alpha}(t,\mathbf{x})|\psi^L_\sigma(q^0;\sigma_q)\rangle$.
The sign of the different terms, subtraction for Majorana and addition for Dirac, is responsible for total lepton number violation in the Majorana case.
To help emphasize the affect of the sign difference we write the expectation value of the total lepton number for the Dirac formulation case,
\begin{equation}
\sum_{\alpha} \lambda^D_{\sigma\rightarrow\alpha}(t,x_2) \simeq
\frac{\sigma_q}{\sqrt{2\pi}}\sum_i |V_{\alpha i}|^2\left[(1+v_{i0})e^{-2\sigma_q^2(x_2-v_{i0}t)^2}+(1-v_{i0})e^{-2\sigma_q^2(x_2+v_{i0}t)^2}\right],
\label{eq:DiracTL}
\end{equation}
where integration over $x_2$ would result in total lepton number conservation for all times.
Next, from Eq.(\ref{Eq:ExpectationValue}), we also take the summation over the lepton family numbers $\alpha$ as follows,
\begin{equation}
\begin{split}
\sum_{\alpha}\lambda^M_{\sigma\rightarrow\alpha}(t,x_2) \simeq & \frac{\sigma_q}{\sqrt{2\pi}}\sum_i | V_{\sigma i}|^2\left[v_{i0}(1+ v_{i0})e^{- 2\sigma_q^2 (x_2-v_{i 0} t)^2}-v_{i0} (1-v_{i0})e^{- 2\sigma_q^2 (x_2+v_{i 0} t)^2} \right.\\
& \left. \hphantom{XXXXXXXXXX} +2(1-v_{i0}^2) e^{- 2\sigma_q^2 (v_{i 0}^2 t^2 + x_2^2)}\cos 2E_i(q^0)t\right].
\end{split} \label{eq:MajoranaTL}
\end{equation}
In Eq.(\ref{eq:DiracTL}) for the Dirac case, there is no oscillation term and the linear density of the total lepton number is always positive.
For a nonrelativistic Majorana neutrino ($v_{i0} \ll 1$) in Eq.(\ref{eq:MajoranaTL}), the oscillation term can dominate and the sign of the linear density can change with respect to time.
\section{Illustration of nonrelativistic differences}
From inspection of the expectation values Eq.(\ref{Eq:ExpectationValue}), Eq.(\ref{Eq:LeftHandedExpectationValue}), and Eq.(\ref{Eq:RightHandedExpectationValue}); and the discussions in sections \ref{sec:Decoherence} and \ref{sec:DiracVsMajorana} we see two key results.
\begin{itemize}
\item Differences between our formulation and the usual QM formulation are negligible under the ultra-relativistic assumption.
\item The neutrino masses $\frac{m_i m_j}{E_i E_j}=\sqrt{1-v_i^2}\sqrt{1-v_j^2}$ suppress any modifications by the Dirac or Majorana mass formulations.
\end{itemize}
To illustrate those results we choose $m_{\text{lightest}}=0.01\text{ eV} < q^0 = 0.2\text{ eV}$ for the linear density expectations of Eqs.(\ref{Eq:ExpectationValue}) and (\ref{Eq:LeftHandedExpectationValue}).
Even for a momentum an order of magnitude greater than the mass, differences between the Majorana and Dirac cases are not visible (Fig. \ref{Fig:3Flavor_Density_02e-mu} and Fig. \ref{Fig:3Flavor_Density_02e-tau}).
\begin{figure}[ht]
\includegraphics[width=\textwidth]{figure1.pdf
\caption{\label{Fig:3Flavor_Density_02e-mu} Time evolution of the linear density for the expectation value of a lepton family number with Majorana, panel (a), or Dirac, panel (b), mass.
We take an arbitrary distance slice at $x_2=2.5$cm.
The initial momentum density is a Gaussian distribution with a width of $\sigma_q = 0.00001$ and a mean momentum of $q^0=0.2$eV.
Normal mass hierarchy is considered.
Oscillation parameters are the best fit values from the NuFIT 5.0 (2020) collaboration\cite{Esteban:2020cvm}.}
\end{figure}
We used the best fit values of the PMNS mixing angles $\theta_{12},\theta_{23}, \text{and }\theta_{13}$, the CP violating phase $\delta_{\text{CP}}$, and the mass squared differences $\Delta m_{21}^2 \text{ and } \Delta m_{31}^2$ from the work of the NuFIT 5.0 (2020) collaboration\cite{Esteban:2020cvm}.
Then, we are left to choose the absolute mass of the lightest neutrino $m_1 \text{ or } m_3$, the neutrino mass hierarchy, normal $m_1<m_2\ll m_3$ or inverted $m_3\ll m_1<m_2$, the value of the mean momentum $q$, the width of the initial Gaussian density $\sigma_q$, the time $t$, the distance $x_2$, and the Majorana phases $\alpha_{21} \text{ and } \alpha_{31}$.
The definition of the Majorana phases is given as $\alpha_{21}=2\arg(V_{e2})$
and $\alpha_{31}=2\arg(V_{\mu 3})$.
\begin{figure}[ht]
\includegraphics[width=\textwidth]{figure2.pdf
\caption{\label{Fig:3Flavor_Density_02e-tau} Similar to Fig.\ref{Fig:3Flavor_Density_02e-mu}, time evolution of the linear density for the expectation value of a lepton family number at $x_2 = 2.5$cm.
However, we now consider the $\sigma=e,\,\alpha=\tau$ lepton family number values.
The initial momentum density is a Gaussian distribution with a width of $\sigma_q = 0.00001$ and a mean momentum of $q^0=0.2$eV.
Normal mass hierarchy is considered.
Oscillation parameters are the best fit values from the NuFIT 5.0 (2020) collaboration\cite{Esteban:2020cvm}.}
\end{figure}
Any changes caused by different Majorana phases are also hidden, because the differences between the Majorana and Dirac cases are not visible for $m_{\text{lightest}}=0.01\text{ eV} < q^0 = 0.2\text{ eV}$.
The decoherence effects discussed in section \ref{sec:Decoherence} from the real exponentials of Eq.(\ref{Eq:ExpectationValue}), Eq.(\ref{Eq:LeftHandedExpectationValue}), and Eq.(\ref{Eq:RightHandedExpectationValue}) cause the oscillations to be localized to a spacetime region.
In Fig. \ref{Fig:DiracContour02}, we show the 2-D spacetime contour for an electron number linear density for the $\sigma = e \rightarrow \alpha = e$ case.
The region where the electron number linear density has a maximum peak of $\sim 0.4$ propagates from the spacetime origin $(t,x_2)=(0,0)$ toward the upper right corner at a constant velocity.
The initial Gaussian shape is maintained during propagation because we assumed no wave packet spreading in the approximation of Eq.(\ref{Eq:IntegralApprox}).
\begin{figure}[ht]
\includegraphics[width=\textwidth]{figure3.pdf
\caption{\label{Fig:DiracContour02} 2-D Spacetime contour of the linear density for the expectation value of the electron family number with a Dirac mass.
The initial momentum density is a Gaussian distribution with a width of $\sigma_q = 0.00001$ and a mean momentum of $q^0=0.2$eV.
Normal mass hierarchy for neutrinos is considered.
Lepton mixing angles, the Dirac CP phase, and the mass squared differences are the reported best fit values from the work of the NuFIT 5.0 (2020) collaboration\cite{Esteban:2020cvm}.}
\end{figure}
We discussed in section \ref{sec:DiracVsMajorana} how the total lepton number becomes violated in the Majorana case, and we illustrate that by decreasing the momentum to $ q^0 = 0.0002 < m_{lightest}=0.01$eV in Fig.\ref{Fig:3Flavor_Density_00002e-mu}.
This leads to the distinction between the Majorana and Dirac cases, in which the Majorana case has negative expectation values.
Both cases feature the small period oscillations of the inset graph because of lines 4 and 5 from Eq.(\ref{Eq:LeftHandedExpectationValue}).
Those lines are not present in the usual QM formulation, so these small period oscillations can distinguish our formulation.
By comparing Fig.\ref{Fig:3Flavor_Density_02e-mu} and Fig.\ref{Fig:3Flavor_Density_00002e-mu}, the oscillations occur for longer times when the average momentum is smaller.
This was discussed in section \ref{sec:Decoherence}, where Eq.(\ref{eq:p1perturbation}) was given as an example.
To recall, the strength of the damping factors depend directly on the velocities. So, a decrease in the velocities lead to longer damping times.
\begin{figure}[ht]
\includegraphics[width=\textwidth]{figure4.pdf
\caption{\label{Fig:3Flavor_Density_00002e-mu} Time evolution of the linear density for the expectation value of a lepton family number at $x_2 = 2.5$cm.
We now consider the initial momentum density to have a mean momentum of $q^0=0.0002$eV for a Gaussian distribution with a width of $\sigma_q = 0.00001$.
Compared to the preceding figures, differences between Majorana and Dirac are visible because $q^0 = 0.0002 < m_{lightest}=0.01$eV.
Normal mass hierarchy is considered.
Oscillation parameters are the reported best fit values from the work of the NuFIT 5.0 (2020) collaboration\cite{Esteban:2020cvm}.
In addition, we choose the Majorana phases to be arbitrary values of $\alpha_{21}=\pi$ and $\alpha_{31}=0.5\pi$.}
\end{figure}
Next, in the distance plane of Fig.\ref{Fig:3Flavor_Distance} the smaller momenta suppress the peak value for the linear density of the lepton family numbers.
This follows directly from Eq.(\ref{eq:p1min}), where for $q^0=0.2$ eV the polynomial is larger than for $q^0=0.0002$ eV.
Thus, the propagation coherence occurs at smaller distances, which is sometimes stated as a smaller coherence length.
\begin{figure}[ht]
\includegraphics[width=\textwidth]{figure5.pdf
\caption{\label{Fig:3Flavor_Distance} Distance evolution of the linear density for the expectation value of a lepton family number.
We have taken different time slices corresponding to when the peak of the linear densities are near $x_2=5.0$cm.
Normal mass hierarchy is considered, and we choose the Majorana phases to be arbitrary values of $\alpha_{21}=\pi$ and $\alpha_{31}=0.5\pi$.
Oscillation parameters are the reported best fit values from the work of the NuFIT 5.0 (2020) collaboration\cite{Esteban:2020cvm}}
\end{figure}
Even with this in mind, at smaller momenta the Dirac and Majorana cases are distinguishable. Similar to Fig.\ref{Fig:3Flavor_Density_00002e-mu}, the Majorana case has negative expectation values and the Dirac case is always positive.
\section{Conclusion}
We have proposed a novel formulation for neutrino oscillations from a QFT point of view.
This formulation can be applied to both relativistic and nonrelativistic energies for neutrinos.
The previous formulation was limited to Majorana neutrinos with a fixed momentum.
We have extended the previous formulation to the Dirac neutrino case and have applied to the case with a momentum distribution.
To study the oscillation of neutrino with a momentum distribution, we have considered the linear lepton number density and the initial state with a Gaussian distribution along a one spacial direction.
After taking the expectation value of the density with the state, we find a result similar to the quantum mechanical wave packet approach used to describe neutrino flavor oscillations.
Our result has additional terms that result in interesting behavior.
In particular, we have studied type of decoherence ascribed to wave packet separation.
We have found how the peak value of lepton number density is suppressed as it propagates.
Even with that decoherence our formulation can distinguish between neutrino mass hierarchy and mass type at nonrelativistic energies.
The mass hierarchies can be distinguished from the short period of the oscillation.
The period is determined by the sum of the rest mass of neutrinos $m_i +m_j$ for nonrelativistic case.
As for mass types, the expectation value of the lepton number density for Majorana neutrinos can take both positive and negative value
while for Dirac neutrino, it does not change the sign.
\begin{acknowledgments}
The work of T.M. is supported by Japan Society for the Promotion of Science (JSPS) KAKENHI Grant Number JP17K05418.
We would like to thank Uma Sankar and Koichi Hamaguchi for useful comments and suggestions during the initial presentations of this work.
\end{acknowledgments}
|
\section{Introduction} \label{sec:introduction}
When we visualize directed graphs (digraphs for short) that model hierarchical relations
with node-link diagrams,
we traditionally turn edge directions into geometric directions by letting each edge point upward.
Aiming for visual clarity, we would like such an upward drawing to be planar,
that is, no two edges should cross~\cite{DETT99}.
If this is possible, the resulting drawing is called an \emph{upward planar drawing}; see \cref{fig:introExample}~(a).
Interestingly, as Di Battista and Tamassia~\cite{DT88} have shown,
every upward planar drawing can be turned into one where each edge is drawn with a single line segment;
such a \emph{straight-line drawing} may however require an exponentially large drawing area.
\begin{figure}[tb]
\centering
\includegraphics{introExamples}
\caption{(a) An upward planar drawing; (b) an orthogonal drawing; (c) an upward planar 2-slope drawing.}
\label{fig:introExample}
\end{figure}
Another important class of drawings are \emph{(planar) orthogonal drawings},
where edges are drawn as sequences of horizontal and vertical line segments~\cite{DETT99}; see \cref{fig:introExample}~(b).
This drawing style is commonly used for schematic drawings such as VLSI circuit design and UML diagrams.
Schematic drawings that allow more than two slopes for edge segments include hexalinear and octilinear drawings,
which find application in metro maps~\cite{NW11}.
In general, the use of only few geometric primitives (such as different slopes)
in a graph drawing facilitates a low visual complexity; a common quality measure for drawings~\cite{Sch15}.
In recent years, the interest in upward planar drawings that use only few different slopes has grown.
For example, among other results, Bekos {et~al.}~\cite{BDDLM18} showed
that every so-called bitonic $st$-graph with maximum degree~$\Delta$
admits an upward planar drawing where every edge has at most one bend
and the edges segments use only~$\Delta$ distinct slopes.
Di Giacomo {et~al.}~\cite{DLM20} provided complementary results
by proving that also every series-parallel digraph admits a 1-bend upward planar drawing
on~$\Delta$ distinct slopes; their drawings also have optimal angular resolution.
Brückner {et~al.}~\cite{BKM19} considered level-planar drawings with a fixed slope set,
that is, upward planar drawings where each vertex is drawn on a predefined integer y-coordinate (its level).
In this paper, we continue this recent trend.
In particular, we study bendless upward planar drawings that use only two different slopes.
An example of such a drawing is shown in \cref{fig:introExample}~(c).
Some of our results also imply results for 1-bend upward planar drawings.
We now define these drawing concepts more precisely and list related work.
\paragraph{Upward planarity.}
An \emph{upward planar drawing} of a directed graph~$G$ is a planar drawing of $G$
where every edge $(u, v)$ is drawn as a monotonic upward curve from $u$ to $v$.
A digraph is called \emph{upward planar} if it admits an upward planar drawing.
Two upward planar drawings of the same digraph are \emph{topologically equivalent} if the left-to-right orderings
of the incoming and outgoing edges around each vertex coincide in the two drawings.
An \emph{upward planar embedding} is an equivalence class of upward planar drawings.
An upward planar digraph is called \emph{upward plane} if it is equipped with an upward planar embedding.
A necessary though not sufficient condition for upward planarity is acyclicity~\cite{BDLM94}.
Moreover, Garg and Tamassia~\cite{GT01} showed that testing upward planarity is NP-complete for general digraphs.
While the digraphs used in their reduction contain vertices with in- and outdegree higher than two,
such vertices can be split into mulitple vertices of maximum in- and outdegree at most two
without losing any of the properties required in their proofs.
It is thus also NP-complete to test upward planarity for digraphs
with in- and outdegree at most two.
On the positive side, there exist several fixed-parameter tractable algorithms for general digraphs~\cite{Cha04,HL06,DGL10}
and polynomial time algorithms for single source digraphs~\cite{BDMT98}, series-parallel digraphs~\cite{DGL10},
outerplanar digraphs~\cite{Pap95}, and triconnected digraphs~\cite{BDLM94}.
Moreover, upward planarity can be decided in polynomial time if the embedding is specified~\cite{BDLM94}.
\paragraph{$\ell$-bend $k$-slope drawings.}
In an \emph{$\ell$-bend drawing} of a graph $G$ each edge is drawn with at most $\ell+1$ line segments;
equivalently, each edge has at most $\ell$ bends.
An \emph{$\ell$-bend $k$-slope drawing} of $G$ is an $\ell$-bend drawing of $G$
where every edge segment has one of at most $k$ distinct slopes.
From now on and if not further specified,
we refer to bendless (or 0-bend) $k$-slope drawings simply with \emph{$k$-slope drawings}.
Note that orthogonal drawings are $2$-slope drawings without a bound on the number of bends.
Tamassia~\cite{Tam87} showed that a planar orthogonal drawing with minimum total number of bends
of a plane graph on $n$ vertices can be computed in~$\ensuremath{\mathcal{O}}(n^2 \log n)$ time.
Rhaman {et~al.}~\cite{RNN03} gave necessary and sufficient conditions for a subcubic plane graph
to admit a bendless orthogonal drawing.
For drawings in 3D, Eppstein~\cite{Epp13} considered bendless orthogonal graph drawings where
two vertices are adjacent if and only if two of their coordinates are equal.
Given a graph $G$, the minimum number of slopes needed for $G$
to admit a $k$-slope drawing is called the \emph{slope number} of $G$~\cite{WC94}.
In the planar setting, this is the \emph{planar slope number} of $G$.
Both these numbers have been studied extensively.
For example, Pach and Pálvölgyi~\cite{PP06} showed that the slope number of graphs with maximum degree 5
can be arbitrarily large. Further results, include bounds on slopes numbers of graph classes
such as trees, 2-trees, planar 3-trees, outerplanar graphs~\cite{DESW07,LLMN13,JJKLTV13,KMW14},
cubic graphs~\cite{MP12}, and subcubic graphs~\cite{DLM18,KMSS18}.
Determining the planar slope number is hard in the existential theory of the reals~\cite{Hof17}.
\paragraph{Upward planar 2-slope drawings.}
The focus of this paper lies on (bendless) upward planar $2$-slope drawings.
We consider only the slope set $\set{-\pi/4,\pi/4}$ and denote it by $\set{\sL, \ensuremath{\nearrow}}$,
since an upward planar 2-slope drawing on any two slopes can be morphed into an upward planar 2-slope drawing
with the slopes $\sL$ and $\ensuremath{\nearrow}$ -- imagine this as (un)skewing a partial grid; see \cref{fig:skewing}.
Note that a natural lower bound on the upward planar slope number of a graph is given by its maximum in- and outdegree.
Hence, we assume that the graphs considered in this paper have maximum in- and outdegree at most two.
\begin{figure}[htb]
\centering
\includegraphics{skewing}
\caption{Two upward planar 2-slope drawings of the same graph on different slope sets -- edge directions are given implicitly.
Using an affine transformation one can transform a drawing on any size-two slope set into one on $\set{\sL, \ensuremath{\nearrow}}$.}
\label{fig:skewing}
\end{figure}
Bachmaier {et~al.}~\cite{BBBHMU09}, Brunner and Matzeder~\cite{BM11}, and Bachmaier and Matzeder~\cite{BM13}
studied straight-line drawings of ordered and unordered rooted trees on $k$-grids for $k \in \set{4, 6, 8}$.
Some of their drawing styles are also upward planar.
A classical results of Crescenzi {et~al.}~\cite{CDP92} shows
that any binary tree with $n$ vertices admits an upward planar 2-slope drawing in $\ensuremath{\mathcal{O}}(n \log n)$ area.
Concerning more complex graphs,
upward planar drawings with few slopes for lattices have been studied by Czyzowicz {et~al.}~\cite{CPR90} and Czyzowicz~\cite{Czy91}.
As mentioned above, Bekos {et~al.}~\cite{BDDLM18} and Di Giacomo {et~al.}~\cite{DLM20}
considered such drawings for $st$-graph and series-parallel graphs
but also allowed bends.
In a companion paper to the current one,
Klawitter and Zink~\cite{KZ21} study upward planar $k$-slope drawings for $k \geq 3$
and among other results show that it is NP-hard to decide whether an outerplanar digraph
admits an upward planar 3-slope drawing.
\paragraph{Phylogenetic networks.}
Our interest in upward planar 2-slope drawings
also stems from the problem of visualizing phylogenetic networks.
\emph{Phylogenetic trees} and \emph{networks} are used to model
the evolutionary history of a set of taxa like species, genes, or languages~\cite{HRS10,Dun14,Ste16}.
The precise definition of phylogenetic networks and their drawing conventions may vary widely depending on the particular use case.
For instance, vertices may have timestamps that should be represented in the drawings or
leaves may be required to be placed on the same height.
In combinatorial phylogenetics the following definition is commonly used~\cite{HRS10}:
A \emph{phylogenetic network} is a rooted digraph where the leaves are labelled bijectively with a set of taxa.
Inner vertices are either \emph{tree vertices} that have indegree one and outdegree two
or \emph{reticulations} that have indegree two and outdegree one; see \cref{fig:phynets}~(a).
A network without reticulations is a \emph{phylogenetic tree}; see \cref{fig:phynets}~(b).
\begin{figure}[htb]
\centering
\includegraphics{phynets}
\caption{(a) A phylogenetic network with two reticulations;
(b) a phylogenetic tree drawn as rectangular cladogram and (c) upward planar with two slopes.}
\label{fig:phynets}
\end{figure}
There exist different drawing styles for phylogenetic trees such as rectangular or circular cladograms~\cite{BBS05,Hus09}.
If the focus is on the topology of the tree (and thus the taxonomy),
a common drawing style is upward planar with 2-slope and all leaves aligned on a line.
Theoretical work to adapt classical drawing styles from phylogenetic trees to phylogenetic networks
has been carried out by Huson {et~al.}~\cite{KH08,Hus09,HRS10}.
A different approach has been taken by Tollis and Kakoulis~\cite{TK16},
who propose a drawing style similar to treemaps for a special class of phylogenetic networks.
There also exist several software tools to draw phylogenetic networks~\cite{HB05,HS12,BDM12,Vau17,SVBDAS21}.
Here we are interested in drawing upward planar phylogenetic networks with two slopes and the additional constraint
that all leaves lie on a horizontal line; see \cref{fig:phynets}~(c).
\paragraph{Contribution.}
In this paper we investigate the following decision problem.
Given a digraph $G$ of in- and outdegree at most two, decide whether it admits an upward planar $2$-slope drawing.
We distinguish the fixed and variable embedding scenario, that is, whether $G$ is already equipped with an upward planar embedding or not.
In the former case,
we give a simple characterisation of when a drawing exists; see \cref{sec:plane}.
By making use of orthogonal drawing algorithms, we also show how to construct a drawing (if it exists) in linear time.
In addition, if no upward planar 2-slope drawing exists, we describe how to obtain an upward planar 1-bend 2-slope drawing of $G$
with minimum number of bends.
For the variable embedding scenario, we check whether graphs of different graph classes admit
an upward planar 2-slope drawing, based on the results of see \cref{sec:plane}.
In \cref{sec:ss}, we show that for a single-source digraph $G$, checking whether an upward planar 2-slope drawing of $G$ exists
can be done based on a single upward planar embedding of $G$.
In the affirmative, a suitable upward planar embedding can be derived and a drawing constructed in linear time.
For series-parallel digraphs (\cref{sec:sp}) and general digraphs (\cref{sec:digraph,sec:nonbi}),
we derive a quartic-time and a fixed-parameter tractable algorithm, respectively.
These algorithms are based on Didimo et al.'s algorithms for upward planarity testing~\cite{DGL10}.
Lastly, we show how to compute 2-slope drawings of upward planar phylogenetic networks,
where all leaves lie on a horizontal line in linear time; see \cref{sec:phynet}.
We conclude with a short discussion and open problems.
\section{Preliminaries} \label{sec:prelim}
Let $G$ be an upward plane digraph with maximum in- and outdegree two.
We assume, without loss of generality, that $G$ is connected and
let $n$ denote the number of vertices of $G$ or the graph currently under consideration.
If a vertex $v$ of $G$ has two incoming edges,
then based on the left-to-right ordering of the edges around $v$,
it is natural to talk about the \emph{left} and the \emph{right} incoming edge of $v$.
If an edge $e$ is the only incoming edge at $v$, we call $e$ the \emph{sole} incoming edge of $v$.
The same holds for outgoing edges; see \cref{fig:angles}~(a).
A \emph{2-slope assignment} $\phi$ is a mapping from the edges of $G$ to the slopes $\ensuremath{\nearrow}$ and $\sL$,
i.e.~$\phi : E(G) \ensuremath{\rightarrow} \set{\sL, \ensuremath{\nearrow}}$.
We say $\phi$ is a \emph{consistent 2-slope assignment} if
\begin{itemize}
\item every left (right) incoming edge of a vertex is assigned the slope $\ensuremath{\nearrow}$ ($\sL$), and
\item every left (right) outgoing edge of a vertex is assigned the slope $\sL$ ($\ensuremath{\nearrow}$).
\end{itemize}
An edge $(u, v)$ that is the sole outgoing edge of $u$ and the sole incoming edge of $v$ may have either slope.
A digraph $G$ together with a consistent 2-slope assignment $\phi$ forms an \emph{upward planar 2-slope representation} $U_G = (G, \phi)$.
To ease reading, we simply write 2-slope representation.
Suppose $G$ contains an edge $e = (u, v)$ that is the left outgoing edge of $u$ and left incoming edge of $v$.
Let $f$ be the face to the right of $e$; see \cref{fig:angles}~(b).
We then call~$e$ a \emph{bad edge} with respect to $f$, since $e$ obstructs a consistent $2$-slope assignment.
The same holds with ``left'' and ``right'' in reversed roles.
\begin{figure}[htb]
\centering
\includegraphics{anglesSwitches}
\caption{(a) The edge $(u, v)$ is the right outgoing edge of $u$ and left incoming edge of $v$, the edge $(v, w)$ is the sole outgoing edge of $v$;
(b) the edge $(u, v)$ is a bad edge with respect to $f$;
(c) $u$ is a left-switch spanning a small angle, $v$ spans a flat angle and is thus not a switch, and $w$ is a sink-switch spanning a large angle. }
\label{fig:angles}
\end{figure}
Let $U_G$ be a $2$-slope representation of $G$.
Let $a = (e_1, v, e_2)$ be a triplet such that $v$ is a vertex of the boundary
of a face $f$ and $e_1$, $e_2$ are incident edges of $v$ that are consecutive on the boundary of $f$
in counterclockwise direction.
The triplet $a$ is called an \emph{angle} of~$f$. We can categorise angles into three groups,
namely, $a$ is
\begin{itemize}
\item a \emph{large angle} if $e_1$ and $e_2$ span a 270$^{\circ}$ angle in $f$,
\item a \emph{small angle} if $e_1$ and $e_2$ span a 90$^{\circ}$ angle in $f$, or
\item a \emph{flat angle} if $e_1$ and $e_2$ span a 180$^{\circ}$ angle in $f$
\end{itemize}
with respect to the slopes assigned to~$e_1$ and~$e_2$.
A vertex~$v$ of~$G$ is called a \emph{local source} with respect to a face~$f$
if~$v$ has two outgoing edges on the boundary of~$f$. A \emph{local sink} is defined analogously.
Furthermore, we call~$v$ a \emph{switch} with respect to~$f$ if the slopes of~$e_1$ and~$e_2$ differ;
for example, every local source is a switch.
We further categorise switches by the angle they span and where they lie on the boundary of~$f$; see \cref{fig:angles}~(c).
A switch~$v$ is a \emph{large switch} if~$e_1$ and~$e_2$ span a large angle at~$f$ and a \emph{small switch} otherwise;
note that there can be no ``flat'' switch.
We call~$v$ a \emph{source-switch} or \emph{sink-switch} if~$v$ is a local sink or local source, respectively.
Otherwise, if~$e_1$ and~$e_2$ have~$f$ to the right~(left), then~$v$ is a \emph{left-switch} (resp.\ \emph{right-switch}).
Note that an inner face~$f$ of~$G$ contains exactly four small switches more than large switches and
that the outer face contains four large switches more than small switches.
An inner (the outer) face~$f$ is \emph{rectangular} if it contains exactly four small (large) switches.
Assume for now that~$G$ is biconnected. The following definitions are illustrated in \cref{fig:splitComponentFlip}.
A \emph{split pair}~$\set{u, v}$ of~$G$ is either a separation pair or a pair of adjacent vertices.
A \emph{split component} of~$G$ with respect to the split pair~$\set{u, v}$ is either an edge~$(u, v)$ (or~$(v, u)$)
or a maximal subgraph~$G'$ of~$G$ such that~$G'$ contains~$u$ and~$v$ and~$\set{u, v}$ is not a split pair of~$G'$.
Let~$G'$ be such a split component with respect to the split pair~$\set{u, v}$.
If~$G'$ is equipped with an upward planar embedding, then we define the \emph{flip} of~$G'$
as the change of the embedding of~$G'$ by reversing the edge ordering of every vertex of~$G'$.
\begin{figure}[htb]
\centering
\includegraphics{splitComponentFlip}
\caption{(a) A biconnected digraph with split pair $\set{u, v}$, (b) which induces three split components; (c) flip of the third split component.}
\label{fig:splitComponentFlip}
\end{figure}
\section{Fixed embedding} \label{sec:plane}
In this section we consider the problem of whether an upward plane digraph admits an upward planar 2-slope drawing.
As noted above, a bad edge obstructs the existence of a consistent 2-slope assignment and thus of a 2-slope representation for $G$.
We show that the absence of any bad edges is not only necessary but also sufficient.
Since upward planar 2-slope drawings are related to orthogonal drawings, we can make use of techniques used to construct them.
The classical algorithm by Tamassia~\cite{Tam87}, which constructs an orthogonal drawing of a plane graph with minimum number of bends,
works in three steps;
refer also to Di Battista {et~al.}~\cite[Chapter 5]{DETT99}.
It starts with a plane graph and constructs a so-called \emph{orthogonal representation}, a description of the shapes of the faces.
The second step, called \emph{refinement}, subdivides each face into rectangles.
Finally, the third step performs a so-called \emph{compaction} -- it assigns coordinates to the vertices with the goal to minimize the area of the drawing.
The technique for construction an orthogonal representation is not suitable for the construction of an upward planar 2-slope drawing,
as it does not preserve the upwardness of edges.
However, with a 2-slope representation already given, we can adopt the refinement algorithm by Tamassia
and the compaction algorithm by Di Battista~{et~al.}~for our purposes.
In the following lemma we sketch such a modified version of Tamassia's algorithm that refines the faces of a 2-slope representation.
For this, recall that a switch is a triplet $(e_1, v, e_2)$ consisting of a vertex and its two incident edges along a face in counterclockwise order
where $e_1$ and $e_2$ have different slopes.
\begin{lemma} \label{clm:plane:refinement}
Let $G$ be an upward plane digraph on $n$ vertices with 2-slope representation $U_G = (G, \phi)$.
Then, in $\ensuremath{\mathcal{O}}(n)$ time, $G$ can be refined into a digraph $\bar{G}$ that contains only rectangular faces
and such that $G$ is a topological minor of $\bar{G}$ respecting $\phi$.
\end{lemma}
\begin{proof}
If every face of $G$ is already rectangular, then we are done and $\bar{G} = G$.
Assume this is not the case and let $f$ be non-rectangular inner face of $G$.
We describe how to refine $f$ into rectangular faces.
First, traverse the boundary of $f$ counterclockwise
and store in each switch pointers to its preceding and subsequent switch.
Next, starting at any switch, traverse the circular sequence of switches in counterclockwise order.
Let $u$ be the first encountered large switch that is preceding a small switch $v$.
Note that such $u$ exists since $f$ is not rectangular but contains at least four small switches.
Without loss of generality, assume $u$ is a sink-switch. (The cases when $u$ is a source-, left-, or right-switch work analogously.)
Let $w$ be the subsequent switch of $v$.
If $w$ is a large switch, then add a vertex $x$ and the edges $(u, x)$ and $(w, x)$ with slopes $\ensuremath{\nearrow}$ and $\sL$, respectively;
see \cref{fig:augmentationStep} (a).
Otherwise, if $w$ is a small switch (and thus a right-switch), then subdivide the outgoing edge of $w$ with a new vertex $x$,
and add the edge $(u, x)$ with slope $\ensuremath{\nearrow}$. Assign the slope $\sL$ to the two edges resulting from the subdivision.
This ensures that $\phi$ is respected by $G$ as topological minor of $\bar{G}$; see \cref{fig:augmentationStep} (b).
In either of the two case, the result is a rectangular face $f_1$
and a face $f_2$ with one less small and one less large switch than $f$.
Let $f_2$ now take the role of $f$ and store the preceding switch of $u$ and the subsequent switch of $w$
as preceding and subsequent switches of $x$, respectively.
If $x$ is a small switch, continue the traversal with the switch preceding $x$ instead of with $x$.
Therefore, if a large switch precedes $u$ in $f$, the process directly continues with a refinement step
without having to potentially traverse the whole circular sequence first.
Stop when only four switches are left and when $f$ is thus rectangular.
Note that this process runs in linear time in terms of the size of $f$ and that it only adds as
many new vertices as $f$ contains large switches.
\begin{figure}[htb]
\centering
\includegraphics{augmentationStepWithExample}
\caption{How to refine non-rectangular faces of $G$ to obtain $\bar{G}$
when a large switch is followed (a) by a small and a large switch or (b) by two small switches.
On the right, the large switch $u_4$ precedes the large switch $u_3$ and is thus processed after $u_3$.}
\label{fig:augmentationStep}
\end{figure}
Repeat this step for every non-rectangular face of $G$ including the outer face.
Let $\bar{G}$ be the resulting digraph, which by construction has $G$ as topological minor and a size in $\ensuremath{\mathcal{O}}(n)$.
Furthermore, since the boundary of every face of $G$ was traversed only twice,
this refinement algorithm runs in $\ensuremath{\mathcal{O}}(n)$ time.
\end{proof}
We can now prove the main theorem of this section.
\begin{theorem} \label{clm:plane}
Let $G$ be an upward plane digraph with $n$ vertices. Then the following statements are equivalent.
\begin{enumerate}[leftmargin=*,label=(F\arabic*)]
\item \label{case:f1} $G$ admits an upward planar $2$-slope drawing.
\item \label{case:f2} $G$ admits a $2$-slope representation $U_G$.
\item \label{case:f3} $G$ contains no bad edge.
\end{enumerate}
Moreover, there exists an $\ensuremath{\mathcal{O}}(n)$-time algorithm that tests if $G$ satisfies \ref{case:f3}
and constructs an upward planar 2-slope drawing of $G$ in the affirmative case.
\end{theorem}
\begin{proof}
Note that \ref{case:f1} implies \ref{case:f2} and \ref{case:f2} implies \ref{case:f3}.
We first show that \ref{case:f3} implies \ref{case:f2} and then how to construct a drawing \ref{case:f1} from \ref{case:f2}.
Whether $G$ contains a bad edge can easily be checked in $\ensuremath{\mathcal{O}}(n)$ time.
Suppose it does not and thus satisfies \ref{case:f3}.
Construct a consistent 2-slope assignment for $G$ as follows.
Go through all edges of $G$ (in any order).
For an edge $e$, if it is a left (right) incoming edge, assign it $\ensuremath{\nearrow}$ ($\sL$).
Otherwise, if it is a left (right) outgoing edge, assign it $\sL$ ($\ensuremath{\nearrow}$).
Since $e$ is not a bad edge, there is no conflict.
If $e$ is both a sole incoming and a sole outgoing edge, assign it either slope, say $\sL$.
Together with the already given upward planar embedding of $G$,
this slope assignment yields a 2-slope representation $U_G$ of $G$.
Next, we construct an upward planar 2-slope drawing of $G$.
Use \cref{clm:plane:refinement} to obtain a $2$-slope representation $U_{\bar G}$ of an upward planar digraph $\bar G$ in which every face is rectangular
in $\ensuremath{\mathcal{O}}(n)$ time.
Ignoring the slope assignment in $U_{\bar G}$ we basically have a orthogonal representation of a graph where every face is rectangular.
Therefore we can apply the linear-time compaction algorithm by Di Battista {et~al.}~\cite[Theorem~5.3]{DETT99}.
This algorithm assigns edge lengths and computes coordinates
while handling the vertical and orthogonal direction of a orthogonal representation independently.
Hence, applying the algorithm to $U_{\bar G}$, the edges with slopes $\ensuremath{\nearrow}$ and $\sL$ are handled independently and keep their slopes.
As a result, we get an upward planar 2-slope drawing of $\bar G$, which we can reduce to a drawing of $G$.
Since the three steps run in $\ensuremath{\mathcal{O}}(n)$ time each, the claim on the running time follows.
\end{proof}
Suppose we have an upward plane digraph $G$ with $k$ bad edges.
Since $G$ admits no upward planar 2-slope drawing, it is natural to ask whether $G$ admits an upward planar 1-bend 2-slope drawing.
In particular, if such a drawing exist, is it enough to bend only the $k$ bad edges?
Using \cref{clm:plane} we can answer this question affirmatively.
\begin{corollary} \label{clm:plane:bends}
Let $G$ be an upward plane digraph with $n$ vertices, maximum in- and outdegree at most two, and with $k$ bad edges.\\
Then $G$ admits an upward planar 1-bend 2-slope drawing with $k$ bends.
Moreover, without changing the embedding, this is the minimum number of bends that can be achieved.
\end{corollary}
\begin{proof}
Subdivide every bad edge once to obtain a graph $G'$.
Then apply \cref{clm:plane} to obtain an upward planar 2-slope drawing of $G'$.
Since a bad edge $e$ of $G$ is neither a sole incoming nor a sole outgoing edge,
the two edges obtained from $e$ in $G'$ have different slopes.
Hence, by turning every subdivision vertex into a bend we get an upward planar 1-bend 2-slope drawing of $G$.
Since even a single bad edge obstructs a 2-slope representation and since bending a non-bad edge clearly does not help for other bad edges either,
it follows that $k$ bends are also necessary.
\end{proof}
Note that $G$ may admit an upward planar 1-bend 2-slope drawing with less or no bends if the embedding is changed.
\section{Variable embedding} \label{sec:variable}
In this section we consider the problem of whether a given upward planar digraph $G$ of a particular graph class
admits an upward planar $2$-slope drawing under any upward planar embedding.
Before we start with the class of single-source digraphs, we make two general observations.
Let $G$ be an upward planar digraph with maximum in- and outdegree at most two. Suppose $G$ contains a leaf $\ell$.
Note that removing $\ell$ from $G$ does not change whether $G$ admits an upward planar 2-slope drawing.
Moreover, we may reduce $G$ to a digraph without leaves, obtain a drawing of the reduced digraph (if possible),
and then add the leaves to obtain a $2$-slope drawing of $G$.
Removing or later restoring leaves takes only linear time.
While leaves are no obstructions, transitive edges are.
\begin{observation} \label{clm:planar:transitiveEdge}
A transitive edge of an upward planar digraph $G$ is a bad edge in any upward planar embedding of $G$.
\end{observation}
\begin{proof}
Let $e = (u, v)$ be a transitive edge of $G$.
By definition, there is a directed path~$P$ from $u$ to $v$ different from $e$.
Since $P$ may not cross $e$, it enters $v$ from the same side (left or right) as it leaves $u$ in any upward planar embedding of $G$
and hence $e$ is always a bad edge.
\end{proof}
\subsection{Single-source digraphs} \label{sec:ss}
For a single-source digraph $G$, our idea is to first compute an arbitrary upward planar embedding of $G$
with the linear-time algorithm of Bertolazzi {et~al.}~\cite{BDMT98}.
We then check whether there are any bad edges and, if so, whether they can be fixed with small changes to the embedding.
For a characterisation of when such fixes are possible, we need the following lemmata.
\begin{lemma} \label{clm:singleSource:outerFace}
An upward planar single-source digraph $G$ contains no bad edge with respect to the outer face.
\end{lemma}
\begin{proof}
Consider an upward planar embedding of $G$
and suppose $e = (u, v)$ is a bad edge with respect to the outer face $f_0$.
Then $v$ is a sink of $f_0$ that would span a small angle in a 2-slope representation.
This implies that there are two sources for $f_0$ that would span large angles, i.e. that are sources of $G$.
This is a contradiction to $G$ being a single-source digraph.
\end{proof}
\begin{lemma} \label{clm:singleSource:twoBadEdges}
Let $G$ be an upward planar single-source digraph with maximum in- and outdegree at most two.
Then in any upward planar embedding of $G$, there are at most two bad edges with respect to the same face.
\end{lemma}
\begin{proof}
Let $G$ be an upward plane single-source digraph and let $f$ be a face of $G$.
Note that a bad edge is always incident to a sink of a face that would span a small angle in a 2-slope representation.
If there are three or more bad edges with respect to $f$,
then $f$ contains at least two such sinks (like $v_1$ and $v_2$ in \cref{fig:singleSource:twoBadEdges} (a)).
There is then at least one source $u$ for $f$ that would span a large angle.
However, $u$ is also a source of $G$ and since it is not the source for the outer face, it is not the only source of $G$.
This is a contradiction to $G$ being a single-source digraph.
\Cref{fig:singleSource:twoBadEdges} (b) gives an example of a face with two bad edges.
\end{proof}
\begin{figure}[htb]
\centering
\includegraphics{singleSourceTwoBadEdges}
\caption{(a) A face $f$ that has four bad edges, two adjacent to the sinks $v_1$ and $v_2$ of the face each,
but that also contains a source $u$;
(b) a face $f$ with two bad edges $(u_1, v_1)$ and $(u_2, v_1)$ that can be a face of single source digraph.}
\label{fig:singleSource:twoBadEdges}
\end{figure}
We can now characterise when an upward planar single-source digraph admits an upward planar 2-slope drawing.
\begin{theorem} \label{clm:singleSource:characterisation}
Let $G$ be an upward plane single-source digraph with maximum in- and outdegree at most two.
Then $G$ admits an upward planar 2-slope drawing if and only if,
for every bad edge $(u, v)$ in the current upward planar embedding of $G$,
there is a split pair $\set{w, u}$ such that $w$ precedes $u$ and
where exactly the two outgoing edges of $w$ are in the split component of $\set{w, u}$ that does not contain $v$.
\end{theorem}
\begin{proof}
If $G$ contains no bad edge with its current embedding, then \cref{clm:plane} gives us that $G$ admits an upward planar 2-slope drawing.
Otherwise, let $e = (u, v)$ be a bad edge with respect to a face $f$.
By \cref{clm:singleSource:outerFace}, $f$ is not the outer face and thus $e$ not a bridge.
Let $e_u$ be the second outgoing edge of $u$ and $e_v$ the second incoming edge of~$v$.
Note that $e_v$ and $e_u$ are also on the boundary of $f$.
Let $f'$ be the second face with $e$ on its boundary.
Without loss of generality, assume $f'$ is to the left and $f$ to the right of $e$.
Suppose there exists a vertex $w$ that precedes $u$ in $G$ and that forms a split pair with~$u$.
If there are multiple choices for $w$, pick the one closest to $u$.
Note that by this choice, $w$ cannot be a source.
Let $G'$ be the split component of $G$ with respect to $\set{w, u}$ that does not contain $v$.
This is not a single edge since $w$ precedes $u$.
Furthermore, by picking $w$ closest to $u$, we know that at least two incident edges of~$w$ are in $G'$.
Suppose that only the two outgoing edges of $w$ lie inside $G'$.
We now flip $G'$ as shown in \cref{fig:singleSource} (a), that is,
we reverse the order of incoming and outgoing edges for any vertex in $G'$.
Note that we only reverse the order for the outgoing edges of $w$.
As a result, we get that $e$ is no longer a bad edge.
The flip cannot introduce a new bade edge, since $w$ is not a sink or source of $f$ or $f'$.
\begin{figure}[htb]
\centering
\includegraphics{singleSourceCases}
\caption{Three scenarios for a bad edge $e$ with respect to $f$ in a single-source digraph:
(a) the split component $G'$ with respect to the split pair $\set{w, u}$ can be flipped such that $e$ is no longer a bad edge;
(b) $G'$ cannot be flipped; and
(c) there is no split pair $\set{w, u}$.}
\label{fig:singleSource}
\end{figure}
Suppose that an incoming edge of $w$ lies inside of $G'$.
This implies that $G'$ contains a source $s$.
Since $s$ is the only source of $G$ and $f$ not the outer face, we know that $f'$ is thus the outer face.
Therefore, all incoming edges of $w$ have to lie in $G'$.
Since $w$ precedes $u$ and $f$ is to the right of $e$, the left outgoing edge of $w$ lies in $G'$.
This case is illustrated in \cref{fig:singleSource}~(b).
We see that $G'$ cannot be flipped and hence $e$ remains a bad edge in any upward planar embedding of $G$.
The relation of $G'$ and the edges of $w$ is similar when $w$ exists but does not precede $u$.
Furthermore, since all incoming edges of $w$ lie inside of $G'$ but $w$ was picked closest to $u$,
there is no other vertex forming a split pair with $u$ that also precedes $u$.
If $u$ is not part of a split pair (as in \cref{fig:singleSource} (c)), then the triconnectedness implies
that $e_v$, $e$, and $e_u$ always lie on the boundary of the same face and, hence, that $e$ is a bad edge.
We conclude that if a flip is possible for every bad edge as shown in \cref{fig:singleSource} (a),
then the embedding can be changed until no bad edge remains.
\end{proof}
With the characterisation from \cref{clm:singleSource:characterisation},
we can check whether every bad edge can be fixed and if so, perform the necessary flips.
Since split components to flip may be nested, executing simply one flip after the other could be costly.
We now show how to keep the running time linear.
\begin{theorem} \label{clm:singleSource:algorithm}
Let $G$ be an upward planar single-source digraph with $n$ vertices.
An upward planar 2-slope drawing of $G$ can be computed in $\ensuremath{\mathcal{O}}(n)$ time, if one exists.
\end{theorem}
\begin{proof}
First, compute an upward planar embedding of $G$ with the algorithm of Bertolazzi {et~al.}~\cite{BDMT98} in $\ensuremath{\mathcal{O}}(n)$ time.
Second, traverse the boundary of every face to identify all bad edges in $\ensuremath{\mathcal{O}}(n)$ time.
Third, for each bad edge $e = (u, v)$ with respect to a face $f$,
check whether it can be repaired with the condition of \cref{clm:singleSource:characterisation}.
If this is the case, mark the respective $w$ as the beginning and $u$ as the end of a split component.
This can be done by a simple traversal of $f$.
Since by \cref{clm:singleSource:twoBadEdges} any face contains at most two bad edges, this step takes also only $\ensuremath{\mathcal{O}}(n)$ time.
Suppose every bad edge can be fixed.
Then run a BFS on $G$ that starts at the source.
During the traversal, remember the parity of encountered marked beginnings and endings of split components in each vertex.
Use this information to decide for each vertex whether its edge order has to be reversed.
This takes again only linear time and as a result we get an upward planar embedding of $G$ that admits a 2-slope representation~$U_G$.
Finally, apply the algorithm from \cref{clm:plane} on $U_G$ to compute an upward planar 2-slope drawing of $G$.
\end{proof}
Note that in \cref{clm:singleSource:characterisation} whether an edge is bad in any upward planar embedding of $G$ is independent from
whether another edge is bad in any upward planar embedding of $G$.
Hence, we can also use \cref{clm:singleSource:algorithm} and \cref{clm:plane:bends} to minimise the number of bends
in a 1-bend 2-slope drawing of $G$.
\begin{corollary} \label{clm:planar:singleSource:oneBend}
Let $G$ be an upward planar single-source digraph with $n$ vertices and maximum in- and outdegree two.
An upward planar 1-bend 2-slope drawing of $G$ with the minimum number of bends can be computed in $\ensuremath{\mathcal{O}}(n)$ time.
\end{corollary}
\subsection{SPQR-trees and upward spirality} \label{sec:spqr}
Didimo {et~al.}~\cite{DGL05,DGL10} described algorithms to compute upward planar embeddings of biconnected series-parallel and general digraphs.
They then use a result from Healy and Lynch~\cite{HL07} about combining biconnected blocks of upward planar digraphs to get rid of the biconnectivity condition.
We follow their approach closely.
In particular, we also use the notions of SPQR-trees and upward spirality (with the latter tailored to our needs)
on which Didimo et al.'s approach heavily relies on.
We refer to Didimo {et~al.}~\cite{DGL10} for the precise definition of SPQR-trees (for undirected graphs and then derived for digraphs),
though recall the main concepts in this section.
Let $G$ be a biconnected digraph and let $e = (s, t)$ be any edge of $G$ called \emph{reference edge}.
An \emph{SPRQ-tree} $T$ of $G$ with respect to (wrt) $e$ represents a decomposition of $G$ with respect to its triconnected components~\cite{DT89,GM01}.
As such, it also represents all planar embeddings of $G$ with $e$ on the outer face.
Starting with the split pair $\set{s, t}$, the decomposition is constructed recursively on the split pairs of $G$.
More precisely, $T$ is a rooted tree where every node is of type S, P, Q, or R:
Q-nodes represent single edges, S-nodes and P-noes represent \emph{series components} and \emph{parallel components}, and
R-nodes represent triconnected (\emph{rigid}) components; see \cref{fig:SPQRexample} for an example.
Each node $\mu$ in $T$ has associated a biconnected multigraph $\skel(\mu)$, called the \emph{skeleton} of $\mu$,
in which the children of $\mu$ and its reference edge are represented by a virtual edge.
The root of $T$ is a Q-node representing $(s, t)$. The child of $\mu$ is now defined recursively as follows:
\begin{description}
\item[Trivial case.] If $G$ consists of exactly two parallel edges between $s$ and $t$,
then $\mu$ is a Q-node representing the edge $(s, t)$ parallel to the reference edge and the skeleton $\skel(\mu)$ is $G$.
\item[Parallel case.] If the split pair $\set{s, t}$ has three or four split components $G_1, G_2, \ldots, G_{k}$
(more are not possible with our degree restrictions), then $\mu$ is a P-node.
The skeleton $\skel(\mu)$ consists of $k$ parallel edges between $s$ and $t$ that represent
the reference edge $G_1$ and the components $G_i$, $2 \leq i \leq k$.
\item[Series case.] If $\set{s, t}$ induces two split components $e$ and $\bar G$, then $\mu$ is an S-node.
If $\bar G$ is a chain of biconnected components $G_1, \ldots, G_k$ with cut vertices $c_1, \ldots, c_{k-1}$ ($k \geq 2$),
then $\skel(\mu)$ is the cycle $t$, $e$, $s$, $e_1$, $c_1$, $\ldots$, $c_{k-1}$, $e_k$, $t$ where $e_i$ represents~$G_i$.
\item[Rigid case.] Otherwise $\mu$ is an R-node.
Let $\set{s_1, t_1}$, \ldots, $\set{s_k, t_k}$ be the maximal split pairs of $G$ with respect to $\set{s, t}$.
Let $G_i$ be the union of split components of $\set{s_i, t_i}$ except the one containing $e$.
Then $\skel(\mu)$ is obtained from $G$ by replacing each subgraph $G_i$ with $e_i$.
\end{description}
The skeleton of the root consists of two parallel edges, $e$ and a virtual edge representing the rest of the graph.
Each node $\mu$ of $T$ that is not a Q-node has children $\mu_1, \ldots, \mu_k$, in this order such that
$\mu_i$ is a child of $\mu$ based on $G_i \cup e_i$ with respect to $e_i$.
The end vertices of $e_i = (s_i, t_i)$ are called the \emph{poles} of $\mu$, of $\skel(\mu)$, and of $G_{\mu}$.
The \emph{pertinent graph} $G_{\mu}$ for a node $\mu$ of $T$ represents the full subgraph of $G$ in the SPQR-tree rooted at $\mu$.
Refer to \cref{fig:SPQRexample} for an example and note that every edge is represented by a Q-node.
We assume that $T$ is in its \emph{canonical form}, that is, every S-node of $T$ has exactly two children.
The canonical form can be derived from a non-canonical form in linear time~\cite{DGL10}.
\begin{figure}[htb]
\centering
\includegraphics{SPQRexample2}
\caption{A digraph $G$ and its (non-canonical) SPQR-tree wrt the reference edge $\set{s,t}$.
The skeletons of some nodes are depicted with the virtual edge drawn dashed.
The pertinent graph $G_\mu$ of the R-node $\mu$ is highlighted in $G$. }
\label{fig:SPQRexample}
\end{figure}
Assume $G$ is equipped with an st-numbering (based on its underlying undirected graph and with respect to the reference edge $\set{s,t}$).
Let $\mu$ be a node of $T$ with poles $u$ and $v$ such that $u$ precedes $v$ in the st-numbering.
Then $u$ is the \emph{first pole} and $v$ is the \emph{second pole} of $\mu$.
Let $U_{G_\mu}$ be a $2$-slope representation of the pertinent graph $G_\mu$ of a node $\mu$ of $T$.
Let~$u$ and $v$ be the poles of $\mu$ and let $w \in \set{u, v}$.
The \emph{pole category} $t_w$ of the pole $w$ is the way the edges of ${G_\mu}$
are incident to $w$ and which slopes they got assigned under $U_{G_\mu}$.
The sixteen possible pole categories of split components are shown in \cref{fig:poleCategories}.
Two poles~$w$ and $w'$ with pole category $t_w$ and $t_w'$, respectively, are \emph{compatible}
if $t_w$ and $t_w'$ can be combined into a pole category of higher degree.
For example, combining iR and iL gives~iRiL.
\begin{figure}[htb]
\centering
\includegraphics{PoleCategories}
\caption{Pole categories in a 2-slope representation;
gray areas indicate where the interior of the graph lies.}
\label{fig:poleCategories}
\end{figure}
Next we define upward spirality, which measures how much a split component is ``rolled up''.
The pertinent graph $G_\mu$ of a node $\mu$ of $T$ may have 2-slope representations with different spirality.
Let $U_{G_\mu}$ be a 2-slope representation of $G_\mu$ with first pole $u$ and second pole $v$.
Let $P$ be an undirected path in $U_{G_\mu}$.
Two subsequent edges $\set{x, y}$ and $\set{y, z}$ of $P$ define a \emph{left (right) turn}
if $P$ makes a 90$^{\circ}$ clockwise (resp.\ counterclockwise) turn at $y$ according to $U_{G_\mu}$.
Define the \emph{turn number} $\n(P)$ of $P$ as the number of right turns minus the number of left turns of $P$.
Let $P_l$ and $P_r$ be the clockwise and counterclockwise paths from $u$ to $v$ along the outer face, respectively.
We define the \emph{upward $2$-slope spirality} $\sigma$ of $U_{G_\mu}$ as
$\sigma(U_{G_\mu}) = \frac{\n(P_l) + \n(P_r)}{2}\text{.}$
See \cref{fig:spirality} for examples.
Note that the turn number of a path is bounded by its length.
Therefore, the number of possible values for the upward $2$-slope spirality of $U_{G_\mu}$ is in $\ensuremath{\mathcal{O}}(n)$
(and in $\ensuremath{\mathcal{O}}(d)$ where $d$ is the diameter of $U_{G_\mu}$).
\begin{figure}[htb]
\centering
\includegraphics{upwardSpirality}
\caption{Two different $2$-slope representations of the same digraph.
(a) The paths $P_l$ and $P_r$ have turn number $1$ and $0$, and so the upward $2$-slope spirality is $0.5$.
(b) The paths $P_l$ and $P_r$ have turn number $-3$ and $-4$, and so the upward $2$-slope spirality is $-3.5$}
\label{fig:spirality}
\end{figure}
Let $\mu$ be a node of $T$ with first pole $u$ and second pole $v$.
A \emph{feasible tuple} of $\mu$ is a tuple $\tau_\mu = \langle U_{G_{\mu}}, \sigma, t_u, t_v \rangle$
where $U_{G_{\mu}}$ is an upward 2-slope representation of $G_\mu$ with pole categories $t_u$ and $t_v$
and upward 2-slope spirality $\sigma$.
Two feasible tuples of $\mu$ are \emph{spirality equivalent}
if they have the same upward $2$-slope spirality and pole categories.
A \emph{feasible set} ${\mathcal{F}}_\mu$ of $\mu$ is a set of all feasible tuples of $\mu$
such that there is exactly one representative tuple for each class of spirality equivalent tuples of $\mu$.
\subsection{Series-parallel digraphs} \label{sec:sp}
Let $G$ be a biconnected series-parallel digraph, that is, an SPQR-tree (or rather SPQ-tree) of $G$ contains no R nodes.
Our goal is to test the existence of an upward planar embedding and a 2-slope representation of $G$ that does not contain a bad edge.
The idea of the algorithm is as follows.
Pick a reference edge $e$ and construct the respective SPQR-tree $T$ of $G$.
In a post-order traversal of $T$, compute the feasible set of each node.
This is straightforward for Q-nodes.
For an S- or P-node $\mu$, try to combine feasible tuples of the children of $\mu$ to feasible tuples of $\mu$.
The pole categories and upward $2$-slope spirality values make it easier to check
whether a composition admits an upward planar $2$-slope representation.
If this leads to non-empty feasible set for the root of $T$, we can construct a drawing.
Otherwise, we try again with another reference edge.
Let $\mu$ be a Q-node of $T$ with first pole $u$ and second pole $v$
and suppose $G_\mu$ is the edge $(u, v)$ (instead of $(v, u)$).
Then there are only two upward $2$-slope representation of~$G_\mu$,
namely when $(u, v)$ is assigned the slope $\sL$ or the slope $\ensuremath{\nearrow}$. Therefore, ${\mathcal{F}}_\mu$ has size two.
The following two lemmata show how to compute the feasible set of an S-node and P-node of $T$.
\begin{lemma} \label{clm:Snode}
Let $G$ be a biconnected digraph with $n$ vertices and $T$ be an SPQR-tree of $G$.
Let $\mu$ be an S-node of $T$ with children $\mu_1$ and $\mu_2$.
Given the feasible sets ${\mathcal{F}}_{\mu_1}$ and ${\mathcal{F}}_{\mu_2}$,
the feasible set ${\mathcal{F}}_{\mu}$ can be computed in $\ensuremath{\mathcal{O}}(n^2)$.
\end{lemma}
\begin{proof}
To compute a feasible tuple $\tau$ of ${\mathcal{F}}_{\mu}$ we check for all pairs
$\tau_1 \in {\mathcal{F}}_{\mu_1}$ and $\tau_2 \in {\mathcal{F}}_{\mu_2}$ whether they can be combined.
More precisely, let $u$ be the first and $v$ the second pole of $\mu$; refer to \cref{fig:seriesComposition}.
Let $u_i$ be the first and $v_i$ be the second pole of $\mu_i$, $i \in \set{1, 2}$.
Without loss of generality, assume that $u = u_1$, $v_1 = u_2$, and $v_2 = v$.
For each pair of feasible tuples $\tau_1 = \langle U_{G_{\mu_1}}, \sigma_1, t_{u_1}, t_{v_1} \rangle \in {\mathcal{F}}_{\mu_1}$
and $\tau_2 = \langle U_{G_{\mu_2}}, \sigma_2, t_{u_2}, t_{v_2} \rangle \in {\mathcal{F}}_{\mu_2}$
check whether $t_{v_1}$ and $t_{u_2}$ are compatible.
In other words, check whether $G_{\mu_1}$ and $G_{\mu_2}$ can be plugged together at $v_1 = u_2$
under the slope assignments of $U_{G_{\mu_1}}$ and $U_{G_{\mu_2}}$.
If the pole categories are compatible,
the feasible tuple~$\tau = \langle U_{G_{\mu}}, \sigma, t_{u}, t_{v} \rangle$ is given by~$t_u = t_{u_1}$, $t_v = t_{v_2}$,
$U_{G_\mu}$ as the series composition of~$U_{G_{\mu_1}}$ and~$U_{G_{\mu_2}}$ at the common vertex~$u_2 = v_1$,
and where~$\sigma$ can be computed as follows.
For~$i \in \set{1, 2}$, let~$e_l^i$ and~$e_r^i$ be the edges of~$U_{G_{\mu_i}}$ that are incident to~$u_2 = v_1$ and lie on the
clockwise and counterclockwise path from~$u$ to~$v$ along the outer face, respectively; see \cref{fig:seriesComposition} (b).
Note that~$e_l^i$ may coincide with~$e_r^i$.
For $j \in \set{l, r}$, let $\alpha_j$ be $-1$, $1$, or $0$ depending on whether~$e_j^1$ and~$e_j^2$ make a left, right, or no turn, respectively.
The upward~$2$-slope spirality of~$U_{G_\mu}$ is $\sigma = \sigma_1 + \sigma_2 + \frac{\alpha_l + \alpha_r}{2}$
(compare to Lemma 6.4 by Didimo {et~al.}~\cite{DGL10}).
Store $\tau$ in~${\mathcal{F}}_\mu$ if~${\mathcal{F}}_\mu$ contains no feasible tuple that is spirality equivalent to $\tau$.
Since ${\mathcal{F}}_{\mu_1}$ and ${\mathcal{F}}_{\mu_2}$ have~$\ensuremath{\mathcal{O}}(n)$ tuples, ${\mathcal{F}}_{\mu}$ can be computed in $\ensuremath{\mathcal{O}}(n^2)$ time.
\end{proof}
\begin{figure}[t]
\centering
\includegraphics{seriesComposition}
\caption{Illustration of a series composition of split components $G_1$ and $G_2$.
(b) $U_{G_1}$ and $U_{G_2}$ are compatible and the upward $2$-slope spirality can be computed based on $\sigma(U_{G_1})$, $\sigma(U_{G_2})$,
and the turns at $e_l^1$ and $e_l^2$ and at $e_r^1$ and $e_r^2$.
(c) $U_{G_1}$ and $U_{G_2}'$ are not compatible.}
\label{fig:seriesComposition}
\end{figure}
\begin{lemma} \label{clm:Pnode}
Let $G$ be a biconnected digraph with $n$ vertices and $T$ be an SPQR-tree of $G$.
Let $\mu$ be a \mbox{P-node} of $T$ with children $\mu_1, \ldots, \mu_k$ with $k \leq 4$.
Given the feasible sets ${\mathcal{F}}_{\mu_1}, \ldots, {\mathcal{F}}_{\mu_k}$,
the feasible set ${\mathcal{F}}_{\mu}$ can be computed in $\ensuremath{\mathcal{O}}(n)$.
\end{lemma}
\begin{proof}
Let $u$ be the first and $v$ the second pole of $\mu$ (and its children).
Since $u$ and $v$ have at most degree four in $G$,
$\mu$ has at most four children (but at least two).
Furthermore, four children are only possible in the final composition with the reference edge $e$.
We consider the case where $\mu$ has three children. The other two cases work along the same line.
For $i \in \set{1, 2, 3}$, let $G_i$ be the pertinent digraph of $\mu_i$.
Note that $u$ and $v$ have degree one in $G_i$.
We can thus define $e_i$ and $e_i'$ as the edges of $G_i$ incident to $u$ and $v$, respectively.
(If $\mu_i$ is a Q-node, then $e_i = e_i'$.)
Let $e_u$ and $e_v$ be the edges of $G$ incident to $u$ and $v$ respectively that lie outside of $G_\mu$.
(Note that $e_u = e_v$ is only possible in the final composition, where $e_u$ is then also the reference edge.)
We want to construct a $2$-slope representation of~$G_\mu$
such that the order of $e_1, e_2, e_3, e_u$ at $u$
is the reverse order of $e_1', e_2', e_3', e_v$ at $v$; see \cref{fig:parallelComposition} (a).
Furthermore, the half edges representing $e_u$ and~$e_v$ have to be on the outer face.
Suppose we pick a feasible tuple $\tau_1 = \langle U_{G_{1}}, \sigma_1, t_{u_1}, t_{v_1} \rangle \in {\mathcal{F}}_{\mu_1}$.
Then $t_{u_1}$ and $t_{u_2}$ restrict the choices of pole categories of compatible upward $2$-slope representations of~$G_2$ and $G_3$.
Likewise, $\sigma_1$ restricts these choices further; see \cref{fig:parallelComposition} (b) and (c).
More precisely, we observe that the upward $2$-slope spirality $\sigma_2$ and $\sigma_3$ of compatible $U_{G_2}$ and~$U_{G_3}$
differ from $\sigma_1$ by at least two and at most six (compare to Lemma~6.6 by Didimo {et~al.}~\cite{DGL10}).
We thus only have to consider $\ensuremath{\mathcal{O}}(1)$ feasible tuples in ${\mathcal{F}}_{\mu_2}$ and ${\mathcal{F}}_{\mu_3}$ instead of iterating over complete sets.
Storing the feasible tuples in a hash table based on the spirality and pole categories
we can find compatible $\tau_2$ and $\tau_3$ in constant time.
Similar to the series-composition, we can compute the upward $2$-slope spirality of the resulting upward $2$-slope representation $U_{G_\mu}$
based on its categories and~$\sigma_i$, $i \in \set{1, 2, 3}$.
Hence, by iterating once over ${\mathcal{F}}_{\mu_1}$ we can construct all feasible tuples of ${\mathcal{F}}_\mu$ in $\ensuremath{\mathcal{O}}(n)$.
\end{proof}
\begin{figure}[htb]
\centering
\includegraphics{parallelComposition}
\caption{Illustration of a parallel compositions with three children and how picking an upward $2$-slope representation of $G_1$
enforces the spirality of representation of $G_2$ and $G_3$.}
\label{fig:parallelComposition}
\end{figure}
Lastly, for the root composition we check in $\ensuremath{\mathcal{O}}(n)$ time whether the root of $T$, which is a Q-node representing $e$,
can be combined with a feasible tuple of its child.
In the affirmative case, we obtain an upward $2$-slope representation of $G$.
With all compositions described, we can now prove the main theorem of this section.
\begin{theorem} \label{clm:seriesParallel}
Let $G$ be a biconnected series-parallel digraph with $n$ vertices.
There exists an $\ensuremath{\mathcal{O}}(n^4)$-time algorithm that tests if $G$ admits an upward planar 2-slope drawing and, if so,
that constructs such a drawing.
\end{theorem}
\begin{proof}
Let $e$ be an edge of $G$. Compute the (canonical) SPQR-tree $T$ with respect to $e$ of~$G$, which can be done in $\ensuremath{\mathcal{O}}(n)$ time~\cite{GM01,DGL10}.
In a post-order traversal of $T$, the algorithm computes the feasible set for every node of $T$.
If the algorithm arrives at a node with an empty feasible set, its starts with another reference of $G$.
Otherwise, the algorithm stops when it has constructed a feasible tuple for $G$.
We can then use \cref{clm:plane} to construct an upward planar 2-slope drawing.
For one reference edge, this takes at most $\ensuremath{\mathcal{O}}(n^2)$ time per node by \cref{clm:Snode,clm:Pnode}.
and since the size of $T$ is linear in $n$, at most $\ensuremath{\mathcal{O}}(n^3)$ time in total.
The total running time is thus in $\ensuremath{\mathcal{O}}(n^4)$.
\end{proof}
In \cref{sec:nonbi} we explain how to handle non-biconnected series-parallel digraphs.
\subsection{Biconnected digraphs} \label{sec:digraph}
We extend the algorithm for biconnected series-parallel digraphs to general biconnected digraphs following again Didimo {et~al.}~\cite{DGL10}.
The upward planarity check is again combined with finding a 2-slope representation.
Let $G$ be a biconnected digraph.
Let~$T$ be the SPQR-tree of $G$ with respect to a reference edge $e$.
The algorithm computes again the feasible sets of the nodes of $T$ in a post-order traversal.
For Q-, S-, or P-nodes this works as before.
Recall that to compute a feasible tuple of an S-node or P-node
it suffices to look at the pole categories and upward spirality of its children.
For R-node this connection is not as clear and we rely thus on a brute-force approach.
More precisely, we compute the feasible set by considering all possible combinations of tuples for each virtual edge of $\skel(\mu)$ to construct $U_{G_\mu}$.
If substitutions are successful, we have to check upward planarity and the existence of bad edges.
\begin{lemma} \label{clm:Rnode}
Let $G$ be a biconnected digraph with $n$ vertices and $T$ be an SPQR-tree of $G$.
Let $\mu$ be an R-node of $\,T$ with children $\mu_1, \ldots, \mu_k$.
Let $d$ be the diameter of $G_\mu$.
Given the feasible sets ${\mathcal{F}}_{\mu_1}, \ldots, {\mathcal{F}}_{\mu_k}$,
the feasible set ${\mathcal{F}}_{\mu}$ can be computed in $\ensuremath{\mathcal{O}}(d^k n^2)$.
\end{lemma}
\begin{proof}
Note that since $\skel(\mu)$ is triconnected, it has a unique planar embedding (up to mirroring),
which we can compute in $\ensuremath{\mathcal{O}}(n)$.
Note that, by \cref{clm:plane}, if $\skel(\mu)$ contains a bad edge with respect to three non-virtual edges,
then ${\mathcal{F}}_{\mu}$ is empty. Moreover, then $G$ admits no upward planar 2-slope drawing and the main algorithm can stop.
So suppose no such bad edge exists.
Construct a 2-slope representation of $G_\mu$ by substituting virtual edges with the respective 2-slope representations.
More precisely, for all $i \in \set{1, \ldots, k}$,
substitute $e_i$ with the $2$-slope representation $U_{G_{\mu_i}}$ of a feasible tuple in ${\mathcal{F}}_{\mu_i}$.
Let $U_{G_{\mu}}'$ denote the partial upward $2$-slope representation of $G_\mu$ during this process.
At each pole of a child $\mu_i$ in~$U_{G_{\mu}}'$ check
whether the $2$-slope representations of the substituted parts are conflicting.
If this check fails, backtrack and try another feasible tuple.
Suppose that it is successful for all poles of all $\mu_i$.
Then test the upward planarity of $U_{G_\mu}'$ with the flow-based upward planarity algorithm by Bertolazzi {et~al.}~\cite{BDLM94}
where the assignment of switches to faces is given for the substituted parts.
This flow-based algorithm runs in $\ensuremath{\mathcal{O}}(n^2)$ time.
If $U_{G_\mu}'$ is upward planar, check whether it
contains any bad edge and, if not, extend $U_{G_\mu}'$ to a
$2$-slope representation $U_{G_{\mu}}$ of $G_\mu$ (compare to Lemma~8.2 by Didimo {et~al.}~\cite{DGL10}).
Lastly, compute the upward $2$-slope spirality of $U_{G_{\mu}}$ in $\ensuremath{\mathcal{O}}(n)$
to obtain a feasible tuple for ${\mathcal{F}}_\mu$.
By backtracking and trying the remaining feasible tuples of the $\mu_i$,
we complete the computation of ${\mathcal{F}}_\mu$.
Note that $d$ is an upper bound on the upward $2$-slope spirality of a split component
and thus each $\mu_i$ has $\ensuremath{\mathcal{O}}(d)$ feasible tuples.
Therefore the algorithm tries at most $d^k$ combinations of feasible tuples of the $k$ children of $\mu$.
Hence the feasible set of an R-node $\mu$ can be computed in $\ensuremath{\mathcal{O}}(d^k n^2)$,
where the factor $\ensuremath{\mathcal{O}}(n^2)$ comes from the flow based upward planarity test.
\end{proof}
Let $d$ denote the maximum diameter of a split components of $G$
and let $t$ denote the number of nontrivial triconnected components of $G$,
i.e., series components, parallel components, and rigid components.
In total, by \cref{clm:Rnode}, the feasible set of all R-nodes can be computed in $\ensuremath{\mathcal{O}}(d^t n^2)$.
Recall that the time needed to compute the feasible set of an S-node is bounded by the square of possible spirality values and thus in $\ensuremath{\mathcal{O}}(d^2)$.
Since we use the canonical form of SPQR-trees there are $\ensuremath{\mathcal{O}}(n)$ S-nodes.
Therefore, we can compute the feasible sets of all S-nodes of $T$ in $\ensuremath{\mathcal{O}}(d^2n)$ time.
Similarly, the feasible sets of all P-nodes of $T$ can be computed in $\ensuremath{\mathcal{O}}(dt)$ time.
Lastly, iterating over all SPQR-trees of $G$ for the different choices of reference edges adds another factor of $\ensuremath{\mathcal{O}}(n)$.
If a 2-slope representation of $G$ has been found, we apply again \cref{clm:plane} to compute a drawing.
Hence, we get the following theorem.
\begin{theorem} \label{clm:digraph}
Let $G$ be a biconnected digraph with $n$ vertices.
Suppose that $G$ has at most $t$ nontrivial triconnected components, and that each split component has diameter at most $d$.
Then there exists an $\ensuremath{\mathcal{O}}(d^tn^3 + dtn +d^2n^2)$-time algorithm that tests if $G$ admits an upward planar 2-slope drawing and, if so,
that constructs such a drawing of $G$.
\end{theorem}
\subsection{General digraphs}\label{sec:nonbi}
So far we have seen how to test whether a biconnected digraph admits a 2-slope representation.
For a general digraph $G$, even if each of its biconnected components, called a \emph{block}, has a 2-slope representation
we may not be able to join the blocks; see \cref{fig:general:nonMerging}.
In fact, even all blocks of $G$ being upward planar does not imply that $G$ is upward planar~\cite{HL07}.
Nonetheless, following Healy and Lynch~\cite{HL07}, our strategy is to test for each block
if its admits a 2-slope representation under some special conditions
and, in the affirmative, join these representations.
\begin{figure}[htb]
\centering
\includegraphics{unsuccessfulBlockMerging}
\caption{$G$ does not admit a 2-slope representations (a), even though its blocks do (b).}
\label{fig:general:nonMerging}
\end{figure}
When we want to merge the representations of two blocks, we have to take two things into consideration.
Namely, we have to test whether their joined cut vertex $c$ is on the outer face for one of the two blocks
and whether their 2-slope representations fit together at $c$ (just like poles and their pole categories).
Before we get into detail on this, we recall what a block-cut tree is
and explain how it gives a suitable order to process the blocks.
\paragraph{Block-cut tree.}
The \emph{block-cut tree} ${\mathcal{T}}$ of $G$ contains a vertex for each block and for each cut vertex,
and an edge between a cut vertex $c$ and each block that contains $c$.
For $G$ to admit a 2-slope representation,
we must be able to root ${\mathcal{T}}$ at a block (with all edges oriented towards this root block)
such that an edge $\set{B, c}$ between a block $B$ and cut vertex $c$ is oriented towards $c$
only if $B$ admits a 2-slope representation where $c$ is on the outer face
(see \cref{fig:blockCutTree} and compare to Lemma 3 by Chan~\cite{Cha04}).
Note that if we can root ${\mathcal{T}}$ at a block $B'$, then any other block $B$ has exactly one outgoing edge
and $B'$ has only incoming edges.
For a block $B$ with outgoing edge to a cut vertex $c$,
we say $B$ is a \emph{block with respect to (wrt)} $c$.
\begin{figure}[htb]
\centering
\includegraphics{blockCutTree}
\caption{A digraph with six blocks and 2-slope representation (a) and the corresponding rooted block-cut tree (b).}
\label{fig:blockCutTree}
\end{figure}
Suppose we would know how to root ${\mathcal{T}}$.
With a post-order traversal of ${\mathcal{T}}$, we could then try to find
a 2-slope representation for each block $B$ wrt $c$ of ${\mathcal{T}}$
that has $c$ on the outer face.
However, a priori we do not know at which block to root ${\mathcal{T}}$.
Hence, our algorithm works as follows.
\paragraph{Algorithm.}
Given $G$, we can find its cut vertices and blocks and construct its block-cut tree ${\mathcal{T}}$ in $\ensuremath{\mathcal{O}}(n)$ time~\cite{Tar72}.
Since we do not know what block can function as root of~${\mathcal{T}}$ yet,
we start at the leaves of ${\mathcal{T}}$ and work ``inwards''.
Note that a leaf block $B$ has only one edge~$\set{B, c}$ in ${\mathcal{T}}$ to a cut vertex $c$ (unless $B = G$)
and we can thus provisionally direct $\set{B, c}$ to $c$.
This yields a block block wrt a cutvertex -- $B$ wrt $c$.
Furthermore, during the algorithm,
for at least one non-leaf blocks either all or all but one of its neighboring blocks have been handled.
Therefore, during this ad hoc post-order traversal of ${\mathcal{T}}$,
we can ensure that we always have at least one block wrt a cutvertex.
To process a block $B$ wrt $c$, we check
whether $B$ has a 2-slope representation $U_B$
with (i) $c$ on the outer face and (ii) additional constraints
on the angles formed at all other cut vertices of $B$, which we describe in detail below.
If this is the case, then we can finalize the direction of the edge $\set{B, c}$ towards $c$.
Otherwise, if $B$ does not admit such a 2-slope representation, then $B$ has to be the root of ${\mathcal{T}}$.
We then orient all edges of ${\mathcal{T}}$ towards $B$ and continue in the remaining part of ${\mathcal{T}}$.
If we later find that another block also needs to be the root of ${\mathcal{T}}$,
then $G$ does not admit a 2-slope representation.
Furthermore, when we arrive back at $B$,
we have to test whether $B$ admits a 2-slope representation at all.
Note that for a block $B$ wrt $c$, where $c$ has to be on the outer face of the 2-slope representation $U_B$,
the algorithms from the previous two sections only have to consider an SPQR-tree
with an edge incident to $c$ as reference edge.
\paragraph{Angles of cut vertices.}
We now describe the additional constraints
that have to be checked for a block $B$ at all of $B$'s cut vertices.
More precisely, let $B$ be a block wrt $c$
and let $c'$ be any other cut vertex of $B$ if it has any.
Depending on the degree of $c$ (and $c'$) in $B$ and in the neighboring blocks of $B$,
we have the following extra conditions on the angles at $c$ and $c'$ in $U_B$ of $B$.
\begin{itemize}
\item Suppose $c$ (or $c'$) has degree one in $B$. Then $B$ is a single edge,
$c$ is automatically on the outer face in any 2-slope representation of $B$,
and the angle at $c$ in $U_B$ is insignificant.
\item Suppose $c$ has degree two in $B$ and $c$ has degree two in another block $B'$; see \cref{fig:blockJoiningAngles}~(a).
Then $c$ has to have a large angle on the outer face in $U_B$,
since otherwise it would not be able to attach to $U_{B'}$ such that $B'$ is in the outer face of~$B$.
If the same case applies to~$c'$, then $c'$ has to have a large angle in $U_B$, but not necessarily on the outer face.
\item Suppose $c$ has degree two in $B$ and $c$ has degree one in two other blocks $B_1$ and $B_2$.
Then we first test if $B$ admits a 2-slope representation where $c$ has a large angle on the outer face; see \cref{fig:blockJoiningAngles}~(b).
In this case, both $B_1$ and $B_2$ lie in the outer face of $B$.
Otherwise, if $c$ has indegree one and outdegree one in $B$,
we test whether $B$ admits a 2-slope representation where $c$ forms a flat angle on the outer face; see \cref{fig:blockJoiningAngles}~(c).
We test once such that $B_1'$ lies on the outer face of $B$ and once for $B_2'$.
If neither is possible, then $B$ has to be the root.
For $c'$ there are no restrictions under these conditions.
\item The case where $c$ (or $c'$) has degree two in $B$ and degree one in exactly one other block is similar to the previous case but simpler.
\item Suppose $c$ has degree three in $B$; see \cref{fig:blockJoiningAngles}~(d). Then $c$ has to have a flat angle at $c$ on the outer face.
Otherwise $B$ has to be the root.
There are again no restrictions for~$c'$ under these conditions.
\end{itemize}
\begin{figure}[htb]
\centering
\includegraphics{blockJoining}
\caption{Conditions on the angles at $c$ and $c'$ in $U_B$ for block a $B$ wrt $c$.}
\label{fig:blockJoiningAngles}
\end{figure}
These conditions are clearly necessary and, following Healy and Lynch~\cite{HL07}, also sufficient.
Furthermore, they can easily be tested by the algorithms from the previous sections.
More precisely, if $c$ has degree two in $B$, then its two incident edges are either merged in the root composition or a series composition.
Hence, at this step we only keep those 2-slope representations with the desired angle at $c$.
Otherwise, if $c$ has degree three in $B$ and has to form a flat angle,
we take the sole outgoing or sole incoming edge of $c$ in $B$ as the reference edge for the SPQR-tree.
In the root composition we then only allow a merge when there is the desired flat angle at $c$ on the outer face.
\paragraph{Result.}
The total running time for our algorithm to test whether a general digraph admits a 2-slope representation
and thus an upward planar 2-slope drawing
is given by (i) a linear amount for the computation of ${\mathcal{T}}$, (ii) the sum of the checks for each block,
and (iii) a linear amount for merging.
Hence, we get the following results.
\begin{theorem}
Let $G$ be a series-parallel digraph with $n$ vertices.
There exists an $\ensuremath{\mathcal{O}}(n^4)$-time algorithm that tests if $G$ admits an upward planar 2-slope drawing and, if so,
that constructs such a drawing.
\end{theorem}
\begin{theorem}
Let $G$ be a digraph with $n$ vertices.
Let $t$ be the maximum number of nontrivial triconnected components of a block of $G$, and
$d$ be the maximum diameter of a split component of a block of $G$.
Then there exists an $\ensuremath{\mathcal{O}}(d^tn^3 + dtn +d^2n^2)$-time algorithm that tests
if~$G$ admits an upward planar 2-slope drawing and, if so,
that constructs such a drawing of~$G$.
\end{theorem}
\section{Phylogenetic networks} \label{sec:phynet}
Recall from \cref{sec:introduction} that a phylogenetic network is a single-source digraph whose sinks are all leaves
and whose non-sink, non-source vertices have degree three.
In this section we show how to find an upward planar 2-slope drawing of a phylogenetic network~$N$
such that its leaves lie on a horizontal line -- if~$N$ admits such a drawing.
Since we want that all leaves are on the outer face, we first merge them into a single vertex
and then apply the linear-time algorithm of Bertolazzi {et~al.}~\cite{BDMT98} to test whether the resulting digraph~$N'$ is upward planar.
Clearly,~$N'$ is upward planar if and only if~$N$ admits a desired upward planar embedding.
In the affirmative case, let~$N$ now be an upward plane phylogenetic network such that its~$k$ leaves lie on the outer face.
Further assume that~$N$ contains no bad edge or, in this case equivalently, no transitive edge.
In \cref{sec:plane}, we constructed an upward planar 2-slope drawing by implementing the refinement step,
which augments all faces to rectangular faces, and by applying a compaction algorithm~\cite{KKM01}.
In order to obtain a drawing where all leaves lie on the same horizontal line,
we apply this algorithm to the following augmentation~$\bar{N}$ of~$N$.
Let~$l_1, l_2, \ldots, l_k$ be the leaves of~$N$ in clockwise order around the outer face.
Add new vertices~$v_1, v_2, \ldots, v_{k-1}$
and edges~$e_i = (l_i, t_i)$ and~$e'_{i}= (l_{i+1}, v_i)$, $i \in \set{1, \ldots, k-1}$; see \cref{fig:phynet} (a).
Then apply \cref{clm:plane} to $\bar{N}$ to obtain an upward planar 2-slope drawing of~$\bar{N}$ in~$\ensuremath{\mathcal{O}}(n)$ time.
\begin{figure}[htb]
\centering
\includegraphics{phynetAlgo}
\caption{(a) The augmentation $\bar{N}$ of an upward planar phylogenetic network $N$.
(b) An upward planar 2-slope drawing where $l_1$ and $l_2$ have different y-coordinates.
(c) The dual flows $G_l$ (red) and $G_r$ (blue). (d) Propagating the length difference of $e_1$ and $e_1'$ through $G_l$.}
\label{fig:phynet}
\end{figure}
We observe from \cref{fig:phynet} (b) that two vertices $l_i$ and $l_{i+1}$ of $\bar{N}$ (neighboring leaves of~$N$)
have different y-coordinates if and only if $e_i$ and $e_i'$ have different lengths.
This can be fixed by propagating these length differences through the drawing in the following way (\cref{fig:phynet} (c--d)).
Let $G$ be the dual graph of $\bar N$.
Furthermore, define $G_l$ and $G_r$ as the two subgraphs of $G$
with $V(G) = V(G_l) = V(G_r)$
and where $E(G_l)$ and $E(G_r)$ are the dual edges of primal edges with slope $\sL$ and $\ensuremath{\nearrow}$, respectively.
In other words, $G_l$ is the dual graph of $\bar N$ restricted to edges with slope $\sL$.
Direct every edge $e^*$ in $E(G_l)$ ($E(G_r)$) with primal edge $e$
from the left (right) face of $e$ to the right (left) face of $e$.
Assign each dual edge as flow the length of its primal edge.
Split the vertex corresponding to the outer face of $\bar N$
into a source $s$ and $k-1$ sinks $t_1, t_2, \ldots, t_{k-1}$ such that $t_i$ has the incoming edges dual
to the edges $e_i$ and $e'_i$; see \cref{fig:phynet} (c).
These dual graphs can be constructed in linear time.
Next, to adjust the heights of the leaves of $N$,
for every pair $e_i$ and $e_i'$, $i \in \set{1, \ldots, k-1}$, if, say, $e_i'$ is shorter
than $e_i$, propagate the difference as flow backwards towards $s$ through~$G_l$; see \cref{fig:phynet} (d).
With one DFS on $G_l$ and $G_r$ each, all leaves can be handled simultaneously and in linear time.
Lastly, since some edge lengths have been changed, update the coordinates of all vertices in $\bar{N}$
and remove the vertices $v_i$, $i \in \set{ 1, \ldots, k}$, to obtain an upward planar 2-slope drawing of $N$.
The following theorem summarises this section.
\begin{theorem} \label{clm:phynet}
Let $N$ be a phylogenetic network with $n$ vertices and no transitive edge.
If $N$ admits an upward planar drawing with all its leaves on the outer face,
then $N$ admits and upward planar 2-slope drawing such that all its leaves lie on a horizontal line.
Moreover, such a drawing can be constructed in $\ensuremath{\mathcal{O}}(n)$-time.
\end{theorem}
Note that by \cref{clm:planar:singleSource:oneBend} a phylogenetic network with $m$ transitive edges
admits a upward planar 1-bend 2-slope drawing with at most $m$ bends.
\section{Concluding remarks}
When considering the number of slopes in a graph drawing, one typically asks how many different slopes are necessary for a graph of certain graph class.
Here we instead constrained the number of slopes to two and asked what digraphs can then be drawn upward planar.
Our digraphs are thus limited to those that contain no transitive edges and have a small maximum degree.
Beyond that, the difficulty of the problem depends on whether or not an upward planar embedding is given
and on the complexity of the digraph.
We have shown that if the embedding is fixed then the question can be answered and,
in the affirmative, a drawing constructed in linear time.
In this case the problem boils down to whether there is a bad edge for the given embedding and, if not, to adapt algorithms for orthogonal drawings.
However, even if there are bad edges present, allowing each of them to bend once is enough to obtain upward planar 1-bend 2-slope drawing
with minimum number of bends.
We conjecture that it is NP-hard to minimize the drawing area of an upward planar 2-slope drawing just like it is for orthogonal drawings~\cite{Pat01}.
It would be interesting to see a proof for this and how compaction algorithms for orthogonal drawings can be applied to upward planar drawings.
If a given digraph is not embedded yet, we first have to check whether the digraph is upward planar.
For single-source digraph, we have seen that it suffices to find one upward planar embedding, which may then be altered to one without bad edges if it exists.
For series-parallel and general digraphs we reused an approach by Didimo {et~al.}~\cite{DGL10} based on SPQR-trees and upward spirality
to find a quartic time and a fixed-parameter tractable algorithm, respectively.
An important difference is that our algorithm does not only compute upward planar embeddings for nodes of the SPQR-tree but also 2-slope representations.
Through the degree restrictions the algorithm became simpler and can thus also consider other properties.
It would be interesting to see whether the algorithm that computes an upward planar embedding of a single-source digraph can be modified
to directly compute a 2-slope representation.
This research was motivated by drawings of phylogenetic networks.
While we here assumed that a given phylogenetic network is upward planar, this is not a biologically motivated property of phylogenetic networks.
One may argue that phylogenetic networks often have few reticulations (vertices with indegree two or higher),
but even just two reticulations suffice to obstruct upward planarity.
Hence, it would be interesting to have algorithms that can also draw non-upward planar phylogenetic network with two slopes.
The biggest challenges remain for drawings with more than two slopes.
Our feeling is that while the complexity of developing algorithms to draw graphs with two slopes is manageable,
three or more slopes increase the geometric interdependence dramatically.
While the companion paper by Klawitter and Zink~\cite{KZ21} started to investigate this,
we would be happy to see more results on upward planar slope numbers of graphs.
\pdfbookmark[1]{References}{References}
|
\section{Introduction}
\label{sec:intro}
Speech technology has rapidly proliferated and integrated deeply into our daily life \cite{akccay2020speech,10.1145/1873951.1874246,xie2019utterance,braga2019automatic,toth2018speech}. While these applications bring convenience to our life, several growing concerns have gained attention and need to be addressed with care. The first is \textit{privacy} due to concerns of sensitive information leakage: for example, users may not expect to disclose their identity information while using a speech emotion recognition (SER) system; on the other hand, users may not wish to share their emotional condition when being assessed by a speaker recognition (SR) system. Moreover, the collective social norm would create unwanted and often detrimental self-exaggerated issues around equality, e.g., unfair biases toward gender types \cite{tatman2017gender} or race \cite{sap2019risk}, when using data-driven approaches for speech technology. Speech is an informative signal which contains personal sensitive attributes by nature; hence developing appropriate methods either to protect privacy information, such as identity and emotion, or to mitigate the undesired biases, like gender and race, is critical in the current era.
Recently, several works in speech processing have started to address these issues using privacy-aware representation learning. For example, Srivastava et al.\ used adversarial representation learning on automatic speech recognition (ASR) to protect speaker identity \cite{Srivastava2019}, Alouf et al.\ used CycleGAN-based method to generate emotion-less synthesized speech for voice assistant to hide personal affect \cite{aloufi2019emotionless}, Jaiswal et al.\ used adversarial learning to generate gender-invariant representation for identify-free emotion recognition \cite{Jaiswal2020PrivacyEM}, and Xia et al.\ applied adversarial learning to mitigate racial bias in hate speech detection \cite{xia-etal-2020-demoting}. While current state-of-the-art methods concentrate on using adversarial learning, this strategy suffers from several shortcomings. Adversarial method address privacy issues by learning a speech signal space with no targeted sensitive attributes as measured by its ability in fooling a well-trained discriminator that is in charge of classifying sensitive information, e.g., gender and speaker identity. This attribute invariant learning strategy lacks a flexible mechanism to adapt to different criterion of privacy preserving; for example, in some tasks only the ``gender'' attribute may need to be protected while some other tasks would require the ``speaker identity" to be private. For different scenario of interest, one would have to re-train the adversarial network over again.
In this work, instead of taking a `per-attribute' adversarial invariant learning approach, we formulate the problem as devising a learning strategy that would result in attribute-aligned speech representation. The core idea centers on conceptualizing that speech contains a mixture of attributes, \cite{Hsu2017,8462169}, e.g., gender, age, emotion and semantics, etc. By factorizing the entangled information of speech signal into independent attributes with proper attribute-alignment, we can protect particular sensitive information by attribute selection, i.e., masking targeted sensitive attributes, to minimize either privacy-leakage or biased decision. In this paper, we evaluate this idea by targeting two sensitive attributes in speech, i.e., emotion and identity, and our aim is to show that this approach can flexibly achieve privacy-preserving applications by eliminating identity contents in SER or emotion contents in SV at ease.
We propose a framework of flexible attribute masking for speech, inspired by the fair representation learning \cite{pmlr-v97-creager19a}. We aim to learn a layered disentangled speech representation with a backbone of variational autoencoder (VAE) \cite{Kingma2014,DBLP:journals/ftml/KingmaW19}. We specifically propose a layered dropout strategy in a multi-task framework to achieve attribute-alignment, i.e., forces the latent to align in an emotion-related to identity-related order. To further \textit{clean up} the aligned representation knowing that these two attributes are highly correlated \cite{8925044,pappagari2020x}, we add adversarial reversal layer to each task-specific branch. Our strategy provides flexibility in either identity masking or emotion masking to come up with an identity-free latent for privacy-preserving SER or emotionless latent for privacy-preserving SV with a unified learning framework. In this work, we evaluate our method on MSP-Podcast \cite{8003425} for SER and SV tasks using three types of feature, and achieve competitive results on SER (emobase: 52.41\% weighted f-score, 41.14\% EER), and an improvement on SV (netvlad: 34.35\% weighted f-score, 10.91\% EER; x-vector: 34.23\% weighted f-score, 9.63\% EER), compared to the state-of-the-art adversarial learning method.
\begin{figure*}[t]
\centering
\vspace{-9mm}
\includegraphics[width=0.9\linewidth,]{figures/architecture.png}
\caption{An illustration of our proposed method for attribute-aligned representation learning. It includes three blocks: representation learning procedure, LR-VAE, and layered dropout. Notice that $Z_{id-free}$ stands for identity-free representation, $Z_{emo-free}$ stands for emotionless representation.}
\label{fig:model}
\vspace{-5mm}
\end{figure*}
\section{Methodology}
\subsection{Dataset Description}
In this study, we focus on two main tasks, emotion recognition and speaker verification. To evaluate the performance of these two tasks, a large corpus with emotional labels and multiple speakers is needed. Hence, we use the MSP-Podcast database \cite{8003425}, which includes over 1,000 podcast recordings. Each podcast is segmented into speaking turns, where segments with music, overlapped speech, telephone quality speech and background noise are discarded.
In this work, we use data with 5 categorical emotions: neutral, angry, sad, happy and disgust as in \cite{pappagari2020x}. We used the standard splits in Release 1.4 for training, development, and testing, which includes 610 speakers in train set, 30 speakers in development set, and 50 speakers in test set, where each set of speakers are disjoint. The distribution of the 5 emotion classes are: angry: 8.81\%, happiness: 27.10\%, neutral: 53.05\%, sad: 3.95\%, disgust: 7.09\%.
\subsection{Feature Extraction}
In this work, we use three different input features for the two tasks: emobase2010, netvlad embedding, and x-vector embedding to verify the effectiveness of our proposed method. First, we use emobase2010, which is a commonly used feature for SER, as input. It is a 1582 dimensional feature including pitch, loudness, mfcc and spectral, etc. We extract emobase2010 using openSMILE toolkit \cite{10.1145/1873951.1874246}. Further, we extract embeddings commonly used in state-of-the-art speaker verification task, i.e., netvlad \cite{xie2019utterance} and x-vector \cite{pappagari2020x}. The netvlad embedding is extracted using the released pre-trained model \cite{xie2019utterance}, while the x-vector embedding is obtained by training on the Voxceleb2 \cite{Chung2018} using the structure mentioned in \cite{pappagari2020x}.
\begin{table*}[th]
\vspace{-4mm}
\caption{The experiment results are presented in weighted f-scores (\%) and EER (\%) for SER and SV respectively. Notice that PP stands for privacy-preserving, where columns of origin stand for original representation without privacy protection, PP-SER stands for identity-free SER, and PP-SV stands for emotionless SV.}
\label{tab:compare}
\centering
\begin{tabular}{l|c|c|c|c|c|c|c|c|c|c|c}
\toprule
\multicolumn{2}{c|}{\multirow{2}{*}{\textbf{Method}}} &
\multicolumn{1}{c|}{\textbf{DNN}} &
\multicolumn{1}{c|}{\textbf{VAE}} &
\multicolumn{2}{c|}{\textbf{A-VAE}} &
\multicolumn{3}{c|}{\textbf{LR-VAE (w/o adv)}} &
\multicolumn{3}{c}{\textbf{LR-VAE}} \\ \cline{3-12}
\multicolumn{2}{c|}{} & origin & origin & PP-SER & PP-SV & origin & PP-SER & PP-SV & origin & PP-SER & PP-SV \\
\hline\midrule
\multirow{2}{*}{\textbf{emobase}} &
emo & 54.90 & 52.61 & 52.09 & 37.93 & 53.54 & \textbf{52.71} & 40.33 & 52.86 & 52.41 & \textbf{37.54} \\ \cline{2-2}
& id & 14.45 & 12.99 & \textbf{41.49} & 17.16 & 13.27 & 30.64 & 19.16 & 11.77 & 41.14 & \textbf{12.70} \\
\midrule
\multirow{2}{*}{\textbf{netvlad}} &
emo & 52.99 & 49.80 & \textbf{49.24} & 38.92 & 50.11 & 49.04 & \textbf{34.10} & 50.01 & 48.23 & 34.35 \\ \cline{2-2}
& id & 8.14 & 8.37 & 40.66 & 15.25 & 8.20 & 32.39 & 14.51 & 8.53 & \textbf{49.51} & \textbf{10.91} \\
\midrule
\multirow{2}{*}{\textbf{x-vector}} &
emo & 53.24 & 52.60 & 48.95 & \textbf{34.13} & 52.22 & \textbf{51.76} & 36.05 & 52.57 & 51.09 & 34.23 \\ \cline{2-2}
& id & 10.05 & 8.55 & 45.93 & 13.56 & 8.75 & 37.61 & 10.66 & 8.44 & \textbf{49.48} & \textbf{9.63} \\
\bottomrule
\end{tabular}
\vspace{-4mm}
\end{table*}
\subsection{Layered Representation Variational Autoencoder}
We propose a layered-representation variational autoencoder (LR-VAE) to factorize the entangled dimensions contained in speech and arrange these dimensions in an emotion-related to identity-related order. LR-VAE contains two main components, i.e., disentangled representation and layered dropout. We will first describe VAE, i.e., a well-known structure for disentangled learning. Then, we will further detail our layered dropout with adversarial multitask learning to obtain attribute-aligned speech representation.
\subsubsection{Variational Autoencoder (VAE)}
\label{sssec:vae}
In this work, we use disentangled representation learning via VAE to derive a latent node-wise independent representation. VAE model aims to learn the marginal likelihood of a data $\mathbf{x}$, with the objective function:
\vspace{-0.5mm}
\begin{equation}
\label{vae_loss}
\mathcal{L}_{VAE} = \mathbb{E}_{q(z|x)}[\log p(x|z)]-D_{KL}(q(z|x) || p(z))
\vspace{-0.5mm}
\end{equation}
where $D_{KL}(||)$ stands for the non-negative Kullback-Leibler divergence. The KL-divergence term encourages the posterior distribution to be close to an isotropic Gaussian to achieve disentanglement purpose.
\vspace{-0.5mm}
\subsubsection{Layered Dropout with Adversarial Multitask Learning}
\label{sssec:dropout}
In this work, we propose a strategy of layered dropout with a multitask-learning architecture. Multitask learning aims to include both emotion and speaker identity information into the latent codes. Layered dropout is further utilized to force these attributes to align toward both ends of the latent codes resulting in a layered representation. Also, adversarial branches, i.e., gradient reversal layer, are used to additionally `purify' this attribute-aligned representation.
Dropout is a well-known regularization method in deep learning to prevent neural networks from overfitting\cite{10.5555/2627435.2670313}. Layered dropout works in a similar manner but with a different purpose. We propose to use this as a learning mechanism to make each dimension of the latent codes carry different importance to the designated task. In our work, the two tasks are defined as the emotion recognition and the speaker verification. We design a dropout rate function making the probability of dropping decreases (or increases) monotonically for each node of the input layer. This effectively forces the target task's discriminatory information to concentrate on nodes with lower dropout rates.
Let $\mathbf{x}$ denotes the input vector with $N$ dimensions of a layer of a neural network, we define a vector with decreasing preserving rates $\mathbf{p}$ for task of emotion recognition (increasing preserving rates for speaker verification), where $0\le p_{i}\le 1$ for $i \in \{0,\dots,N-1\}$. With layered dropout, the input vector $\mathbf{x}$ of the feed-forward operation is replaced by vector $\mathbf{\widetilde{x}}$, generated by:
\vspace{-1mm}
\begin{equation}
m_i \sim \text{Bernoulli}(p_i)
\end{equation}
\begin{equation}
\mathbf{\widetilde{x}} = \mathbf{m} * \mathbf{x}
\end{equation}
Here, $*$ denotes an element-wise product. $\mathbf{m}$ acts as a mask before the vector $\mathbf{x}$ is fed into the layer. For a dimension $m_{i}$ in the vector $\mathbf{m}$, it's an independent Bernoulli random variable with probability $p_{i}$ being 1, which means to preserve the $i^{th}$ node, and 0 means to drop the $i^{th}$ node. While testing, $\mathbf{W}$ represents for weights of network and the weights are scaled as $\mathbf{W}_{test}=\mathbf{W}*\mathbf{p}^\top$ and inference without dropout, which is same as the vanilla dropout layer.
This layered dropout mechanism alters the dropout rates being applied on both sides of the representation before an emotion (identity) classifier, the latent codes form an aligned emotion-to-identity order from top end to bottom end during the optimization step. Furthermore, we add an auxiliary mechanism of adversarial branches with gradient reversal layers \cite{ganin2015unsupervised} during multitask learning. The goal is to learn cleaner factorized identity-free (emotion-free) representation. After having an attribute-aligned representation, we simply need to mask the dimension representing the particular attribute of interest. For example, to protect identity information in SER, we can simply mask the nodes that have high emotion preservation rates and low speaker identity preservation rates, and vice versa for in SV. Notice that our attribute aligned strategy provides a mechanism to select ``what to protect'' with a single unified learning, which is more efficient than the adversarial method that requires re-training in different scenarios.
\vspace{-2mm}
\section{Experiments}
\subsection{Experiment Setup}
\label{ssec:setup}
The structure of our VAE model is as follows: multi-layer perceptron (MLP) is applied for encoder and decoder. Additionally, fully connected layer is applied to model the mean and log variance of the latent code for the encoder. For multi-task learning, two MLP classifiers are trained for emotion recognition and speaker identification, and two MLP discriminators are trained for adversarial learning by applying gradient reversal layers \cite{ganin2015unsupervised}. We set the learning rate as $5e^{-4}$, and the batch size as 128. Moreover, we add a regularization of $1e^{-6}$ to all weights and biases to stabilize the training process. Rectified Linear Unit (ReLU) is chosen as the activation function. We train the model using Adam optimizer, with $L_{obj}$ as the objective, defined as:
\vspace{-1.5mm}
\begin{equation}
L_{obj}=L_{VAE}+L_{emo}+L_{id}+L_{emo-adv}+L_{id-adv}
\vspace{-1.5mm}
\end{equation}
where $L_{VAE}$ represents the reconstruction error and KL divergence loss as defined in equation \ref{vae_loss}, while $L_{emo}$ and $L_{id}$ represents the cross-entropy loss for emotion recognition and speaker identification; $L_{emo-adv}$ and $L_{id-adv}$ represents the adversarial loss for emotion recognition and speaker identification. Notice that for speaker verification (SV) task, models are trained to predict speaker identity in the training set, i.e., speaker identification, to learn identity-related information during training; while during evaluation, the hidden layer embedding is extracted and apply to speaker verification system.
We evaluate the performance of SER using weighted f-score (WFS), following the experiment setup in \cite{pappagari2020x}, and evaluate the performance of speaker verification by equal error rate (EER). For each feature set, we train a factorized layered representation encoder based on training set, select model using validation set, and test performance on test set. Assuming attackers have access to the training set with encoded representation and labels of speakers. Our goal is to generate a representation such that for the encoded representations with particular sensitive attributes masked, neither attackers nor hosts are able to identify the sensitive attributes while the main task performance is maintained.
\begin{figure*}[htb!]
\vspace{-4mm}
\minipage{0.33\textwidth}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth, width=5.5cm]{figures/curve_emobase.png}
\caption{emobase}
\label{fig:emobase}
\end{subfigure}
\endminipage
\minipage{0.33\textwidth}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=1.0linewidth, width=5.5cm]{figures/curve_netvlad.png}
\caption{netvlad}
\label{fig:netvlad}
\end{subfigure}
\endminipage
\minipage{0.33\textwidth}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth, width=5.5cm]{figures/curve_xvector.png}
\caption{x-vector}
\label{fig:xvector}
\end{subfigure}
\endminipage
\caption{The performance curves in the masking experiment where the y-axis for SER results are weighted f-scores, and EER for SV. For the first row (protect identity), the masking process starts in a bottom up order, while for the second row (protect emotion), we start masking from the top group.}
\label{fig:curve}
\vspace{-5 mm}
\end{figure*}
\subsubsection{Baseline Methods}
\label{sssec:baseline}
The following are the baseline methods of different learning strategies that we use to compare with LR-VAE. Notice that privacy-preserving (PP) on LR-VAE are done by masking the dimension of particular sensitive attributes in the latent codes. \\
{\bf DNN}: A model conducted by fully connected layers to obtain the baseline performance on SER and SV for each feature.\\
{\bf VAE}: A vanilla VAE trained by multi-task learning on SER and SV tasks. \\
{\bf A-VAE}: A VAE trained for single task (SER or SV) with adversarial learning (reverse gradient) on the other task (SV or SER). \\
{\bf LR-VAE (w/o adv)}: A model similar to our proposed LR-VAE, but trained without adding the adversarial branch.
\subsection{Result and Analysis}
\label{ssec:analysis}
\subsubsection{Sensitive Attribute Protection}
\label{sssec:performance}
Note that all the comparison are presented in absolute points in this section. For privacy-preserving speech emotion recognition (PP-SER), we aim to protect user's identity information while preserving the emotion recognition performance. As shown in the PP-SER columns in table \ref{tab:compare}, our proposed LR-VAE achieves the better privacy preserving performance on x-vector and netvlad and similar result on emobase comparing to A-VAE. It shows that our proposed method is able to obtain a competitive emotion recognition performance (0.32\% WFS higher on emobase, 1.01\% WFS lower on netvlad, and 2.14\% WFS higher on x-vector), with better improvements in protecting speaker identity (only 0.35\% EER worse on emobase, 8.85\% EER better on netvlad, and 3.55\% EER better on x-vector).
On the other hand, to achieve emotion-protected speaker verification (PP-SV), we aim to reduce users' emotional information in the speech while preserving the speaker verification performance. As shown in PP-SV columns in table \ref{tab:compare}, our proposed LR-VAE achieves the best emotion protection performance on all three features comparing to A-VAE. It shows that our proposed method could better maintain the speaker verification performance (4.46\% EER better on emobase, 4.34\% EER better on netvlad, and 3.93\% EER better on x-vector), while achieving state-of-the-art emotion-related attributes protection (0.39\% WFS better on emobase, 4.57\% WFS better on netvlad, and 0.10\% WFS worse on x-vector)
We first study the baseline DNN results shown in the column, DNN, in table \ref{tab:compare}. The promising performance show that regardless of features, it contains both emotion and identity information. It reinforces the current concerns that speech contains many personal attributes that users may not want to reveal. Then, we compare the DNN results to VAE results shown in the column, VAE origin, in table \ref{tab:compare}. We do see that there is a slight performance drop in emotion recognition potentially due to the information loss caused by kl-divergence loss in VAE training for factorization, which is a trade-off between disentanglement and reconstruction. This factorization VAE is however a key backbone in achieving our attribute-aligned representation.
To study how adversarial branches work in our framework, we compare LR-VAE results to LR-VAE(w/o adv). It shows that without adversarial learning in explicitly purifying the emotion-related (identity-related) dimension to identity-free (emotion-free), the representation learned is not ``clean'' enough. Hence, while LR-VAE(w/o adv) also achieves competitive results on main tasks, the sensitive attribute-preserving results are usually worse. This also demonstrates that the emotion-related (identity-related) attributes may contains identity (emotion) information if not explicitly cleaned.
\vspace{-2mm}
\subsubsection{Analysis of Aligned Attributes}
\label{sssec:concentration}
In this section, we further discuss the effectiveness of layered dropout that align attribute-specific information to both ends of the latent codes. We conduct an experiment with the following procedure: we encode the input features into latent codes; next, we divide the latent dimensions into 32 groups; then, for each step, we mask one additional group of latent codes, and train two models, one for emotion recognition, and the other for speaker verification. We compare the performance curve of LR-VAE and A-VAE to observe how layered dropout influence the discriminatory power of the chosen latent code dimension.
We first study the privacy-preserving speech emotion recognition task. The results are shown in the upper row of figure \ref{fig:curve}. In this experiment, we start masking from the bottom of the latent code, which contains more identity-related attributes, to the top of the latent code (more emotion-related attributes). As the procedure moves on, the speaker verification performance steadily decreases (EER increases) until the masking process reaches the middle part of the latent code, where it results in a high EER indicating the point where we achieve an identity-free representation. We can also see that EER curves of LR-VAE and A-VAE intersects, which shows that the masked LR-VAE latent can better eliminate the identity-related attributes. On the other hand, we also observe that the emotion recognition performance slightly decreases toward the ending portion of masking process due to a significant reduction in the node dimension, though A-VAE has an even earlier performance drop.
Next, we study the emotion-protection speaker verification task. The results are shown in the lower row of figure \ref{fig:curve}. In this part of experiment, we start masking from the top of the latent code, similar to the previous procedure, but in a reverse order. As the progress moves on, the emotion recognition performance steadily decreases (weighted f-score decreases), and finally reaches to a similar result comparing to A-VAE. On the other hand, we can see that the speaker verification performance of LR-VAE is better preserved comparing to A-VAE, i.e., the EER curve of LR-VAE is lower in the beginning and increases slower comparing to A-VAE.
\vspace{-2mm}
\section{Conclusions and Future works}
In this paper, we propose a novel disentangled layered speech representation learning that can flexibly preserve sensitive attribute in a unified single training architecture. Compared with other methods, our method achieves a competitive performance on identity-free SER and an improvement on emotionless SV. Also, we show that our proposed method help in pushing the emotion and identity information toward the both ends of the latent codes, and this strategy provides a flexible mechanism to select the target sensitive attributes to protect. Moreover, our attribute aligned learning strategy reduce the training and memory cost as we require only single process and single model to achieve competitive privacy-preserving results on SER and SV against adversarial training, which requires training twice and two models.
In the future, we will generalize our attribute aligned representation from two specific task to general multi-attributes scenarios. We could utilize the middle portion of the latent codes to capture other information about the speaker, e.g., gender, personality, semantics, etc, in order to provide a more complete profile on this factorized speech representation. Moreover, as the disentanglement achieved by kl divergence loss causes information loss, different factorization methods may be applied to enhance our representation capacity.
\bibliographystyle{IEEEtran}
\newlength{\bibitemsep}\setlength{\bibitemsep}{.5\baselineskip plus .05\baselineskip minus .05\baselineskip}
\newlength{\bibparskip}\setlength{\bibparskip}{0pt}
\let\oldthebibliography\thebibliography
\renewcommand\thebibliography[1]{%
\oldthebibliography{#1}%
\setlength{\parskip}{\bibitemsep}%
\setlength{\itemsep}{\bibparskip}%
}
|
\section{\label{sec:level1}Introduction}
Systems with strong spin-orbit coupling (SOC) effects have attracted significant attention in the field of spintronics, owing to their potential to generate a large spin current.
However, the spin of these systems is expected to relax quickly, owing to the SOC effect \cite{Elliott1954,Yafet1963,Emoto2016}.
Therefore, it is essential to clarify the criteria to obtain a long spin relaxation length to realize spintronics.
A potential method to evaluate the spin relaxation length is weak localization (WL) analysis using the quantum correction effect \cite{Bergman1982}.
It is well known that the quantum correction effect is described by the Hikami-Larkin-Nagaoka (HLN) theory \cite{HLN1980},
which has been widely used to evaluate the spin relaxation length in systems with strong SOC \cite{Assaf2013,Deorani2014,Peres2014}.
Crystal atoms that have a strong SOC can be called ``SOC lattices,'' which are different from the case in which impurities have a strong SOC. The SOC lattice system is described by the Hamiltonian equivalent of the Dirac electron system \cite{Wolff1964,Fuseya2015,Hayasaka2016}.
Thus far, the quantum correction effects have been investigated in many Dirac fermion systems such as graphene and surface states of topological insulators \cite{Suzuura2002, McCann2006, ShanWenYu2012, LuHaiZhou2011}. In these Dirac fermion systems, weak antilocalization (WAL) occurs owing to Berry phase $\pi$ effects.
A remarkable feature of an SOC lattice system is its intraband and interband spin hybridization owing to the SOC effect.
The effect of the band-spin hybridization changes the impurity scattering process compared with the case of free electrons.
Even if the impurity potential is diagonal for band and spin indices in the SOC lattice system, the matrix elements between eigenstates with different energies are non-zero elements.
According to Fermi's golden rule, transitions between states with different energies are forbidden by the energy conservation law.
However, in processes that involve higher-order scattering, such as the localization problem, non-adiabatic transitions with different energies are virtually allowed.
The importance of non-adiabatic transitions has been discussed in the context of the anomalous Hall effect \cite{Sinitsyn2007}.
When non-adiabatic transition processes are included, the understanding of the Berry phase based on the adiabatic picture does not hold, and thus,
it is not obvious whether the WAL completely disappears or partially remains. In addition, there is a lack of understanding of the difference between the spin relaxation length evaluated by the HLN theory and that based on the Dirac system.\\
\ \ The quantum correction effect is calculated by solving the Bethe-Salpeter equation, which involves correlations between two particles; thus, it has the square of the degrees of freedom of an individual particle.
When all transition processes are considered in the SOC lattice system, the Bethe-Salpeter equation becomes a ${\rm 16\times 16}$ matrix. Therefore, 16 Cooperons are naively expected to contribute to the quantum correction effect.
In such multiple-degree-of-freedom systems, understanding the experimental results of the quantum correction effects often requires a highly sophisticated interpretation \cite{Eda2016}. Alternatively, it relies on simplification, such as assuming only a single WAL channel by using the HLN formula \cite{Chen2011,Hirahara2014, Akiyama2016}. When a single WAL channel is used, the spin relaxation length is assumed to be sufficiently short. In this case, only the phase relaxation length can be experimentally obtained. \\
\ \ In this study, we consider non-adiabatic transitions in the quantum correction effect using the two-dimensional Wolff Hamiltonian, which is an effective model of the two-dimensional SOC lattice system, such as the {\it L}-point of Bi and PbTe \cite{Wolff1964,Fuseya2015,Hayasaka2016}.
We show that only intraband triplet and interband singlet Cooperons contribute to the quantum correction effect.
By incorporating virtual non-adiabatic transitions, the Cooperon contribution of the interband singlet that leads to WAL is suppressed.
In the weak magnetic field and WAL regimes, we show that the WAL effect increases with an increasing spin relaxation length, in contrast to the HLN theory.
We also show that the quantum correction effect on electrical conductivity clearly depends on the spin relaxation length when compared with the HLN theory.
We demonstrate the WL analysis in Bi thin films and obtain the spin relaxation and phase relaxation lengths.
\section{Model}
For a two-dimensional SOC lattice system, we consider the following Hamiltonian:
\begin{align}
{\cal H}={\cal H}_{0}+V(\bm{r}),
\end{align}
where ${\cal H}_{0}$ is the Wolff Hamiltonian, and $V(\bm{r})$ is the impurity potential. They are given by
\begin{align}
{\cal H}_{0}=\left[
\begin{array}{cc}
\Delta & i\hbar\gamma\bm{\sigma} \cdot \bm{k}\\
-i\hbar\gamma\bm{\sigma} \cdot \bm{k} &-\Delta\\
\end{array}
\right],
\end{align}
\begin{align}
V(\bm{r})=u_0\sum_{i}\delta(\bm{r}-\bm{R}_{i}),
\end{align}
where $\hbar$ is Planck's constant, $\bm{k}=(k_{x}, k_{y})$ is the wavenumber vector, $2\Delta$ is the band gap, $\gamma$ is the band parameter, $\sigma$ is the Pauli matrix, $u_{0}$ is the strength of the impurity potential, and $\bm{R}_{i}$ is the impurity position.
The basis of the Wolff Hamiltonian is $\{\ket{c\uparrow},\ket{c\downarrow},\ket{v\uparrow},\ket{v\downarrow}\}$,
where $\up,\dn$ are the spin degrees of freedom, and $c,v$ are the conduction and valence band degrees of freedom, respectively.
The energy eigenvalues of the Wolff Hamiltonian are $\pm E_{\bm{k}}=\pm\sqrt{\Delta^2+\gamma^2 k^2}$.
For simplicity, we consider the impurity potential to be diagonal for band and spin indices.
The plane wave solutions for $+E_{\bm{k}}$ are
\begin{align}
\ket{1,\bm{k}}
&=
\frac{N_{\bm{k}}}{2}\left(
\begin{array}{c}
1\\
e^{i\phi_{\bm{k}}}\\
-iY\\
-iYe^{i\phi_{\bm{k}}}
\end{array}
\right),
\end{align}
\begin{align}
\ket{2,\bm{k}}
&=
\frac{N_{\bm{k}}}{2}\left(
\begin{array}{c}
-e^{-i\phi_{\bm{k}}}\\
1\\
-iYe^{-i\phi_{\bm{k}}}\\
iY
\end{array}
\right),
\end{align}
where $Y=\hbar\gamma|\bm{k}|/(E_{\bm{k}}+\Delta)$, $N_{\bm{k}}=\sqrt{(\Delta+E_{\bm{k}})/(2E_{\bm{k}})}$,
$k_x=|\bm{k}|{\rm cos}\phi_{\bm{k}}$, and $k_y=|\bm{k}| {\rm sin}\phi_{\bm{k}}$. The plane wave solutions for $-E_{\bm{k}}$ are
\begin{align}
\ket{3,\bm{k}}
&=
\frac{N_{\bm{k}}}{2}\left(
\begin{array}{c}
-iY\\
-iYe^{i\phi_{\bm{k}}}\\
1\\
e^{i\phi_{\bm{k}}}
\end{array}
\right),
\end{align}
\begin{align}
\ket{4,\bm{k}}
&=
\frac{N_{\bm{k}}}{2}\left(
\begin{array}{c}
-iYe^{-i\phi_{\bm{k}}}\\
iY\\
-e^{-i\phi_{\bm{k}}}\\
1
\end{array}
\right).
\end{align}
Considering the lowest order Born scattering by the impurity potential, the relaxation time is defined as
\begin{align}
\frac{1}{\tau}&=-2{\rm Im}\Sigma^{\rm R/A}(E_{F})=-2\sum_{\bm{k}'}\average{VG^{\rm R/A}_{0}(\bm{k}')V}_{\rm imp}\nonumber\\
&=2\pi nu_{0}^2\rho_{0},
\end{align}
where $n$ is the impurity concentration, $\rho_{0}$ is the density of states per spin degree of freedom, and $\average{\cdots}_{\rm imp}$ denotes the configuration average of impurities.
$G_{\alpha0}^{\rm R/A}(\bm{k})=(E_F-E_{\alpha{\bm{k}}}\pm i\delta)^{-1}$ is the non-perturbed single-particle Green's function, where
$E_{1\bm{k}}=E_{2\bm{k}}=E_{\bm{k}}$ and $E_{3\bm{k}}=E_{4\bm{k}}=-E_{\bm{k}}$.
Hereafter, we consider the positive and negative energy eigenstates. The Fermi energy is now sufficiently higher than $\tau^{-1}$, i.e., $E_{F}\tau\gg1$. Even under this condition, the electrons in the negative energy state contribute quantitatively to the quantum correction effect. The case where non-adiabatic transition processes appear is described in Appendix A.
\section{Quantum correction effects on electric and magnetic conductivity}
The effect of quantum correction on the electrical conductivity $\delta\sigma^{(0)}$ (Fig. \ref{Fig1}(a)) is given by \cite{FukuyamaRev1985,Efros_book1985,PALeeRev1985}:
\begin{figure}[t!]
\begin{center}
\includegraphics[scale=0.30, bb=150 50 650 680]{Fig1.eps}
\end{center}
\caption[Quantum correction effect on electrical conductivity.]{Feynman diagrams of the quantum correction effects on electrical conductivity. (a) Quantum corrections of electrical conductivity. (b), (c) Diagrams incorporating the effect of the corrections to the Cooperon. (d), (e) Vertex correction and Bethe-Salpeter equation, respectively.}
\label{Fig1}
\end{figure}
\begin{align}
\label{kubo}
\delta\sigma^{(0)}(L)&=\frac{e^2\hbar}{2\pi}\sum_{\bm{k}}\tilde{v}^x_{\alpha\omega}(\bm{k})\tilde{v}_{\xi\beta}^{x}(-\bm{k})\nonumber\\
&\times G^{\rm R}_{\alpha}(\bm{k})G^{\rm R}_{\beta}(-\bm{k})G^{\rm A}_{\omega}(\bm{k})G^{\rm A}_{\xi}(-\bm{k})\nonumber\\
&\times \sum_{\bm{q}}\Gamma^{\alpha\beta}_{\xi\omega}(\bm{k},\bm{k},\bm{q}),
\end{align}
where the Einstein summation convention for Greek indices is used.
$G^{\rm R/A}_{\alpha}(\bm{k})=(E_F-E_{\alpha\bm{k}}\pm\rm i\hbar/(2\tau))^{-1}$ denotes the impurity-averaged Green's function.
$\tilde{v}^{x}(\bm{k})$ is the velocity operator with a vertex correction of the form:
\begin{align}
\label{vertex}
&\tilde{v}^{x}_{\beta\alpha}(\bm{k})=v^{x}_{\beta\alpha}+\sum_{\bm{k}'}G^{A}_{\alpha'}(\bm{k}')G^{R}_{\beta'}(\bm{k}')\nonumber\\
&\times \average{\bra{\beta,\bm{k}}V(\bm{r})\ket{\beta',\bm{k}'}
\bra{\alpha', \bm{k}'}V(\bm{r})\ket{\alpha,\bm{k}}}_{\rm imp}\nonumber\\
&\times \tilde{v}^{x}_{\beta'\alpha'}(\bm{k}').
\end{align}
The bare velocity operator $\hat{v^{x}}$ is given by
\begin{align}
\hat{v^{x}}=\frac{1}{\hbar}\frac{\partial {\cal H}_{0}}{\partial k_{x}}=
\left[
\begin{array}{cc}
0 & i\gamma {\sigma_x}\\
-i\gamma{\sigma_x} & 0\\
\end{array}
\right].
\end{align}
Equation (\ref{vertex}) can be solved by assuming a solution of the following form: $\tilde{v}^{x}_{\beta\alpha}=\eta_{v}v^{x}_{\beta\beta}\delta_{\beta\alpha}$,
where $\eta_{v}=2\lambda^2/(\lambda^2+1)$ and $\lambda=E_{\rm F}/\Delta$.
The quantum correction effect is given by the divergent contribution of $\Gamma^{\alpha\beta}_{\xi\omega}(\bm{k},\bm{k},\bm{q})$ in the particle-particle ladder type scattering process, which causes Cooper instability.
$\Gamma^{\alpha\beta}_{\xi\omega}(\bm{k},\bm{k},\bm{q})$ is given as a solution to the following Bethe-Salpeter equation \cite{GarateIon2009,GarateIon2012}:
\begin{align}
\label{eq12}
\Gamma^{mn}_{m'n'}(\bm{q})&=nu_{0}^2\delta_{mn}\delta_{m'n'}\nonumber\\
&+nu_{0}^2\sum_{\bm{k},\ell,\ell'}G^{R}_{lm}(\bm{k})G^{A}_{l'm'}(\bm{q}-\bm{k})\Gamma^{ln}_{l'n'}(\bm{q}).
\end{align}
Here, we change the basis of $\Gamma^{\alpha\beta}_{\xi\omega}$ to $\Gamma^{mn}_{m'n'}(\bm{q})$ using the following relation:
\begin{align}
\label{ansatz}
\Gamma^{\alpha\beta}_{\xi\omega}(\bm{k}_{1},\bm{k}_2,\bm{q})=\sum_{n,n',m,m'}\braket{\beta,\bm{k}_{2}|n}\braket{\omega,\bm{q}-\bm{k}_{2}|n'}\nonumber\\
\braket{m|\alpha,\bm{k}_{1}}\braket{m'|\xi,\bm{q}-\bm{k}_{1}}\Gamma^{mn}_{m'n'}(\bm{q}).
\end{align}
The basis denoted by $\ket{n}$, $\ket{n'}$, $\ket{m'}$, $\ket{m'}\in \{1,2,3,4\}$ corresponds to $\{\ket{c\up}, \ket{c\dn}, \ket{v\up}, \ket{v\dn}\}$.
$G^{R/A}_{lm}(\bm{k})$ is given by
\begin{align}
\label{eq14}
G^{R/A}_{lm}(\bm{k})=\braket{\alpha,\bm{k}|m}G^{R/A}_{\alpha}(\bm{k})\braket{l|\alpha,\bm{k}}.
\end{align}
Details of the calculations are given in Appendix B.
the components of $\Gamma$ have the following form:
\begin{align}
\label{5510}
\Gamma^{11}_{11}+\Gamma^{22}_{22}&=\bra{c\up}\otimes\bra{c\up}\hat{\Gamma}\ket{c\up}\otimes\ket{c\up}\nonumber\\
&+\bra{c\dn}\otimes\bra{c\dn}\hat{\Gamma}\ket{c\dn}\otimes\ket{c\dn}\nonumber\\
&=\frac{32nu_{0}^2\lambda^2\pi}{(\lambda^4+\lambda^2+2)v_{F}^2\tau^2q^2+4(\lambda^2-1)},
\end{align}
\begin{align}
\label{5511}
\Gamma^{33}_{33}+\Gamma^{44}_{44}&=\bra{v\up}\otimes\bra{v\up}\hat{\Gamma}\ket{v\up}\otimes\ket{v\up}\nonumber\\
&+\bra{v\dn}\otimes\bra{v\dn}\hat{\Gamma}\ket{v\dn}\otimes\ket{v\dn}\nonumber\\
&=\frac{32nu_{0}^2\lambda^2\pi}{(\lambda^4+\lambda^2+2)v_{F}^2\tau^2q^2+4(\lambda^2-1)},
\end{align}
\begin{align}
\label{5512}
&\Gamma^{11}_{44}-\Gamma^{14}_{41}-\Gamma^{41}_{14}+\Gamma^{44}_{11}\nonumber\\
&=(\bra{c\up}\otimes\bra{v\dn}-\bra{v\dn}\otimes\bra{c\up})\hat{\Gamma}\nonumber\\
&\times(\ket{c\up}\otimes\ket{v\dn}-\ket{v\dn}\otimes\ket{c\up})\nonumber\\
&=\frac{8\pi nu_{0}^2\lambda^2}{(2\lambda^2-1)q^2\tau^2v_{F}^2+2},
\end{align}
\begin{align}
\label{5513}
&\Gamma^{22}_{33}-\Gamma^{23}_{32}-\Gamma^{32}_{23}+\Gamma^{33}_{22}\nonumber\\
&=(\bra{c\dn}\otimes\bra{v\up}-\bra{v\up}\otimes\bra{c\dn})\hat{\Gamma}\nonumber\\
&\times(\ket{c\dn}\otimes\ket{v\up}-\ket{v\up}\otimes\ket{c\dn})\nonumber\\
&=\frac{8\pi nu_{0}^2\lambda^2}{(2\lambda^2-1)q^2\tau^2v_{F}^2+2},
\end{align}
where $v_{F}$ is the Fermi velocity.
Here, we do not explicitly describe all the components because the other components of $\Gamma$ are zero, or they vanish in the summation in Eq. (\ref{kubo}).
Equations (\ref{5510}) and (\ref{5511}) are intraband triplets, and equations (\ref{5512}) and (\ref{5513}) are interband singlets.
\begin{table*}
\caption{Components of $\Gamma$ with Cooper instability.}
\label{table:Gamma34}
\centering
\begin{ruledtabular}
\begin{tabular}{clll}
& Triplet & Singlet& \\
\hline
SOC lattice &$\ket{c\uparrow}\otimes\ket{c\uparrow}$, $\ket{c\downarrow}\otimes\ket{c\downarrow}, $& $\ket{c\uparrow}\otimes\ket{v\downarrow}-\ket{v\downarrow}\otimes\ket{c\uparrow}$, & \\
(with non-adiabatic transitions) &$\ket{v\uparrow}\otimes\ket{v\uparrow}$, $\ket{v\downarrow}\otimes\ket{v\downarrow}$ & $\ket{c\downarrow}\otimes\ket{v\uparrow}-\ket{v\uparrow}\otimes\ket{c\downarrow}$ & \\
\\
SOC lattice \cite{Hayasaka2020} &$\ket{c\uparrow}\otimes\ket{c\uparrow}$, $\ket{c\downarrow}\otimes\ket{c\downarrow}$& $\ket{c\uparrow}\otimes\ket{v\downarrow}-\ket{v\downarrow}\otimes\ket{c\uparrow}$,\\
(without non-adiabatic transitions)& & $\ket{c\downarrow}\otimes\ket{v\uparrow}-\ket{v\uparrow}\otimes\ket{c\downarrow}$ & \\
\\
HLN \cite{HLN1980} & $\ket{c\uparrow}\otimes\ket{c\uparrow}$, $\ket{c\downarrow}\otimes\ket{c\downarrow}$, & $\ket{c\uparrow}\otimes\ket{c\downarrow}-\ket{c\downarrow}\otimes\ket{c\uparrow}$ & \\
($k_{z}\neq 0$) &$\ket{c\uparrow}\otimes\ket{c\downarrow}+\ket{c\downarrow}\otimes\ket{c\uparrow}$ & & \\
\end{tabular}
\end{ruledtabular}
\end{table*}
We obtain the quantum corrections to the electrical conductivity as follows:
\begin{align}
&\delta\sigma^{(0)}(L)=-\frac{e^2}{2\pi^2 \hbar}\sum_{i={cc,vv,s}}\alpha_i {\rm log}\frac{\ell_0^{-2}+\ell_{i}^{-2}}{L^{-2}+\ell_i^{-2}},
\end{align}
where $\alpha_{cc}=(\lambda+1)^2/(\lambda^4+\lambda^2+2)$, $\alpha_{vv}=(\lambda-1)^2/(\lambda^4+\lambda^2+2)$, $\alpha_{s}=-(\lambda^2-1)/[2(2\lambda^2-1)]$,
$\ell_{cc}^{-2}=\ell_{vv}^{-2}=2(\lambda^2-1)/(\lambda^4+\lambda^2+2)\ell_{0}^{-2}$, $\ell_{s}^{-2}=2/(2\lambda^2-1)\ell_{0}^{-2}$, $\ell_{0}=\sqrt{D_{0}\tau}$, and $D_0=v_{F}^2\tau/2$.
$cc$, $vv$, and $s$ represent conduction-intraband triplets, valence-intraband triplets, and an interband singlet, respectively.
Only intraband triplets and the interband singlet, which is qualitatively equivalent to previous results that do not consider non-adiabatic transitions \cite{Hayasaka2020}, remain.
However, valence-intraband triplets ($\ket{v\up}\otimes\ket{v\up}$ and $\ket{v\dn}\otimes\ket{v\dn}$) also make a divergent contribution.
Table \ref{table:Gamma34} summarizes the components of $\Gamma$ with Cooper instability.
Notably, the intraband triplet ($\ket{c\up}\otimes\ket{c\dn}+\ket{c\dn}\otimes\ket{c\up}$) and intraband singlet ($\ket{c\up}\otimes\ket{c\dn}-\ket{c\dn}\otimes\ket{c\up}$) do not appear, even when transitions between all energy eigenstates are considered.
The absence of the intraband triplet and intraband singlet may be a property that is specific to the exact two-dimensional system with $k_{z}=0$.
Even in the HLN theory with $k_{z}=0$, the intraband singlet and intraband triplet cancel each other out and do not contribute to the quantum correction effect of the electrical conductivity. In the case of the SOC lattice system with $k_{z}=0$, the same cancelation is expected to occur.
It should be noted that in Ref. \cite{GarateIon2012}, the negative energy states are neglected.
If we ignore the negative energy state when solving Eq. (\ref{eq12}) as in Ref. \cite{GarateIon2012}, the result of Ref. \cite{Hayasaka2020} is reproduced.
Using the lowest-order Cooperon correction, the quantum correction effect on the electrical conductivity (Fig. \ref{Fig1}(b, c)) can be expressed as follows:\begin{align}
\label{556}
&\delta\sigma^{(1)}_{a}=\frac{e^2\hbar}{2\pi}\sum_{\bm{k},\bm{k}_1,\bm{q}}\tilde{v}_{\alpha'\omega}^x(\bm{k})\tilde{v}_{\xi\beta'}^x(-\bm{k}_1)\nonumber\\
&\times G_{\alpha'}^{R}(\bm{k})G_{\alpha}^{R}(\bm{k}_1)G_{\beta}^{R}(-\bm{k})G_{\beta'}^{R}(-\bm{k}_1)\nonumber\\
&\times G_{\xi}^{A}(-\bm{k}_1)G_{\omega}^{A}(\bm{k})
\Gamma^{\alpha\beta}_{\xi\omega}(\bm{q})\nonumber\\
&\average{\braket{\alpha, \bm{k}_1|V(\bm{r})|\alpha',\bm{k}}\braket{\beta',-\bm{k}_1|V(\bm{r})|\beta,-\bm{k}}}_{\rm imp}.
\end{align}
\begin{align}
\label{557}
&\delta\sigma^{(1)}_{b}=\frac{e^2\hbar}{2\pi}\sum_{\bm{k},\bm{k}_1,\bm{q}}\tilde{v}^x_{\alpha\omega'}(\bm{k})\tilde{v}^x_{\xi'\beta}(-\bm{k}_{1})\nonumber\\
&\times G_{\alpha}^{R}(\bm{k})G_{\beta}^{R}(-\bm{k}_1)
G_{\xi'}^{A}(-\bm{k}_1)G_{\xi}^{A}(-\bm{k})\nonumber\\
&\times G_{\omega}^{A}(\bm{k}_1)G_{\omega'}^{A}(\bm{k})
\Gamma^{\alpha\beta}_{\xi\omega}(\bm{q})\nonumber\\
&\times \average{\braket{\xi, -\bm{k}, |V(\bm{r})|\xi', -\bm{k}_1}\braket{\omega', \bm{k}|V(\bm{r})|\omega, \bm{k}_1}}_{\rm imp}.
\end{align}
The quantum correction $\delta\sigma_{\rm W}$ to the electrical conductivity, considering all contributions from Fig. 1(a), (b), and (c), becomes
\begin{align}
\delta\sigma_{\rm W}&=\delta\sigma^{(0)}+\delta\sigma^{(1)}_a+\delta\sigma^{(1)}_b\nonumber\\
&=-\frac{e^2}{2\pi^2\hbar}\eta_{v}^2\sum_{i=cc,vv,s}(1+2\eta_{H,i})\alpha_i {\rm log}\frac{\ell_0^{-2}+\ell_{i}^{-2}}{L^{-2}+\ell_i^{-2}}.
\end{align}
where $\eta_{H,cc}=-(\lambda-1)/(4\lambda)$, $\eta_{H,vv}=-(\lambda+1)/(4\lambda^2)$ and $\eta_{H,s}=-1/4$.
The electrical conductivity in a magnetic field $\delta\sigma_{\rm W}(B)=\delta\sigma^{(0)}(B)+\delta\sigma^{(1)}_a(B)+\delta\sigma^{(1)}_b(B)$ is obtained as follows:
\begin{align}
\label{eqMC}
\delta&\sigma_{\rm W}(B)=-\frac{e^2}{2\pi^2\hbar}\eta_{v}^2\sum_{i=cc,vv,s}\alpha_{i}(1+2\eta_{H,i})\nonumber\\
&\times\Biggl[\psi\left(\frac{1}{2}+\frac{\ell_{B}^2}{\ell_{0}^2}+\frac{\ell_{B}^2}{\ell_i^2}+\frac{\ell_{B}^2}{\ell_{\phi}^2}\right)-\psi\left(\frac{1}{2}+\frac{\ell_{B}^2}{\ell_i^2}+\frac{\ell_{B}^2}{\ell_{\phi}^2}\right)\Biggr],
\end{align}
where $\psi$ is the digamma function and $\ell_{B}=\sqrt{\hbar/4eB}$ is the magnetic length of the electron pair.
The magnetic field dependence of the quantum correction effect on the electrical conductivity is plotted in Fig. \ref{fig:MC}. By changing $E_{\rm F}/\Delta$, a WL-WAL crossover occurs.
\begin{figure}[t!]
\begin{center}
\includegraphics[scale=0.33, bb=150 50 650 550]{Fig2.eps}
\end{center}
\caption[Magnetic field dependence of $\delta \sigma_{\rm W}$]{Magnetic field dependence of $\Delta\sigma(B)=[\delta\sigma_{\rm W}(B)-\delta\sigma_{\rm W}(0)]/(e^2/2\pi^2\hbar)$. }
\label{fig:MC}
\end{figure}
\begin{figure}[t!]
\begin{center}
\includegraphics[scale=0.33, bb=150 50 650 550]{Fig3.eps}
\end{center}
\caption{Dependence of $\alpha_{cc}$, $\alpha_{vv}$, and -$\alpha_{s}$ on $E_{F}/\Delta$. }
\label{alpha5}
\end{figure}
The WL-WAL crossover is characterized by the strength of the intraband triplet $\alpha_{cc}+\alpha_{vv}$ and interband singlet $\alpha_{s}$. Figure 3 shows the dependence of $\alpha$ on $E_{\rm F}/\Delta$.
The conduction-intraband triplet $\alpha_{cc}$ and interband singlet $\alpha_{s}$ dominate for $E_{\rm F}/\Delta\sim 1$ and $E_{\rm F}/\Delta\rightarrow\infty$, respectively.
The valence-intraband triplet $\alpha_{vv}$ exhibits a gentle peak in the intermediate $E_{\rm F}/\Delta$ region.
Consequently, the WL-WAL crossover occurs at $E_{F}/\Delta \sim 3$. This value agrees with previous results \cite{Hayasaka2020}.
The interband singlet $\alpha_{s}$ becomes $1/4$ as $ E_{\rm F}/\Delta \rightarrow \infty$. This value is smaller than $\alpha_{s}=1/2(E_{\rm F}/\Delta \rightarrow \infty)$ when only adiabatic processes are considered \cite{Hayasaka2020}; thus, the WAL effect is suppressed.
This can be intuitively understood from the matrix elements of impurity scattering.
During the transition of an electron from the $\ket{1,\bm{k}}$ state to the $\ket{1,\bm{k}'}$ state, owing to impurities,
the matrix elements of the impurity scattering become
\begin{align}
\label{20}
\braket{1,\bm{k}'|V(\bm{r})|1,\bm{k}}\propto(1+Y^2)(1+e^{i(\phi-\phi')}).
\end{align}
The backscattering process can be obtained by making the following substitution: $\bm{k}'\rightarrow-\bm{k}$, that is, $\phi'\rightarrow\phi+\pi$; thus, equation (\ref{20}) becomes zero. This is similar to what is well known for graphene and the surface states of topological insulators, which indicates that the backscattering process is suppressed \cite{Ando1998, Suzuura2002, LuHaiZhou2011, ShanWenYu2012}. Thus, this process contributes to the WAL.
For the transition from $\ket{1,\bm{k}}$ to $\ket{2,\bm{k}'}$, the following relationship holds:
\begin{align}
\label{21}
\braket{2,\bm{k}'|V(\bm{r})|1,\bm{k}}\propto(1-Y^2)(e^{i\phi'}-e^{i\phi}).
\end{align}
For $E_{\rm F}/\Delta\rightarrow\infty$, $Y\rightarrow 1$; thus, this process also contributes to the WAL in the large $E_{\rm F}/\Delta$ region.
In contrast, the non-adiabatic transition process from $\ket{1,\bm{k}}$ to $\ket{4,\bm{k}'}$ becomes
\begin{align}
\label{23}
\braket{4,\bm{k}'|V(\bm{r})|1,\bm{k}}\propto 2iYe^{i\phi'}-2iYe^{i\phi}.
\end{align}
Equation (\ref{23}) does not contribute to the WAL as an adiabatic process (Equations (\ref{20}) and (\ref{21})).
Therefore, when this process is considered, the sum of all transition probabilities is conserved, which weakens the effect of WAL compared with the case for which this process is not considered.
\section{Spin relaxation length}
\begin{figure}[t!]
\begin{center}
\includegraphics[scale=0.33, bb=150 50 650 550]{Fig4.eps}
\end{center}
\caption[Comparison of $\delta\sigma_{\rm W}$ and $\delta\sigma_{\rm HLN}$]{Dependence of $\Delta\sigma(B)=[\delta\sigma(B)-\delta\sigma(0)]/(e^2/2\pi^2\hbar)$ on the magnetic field. The solid lines correspond to $\delta\sigma_{\rm W}$, and the dotted line corresponds to $\delta\sigma_{\rm HLN}$. In $\delta\sigma_{\rm W}$, $E_{F}/\Delta=10$, $\ell_{\phi}=1000{\rm nm}$, and ${\ell_{0}=20,25,30,35,40{\rm nm}}$. We used $\ell_{{\rm sf},{\rm W}}\sim\ell_{0}/\sqrt{2}$. For $\delta\sigma_{\rm HLN}$, we set $\ell_{\phi}=300{\rm nm}$, $\ell_{0}=20{\rm nm}$, $\ell_{\rm so}=20,40{\rm nm}$, and we used $\ell_{{\rm sf},{\rm HLN}}=\sqrt{3}/2\ell_{\rm so}$.}
\label{WvsH}
\end{figure}
In this section, we compare the evaluation of the spin relaxation length with that of the HLN theory.
The spin relaxation length can be estimated by fitting the formula for the quantum correction effect on the electrical conductivity in the weak-field region ($\ell_{B}\gg\ell_{0}$).
In the SOC lattice system, the coupling of spin and momentum results in a simultaneous relaxation of the spin with the relaxation of the momentum.
Therefore, the spin relaxation length is related to $\ell_{0}$, and it can be evaluated as $\ell_{{\rm sf},{\rm W}}\sim\ell_{0}/\sqrt{2}$.
$\sqrt{2}$ arises from doubling of the spin. The HLN formula is given by \cite{HLN1980}:
\begin{align}
\label{NormalHLN}
&\delta\sigma_{\rm HLN}(B)=-\frac{e^2}{2\pi^2\hbar}\Biggl[\frac{3}{2}\Biggl\{\Psi\Bigg(\frac{1}{2}+\frac{\ell_B^2}{\ell_0^2}+\frac{\ell_B^2}{\ell_{\rm so}^2}+\frac{\ell_B^2}{\ell_{\phi}^2}\Biggr)\nonumber\\
&-\Psi\Bigg(\frac{1}{2}+\frac{4}{3}\frac{\ell_B^2}{\ell_{\rm so}^2}+\frac{\ell_B^2}{\ell_{\phi}^2}\Biggr)\Biggr\}
-\frac{1}{2}\Biggl\{\Psi\Biggl(\frac{1}{2}+\frac{\ell_B^2}{\ell_{0}^2}+\frac{\ell_B^2}{\ell_{\rm so}^2}+\frac{\ell_B^2}{\ell_{\phi}^2}\Biggr)\nonumber\\
&-\Psi\Biggl(\frac{1}{2}+\frac{\ell_B^2}{\ell_{\phi}^2}\Biggr)\Biggr\}\Biggr].
\end{align}
The relation between the spin relaxation time and the spin-orbit relaxation time is given by $1/\tau_{s}=4/(3\tau_{\rm so})$ \cite{DasSarma2004}.
Therefore, the spin relaxation length in the HLN theory is given by $\ell_{{\rm sf},{\rm HLN}}=\sqrt{3}/2\ell_{\rm so}$.
Figure 4 presents a plot of $\delta\sigma_{\rm W}$ and $\delta\sigma_{\rm HLN}$.
Compared with $\delta\sigma_{\rm HLN}$, $\delta\sigma_{\rm W}$ shows a sharp change in the WAL effect with respect to the change in the spin relaxation length.
This is because the instability of the interband singlet, which leads to the WAL effect, is essentially inseparable from the SOC effect.
In fact, the interband singlet contains information about the spin relaxation through $\ell_{0}$.
In the case of the HLN theory, the intraband singlet is not affected by the SOC effect. As a result, the intraband singlet does not contain the spin relaxation length, but only the phase relaxation length $\ell_\phi$.
Therefore, if $\ell_\phi$ is constant, $\delta\sigma_{\rm HLN}$ slightly changes in the weak-field region with respect to the change in the spin relaxation length.
A conventional WL analysis using the HLN theory suggests that the WAL effect can be observed only when the spin relaxation length is sufficiently short.
Furthermore, as previously mentioned, the conductivity of the HLN theory is almost independent of the spin relaxation length in the weak magnetic field and the WAL regime; hence, the quantum correction effect was analyzed using only the intraband singlet \cite{Chen2011,Hirahara2014, Akiyama2016}. In this analysis, only the phase relaxation length can be obtained. However, the WAL effect in the SOC lattice system is enhanced by an increasing spin relaxation length, and it is more sensitive to changes in the spin relaxation length than $\delta\sigma_{\rm HLN}$.
Therefore, $\delta\sigma_{\rm W}$ can extract more rich information than $\delta\sigma_{\rm HLN}$.
\section{WL analysis}
\begin{figure}[t!]
\begin{center}
\includegraphics[scale=0.32, bb=150 50 720 600]{Fig5.eps}
\end{center}
\caption[Fitting of experimental data]{The temperature dependence of conductivity in the WAL regime for a 20 bilayer (BL) Bi film (These values are extracted from Ref. \cite{Hirahara2014}). Solid lines represent $3\times( \delta\sigma_{\rm W}(B)-\delta\sigma_{\rm W}(0))$. }
\label{WLanalysis}
\end{figure}
Finally, we perform the WL analysis for Bi thin films . According to the recent experiments of Aitani et al., the transport properties of 20 bilayer (BL) Bi thin films are dominated by bulk properties rather than surface states \cite{Hirahara2014}. We focus on the 20 BLs to ignore the effect of the surface states.
The Fermi energy and band gap are given by $E_{F}=35.3\ {\rm meV}$ and $\Delta=7.7\ {\rm meV}$, respectively \cite{Fuseya2012}. Therefore, $\lambda=4.6$.
As Bi has three equivalent electron surfaces \cite{Kamran2011}, we can assume that these contributions are additive and multiply the coefficient of $\delta \sigma_{\rm W}$ by three.
In the fitting procedure, a somewhat technical method is used owing to the complex parameter space resulting from the nonlinear functions \cite{Peres2014}.
We use the following procedure to reduce this difficulty: We assume that the phase relaxation length is infinite and determine the spin relaxation length at 0.8 K. As impurity scattering rather than electron-lattice scattering is dominant at low temperatures, the phase relaxation length is used as a fitting parameter for 1.4 K--4.0 K and used as the fixed value of the spin relaxation length obtained at 0.8 K.
Note that in Ref. \cite{Hirahara2014}, although the magnet conductivity is studied up to 0.2 T, the classical contribution $\propto B^2$ overlaps with the diffusive contribution at a higher magnetic field \cite{Assaf2013}.
To neglect this contribution, we limited the range up to 0.03 T in the fitting procedure.
The results of the fitting and the fitting parameters are shown in Fig. 5 and Table \ref{table2}, respectively.
In this manner, the spin relaxation length can be determined as $\ell_{\rm sf,W}=44.9\ {\rm nm}$ in the 20 BL Bi film.
\begin{table}
\caption{spin relaxation and phase relaxation lengths.}
\label{table2}
\centering
\begin{ruledtabular}
\begin{tabular}{clll}
& $\ell_{\rm sf,W}$ & $\ell_{\phi}$ & \\
\hline
0.8 K & 44.9 nm & $ \infty $ & \\
1.4 K & & 336 nm & \\
2.0 K & & 300 nm & \\
2.8 K & & 183 nm & \\
3.3 K & & 129 nm & \\
4.0 K & & 112 nm & \\
\end{tabular}
\end{ruledtabular}
\end{table}
\section{Conclusion}
We investigated the quantum correction effect based on an effective model of the SOC lattice system.
We showed that the WAL effect is suppressed when non-adiabatic transitions are considered, compared with the case in which only adiabatic transitions are considered.
We found that only intraband triplets and interband spin singlets contribute to Cooper instability, even if non-adiabatic transitions are included.
This significantly simplifies the interpretation of the experimental results of the quantum correction effects in SOC lattice systems.
The WAL effect in the SOC lattice system is sensitive to changes in the spin relaxation length, and it increases for longer spin relaxation lengths in contrast to the HLN theory. We expect that our results on the quantum correction effect will be useful for the quantitative evaluation of the spin relaxation length in SOC lattice systems.
\begin{acknowledgments}
The author would like to thank Y. Fuseya for commenting on the manuscript.
\end{acknowledgments}
|
\section{Introduction}
Segmentation tasks based on remotely-sensed imagery rely heavily on the availability of good quality semantic or
polygon labels to be useful \citep{wang:weaklysupervised2020, zhu:pickqualityeval2019}.
Specifically, \cite{ahn:semanticaffinity2018} notes that the performance of ConvNets for segmentation tasks
is largely dependent on the amount of time spent creating training labels. These tasks can be as simple as instance counting
\cite{li:oilpalm2017,neupane:bananacount2019, osco:citruscount2020,yao:treecount2021} or object identification \citep{cha:crackdetection2017,gulgec:damagedetect2017}.
However, in practice these labels tend not to be readily available outside of established tasks
such as building footprint extraction or vehicle detection, so it is often necessary
to spend significant manual effort to digitize these polygons \citep{wang:weaklysupervised2020}.
This drastically increases the cost to build segmentation models for practical applications, and
presents a major barrier for smaller organizations.
Several approaches have been proposed in response to this issue such as weak-label learning
\citep{ahn:interpixel2019,laradji:pointsuper2019,paul:domainadapt2020, wang:weaklysupervised2020}
and self-supervision via a contrastive objective in Siamese or Triplet networks
\citep{caron:emerging2021, dhere:selfsupseg2021, jaiswal:surveyselfsup2020,leyva:contrastopt2021}.
However, these approaches come with their own problems. Models based on weak
labels for instance are challenging to evaluate since ground truth polygons
are not necessarily meaningful labels.
In the case of aerial imagery, a model which fixates on salient features of an object such as a
chimney or a wind shield should not be penalized for having a non-intuitive interpretation of what
best defines a building or a vehicle.
This paper contributes a novel approach to semantic segmentation with weak labels by incorporating
constraints in the segmentation model $S$ through an adversarial objective,
consisting of an object localization discriminator $D_1$ and a contextual discriminator $D_2$.
In our framework, $D_1$ discriminates between the true image $I_R$ and a fake
image $I_{F_1}$, where $I_{F_1}$ is a spatially similar sample in the training set with
the positive object superimposed to it. By treating the segmentation model as
a generator, and implementing an additional discriminator, we propagate learnable
gradients back to the generator. This allows the segmentation model to produce masks
which when used to superimpose the original image into neighboring contexts, produce
believable “generated” images. The full formulation of the discriminator and the superimpose function
$f$ is discussed in Section \ref{subsection:adversarial}. $D_2$ is an additional contextual discriminator
which we introduce in order to constrain artifacts that arise from having only one discriminator.
We show that this method produces high quality segmentations without having to consider imagery
gradients and uses only buffered object polygon centroids as inputs. The major contribution of this paper is a new form of
semi-supervised segmentation model that can be applied to segment objects using weak labels, such as those
provided by manual annotation or object detection models which only provide bounding boxes. Future extensions
of this work will address the localization task, providing end-to-end instance segmentation from weak labels and
low requirements for the number of annotations to reach acceptable performance.
\section{Related Work}
\begin{figure}[!t]
\begin{center}
\setlength{\tabcolsep}{1pt}
\begin{tabular}{ccccccc}
\includegraphics[width=.14\linewidth]{figs/bangers/planes/l23_P2P/Ir_6.png} &
\includegraphics[width=.14\linewidth]{figs/bangers/planes/l23_P2P/y_tilde_6.png} &
\includegraphics[width=.14\linewidth]{figs/bangers/planes/l210_UNET101/y_hat_4.png} &
\includegraphics[width=.14\linewidth]{figs/bangers/planes/l211_UNET50/y_hat_4.png} &
\includegraphics[width=.14\linewidth]{figs/aircraft_unet50.png} &
\includegraphics[width=.14\linewidth]{figs/bangers/planes/l23_P2P/y_hat_7.png} &
\includegraphics[width=.14\linewidth]{figs/bangers/planes/l23_P2P/y_6.png} \\
\includegraphics[width=.14\linewidth]{figs/bangers/woolsey/l13_P2P/Ir_2.png} &
\includegraphics[width=.14\linewidth]{figs/bangers/woolsey/l13_P2P/y_tilde_2.png} &
\includegraphics[width=.14\linewidth]{figs/bangers/woolsey/l110_UNET18/y_hat_1.png} &
\includegraphics[width=.14\linewidth]{figs/bangers/woolsey/l111_UNET50/y_hat_1.png} &
\includegraphics[width=.14\linewidth]{figs/woolsey_unet50.png} &
\includegraphics[width=.14\linewidth]{figs/bangers/woolsey/l13_P2P/y_hat_2.png} &
\includegraphics[width=.14\linewidth]{figs/bangers/woolsey/l13_P2P/y_2.png} \\
(a) & (b) & (c) & (d) & (e) & (f) & (g) \\
\end{tabular}
\end{center}
\caption[Predicted segmentation masks from the Aircraft and Woolsey datasets]{
Predicted segmentation masks from the Aircraft and Woolsey datasets using fully-supervised
UNet models with ResNet backbones (c-e) and our semi-supervised P2P model (f). All generated masks
come from training with Dice loss, besides (e), which is trained with cross entropy loss for comparison.
(a) Input image to each model. (b) Pseudo label $\tilde{y}$ used to train P2P. (c) Output from
a UNet-101 baseline model. (d) Output from UNet-50. (e) Output from UNet-50 trained with cross-entropy.
(f) Output from P2P (ours). (g) Ground Truth labels.
\label{fig:bangers} }
\end{figure}
Points2Polygons shares ideas in common with four general types of segmentation models which
we will discuss here:
\paragraph{Footprint extraction models}
We make a reference to building footprint extraction models because these models
aim to extract (either via semantic or instance segmentation) complex geometries
that come from a much more complex distribution space - in terms of shape and texture variance -
than features such
as vehicles or trees \citep{bittner:footprintdsm2016, chawda:extractfootprints2018,zhu:mapnet2019}.
Building footprints are also a class of features which are
conditionally dependent on contextual information; a building is much more likely
to be found in a suburban context next to trees or driveways as opposed to in
the desert.
Existing approaches however rely heavily on segmentation labels
\citep{chawda:extractfootprints2018, zhu:mapnet2019} and primarily focus on improving
the accuracy from a supervised learning approach given large amounts of
training data, such as in the SpaceNet \citep{vanetten:spacenet2018} and WHU \citep{xia:dota2017} datasets.
This is unfortunately not readily available for many worthwhile segmentation tasks
such as well pad extraction for localizing new fracking operations or airplane detection
for civil and remote sensing applications.
\paragraph{Semi-supervised or unsupervised methods}
Other methods such as \citep{ahn:semanticaffinity2018, ahn:interpixel2019, laradji:pointsuper2019, paul:domainadapt2020,wang:weaklysupervised2020}
and \citep{dai:boxsup2015} attempt to address the scarcity of existing ground truth labels by
adopting a semi-supervised or unsupervised approach. These include methods which
attempt to perform segmentation based on a single datapoint as demonstrated in \cite{wang:weaklysupervised2020}.
However approaches such as these only factor in low-frequency information and are designed for
landcover classification. \citet{laradji:pointsuper2019} utilizes point-based
pseudo labels to segment objects of interest in imagery using a network with a localization
branch and an embedding branch to generate full segmentation masks. While similar in nature to P2P,
\citep{laradji:pointsuper2019} optimizes a squared exponential distance metric between same-class
pixels in embedding space to generate segmentation masks, whereas P2P frames the segmentation task
as a simple image-to-image translation problem with localization constraints
enforced by an adversarial objective. Our approach thus circumvents the need for pair-wise
calculations from a similarity function with the additional benefit
of a separately tunable constraint.
\cite{NEURIPS2019_32bbf7b2} attempts to solve segmentation problems in a completely unsupervised setting using
an adversarial architecture to segment regions that can be clipped and redrawn using a generator network such that
the generated images is aligned with the original. P2P goes further by using easy-to-obtain supervised labels that
consist of points within individual instances.
An alternate approach to weak supervision for segmentation is to utilize image-level
class labels in conjunction with class attention maps \citep{ahn:interpixel2019,oquab:freelocalization2015}.
Besides the low labeling cost, the benefit of this approach is that the class
boundary for the object of interest is informed by contextual information in the rest of the image, in contrast
to pixel-level supervision, where the model is encouraged to learn shape and
color cues in the object of interest.
P2P incorporates image-level information by learning characteristics of an object's context via
a second contextual discriminator, $D_2$, as described in Section \ref{subsection:adversarial}.
\paragraph{Adversarial methods}
One of the primary difficulties of learning segmentation from weak labels is
that it is difficult to evaluate the quality of the segmentation model in the
absence of ground-truth polygons.
Traditional methods tackle this challenge by introducing priors or
domain-specific constraints, such as in superpixel-based models \citep{csillik:slicremotesensing2017, fiedler:powerslic2020, hartley:superpixelcnn2019,ren:gslicr2018}.
However this approach is difficult to generalize to cases where the landscape changes
or in the presence of covariate shift introduced via a new dataset.
In these cases, it may be necessary to adjust model hyperparameters, such as is commonly done in
experiments with production-level segmentation pipelines
in order to compensate for these new domains which in itself can be very time-intensive.
We draw inspiration from works that leverage Generative Adversarial Networks (GANs) which aim to introduce a
self-imposed learning objective via a discriminator $D$ which models and attempts to
delineate the difference in distribution between a real and fake input set. In particular,
\citet{luc:semseg2016} uses an adversarial network to enforce higher-order consistency
in the segmentation model such that shape and size of label regions are considered, while
\citet{souly:semigan2017} use GANs to create additional training data.
This is a form of self-supervision which we are able to incorporate into P2P. By introducing
discriminators, we get past the need to impose strong human priors and instead allow
the model to determine whether or not the segmentation output is “realistic”.
In effect, our optimization objective enforces contextual consistency
through a $\min\max$ formulation. We additionally modify the vanilla GAN
objective in order to prevent high model bias,
which we outline under Contextual Similarity in Section \ref{subsection:context}
\section{Methodology}
\begin{figure}[!t]
\centering
\begin{tikzpicture}[scale=.70]
\begin{scope}[main node/.style={thick,draw}]
\node [main node](y_tilde) at (0,0) {$\tilde{y}$};
\node [style={circle,thick,draw}](G) at (2, 0) {$S$};
\node [main node](y_hat) at (4.25, 1) {$\hat{y}$};
\node [main node](y_prime_c) at (4.25, -1) {$\hat{y}^c$};
\node [style={circle,thick,draw}](D1) at (9.25, 2) {$D_1$};
\node [style={circle,thick,draw}](D2) at (9.25, -2) {$D_2$};
\node [main node](I_R) at (2, 2) {$I_{R}$};
\node [main node](I_F1) at (7.5, 1) {$I_{F_1}$};
\node [main node](I_F2) at (7.5, -1) {$I_{F_2}$};
\node [main node](x_bar)at (2, -2) {$I_{ctx}$};
\node (f1) at (6, 1) {$f$};
\node (f2) at (6, -1) {$f$};
\end{scope}
\begin{scope}[>={Stealth[black]},
every node/.style={fill=white,circle},
every edge/.style={draw=black,very thick}]
\path [->] (y_tilde) edge (G);
\path [->] (G) edge (y_hat);
\path [->] (I_F1) edge (D1);
\path [->] (y_hat) edge (y_prime_c);
\path [->] (y_hat) edge (f1);
\path [->] (f1) edge (I_F1);
\path [->] (I_R) edge [bend left] (D1);
\path [->] (I_R) edge (f1);
\path [->] (I_R) edge (f2);
\path [->] (y_prime_c) edge (f2);
\path [->] (x_bar) edge [bend right](D2);
\path [->] (x_bar) edge (f2);
\path [->] (x_bar) edge (f1);
\path [->] (f2) edge (I_F2);
\path [->] (I_F2) edge (D2);
\path [->] (I_R) edge (G);
\end{scope}
\end{tikzpicture}
\caption{Our modified adversarial objective. Our segmentation model
$S$ takes in pseudo label $\tilde{y}$ and $I_R$ and returns a predicted
segmentation mask $\hat{y}$, which is then used to generate fake images
$I_{F_1}$ and $I_{F_2}$, which are used to train the discriminators $D_1$
and $D_2$. See Section \ref{subsection:adversarial} for more details on this
procedure.}
\end{figure}
\subsection{Adversarial Objective}\label{subsection:adversarial}
This is the vanilla GAN objective:
\begin{equation}\label{gan_objective}
\min_{G} \max_{D} \loss(G, D) = \EX_{x \sim p_{data}}
[\log D(x)] + \EX_{z \sim p(z)}[\log(1 - D(G(z)))]
\end{equation}
We modify the vanilla GAN objective by introducing an additional
discriminator which forces the generator to include features which contribute
more to the object of interest than its background. P2P uses an adversarial objective using two discriminators, $D_1$ and $D_2$.
The first discriminator $D_1$, discriminates between the real input
$I_{R}$ and a fake variant of the input, $I_{F_1}$, which consists of the object of
interest superimposed to a spatially similar training sample, $I_{ctx}$.
The second discriminator $D_2$ uses a similar approach, but with a
different fake image, $I_{F_2}$. For this discriminator, the negative area
around the object of interest is used to create the fake image.
Fake images used in our GAN formulation are generated by combining the predicted
segmentation mask $\hat{y}$ with our original image $I_R$ and context image $I_{ctx}$
through a process equivalent to alpha blending \citep{porter:alphablend1984}.
Formally, the process to generate a fake sample is
\begin{equation}
f(I, I_{ctx}, y) = y \odot I + y^c \odot I_{ctx}
\end{equation}
where $I, I_{ctx} \in T$, $\odot$ is the Hadamard product and $y$ is a
segmentation mask. Both $I$ and $I_{ctx}$ are samples from the training dataset $T$, where
$I_{ctx}$ is found by iterating through adjacent tiles to $I$ until a tile
without a training label appears, and is chosen to be $I_{ctx}$. Figure
\ref{fig:contextop} describes the procedure to find $I_{ctx}$ in more detail.
In P2P's case, $I \equiv I_R$ is the "real" input image fed into the discriminators
and $\hat{y}$ is used in place of $y$ for generating $I_{F_1}$, while $\hat{y}^c = (1 - \hat{y})$ is used
for generating $I_{F_2}$. Therefore, the fake images are generated using
\begin{align}
I_{F_1} &= f(I_R, I_{ctx}, \hat{y}) = \hat{y} \odot I_R + \hat{y}^c \odot I_{ctx}\\
I_{F_2} &= f(I_R, I_{ctx}, \hat{y}^c) = \hat{y}^c \odot I_R + \hat{y} \odot I_{ctx}
\end{align}
The operation $f(\cdot)$ used to generate the fake examples can be thought of as a superimpose operation
that uses the predicted segmentation mask to crop the object of interest (or its background)
and paste it in a nearby training sample.
Given the above definitions for our real and fake images for $D_1$ and $D_2$, we arrive at the
following objective:
\begin{align}
\min_G\max_D\loss(G, D) &= \min_G\max_D \EX_{\chi \sim p_{data}}
(2 \log(1 - D_1(G_1)) + 2 \log(1 - D_2(G_2)) \nonumber \\
&+ \log(D_1(I_R)) + \log(D_2(I_{ctx})) + \loss_{loc} ) \label{eq:ganobjective}
\end{align}
where $\chi = (I_R, I_{ctx})$, $D = (D_1, D_2)$ and $G = (G_1, G_2)$, $G_1 = f(\hat{y},I_R, I_{ctx})$,
and $G_2 = f(\hat{y}^c,I_R, I_{ctx})$. Note that $G_1$ and $G_2$ share weights;
they are both functions of the segmentation model $S(I_R, \tilde{y}) = \hat{y}$,
although $G_2$ uses the complement of $\hat{y}$, namely $\hat{y}^c$, to generate fake samples.
The loss function $\loss(G,D)$ for P2P is the sum of $D_1$ and $D_2$'s binary cross entropy loss,
a generator loss $\loss_G$, and a localization loss $\loss_{loc} = \min(\tilde{y}\cdot \log(\hat{y}) + (1 - \tilde{y})
\cdot \log(1-\hat{y}), \rho)$ that is included in $\loss_G$.
The localization loss term is a thresholded version of binary cross entropy,
where $\rho$ is the threshold parameter that controls the amount of
influence that $\loss_{loc}$ has on the overall loss function. This parameter is chosen with
a hyperparameter search, specifically ASHA \citep{li:asha2018}. For more details on how the
objective function is derived, please see the Appendix.
\subsection{Contexts} \label{subsection:context}
\begin{figure}[t!]
\centering
\begin{tikzpicture}[scale=0.65]
\begin{axis}[hide axis, xmin=0.0, xmax=7.0, ymin=0.0, ymax=7.0]
\draw[help lines] (axis cs:0, 0) grid [step = 1] (axis cs:7,7);
\draw[cyan!65, thick] (3, 6) grid [step=1] (5, 7);
\draw[cyan!65, thick] (1, 4) grid [step=1] (3, 5);
\draw[cyan!65, thick] (2, 3) rectangle (3, 4);
\draw[cyan!65, thick] (5, 3) rectangle (6, 4);
\draw[cyan!65, thick] (1, 2) rectangle (2, 3);
\draw[cyan!65, thick] (4, 2) grid [step=1] (6, 3);
\draw[orange, thick] (1, 6) grid [step=1] (3, 7);
\draw[orange, thick] (5, 6) rectangle (6, 7);
\draw[orange, thick] (1, 5) grid [step=1] (6, 6);
\draw[orange, thick] (4, 4) grid [step=1] (6, 5);
\draw[orange, thick] (1, 3) rectangle (2, 4);
\draw[orange, thick] (3, 3) grid [step=1] (5, 4);
\draw[orange, thick] (2, 2) grid [step=1] (4,3);
\draw[violet!85, thick] (3, 4) rectangle (4, 5);
\draw [fill=orange!25, fill opacity=0.4, draw=none, thick] (1,6) rectangle (3,7);
\draw [fill=orange!25, fill opacity=0.4, draw=none, thick] (5,6) rectangle (6,7);
\draw [fill=orange!25, fill opacity=0.4, draw=none, thick] (1,5) rectangle (6,6);
\draw [fill=orange!25, fill opacity=0.4, draw=none, thick] (4,4) rectangle (6,5);
\draw [fill=orange!25, fill opacity=0.4, draw=none, thick] (1,3) rectangle (2,4);
\draw [fill=orange!25, fill opacity=0.4, draw=none, thick] (3,3) rectangle (5,4);
\draw [fill=orange!25, fill opacity=0.4, draw=none, thick] (2,2) rectangle (4,3);
\draw [fill=cyan!25, fill opacity=0.6, draw=none, thick] (3,6) rectangle (5,7);
\draw [fill=cyan!25, fill opacity=0.6, draw=none, thick] (1,4) rectangle (3,5);
\draw [fill=cyan!25, fill opacity=0.6, draw=none, thick] (2,3) rectangle (3,4);
\draw [fill=cyan!25, fill opacity=0.6, draw=none, thick] (5,3) rectangle (6,4);
\draw [fill=cyan!25, fill opacity=0.6, draw=none, thick] (1,2) rectangle (2,3);
\draw [fill=cyan!25, fill opacity=0.6, draw=none, thick] (4,2) rectangle (6,3);
\draw [fill=violet!65, fill opacity=0.4, draw=none, thick] (axis cs:3,4) rectangle (axis cs:4,5);
\draw[thick, white] (0,0) rectangle (7, 7);
\draw[thick, white] (1,2) rectangle (6, 7);
\draw[ultra thick, white] (2,3) rectangle (5, 6);
\draw (0,0) rectangle (7, 7);
\draw[thick] (1, 2) rectangle (6, 7);
\draw[thick, dashed] (axis cs:2, 3) rectangle (axis cs:5, 6);
\addplot[only marks, thick, mark=*, mark size =5.5pt, color = gray, fill = gray!20] table[x=x,y=y,col sep=comma]{data/context_data.csv};
\end{axis}
\end{tikzpicture}
\caption{Contexts (blue tiles) are discovered by expanding out from an origin tile, denoted in violet.
Origin tiles are tiles containing positive samples (orange) which are indexed during preprocessing
then sampled at random during training. We first index the surrounding 8 tiles from the origin.
Positive tiles $I_R$ are denoted in orange while contexts $I_{ctx}$ (negative tiles) are denoted in blue.
If the number of contexts do not meet the minimum context number, we continuously
expand the search space until a sufficient number of positive tiles are met.}
\label{fig:contextop}
\end{figure}
To support the generator in producing realistic fake outputs without being
overpowered by the discriminator, we need to place these outputs in contexts that
are realistic and do not expose information which might signal to the discriminator
that these outputs are fake. As an example we do not want to superimpose objects against a purely black or noisy background. However we do want to place
these buildings in a suburban neighborhood with trees and roads. We
define the term context as an image tile in close proximity to a tile in question
which also does not contain a positive sample.
P2P requires each context to be extracted from a positive tile’s spatial neighborhood
to maintain semantic similarity with the rest of the inputs. We provide users the
flexibility to choose the number of contexts that are aggregated into each batch at
training time. This ensures a diverse selection of background characteristics for the
superimpose function. To facilitate dynamic tile fetching during training without
heavy computational overheads, we map each positive tile to its neighboring contexts
during a preprocessing step (see Figure \ref{fig:contextop}). The output of this step is a dictionary
that maps each positive tile to a set number of neighboring contexts which is then
used by P2P to quickly access these tiles for training.
\subsection{Object Localization}
Since the discriminators do not actually care about whether or not the extracted
features exhibit sufficient localization behavior, only that these outputs look
“realistic”, we reinforce the generator with a localization objective. This
objective makes use of the input point labels, which we term pseudo labels, and creates a buffer around these
points against which the output segmentation masks are evaluated. In particular,
we try to minimize the ratio between the amount of positive pixels in the output
segmentation mask and this buffer. We also set a minimum threshold to how low
this ratio can go to prevent the model from overweighting the localization objective.
\section{Implementation}
\begin{figure}[!t]
\centering
\begin{tikzpicture}[node distance=1cm]
\node (f1) {$f_1:$};
\node (y_hat) [label=below:$\hat{y}$, right of=f1] {\includegraphics[width=.1\textwidth]{figs/superimpose/y_hat.PNG}};
\node (mult) [right of=y_hat]{$\odot$};
\node (I_R) [label=below:$I_R$, right of=mult] {\includegraphics[width=.1\textwidth]{figs/superimpose/Ir.PNG}};
\node (plus) [right of=I_R] {$+$};
\node (y_hat_c) [label=below:$\hat{y}^c$, right of=plus] {\includegraphics[width=.1\textwidth]{figs/superimpose/y_hat_c.png}};
\node (mult2) [right of=y_hat_c]{$\odot$};
\node (I_ctx) [label=below:$I_{ctx}$, right of=mult2] {\includegraphics[width=.1\textwidth]{figs/superimpose/Ictx.PNG}};
\node (equals) [right of=I_ctx] {$=$};
\node (I_F1) [label=below:$I_{F_1}$, right of=equals] {\includegraphics[width=.1\textwidth]{figs/superimpose/F1.PNG}};
\end{tikzpicture}
\\
\begin{tikzpicture}
\node (f2) {$f_2:$};
\node (y_hat_prime) [label=below:$\hat{y}^c$, right of=f2] {\includegraphics[width=.1\textwidth]{figs/superimpose/y_hat_c.png}};
\node (mult) [right of=y_hat_prime]{$\odot$};
\node (I_R) [label=below:$I_R$, right of=mult] {\includegraphics[width=.1\textwidth]{figs/superimpose/Ir.PNG}};
\node (plus) [right of=I_R] {$+$};
\node (y_hat) [label=below:$\hat{y}$, right of=plus] {\includegraphics[width=.1\textwidth]{figs/superimpose/y_hat.PNG}};
\node (mult2) [right of=y_hat]{$\odot$};
\node (I_ctx) [label=below:$I_{ctx}$, right of=mult2] {\includegraphics[width=.1\textwidth]{figs/superimpose/Ictx.PNG}};
\node (equals) [right of=I_ctx] {$=$};
\node (I_F2) [label=below:$I_{F_2}$, right of=equals] {\includegraphics[width=.1\textwidth]{figs/superimpose/F2.PNG}};
\end{tikzpicture}
\caption{The general procedure for generating fake images from an input sample $I_R$ and
predicted segmentation mask $\hat{y}$. The GAN training procedure is designed to generate segmentation
masks $\hat{y}$ that create believable ``fake'' examples.}
\end{figure}
In preprocessing, we divide up the input rasters into multiple image chips. For
each positive image chip, we search its 8 neighboring contexts in order to index
any negative chips (chips which do not contain positive samples). If these cannot
be found in the immediate contexts, we expand our search to the next 16 contexts
and the next 32 contexts etc. For every positive chip we store a number of its
negative neighbors (up to $k$ neighbors) in the context dictionary.
In the training step, we jointly train the segmentation model (generator)
along with two discriminators.
We sample a batch of positive chips and feed them to the generator to produce
output masks. These are then used to ``transplant'' the positive features from
the original positive context into neighboring negative contexts to produce
fake positives. Both the real and fake positives are fed into $D_1$ to evaluate
the positive transplants.
To prevent the generator from overpowering the discriminator using salient
but ill-segmented features, we also produce negative contexts. The real
negative contexts are the original neighbors while the fake negative contexts
are created by transplanting features from the positive context to its
neighboring negatives using the inverse of the output mask. These are fed
to $D_2$.
Finally at inference time, we simply extract the raw outputs of the generator.
\section{Experiments}
\begin{table}[b!]
\caption{Fully-supervised benchmark comparison. }
\label{benchmark_comparisons}
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{l|llll|llll|llll|llll|}
\toprule
& \multicolumn{4}{c}{Dice} & \multicolumn{4}{c}{Jaccard} & \multicolumn{4}{c}{Recall} & \multicolumn{4}{c}{Precision}\\
\midrule
Model & \rotate{Well Pads} & \rotate{Airplanes} & \rotate{Woolsey} & \rotate{SpaceNet}
& \rotate{Well Pads} & \rotate{Airplanes} & \rotate{Woolsey} & \rotate{SpaceNet}
& \rotate{Well Pads} & \rotate{Airplanes} & \rotate{Woolsey} & \rotate{SpaceNet}
& \rotate{Well Pads} & \rotate{Airplanes} & \rotate{Woolsey} & \rotate{SpaceNet} \\
\midrule
UNet-50 & $0.33$ & $0.65$ & $0.46$ & $0.47$
& $0.20$ & $0.49$ & $0.31$ & $0.31$
& $\textbf{1.00}$ & $\textbf{1.00}$ & $\textbf{0.99}$ & $\textbf{1.00}$
& $0.20$ & $0.49$ & $0.31$ & $0.31$\\
UNet-101 & $0.21$ & $0.46$& $0.57$ & $0.47$ & $0.12$ & $0.30$ & $0.41$
& $0.32$ & $\textbf{1.00}$ & $\textbf{1.00}$ & $0.85$ & $\textbf{1.00}$ & $0.12$ & $0.30$ & $0.44$ & $0.32$\\
UNet-50 (CE) & $\textbf{0.80}$ & $\textbf{0.90}$ & $\textbf{0.81}$ & $\textbf{0.90}$
& $\textbf{0.68}$ & $\textbf{0.82}$ & $\textbf{0.70}$ & $\textbf{0.82}$
& $0.90$ & $0.86$ & $0.80$ & $0.89$
& $\textbf{0.75}$ & $\textbf{0.94}$ & $\textbf{0.85}$ & $\textbf{0.92}$\\
P2P (ours) & $0.65$ & $0.64$ & $0.66$ & $0.62$ & $0.49$
& $0.48$ & $0.51$ & $0.46$ & $0.93$ & $0.67$ & $0.74$ & $0.60$ & $0.52$ & $0.64$ & $0.64$ & $0.68$ \\
\bottomrule
\end{tabular} }
\end{table}
\subsection{Datasets}
In our experiments, we make use of five datasets to train P2P and assess its performance.
This includes two small datasets for well pads and airplanes as described in the Appendix,
and one medium dataset for building footprint extraction which we
sourced from USAA. We also train and benchmark our model against the SpaceNet Challenge
AOI2 Las Vegas dataset \citep{vanetten:spacenet2018}. Additionally, as part of our ablative studies we also
create a smaller version of this dataset which we name Vegas Lite.
\paragraph{Well pad dataset}
We construct a well pad dataset using NAIP: Natural Color imagery at 1.5m resolution.
This dataset is created using ArcGIS Pro and contains 9 training rasters and 1 testing
raster. There are 2909 training samples and 426 testing samples, which makes this a
small-to-medium sized dataset. The labels are bounding geometries of visible well pads,
which contain both active and disused examples.
\paragraph{Airplane dataset}
We construct an airplane dataset using NAIP: Natural Color imagery at 0.5m resolution.
This dataset is created using ArcGIS Pro and contains 2 training rasters and 1 testing
raster. Given there are 335 training samples and 40 testing samples, this constitutes
a small dataset and what can be realistically expected in a real-world proof-of-concept
application. The labels incorporate a variety of plane sizes from 3 locations in the
United States: Southern California Logistics Airport, Roswell International Air
Center and Phoenix Goodyear Airport.
\paragraph{Woolsey Fire dataset}
We make use of the DataWing imagery from USAA that was captured after the 2018
Woolsey Fires for damage assessment. These are 0.3m resolution rasters with
8682 training samples distributed across 6 training rasters and 572 samples on
1 testing raster. This is a medium sized dataset and contains a large number of
negative samples (of damaged houses) in a variety of terrains.
\paragraph{SpaceNet Challenge: Las Vegas dataset}
In order to benchmark our model against popular datasets for remote sensing,
we make use of the SpaceNet Las Vegas (AOI2) dataset \citep{vanetten:spacenet2018}. Both training and
testing sets are publicly available however the testing set does
not contain ground truth geometries. We therefore further split the training
set into 2905 image tiles for training (tiles numbered 1325 and above) and 712 image
tiles for testing (tiles up to, but not including, 1325).
\paragraph{Vegas Lite dataset}
We further reduce the size of the Las Vegas dataset in order to perform ablative
tests on a smaller set of samples. This we call the Vegas Lite dataset and is simply
the full Las Vegas Dataset without any tiles numbered 3000 upwards in the training
set and 500 upwards in the testing set. This dataset contains 970 tiles in the
training set and 268 tiles in the testing set. These are roughly 33\% and 38\%
the size of the original training and testing sets respectively, assuming samples
are equally distributed across tiles.
We use the Vegas Lite dataset to understand the importance of contextual similarity
between positive and negative chips. We do so by introducing a series of negative
rasters with varying contextual similarities in our ablative tests.
\subsection{Performance Study}
We evaluate our model based on several real-world feature segmentation tasks
in order show the efficacy of our proposed approach.
\paragraph{P2P vs Baseline Comparison}
We compare P2P which uses weak labels $\tilde{y}$ against 3 fully supervised segmentation
models which make use of ground truth polygon labels $y$. Among baselines trained with Dice loss, we find
that UNet-50 produces the best results on all except the well pads dataset,
where UNet-18 performs the best out of the baseline models. In the case of well pad extraction, an increase in the
number of parameters in the model results in lower performance, suggesting that
a low-complexity problem such as well pad extraction could be over-parameterized in this
approach. A considerable improvement is made to the performance of baseline models with full labels
when trained with cross-entropy loss, which produces better learnable gradients for these datasets.
A survey of the visual outputs show that these baselines generally produce segmentation
masks that are significantly larger than the object outlines which is not the
case for P2P. We further clarify that these P2P results consider the optimal
localization guidance parameters, however it is not difficult to select
a uniform guidance parameter that would outperform all baselines (7000 for
example).
\subsection{Ablative Tests}
\paragraph{Negative Discriminator}
We explore the impact of a second negative discriminator on testing accuracy
through all 4 datasets. We define this second discriminator $D_2$ as a UNet-50 (same
as $D_1$) binary classifier which predicts the probability that its input comes
from the distribution of real negatives. Negative samples associated with a
specific positive sample are drawn from the positive sample’s neighboring contexts.
This is done during preprocessing as discussed in Section 3.4. We then employ the context
shift function $f$ to superimpose the positive sample $I_R$
to its context $I_{ctx}$ using the inverse of the labels. This creates our fake negative sample $I_{F_2}$,
which we jointly feed to $D_2$ along with the real context $I_{ctx}$.
In experimentation, we find that implementing $D_2$ significantly improves model
performance across all but one of the datasets. In the well pads
dataset, we find that introducing $D_2$ incurs a small performance cost. This is
possibly due to the simple nature of the segmentation task where having additional
tunable parameters potentially lead to model overfitting.
By observing the quantitative and visual outputs we identify that the lack of $D_2$
generally leads to segmentation masks with much lower precision and higher recall.
\begin{table}[h!]
\caption{Negative Discriminator $D_2$ ablative study.}
\label{table:d2_ablative}
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{l|lll|lll|lll|lll|}
\toprule
& \multicolumn{3}{c}{Dice} & \multicolumn{3}{c}{Jaccard} & \multicolumn{3}{c}{Recall} & \multicolumn{3}{c}{Precision}\\
\midrule
Model & \rotate{Well Pads} & \rotate{Airplanes} & \rotate{Woolsey}
& \rotate{Well Pads} & \rotate{Airplanes} & \rotate{Woolsey}
& \rotate{Well Pads} & \rotate{Airplanes} & \rotate{Woolsey}
& \rotate{Well Pads} & \rotate{Airplanes} & \rotate{Woolsey} \\
\midrule
P2P w/o $D_2$
& $\textbf{0.67}$ & $0.23$& $0.43$
& $\textbf{0.51}$ & $0.13$ & $0.28$
& $0.92$ & $\textbf{0.96}$ & $\textbf{0.90}$
& $\textbf{0.54}$ & $0.13$ & $0.30$ \\
P2P w/ $D_2$
& $0.65$ & $\textbf{0.64}$ & $\textbf{0.66}$
& $0.49$ & $\textbf{0.48}$ & $\textbf{0.51}$
& $\textbf{0.93}$ & $0.67$ & $0.74$
& $0.52$ & $\textbf{0.64}$ & $\textbf{0.64}$ \\
\bottomrule
\end{tabular} }
\end{table}
\paragraph{Localization Guidance}
We show that a purely unsupervised approach without any guidance on object localization
does not produce good results in practice. In order to guide the
model during training, we create a buffer around object centroids at training time
and incur a penalty on the generator whenever predicted masks do not cover these
buffered pixels. We then threshold this loss to prevent this localization loss
from overwhelming the generator. A failure case would be the model treating the
buffered centroids as the ground truth labels under a fully supervised objective.
By thresholding, we ensure that the model is only concerned with localization at
a rough scale.
To qualify the optimal buffer size for a variety of datasets we trained P2P using
a variety of preset buffer sizes. This is a variable we term centroid
size multiplier (csm). The exact formulation for localization loss and additional visual results can be found
in the Appendix.
We also recognize that a shortfall of this method is that it is less suitable for
detecting objects that do not share the same size. This is of lesser concern for
remote sensing use cases but should be taken into consideration when applying this
model to general-purpose segmentation tasks.
\begin{table}[h!]
\caption{Localization Guidance ablative study for various $\textmd{csm}$ values.}
\label{table:localization_ablative}
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{l|lll|lll|lll|lll|}
\toprule
& \multicolumn{3}{c}{Dice} & \multicolumn{3}{c}{Jaccard} & \multicolumn{3}{c}{Recall} & \multicolumn{3}{c}{Precision}\\
\midrule
Model (csm) & \rotate{Well Pads} & \rotate{Airplanes} & \rotate{Woolsey}
& \rotate{Well Pads} & \rotate{Airplanes} & \rotate{Woolsey}
& \rotate{Well Pads} & \rotate{Airplanes} & \rotate{Woolsey}
& \rotate{Well Pads} & \rotate{Airplanes} & \rotate{Woolsey} \\
\midrule
P2P ($6000$) & $0.27$ & $0.15$& $0.04$
& $0.16$ & $0.09$ & $0.02$
& $0.72$ & $0.17$ & $0.02$
& $0.17$ & $0.18$ & $0.17$ \\
P2P ($7000$) & $0.55$ & $\textbf{0.64}$ & $\textbf{0.66}$
& $0.39$ & $\textbf{0.48}$ & $\textbf{0.51}$
& $\textbf{0.87}$ & $0.67$ & $0.74$
& $0.42$ & $0.64$ & $0.64$ \\
P2P ($8000$) & $0.64$ & $0.54$ & $0.54$
& $0.48$ & $0.38$ & $0.38$
& $0.84$ & $0.78$ & $0.49$
& $0.54$ & $0.42$ & $0.64$ \\
P2P ($9000$) & $0.07$ & $0.56$ & $0.64$
& $\textbf{0.50}$ & $0.39$ & $0.48$
& $0.78$ & $\textbf{0.86}$ & $\textbf{0.82}$
& $0.60$ & $0.43$ & $0.55$ \\
P2P ($10000$) & $\textbf{0.65}$ & $0.51$ & $0.65$
& $0.49$ & $0.35$ & $0.50$
& $0.52$ & $0.37$ & $0.62$
& $\textbf{0.93}$ & $\textbf{0.86}$ & $\textbf{0.74}$ \\
\bottomrule
\end{tabular} }
\end{table}
\paragraph{Contextual Similarity}\label{subsection:contextual_ablative}
In order to understand the importance of applying semantically coherent
negative contexts to $f$, we try out different types of context images in
training P2P on the Vegas Lite dataset. We compare the performance of the
Original Context image (an image that is spatially similar to real image
tiles) with transformed context rasters that consist of zero values (Blank
Context), chromatic aberrations (Red Context) and random rasters drawn from
a Gaussian distribution (Noise Context). We note that that the introduction
of strong signals indicative of contexts such as in the case of Red Context
results in poor model performance whereas spatially similar contexts produce
the highest Dice/Jaccard scores. One outlier example is the Noise Context,
which produces equivalently good Dice/Jaccard scores as the Original but instead
has much lower precision and higher recall. This is indicative of high model bias.
We further highlight the similarity between this and removing the contextual discriminator
which also results in lower precision and higher recall. Noisy contexts can therefore be
interpreted as effectively weakening the learning signal from the contextual discriminator.
\begin{table}[h!]
\caption{Contextual similarity ablative study using only the Vegas Lite dataset.}
\label{table:context_ablative}
\centering
\begin{tabular}{l|l|l|l|l}
\toprule
Transformation & Dice & Precision & Recall & Jaccard\\
\midrule
Original Context & $\textbf{0.66}$ & $\textbf{0.79}$ & $0.58$ & $\textbf{0.50}$ \\
Blank Context & $0.60$ & $0.56$ & $0.70$ & $0.44$ \\
Red Context & $0.56$ & $0.57$ & $0.61$ & $0.40$ \\
Noise Context & $\textbf{0.66}$ & $0.57$ & $\textbf{0.84}$ & $\textbf{0.50}$ \\
\bottomrule
\end{tabular}
\end{table}
\section{Conclusion \& Future Work}
In this paper, we introduced Points2Polygons, a semantic segmentation model that uses an adversarial
approach consisting of two discriminators to learn segmentation masks using weak labels.
We show that a generative adversarial network can be used to perform segmentation
tasks with good performance across a wide variety of small datasets where only weak
point labels are provided by introducing the contextual superimpose function.
Future work includes evaluating our weak label approach to other forms of imagery
outside of remote sensing. The remote sensing perspective on this problem is important
since spatially neighboring
tiles can be used to construct new contexts to improve the models robustness, but perhaps the same
technique could be applied in different settings. While this approach is powerful by itself to produce
segmentation using simple labels, it would be more powerful when combined in a object detection setting
to provide instance segmentation with weak labels.
\begin{ack}
We would like to thank our colleague Mansour Raad for his domain expertise that made this research possible, as well as
Omar Maher, Ashley Du and Shairoz Sohail for reviewing and providing comments on a draft of this paper.
\end{ack}
|
\section{Conclusion}
\label{sec:conclusion}
In this work, we investigate the advantageous diversity of the ensemble model against adversarial attacks. By studying the robustness of ensemble DNNs and the Pearson correlation coefficient among models trained with filters, we propose the "minimum correlation coefficients" principle for choosing filters, which is instrumental in building the ensemble defense.
Beyond existing ensemble defenses, we consider the diversity of ensemble models with a new perspective. We obtain the diversity from the filtered training data and confirm it experimentally. We observe that our ensemble model without adversarial information is more robust against adversarial attacks than adversarial training models.
Our discovery not only contributes to proposing a decent robust ensemble model but also supplies data diversity. As our future work, it is interesting to study further how much robustness we could gain from data diversity and model ensemble. We are also considering extending our framework to larger datasets like ImageNet and training our sub-models using different network structures.
\section{Experimental Evaluation}
\label{sec:experiment}
\old{In this section, the experimental results are provided to demonstrate the performance of our method against adversarial perturbations. We evaluate the correlation between filters stated in Sect.~\ref{Sect:Correlation}. The adversarial examples of the network trained with original data are generated and tested on the network trained with filtered data to measure the transferability empirically. We choose two filters accordingly for the ensemble model and compare the adversarial robustness with the previous ensemble defense and the adversarial training.}
\hw{In this section, we demonstrate the experimental results to support our inference and our method. We first give a brief introduction of the experimental settings, and then show our study of the correlation between filters in Section~\ref{Sect:Correlation}. We measure the robustness of our models with respect to transferability, via calculating the accuracy of our models on adversarial examples produced by attacking the original network. In the end, two front filters are chosen by `minimum correlation coefficients' to constitute the ensemble defense. We compare its robustness with adversarial training. We generate adversarial examples under various attacks implemented in FoolBox~\cite{rauber2017foolbox,rauber2017foolboxnative}. The FoolBox version is 3.31, and the license is MIT license. All experiments are conducted on a Windows 10 laptop with Intel i7-9750H, GTX 2060, and 16G RAM.}
\paragraph{Dataset and Network} We train our models on the CIFAR-10~\cite{krizhevsky2009learning} dataset under the MIT license. The CIFAR-10 dataset contains $60\,000$ RGB images in total with ten exclusive classes. We train our ResNet18~\cite{he2016deep} models on the $50\,000$ training images and test them on the $10\,000$ testing images. We use the stochastic gradient descent optimizer for training with $0.1$, $0.01$, and $0.001$ as the learning rate successively.
\subsection{The Optimal Filter Combination for Ensemble}
\old{Our principle for choosing filters is to pick the least correlated filters with the minimum correlation coefficients. We first calculate the correlation coefficients for each pair of the candidate filters and then evaluate the sub-models trained with different filters using the transfer-based attack.}
\hw{We analyze the Pearson correlation coefficients for each pair of the candidate filters and pick the least correlated filters for the ensemble with the minimum correlation coefficients. We also evaluate the robustness of sub-models on adversarial examples produced on the original network.}
\old{\paragraph{Statistical Correlation Analysis for Filters} We apply noise of size $\epsilon \leq 20/255$ to 100 images randomly picked from the test set and measure the correlation using the mentioned Pearson correlation coefficient~\eqref{eq:corre}. The result is shown in Figure~\ref{fig:correlation}. The high-pass filtered data is most correlated to the original data, and the grayscale filtered is the second. It infers that the adversarial example of the network trained with the original data may be easy to transfer to the network trained with these filtered data. The downsizing filter is strongly correlated to the low-pass filter, implying that it is probable to attack these two filters together. It is interesting to see that the 16 colour reduction filter shows little correlation to the low-pass filter, and they have a relatively low correlation to the original data. It gives us the confidence to use the low-pass and the 16 colour reduction filter with the original data to build a robust ensemble model.}
\hw{\paragraph{Statistical Correlation Analysis for Filters} We apply noise of size $\epsilon \leq 20/255$ to $100$ images randomly picked from the test set and evaluate Pearson correlation coefficient according to~\eqref{eq:corre}.
As reported by Figure~\ref{fig:correlation}, the correlation coefficient between the high-pass filtered data and the original inputs is the largest, i.e., $0.90$. The grayscale filtered data gets $0.47$, the second to the original inputs. It indicates that adversarial examples produced on the original network easily transfer to the sub-models with these filters. The downsizing filter
strongly correlates With the low-pass filter, implying that they are probably deceived by the same adversarial examples.
The $16$ color reduction filter shows little correlation, i.e., $0.02$, to the low-pass filter, and they both have a relatively low correlation to the original data, i.e., $0.13$ and $0.30$. According to minimum correlation coefficients, the robust ensemble model includes the original network and the two sub-models trained with the low-pass filter and the $16$ color reduction filter.}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.5\textwidth]{fig/fig_correlation_tri.eps}
\caption{The Pearson correlation coefficient between the sensitivity of different filtered data. The low-pass and the $16$ color reduction filter is marked yellow since they have the smallest coefficient.}
\label{fig:correlation}
\end{figure}
\old{\paragraph{Transfer-based Attack Analysis} We generate adversarial examples of the original network under the FGSM and the PGD attack with different distances and test them on the network trained with filtered data. The result is in Fig.~\ref{fig:transfAtt}. The outcome shows that the low-pass, colour quantization, and downsizing filters have a comparably better defence against the transfer-based attack. On the contrary, the grayscale and the high-pass filters are vulnerable to the transfer-based attack. So the corresponding two filters are not competitive to other filters since we want to build a robust ensemble model. The result of the transfer-based attack agrees with the previous correlation analysis.}
\hw{\paragraph{Transfer-based Attack Analysis} We generate adversarial examples against the original network by FGSM and PGD attacks with different values of attacking radius $\epsilon$ and test them on the sub-models trained with filtered data. In Figure~\ref{fig:transfAtt}, the sub-models with the low-pass, color quantization, and downsizing filters perform better than the original network against both FGSM and PGD attacks. The accuracy of these sub-models remains above $50\%$ when the attacking radius is $20/255$ under the FGSM attack. The PGD attack is more powerful against the original network and drops its accuracy to nearly $0$. The sub-model with a downsizing filter has the lowest accuracy among the three sub-models, which is $62.39\%$ under the PGD attack. However, the sub-models with the grayscale and the high-pass filters are vulnerable to the transfer-based attack. These sub-models have lower accuracy than the original network under the FGSM attack with $\epsilon \geq 10/255$. It is consistent with our analysis in Section~\ref{Sect:Correlation}, which suggests that these two filters should not be part of the ensemble. }
\old{
\begin{figure}[htbp]
\centering
\subfloat{
\includegraphics[width=0.46\linewidth]{fig/fig_FGSM.eps}}
\qquad
\subfloat{
\includegraphics[width=0.46\linewidth]{fig/fig_PGD.eps}}
\caption{Transfer Attck by FGSM (left) and PGD (right) }
\label{fig:transfAtt}
\end{figure}
}
\hw{
\begin{figure}[htbp]
\centering
\begin{tabular}{c c}
\includegraphics[width=0.46\linewidth]{fig/fig_FGSM_2.eps} &
\includegraphics[width=0.46\linewidth]{fig/fig_PGD_2.eps}\\
(a) FGSM & (b) PGD
\end{tabular}
\caption{The transferablity of the adversarial examples against the original network generated by FGSM (left) and PGD (right). Note that the accuracy of the original network is near $0$ under the PGD attack.}
\label{fig:transfAtt}
\end{figure}
}
\old{\subsection{Comparison with Different Ensemble Methods}
According to the results of the correlation analysis and the transfer-based attack, we choose the 16 colour reduction and the low-pass filter for the ensemble model. We also include a model trained with original data for the ensemble to maintain state-of-the-art accuracy on clean data. In this section, we compare the adversarial robustness of our ensemble model with the previous ensemble defence in \cite{ensembledefense} and the worst-case ensemble in our setting. We choose the BPDA attack based on the BIM attack to attack the ensemble model with filters. The number of iterations is set to be $20$, and the step size is equal to $\epsilon/10$. The result is shown in Fig.~\ref{fig:otherensemble}.}
\hw{\subsection{Comparison with Different Ensemble Methods}
In this section, we compare the adversarial accuracy of our ensemble model with different ensemble models. The details of each ensemble model are as follows:
\begin{itemize}
\item \textbf{Minimum correlated Ensemble} According to statistical correlation analysis for filters, we choose the $16$ color reduction filter and the low-pass filter for the ensemble, which have the lowest correlation. The ensemble includes the original network to maintain state-of-the-art accuracy on clean data.
\item \textbf{Maximum correlated Ensemble} The worst-case situation suggested by statistical correlation analysis is to constitute the ensemble with the original network and the two sub-models with the high-pass and grayscale filters, whose correlation is the highest as shown in Figure~\ref{fig:correlation}.
\item \textbf{Gaussian Noise Ensemble~\cite{strauss2017ensemble}} The ensemble of models trained with Gaussian noise is the most robust model compared to the other ensemble methods~\cite{strauss2017ensemble}. We train three sub-models with Gaussian noise to build the ensemble model.
\end{itemize}
We compare the minimum correlated model to the Gaussian noise ensemble model and the maximum correlated ensemble model. We choose the BPDA attack based on the BIM to attack ensemble models. We use the sum of the gradient of sub-models to attack the ensemble model as a whole. The number of iterations is $20$, and the step size is $\epsilon/10$. Hereafter, the vote-based ensemble follows the voting mechanism described in Section~\ref{sec:method}, and the score-based ensemble outputs the class with the maximum average score.}
\hw{According to Figure~\ref{fig:otherensemble}(a), when the disturbance is $5/255$, $10/255$, $15/255$, and $20/255$, the score-based accuracy of the minimum correlated ensemble model is $27.52\%$, $22.4\%$, $14.33\%$, and $8.59\%$ higher than the Gaussian noise model, and is $25.54\%$, $15.46\%$, $9.53\%$, and $5.59\%$ higher than the maximum correlated model, respectively.
The minimum correlated ensemble model also has higher vote-based adversarial accuracy than the Gaussian ensemble model and the maximum correlated ensemble model.
It agrees with the previous analysis that the ensemble model with less correlated sub-models obtains better adversarial robustness. }
\old{Figure~\ref{fig:otherensemble}(b) demonstrates the accuracy of sub-models when attacking the ensemble model as a whole. When we attack the ensemble model trained with Gaussian noise, the accuracy of its three sub-models decreases in a similar pattern. However, the sub-models of our ensemble model perform differently. The accuracy of the sub-model with the low-pass filter stabilizes when the radius is larger than $5$. The accuracy of the sub-model with the 16 color filter decreases but still above the curves of sub-models of Gaussian ensemble. The original network performs the worst. This result justifies that our ensemble model improves robustness against adversarial attacks by introducing advantageous diversity.}
\hw{Figure~\ref{fig:otherensemble}(b) demonstrates the accuracy of sub-models when the ensemble model is attacked as a whole. When we attack the Gaussian noise ensemble model, the accuracy differences among its sub-models are close, which is $6.09\%$ at most. Also, the accuracy of its three sub-models decreases in a similar pattern. However, the sub-models of the minimum correlated ensemble model perform differently. The accuracy of the sub-model with the low-pass filter stabilizes when the radius is larger than $5/255$. The accuracy of the sub-model with the 16 color filter and the original network decreases with a relatively larger accuracy difference of $8.48\%$ at least. It justifies that our minimum correlated ensemble model improves robustness against adversarial attacks by introducing advantageous diversity.}
\old{
\begin{figure}[t]
\centering
\subfloat{
\includegraphics[width=0.46\linewidth]{fig/fig_vsGauss.eps}}
\qquad
\subfloat{
\includegraphics[width=0.46\linewidth]{fig/fig_sixlinesGauss.eps}}
\caption{Comparison with Gauss Ensembling.}
\label{fig:otherensemble}
\end{figure}
}
\hw{
\begin{figure}[htbp]
\centering
\begin{tabular}{c c}
\includegraphics[width=0.46\linewidth]{fig/fig_vsGauss_2.eps} &
\includegraphics[width=0.46\linewidth]{fig/fig_sixlinesGauss_2.eps}\\
(a) & (b)
\end{tabular}
\caption{Adversarial accuracy of minimum correlated ensemble model,
the maximum correlated ensemble model and the Gaussian noise ensemble model under the BPDA attack. The sub-model accuracy of the minimum correlated ensemble model and the Gaussian noise ensemble model is presented on the right.}
\label{fig:otherensemble}
\end{figure}
}
\subsection{Comparison with Adversarial Training}
Adversarial training is one of the most effective methods to improve the robustness of a DNN. We use the method proposed in \cite{shafahi2019adversarial} and compare the robustness of our minimum correlated ensemble model with the adversarial training. The adversarial training procedure takes $4$ iterations with the maximal perturbation size $\epsilon=8/255$.
\old{\paragraph{Comparison with a single Adversarial Training Model}
We first compare our ensemble model with one single adversarial training model. The grey line in the left part of Fig.\ref{fig:ATcompare} shows the adversarial robustness of a single adversarially trained model. It is interesting to see that our ensemble model has better adversarial accuracy under all perturbations. Note that every sub-model in our ensemble model has no defence mechanisms acting on the network. In other words, our ensemble-based defence can build robust models competing with adversarial training without manipulating the network.
}
\hw{\paragraph{Comparison with a single Adversarial Training Model}
We first compare our ensemble model with one single adversarial training model. The grey line in Figure~\ref{fig:ATcompare}(a) shows the adversarial robustness of a single adversarially trained model. Our ensemble model has better adversarial accuracy under all perturbations. The score-based accuracy of our method is $12.53\%$ higher than the single adversarial trained model at $\epsilon=10$. It is worth highlighting that every sub-model in our ensemble model has no defense mechanisms acting on the network. In other words, our ensemble-based defense can build robust models competing with adversarial training without manipulating the network.}
\paragraph{Comparison with Adversarial Training Ensembles}
We compare our method with the ensemble of three independent adversarially trained models. The orange lines in Figure~\ref{fig:ATcompare}(a) show the performance of the ensemble model using adversarial training. Remarkably, our score-based ensemble model has better accuracy than its counterpart with adversarial training when the perturbation size is large, i.e., $\epsilon \geq 10/255$. Meanwhile, our vote-based ensemble is very close to the one with adversarial training. Our ensemble model has a comparable defense to the ensemble of adversarially trained sub-models.
The orange lines in the right part of Figure~\ref{fig:ATcompare}(b) depict the accuracy of adversarially trained sub-models when the ensemble model is attacked as a whole. Comparing with the sub-models trained with Gaussian noise in Figure~\ref{fig:otherensemble}, adversarial training does not significantly improve the diversity between sub-models. The accuracy difference between sub-models is still relatively small, and that means the attacking methods can affect different sub-models simultaneously.
Since the adversarial training works on the network level and our method works on the data level, it is natural to combine these two methods. We demonstrate our ensemble model with adversarial training in Figure~\ref{fig:ATcompare}(a) using the cyan lines. Our ensemble model with adversarial training reaches a better robustness performance in both score-based and vote-based settings. The vote-based ensemble model achieves $71.53\%$ accuracy at $\epsilon = 5/255$. Conclusively, we build an ensemble model with high adversarial robustness using both our filter-based defense and adversarial training.
\old{
\begin{figure}[t]
\centering
\subfloat{
\includegraphics[width=0.46\linewidth]{fig/fig_vsESMAT.eps}}
\qquad
\subfloat{
\includegraphics[width=0.46\linewidth]{fig/fig_sixlines.eps}}
\caption{Comparison with Adversarial Training.}
\label{fig:ATcompare}
\end{figure}
}
\hw{
\begin{figure}[htbp]
\centering
\begin{tabular}{c c}
\includegraphics[width=0.46\linewidth]{fig/fig_vsESMAT_2.eps} &
\includegraphics[width=0.46\linewidth]{fig/fig_sixlines_2.eps}\\
(a) & (b)
\end{tabular}
\caption{Adversarial accuracy of the single adversarial trained (AT) model, ensemble model with adversarial training, our ensemble model with and without adversarial training under the BPDA attack. The sub-model accuracy of our ensemble model and the ensemble model with adversarial training is presented on the right.}
\label{fig:ATcompare}
\end{figure}
}
\section{Submission of papers to NeurIPS 2021}
\newpage
\section*{Checklist}
\begin{enumerate}
\item For all authors...
\begin{enumerate}
\item Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
\answerYes{}
\item Did you describe the limitations of your work?
\answerYes{}
\item Did you discuss any potential negative societal impacts of your work?
\answerNo{It should have no potential negative societal impacts.}
\item Have you read the ethics review guidelines and ensured that your paper conforms to them?
\answerYes{}
\end{enumerate}
\item If you are including theoretical results...
\begin{enumerate}
\item Did you state the full set of assumptions of all theoretical results?
\answerYes{}
\item Did you include complete proofs of all theoretical results?
\answerYes{The proof of the lemma 1 is in the supplementary material.}
\end{enumerate}
\item If you ran experiments...
\begin{enumerate}
\item Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)?
\answerYes{We include an url to our code and data in the supplemental material.}
\item Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)?
\answerYes{The data splits and the training parameters are discussed in Section~\ref{sec:experiment}.}
\item Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)?
\answerNA{}
\item Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)?
\answerYes{We introduce our experimental environments at the beginning of Section~\ref{sec:experiment}.}
\end{enumerate}
\item If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
\begin{enumerate}
\item If your work uses existing assets, did you cite the creators?
\answerYes{}
\item Did you mention the license of the assets?
\answerYes{The CIFAR-10 dataset and the FoolBox are under MIT license.}
\item Did you include any new assets either in the supplemental material or as a URL?
\answerNA{}
\item Did you discuss whether and how consent was obtained from people whose data you're using/curating?
\answerNA{}
\item Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content?
\answerNA{}
\end{enumerate}
\item If you used crowdsourcing or conducted research with human subjects...
\begin{enumerate}
\item Did you include the full text of instructions given to participants and screenshots, if applicable?
\answerNA{}
\item Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable?
\answerNA{}
\item Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation?
\answerNA{}
\end{enumerate}
\end{enumerate}
\section{Introduction}
Recent research reveals that deliberately crafted adversarial perturbations succeed in leading Deep Neural Networks (DNNs) to make wrong predictions, not only when attackers are aware of the architecture of DNNs, i.e., white-box setting, but also when they only have access to the input-output pairs of DNNs, i.e., black-box setting. This discovery exposes the potential danger in existing machine learning applications and encourages defenses against adversarial attacks.
These defenses are divided into two families, namely reactive defenses and proactive defenses. Reactive defenses~\cite{xu2017feature,guo2017countering,Sun_2019_CVPR,xie2017mitigating} aim to gain robustness by introducing an extra element to recognize or remove the adversarial context. Proactive defenses~\cite{goodfellow2014explaining,madry2017towards,lyu2015unified,chen2019improving} attempt to build networks inherently robust to adversarial attacks.
Transformations~\cite{xu2017feature,guo2017countering}, as a typical reactive approach, remove adversarial effects via applying simple filters. It is cheap but performs poorly against strong attacks, e.g., PGD~\cite{madry2017towards}, C\&W~\cite{carlini2017towards} and DeepFool~\cite{moosavi2016deepfool}. To augment the performance, randomness~\cite{raff2019barrage,prakash2018deflecting} and representation~\cite{moosavi2018divide,buckman2018thermometer,liu2019feature} are introduced into transformation. Transformation gains robustness but loses accuracy, since the original images are altered when it discards adversarial context. The networks learn from original data and cannot recognize the distorted information.
Adversarial Training~\cite{goodfellow2014explaining,madry2017towards}, as a proactive defense, augments the training process with adversarial images such that the network learns the relative knowledge. To produce adversarial images, adversarial training employees a special attack. Since different attacks have different preferences, the model is vulnerable to unseen attacks.
The ensemble is a solution to amend this drawback. It consists of several sub-models which learn from similar but different training sets, for instance, applying Gaussian noise to inputs and bootstrap~\cite{strauss2017ensemble} to augment robustness. Incorporating randomness stabilizes the performance of ensemble models. Random Self-Ensemble (RSE)~\cite{liu2018towards} adds random noise layers to prevent strong gradient-based attacks.
Diversity is essential to ensembles. To increase the diversity of adversarial examples during the training, ensemble adversarial training~\cite{tramer2017ensemble} adds adversarial examples transferred from other pre-trained models. Ensemble-of-specialists~\cite{abbasi2017robustness} multiplies adversarial examples targeted over different incorrect labels. This defense is confirmed not robust enough~\cite{he2017adversarial}. Other than data augmentation, Adaptive Diversity Promoting (ADP)~\cite{pang2019improving} and Diversity Training~\cite{kariyappa2019improving} design a regularizer to encourage diversity. However, these methods either fail in defending against strong attacks or are too expensive because it needs too many sub-models to achieve a decent diversity. So the question arises:
\begin{quote}
What is the advantageous diversity to improve the ensemble defenses against adversarial attacks?
\end{quote}
\paragraph{Contributions.} This work investigates the answer to this question. Inspired by the transformation defenses, we train sub-models with different front filters, such as dimension reduction, color quantization, and frequency filter. The model trained on a particular front filter is sensitive to a specific type of distortion. These front filters distort adversarial contexts. At the same time, training with transformed data allows models to learn and maintain accuracy on them.
We analyze the Pearson correlation coefficient among the models and the performance of the models and their ensemble. We infer that the sub-models with weakly correlated sensitivity constitute a more robust ensemble, and propose a simple and powerful defense framework for ensemble models based on the inference. Finally, the experimental results demonstrate that the proposed method improves the robust of the network against adversarial examples.
In the rest of the paper, we first introduce some basic notations and adversarial attacks in Section~\ref{sec:preliminary}. After elaborating the framework and the theoretical analysis in Section~\ref{sec:method}, we present experimental analysis and evaluation in Section~\ref{sec:experiment}. Finally, we conclude this paper in Section~\ref{sec:conclusion}.
\section{Methodology}
\label{sec:method}
In this section, we state the structure of our ensemble framework. A filter is an image transformation that extracts some important features of the original image.
We embed a filter in each sub-model as the core component,
which provides the diversity of sub-models. Then we analyze the relationship between the correlations of the filters and the local robustness, which induces a principle of choosing the optimal filter combination. This improves the ensemble defense against adversarial attacks.
\subsection{Filter-based Ensemble}
\old{The structure of our ensemble model is shown in Figure~\ref{fig:stucture}.}
\begin{figure}[ht]
\centering
\centering
\includegraphics[width=0.95\linewidth]{fig/fig_structure_2.eps}
\caption{An illustration of the structure of our ensemble model.
\label{fig:stucture}
\end{figure}
The structure of an ensemble model is shown in Figure~\ref{fig:stucture}. The model consists of several different sub-models, each of them equipped with a DNN and a front filter.
\hw{As shown in Figure~\ref{fig:stucture},}
an input image is pre-processed by three
different filters, and the obtained results are then respectively fed into three DNNs, which classify them and output the classification label individually. In the end, the results of each sub-models are combined by a voting mechanism.
\hw{Formally, a certain sub-model applies a front filter, denoted by $\zeta(\bm{x}):{\mathbb{R}}^m \to {\mathbb{R}}^{s}$, on the original inputs, and the DNN model follows the filter by $ f (\bm{x}):{\mathbb{R}}^{s} \to {\mathbb{R}}^n$.}
Then an ensemble model with $k$ sub-models $(\zeta_i,f_i)_{i=1}^k$ can be expressed as
\[
\mathcal{E}(\bm{x})= \mathrm{Vote} ( \{C_{ f_i } \circ \zeta_i (\bm{x})\}_{i=1}^k )\;,
\]
where the function $\mathrm{Vote}$ outputs the mode of the results of the sub-models, i.e., the classification label which appears the most times.
We call a input $\bm{x}$ stable for an ensemble model $\mathcal{E}$ if the output labels of all the sub-models in $\mathcal{E}$ are consistent. It is easy to obtain the following proposition.
\begin{proposition}\label{prop:attboth}
The ensemble model $\mathcal{E}$ cannot be attacked at a stable input $\bm{x}$ by the perturbation $\bm{\delta}$ if
any two sub-models are not attacked simultaneously at their respective inputs $\zeta(\bm{x})$ by the perturbations $\zeta(\bm{x}+\bm{\delta})-\zeta(\bm{x})$.
\end{proposition}
\old{To defend against the adversarial attack, we hope to build a more robust ensemble model. Unlike a lot of recent defense approaches, we don not focus on the training skills for the sub-models to gain a good defense effect.
On the contrary, our main idea is enhance the diversity of the sub-models --- using differentiated front filters to pass the partial feature of the inputs, which makes the perturbation generated by adversarial attacks cannot affect every DNN ensembled.}
\hw{To defend against an adversarial attack, we propose to build a more robust ensemble model.
Unlike network-based defenses like adversarial training, we do not focus on training skills for the sub-models to improve robustness.
On the contrary, our principle is to enhance the diversity of sub-models by extracting partial features of inputs using differentiated front filters. Since it is hard for an adversarial attack to effectively affect all sub-models at the same time, the ensemble model achieves better robustness from the diversity of front filters.}
So, the key to establish this ensemble model is how to choose a proper filter combination that provide both accuracy and robustness.
In the following, we explain how we gain a more robust ensemble model through the relation among the filters.
\subsection{Low Correlation Implies Strong Robustness}
\label{Sect:Correlation}
In this subsection, we give a theoretical description on the intuition that a low correlation of the sensitivity of two filters implies a more robust ensemble model under the assumption that the filters are of high quality, and this will guide us to choose the optimal filter combination from the candidates.
For an input $\bm{x}\in{\mathbb{R}}^m$ and a perturbation $\bm{\delta}\in{\mathbb{R}}^m$,
we define a function $r:{\mathbb{R}}^m \times {\mathbb{R}}^m \to {\mathbb{R}}_{\ge 0}$ as
\begin{equation}\label{eq:rsensi}
r(\bm{x},\bm{\delta})=\|\zeta(\bm{x}+\bm{\delta})-\zeta(\bm{x})\|_2
\end{equation}
to measure the sensitivity of a filter $\zeta$, i.e., the $L_2$-norm of the
perturbation affecting the input of the DNN in the sub-model.
Considering $r$ as a random variable,
we invoke the \emph{Pearson correlation coefficient}
to evaluate the correlation of the sensitivity of two filters,
which is expressed as
\begin{equation}\label{eq:corre}
\rho_{r_1,r_2}=\frac{{\mathrm{E}}[r_1r_2]-{\mathrm{E}}[r_1]{\mathrm{E}}[r_2]}{\sqrt{\mathrm{Var}[r_1]\mathrm{Var}[r_2]}}\;.
\end{equation}
We assume that the filters in the ensemble model are of high quality, which means that the interpretation of the difference of two images by each filter sincerely reflects their semantics difference in statistics, i.e., the random variables $r_1$ and $r_2$ are identically distributed. Under this assumption,
the equation~\eqref{eq:corre} indicates that ${\mathrm{E}}[r_1r_2]$ is monotonically increasing w.r.t. $\rho_{r_1,r_2}$.
For a certain DNN $ f $ and an input $\bm{x}$ classified into label $\labell$, we define the score difference by $\Delta_{ f }(\bm{x})= \min_{i \neq \labell}\left( f _\labell(\bm{x}) -
f _i(\bm{x})\right)$.
Then the robust radius at an input $\bm{\bar{x}}$ can be
estimated according to the following lemma.
\begin{lemma}[\cite{faoc}]
Consider a DNN defined by $ f (\bm{x}):{\mathbb{R}}^m\to{\mathbb{R}}^n$, whose Lipschitz constant
is $\mathcal{L}_{ f }$. Then for an input $\bm{\bar{x}}$, the DNN is robust
in $B_2(\bm{\bar{x}}, r)$ with $r < \tfrac{\Delta_{ f }(\bm{\bar{x}})}{\sqrt{2}\mathcal{L}_{ f }}$.
\end{lemma}
Then, we can infer that
two DNNs $ f _1$ and $ f _2$ cannot be attacked simultaneously at $\bm{\bar{x}}_1$ and $\bm{\bar{x}}_2$ by the perturbations $\bm{\delta}_1$ and $\bm{\delta}_2$ respectively,
if
\begin{equation}\label{eq:RobProd}
\|\bm{\delta}_1\|_2\|\bm{\delta}_2\|_2 < \frac{\Delta_{ f _1}(\bm{\bar{x}}_1) \Delta_{ f _2}(\bm{\bar{x}}_2)}
{2\mathcal{L}_{ f _1}\mathcal{L}_{ f _2}}\;.
\end{equation}
In our framework, $\bm{\bar{x}}_1$ and $\bm{\bar{x}}_2$ are the input processed by two filters, and $\|\bm{\delta}_1\|_2$ and $\|\bm{\delta}_2\|_2$ are the sensitivity calculated by \eqref{eq:rsensi}.
It is clear that the right part in the inequality~\eqref{eq:RobProd} is determined by the structure and parameters of the DNNs, while the left part is determined by the front filters.
The expectation $\mathrm E[r_1,r_2]$ is a statistical description of the item $\|\bm{\delta}_1\|_2\|\bm{\delta}_2\|_2$ in \eqref{eq:RobProd}: A small expectation $\mathrm E[r_1,r_2]$ imples that the value of $\|\bm{\delta}_1\|_2\|\bm{\delta}_2\|_2$ tends to be small statistically.
Consequently, by combining \eqref{eq:corre}, \eqref{eq:RobProd} and Proposition~\ref{prop:attboth},
\old{we infer that the filters with independent sensitivity
implies a more robust ensemble model.}
we infer that low correlation of the sensitivity among filters
implies strong robustness of ensemble models.
This leads to our principle as `{\bf minimum correlation coefficients}' for choosing filter combinations, i.e., to optimize the robustness of our ensemble model, we choose the filters among which the correlation is the weakest.
\begin{remark}
Note that the cosine similarity of $\zeta(\bm{x}+\bm{\labell})-\zeta(\bm{x})$ between two filters
is also a measure of their correlation.
However, we only consider the sensitivity from the perspective of magnitudes of perturbation vectors generated by filters, because the gradient of the entire sub-model depends on both the filter and the DNN, so we do not choose to analyze their directions without considering the following DNNs.
\end{remark}
\subsection{Filter Candidates} The original image is prop-processed by a filter before it is sent to the network. Therefore, some information is discarded and thus the overall entropy of the image
is reduced. It is also regarded as a manual feature extraction procedure that extracts the most important features that benefit the task. It is generally harder to attack the filtered image because there is less information that the attacking methods can utilize. The filters we use are categorized into the following four classes.
\paragraph{Dimension Reduction}
The easiest way to reduce the entropy of an image is to reduce its dimensionality. Color pictures have three dimensions, i.e., length, width, and color channels. It is simple to reduce the first two dimensions by downsizing an image. Grayscale transformation can compress the color channels into one grayscale channel. Generally, downsizing and grayscale transformation preserve the overview of the original image with certain loss of details. Bilinear interpolation is used in downsizing filters. We use the ITU-R BT.601{\footnote{BT.601 : Studio encoding parameters of digital television for standard 4:3 and wide screen 16:9 aspect ratios: \url{https://www.itu.int/rec/R-REC-BT.601-7-201103-I/en}}} luma transformation for the grayscale filter.
\paragraph{Color Quantization} Another way to reduce the complexity of an image is to reduce its number of colors. The size of a CIFAR-10 image is $32\times32$. It may have $1\,024$ colors at most. But it can still be recognizable using much fewer colors.
The fast octree algorithm~\cite{fastoctree} is adopted to reduce the colors of the image. A full-size octree with a depth of seven can be used to partition the RGB color space. The octree subdivides the colorspace into eight octants recursively. Each leaf node of the octree represents an individual color. The fast octree algorithm builds the tree according to the color of a given image and merges the leaf nodes when the number of colors overflows.
\paragraph{Frequency Filters} In digital image processing, frequency filters are commonly applied to extract useful features from pictures. The high-frequency features are usually the noise and the details of the original image, and the low-frequency features are often its overview. The high-pass filters suppress the low-frequency features, and the low-pass filters do the opposite.
Our low-pass and high-pass filters are based on the discrete Fourier Transform.
Via shifting the low-frequency part to the center of the spectrum and multiplying it by a Gaussian mask (high-pass mask) element-wise, we obtain the low-pass (high-pass) filtered image.
\paragraph{Data Discretization}
Inputs for a DNN can be any real value, while the 8-bit RGB color model takes integer values in the range of $[0,255]$. To keep the practical meaning of an input, real numbers are approximated by its closest integer. This is essential since the DNN for image classification should have actual image data instead of arbitrary inputs. When we use iterative methods to attack the DNN, discretization can help generate practical adversarial examples. In our ensemble model, every sub-model trained with the original data is equipped with a discretization filter.
\section{Preliminary}
\label{sec:preliminary}
\old{We recall some basic notions on deep neural networks and abstract interpretation.
For a vector $\bar x \in \mathbb{R}^n$, we use $x_i$ to denote its $i$-th entry.
For a matrix $W \in \mathbb R ^ {m \times n}$, $W_{i,j}$ denotes the entry in its $i$-th row and $j$-th column.}
\hw{In this section, we first state some basic notions of DNNs and the definition of DNN robustness in a local region.
Then we recall the norm-based robustness region, as well as the Lipschitz constant of DNNs. After that, we give a brief introduction to a few existing attacks used in our experiments.}
\newcommand{\mathbb{P}}{\mathbb{P}}
\newcommand{\operatorname{sign}}{\operatorname{sign}}
\def \labell {\ell}
\subsection{Deep Neural Network and Local Robustness}
\old{Our work concentrates on deep neural networks, which can be characterized as a function $ f :\mathbb{R}^m\to\mathbb{R}^n$.
For classification tasks, a DNN usually chooses the output dimension with the largest score, i.e., $C_{ f }(\bm{x}):=\arg\max_{1\leq i\leq n}f_i(\bm{x})$, as its output label.}
\hw{Our work concentrates on the image classification task.
A DNN, which can be characterized as a function $ f :\mathbb{R}^m\to\mathbb{R}^n$, usually gives prediction by maximizing the output vector, i.e., $C_{ f }(\bm{x}):=\arg\max_{1\leq i\leq n}f_i(\bm{x})$, where $\bm{x} \in \mathbb{R}^m$ represents an image. To optimize the network, we minimize the cost function $\mathbb{J}( f ,\bm{x},\labell)$, in which $\labell$ is the ground truth.}
Intuitively, the local robustness of a DNN ensures the consistency of its behavior of a given input under certain perturbations, and a strict robustness condition ensures that there is no adversarial example around an input $\bm{x}$. Formally, the local robustness of a DNN can be defined as below.
\begin{definition}[DNN robustness]
\label{def:localrobustness}
Given a DNN $ f :\mathbb{R}^m \to \mathbb{R}^n$ and an input region $B\subset\mathbb{R}^m$, we say that $ f $ is (locally) \emph{robust} in $B$ if for any $\bm{x},\bm{x}'\in B$, we have $C_{ f }(\bm{x})=C_{ f }(\bm{x}')$.
\end{definition}
In a typical way, the region $B$ here is usually defined by the neighborhood of an input, where $L_p$-norm balls are commonly used.
As the case of the $L_2$,
the neighborhood of an input $\bm{\bar{x}}$ bounded by the $L_2$-norm can be described as an $L_2$ ball: The $L_2$ (closed) ball with the center $\bm{\bar{x}} \in \mathbb R^n$ and the radius $r>0$ is defined as $B_2(\bm{\bar{x}},r)=\{\bm{x} \in \mathbb R^n \mid \|\bm{x}-\bm{\bar{x}}\|_2 \le r\}$.
\paragraph{Lipschitz constant of DNNs}
The Lipschitz constant of a function is a measure to indicates the maximum ratio between variations in the output space and variations in the input space.
In~\cite{RHK2018}, a DNN $f:\mathbb R^m \to \mathbb R^n$ is proved to be Lipschitz continuous. Namely, there exists $\mathcal{L} >0$, s.t. for any $ \bm{x}, \bm{x}' \in \mathbb R^m$,
\[
\| f ( \bm{x})- f (\bm{x}')\|_2 \le \mathcal{L} \cdot \|\bm{x}-\bm{x}'\|_2,
\]
and here $\mathcal{L}$ is called a Lipschitz constant of $f$. Generally, DNNs with a smaller Lipschitz constant are likely to be more robust.
\subsection{Adversarial Attacks}
Adversarial attacking methods attempt to find an imperceptible perturbation leading to misclassification, also regarded as a testing method for network robustness. We present several fundamental untargeted attacks widely used in existing literature. Hereafter, we denote a potential adversarial example as $\bm{x}'$, the adversarial perturbation as $\bm{\delta}$, and the gradient calculated from the cost function to input as $\nabla_{\bm{x}} \mathbb{J}(\bm{x})$.
\paragraph{Fast Gradient Sign Method (FGSM)}
Fast Gradient Sign Method (FGSM) by Goodfellow and his colleagues in 2014 ~\citep{goodfellow2014explaining} simply uses the one-step gradient to generate the perturbation:
\begin{equation}
\bm{x}' = \bm{x} + \bm{\delta} = \bm{x} - r \operatorname{sign} ( \nabla_{\bm{x}} \mathbb{J}(\bm{x}) ).
\label{eq:fgsm}
\end{equation}
It is the perturbation that minimizes the first-order objective function for the constraint $\| \bm{\delta}\|_\infty = r$.
\old{\paragraph{Basic Iterative Method (BIM)}
It is possible to refine FGSM iteratively. BIM~\citep{kurakin2016physical} is the iterative version of FGSM. BIM initializes $\bm{x}'_0 := \bm{x}$ and then iterates by progressing in the opposite direction of the gradient with stepsize $\alpha$. The recurrence is therefore:
\begin{equation}
\bm{x}'_{i+1} := \mathbb{P}_{B_\infty(\bm{x}, r)}(\bm{x}'_i - \alpha \operatorname{sign} \nabla_{\bm{x}} \mathbb{J}(\bm{x}'_i)).
\label{eq:ifgsm}
\end{equation}
Here, the ball $B_\infty(\bm{x}, r)$ of $L_\infty$-norm is centered in $\bm{x}$ with radius $r>\alpha$. Projection on the ball $B_\infty(\bm{x}, r)$ is defined as
\begin{align}
\mathbb{P}_{B_\infty(\bm{x}, r)}(\bm{x}') := \bm{x} + clip_{[-r,r]}(\bm{x}' - \bm{x}), \old{\textcolor{red}{What~does~'clip'~mean?}}
\end{align}
where $clip_{[-r,r]}(\cdot)$ denotes the clipping input into range $[-r,r]$ element-wise.
Thus, when the current solution remains inside the ball, the projection is not active. While the iterations calculate perturbations that get outside the ball, then the projection brings them back to its surface.}
\old{\paragraph{DeepFool}
DeepFool~\citep{moosavi2016deepfool} employs an objective function in order to modify the adversarial image into the most likely class. It seems easier to make $\bm{x}$ negative when the objective function has a positive but low value at $\bm{x}$.
In the first order for a targeted class, the minimum distortion necessary in $L_2$-norm is reached when $\bm{\delta}\propto - \nabla_{\bm{x}}\mathbb{J}(\bm{x}, \labell)$ with:
\begin{equation}
\| \bm{\delta}\| = \frac{\mathbb{J}(\bm{x},\labell)}{\|\nabla_{\bm{x}}\mathbb{J}(\bm{x},\labell)\|}.
\end{equation}
It is best to target the class $\labell$ that will cause the addition of the smallest distortion.}
\paragraph{Projected Gradient Descent (PGD)}
PGD initializes $\bm{x}'_0 := \bm{x}$ and then iterates by progressing in the opposite direction of the gradient with stepsize $\alpha$. The accumulated distortion are projected onto an $L_p$-norm ball~\citep{madry2017towards}:
\begin{equation}
\bm{x}'_{i+1} := \mathbb{P}_{B_p(\bm{x}, r)}(\bm{x}'_i - \alpha \mathsf{n}(\nabla_{\bm{x}} \mathbb{J}(\bm{x}'_i))),
\label{eq:pgd}
\end{equation}
where $\mathsf{n}(\bm{x}) := \bm{x} / \| \bm{x} \|_p$, and
\begin{align}
\mathbb{P}_{B_p(\bm{x}, r)}(\bm{x}') :=
\begin{cases}
\bm{x}', & \text{if~}\| \bm{x}' - \bm{x} \|_p < r,\\
\bm{x} + r \mathsf{n}({\bm{x}' - \bm{x}}), & \text{otherwise}.
\end{cases}
\end{align}
The $L_p$-norm ball used for projection is $B_p(\bm{x}, r)$, centered in $\bm{x}$ with radius $r$. Again, the attack does not end when $\bm{x}'_i$ hits the boundary of the ball $B_p(\bm{x}, r)$ for the first time. It continues and seeks to minimize the objective function while remaining on the sphere.
\paragraph{Basic Iterative Method (BIM)}
BIM~\citep{kurakin2016physical}, as an iterative version of FGSM, it employees the sign of gradients of network iteratively with stepsize $\alpha$ to update adversarial perturbations. The principle of BIM is similar to a $L_\infty$-norm version of PGD. All the pixels in the adversarial example are clipped into range $[-r,r]$, i.e., adversarial perturbations are resized within the surface of an $L_\infty$-norm ball with radius $r$. The main difference is that PGD utilizes a random initialization and uses the gradients directly.
\paragraph{Backward Pass Differentiable Approximation (BPDA)}
BPDA~\citep{athalye2018obfuscated} allows attackers to generate adversarial perturbations targeted at the network with defenses as a whole. BPDA approximates derivatives by computing the forward pass normally and computing the backward pass using a differentiable approximation of the defense function. For instance, if it is impossible to calculate gradients through the transformation, BPDA generates adversarial examples by including the transformation during the forward pass and
replaces the transformation with an identity function during the backward pass under the assumption that the transformation output is close to the original input.
|
\section{I. Introduction}
One-dimensional material systems continue attracting considerable attention due to their rich physical properties, where the charge, spin, orbital, and lattice degrees of freedom are intertwinned in a reduced dimensional phase space~\cite{Yunoki:prl,Dagotto:rmp94,Grioni:JPCM,Monceau:ap}. Remarkable physical phenomena have been reported in different one-dimensional systems, such as high critical temperature superconductivity in copper chains and ladders~\cite{cu-ladder1,cu-ladder2,cu-ladder3}, ferroelectricity~\cite{Park:prl,Lin:prm,Zhang:prb21}, spin block states~\cite{Zhang:prbblock,Herbrych:prbblock}, nodes in spin density ~\cite{Lin:prb21}, orbital ordering ~\cite{Pandey:prb21,Lin:prm21}, orbital-selective Mott phases~\cite{Patel:osmp,Herbrych:osmp}, and charge density wave or Peierls distortions~\cite{Zhang:prb21,Zhang:prbcdw,Gooth:nature,Zhang:arxiv}.
Because superconductivity at high pressure was reported a few years ago in the two-leg ladder compounds BaFe$_2$$X_3$ ($X$ = S, Se)~\cite{Takahashi:Nm,Ying:prb17} with electronic density $n = 6$, the iron ladders became interesting one-dimensional systems to research high-temperature iron-based superconductors~\cite{Yamauchi:prl15,Zhang:prb17,Zhang:prb18,Zheng:prb18,Zhang:prb19,Zhang:prb19,Materne:prb19,Pizarro:prm,Wu:prb19,Craco:prb20}. BaFe$_2$S$_3$ displays a stripe-type antiferromagnetic (AFM) order below $12$~K, involving AFM legs and ferromagnetic (FM) rungs (this state is called CX)~\cite{Takahashi:Nm,chi:prl}. In addition, BaFe$_2$Se$_3$, namely replacing S by Se, displays an exotic AFM state with $2\times2$ FM blocks coupled antiferromagnetically along the long ladder direction below $256$~K under ambient conditions~\cite{Caron:Prb,Caron:Prb12}. By applying hydrostatic pressure, both systems display an insulator-metal transition~\cite{Zhang:prb17,Zhang:prb18,Materne:prb19,Ying:prb17}, followed by superconductivity at $P$ $\sim 11$ Gpa~\cite{Takahashi:Nm,Ying:prb17}. Furthermore, an OSMP state was found in BaFe$_2$Se$_3$ by neutron experiments at ambient pressure~\cite{mourigal:prl15}. This state was theoretically predicted before experimental confirmation by using the density matrix renormalization group (DMRG) method based on a multi-orbital Hubbard model~\cite{osmp1,osmp2}. These developments in the area of two-leg iron ladder systems naturally introduce a simple question: can iron chains, as opposed to ladders, also display similar physical properties?
Some iron chalcogenide chains $A$Fe$X_2$ ($A$ = K, Rb, Cs and Tl, $X$ = S or Se) have already been prepared experimentally~\cite{Seidov:prb01,Seidov:prb16}. Neutron diffraction experiments revealed that the magnetic coupling along the chain direction is AFM with dominant $\pi$ wavevector ($\uparrow$-$\downarrow$-$\uparrow$-$\downarrow$)~\cite{Bronger:jssc,Seidov:prb16}, similar to the CX-AFM state in BaFe$_2$S$_3$. But in $A$Fe$X_2$ compounds there are 5 electrons in the $3d$ iron orbitals, corresponding to valence Fe$^{\rm 3+}$. At this electronic density $n = 5$, the AFM phase with the $\uparrow$-$\downarrow$-$\uparrow$-$\downarrow$ configuration was observed in a large portion of the magnetic phase diagram when using the five-orbital Hubbard model for iron chains studied via the real-space Hartree-Fock (HF) approximation~\cite{Luo:prb14}. However, these HF calculations reported a much richer phase diagram for iron chains at electronic density $n = 6$. Considering that the atomic electronic density $n=6$ is the same as in iron planar and ladder
superconductors~\cite{Dagotto:Rmp}, then iron chains with electronic density $n = 6$ could be potential candidates to achieve a similar superconducting state.
Na$_2$FeSe$_2$ with $n=6$ was considered as a candidate~\cite{Stuble:jssc}. Recent DMRG calculations~\cite{Pandey:prb} found a stable region of staggered spin order in the phase diagram (with $\uparrow$-$\downarrow$-$\uparrow$-$\downarrow$ order) at low Hund coupling $J_H$/U, while block phases ($\uparrow$-$\uparrow$-$\downarrow$-$\downarrow$) dominate at larger $J_H$/U. Another interesting iron chain with $n=6$ $Ln$$_2$O$_2$FeSe$_2$ ($Ln$ = Ce, La) was prepared experimentally~\cite{McCabe:cc,McCabe:prb}. In addition, OSMP and Hund physics were discussed in this compound by using DMRG-based calculations based on the Hubbard model~\cite{Lin:osmp}.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig1.pdf}
\caption{Schematic crystal structure of the high-pressure Ba$_2$FeS$_3$ conventional cell (green = Ba; brown = Fe; yellow = S). (a) Sketch of the $ac$ plane along $b$ direction. (b) FeS$_4$ chain along the $b$-axis. (c) Inter chains magnetic exchange couplings on the $ac$ plane. }
\label{Fig1}
\end{figure}
Very recently, a new $n=6$ iron chalcogenide Ba$_2$FeS$_3$ (note this is a 213 formula, unlike the 123 of ladders) was synthesized under high-pressure (HP) and high-temperature conditions~\cite{Guan:jac}. A long-range AFM transition was reported at $\sim 56$~K, and the magnetic susceptibility curve exhibited a round hump behavior until $110$~K~\cite{Guan:jac}. As shown in Fig.~\ref{Fig1}(a), Ba$_2$FeS$_3$ (HP) has an orthorhombic structure with space group $Pnma$ (No. $62$). In the crystal structure of Ba$_2$FeS$_3$ (HP), there are four FeS$_4$ chains along the $b$-axis, where nearest-neighbor irons are connected by sulfur atoms along the chain direction [see Fig.~\ref{Fig1}(b)]. The nearest-neighbor (NN) Fe-Fe bond is $4.30$~\AA, corresponding to the lattice constant along the $b$-axis ~\cite{Guan:jac}. In addition, the NN and next-nearest-neighbor (NNN) interchain distances are $5.52$~\AA ~and $5.86$~\AA, respectively, as displayed in Fig.~\ref{Fig1}(c). Based on the crystal structure, the Ba$_2$FeS$_3$ (HP) phase displays quasi one-dimensional characteristics, suggesting
that the chain direction plays the dominant role in the physical properties.
To better understand the electronic and magnetic structures, here both first-principles DFT and DMRG methods were employed to investigate Ba$_2$FeS$_3$ at high pressure. First, the {\it ab initio} DFT calculations indicated a strongly anisotropic
electronic structure for Ba$_2$FeS$_3$ (HP), in agreement with its anticipated one-dimensional geometry. Furthermore, based on DFT calculations, we found the staggered spin order was the most likely magnetic ground state, with the coupling along the chain direction dominanted by the $\pi$ wavevector. Based on the Wannier functions obtained from first-principle calculations, we obtained the relevant hopping amplitudes and on-site energies for the iron atoms. Next, we constructed a multi-orbital Hubbard model for the iron chains. Based on the DMRG calculations, we calculated the ground-state phase diagram varying the on-site Hubbard repulsion $U$ and the on-site Hund coupling $J_H$. The staggered AFM with $\uparrow$-$\downarrow$-$\uparrow$-$\downarrow$ was found to be dominant in a robust portion of the phase diagram, in agreement with DFT calculations. In addition, OSMP physical properties were found in the regime of intermediate Hubbard coupling strengths. Eventually, at very large $U/W$, the OSMP is replaced by the Mott insulating (MI) phase.
\section{II. Method}
\subsection{A. DFT Method}
In the present study, the first-principles DFT calculations were performed with the projector augmented wave (PAW) method, as implemented in the Vienna {\it ab initio} simulation package (VASP) code~\cite{Kresse:Prb,Kresse:Prb96,Blochl:Prb}. Here, the electronic correlations were considered by using the generalized gradient approximation (GGA) with the Perdew-Burke-Ernzerhof functional~\cite{Perdew:Prl}. The plane-wave cutoff was $500$ eV. The k-point mesh was $8\times16\times4$ for the non-magnetic calculations, which was accordingly adapted for the magnetic calculations. Note that we tested explicitly that this $k$-point mesh already leads to converged energies. Furthermore, the local spin density approach (LSDA) plus $U_{\rm eff}$ with the Dudarev format was employed~\cite{Dudarev:prb} in the magnetic DFT calculations. Both the lattice constants and atomic positions were fully relaxed with different spin configurations until the Hellman-Feynman force on each atom was smaller than $0.01$ eV/{\AA}. In addition to the standard DFT calculations, the maximally localized Wannier functions (MLWFs) method was employed to fit the Fe $3d$'s bands by using the WANNIER90 packages~\cite{Mostofi:cpc}. All the crystal structures were visualized with the VESTA code~\cite{Momma:vesta}.
\subsection{B. Multi-orbital Hubbard Model}
To better understand the magnetic behavior of the quasi-one-dimensional Ba$_2$FeS$_3$ in the dominant chain direction, an effective multi-orbital Hubbard model was constructed. The model studied here includes the kinetic energy and interaction energy terms $H = H_k + H_{int}$. The tight-binding kinetic portion is described as:
\begin{eqnarray}
H_k = \sum_{\substack{i\sigma\gamma\gamma'}}t_{\gamma\gamma'}(c^{\dagger}_{i\sigma\gamma}c^{\phantom\dagger}_{i+1\sigma\gamma'}+H.c.)+ \sum_{i\gamma\sigma} \Delta_{\gamma} n_{i\gamma\sigma},
\end{eqnarray}
where the first part represents the hopping of an electron from orbital $\gamma$ at site $i$ to orbital $\gamma'$ at the NN site $i+1$, using a chain of length $L$. $\gamma$ and $\gamma'$ represent the three different orbitals.
The (standard) electronic interaction portion of the Hamiltonian is:
\begin{eqnarray}
H_{int}= U\sum_{i\gamma}n_{i\uparrow \gamma} n_{i\downarrow \gamma} +(U'-\frac{J_H}{2})\sum_{\substack{i\\\gamma < \gamma'}} n_{i \gamma} n_{i\gamma'} \nonumber \\
-2J_H \sum_{\substack{i\\\gamma < \gamma'}} {{\bf S}_{i\gamma}}\cdot{{\bf S}_{i\gamma'}}+J_H \sum_{\substack{i\\\gamma < \gamma'}} (P^{\dagger}_{i\gamma} P_{i\gamma'}+H.c.).
\end{eqnarray}
The first term is the intraorbital Hubbard repulsion. The second term is the electronic repulsion between electrons at different orbitals where the standard relation $U'=U-2J_H$ is assumed due to rotational invariance. The third term represents the Hund's coupling between electrons occupying the iron $3d$ orbitals. The fourth term is the pair hopping between different orbitals at the same site $i$, where $P_{i\gamma}$=$c_{i \downarrow \gamma} c_{i \uparrow \gamma}$.
To solve the multi-orbital Hubbard model, and obtain the magnetic properties of Ba$_2$FeS$_3$ (HP) along the $b$-axis direction including quantum fluctuations, the many-body technique was employed based on the DMRG method~\cite{white:prl,white:prb}, where specifically we used the DMRG++ software~\cite{Alvarez:cpc}. In our DMRG calculations, we employed a $16$-sites cluster chain with open-boundary conditions (OBC). Furthermore, at least $1400$ states were kept and up to $21$ sweeps were performed during our DMRG calculations. In addition, the electronic filling $n = 4$ in the three orbital was considered. This electronic density (three electrons in four orbitals) is widely used in the context of iron superconductors with Fe$^{\rm 2+}$ valence (n = 6)~\cite{osmp1,Luo:prb10}. The common rationalization to justify this density is to consider one orbital doubly occupied and one empty, and thus both can be discarded. This leads to four electrons in the remaining three orbitals, providing a good description of the physical properties for the real iron systems with $n= 6$~\cite{osmp1,Daghofer:prb10,Rin:prl}.
In the tight-binding term, we used the Wannier function basis \{$d_{xz}$, $d_{x^2-y^2}$, $d_{xy}$\}, here referred to as $\gamma$ = \{0, 1, 2\}, respectively. We only considered the NN hopping matrix:
\begin{equation}
\begin{split}
t_{\gamma\gamma'} =
\begin{bmatrix}
0.012 & 0.045 & 0.080 \\
0.045 & 0.112 & -0.018 \\
-0.080 & 0.018 & 0.238
\end{bmatrix}.\\
\end{split}
\end{equation}
All the hopping matrix elements are given in eV units. $\Delta_{\gamma}$ is the crystal-field splitting of orbital $\gamma$. Specifically, $\Delta_{0} =-0.339$, $\Delta_{1} = 0.047$, and $\Delta_{2} = -0.127$ (the Fermi level is considered to be zero). The total kinetic energy bandwidth $W$ is 1~eV. More details about the Wannier functions and hoppings can be found in APPENDIX A.
\section{III. DFT results}
\subsection{A. Non-magnetic state}
Before addressing the magnetic properties, let us discuss the electronic structures of the non-magnetic state of Ba$_2$FeS$_3$ (HP) based on the experimentally available structural properties~\cite{Guan:jac}. At high pressure, the lattice constants are $a = 8.683$ ~\AA, $b = 4.297$ \AA~and $c = 17.025$ \AA, respectively.
First, we present the density-of-states (DOS) of the non-magnetic state of Ba$_2$FeS$_3$ (HP), displayed in Fig.~\ref{Fig2}(a). Near the Fermi level, the electronic density is mainly contributed by the iron $3d$ orbitals, where the $p-d$ hybridization between Fe $3d$ and S $3p$ states is weak. Furthermore, the Fe $3d$ bands of Ba$_2$FeS$_3$ (HP) are located in a relatively small range of energy from $-1$ to $1$ eV, while the S $3p$ bands are located at a deeper energy level from $-5$ eV to $-2$ eV. In iron ladders~\cite{Zhang:prb19}, the $p-d$ hybridization was reported to be stronger than in the Ba$_2$FeS$_3$ (HP) chain under investigation here. According to the DOS of Ba$_2$FeS$_3$ (HP), the charge transfer gap $\Delta$ = $\varepsilon_d$ - $\varepsilon_p$ is large, indicating Ba$_2$FeS$_3$ (HP) is a Mott-Hubbard system. Thus, the Fe-S hybridization is smaller than that in iron ladders.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig2.pdf}
\caption{ (a) Density-of-states near the Fermi level of Ba$_2$FeS$_3$ (HP) for the non-magnetic phase (black = Total; blue = Ba; red = Fe; cyan = S). (b) Band structures of Ba$_2$FeS$_3$ (HP) for the non-magnetic state. The Fermi level is shown with a dashed horizontal line. The coordinates of the high-symmetry points in the Brillouin zone (BZ) are $\Gamma$ = (0, 0, 0), X = (0.5, 0, 0), S = (0.5, 0.5, 0), Y = (0, 0.5, 0), Z = (0, 0, 0.5), U = (0.5, 0, 0.5), R = (0.5, 0.5, 0.5). Note that all the high-symmetry points are in scaled units, corresponding to the units of $2\pi$/s, ($s = a, ~b$ or $c$).}
\label{Fig2}
\end{figure}
The result of the previous paragraph can be understood intuitively. First, in the dominant Fe chain of Ba$_2$FeS$_3$ (HP), the Fe-Fe bond along the chain is about $4.30$ \AA, larger than the corresponding number for the iron ladder ($\sim$ $2.64$ \AA) with $n = 6$~\cite{chi:prl}. Second, there is only one S atom connecting NN Fe atoms (with Fe-S bond $\sim$ $2.44$ ~\AA) in the iron chains of Ba$_2$FeS$_3$ (HP). On the other hand, in iron ladders with $n = 6$~\cite{chi:prl}, there are two S atoms connecting the NN Fe atoms along the leg direction (with Fe-S bonds being $2.29$ and $2.27$~\AA). Considering those differences of structural geometries as compared to iron ladders, the overlap of Fe and S atoms of Ba$_2$FeS$_3$ (HP) would be weaker than that of iron ladders, resulting in a weaker $p-d$ hybridization in the Ba$_2$FeS$_3$ (HP) chain.
The projected band structures of Ba$_2$FeS$_3$ (HP) are displayed in Fig.~\ref{Fig2}(b). It is clearly shown that the band is more dispersive from Y to $\Gamma$ along the chains than along other directions, such as
$\Gamma$ to X along the $a$-axis,
which is compatible with the presence of quasi-one-dimensional chains along the $k_y$ axis. In this case, the intrachain coupling should play the key role in magnetism and other physical properties.
\subsection{B. Magnetism}
To qualitatively represent the magnetism of Ba$_2$FeS$_3$ (HP), a simple classical Heisenberg model with three magnetic exchange couplings $J$ was introduced to described phenomenologically this system:
\begin{eqnarray}
\nonumber H&=&-J_1\sum_{<ij>}\textbf{S}_i\cdot\textbf{S}_j-J_2\sum_{[kl]}\textbf{S}_k\cdot\textbf{S}_l\\
&&-J_3\sum_{\{mn\}}\textbf{S}_m\cdot\textbf{S}_n,
\end{eqnarray}
where $J_1$ is the intrachain exchange interactions between NN Fe-Fe spin pairs, while $J_2$ and $J_3$ are the interchain exchange interactions between two NN iron chains, corresponding to two different interchain Fe-Fe distances, as displayed in Fig.~\ref{Fig1}(c). By mapping the DFT energies of different magnetic
configurations~\cite{Jcontext}, based on the experimental lattice structure, we obtained the coefficients of different $J$'s as a function of Hubbard $U_{\rm eff}$ in Fig.~\ref{Fig3}. As expected, $J_1$ is the dominant one, indicating that the intrachain magnetic coupling plays the key role in this system. Based on these calculated $J$'s, the magnetic coupling along the chain favours AFM, while the interchain couplings are quite weak.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig3.pdf}
\caption{Magnetic exchange couplings as a function of the Hubbard $U_{\rm eff}$ coupling.}
\label{Fig3}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig4.pdf}
\caption{(a) Sketch of some possible AFM patterns studied here. Spin up and down are distinguished by brown and blue, respectively. (b-d) DFT results for Ba$_2$FeS$_3$ (HP) as a function of $U_{\rm eff}$. (b) Energies (per Fe) of various magnetic orders are indicated. The FM configuration is taken as the reference. (c) Local magnetic moments of Fe, integrated within the default Wigner-Seitz sphere as specified by VASP. (d) Band gaps for the many states analyzed.}
\label{Fig4}
\end{figure}
To better understand the possible magnetic configurations, we also considered several AFM configurations in a $1\times2\times1$ supercell, as shown in Fig.~\ref{Fig4}(a). Both the lattice constants and atomic positions were fully relaxed for those different spin configurations. First, the AF2 magnetic order always has the lowest energy among all tested candidates, independently of $U_{\rm eff}$, as shown in Fig.~\ref{Fig4}(b). Furthermore, the energies of the AF1 and AF3 orders are close to the energy of the AF1 state, indicating a quite weak $J_3$ coupling, in agreement with our previous discussion using the Heisenberg model.
Moreover, the calculated local magnetic moment per Fe is displayed in Fig.~\ref{Fig4}(c), for different possible magnetic configurations. With increasing $U_{\rm eff}$, the moment of Fe in the AF2 state increases from $3.12$ to $3.55$ $\mu_{\rm B}$/Fe, which is higher than those calculated for the CX-AFM type configuration in iron ladders with $n=6$~\cite{Zhang:prb17}. As shown in Fig.~\ref{Fig4}(d), all AFM orders are insulating and the gap increases with $U_{\rm eff}$, as expected.
According to the calculated DOS for the AF2 state, the bands near the Fermi level are mostly contributed by Fe's $3d$ orbitals and the Fe atom is in the high spin configuration. Furthermore, the DOS plot indicates a Mott transition behavior [Fig.~\ref{Fig5}]. Our calculated band gap for the case $U_{\rm eff} = 1$ eV in the AF2 state is about $0.62$ eV, which is very close to the experimental gap obtained from fitting the resistivity versus $1/T$ curve ($\sim$ $0.676$ eV)~\cite{Guan:jac}.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig5.pdf}
\caption{DOS for the AF2 state obtained using LSDA + $U_{\rm eff}$ (= 1 eV) DFT caculations.}
\label{Fig5}
\end{figure}
In summary of our DFT results, we found a strongly anisotropic quasi-one-dimensional electronic band structure, corresponding to its dominant chain geometry. In addition,we found the AF2 magnetic order is the most likely ground state, where the interchain coupling dominates. Furthermore, our calculations also indicated this system is a Mott Hubbard system with a Mott gap.
\section{IV. DMRG results}
As discussed in the DFT section, the chain direction is the most important for the physical properties of Ba$_2$FeS$_3$ (HP). Using DFT+$U$ calculations, we obtained a strong Mott insulating AFM phase. However, in one-dimensional systems, quantum fluctuations are important at low temperatures. Because DFT neglects fluctuations, here we employed the many-body DMRG method to incorporate the quantum magnetic couplings in the dominant chain, where quantum fluctuations are needed to fully clarify the true magnetic ground state properties. In fact, in previous well-studied iron 1D ladders and chains, those quantum fluctuations were found to be crucial to understand the magnetic properties~\cite{mourigal:prl15,Herbrych:osmp}. It also should be noticed that the DMRG method has proven to be a powerful technique for discussing low-dimensional interacting systems ~\cite{Schollw:rmp05,Stoudenmire:ARCMP}.
As discussed before, here we consider the effective multi-orbital Hubbard model with four electrons in three orbitals per site (more details can be found in Section II-B), corresponding to the electronic density per orbital $n = 4/3$. Note that this electronic density is widely used in the context of iron low dimensional compounds with DMRG technology, where the ``real'' iron is in a valence Fe$^{\rm 2+}$, corresponding to six electrons in five orbitals per site~\cite{osmp1,Luo:prb10}. To understand the physical properties of this system, we measured several observables based on the DMRG calculations.
The spin-spin correlation in real space are defined as
\begin{eqnarray}
S(r)=\langle{{\bf S}_i \cdot {\bf S}_j}\rangle,
\end{eqnarray}
with $r=\left|{i-j}\right|$, and the spin structure factor is
\begin{eqnarray}
S(q)=\frac{1}{L}\sum_{r}e^{-iqr}S(r).
\end{eqnarray}
The site-average occupancy of orbitals is
\begin{eqnarray}
n_{\gamma}=\frac{1}{L}{\langle}n_{i\gamma\sigma}\rangle.
\end{eqnarray}
The orbital-resolved charge fluctuation is defined as:
\begin{eqnarray}
\delta{n_{\gamma}}=\frac{1}{L}\sum_{i}({\langle}n_{\gamma,i}^2\rangle-{\langle}n_{\gamma,i}{\rangle}^2).
\end{eqnarray}
The local spin-squared, averaged over all sites, is
\begin{eqnarray}
\langle S^2 \rangle=\frac{1}{L}\sum_{i}\langle{{\bf S}_i \cdot {\bf S}_i}\rangle.
\end{eqnarray}
As already explained, the hopping amplitudes were obtained from the {\it ab initio} DFT calculations for Ba$_2$FeS$_3$ (HP) (see APPENDIX A for details). Furthermore, based on the spin-spin correlation and spin structure factor, we calculated the phase diagram of the Ba$_2$FeS$_3$ iron chain with increasing $U/W$ at different Hund couplings $J_H$/$U$, using primarily a system size $L$ = $16$. In the following, we will discuss our main DMRG results at $J_H$/$U$ = 0.25 because this robust $J_H$/$U$ value is believed to be physically realistic
for iron-based superconductors~\cite{Luo:prb10}.
\subsection{A. Staggered AFM phase}
Based on the DMRG measurements of the spin-spin correlation and spin structure factor, we found the paramagnetic phase (PM) at small $U/W$, followed by a robust canonical staggered AFM phase with $\uparrow$-$\downarrow$-$\uparrow$-$\downarrow$ configuration.
Figure~\ref{Fig6}(a) shows the spin-spin correlation $S(r)$=$\langle{{\bf S}_i \cdot {\bf S}_j}\rangle$ vs distance $r$, for different values of $U/W$. The distance is defined as $r=\left|{i-j}\right|$, with $i$ and $j$ site indexes. At small Hubbard interaction $U/W$ $\textless$ 0.6, the spin correlation $S(r)$ decays rapidly vs. distance $r$, indicating paramagnetic behavior, as shown in Fig.~\ref{Fig6}(a) (see result at $U/W$ = 0.3). By increasing $U/W$, the system transitions to the canonical staggered AFM phase with the $\uparrow$-$\downarrow$-$\uparrow$-$\downarrow$ configuration in the whole region of our study ($U/W \le 12$). As shown in Fig.~\ref{Fig6}(b), the spin structure factor $S(q)$ displays a sharp peak at $q = \pi$, corresponding to the canonical staggered AFM phase, consistent with our DFT calculations. In addition, we also calculated the spin-spin correlation $S(r)$ and spin structure factor $S(q)$ using a larger cluster $L = 24$, as shown in Figs.~\ref{Fig6}(c-d). Those results are similar to the results of $L = 16$, indicating that our conclusions of having a canonical staggered AFM phase with $\pi$ vector dominating in the phase diagram is robust against changes in $L$. Note that in one dimension,
quantum fluctuations prevent full long-range order. Thus, the tail of the spin-spin correlations have a smaller value for $L=24$ than for $L=16$. But the staggered order tendency is clear in both cases.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig6.pdf}
\caption{(a) Spin-spin correlation $S(r)=\langle{{\bf S}_i \cdot {\bf S}_j}\rangle$ (with $r=\left|{i-j}\right|$ in real space) and (b) the spin structure factor $S(q)$, both at different values of $U/W$, and all at $J_H$/$U$ = 0.25. We use a chain with $L = 16$. (c) Spin-spin correlation $S(r)=\langle{{\bf S}_i \cdot {\bf S}_j}\rangle$ (with $r = \left|{i-j}\right|$ in real space) and (d) the spin structure factor $S(q)$, at different values of $U/W$ and fixed $J_H$/$U$ = 0.25, for $L = 24$. (e) Magnetic phase diagram for $J_H$/$U$ = 0.25.}
\label{Fig6}
\end{figure}
In the range of $U/W$ we studied, we did not observe any other magnetic ordering tendencies, suggesting the AFM coupling($\uparrow$-$\downarrow$-$\uparrow$-$\downarrow$) is quite stable. This is physically reasonable, considering known facts about the Hubbard model. In the Ba$_2$FeS$_3$ (HP) system, the iron $3d$ orbitals are mainly located in a small energy region and with small bandwidth ($\sim 1$ eV), as shown in Fig.~\ref{Fig2}. By introducing the on-site Hubbard $U$ interaction on Fe sites, the $3d$ orbitals would be easily localized in the Fe sites because the bandwidth is narrow. In this case, the standard superexchange Hubbard spin-spin interaction dominates, leading the spins to order antiferromagnetically along the chain. Note that one orbital ($\gamma=2$) clearly has the largest hopping amplitude from the DFT results, thus this orbital leads in the formation of the AFM order. Due to the large Fe-Fe distance ($\sim 4.3$ \AA) and the special FeS$_4$ chain geometry, in the Ba$_2$FeS$_3$ (HP) system the electrons of the iron $3d$ states are localized with weak $p-d$ hybridization, dominating the superexchange mechanism. Hence, our DMRG results indicating the dominance of the $\uparrow$-$\downarrow$-$\uparrow$-$\downarrow$ configuration are in agreement with our DFT calculations.
\subsection{B. Charge fluctuations}
The site-average occupancy of different orbitals $n_{\gamma}$ vs $U/W$ is shown in Fig.~\ref{Fig7}, for a typical value of $J_H/U$. At small $U/W$ ($\textless 0.6$), a metallic weakly-interacting state is found, with non-integer $n_{\gamma}$ values. In the other extreme of much larger $U/W$, the population of orbital $\gamma = 0$ reaches $2$, and this orbital decouples from the system. Furthermore, the other two orbitals $\gamma = 1$ and $\gamma = 2$ reach population $1$, leading to two half-occupied states. In this extreme $U/W$ case ($n_0$ = 2, $n_1$ =1 and $n_2$ = 1), the system is in a Mott insulator staggered AFM state.
In addition, the average value of the local spin-squared averaged over all sites $\langle{{S}}^2\rangle$ is also displayed in Fig.~\ref{Fig7}, varying $U/W$. The strong local magnetic moments are fully developed with spin magnitude S $\sim 1$, corresponding to four electrons in three orbitals at very large $U/W$.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig7.pdf}
\caption{Orbital-resolved occupation number $n_{\gamma}$, averaged value of the total spin-squared $\langle{{S}}^2\rangle$ vs. $U/W$, at $J_{H}/U = 1/4$. Here, we used a $16$-sites cluster chain with NN hoppings for four electrons in three orbitals.}
\label{Fig7}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig8.pdf}
\caption{Charge fluctuations $\delta{n_{\gamma}}=\frac{1}{L}\sum_{i}({\langle}n_{\gamma,i}^2\rangle-{\langle}n_{\gamma,i}{\rangle}^2)$ vs $U/W$ at $J_{H}/U = 1/4$. Here, we used a $16$-sites cluster chain with NN hoppings for four electrons in three orbitals.}
\label{Fig8}
\end{figure}
To better understand the characteristics of metallic vs insulating behavior in this system, we have also studied the charge fluctuations $\delta{n_{\gamma}}$ for different orbitals, as shown in Fig.~\ref{Fig8}. In the small-$U$ paramagnetic phase ($U/W \textless 0.6$), the system is metallic due to weak interactions. Increasing $U/W$, the charge fluctuations of different orbitals are considerable at intermediate Hubbard coupling strengths, indicating strong quantum fluctuations along the chains. As $U/W$ increases further, the charge fluctuations of $\gamma = 1$ rapidly reaches zero, leading to localized orbital characteristics, while the $\gamma = 2$ orbital still has larger fluctuations with some itinerant electrons. In this case, this intermediate regime corresponds to the OSMP state. At even larger $U/W$ ($\gtrsim 4.5$), the charge fluctuations of the different orbitals are suppressed to nearly zero. Thus, the system becomes fully insulating at very large $U/W$, with two half-filled orbitals ($\gamma = 1$ and $\gamma = 2$) and one fully occupied orbital ($\gamma = 0$) [see Fig.~\ref{Fig7}], as already explained. Here, the charge fluctuations are totally suppressed by the electronic correlations.
\subsection{C. Orbital-selective Mott phase}
Let us now focus on the intermediate regime of OSMP. As displayed in Fig.~\ref{Fig7}, at intermediate Hubbard coupling strengths, the system displays OSMP behavior. In this regime, the $\gamma = 1$ orbital population reaches $1$, indicating localized electronic characteristics, while the other two orbitals have non-integer electronic density, leading to metallic electronic features. Furthermore, we also compare these results with a larger system site $L = 24$ (see APPENDIX B), indicating the conclusion is robust against changes in $L$. Although the site-average occupancy is $1$ (see Fig.~\ref{Fig7}), the $\gamma = 1$ orbital has some charge fluctuations in the region $0.6 \lesssim U/W \textless 2.0$, as shown in Fig.~\ref{Fig8}. Above $U/W = 2.0$, the charge fluctuations of the $\gamma = 1$ orbital remain zero, indicating full localized behavior, while the other two orbitals still have finite values for the charge fluctuations until a larger $U/W \sim 5$.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig9.pdf}
\caption{(a-c) Single-particle spectra $A_{\gamma}$($q, \omega$) for different orbitals at $U/W$ = 1 and $J_H/U = 0.25$. (d-f) Orbital-resolved PDOS} $\rho_{\gamma}$($\omega$) for different orbitals at $U/W = 1.0$ and $J_H/U = 0.25$.
\label{Fig9}
\end{figure}
To better understand the OSMP region, we calculated the single-particle spectra $A_{\gamma}$($q, \omega$) and the orbital-resolved projected density of states (PDOS) $\rho_{\gamma}(\omega)$ vs. frequency $\omega$ by using the dynamical DMRG, where the dynamical correlation vectors were obtained using the Krylov-space approach~\cite{Kuhner:prb,Nocera:pre}. Here, the broadening parameter $\eta = 0.1$ was chosen in our DMRG calculations. The chemical potential is obtained from $\mu = (E_{N+1} - E_{N-1})/2$, where $E_N$ is the ground state energy of the $N$-particle system. The single-particle spectra $A_{\gamma}(q, \omega)= A_{\gamma}(q, \omega \textless \mu)+ A_{\gamma}(q, \omega \textgreater \mu)$ is calculated from the portions of the spectra below and above $\mu$, respectively.
\begin{eqnarray}
&&\nonumber A_{\gamma}(q, \omega \textless \mu)=\frac{-1}{L\pi} \sum_{j}e^{ijq}\\
&& Im \left< \Psi_{\rm GS} \left|c_{j,\gamma} \frac{1}{\omega -H + E_G+i\eta}c_{L/2,\gamma} \right|\Psi_{\rm GS}\right>,
\end{eqnarray}
\begin{eqnarray}
&&\nonumber A_{\gamma}(q, \omega \textgreater \mu)=\frac{-1}{L\pi} \sum_{j}e^{ijq}\\
&& Im \left< \Psi_{\rm GS} \left| c_{j,\gamma}^{\dagger} \frac{1}{\omega +H -E_G+i\eta}c_{L/2,\gamma}^{\dagger} \right|\Psi_{\rm GS} \right>,
\end{eqnarray}
where $j$ is a site, $c_{j,\gamma} = \sum_{\sigma}c_{j,\gamma,\sigma}$ is the fermionic anihilation operator. while $c^{\dagger}_{i,\gamma} = \sum_{\sigma}c^{\dagger}_{j,\gamma,\sigma}$ is the creation operator, $E_G$ is the ground state energy, and $\Psi_{\rm GS}$ is the ground-state wave function of the system.
The corresponding orbital-resolved PDOS $\rho_{\gamma}(\omega)$ was defined as:
\begin{eqnarray}
\rho_{\gamma}(\omega ) = \frac{-1}{\pi}\sum_{q}Im A_{\gamma}(q, \omega).
\end{eqnarray}
where $A_{\gamma}$($q, \omega$) is a single-particle Green's function of the ${\gamma}$ orbital electrons.
We calculated the single-particle spectra $A_{\gamma}$($q, \omega$) and PDOS $\rho_{\gamma}(\omega)$ at $J_{H}/U = 0.25$ and $U/W = 1.0$, as displayed in Fig.~\ref{Fig9}. The ${\gamma} = 0$ and ${\gamma} = 2$ orbitals present a metallic behavior, suggesting the electrons are itinerant. Meanwhile, the ${\gamma} = 1$ orbital displays the Mott transition behavior with the pseudogap characteristic, where there are still some finite charge fluctuations in this orbital. In addition, we also present the single-particle spectra $A_{\gamma}$($q, \omega$) and PDOS $\rho_{\gamma}(\omega)$ for $U/W = 2$ in Fig.~\ref{Fig10}. It is clearly shown that the ${\gamma} = 1$ orbital has a Mott gap, while the other two orbitals have some electronic bands crossing the Fermi level, indicating itinerant electronic behavior.
Hence, in this regime of intermediate Hubbard coupling strengths, the coexistence of localized and itinerant carriers supports the OSMP picture. This OSMP is related to having special conditions in the system, such as different bandwidth and crystal fields, as well as intermediate electronic correlation. Here, the ${\gamma} = 1$ orbital is easier to be localized by Hubbard $U$ than the ${\gamma} = 2$ orbital due to different bandwidths. The OSMP physics has been extensively discussed in experimental and theoretical works in low-dimensional iron systems with electronic density $n = 6$, such as the iron ladders BaFe$_2$Se$_3$~\cite{mourigal:prl15,osmp1,osmp2} and the iron pnictides/chalcogenides superconductors~\cite{OSMP,Yi:prl}. Here, our DMRG results indicate this interesting OSMP state may also appear in the Ba$_2$FeS$_3$ iron chain system (electronic density $n = 6$), and they thus deserve more experimental studies. As the electronic correlation $U/W$ increases, all the orbitals eventually become Mott-localized with the electronic occupancies ($n_0$ = 2, $n_1$ =1 and $n_2$ = 1), as displayed in Fig.~\ref{Fig7}. Then, the MI phase eventually suppresses the OSMP at very large Hubbard coupling.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig10.pdf}
\caption{(a-c) Single-particle spectra $A_{\gamma}$($q, \omega$) for different orbitals at $U/W$ = 1 and $J_H/U = 0.25$. (d-f) Orbital-resolved PDOS $\rho_{\gamma}$($\omega$) for different orbitals, at $U/W = 2$ and $J_H/U = 0.25$.}
\label{Fig10}
\end{figure}
In addition, we also calculated the entanglement entropy to better understand the OSMP-MI phase transition, using the Von Neumann form~\cite{Calabrese:JSM,Eisert:rmp}. As shown in Fig.~\ref{Fig11}, there are three regimes here, corresponding to PM, AFM-OSMP, and AFM-MI states, which is qualitatively in agreement with our results via the spin-spin correlation $S(r)$ and charge fluctuations $\delta{n_{\gamma}}$. At $U/W \geq 0.4$, $S_{\rm VN}$ begins to drop rapidly, corresponding to the PM to AFM-OSMP phase transition. At $U/W \geq 4.5$, $S_{\rm VN}$ smoothly converges to a constant. In fact, this convergence does not reflect on the spin-spin correlation $S(r)$ because the magnetic order does not change from the AFM-OSMP state to the AFM-MI phase. The main difference between AFM-MI and AFM-OSMP relies on the electronic density i.e. whether is localized or not. In this case, this difference between those two states can be reflected in the charge fluctuations $\delta{n_{\gamma}}$, where all the orbitals eventually with increasing $U/W$ become Mott-localized leading to insulating behavior starting approximately at $U/W \sim 4.5$ (see Fig.~\ref{Fig7}). It also should be noticed that finite lattice size effects and the use of a limited number of states in DMRG would affect the specific boundary values of this regime change from delocalized to localized electrons. But the presence of three different regimes in this model was established via the entanglement entropy, qualitatively agreeing with our other DMRG results. Since the two states (AFM-OSMP and AFM-MI) involved in the discussion are both AFM, we believe that the transition from OSMP to MI is not a sharp true phase transition involving a singularity in some quantity (see Fig.~\ref{Fig11}). Hence, we believe it can be better described as a ``rapid crossover'' from AFM-OSMP to AFM-MI.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig11.pdf}
\caption{Von Neumann entanglement entropy ($S_{\rm VN}$) for the three-orbital chain model as a function of $U/W$ at $J_H$/$U$ = 0.25. inset: derivative of $S_{\rm VN}$}
\label{Fig11}
\end{figure}
\subsection{D. Additional Results}
As shown in Fig.~\ref{Fig12}, we present the spin-spin correlation $S(r)$ for several values of $U/W$, at different Hund couplings $J_H$/$U$ = $0.10$, $0.15$, and $0.20$. As the electronic correlation $U/W$ increases, the staggered AFM phase with $\pi$ vector becomes dominant in the entire region, at least within the range we studied. In fact, this staggered AFM order ($\pi$ vector) was also observed in a large regime of the phase diagram in previous mean-field calculations~\cite{Luo:prb14}, although by using different hoppings.
Due to its unique geometric chain configuration, this system displays strong Hubbard superexchange interaction along the chain, which is different from other iron-based chains or ladders. It should be noted that several interesting phases (i.e. block-type $\uparrow$-$\uparrow$-$\downarrow$-$\downarrow$ and FM phases) were found in our previous DMRG phase diagram for a chain system~\cite{Rin:prl,Pandey:prb,Lin:osmp}. Previous work~\cite{Rin:prl} suggests the block-type AFM could be stable due to the competition between the $J_H$ and superexchange interaction. The $J_H$ favors FM ordering, corresponding to the double-exchange interaction in manganites~\cite{Dagotto:rp}, while the superexchange interaction favors AFM ordering. However, in the Ba$_2$FeS$_3$ system of our focus here, the weak $p-d$ hybridization suppresses the double exchange interaction. Thus, the superexchange Hubbard interaction is dominant, leading to robust staggered AFM order. Again, we believe this is because only one of the orbitals has a robust intraorbital hopping, thus dominating the physics.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig12.pdf}
\caption{Spin-spin correlation $S(r)=\langle{{\bf S}_i \cdot {\bf S}_j}\rangle$ (with $r=\left|{i-j}\right|$) in real space for different values of $U/W$ at (a) $J_H$/$U$ = 0.20, (b) $J_H$/$U$ = 0.15, and (c) $J_H$/$U$ = 0.10.}
\label{Fig12}
\end{figure}
In addition, the magnetic phase diagram was calculated varying $J_H$/$U$ and $U/W$, based on the DMRG results (spin-spin correlation $S(r)$ and charge fluctuations $\delta{n_{\gamma}}$). We found three dominant regimes, involving metallic PM, AFM-OSMP, and AFM-MI phases, as shown in Fig.~\ref{Fig13}. Note that the boundaries coupling values should be considered only as crude approximations. However, the existence of the three regions shown was clearly established, even if the boundaries are only rough estimations. We believe our theoretical phase diagram should encourage a more detailed experimental study of iron chalcogenide compounds or related systems.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig13.pdf}
\caption{DMRG phase diagram of the three-orbital Hubbard model varying $U/W$ and $J_H/U$, using a $L = 16$ chain. Different phases are indicated, with the PM, AFM-OSMP, AFM-MI phasese. Small solid circles indicate specific values of data points that were investigated
with DMRG calculations.}
\label{Fig13}
\end{figure}
If the NN distance could be reduced by considering chemical doping or strain effects, the crystal-field splitting and the hybridization would increase. Then, it may be possible to achieve some interesting magnetic phases in this system, as discussed in Refs.{~\cite{Luo:prb14,Rin:prl}}. This maybe a possible direction for further experimental or theoretical studies working on this material or similar variations obtained by altering the 213 chemical formula.
In summary of our DMRG results, after the paramagnetic regime of weak coupling we found the AFM state with $\uparrow$-$\downarrow$-$\uparrow$-$\downarrow$ configuration in our three-orbital Hubbard model, at the robust range of $U/W$ and $J_H$/$U$ that we studied. At intermediate Hubbard coupling strengths, this system displayed OSMP behavior, while the OSMP was suppressed by MI phase at very large $U/W$.
\section{V. Conclusions}
In this publication, we systematically studied the compound Ba$_2$FeS$_3$ (HP) by using first-principles DFT and also DMRG calculations. A strongly anisotropic one-dimensional electronic band structure was observed in the non-magnetic phase, corresponding to its dominant chain geometry. The magnetic coupling along the chain was found to be the key ingredient for magnetism. The staggered magnetic state with a Mott gap was found to be the most likely magnetic ground state among all the candidates studied. Based on the Wannier functions calculated from DFT, we obtained the nearest-neighbor hopping amplitudes and on-site energies for the iron atoms. Then, a multi-orbital Hubbard model for the iron chain was constructed and studied by using the many-body DMRG methodology, considering quantum fluctuations. Based on the DMRG calculations, we obtained a dominant staggered AFM state ($\uparrow$-$\downarrow$-$\uparrow$-$\downarrow$). This staggered $\uparrow$-$\downarrow$ AFM with $\pi$ vector was found in a robust portion of the phase diagram at many values of $U/W$ and $J_H/U$, in agreement with DFT calculations. At intermediate Hubbard coupling strengths, this system displayed orbital-selective Mott phase behavior, corresponding to one localized orbital and two itinerant metallic orbitals, the latter with nonzero charge fluctuations. At larger $U/W$, the system crossovers to a Mott insulating state ($n_0$ = 2, $n_1$ =1 and $n_2$ = 1) with one double occupied orbital ($\gamma = 0$) and two half occupied orbitals ($\gamma = 1$ and $\gamma = 2$).
\section{Acknowledgments}
The work of Y.Z., L.-F.L., A.M. and E.D. is supported by the U.S. Department of Energy (DOE), Office of Science, Basic Energy Sciences (BES), Materials Sciences and Engineering Division. G.A. was partially supported by the scientific Discovery through Advanced Computing (SciDAC) program funded by U.S. DOE, Office of Science, Advanced Scientific Computing Research and BES, Division of Materials Sciences and Engineering. The calculations were carried out at the Advanced Computing Facility (ACF) of the University of Tennessee, Knoxville.
\section{APPENDIX}
\subsection{A. Hoppings}
Here, we focus only on the iron chain since the intrachain coupling is the key aspect to understand the physical properties of Ba$_2$FeS$_3$ (HP). Thus, we used the MLWFs to fit the DFT bands along the $b$-axis (Y-$\Gamma$), corresponding to the quasi-one-dimensional electronic characteristics of Ba$_2$FeS$_3$ (HP), as displayed in Fig.~\ref{Fig14}(a). Based on the Wannier fitting results, we deduced the hopping parameters and on-site matrix.
Considering the computational limitation of the DMRG method, we constructed a three-orbital model involving the orbital basis $d_{xz}$, $d_{x^2-y^2}$ and $d_{xy}$ for the iron chain, readjusted to properly fit the band structure after reducing the original five orbitals to three. The three-orbital tight-binding bands agree qualitatively well with the DFT band structure, as displayed in Fig.~\ref{Fig14}(b).
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig14a.pdf}
\includegraphics[width=0.48\textwidth]{Fig14b.pdf}
\caption{(a) DFT and Wannier bands of Ba$_2$FeS$_3$ (HP) along the FeS$_4$ chain direction ($b$-axis). (b) Three-orbital tight-binding model with nearest-neighbor hoppings along the $b$ axis. The BZ points are Y = (0, 0.5, 0) and $\Gamma$ = (0, 0, 0). Note that Y is in scaled units, corresponding to the units of $2\pi$/b.}
\label{Fig14}
\end{figure}
Based on the Wannier fitting, we obtained four on-site matrices for the four Fe atoms in a unit cell, using the basis \{$d_{z^2}$, $d_{xz}$, $d_{yz}$, $d_{x^2-y^2}$, $d_{xy}$\}.
\begin{equation}
\begin{split}
t_{onsite}^1 =
\begin{bmatrix}
d_{z^2} & d_{xz} & d_{yz} & d_{x^2-y^2} & d_{xy} \\
3.812 & -0.063 & 0.000 & 0.075 & 0.000 \\
-0.063 & 3.628 & 0.000 & 0.183 & 0.000 \\
0.000 & 0.000 & 3.509 & 0.000 & -0.054 \\
0.075 & 0.183 & 0.000 & 3.644 & 0.000 \\
0.000 & 0.000 & -0.054 & 0.000 & 3.618
\end{bmatrix},\\
\end{split}
\label{onsite1}
\end{equation}
\begin{equation}
\begin{split}
t_{onsite}^2 =
\begin{bmatrix}
3.812 & -0.063 & 0.000 & 0.075 & 0.000 \\
-0.063 & 3.628 & 0.000 & 0.183 & 0.000 \\
0.000 & 0.000 & 3.509 & 0.000 & -0.054 \\
0.075 & 0.183 & 0.000 & 3.644 & 0.000 \\
0.000 & 0.000 & -0.054 & 0.000 & 3.618
\end{bmatrix},\\
\end{split}
\label{onsite2}
\end{equation}
\begin{equation}
\begin{split}
t_{onsite}^3 =
\begin{bmatrix}
3.812 & 0.063 & 0.000 & 0.075 & 0.000 \\
0.063 & 3.628 & 0.000 & -0.183 & 0.000 \\
0.000 & 0.000 & 3.509 & 0.000 & 0.054 \\
0.075 & -0.183 & 0.000 & 3.644 & 0.000 \\
0.000 & 0.000 & 0.054 & 0.000 & 3.618
\end{bmatrix},\\
\end{split}
\label{onsite3}
\end{equation}
\begin{equation}
\begin{split}
t_{onsite}^4 =
\begin{bmatrix}
3.812 & 0.063 & 0.000 & 0.075 & 0.000 \\
0.063 & 3.628 & 0.000 & -0.183 & 0.000 \\
0.000 & 0.000 & 3.509 & 0.000 & 0.054 \\
0.075 & -0.183 & 0.000 & 3.644 & 0.000 \\
0.000 & 0.000 & 0.054 & 0.000 & 3.618
\end{bmatrix}.\\
\end{split}
\label{onsite4}
\end{equation}
Furthermore, we also obtained four nearest-neighbors hopping matrices along the $b$-axis, corresponding to the four Fe atoms in a unit cell.
\begin{equation}
\begin{split}
t_{\vec{b}}^1 =
\begin{bmatrix}
0.057 & -0.094 & 0.071 & 0.048 & 0.011 \\
-0.094 & -0.003 & -0.019 & 0.020 & -0.083 \\
-0.071 & 0.019 & -0.016 & 0.055 & -0.132 \\
0.048 & 0.020 & -0.055 & 0.169 & 0.022 \\
-0.011 & 0.083 & -0.132 & -0.022 & 0.172
\end{bmatrix},\\
\end{split}
\label{hopping1}
\end{equation}
\begin{equation}
\begin{split}
t_{\vec{b}}^2 =
\begin{bmatrix}
0.057 & -0.094 & -0.071 & 0.048 & -0.011 \\
-0.094 & -0.003 & 0.019 & 0.020 & 0.083 \\
0.071 & -0.019 & -0.016 & -0.055 & -0.132 \\
0.048 & 0.020 & 0.055 & 0.169 & -0.022 \\
0.011 & -0.083 & -0.132 & 0.022 & 0.172
\end{bmatrix},\\
\end{split}
\label{hopping2}
\end{equation}
\begin{equation}
\begin{split}
t_{\vec{b}}^3 =
\begin{bmatrix}
0.057 & 0.094 & 0.071 & 0.048 & -0.011 \\
0.094 & -0.003 & 0.019 & -0.020 & -0.083 \\
-0.071 & -0.019 & -0.016 & 0.055 & 0.132 \\
0.048 & -0.020 & -0.055 & 0.169 & -0.022 \\
0.011 & 0.083 & 0.132 & 0.022 & 0.172
\end{bmatrix},\\
\end{split}
\label{hopping3}
\end{equation}
\begin{equation}
\begin{split}
t_{\vec{b}}^4 =
\begin{bmatrix}
0.057 & 0.094 & -0.071 & 0.048 & 0.011 \\
0.094 & -0.003 & -0.019 & -0.020 & 0.083 \\
0.071 & 0.019 & -0.016 & -0.055 & 0.132 \\
0.048 & -0.020 & 0.055 & 0.169 & 0.022 \\
-0.011 & -0.083 & 0.132 & -0.022 & 0.172
\end{bmatrix}.\\
\end{split}
\label{hopping4}
\end{equation}
As shown above, there are some non-zero off-diagonal elements in the on-site matrices, indicating the constructed MLWFs orbitals are not exactly orthogonal
to one other. Hence, we introduced a unitary matrix transformation
to reconstruct the effective on-site and hopping matrices:
\begin{equation}
\begin{split}
U =
\begin{bmatrix}
0.881 & -0.246 & 0.000 & -0.406 & 0.000 \\
0.131 & -0.696 & 0.000 & 0.706 & 0.000 \\
0.000 & 0.000 & 0.925 & 0.000 & -0.381 \\
0.456 & 0.675 & 0.000 & 0.580 & 0.000 \\
0.000 & 0.000 & 0.381 & 0.000 & 0.925
\end{bmatrix},\\
\end{split}
\label{unit}
\end{equation}
As discussed in the main text, the Ba$_2$FeS$_3$ (HP) is a quasi-one-dimensional system, where the physical properties are primarily contributed by the intrachain coupling. Hence, we just considered one iron chain and NN hopping in our DMRG calculations. The reconstructed on-site and hopping matrices are:
\begin{equation}
\begin{split}
t_{onsite}^1 =
\begin{bmatrix}
d_{z^2} & d_{xz} & d_{yz} & d_{x^2-y^2} & d_{xy} \\
3.841 & 0.000 & 0.000 & 0.000 & 0.000 \\
0.000 & 3.428 & 0.000 & 0.000 & 0.000 \\
0.000 & 0.000 & 3.487 & 0.000 & 0.000 \\
0.000 & 0.000 & 0.000 & 3.814 & 0.000 \\
0.000 & 0.000 & 0.000 & 0.000 & 3.640
\end{bmatrix},\\
\end{split}
\label{onsiteR}
\end{equation}
\begin{equation}
\begin{split}
t_{\vec{b}}^1 =
\begin{bmatrix}
0.098 & 0.119 & 0.035 & -0.006 & -0.005 \\
0.119 & 0.012 & -0.011 & 0.045 & 0.080 \\
-0.035 & 0.011 & -0.082 & 0.087 & -0.028 \\
-0.006 & 0.045 & -0.087 & 0.112 & -0.018 \\
0.005 & -0.080 & -0.028 & 0.018 & 0.238
\end{bmatrix}.\\
\end{split}
\label{hoppingR}
\end{equation}
Here, we used the three orbitals \{$d_{xz}$, $d_{x^2-y^2}$, $d_{xy}$\} in our calculations, corresponding to the electronic density per orbital $n = 4/3$. As explained before, this electronic density is widely used in the context of iron low-dimensional compounds with DMRG technology, where the ``real'' iron is in a valence Fe$^{\rm 2+}$, corresponding to six electrons in five orbitals per site~\cite{osmp1,Luo:prb10}. In our DMRG calculations, the on-site and hopping matrices are:
\begin{equation}
\begin{split}
t_{onsite} =
\begin{bmatrix}
d_{xz} & d_{x^2-y^2} & d_{xy} \\
3.428 & 0.000 & 0.000 \\
0.000 & 3.814 & 0.000 \\
0.000 & 0.000 & 3.640
\end{bmatrix},\\
\end{split}
\label{onsiteF}
\end{equation}
\begin{equation}
\begin{split}
t_{\gamma\gamma'} =
\begin{bmatrix}
0.012 & 0.045 & 0.080 \\
0.045 & 0.112 & -0.018 \\
-0.080 & 0.018 & 0.238
\end{bmatrix}.\\
\end{split}
\end{equation}
\subsection{B. DMRG results for $L = 24$}
As displayed in Fig.~\ref{Fig15}, we show the site-averaged occupancy of different orbitals $n_{\gamma}$ vs $U/W$ for $L = 24$, at the typical value of $J_H/U$. Those results are similar to the results of $L = 16$ (Fig.~\ref{Fig7}), indicating that our results are robust against changes in $L$ (small size effects).
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig15.pdf}
\caption{Orbital-resolved occupation number $n_{\gamma}$, mean-value of the total spin-squared $\langle{{S}}^2\rangle$, at different values of $U/W$ and $J_{\rm H}/U = 1/4$. Here, we used a $24$-sites cluster chain with nearest-neighbor hoppings for four electrons in three orbitals.}
\label{Fig15}
\end{figure}
|
\section{Introduction}
Recommender systems have been an important tool for online commerce platforms to assist users in filtering for the best content while shaping their interests at the same time. To achieve this objective, collaborative filtering has been among the most popular techniques and was successfully deployed for decades. However as shown in the recent literature, classical collaborative filtering algorithms often lead to the problem of over-specialization \cite{adamopoulos2014over}, filter bubbles \cite{pariser2011filter,nguyen2014exploring} and user boredom \cite{kapoor2015like,kapoor2015just}. To make things worse, users might get annoyed if they are recommended similar items repeatedly in a short period of time.
To address these issues, researchers have proposed to incorporate unexpectedness metric in recommendation models \cite{murakami2007metrics}, the goal of which is to provide novel, surprising and satisfying recommendations. Unlike diversity, which only measures dispersion among recommended items, unexpectedness measures deviations of recommended items from user expectations and thus captures the concept of user surprise and allows recommender systems to break from the filter bubble. It has been shown in \cite{adamopoulos2014discovering,adamopoulos2015unexpectedness} that unexpectedness leads to significant increase of user satisfaction. Therefore, researchers have introduced various recommendation algorithms to optimize the unexpectedness metric and achieved strong recommendation performance \cite{adamopoulos2015unexpectedness,zhang2012auralist,lu2012serendipitous,li2019latent3}.
However, very few of them have been successfully applied to industrial applications or achieved significant improvements in real business settings. This is the case for the following reasons. First, to improve unexpectedness of recommended items, previous models often have to sacrifice the business-oriented metrics \cite{adomavicius2008overcoming,zhou2010solving}, such as CTR (Click-Through-Rate) and GMV (Gross-Merchandise-Volume). Second, most of the proposed unexpected recommendation algorithms are not scalable and therefore cannot be deployed in large-scale industrial applications. Third, the lack of personalization and session-based design of recommender systems reduces user satisfaction and their online browsing performance metrics. Thus, it is important to overcome these problems and intelligently deploy unexpected recommender systems in real-world applications.
Note that, while working on unexpected recommendations, it is crucial to address the problem of heterogeneous user preferences by focusing on providing personalized recommendations according to these preferences. For example, some people tend to be `'variety-seekers'' \cite{mcalister1982variety} and are more willing to click on novel item recommendations, while others prefer to stay in their own comfort zones and are in favor of familiar item recommendations. In addition, even the same person might have different perceptions of unexpectedness in different contexts, which also motivates us to include session-based information into the design of an unexpected recommender system. For example, it is more reasonable to recommend the next episode of a TV series to the user who has just finished the first episode, instead of recommending new types of videos to that person. On the other hand, if the user has been binge-watching the same TV series in one night, it is better to recommend something different to him or her.
To address these concerns, in this paper we propose the \textit{Personalized Unexpected Recommender System} (PURS) that incorporates unexpectedness into recommendations in an end-to-end scalable manner. Specifically, we model user interests as \textit{clusters} of the previously consumed items in the \textit{latent} space and subsequently calculate unexpectedness as the weighted distance between a new item and the clusters of these interests. Furthermore, we utilize sequence modeling and the self-attention mechanism to capture personalized and session-based user perception of unexpectedness. We also propose a novel unexpected activation function to adjust the network output for better unexpected recommendation performance. To provide unexpected recommendations, we construct a hybrid utility function by combining the aforementioned unexpectedness measure with estimated business performance metrics, especially the click-through-rate. Finally, extensive offline experiments on three real-world datasets illustrate that the proposed model consistently and significantly outperforms the state-of-the-art baseline approaches in both accuracy and novelty measures. We also conduct an online A/B test on a major video platform Alibaba-Youku, where our model achieves significant improvements over the current production model. Based on these positive results, the proposed model is in the process of being moved into production in the company.
This paper makes the following research contributions. It
(1) presents a novel PURS model that efficiently and effectively incorporates unexpectedness into recommendations;
(2) proposes to use the self-attention mechanism to model personalized and session-based user perception of unexpectedness;
(3) proposes to calculate unexpectedness as the distance between a new item and clusters of user interests in the latent space;
(4) presents a novel unexpected activation function to optimize performance of unexpected recommendations;
(5) presents extensive offline experiments and an online A/B test that empirically demonstrate the strengths of the PURS model.
\section{Related Work}
In this section, we discuss prior literature related to our work in three categories: unexpected recommendations, deep-learning based recommendations and personalized \& session-based recommendations.
\subsection{Unexpected Recommendations}
As discussed in the previous section, to overcome the problem of filter bubbles and user boredom, researchers have proposed to optimize beyond-accuracy objectives, including unexpectedness, serendipity, novelty, diversity and coverage \cite{murakami2007metrics,ge2010beyond}. Note that, these metrics are closely related to each other, but still different in terms of definition and formulation. Specifically, serendipity measures the positive emotional response of the user about a previously unknown item and indicates how surprising these recommendations are to the target users \cite{shani2011evaluating,chen2019serendipity}; novelty measures the percentage of new recommendations that the users have not seen before or known about \cite{mcnee2006being}; diversity measures the variety of items in a recommendation list, which is commonly modeled as the aggregate pairwise similarity of recommended items \cite{ziegler2005improving}; coverage measures the degree to which recommendations cover the set of available items \cite{herlocker2004evaluating}.
Among them, unexpectedness has attracted particular research interests, for it is shown to be positively correlated with user experience \cite{adamopoulos2015unexpectedness,chen2019serendipity}. It also overcomes the overspecialization problem \cite{adamopoulos2015unexpectedness,iaquinta2010can}, broadens user preferences \cite{herlocker2004evaluating,zhang2012auralist,zheng2015unexpectedness} and increases user satisfaction \cite{adamopoulos2015unexpectedness,zhang2012auralist,lu2012serendipitous}. Unexpectedness measures those recommendations that are not included in the users' previous purchases and depart from their expectations. Different from the diversity measure, unexpectedness is typically defined as the distance between the recommended item and the set of expected items in the feature space \cite{adamopoulos2015unexpectedness}. However, as pointed out in \cite{li2019latent3,li2020latent}, it is difficult to properly construct the distance function in the feature space, thus the authors propose to calculate the distance of item embeddings in the latent space instead.
Researchers have thus proposed multiple recommendation models to improve novelty measures, including Serendipitous Personalized Ranking (SPR) \cite{lu2012serendipitous} that extends traditional personalized ranking methods by discounting item popularity in AUC optimization; Auralist \cite{zhang2012auralist} that balances the accuracy and novelty measures simultaneously using topic modeling; Determinantal Point Process (DPP) \cite{gartrell2017low,chen2018fast} that utilizes the greedy MAP inference to diversify the recommendation results; HOM-LIN \cite{adamopoulos2015unexpectedness} that use a hybrid utility function of estimated ratings and unexpectedness to provide recommendations; and Latent Modeling of Unexpectedness \cite{li2019latent3,li2020latent}. All of these models have successfully improved the unexpectedness measure of recommendations.
However, these prior literature aim to provide uniform unexpected recommendations to all users under all circumstances, without taking into account user-level heterogeneity and session-based information, thus might not reach the optimal recommendation performance. In addition, those models have the scalability issue and might not work well in the large-scale industrial settings. In this paper, we propose a novel deep-learning based unexpected recommender system to provide personalized and session-based unexpected recommendations in an efficient manner.
\subsection{Deep-Learning based Recommendations}
Another body of related work is around deep-learning based approaches to extract user preferences and provide recommendations. Compared with classical recommender systems, deep-learning based models are capable of conducting representation learning, sequence modeling and nonlinear transformation to possess high level of flexibility \cite{zhang2019deep}. Some representative work include \cite{shi2017survey,shi2018heterogeneous,dong2017metapath2vec} that construct heterogeneous information network embeddings to model complex heterogeneous relations between users and items; \cite{rumelhart1985learning,he2017neural} that apply deep neural network (DNN) techniques to obtain semantic-aware representations of users and items for efficient learning and recommendation; and \cite{hinton2006reducing,sedhain2015autorec} that utilize autoencoder (AE) techniques to map the features of users and items into latent embeddings and model their relationships accordingly.
In this paper, we adopt deep-learning based approaches to model unexpectedness and subsequently provide unexpected recommendations, thus obtaining all these advantages discussed above.
\subsection{Personalized and Session-based Recommendations}
Note that different users might have different preferences towards the same recommendation, and even the same user might have different opinions towards the same recommendations on different sessions. Therefore, personalized and session-based recommender systems constitute important tools for providing satisfying recommendations. Specifically, these models take user behavior sequences into account to learn user behavior patterns and the shift of user preferences from one transaction to another \cite{wang2019survey}. Recent work on this field include DIN \cite{zhou2018deep}, DSIN \cite{feng2019deep}, DIEN \cite{zhou2019deep}, DeepFM \cite{guo2017deepfm}, Wide \& Deep \cite{cheng2016wide} and PNN \cite{qu2016product} that combine user features, item features and user historic behaviors to provide useful recommendations.
In this paper, we incorporate personalized and session-based information into the model design to provide unexpected recommendations and improve user satisfaction.
\section{Unexpectedness}
\subsection{Modeling of Unexpectedness}
In \cite{adamopoulos2015unexpectedness}, the authors define unexpectedness of a recommended item as the distance between the new item and the closure of previous consumptions in the \textit{feature} space. As discussed in the related work section, this approach might not achieve optimal performance, for the distance function is typically difficult to define in the feature space. Therefore in \cite{li2019latent3}, unexpectedness is constructed in the \textit{latent} space instead of the feature space, which is calculated as the euclidean distance to the entire latent closure of previous consumptions. This latent modeling approach of unexpectedness manages to achieve strong recommendation performance and improve novelty of recommendations without losing accuracy measures.
Note however, that many users have diverse interests, thus making it unreasonable to use one single closure of consumed items to model user interests. For example, in the movie recommendation applications, a user might have several different types of movie interests, including documentaries, fiction movies, comedies, animation and so on. Therefore, it is suboptimal to use one single closure to model wide interests and might be more appropriate to model user interests in separate clusters instead. In addition, as shown in Figure \ref{definition}, if we take one single closure including all items that the user has consumed as the expected set, it might lead to the construction of an extremely large expected set and might accidentally include some unexpected items into the expected set as well. However, if we cluster the user's historic consumptions based on the diverse interests, it would be easier to identify the behavior patterns within each cluster and model the expectation sets accordingly. It also encourages those unexpected recommendations that could bridge the diverse interests of the same user, as illustrated in Figure \ref{definition}.
\begin{figure*}
\centering
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{picture/unexp.png}
\caption{Item Embeddings}
\end{subfigure}%
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{picture/unexp1.jpg}
\caption{One Single Latent Closure}
\end{subfigure}%
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{picture/unexp2.jpg}
\caption{Clustering of Latent Closure}
\end{subfigure}%
\caption{Comparison of unexpectedness modeling in the latent space. Blue points stand for the available items; Orange points represent the consumed items; Green point refers to the new recommended item. We propose to define unexpectedness as the distance between the new item and clusters of latent closure generated by all previous consumptions.}
\label{definition}
\end{figure*}
Therefore, we propose to conduct clustering on embeddings of previous consumptions and form user interest clusters accordingly. We select the Mean Shift algorithm \cite{cheng1995mean} to identify the clusters of historic behaviors automatically in the latent space for the following reasons. First, it is an unsupervised clustering algorithm, therefore we do not have to explicitly specify the number of interest clusters for each user, as we generally do not have that information as well. Mean Shift algorithm is capable of optimally selecting the best number of clusters without manual specification. Second, it is powerful for the analysis of a complex multi-modal feature space for recommendation applications, and it can also delineate arbitrarily shaped clusters in it \cite{comaniciu2002mean}.
Mean Shift clustering utilizes an iterative process to locate the maxima of a density function given discrete data sampled from that function. To handle the clustering procedure in our model, we denote the kernel function as $K(x_i - x)$ given an initial estimate $x_{i}$ and observation $x$. This kernel function determines the weight of nearby points for re-estimation of the mean, which is typically modeled as a Gaussian distribution
\begin{equation}
K(x_{i}-x)=e^{-c||x_{i}-x||^{2}}
\end{equation}
The weighted mean of the density in the window determined by $K$ is calculated as
\begin{equation}
m(x) = \frac{\sum_{x_i \in N(x)} K(x_i - x) x_i }{\sum_{x_i \in N(x)} K(x_i - x)}
\end{equation}
where $N(x)$ is the neighborhood of $x$. The mean-shift algorithm will reset $x$ as $m(x)$, and repeat the estimation process until $m(x)$ converges.
For each user $u$, we extract the historic behavior sequence as $\{i_{1}, i_{2}, \cdots, i_{n}\}$ and their corresponding embeddings in the latent space as $\{w_{1}, w_{2}, \cdots, w_{n}\}$ through sequence modeling. We subsequently apply Mean Shift algorithm to cluster these embeddings into user interest clusters as $\{C_{1}, C_{2}, \cdots, C_{N}\}$. For each new item recommendation $i_{*}$ for user $u$, unexpectedness is hereby modeled as the weighted average distance between each cluster $C_{k}$ and embedding of the new item $w_{*}$:
\begin{equation}
unexp_{i_{*},u} = \sum_{k=1}^{N} d(w_{*}, C_{k})\times\frac{|C_{k}|}{ \sum_{k=1}^{N} |C_{k}|}
\end{equation}
\subsection{Unexpected Activation Function}
Though it is natural to directly include the unexpectedness obtained from the previous section into the utility function, it is suboptimal to do so, as our goal is to provide unexpected yet relevant and useful recommendations. If we explicitly combine unexpectedness into the hybrid utility function, it will tend to recommend items with very high unexpectedness, which are likely to be either irrelevant or even absurd to the user. In \cite{adamopoulos2015unexpectedness}, the authors propose to use a unimodal function to adjust the unexpectedness input to the utility function.
However, unimodality alone is not enough for any function to serve as the unexpected activation function, as we also need to balance between unexpectedness and relevance objectives. For example, the unimodal Gaussian function does not obatin the optimal recommendation performance, as shown in Section 6. We consequently propose that the unexpected activation function $f(\cdot)$ should satisfy the following mathematical properties:
\begin{enumerate}
\item \textbf{Continuity} For two items $i_{1}$ and $i_{2}$, if their unexpectedness towards user $u$ are close enough ($|unexp_{u,i_{1}}-unexp_{u,i_{2}}|<\epsilon$), then their unexpectedness output to respective utility should also be close ($|f(unexp_{u,i_{1}})-f(unexp_{u,i_{2}})|<\delta$), which implies the continuity requirement of $f(\cdot)$.
\item \textbf{Boundedness} Note that the utility function $utility_{u,i}$ for any user $u$, item $i$ should be bounded, therefore the corresponding unexpectedness output and the unexpected activation function should also be bounded ($|f(\cdot)|<\infty$). In addition, when unexpectedness goes to zero or infinity, it suggests the new item is either too similar to previous consumptions or totally irrelevant, thus its contribution to utility function should be negligible ($lim_{x\to 0} f(x) = 0$, $lim_{x\to \infty} f(x) = 0$).
\item \textbf{Unimodality} For the optimization convenience, it is ideal to have a unimodal unexpectedness output to the utility function instead of multi-peaks which require additional effort for the recommendation model to balance between unexpectedness and relevance, as discussed in \cite{adamopoulos2015unexpectedness}. Therefore, we need to select a unimodal function as the activation function.
\item \textbf{Short-Tailed} To provide unexpected yet relevant recommendations, it is important to note that relevance decreases very fast after unexpectedness increases above certain threshold, which indicates the use of a short-tailed or sub-Gaussian distribution \cite{park2008long}. Specifically, a sub-Gaussian distribution is a probability distribution with strong tail decay, whose tails are dominated by or decay at least as fast as the tails of a Gaussian. Therefore, the activation function should follow the sub-Gaussian distribution, the kurtosis of which should be less than 3 \cite{buldygin1980sub}.
\end{enumerate}
Many functions satisfy the properties above and thus become potential candidates for the unexpected activation function. To simplify the model and accelerate the optimization process, in this paper we choose a popular solution $f(x)=x*e^{-x}$ from the Gamma function \cite{davis1959leonhard} as the unexpected activation function, which satisfies all four required mathematical properties--it is a continuous, bounded and unimodal function with kurtosis value 1.5. As we show in Section 6.3, the unexpected activation function contributes significantly to the superior recommendation performance.
\begin{figure}
\centering
\includegraphics[width=0.3\textwidth]{picture/function.png}
\caption{Unexpected Activation Function}
\label{function}
\end{figure}
\section{Personalized Unexpected Recommender System}
\subsection{Overview}
To provide unexpected recommendations, we propose to use the following hybrid utility function for user $u$ and item $i$
\begin{equation}
Utility_{u,i} = r_{u,i} + f(unexp_{u,i})*unexp\_factor_{u,i}
\end{equation}
which consists of the following components:
\bm{$r_{u,i}$} that represents the click-through-rate estimation for user $u$ towards item $i$ based on their features and past behaviors.
\bm{$unexp_{u,i}$} that represents the unexpectedness of item $i$ towards user $u$, as introduced in the previous section.
\bm{$unexp\_factor_{u,i}$} that represents the personalized and session-based perception of unexpectedness for user $u$ towards item $i$.
\bm{$f(\cdot)$} that stands for the activation function for unexpectedness in order to effectively and efficiently incorporate this piece into the utility function, as introduced in the previous section.
The proposed recommendation model is presented in Figure \ref{model}, which consists of two components: the base model, which estimates the click-through-rate of certain user-item pairs, as we will discuss in this section; and the unexpected model, which captures unexpectedness of the new recommendation as well as user perception towards unexpectedness, as we have discussed in the last section. The unexpected recommendations are provided through joint optimization of the utility function.
\begin{figure*}
\centering
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{picture/model1.jpg}
\caption{Base}
\end{subfigure}%
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{picture/model2.jpg}
\caption{Unexpected}
\end{subfigure}%
\caption{Overview of the proposed PURS model. The base model estimates the click-through-rate of certain user-item pairs, while the unexpected model captures unexpectedness of the new recommendation as well as user perception towards unexpectedness.}
\label{model}
\end{figure*}
\subsection{User and Item Embeddings}
To effectively identify user interests and provide corresponding item recommendations, it is crucial to capture the feature-level information about the users and the items, which can be used for recommendation purposes in many different ways.
In addition, the intrinsic nature of users and items play an important role in determining the success of the recommendations. Some people, for example heavy users of the online video platform, tend to trust more on the recommendations provided by the platform. Thus for a certain recommendation for them, the utility function tend to be higher. On the other hand, the quality of an item is also crucial for the click-through-rate estimation, for the better quality of recommended items would lead to an increase of user satisfaction.
In this paper, we capture this user and item information in the form of \textit{embeddings} in the latent space and utilize the deep-learning based autoencoder approach to obtain these user and item embeddings and also to capture their interactions in the latent space. Specifically, we denote the explicit features for user $u$ and item $i$ as $W_{u} = [w_{u_{1}}, w_{u_{2}}, \cdots, w_{u_{m}}]$ and $W_{i} = [w_{i_{1}}, w_{i_{2}}, \cdots, w_{i_{n}}]$ respectively, where $m$ and $n$ stand for the dimensionality of user and item feature vectors. The goal is to train two separate neural networks: the encoder network that maps feature vectors into latent embeddings, and the decoder network that reconstructs feature vectors from latent embeddings. Due to effectiveness and efficiency of the training process, we represent both the encoder and the decoder as multi-layer perceptron (MLP). The MLP network learns the hidden representations by optimization reconstruction loss $L$:
\begin{equation}
L = ||W_{u}-MLP_{dec}^{u}(MLP_{enc}^{u}(W_{u}))||
\end{equation}
\begin{equation}
L = ||W_{i}-MLP_{dec}^{i}(MLP_{enc}^{i}(W_{i}))||
\end{equation}
where $MLP_{enc}$ and $MLP_{dec}$ represents the MLP network for encoder and decoder respectively. The MLP network is separately trained for obtaining user embedding and item embeddings, and is updated through back-propagation from the utility function optimization.
\subsection{Click-Through-Rate Estimation ($r_{u,i}$) using Self-Attentive GRU}
For a specific recommended item $i$, our goal is to predict whether user $u$ would click on this recommendation or not, which largely depends on the matching between the content of item $i$ and the interest of user $u$. This indicates the importance of precisely inferring user preferences from historic behaviors.
We denote the previous consumption of user $u$ as the sequence $P_{u} = [i_{u_{1}}, i_{u_{2}}, \cdots, i_{u_{K}}]$ The click-through-rate prediction model obtains a fixed-length representation vector of user interests by pooling all the embedding vectors over the user behavior sequence $P_{u}$. We follow the idea of sequence modeling and utilize the bidirectional GRU \cite{cho2014learning} neural networks to obtain the sequence embeddings. It is important to use the recurrent neural network to model user interests, for it is capable of capturing the time information and the order of user purchase, as more recent behavior would naturally have a higher impact on the current recommendation than previous actions. In addition, compared with other recurrent models like RNN or LSTM, GRU is computationally more efficient and better extracts semantic relationships \cite{chung2014empirical}.
During the training process, we first map the behavior sequence to the corresponding item embeddings obtained in the previous stage. To illustrate the GRU learning procedure, we denote $W_{z}$,$W_{r}$,$U_{z}$ and $U_{r}$ as the weight matrices of current information and the past information for the update gate and the reset gate respectively. $x_{t}$ is the behavior embedding input at timestep $t$, $h_{t}$ stands for the output user interest vector, $z_{t}$ denotes the update gate status and $r_{t}$ represents the status of reset gate. The hidden state at timestep $t$ could be obtained following these equations:
\begin{equation}
z_t = \sigma_g(W_{z} x_t + U_{z} h_{t-1} + b_z)
\end{equation}
\begin{equation}
r_t = \sigma_g(W_{r} x_t + U_{r} h_{t-1} + b_r)
\end{equation}
\begin{equation}
h_t = (1-z_t) \circ h_{t-1} + z_t \circ \sigma_h(W_{h} x_t + U_{h} (r_t \circ h_{t-1}) + b_h)
\end{equation}
Note that, each historic consumption might have different influence on the current recommendation. For example, if a user have watched the documentary ``Deep Blue'' from BBC and is very satisfied with the viewing experience, the user will be more likely to accept the current recommendation of documentary ``Earth'', also from BBC. Meanwhile, the historic record of watching James Bond might have relatively smaller influence towards the acceptance of that documentary recommendation. Therefore, to capture the item-level heterogeneity in the behavior sequence, we propose to incorporate self-attentive mechanism \cite{shaw2018self} during the sequence modeling process. Typically, each output element $s_{t}$ is computed as weighted sum of a linearly transformed input elements
\begin{equation}
s_{t}=\sum_{i=1}^{n}\alpha_{ti}(x_{i}W^{t})
\end{equation}
Each weight coefficient $\alpha_{ti}$ is computed using a softmax function
\begin{equation}
\alpha_{ti}=\frac{\exp{e_{ti}}}{\sum_{i=1}^{n}\exp{e_{ti}}}
\end{equation}
where $e_{ti}$ is computed using a compatibility function that compares two input elements $x_{t}$ and $x_{i}$ correspondingly.
By iteratively calculating hidden step throughout every time step, we obtain final hidden state at the end of the behavior sequence, which constitutes the user interest embeddings $R_{u}$ that captures the latent semantic information of the user's historic actions. To provide the click-through-rate estimation, we concatenate the obtained user interest embeddings $R_{u}$ with user embeddings $E_{u}$ and item embeddings $E_{i}$ extracted in the previous section and feed into a MLP network to get the prediction: $r_{u,i} = MLP(R_{u};E_{u};E_{i})$
\subsection{Unexpected Factor ($unexp\_factor_{u,i}$) using Self-Attentive MLP}
As we have discussed in the previous section, different people might have difference preferences towards unexpected recommendations, and their perception of unexpectedness is also influenced by the session-based information. Therefore, we need to take the user's historic actions in account when computing $unexp\_factor_{u,i}$ and provide personalized session-based recommendations.
We denote the previous consumption of user $u$ as the sequence $P_{u} = [i_{u_{1}}, i_{u_{2}}, \cdots, i_{u_{K}}]$ Following the idea of session-based recommendation \cite{ludewig2018evaluation}, instead of using the entire sequence to measure the factor of unexpectedness, we propose to only use a window of purchased items to identify the factor. The specific length of the window $K$ is a hyperparameter that we can adjust to get the optimal performance.
The most recent user actions in the window will then be extracted and serve as the input to the MLP network. To capture the heterogeneity of the extracted historic behavior towards current unexpected recommendations, we utilize the structure of local activation unit \cite{zhou2018deep} to determine whether the embedding of each item will be fed into the network. Instead of expressing the user's diverse interests with the same network structure, local activation unit is capable of adaptively calculating the relevance of historical behaviors towards current candidate recommendations.
Specifically, local activation unit performs a weighted sum pooling to adaptively calculate the activation stage of each behavior embedding and generate one single representation. We denote the sequence of item embeddings for user $u$ in the session-window as $P_{u} = [E_{i_{1}}, E_{i_{2}}, \cdots, E_{i_{K}}]$ For user $u$ and item $i$, the unexpected factor $unexp\_factor_{u,i}$ for this user-item pair will be calculated as
\begin{equation}
unexp\_factor_{u,i} = MLP(E_{u};\sum_{j=1}^{K}a(E_{u}, E_{i_{j}}, E_{i})E_{i_{j}};E_{i})
\end{equation}
where $a(\cdot)$ is a feed-forward network with output as the activation weight for each past purchase.
\section{Experiments}
In this section, we introduce extensive experiments that validate the superior recommendation performance of the proposed model in terms of both accuracy and novelty measures. The hyperparameters are optimized through Bayesian optimization \cite{snoek2012practical}. For all experiments, we select $K=10$ and use SGD as the optimizer with learning rate 1 and exponential decay 0.1. The dimensionality of embedding vector is 32. Layers of MLP is set by $32\times 64\times 1$. The batch size is set as 32. The codes are publicly available\footnote{Codes are available at \url{https://github.com/lpworld/PURS}}.
\subsection{Data}
We implement our model on three real-world datasets: the Yelp Challenge Dataset \footnote{https://www.yelp.com/dataset/challenge}, which contains check-in information of users and restaurants; the MovieLens Dataset \footnote{https://grouplens.org/datasets/movielens/}, which contains informations of user, movies and ratings; and the Youku dataset collected from the major online video platform Youku, which contains rich information of users, videos, clicks and their corresponding features. We list the descriptive statistics of these datasets in Table \ref{statisticalnumber}. For the click-through-rate prediction purposes, we binarize the ratings in Yelp and MovieLens datasets using the threshold 3.5 and transfer them into labels of click and non-click.
\begin{table}[h]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
Dataset & \textbf{Yelp} & \textbf{MovieLens} & \textbf{Youku}\\ \hline
\# of Ratings & 2,254,589 & 19,961,113 & 1,806,157 \\ \hline
\# of Users & 76,564 & 138,493 & 46,143 \\ \hline
\# of Items & 75,231 & 15,079 & 53,657 \\ \hline
Sparsity & 0.039\% & 0.956\% & 0.073\% \\ \hline
\end{tabular}
\caption{Descriptive Statistics of Three Datasets}
\label{statisticalnumber}
\end{table}
\subsection{Baselines and Evaluation Metrics}
To illustrate that the proposed model indeed provide unexpected and useful recommendations at the same time, we select two groups of state-of-the-art baselines for comparison: click-through-rate prediction models and unexpected recommendation models. The first category includes:
\begin{itemize}
\item \textbf{DIN \cite{zhou2018deep}} Deep Interest Network designs a local activation unit to adaptively learn the representation of user interests from historical behaviors with respect to a certain item.
\item \textbf{DeepFM \cite{guo2017deepfm}} DeepFM combines the power of factorization machines for recommendation and deep learning for feature learning in a new neural network architecture.
\item \textbf{Wide \& Deep \cite{cheng2016wide}} Wide \& Deep utilizes the wide model to handle the manually designed cross product features, and the deep model to extract nonlinear relations among features.
\item \textbf{PNN \cite{qu2016product}} Product-based Nerual Network model introduces an additional product layer to serve as the feature extractor.
\end{itemize}
The second baseline category includes:
\begin{itemize}
\item \textbf{SPR \cite{lu2012serendipitous}} Serendipitous Personalized Ranking extends traditional personalized ranking methods by considering item popularity in AUC optimization.
\item \textbf{Auralist \cite{zhang2012auralist}} Auralist is a personalized recommendation system that balances between the desired goals of accuracy, diversity, novelty and serendipity simultaneously.
\item \textbf{DPP \cite{chen2018fast}} The Determinantal Point Process utilizes a fast greedy MAP inference approach to generate relevant and diverse recommendations.
\item \textbf{HOM-LIN \cite{adamopoulos2015unexpectedness}} HOM-LIN is the state-of-the-art unexpected recommendation algorithm, which provides recommendations through the hybrid utility function.
\end{itemize}
In addition, we select the following popular accuracy and novelty metrics for the evaluation process: \textbf{AUC}, which measures the goodness of recommendation order by ranking all the items with predicted CTR and comparing with the click information; \textbf{HR@10}, which measures the number of clicks in top 10 recommendations; \textbf{Unexpectedness}, which measures the recommendations to users of those items that are not included in their consideration sets and depart from what they would expect from the recommender system and is calculated as Equation (3); and \textbf{Coverage}, which measures as the percentage of distinct items in the recommendation over all distinct items in the dataset.
\section{Results}
\subsection{Recommendation Performance}
We implement the proposed PURS model and baseline methods in three real-world datasets. We conduct the time-stratified cross-validation with different initializations and report the average results over these experiments. As shown in Table \ref{unexp} and Figure \ref{comparison}, our proposed model consistently and significantly outperforms all other baselines in terms of both accuracy and unexpectedness measures across three datasets. Compare to the second-best baseline approach, we witness an increase of 2.75\% in AUC, 6.97\% in HR@10, 24.64\% in Unexpectedness and 43.71\% in Coverage measures. Especially in Youku Dataset where rich user behavior sequences in real business setting are available, PURS achieves the most significant improvement over other models. We also observe that all deep-learning based approaches performs significantly better than feature-based methods, which demonstrates the effectiveness of latent models. We conclude that our proposed approach achieves state-of-the-art unexpected recommendation performance and indeed provide satisfying and novel recommendations simultaneously to the target user.
\begin{table*}
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{|c|cccc|cccc|cccc|} \hline
\multirow{2}{*}{Algorithm} & \multicolumn{4}{c|}{Youku} & \multicolumn{4}{c|}{Yelp} & \multicolumn{4}{c|}{MovieLens} \\ \cline{2-13}
& AUC & HR@10 & Unexp & Coverage & AUC & HR@10 & Unexp & Coverage & AUC & HR@10 & Unexp & Coverage \\ \hline
\textbf{PURS} & \textbf{0.7154*} & \textbf{0.7494*} & \textbf{0.0913*} & \textbf{0.6040*} & \textbf{0.6723*} & \textbf{0.6761*} & \textbf{0.2401*} & \textbf{0.7585} & \textbf{0.8090*} & \textbf{0.6778*} & \textbf{0.2719*} & \textbf{0.3732*} \\ \hline
DIN & 0.6957 & 0.6972 & 0.0688 & 0.1298 & 0.6694 & 0.6702 & 0.0391 & 0.6934 & 0.7021 & 0.6485 & 0.0887 & 0.2435 \\
DeepFM & 0.5519 & 0.5164 & 0.0333 & 0.2919 & 0.6396 & 0.6682 & 0.0412 & 0.6044 & 0.7056 & 0.6169 & 0.1275 & 0.3098 \\
Wide\&Deep & 0.6807 & 0.6293 & 0.0472 & 0.3400 & 0.6698 & 0.6693 & 0.0392 & 0.7580 & 0.7940 & 0.6333 & 0.0944 & 0.3432 \\
PNN & 0.5801 & 0.5667 & 0.0593 & 0.1860 & 0.6664 & 0.6692 & 0.0391 & 0.7548 & 0.7140 & 0.6382 & 0.1318 & 0.3665 \\ \hline
HOM-LIN & 0.5812 & 0.5493 & 0.0602 & 0.4284 & 0.6287 & 0.6490 & 0.1433 & 0.5572 & 0.7177 & 0.5894 & 0.1116 & 0.1525 \\
Auralist & 0.5319 & 0.5250 & 0.0598 & 0.3990 & 0.6428 & 0.6104 & 0.1434 & 0.5442 & 0.6988 & 0.5710 & 0.1010 & 0.1333 \\
SPR & 0.5816 & 0.5156 & 0.0739 & 0.4668 & 0.6364 & 0.6492 & 0.1438 & 0.5849 & 0.7059 & 0.6122 & 0.1396 & 0.1728 \\
DPP & 0.6827 & 0.5777 & 0.0710 & 0.4702 & 0.5940 & 0.6414 & 0.1330 & 0.5072 & 0.7062 & 0.5984 & 0.1602 & 0.1908 \\ \hline
\end{tabular}
}
\caption{Comparison of recommendation performance in three datasets. The first block contains baselines for click-through-rate optimization, while the second block contains baselines for unexpectedness optimization. `*' represents statistical significance at the 0.95 level.}
\label{unexp}
\end{table*}
\begin{figure*}[!]
\centering
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{picture/scores_par_youku.png}
\caption{Youku Dataset}
\end{subfigure}%
~
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{picture/scores_par_yelp.png}
\caption{Yelp Dataset}
\end{subfigure}%
~
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{picture/scores_par_movielen.png}
\caption{MovieLen Dataset}
\end{subfigure}%
\caption{Comparison of recommendation performance in terms of accuracy and unexpectedness measures in three datasets.}
\label{comparison}
\end{figure*}
\subsection{Ablation Study}
As discussed in the previous section, PURS achieves significant improvements over other baselines. These improvements indeed come from incorporating the following four components into the design of recommendation model: \textbf{Unexpectedness}, which aims at providing novel and satisfying recommendations; \textbf{Unexpected Activation Function}, which adjusts the input of unexpectedness into the utility function; \textbf{Personalized and Session-Based Factor}, which captures the user and session-level heterogeneity of perception towards unexpectedness; and \textbf{Clustering of Behavior Sequence}, which extracts the diverse user interests and constructs user expectations.
In this section, we conduct the ablation study to justify the importance of each factor. Specifically, we compare the proposed model with the following variations:
\begin{itemize}
\item \textbf{PURS-Variation 1 (Gaussian Activation)} This model users Gaussian distribution to serve as the unexpected activation function in the original design, and provides recommendations based on the following utility function $Utility_{u,i} = b_{u} + b_{i} + r_{u,i} + exp(-unexp_{u,i}^2)*unexp\_factor_{u,i}$
\item \textbf{PURS-Variation 2 (No Activation)} This model removes the unexpected activation function in the original design, and provides recommendations based on the following utility function $Utility_{u,i} = b_{u} + b_{i} + r_{u,i} + unexp_{u,i}*unexp\_factor_{u,i}$
\item \textbf{PURS-Variation 3 (No Unexpectedness Factor)} This model removes the unexpected activation function in the original design, and provides recommendations based on the following utility function $Utility_{u,i} = b_{u} + b_{i} + r_{u,i} + f(unexp_{u,i})$
\item \textbf{PURS-Variation 4 (No Unexpectedness)} This model removes the unexpected activation function in the original design, and provides recommendations based on the following utility function $Utility_{u,i} = b_{u} + b_{i} + r_{u,i}$
\item \textbf{PURS-Variation 5 (Single Closure of User Interest)} This model provides the unexpected recommendations using the same utility function, but instead remove the clustering procedure and model unexpectedness following \cite{li2019latent3} as $unexp_{i_{*},u} = d(w_{*}, C_{u})$ where $C_{u}$ is the entire latent closure generated by all past transactions of user $u$.
\end{itemize}
As shown in Table \ref{ablation}, if we remove any of these four components out of the recommendation model, we will witness significant loss in both accuracy and unexpectedness measures, especially in coverage metric. Therefore, the ablation study demonstrates that the superiority of our proposed model really comes from the combination of four novel components that all play significant role in contributing to satisfying and unexpected recommendations.
\begin{table*}
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{|c|cccc|cccc|cccc|} \hline
\multirow{2}{*}{Algorithm} & \multicolumn{4}{c|}{Youku} & \multicolumn{4}{c|}{Yelp} & \multicolumn{4}{c|}{MovieLens} \\ \cline{2-13}
& AUC & HR@10 & Unexp & Coverage & AUC & HR@10 & Unexp & Coverage & AUC & HR@10 & Unexp & Coverage \\ \hline
\textbf{PURS} & \textbf{0.7154*} & \textbf{0.7494*} & \textbf{0.0913*} & \textbf{0.6040*} & \textbf{0.6723*} & \textbf{0.6761*} & \textbf{0.2401*} & \textbf{0.7585*} & \textbf{0.8090*} & \textbf{0.6778*} & \textbf{0.2719*} & \textbf{0.3732*} \\ \hline
PURS-Variation 1 & 0.6826 & 0.7292 & 0.0828 & 0.5548 & 0.6682 & 0.6602 & 0.1298 & 0.7292 & 0.7757 & 0.6419 & 0.1812 & 0.3350 \\
PURS-Variation 2 & 0.7067 & 0.7148 & 0.0707 & 0.3026 & 0.6692 & 0.6630 & 0.0412 & 0.7580 & 0.8041 & 0.6585 & 0.2471 & 0.3580 \\
PURS-Variation 3 & 0.7036 & 0.6720 & 0.0762 & 0.1522 & 0.6508 & 0.6692 & 0.0391 & 0.7583 & 0.7666 & 0.6460 & 0.0888 & 0.2792 \\
PURS-Variation 4 & 0.7063 & 0.5628 & 0.0688 & 0.1298 & 0.6702 & 0.6702 & 0.0391 & 0.6934 & 0.7586 & 0.6331 & 0.0887 & 0.2435 \\
PURS-Variation 5 & 0.7038 & 0.7080 & 0.0477 & 0.2042 & 0.6701 & 0.6700 & 0.0395 & 0.7580 & 0.7715 & 0.6596 & 0.1727 & 0.3561 \\ \hline
\end{tabular}
}
\caption{Ablation Study of recommendation performance in three datasets. `*' represents statistical significance at the 0.95 level.}
\label{ablation}
\end{table*}
\subsection{Improving Accuracy and Novelty Simultaneously}
In Table \ref{unexp}, we observe that unexpectedness-oriented baselines generally achieve better performance in unexpected measures, but at the cost of losing accuracy measures, when comparing to the CTR-oriented baselines. This observation is in line with the prior literature \cite{adomavicius2008overcoming,zhou2010solving} discussing the trade-off between these two objectives. However, our proposed PURS model manages to surpass baselines models in both \textbf{accuracy} and \textbf{novelty} measures at the same time. In addition, PURS is capable of improving AUC and unexpectedness metrics simultaneously throughout the training process, as shown in Figure \ref{tradeoff}.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{picture/tradeoff.png}
\caption{Improving accuracy and unexpectedness measures simultaneously in each training epoch of PURS in the MovieLens dataset}
\label{tradeoff}
\end{figure}
\subsection{Scalability}
To test for scalability, we provide recommendations using PURS with aforementioned parameter values for the Yelp, MovieLens and Youku datasets with increasing data sizes from 100 to 1,000,000 records. As shown in Figure \ref{scalability}, we empirically observe that the proposed PURS model scales linearly with increase in number of data records to finish the training process and provide unexpected recommendations accordingly. The training procedure comprises of obtaining user and item embeddings and jointly optimizing the utility function. The optimization phase is made efficient using batch normalization \cite{ioffe2015batch} and distributed training. As our experiments confirm, PURS is capable of learning network parameters efficiently and indeed scales well.
\begin{table*}
\centering
\begin{tabular}{|c|cccccc|} \hline
& VV & TS & ID & CTR & Unexpectedness & Coverage \\ \hline
Improvement & +3.74\%* & +4.63\%* & +4.13\%* & +0.80\%* & +9.74\%* & +1.23\%* \\ \hline
\end{tabular}
\caption{Unexpected recommendation performance in online A/B test: performance increase compared to the current model. `*' represents statistical significance at the 0.95 level.}
\label{online}
\end{table*}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{picture/scalability.jpg}
\caption{Scalability of PURS on the Yelp, MovieLens and Youku datasets with increasing data sizes from 100 to 1,000,000 records.}
\label{scalability}
\end{figure}
\section{Online A/B Test}
We conduct the online A/B test at Alibaba-Youku, a major video recommendation platform from 2019-11 to 2019-12. During the testing period, we compare the proposed PURS model with the latest production model in the company. We measure the performance using standard business metrics: \textbf{VV} (Video View, average video viewed by each user), \textbf{TS} (Time Spent, average time that each user spends on the platform), \textbf{ID} (Impression Depth, average impression through one session) and \textbf{CTR} (Click-Through-Rate, the percentage of user clicking on the recommended video). We also measure the novelty of the recommended videos using the unexpectedness and coverage measures described in Section 5.2. We present the results in Table \ref{online} that demonstrates significant and consistent improvements over the current model in all the four business metrics, and these improvements indeed come from providing more unexpected recommendations, as demonstrated in Section 6. The proposed model is in the process of being deployed by the company.
\section{Conclusions}
In this paper, we present the PURS model that incorporates unexpectedness into the recommendation process. Specifically, we propose to model user interests as clusters of embedding closures in the latent space and calculate unexpectedness as the weighted distance between the new items and the interest clusters. We utilize the sequence modeling and the self-attention mechanism to capture personalized and session-based user perception of unexpectedness. We also propose a novel unexpected activation function to achieve better unexpected recommendation performance. We subsequently combine the CTR estimation with the degree of unexpectedness to provide final recommendations. Extensive offline and online experiments illustrate superiority of the proposed model.
As the future work, we plan to study the impact of unexpected recommendations on user behaviors. Also, we plan to model the user interests in a more explicit manner to provide better unexpected recommendations.
\balance
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction} \label{sec:intro}
The slice filtration, a filtration of genuine \(G\)-spectra, was developed by Hill, Hopkins, and Ravenel in their solution to the Kervaire invariant-one problem \cite{HHR} and is a generalization of Dugger's filtration \cite{D}. It was modeled after Voevodsky's motivic slice filtration \cite{V} and is an equivariant generalization of the Postnikov tower. Rather than decomposing a \(G\)-spectrum into Eilenberg-Mac Lane specra, as does the Postnikov tower, instead the slice filtration decomposes a \(G\)-spectrum into ``\(n\)-slices''.
There is a complete characterization of all \(n\)-slices where \(-1\leq n\leq 1\), listed in \cref{charac:slice:01-1}. This, combined with \cref{rho:susp:commutes}, characterizes all slices in degrees congruent to \(-1\), 0, or 1, modulo the order of \(G\). For \(G=C_2\times C_2\), we are then only missing the \((4k+2)\)-slices. In \cref{sec:2slice} we finish this characterization with the following result.
\textit{\cref{2slice:characterization}: Suppose the only nontrivial homotopy Mackey functors of a \((C_2\times C_2)\)-spectrum \(X\) are \(\ul\pi_1(X)\) and \(\ul\pi_2(X)\) where certain maps in each Mackey functor are injective. Then \(X\) is a 2-slice. Conversely, if \(X\) is a 2-slice, then its only nontrivial homotopy Mackey functors are \(\ul\pi_1(X)\) and \(\ul\pi_2(X)\) where \(\Si{2} H\ul\pi_2(X)\) is a 2-slice and \(\Si{1} H\ul\pi_1(X) \in [2,4]\).}
Much work has been done computing the slices of certain \(RO(G)\)-graded suspensions of \(H_G\underline{\Z}\) including \(G=C_{p^n}\) by \cite{HHR2} and \cite{Y}, and \mbox{\(G=D_{2p}\)} by \cite{Z2}. \cite{GY} computes the slices of \(\Si{n} H_K \underline{\F}\) where \(K=C_2\times C_2\). Most of these slices are \(RO(K)\)-graded suspensions of \(H\ul\pi_{-i}(\Si{-k\rho_K} H_K\underline{\F})\) for \(i\) in the range \([k+3,4k]\). We primarily focus on the slices of \(\Si{\pm n} H_K\underline{\Z}\). Although we have cofiber sequences relating \(\Si{n} H_K\underline{\Z}\) to \(\Si{n} H_K\underline{\F}\), we can only recover some information about the former from the latter.
As for the slices of \(\Si{\pm n} H_K\underline{\Z}\), the main result can be summarized as follows:
\textit{Main Result: For \(n<0\), all nontrivial slices of \(\Si{n} H_K\underline{\Z}\) are given by:
\begin{align*}
P^{i}_{i}(\Si{n} H_K\underline{\Z}) \simeq \Si{-V} H_K\ul M
\end{align*}
where \(i\) is in the range \([4n,n]\).
For \(0\leq n\leq 5\), \(\Si{n} H\underline{\Z}\) is an \(n\)-slice. Finally, for \(n>5\),
\begin{align*}
P^i_i(\Si{n} H_K\underline{\Z}) \simeq \Si{V} H_K\ul M
\end{align*}
where \(i\) is in the range \([n,4(n-4)]\).
The representations \(V\) and Mackey functors \(\ul M\) are given in \cref{-nslice:K:Z}, \cref{-4kslices:K:Z}, \cref{-4k-2slices:K:Z}, \cref{slices:K:n:Z}, \cref{slices:K:4k:Z}, and \cref{slices:4k+2:K:Z}.}
The paper is organized as follows. In \cref{sec:background}, we review the slice filtration and relevant dualities. The story for \(K\) must restrict to the corresponding results for \(C_2\), so we review these results in \cref{sec:C2}. \cref{sec:K:Mackey} provides us with the main Mackey functors for \(K\) and some pertinent results for \cref{sec:2slice}, in which we characterize all 2-slices over \(K\). We provide some slice towers in \cref{sec:cotowers:K:Z} and describe the slices of \(\Si{-n} H_K\underline{\Z}\) in
\cref{sec:-nslice:K:Z}. In \cref{sec:slices:n:K:Z}, we use Brown-Comenetz duality and the slices of \(\Si{-n} H\underline{\Z}\) to obtain the slices of \(\Si{n} H_K\underline{\Z}\). We then compute the homotopy Mackey functors of the slices of \(\Si{\pm n} H_K\underline{\Z}\) in \cref{sec:htpycomp}. Finally, we provide some examples of the slice spectral sequence for \(\Si{\pm n} H_K\underline{\Z}\) in \cref{sec:SSS}.
The author is grateful for the guidance of Bert Guillou and some helpful conversations with Vigleik Angelveit. Figures \ref{fig:SSS:-1..-5:HZ}, \ref{fig:SSS:-7..-9:HZ}, \ref{fig:SSS:6..8:HZ}, \ref{fig:SSS:10..11:HZ}, \ref{fig:SSS:12:HZ}, and \ref{fig:SSS:14:HZ} were created using Hood Chatham's \texttt{spectralsequences} package.
\section{Background} \label{sec:background}
In this section we give background for the slice filtration as well as Brown-Comenetz and Anderson duality, and \(K\)-representations. Here, except for \cref{subsec:signrep}, \(G\) is any finite group.
\subsection{The Slice Filtration} \label{slice_filt_intro}
\phantom{a}
We start with a brief review of the equivariant slice filtration. For more details see \cite[Section 4]{HHR}.
\begin{defn} \label{def:tau_n}
Let \(\text{Sp}^G\) be the category of genuine \(G\)-spectra. Let \(\tau^G_{\geq n} \subseteq \text{Sp}^G\) be the localizing subcategory generated by \(G\)-spectra of the form \(\Si{\infty}_G G_+ \wedge_H S^{k\rho_H}\) where \(H\leq G\), \(\rho_H\) is the regular representation of \(H\), and \(k\abs{H} \geq n\). We write \(X\geq n\) to mean that \(X\in \tau^G_{\geq n}\).
\end{defn}
\begin{defn} \label{def:lessthan}
We say that \(X<n\) if
\begin{align*}
[S^{k\rho_H+r}, X]^H = 0
\end{align*}
for all \(r\geq 0\) and all subgroups \(H\leq G\) such that \(k\abs{H}\geq n\).
\end{defn}
\begin{thm}{\cite[Corollary 2.9, Theorem 2.10]{HY}} \label{thm:atleastn}
Let \(n\geq 0\). Then \(X\geq n\) if and only if
\begin{align*}
\pi_k(X^H) = 0 \quad \text{for } \quad k<\frac{n}{\abs{H}}.
\end{align*}
\end{thm}
\begin{prop} \label{charac:slice:01-1}
\phantom{a}
\begin{enumerate}
\item \cite[Proposition 4.50]{HHR} \(X\) is a 0-slice if and only if \(X\simeq H\ul M\) for some \(\ul M\in \text{Mack}(G)\).
\item \cite[Proposition 4.50]{HHR} \(X\) is a 1-slice if and only if \(X\simeq \Si{1} H\ul M\) for some \(\ul M\in \text{Mack}(G)\) with injective restrictions.
\item \cite[Theorem 6-4]{U} \(\Si{-1} H\ul M\) is a \((-n)\)-slice iff \(\ul M\) has surjective transfers for \(\abs{H}\geq n\) and \(M(G/H) = 0\) for \(\abs{H}<n\).
\end{enumerate}
\end{prop}
It is important to note that \cite{HHR} uses the original slice filtration whereas we employ the regular slice filtration from \cite{U}. Except for an indexing difference of one, the results are the same.
\begin{prop}{\cite[Corollary 4.25]{HHR}} \label{rho:susp:commutes}
For any \(k\in\mathbb{Z}\),
\begin{align*}
P^{k+\abs{G}}_{k+\abs{G}} (\Si{\rho} X) &\simeq \Si{\rho} P^k_k(X).
\end{align*}
\end{prop}
That is, suspension by the regular representation commutes with the slice filtration.
Given some surjection of groups \(\phi_N:G\rightarrow G/N\) where \(N\unlhd G\), there is a geometric pullback functor \(\phi^*_N : \text{Sp}^{G/N} \rightarrow \text{Sp}^G\) \cite[Definition 4.1]{H}.
\begin{prop}{\cite[Conjecture 4.11]{H}, \cite[Corollary 4-5]{U}} \label{pullback:slice}
Let \(N\unlhd G\). If the \((G/N)\)-spectrum \(X\) is a \(k\)-slice over \(G/N\), then \(\phi^*_N X\) is a \(k[G:N]\)-slice over \(G\).
\end{prop}
\subsection{Brown-Comenetz and Anderson Duality} \label{BCandA:duality}
\phantom{A}
As in \cite{HS}, we write \(I_{\mathbb{Q}/\mathbb{Z}}\) to indicate the representing \(G\)-spectrum of the cohomology theory \(X\mapsto \text{Hom}(\pi_{-\ast}^G X, \mathbb{Q}/\mathbb{Z})\). The Brown-Comenetz dual of \(X\) is then defined to be the function \(G\)-spectrum \(F(X, I_{\mathbb{Q}/\mathbb{Z}})\). Similarly, \(I_{\mathbb{Q}}\) represents \(X\mapsto \text{Hom}(\pi_{-\ast} X, \mathbb{Q})\) and \(I_\mathbb{Q} X = F(X, I_{\mathbb{Q}})\). Finally, the Anderson dual of \(X\) is \(I_{\mathbb{Z}} X = F(X,I_\mathbb{Z})\), where \(I_\mathbb{Z}\) is the fiber of the natural map \(I_{\mathbb{Q}} \rightarrow I_{\mathbb{Q}/\mathbb{Z}}\).
\begin{exmp}{\cite[Section 3A]{GM}, \cite{HS}}
For a torsion Eilenberg-MacLane spectrum \(H\ul M\),
\begin{align*}
I_\mathbb{Z} H\ul M &\simeq \Si{-1} H\ul M^*
\end{align*}
and
\begin{align*}
I_{\mathbb{Q}/\mathbb{Z}} H\ul M &\simeq H\ul M^*.
\end{align*}
One should note that \cite{HS} deals with non-equivariant spectra and \cite[Section 3A]{GM} refers specifically to \(H\ul M\) as an \(\mathbb{F}_2\)-torsion spectrum. This example, however, follows easily from their work and \cite[Corollary I.7.3]{U2}. See \cite[Section 3A, Section 3B]{GM} for a more detailed discussion of equivariant Anderson duality.
\end{exmp}
\begin{prop}(\cite[Theorem I.7.7, Theorem I.7.8]{U2}) \label{duality:prop}
For a spectrum \(X\),
\begin{align*}
X\geq n &\Leftrightarrow I_{\mathbb{Q}/\mathbb{Z}} X\leq -n.
\end{align*}
In particular,
\begin{align*}
P^n_k I_{\mathbb{Q}/\mathbb{Z}} X &\cong I_{\mathbb{Q}/\mathbb{Z}} P^{-k}_{-n} X.
\end{align*}
\end{prop}
That is, the Brown-Comenetz dualization functor dualizes slice status.
\subsection{\texorpdfstring{\(K\)}{K4}-Representations} \label{subsec:signrep}
\phantom{a}
Recall that \(\sigma\) is the one-dimensional sign representation of \(C_2\). The sign representations of \(K\) are then the pullbacks \(p_1^*\sigma\), \(m^*\sigma\), and \(p_2^*\sigma\) where \(p_1\), \(m\), and \(p_2\) are
\begin{equation*}
\begin{tikzcd}[ampersand replacement=\&]
\& C_2 \arrow[dr, "\sigma"] \\
C_2\times C_2 \arrow[ur, "p_1"] \arrow[r, "m"] \arrow[dr, "p_2"] \& C_2 \arrow[r, "\sigma"] \& \mathbb{R} \\
\& C_2 \arrow[ur, "\sigma"] \&
\end{tikzcd}
\end{equation*}
As in \cite{GY}, we will write
\begin{align*}
\alpha=p_1^*\sigma, \qquad \beta=p_2^*\sigma, \qquad \text{and } \qquad \gamma=m^*\sigma.
\end{align*}
The regular representation of \(K\) is then \(\rho_{K} = 1 + \alpha + \beta + \gamma\).
\section{Review of \texorpdfstring{\(C_2\)}{C2}} \label{sec:C2}
The Lewis diagram for a \(C_2\)-Mackey functor takes the form
\begin{equation*}
\begin{tikzcd}
\ul M(C_2) \arrow[dd, color = black, bend right = 20] \\ \\
\ul M(e) \arrow[uu, color = black, bend right = 20]
\end{tikzcd}
\end{equation*}
where \(\ul M(e)\) has a \(C_2/e\) action. Arrows going down the diagram are called restrictions and arrows going up are called transfers.
For the sake of clarity in large diagrams, in a general Mackey functor \(\ul M\) we will henceforward denote \(\ul M(H)\) by \(M_H\).
\begin{prop} \label{C2:2slice}
A \(C_2\)-spectrum \(Y\) is a 2-slice over \(C_2\) if and only if the only nontrivial homotopy Mackey functors are \(\ul\pi_1(Y)\) and \(\ul\pi_2(Y)\), where
\begin{enumerate}
\item The restriction map \(\res{C_2}{e}: \pi_2^{C_2}(Y) \rightarrow \pi_2^e(Y)\) is injective \label{2slice:C2:M}
\item \(\pi_1^e(Y) = 0\). \label{2slice:C2:N}
\end{enumerate}
That is, both \(\Si{2} H\ul\pi_2(Y)\) and \(\Si{1} \ul\pi_1(Y)\) are 2-slices.
\end{prop}
\begin{proof}
Let \(\ul M=\ul\pi_2(Y)\) and \(\ul N=\ul\pi_1(Y)\) and suppose that they are the only nontrivial homotopy Mackey functors of \(Y\). We show that \(\Si{2} H\ul M\) and \(\Si{1} H\ul N\) are 2-slices if and only if \cref{2slice:C2:M} and \cref{2slice:C2:N} hold.
First, consider \(\Si{2} H\ul M\). We immediately see that this spectrum is at least two. As for \(\Si{2} H\ul M\leq 2\), the cofiber sequence
\begin{align*}
(S^{-\sigma} \rightarrow S^0 \rightarrow C_2/e_+) \wedge \Si{2} H\ul M
\end{align*}
provides the homotopy
\begin{center}
\begin{tikzpicture}
\node at (-4.25,1) {\(\ul\pi_2 (\Si{-\sigma} H\ul M) \quad =\)};
\node (lMC) at (-1.5,2) {\(\ker(\res{C_2}{e})\)};
\node (lMe) at (-1.5,0) {\(0\)};
%
\node at (0,1) {and};
%
\node at (2.5,1) {$\ul\pi_1 (\Si{-\sigma} H\ul M) \quad =$};
\node (MC) at (5,2) {$M_e / \text{im } \res{C_2}{e}$};
\node (Me) at (5,0) {$\mathbb{Z}_\sigma \otimes M_e$};
%
\arres{MC}{Me}{0}{0}{}
\artr{Me}{MC}{0}{0}{}
\end{tikzpicture}
\end{center}
Then the cofiber sequences
\begin{align*}
(S^{-k\sigma} \rightarrow S^{-(k-1)\sigma} \rightarrow C_2/e_{+}\wedge S^{-(k-1)\sigma})\wedge \Si{2} H\ul M
\end{align*}
reveal that \(\ul\pi_2^{C_2}(\Si{-k\sigma} \Si{2} H\ul M) = \ker(\res{C_2}{e})\) for all \(k\geq 2\).
By definition, \(\Si{2} H\ul M\leq 2\) if and only if
\(\ul\pi_2^{C_2}(\Si{-k\sigma} \Si{2} H\ul M) = 0\) for all \(k\geq 2\). Consequently, \(\Si{2} H\ul M\) is a 2-slice if and only if \(\res{C_2}{e}\) is injective .
As for \(\Si{1} H\ul N\), by \cref{thm:atleastn}, we find that \(Y\geq 2\) if and only if \(N_{e} = 0\). But then \(\Si{1} H\ul N\) is the pullback \(\phi^*_K \Si{1} HN_{e}\) and consequently, a 2-slice by \cref{pullback:slice}.
Conversely, assume \(Y\) is a 2-slice. By \cref{thm:atleastn}, we know \(\ul\pi_1^e(Y) = 0\) and \(\ul \pi_k(Y) = \ul 0\) for \(k\leq 0\). For \(k\geq 3\), the slice status of \(Y\) and \cref{def:lessthan} dictate that \(\ul \pi_k(Y) = \ul 0\). Consequently, the only nontrivial homotopy Mackey functors of \(Y\) and \(\ul\pi_1(Y)=\ul N\) and \(\ul\pi_2(Y)=\ul M\). Thus, we have a fiber sequence \(\Si{2} H\ul M\rightarrow Y\rightarrow \Si{1} H\ul N\).
If either \(\Si{2} H\ul M\) or \(\Si{1} H\ul N\) is trivial, the result follows from above. So assume that both are nontrivial. Because \(N_e = 0\), \(\Si{1} H\ul N\) is a 2-slice. In particular, \(\Si{1} H\ul N\leq 2\), and since \(Y\leq 2\), we have that \(\Si{2} H\ul M\leq 2\). Consequently, as \(\Si{2} H\ul M\geq 2\), we have that \(\Si{2} H\ul M\) is a 2-slice.
\end{proof}
From \cite{GY}, we will see the \(C_2\)-Mackey functors in \cref{tab-oldC2Mackey}.
\begin{table}[h]
\caption[Familiar {$C_2$}-Mackey functors.]{Familiar $C_2$-Mackey functors.}
\label{tab-oldC2Mackey}
\begin{center}
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c|c|}
\hline
$\underline{\Z}$
& $\underline{\Z}^*$
& ${\underline{\widehat{f}}}$
\\ \hline
$\MackC{\mathbb{Z}}{\mathbb{Z}}{blue}{orange}{1}{2}$
& $\MackC{\mathbb{Z}}{\mathbb{Z}}{orange}{blue}{2}{1}$
& $\MackC{0}{\mathbb{Z}_\sigma}{white}{white}{}{}$
\\
\hline
$\underline{\F}$
& $\underline{\F}^*$
& $\ul f$
\\ \hline
$\MackC{\mathbb{F}_2}{\mathbb{F}_2}{black}{white}{1}{}$
& $\MackC{\mathbb{F}_2}{\mathbb{F}_2}{white}{black}{}{1}$
& $\MackC{0}{\mathbb{F}_2}{white}{white}{}{}$
\\
\hline
$\underline{g}$ && \\
$\MackC{\mathbb{F}_2}{0}{white}{white}{}{}$ && \\ \hline
\end{tabular} }
\end{center}
\end{table}
\begin{prop}[\cite{GY}] \label{C2:equiv}
There are equivalences
\begin{enumerate}
\item \(\Si{4} H_{C_2}\underline{\Z} \simeq \Si{2\rho} H_{C_2}\underline{\Z}^*\)
\item \(\Si{2} H_{C_2}\ul f \simeq \Si{\rho} H_{C_2}\underline{\F}^*\)
\item \(\Si{1} H_{C_2}\underline{g} \simeq \Si{\rho} H_{C_2}\underline{g}\) \label{C2:g:2nslice}
\end{enumerate}
\end{prop}
Note, in particular, that \cref{C2:g:2nslice} makes \(\Si{n} H_{C_2}\underline{g}\) a \(2n\)-slice for any \(n\in \mathbb{Z}\).
\begin{prop} \label{ksigma+r:homotopy:C2:Z}
For \(k,r\geq 0\),
\begin{align*}
\ul \pi_i (\Si{k\sigma + r} H_{C_2}\underline{\Z}) &= \left\lbrace \begin{array}{ll}
\underline{\Z} & i=k+r, \text{ } k \text{ even} \\
{\underline{\widehat{f}}} & i=k+r, \text{ } k \text{ odd} \\
\underline{g} & i\in[r,k+r-1], \text{ } i\equiv r\pmod 2
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
We calculate \(\ul \pi_i (\Si{k\sigma} H_{C_2}\underline{\Z})\) and then shift the degrees by \(r\). The result follows by induction on \(j\geq 1\) using the cofiber sequence
\begin{align*}
\Si{(j-1)\sigma+2} H\underline{\Z} \simeq \Si{j\sigma} H\underline{\Z}^* \rightarrow \Si{j\sigma} H\underline{\Z} \rightarrow \Si{j\sigma} H\underline{g}\simeq H\underline{g}.
\end{align*}
\end{proof}
\subsection{Slices of \texorpdfstring{\(\Si{\pm n} H_{C_2}\underline{\Z}\)}{HZ}} \label{towers:C2:Z:V} \hspace{2in}
Because the slices for \(\Si{\pm n} H_{K}\underline{\Z}\) over \(K\) restrict to the corresponding slices over \(C_2\), we must know these slices over \(C_2\). They are as follows.
\begin{prop} \label{C2:Z:slice:0-6}
\(\Si{n} H_{C_2}\underline{\Z}\) is an \(n\)-slice for \(0\leq n\leq 6\).
\end{prop}
\begin{proof}
For \(0\leq n\leq 2\), this follows from \cref{charac:slice:01-1} and \cref{C2:2slice}. By the same results, \(\Si{n} H\underline{\Z}^*\) is an \(n\)-slice for \(0\leq n\leq 2\). Furthermore, \(\Si{-1} H\underline{\Z}^*\) is a \((-1)\)-slice by \cref{charac:slice:01-1}.
Then, by \cref{C2:2slice} and \cref{rho:susp:commutes},
\(\Si{n} H\underline{\Z} \simeq \Si{n-4+2\rho} H\underline{\Z}^*\) is a \(n\)-slice for \(3\leq n\leq 6\).
\end{proof}
\begin{prop} \label{tower:C2:Z:n}
Let \(n\geq 7\) and take \(r\equiv n\pmod 4\) with \(3\leq r\leq 6\). The slice tower of \(\Si{n} H_{C_2}\underline{\Z}\) is
\begin{equation*}
\begin{tikzcd}
\arrow[r] P^{2n-6}_{2n-6} = \Si{n-3} H\underline{g} &
\arrow[d] \Si{n} H\underline{\Z} \\
\arrow[r] P^{2n-10}_{2n-10} = \Si{n-5} H\underline{g} &
\arrow[d] \Si{2\rho + (n-4)} H\underline{\Z} \\
& \arrow[d] \vdots \\
\arrow[r] P^{n+r-2}_{n+r-2} = \Si{\frac{n+r}{2}-1} H\underline{g} & \arrow[d] \Si{\frac{n-r-4}{2}\rho+(r+4)} H\underline{\Z} \\
& P^n_n = \Si{\frac{n-r}{2}\rho + r} H\underline{\Z}
\end{tikzcd}
\end{equation*}
\end{prop}
\begin{proof}
The exact sequence \(\underline{\Z}^* \rightarrow \underline{\Z} \rightarrow \underline{g}\) and the homotopy equivalence \(\Si{n} H\underline{\Z} \simeq \Si{n-4+2\rho} H\underline{\Z}^*\) provide the fiber sequence
\begin{align*}
\Si{n-3} H\underline{g} \rightarrow \Si{n} H\underline{\Z} \rightarrow \Si{2\rho + (n-4)} H\underline{\Z}.
\end{align*}
We then augment this sequence with its appropriate \(2\rho\) suspensions until \(\Si{\frac{n-r}{2}\rho + r} H\underline{\Z}\), a slice, is reached.
\end{proof}
\begin{cor} \label{2kslice:C2:HZ}
Let \(n\geq 7\) and take \(r\equiv n\pmod 4\) with \(3\leq r\leq 6\). The \((2k)\)-slices of \(\Si{n} H_{C_2}\underline{\Z}\) are
\begin{align*}
P^{2k}_{2k}( \Si{n} H_{C_2}\underline{\Z} ) &\simeq \Si{k} H_{C_2}\underline{g}
\end{align*}
for \(k\equiv n+1\pmod 2\) and \(k\in \left[\frac{n+r}{2}-1,\ldots, n-3\right]\).
\end{cor}
\begin{prop} \label{tower:C2:Z:-n}
Let \(n\geq 1\) and take \(r\equiv n\pmod 4\) with \(1\leq r\leq 4\). The slice cotower of \(\Si{-n} H_{C_2}\underline{\Z}\) is
\begin{equation*}
\begin{tikzcd}
\arrow[d] \Si{-\frac{n-r}{2}\rho-r} H\underline{\Z}^* = P^{-n}_{-n} & \\
\arrow[d] \arrow[r] \Si{-\frac{n-r}{2}\rho -r} H\underline{\Z} & \Si{-\frac{n+r}{2}} H\underline{g} = P^{-n-r}_{-n-r} \\
\arrow[d] \vdots & \\
\arrow[d] \arrow[r] \Si{-2\rho -n+4} H\underline{\Z} & \Si{-n+2} H\underline{g} = P^{-2n+4}_{-2n+4} \\
\arrow[r] \Si{-n} H\underline{\Z} & \Si{-n} H\underline{g} = P^{-2n}_{-2n}
\end{tikzcd}
\end{equation*}
\end{prop}
\begin{proof}
The exact sequence \(\underline{\Z}^* \rightarrow \underline{\Z} \rightarrow \underline{g}\) and the homotopy equivalence \(\Si{-n} H\underline{\Z} \simeq \Si{2\rho-n-4} H\underline{\Z}\) provide the cofiber sequence
\begin{align*}
\Si{-2\rho-n+4} H\underline{\Z} \rightarrow \Si{-n} H\underline{\Z} \rightarrow \Si{-n} H\underline{g}.
\end{align*}
We then augment this sequence with its appropriate \(-2\rho\) suspensions until \(\Si{-\frac{n-r}{2}\rho-r} H\underline{\Z}\), a slice, is reached.
\end{proof}
\begin{cor} \label{-2kslice:C2:HZ}
Let \(n\geq 1\) and take \(r\equiv n\pmod 4\) with \(1\leq r\leq 4\). The \((-2k)\)-slices of \(\Si{-n} H_{C_2}\underline{\Z}\) are
\begin{align*}
P^{-2k}_{-2k}( \Si{-n} H_{C_2}\underline{\Z} ) &\simeq \Si{-k} H_{C_2}\underline{g}
\end{align*}
for \(k\equiv n\pmod 2\) and \(k\in \left[\frac{n+r}{2},n\right]\).
\end{cor}
\section{\texorpdfstring{\(K\)}{K}-Mackey Functors \label{sec:K:Mackey} \hspace{2in}}
Recall that \(K=C_2\times C_2\) and let \(L\), \(D\), and \(R\) be the left, diagonal, and right subgroups of \(K\), respectively. The Lewis diagram for a \(K\)-Mackey functor takes the form
\begin{center}
\begin{tikzpicture}
\node (MK) at (0,4) {$M_K$};
\node (ML) at (-2,2) {$M_L$};
\node (MD) at (0,2) {$M_D$};
\node (MR) at (2,2) {$M_R$};
\node (Me) at (0,0) {$M_e$};
\arres{MK}{ML}{0}{0}{}
\arres{MK}{MD}{0}{0}{}
\arres{MK}{MR}{0}{0}{}
\artr{ML}{MK}{0}{0}{}
\artr{MD}{MK}{0}{0}{}
\artr{MR}{MK}{0}{0}{}
\arres{ML}{Me}{0}{0}{}
\arres{MD}{Me}{0}{0}{}
\arres{MR}{Me}{0}{0}{}
\artr{Me}{ML}{0}{0}{}
\artr{Me}{MD}{0}{0}{}
\artr{Me}{MR}{0}{0}{}
\end{tikzpicture}
\end{center}
with a Weyl group action \(W_G(H) \circlearrowright M_H\) at each level. If we do not indicate the Weyl group actions for a specific Mackey functor, then they are trivial.
A number of Mackey functors from \cite{GY} will make their appearance.
\begin{table}[h]
\caption[Familiar {$K$}-Mackey functors.]{Familiar $K$-Mackey functors.}
\label{tab-oldKMackey}
\begin{center}
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c|}
\hline
$\ul {mg}$
& $\ul {m}$
\\ \hline
$\MackKupperR{\mathbb{F}_2^2}{\mathbb{F}_2}{black}{p_1}{\nabla}{p_2}$
& $\MackKupperR{\mathbb{F}_2}{\mathbb{F}_2}{black}{1}{1}{1}$
\\
\hline
$\phi_{LDR}^*\underline{\F}$
& $\underline{g}^n$
\\ \hline
$\MackKupperR{\mathbb{F}_2^3}{\mathbb{F}_2}{black}{p_1}{p_2}{p_3}$
& $\MackKgn{\mathbb{F}_2^n}$
\\
\hline
\end{tabular} }
\end{center}
\end{table}
We will also see the duals of the Mackey functors in \cref{tab-oldKMackey}.
\begin{prop}(\cite[Proposition 4.8]{GY}) \label{m:mg:equiv}
There are equivalences
\begin{align}
\Si{-\rho} H_K\ul m &\simeq \Si{-2} H_K\ul{mg}^* \\
\Si{-2} H_K\ul m^* &\simeq \Si{-\rho} H_K\ul{mg}
\end{align}
\end{prop}
We will also see the new Mackey functors in \cref{tab-newKMackey}.
\begin{table}[h]
\caption[{New $K$}-Mackey functors.]{New $K$-Mackey functors.}
\label{tab-newKMackey}
\begin{center}
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c|}
\hline
$\underline{\Z}$
& $\underline{\Z}^*$
\\ \hline
$\MackKall{\mathbb{Z}}{blue}{blue}{orange}{orange}$
& $\MackKall{\mathbb{Z}}{orange}{orange}{blue}{blue}$
\\ \hline
$\underline{\Z}(2,1)$
& $\underline{\Z}(2,1)^*$
\\ \hline
$\MackKall{\mathbb{Z}}{orange}{blue}{blue}{orange}$
& $\MackKall{\mathbb{Z}}{blue}{orange}{orange}{blue}$
\\ \hline
$\underline{\mathbb{M}}$ & \\ \hline
${\begin{tikzcd}[ampersand replacement=\&, bend right=2.5ex, column sep=scriptsize]
\& \mathbb{Z}/4 \ar[dl, color=blue] \ar[d, color=blue] \ar[dr,color=blue] \& \\
\mathbb{Z}/2 \ar[ur, color=orange] \& \mathbb{Z}/2 \ar[u, color=orange] \& \mathbb{Z}/2 \ar[ul, color=orange] \\
\& 0 \&
\end{tikzcd}}$
& \\ \hline
\end{tabular} }
\end{center}
\end{table}
In \cref{tab-newKMackey}, \(\underline{\Z}^*\) is the dual to the constant Mackey functor \(\underline{\Z}\) and \(\underline{\Z}(2,1)^*\) is the dual to \(\underline{\Z}(2,1)\), so named for consistency with \cite{Y}. The Mackey functor \(\underline{\mathbb{M}}\) results from the injection \(\underline{\Z}^*\hookrightarrow \underline{\Z} \twoheadrightarrow \underline{\mathbb{M}}\). In each Mackey functor, the blue arrows are multiplication by one and the orange arrows are multiplication by two.
\begin{prop} \label{equivalence:ZZ*}
We have the equivalence
\begin{align*}
\Si{-\rho} H_K\underline{\Z} &\simeq \Si{-4} H_K\underline{\Z}^*.
\end{align*}
\end{prop}
\begin{proof}
This follows from \cref{homotopy:-rho:M}.
\end{proof}
\subsection{Induced Mackey Functors} \label{Induction_Formula}
\phantom{a}
We will now give an explicit description of \(K\)-Mackey functors induced from the cyclic subgroups.
We recall the following standard result.
\begin{lem}[Shearing isomorphism] \label{shearingiso}
Let \(M\) be a \(\mathbb{Z}[G]\)-module and \(H\) a subgroup of \(G\). Then we have an isomorphism of \(\mathbb{Z}[G]\)-modules,
\begin{center}
\begin{tikzpicture}
%
\node (L) at (-1.5,0) {$\mathbb{Z}[G]\otimes_{H} M$};
\node (M) at (0,0) {$\cong$};
\node (R) at (1.5,0) {$\mathbb{Z}[G/H]\otimes M$};
\node (loop1) at (-.8,0) {$\phantom{A}$};
\node (loop2) at (2.3,0) {$\phantom{A}$};
%
\draw[->,out=240,in=300,loop, looseness=4] (loop1) to
node[below] {\scalebox{0.6}{$H$}} (loop1);
\draw[->,out=240,in=300,loop, looseness=4] (loop2) to node[below] {\scalebox{0.6}{$G$}} (loop2);
\end{tikzpicture}
\end{center}
where \(G\) acts on the left of \(\mathbb{Z}[G]\) and diagonally on \(\mathbb{Z}[G/H]\otimes M\) via the map \(g\otimes m \mapsto gH \otimes g\cdot m\).
\end{lem}
Let \(\ul N\) be a \(C_2\)-Mackey functor. Take \(h\inK\) and set \(\Delta_h:N_e \rightarrow \mathbb{Z}[K] \otimes_{\mathbb{Z}[L]} N_e\) and \(\nabla_h:\mathbb{Z}[K] \otimes_{\mathbb{Z}[L]} N_e \rightarrow N_e\) to be \(n\mapsto (e+h)\otimes n\) and \(e\otimes n,h\otimes n\mapsto n\), respectively.
Then \(\uparrow^K_L \ul N\) is
\begin{equation*}
\raisebox{12ex}{$\uparrow^K_L \ul N \quad \cong$} \quad
\begin{tikzpicture}
\node (MK) at (0,4) {$N_{C_2}$};
\node (ML) at (-2,2) {$\mathbb{Z}[K/L] \otimes N_{C_2}$};
\node (MD) at (0,2) {$N_e$};
\node (MR) at (2,2) {$N_e$};
\node (Me) at (0,0) {$\mathbb{Z}[K] \otimes_{\mathbb{Z}[L]} N_e$};
\arres{MK}{ML}{0}{0}{$\Delta$}
\arres{MK}{MD}{-6}{-16}{$\res{C_2}{e}$}
\arres{MK}{MR}{-3}{-15}{$\res{C_2}{e}$}
\artr{ML}{MK}{0}{0}{$\nabla$}
\artr{MD}{MK}{0}{0}{$\tr{C_2}{e}$}
\artr{MR}{MK}{0}{0}{$\tr{C_2}{e}$}
\draw[->,black,bend right=2ex] (ML) to node[sloped, anchor=center, xshift={0pt},yshift={-5pt}] {\scalebox{0.5}{$\text{id}\otimes \res{C_2}{e}$}} (Me);
\arres{MD}{Me}{-5}{0}{$\Delta_d$}
\arres{MR}{Me}{0}{0}{$\Delta_r$}
\draw[->,black,bend right=2ex] (Me) to node[sloped, anchor=center, xshift={0pt},yshift={-5pt}] {\scalebox{0.5}{$\text{id} \otimes \tr{C_2}{e}$}} (ML);
\artr{Me}{MD}{0}{0}{$\nabla_d$}
\artr{Me}{MR}{0}{0}{$\nabla_r$}
\end{tikzpicture}
\end{equation*}
The Weyl group \(W_K(D)\) acts on \(\uparrow_L^K \ul{N}(D)= N_e\) via the isomorphism \(W_K(D)\cong C_2\) and the given action of \(C_2 = W_{C_2}(e)\) on \(N_e\). Similarly for the action of \(W_{K}(R)\) on \(N_e\). As for \(L\), \(W_K(L)\) acts only on \(\mathbb{Z}[K/L]\). Finally, \(W_K(e) = K\) acts on the \(\mathbb{Z}[K]\) factor of \(\mathbb{Z}[K]\otimes_{\mathbb{Z}[L]} N_e\).
We are now going to define the unit map \(\ul{M}\rightarrow \uparrow^K_{L} \downarrow^K_L \ul M\). The pullback along \(K\twoheadrightarrow K/L\cong C_2\) of \(\mathbb{Z}\hookrightarrow \mathbb{Z}[C_2]\) is an inclusion \(\mathbb{Z}\hookrightarrow \mathbb{Z}[K/L]\). Tensoring with \(M_e\) gives
\begin{equation*}
\begin{tikzcd}
M_e \arrow[r, hook, "i_L"] & \mathbb{Z}[K/L]\otimes M_e.
\end{tikzcd}
\end{equation*}
We will also use \(i_L\) to denote the inclusion of \(K/L\) fixed points
\begin{equation*}
\begin{tikzcd}
M_L \arrow[r, hook, "i_L"] & \mathbb{Z}[K/L]\otimes M_L.
\end{tikzcd}
\end{equation*}
Consider
\begin{align*}
\nabla^r: \mathbb{Z}[K/L]\otimes M_e \rightarrow M_e \qquad \text{and} \qquad \Delta^r:M_e \rightarrow \mathbb{Z}[K/L]\otimes M_e.
\end{align*}
Let \(\nabla^r\) be the action map and define \(\Delta^r\) by \(m\mapsto e\otimes m + r\otimes r\cdot m\). Then, for \(\ul M\in \text{Mack}(K)\), the map \(\ul M \rightarrow \uparrow^K_{L} \downarrow^K_L \ul M\) is
\begin{equation*}
\begin{tikzpicture}
\node (MK) at (0,4) {$M_K$};
\node (ML) at (-2,2) {$M_L$};
\node (MD) at (0,2) {$M_D$};
\node (MR) at (2,2) {$M_R$};
\node (Me) at (0,0) {$M_e$};
%
\node (rMK) at (7,4) {$M_L$};
\node (rML) at (5,2) {$\mathbb{Z}[K/L] \otimes M_L$};
\node (rMD) at (7,2) {$M_e$};
\node (rMR) at (9,2) {$M_e$};
\node (rMe) at (7,0) {$\mathbb{Z}[K/L] \otimes M_e$};
\arres{MK}{ML}{0}{0}{}
\arres{MK}{MD}{0}{0}{}
\arres{MK}{MR}{0}{0}{}
\artr{ML}{MK}{0}{0}{}
\artr{MD}{MK}{0}{0}{}
\artr{MR}{MK}{0}{0}{}
\arres{ML}{Me}{0}{0}{}
\arres{MD}{Me}{0}{0}{}
\arres{MR}{Me}{0}{0}{}
\artr{Me}{ML}{0}{0}{}
\artr{Me}{MD}{0}{0}{}
\artr{Me}{MR}{0}{0}{}
\arres{rMK}{rML}{0}{0}{$\Delta^r$}
\arres{rMK}{rMD}{0}{0}{}
\arres{rMK}{rMR}{0}{0}{}
\artr{rML}{rMK}{0}{0}{$\nabla^r$}
\artr{rMD}{rMK}{0}{0}{}
\artr{rMR}{rMK}{0}{0}{}
\draw[->,black,bend right=2ex] (rML) to node[sloped, anchor=center, xshift={0pt},yshift={-5pt}] {\scalebox{0.7}{$\text{id}\otimes r\cdot \res{L}{e}$}} (rMe);
\arres{rMD}{rMe}{-5}{2}{$\Delta^d$}
\arres{rMR}{rMe}{0}{0}{$\Delta^r$}
\draw[->,black,bend right=2ex] (rMe) to node[sloped, anchor=center, xshift={0pt},yshift={5pt}] {\scalebox{0.7}{$\text{id}\otimes r\cdot \tr{L}{e}$}} (rML);
\artr{rMe}{rMD}{0}{0}{$\nabla^d$}
\artr{rMe}{rMR}{0}{0}{$\nabla^r$}
%
\draw[->] (MK) to node[above, xshift={-1pt}] {\scalebox{0.7}{$\res{K}{L}$}} (rMK);
\draw[->] (MR) to node[above, xshift={-1pt}] {\scalebox{0.7}{$\begin{pmatrix}i_L \\ \res{D}{e} \\ \res{R}{e} \end{pmatrix}$}} (rML);
\draw[->] (Me) to node[above, xshift={-1pt}] {\scalebox{0.7}{$i_L$}} (rMe);
\end{tikzpicture}
\end{equation*}
Again, \(W_K(D)\) acts on \(M_e\) via the \(C_2\)-action \(W_{C_2}(e)\circlearrowright M_e\). The same applies for \(W_K(R)\circlearrowright M_e\). Note we have used \cref{shearingiso} to rewrite the bottom group in \(\uparrow^K_L \downarrow^K_L \ul M\). We now have a diagonal action \(K/L\) on \(\mathbb{Z}[K/L]\otimes M_L\). Similarly, \(K/e\) acts diagonally on \(\mathbb{Z}[K/L]\otimes M_e\).
\begin{exmp}
For \(\ul M=\underline{\Z}\), $\uparrow_L^K \underline{\Z}$ and $\underline{\Z} \rightarrow \uparrow_L^K \underline{\Z}$ are as follows.
\begin{equation*}
\begin{tikzpicture}
\node (MK) at (0,4) {$\mathbb{Z}$};
\node (ML) at (-2,2) {$\mathbb{Z}$};
\node (MD) at (0,2) {$\mathbb{Z}$};
\node (MR) at (2,2) {$\mathbb{Z}$};
\node (Me) at (0,0) {$\mathbb{Z}$};
%
\node (rMK) at (7,4) {$\mathbb{Z}$};
\node (rML) at (5,2) {$\mathbb{Z}[K/L] \otimes \mathbb{Z}$};
\node (rMD) at (7,2) {$\mathbb{Z}$};
\node (rMR) at (9,2) {$\mathbb{Z}$};
\node (rMe) at (7,0) {$\mathbb{Z}[K/L] \otimes \mathbb{Z}$};
\draw[->,blue,bend right=2ex] (MK) to (ML);
\draw[->,blue,bend right=2ex] (MK) to (MD);
\draw[->,blue,bend right=2ex] (MK) to (MR);
\draw[->,orange,bend right=2ex] (ML) to (MK);
\draw[->,orange,bend right=2ex] (MD) to (MK);
\draw[->,orange,bend right=2ex] (MR) to (MK);
\draw[->,blue,bend right=2ex] (ML) to (Me);
\draw[->,blue,bend right=2ex] (MD) to (Me);
\draw[->,blue,bend right=2ex] (MR) to (Me);
\draw[->,orange,bend right=2ex] (Me) to (ML);
\draw[->,orange,bend right=2ex] (Me) to (MD);
\draw[->,orange,bend right=2ex] (Me) to (MR);
\arres{rMK}{rML}{0}{0}{$\Delta^r$}
\draw[->,blue,bend right=2ex] (rMK) to (rMD);
\draw[->,blue,bend right=2ex] (rMK) to (rMR);
\artr{rML}{rMK}{0}{0}{$\nabla^r$}
\draw[->,orange,bend right=2ex] (rMD) to (rMK);
\draw[->,orange,bend right=2ex] (rMR) to (rMK);
\draw[->,blue,bend right=2ex] (rML) to node[sloped, anchor=center, xshift={0pt},yshift={-5pt}] {\scalebox{0.7}{$\text{id}\otimes r\cdot 1$}} (rMe);
\arres{rMD}{rMe}{-5}{2}{$\Delta^d$}
\arres{rMR}{rMe}{0}{0}{$\Delta^r$}
\draw[->,orange,bend right=2ex] (rMe) to node[sloped, anchor=center, xshift={0pt},yshift={5pt}] {\scalebox{0.7}{$\text{id}\otimes r\cdot 2$}} (rML);
\artr{rMe}{rMD}{0}{0}{$\nabla^d$}
\artr{rMe}{rMR}{0}{0}{$\nabla^r$}
%
\draw[->] (MK) to node[above, xshift={-1pt}] {\scalebox{0.7}{$1$}} (rMK);
\draw[->] (MR) to node[above, xshift={-1pt}] {\scalebox{0.7}{$\begin{pmatrix}i_L \\ 1 \\ 1 \end{pmatrix}$}} (rML);
\draw[->] (Me) to node[above, xshift={-1pt}] {\scalebox{0.7}{$i_L$}} (rMe);
\end{tikzpicture}
\end{equation*}
Again, blue denotes multiplication by one and orange multiplication by two.
\end{exmp}
\section{2-slice Characterization \hspace{2in}} \label{sec:2slice}
Consider a \(K\)-spectrum \(X\). Then by \cref{rho:susp:commutes},
\begin{align*}
P^n_n X &\simeq \Si{\frac{n-r}{4}\rho} P^r_r \left( \Si{-\frac{n-r}{4}\rho} X \right)
\end{align*}
where \(n\equiv r\pmod 4\) and \(0\leq r\leq 3\). Thus, to know the slices of \(X\), we need only know the 0-, 1-, 2-, and 3-slices of certain suspensions of \(X\). \cref{charac:slice:01-1} and \cref{rho:susp:commutes} characterize the 0-, 1-, and 3-slices. We now characterize the 2-slices.
\begin{prop} \label{Homotopy_Reduction}
Let \(G\) be a finite group, \(H\) an index two subgroup of \(G\), and \(\sigma^H\) the sign representation from \(G\rightarrow G/H\).
For a \(G\)-spectrum \(X\), if \(\downarrow^G_H \ul \pi_{n+1}(X) = \ul 0 = \downarrow^G_H \ul \pi_n(X)\), the natural map \(\Si{-\sigma^H} X \rightarrow X\) induces an isomorphism on \(\ul\pi_{n}\). In particular, if \(\pi_{n+1}^H(X) = 0 = \pi_n^G(X)\), then \(\pi_n^G(\Si{-\sigma^H} X) = 0\).
\end{prop}
\begin{proof}
Because \(\sigma^H\) is one-dimensional we may construct \(\Si{-\sigma^H} X\) with the cofiber sequence \((S^{-\sigma^H} \rightarrow S^0 \rightarrow G/H_+) \wedge X\). The resulting long exact sequence in homotopy is
\begin{align} \label{les:reduction}
\uparrow^G_H \downarrow^G_H \ul\pi_{n+1}(X) \rightarrow \ul\pi_n(\Si{-\sigma^H} X) \rightarrow \ul\pi_n(X) \rightarrow \uparrow^G_H \downarrow^G_H \ul\pi_{n}(X).
\end{align}
As \(\uparrow^G_H \downarrow^G_H \ul\pi_{n+1}(X) = \ul 0 = \uparrow^G_H \downarrow^G_H \ul \pi_n(X)\),
we find that \(\ul\pi_n(\Si{-\sigma^H} X) \cong \ul\pi_n(X)\).
Now suppose that \(\pi_{n+1}^H(X) = 0 = \pi_n^G(X)\). Because \(\pi^H_{n+1}(X)\) is the value of \(\uparrow^G_H \downarrow^G_H \ul\pi_{n+1}(X)\) at the orbit \(G/H\), the left three terms of \cref{les:reduction} prove the exact sequence
\begin{align*}
0 \rightarrow \pi_n^G(\Si{-\sigma^H} X) \rightarrow 0.
\end{align*}
Consequently, \(\pi_n^G(\Si{-\sigma^H} X) = 0\).
\end{proof}
\begin{cor} \label{EM:homotopy:reduction}
Let \(H\ul M\) be an Eilenberg-MacLane \(G\)-spectrum and \(V \cong \mathbf{s} \oplus \bigoplus_{i=1}^r \sigma^{H_i}\) be a real representation with \(s\) copies of the trivial representation and each \(\sigma^{H_i}\) the sign representation from \(G\rightarrow G/H_i\), where \(H_i\) is an index two subgroup of \(G\). Then \(\Si{-V} H\ul M\) does not have nontrivial homotopy above degree \(-s\).
\end{cor}
\begin{proof}
First apply \cref{les:reduction} to \(X_1 := \Si{-\sigma^{H_1}} H\ul M\) with \(n\geq 1\). Then repeat for \mbox{\(X_i:= \Si{-\sigma^{H_i}} X_{i-1}\)} where \(2\leq i\leq r\). Then \(\Si{-V} H\ul M \simeq \Si{-s} X_{r}\) has no homotopy above degree \(-s\).
\end{proof}
Recall from \cref{subsec:signrep} that \(\alpha\), \(\beta\), and \(\gamma\) are the sign representations of \(K\).
\begin{lem} \label{homotopy:-beta:M}
Let \(\ul M\) be a \(K\)-Mackey functor. The nontrivial homotopy of \(\Si{-\beta} H\ul M\) is
\begin{center}
\begin{tikzpicture}[scale=0.8, every node/.style={scale=0.75}, bend right=2ex]
\node (pi0) at (-8,2) {$\ul \pi_0 \quad =$};
\node (lMK) at (-5,4) {$\ker \res{K}{L}$};
\node (lML) at (-7,2) {$0$};
\node (lMD) at (-5,2) {$\ker \res{D}{e}$};
\node (lMR) at (-3,2) {$\ker \res{R}{e}$};
\node (lMe) at (-5,0) {$0$};
%
%
\begin{scope}[xshift=-3cm]
\node (pi-1) at (3,2) {$\ul \pi_{-1} \quad =$};
\node (MK) at (7,4) {$M_L / \text{im } \res{K}{L}$};
\node (ML) at (5,2) {$\mathbb{Z}_\sigma^L\otimes M_L$};
\node (MD) at (7,2) {$M_e / \text{im } \res{D}{e}$};
\node (MR) at (9,2) {$M_e / \text{im } \res{R}{e}$};
\node (Me) at (7,0) {$\mathbb{Z}_\sigma^L\otimes M_e$};
\end{scope}
\draw[->,black] (lMK) to (lMR);
\draw[->,black] (lMR) to (lMK);
\draw[->,black] (lMK) to (lMD);
\draw[->,black] (lMD) to (lMK);
\arres{MK}{ML}{-2}{0}{$\varphi_r^{K L}$}
\draw[->,black] (MK) to (MD);
\draw[->,black] (MK) to (MR);
\artr{ML}{MK}{0}{0}{$\pi$}
\draw[->,black] (MD) to (MK);
\draw[->,black] (MR) to (MK);
\draw[->,black] (ML) to node[sloped,anchor=center,xshift={0pt},yshift={-5pt}] {\scalebox{0.7}{$1\otimes r\cdot \res{L}{e}$}} (Me);
\draw[->,black] (Me) to node[sloped,anchor=center,xshift={0pt},yshift={5pt}] {\scalebox{0.7}{$1\otimes r\cdot \tr{L}{e}$}} (ML);
\arres{MD}{Me}{-6}{0}{$\varphi_r^{De}$}
\arres{MR}{Me}{1}{2}{$\varphi_r^{Re}$}
\artr{Me}{MD}{0}{0}{$\pi$}
\artr{Me}{MR}{0}{0}{$\pi$}
\end{tikzpicture}
\end{center}
Here, \(\varphi_r^{K L}\) is induced by \(\Delta^r\) in the square
\begin{equation*}
\begin{tikzcd}
M_L \arrow[r, "\pi"] \arrow[d, "\Delta^r"] & M_L/ \text{im } \res{K}{L} \arrow[d, "\varphi_r^{K L}"] \\
\mathbb{Z}[K/L] \otimes M_L \arrow[r, "q_L"] & \mathbb{Z}_\sigma^L\otimes M_L
\end{tikzcd}
\end{equation*}
and is given by \(\varphi_r^{KL}(m) = m-r\cdot m\). The maps \(\varphi_r^{De}:M_e/\text{im } \res{D}{e}\rightarrow \mathbb{Z}_\sigma^L\otimes M_e\) and \(\varphi_r^{Re}:M_e/\text{im } \res{R}{e}\rightarrow \mathbb{Z}_\sigma^L\otimes M_e\) are defined similarly.
\end{lem}
\begin{proof}
We have the cofiber sequence \((S^{-\beta} \rightarrow S^0 \rightarrow K/L_+)\wedge H\ul M\). The result then follows from the description of the map \(\ul M \rightarrow \uparrow^K_L\downarrow^K_L \ul M\) given in \cref{Induction_Formula}.
\end{proof}
\begin{prop} \label{homotopy:-rho:M}
Let \(\ul M\) be a \(K\)-Mackey functor. The nontrivial homotopy of \(\Si{-\alpha-\beta-\gamma} H\ul M\) is
\begin{tikzpicture}[scale=0.8, every node/.style={scale=0.75}]
\node (pi0) at (-9,2) {$\ul \pi_0 \quad =$};
%
\node (lMK) at (-5,4) {$\ker (\res{K}{L} + \res{K}{D} + \res{K}{R})$};
\node (lML) at (-7,2) {$0$};
\node (lMD) at (-5,2) {$0$};
\node (lMR) at (-3,2) {$0$};
\node (lMe) at (-5,0) {$0$};
%
%
%
\begin{scope}[xshift=-3cm]
\node (pi-1) at (3,2) {$\ul \pi_{-1} \quad =$};
%
\node (MK) at (7,4) {$E_2$};
\node (ML) at (5,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^L &\otimes \\ \ker & \res{L}{e}
\end{aligned}$}};
\node (MD) at (7,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^D &\otimes \\ \ker & \res{D}{e}
\end{aligned}$}};
\node (MR) at (9,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^R &\otimes \\ \ker & \res{R}{e}
\end{aligned}$}};
\node (Me) at (7,0) {$0$};
\end{scope}
%
\arres{MK}{ML}{0}{0}{}
\arres{MK}{MD}{0}{0}{}
\arres{MK}{MR}{0}{0}{}
\artr{ML}{MK}{0}{0}{}
\artr{MD}{MK}{0}{0}{}
\artr{MR}{MK}{0}{0}{}
\end{tikzpicture}
\begin{tikzpicture}[scale=0.8, every node/.style={scale=0.75}]
\node (pi-2) at (-9,2) {$\ul \pi_{-2} \quad =$};
%
\node (lMK) at (-5,4) {$E_3$};
\node (lML) at (-7,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^L \otimes& \\ \ker\varphi_r^{Le}
\end{aligned}$}};
\node (lMD) at (-5,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^D \otimes& \\ \ker \varphi_r^{De}
\end{aligned}$}};
\node (lMR) at (-3,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^R \otimes& \\ \ker\varphi_r^{Re}
\end{aligned}$}};
\node (lMe) at (-5,0) {$0$};
%
\arres{lMK}{lML}{0}{0}{}
\arres{lMK}{lMD}{0}{0}{}
\arres{lMK}{lMR}{0}{0}{}
\artr{lML}{lMK}{0}{0}{}
\artr{lMD}{lMK}{0}{0}{}
\artr{lMR}{lMK}{0}{0}{}
%
%
\begin{scope}[xshift=-3cm]
\node (pi-3) at (3,2) {$\ul \pi_{-3} \quad =$};
%
\node (MK) at (7,4) {$(M_e / \text{im } \varphi_r^{De}) / \text{im } \varphi_r^{R,L}$};
\node (ML) at (5,2) {$M_e / \text{im } \varphi_d^{Le}$};
\node (MD) at (7,2) {$M_e / \text{im } \varphi_r^{De}$};
\node (MR) at (9,2) {$M_e / \text{im } \varphi_r^{Re}$};
\node (Me) at (7,0) {$M_e$};
\end{scope}
%
\arres{MK}{ML}{0}{0}{$\varphi_d$}
\arres{MK}{MD}{-5}{-10}{$\varphi_l$}
\arres{MK}{MR}{0}{-14}{$\varphi_d$}
\artr{ML}{MK}{0}{0}{$\pi$}
\artr{MD}{MK}{0}{0}{$\pi$}
\artr{MR}{MK}{0}{0}{$\pi$}
\arres{ML}{Me}{-15}{-5}{$\varphi_{l}^{dLe}$}
\arres{MD}{Me}{-10}{5}{$\varphi_{d}^{rDe}$}
\arres{MR}{Me}{0}{0}{$\varphi_{l}^{rRe}$}
\artr{Me}{ML}{0}{0}{$\pi$}
\artr{Me}{MD}{0}{0}{$\pi$}
\artr{Me}{MR}{0}{0}{$\pi$}
\end{tikzpicture}
where \(E_1\), \(E_2\), and \(E_3\) are extensions
\begin{equation*}
\begin{tikzcd}[row sep=small]
M_e/\text{im } \varphi_r^{De} \arrow[r] & E_3 \arrow[r] & (M_e/\text{im } \res{R}{e})/\text{im } \res{L}{e} \\
\ker\res{D}{e} \arrow[r] & E_2 \arrow[r] & E_1 \\
\ker\res{R}{e} \arrow[r] & E_1 \arrow[r] & M_L/\text{im } \res{K}{L}
\end{tikzcd}
\end{equation*}
Let \(\varphi_h^{*}:\overline{M_e} \rightarrow \overline{M_e}'\) be one of the maps shown. Then \(\varphi_h^{*}(m) = m-h\cdot m\). Additionally, \((M_e/\text{im } \varphi_r^{De})/\text{im } \varphi_r^{R,L}\) is the cokernel of the map
\begin{equation*}
\begin{tikzcd}
(M_e/\text{im } \res{R}{e})/\text{im } \res{L}{e} \arrow[r, "\varphi_r^{R,L}"] & M_e/\text{im } \varphi_r^{De}.
\end{tikzcd}
\end{equation*}
\end{prop}
\begin{proof}
\cref{homotopy:-beta:M} supplies the homotopy for \(\Si{-\beta} H\ul M\). We continue constructing \(\Si{-\alpha-\beta-\gamma} H\ul M\) iteratively. The cofiber sequence
\begin{align*}
(S^{-\alpha} \rightarrow S^0 \rightarrow K/R_+) \wedge \Si{-\beta} H\ul M
\end{align*}
results in the homotopy of \(\Si{-\alpha-\beta} H\ul M\)
\begin{tikzpicture}[scale=0.8, every node/.style={scale=0.75}]
\node (pi0) at (-9,2) {$\ul \pi_0 \quad =$};
%
\node (lMK) at (-5,4) {$\ker (\res{K}{L} + \res{K}{R})$};
\node (lML) at (-7,2) {$0$};
\node (lMD) at (-5,2) {$\ker \res{K}{D}$};
\node (lMR) at (-3,2) {$0$};
\node (lMe) at (-5,0) {$0$};
%
\arres{lMK}{lMD}{0}{0}{}
\artr{lMD}{lMK}{0}{0}{}
%
%
\begin{scope}[xshift=-3cm]
\node (pi-1) at (3,2) {$\ul \pi_{-1} \quad =$};
\node (MK) at (7,4) {$E_1$};
\node (ML) at (5,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^L &\otimes \\ \ker & \res{L}{e}
\end{aligned}$}};
\node (MD) at (7,2) {\scalebox{0.9}{$\ker\varphi_r^{De}$}};
\node (MR) at (9,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^R &\otimes \\ \ker & \res{R}{e}
\end{aligned}$}};
\node (Me) at (7,0) {$0$};
\end{scope}
%
\arres{MK}{ML}{0}{0}{}
\arres{MK}{MD}{0}{0}{}
\arres{MK}{MR}{0}{0}{}
\artr{ML}{MK}{0}{0}{}
\artr{MD}{MK}{0}{0}{}
\artr{MR}{MK}{0}{0}{}
\end{tikzpicture}
\begin{tikzpicture}[scale=0.8, every node/.style={scale=0.75}]
\node (pi-2) at (-9,2) {$\ul \pi_{-2} \quad =$};
%
\node (lMK) at (-5,4) {$(M_e/\text{im } \res{D}{e})/ \text{im } \res{L}{e}$};
\node (lML) at (-7,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^L &\otimes \\ M_e / &\text{im } \res{L}{e}
\end{aligned}$}};
\node (lMD) at (-5,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^D &\otimes \\ M_e / &\text{im } \varphi_r^{De}
\end{aligned}$}};
\node (lMR) at (-3,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^R &\otimes \\ M_e / &\text{im } \res{R}{e}
\end{aligned}$}};
\node (lMe) at (-5,0) {$\mathbb{Z}_\sigma^D \otimes M_e$};
%
\arres{lMK}{lML}{0}{0}{$\varphi_r^{D,L}$}
\arres{lMK}{lMD}{-8}{-17}{$\varphi_r^{D,L}$}
\arres{lMK}{lMR}{0}{-23}{$\varphi_d^{D,L}$}
\artr{lML}{lMK}{-11}{0}{$\pi$}
\artr{lMD}{lMK}{0}{0}{$\pi$}
\artr{lMR}{lMK}{0}{0}{$\pi$}
\arres{lML}{lMe}{-10}{-8}{$\varphi_d^{Le}$}
\arres{lMD}{lMe}{-6}{3}{$\varphi_d^{rDe}$}
\draw[->,black,bend right=2ex] (lMR) to node[sloped, anchor=center, xshift={0pt},yshift={-5pt}] {\scalebox{0.7}{$1\otimes \varphi_r^{Re}$}} (lMe);
\artr{lMe}{lML}{0}{0}{$\pi$}
\artr{lMe}{lMD}{-2}{0}{$\pi$}
\draw[->,black,bend right=2ex] (lMe) to node[sloped, anchor=center, xshift={0pt},yshift={-5pt}] {\scalebox{0.7}{$1\otimes \pi$}} (lMR);
\end{tikzpicture}
Finally, the cofiber sequence \((S^{-\gamma} \rightarrow S^0 \rightarrow K/D_+) \wedge \Si{-\alpha-\beta} H\ul M\) provides the result.
\end{proof}
\begin{lem} \label{Kernel_Equivalences}
Let \(\phi:G\rightarrow N\) and \(\psi:G\rightarrow M\) be group homomorphisms. Then
\begin{align*}
\ker(\phi, \psi) = \ker(\phi) \cap \ker(\psi) = \ker( \phi\vert_{\ker(\psi)})
\end{align*}
where \((\phi,\psi):G\rightarrow N\oplus M\) is defined by \(g\mapsto \phi(g)\oplus \psi(g)\).
\end{lem}
\begin{prop} \label{2slice:equivalences:si2}
Let \(\ul M\) be a \(K\) Mackey functor where the restrictions \(\res{L}{e}\), \(\res{D}{e}\), and \(\res{R}{e}\) are injective. The following are equivalent.
\begin{enumerate}[(A)]
\item The sequence
\begin{align*}
M_K \xrightarrow{\res{K}{L}+\res{K}{D}+\res{K}{R}} M_L \oplus M_D \oplus M_R \xrightarrow{\left(\begin{smallmatrix} -\res{L}{e} & \res{L}{e} & 0 \\ \res{D}{e} & 0 & -\res{D}{e} \\ 0 & -\res{R}{e} & \res{R}{e} \end{smallmatrix}\right)} M_e^3
\end{align*}
is exact. \label{2slice:si2:exact1}
\item \(\text{im } \res{K}{H_1} = (\res{H_1}{e})^{-1}(\text{im } \res{H_2}{e}) \cap (\res{H_1}{e})^{-1}(\text{im } \res{H_3}{e})\) where \(H_1\), \(H_2\), \(H_3\) are distinct order two subgroups of \(K\). \label{2slice:res:injective}
This equality is represented by the diagram below.
\begin{equation*}
\begin{tikzcd}
\text{im } \res{K}{H_1} \arrow[d, hook] \ar[r] & \text{im } \res{H_2}{e} \cap \text{im } \res{H_3}{e} \ar[d, hook] \\
M_{H_1} \ar[r, "\res{H_1}{e}"] & M_e
\end{tikzcd}
\end{equation*}
\end{enumerate}
\end{prop}
\begin{proof}
Without loss of generality, let \(H_1=L\), \(H_2=D\), and \(H_3=R\). For convenience, set \(I = (\res{L}{e})^{-1}(\text{im } \res{D}{e}) \cap (\res{L}{e})^{-1}(\text{im } \res{R}{e})\).
\cref{2slice:si2:exact1} \(\Rightarrow\) \cref{2slice:res:injective}:
Let \(x\in \text{im } \res{K}{L}\). Then we have \(k\in M_K\) such that \(\res{K}{L}(k) = x\). Additionally, \(\res{L}{e}(x) = \res{D}{e} \res{K}{D} (k) = \res{R}{e} \res{K}{R} (k)\). It follows that \(x\in I\).
Now suppose \(x\in I\). We then have \(y\in D\) and \(z\in R\) such that
\begin{align*}
\res{L}{e}(x) = \res{D}{e}(y) = \res{R}{e}(z).
\end{align*}
Thus, \((x,y,z) \in \ker \left(\begin{smallmatrix} -\res{L}{e} & \res{L}{e} & 0 \\ \res{D}{e} & 0 & -\res{D}{e} \\ 0 & -\res{R}{e} & \res{R}{e} \end{smallmatrix}\right)\). Consequently, we have \(x = \res{K}{L}(k)\) for some \(k\in M_K\); that is, \(x\in \text{im } \res{K}{L}\).
\cref{2slice:res:injective} \(\Rightarrow\) \cref{2slice:si2:exact1}:
Let \((x,y,z)\in \ker \left(\begin{smallmatrix} -\res{L}{e} & \res{L}{e} & 0 \\ \res{D}{e} & 0 & -\res{D}{e} \\ 0 & -\res{R}{e} & \res{R}{e} \end{smallmatrix}\right)\). Then \(\res{L}{e}(x) = \res{D}{e}(y) = \res{R}{e}(z)\), so \(x\in \text{im } \res{K}{L}\). So there is some \(k\in M_K\) such that \(x = \res{K}{L}(k)\). Because \(\res{D}{e}\) and \(\res{R}{e}\) are injective, it must be that \(y = (\res{D}{e})^{-1}(\res{L}{e}(x))\) and \(z = (\res{R}{e})^{-1}(\res{L}{e}(x))\). Hence, \cref{2slice:si2:exact1} is exact.
\end{proof}
\begin{prop} \label{2slice:characterization:si1}
The \(K\)-spectrum \(\Si{1} H\ul N\) is a 2-slice if and only if \(N_e = 0\) and \(N_K \rightarrow N_L \oplus N_D \oplus N_R\) is injective.
\end{prop}
\begin{proof}
By \cref{thm:atleastn}, we find that \(\Si{1} H\ul N\geq 2\) if and only if \(N_e = 0\). Consequently, going forward we may assume \(N_e = 0\).
Now \(\Si{1} H\ul N\leq 2\) if and only if \([S^{k\rho_H+r}, \Si{2} H\ul N]^H=0\) for all \(k\geq \frac{3}{\abs{H}}\) and \(r\geq 0\).
Because \(N_e = 0\), \cref{C2:2slice} implies that \(i_H^*(\Si{1} H\ul N)\) is a 2-slice, where \(H\) is an order two subgroup of \(K\). Thus, to finish the equivalence, we only need consider
\begin{align*}
[S^{k\rho_K+r}, \Si{1} H\ul N]^K = [S^{k+r-1}, \Si{-k\overline{\rho}_K} H\ul N]^K
\end{align*}
for all \(k\geq 1\) and \(r\geq 0\).
From \cref{EM:homotopy:reduction}, we need only concern ourselves with \([S^0, \Si{-k\overline{\rho}_K} H\ul N]\). From \cref{homotopy:-rho:M},
\begin{align*}
\ul \pi_0(\Si{-\overline\rho} H\ul N) &= \phi^*_K( \ker \res{K}{L} \cap \ker \res{K}{D} \cap \ker \res{K}{R}).
\end{align*}
Therefore \cref{Homotopy_Reduction} shows that \(\ul \pi_0(\Si{-\overline\rho} H \ul N)\) vanishes if and only if \(\ul \pi_0(\Si{-k\overline\rho} H \ul N)\) vanishes for all \(k\geq 1\). Hence, \(\Si{1} H\ul N\leq 2\) if and only if \(\ker \res{K}{L} \cap \ker \res{K}{D} \cap \ker \res{K}{R} = \{0\}\). By \cref{Kernel_Equivalences} this is equivalent to \(N_K \rightarrow N_L \oplus N_D \oplus N_R\) being injective.
\end{proof}
\begin{prop} \label{2slice:characterization:si2}
The spectrum \(\Si{2} H\ul M\) is a 2-slice if and only if all restrictions of \(\ul M\) are injective and the sequence
\begin{align*}
M_K \xrightarrow{\res{K}{L}+\res{K}{D}+\res{K}{R}} M_L \oplus M_D \oplus M_R \xrightarrow{\left(\begin{smallmatrix} -\res{L}{e} & \res{L}{e} & 0 \\ \res{D}{e} & 0 & -\res{D}{e} \\ 0 & -\res{R}{e} & \res{R}{e} \end{smallmatrix}\right)} M_e^3
\end{align*}
is exact.
\end{prop}
\begin{proof}
Note that \(\Si{2} H\ul M \geq 2\), so we just need to show that \(\Si{2} H\ul M\leq 2\). Now \(\Si{2} H\ul M\leq 2\) if and only if \([S^{k\rho_H+r}, \Si{2} H\ul M]^H=0\) for all \(k\geq \frac{3}{\abs{H}}\) and \(r\geq 0\). For \(H\) an order two subgroup of \(K\), \(i_H^* \Si{2} H\ul M\leq 2\) if and only if \(\res{H}{e}\)
is injective. Consequently, going forward we may assume \(\res{L}{e}\), \(\res{D}{e}\), and \(\res{R}{e}\) are injective.
To finish the equivalence, we only need consider
\begin{align*}
[S^{k\rho_K+r}, \Si{2} H\ul M]^K = [S^{k+r-2}, \Si{-k\overline{\rho}_K} H\ul M]^K
\end{align*}
for all \(k\geq 1\) and \(r\geq 0\). By \cref{EM:homotopy:reduction}, it is enough to examine \([S^0, \Si{-k\overline{\rho}_K} H\ul M]\) and \([S^{-1}, \Si{-k\overline{\rho}_K} H\ul M]\). From \cref{homotopy:-rho:M},
\begin{align*}
\ul \pi_0(\Si{-\overline\rho} H\ul M) &= \phi^*_K( \ker \res{K}{L})
\end{align*}
and
\begin{align*}
\ul \pi_{-1}(\Si{-\overline\rho} H\ul M) &= \phi^*_K \left( \frac{(\res{L}{e})^{-1}(\text{im }\res{D}{e}) \cap (\res{L}{e})^{-1}(\text{im } \res{R}{e})}{\text{im } \res{K}{L}} \right).
\end{align*}
Note that \cref{homotopy:-rho:M} states
\begin{align*}
\ul \pi_0(\Si{-\overline\rho} H\ul M) = \phi^*_K( \ker \res{K}{L} \cap \ker\res{K}{D} \cap\ker\res{K}{R}),
\end{align*}
but because the lower restrictions are injective, these kernels coincide.
\cref{Homotopy_Reduction} then yields that
\begin{align*}
\ul\pi_0^G(\Si{-k\overline\rho} H\ul M) &\cong \ul\pi_0^G(\Si{-\overline\rho} H\ul M) = \ker \res{K}{L} \\
\ul\pi_{-1}^G(\Si{-k\overline\rho} H\ul M) &\cong \ul\pi_{-1}^G(\Si{-\overline\rho} H\ul M) = \frac{(\res{L}{e})^{-1}(\text{im }\res{D}{e}) \cap (\res{L}{e})^{-1}(\text{im } \res{R}{e})}{\text{im } \res{K}{L}}
\end{align*}
for all \(k\geq 2\). By \cref{def:lessthan}, we find that \(\Si{2} H\ul M\leq 2\) if and only if these two homotopy groups vanish. Hence, \(\Si{2} H\ul M\leq 2\) if and only if \(\ker \res{K}{L}=\{0\}\) and \(\text{im } \res{K}{L} = (\res{L}{e})^{-1}(\text{im }\res{D}{e}) \cap (\res{L}{e})^{-1}(\text{im } \res{R}{e})\). Because the homotopy of \(\Si{-\alpha-\beta-\gamma} H\ul M\) is invariant under the order of construction -- that is, whether \(H\ul M\) is suspended by say \(-\alpha\) or \(-\beta\) first -- we find that all upper restrictions must be injective and that \cref{2slice:res:injective} must hold.
\end{proof}
\begin{thm} \label{2slice:characterization}
Suppose the only nontrivial homotopy Mackey functors of a \(K\)-spectrum \(X\) are \(\ul\pi_1(X)\) and \(\ul\pi_2(X)\) where
\begin{enumerate}
\item All restrictions of \(\ul\pi_2(X)\) are injective and the sequence
\begin{align*}
\pi_2^K(X) \xrightarrow{\res{K}{L}+\res{K}{D}+\res{K}{R}} \pi_2^L(X) \oplus \pi_2^D(X) \oplus \pi_2^R(X) \xrightarrow{\left(\begin{smallmatrix} -\res{L}{e} & \res{L}{e} & 0 \\ \res{D}{e} & 0 & -\res{D}{e} \\ 0 & -\res{R}{e} & \res{R}{e} \end{smallmatrix}\right)} \pi_2^e(X)^3
\end{align*}
is exact. \label{2slice:thm:M}
\item \(\pi_1^e(X) = 0\). \label{2slice:thm:N0}
\item \(\pi_1^K(X) \rightarrow \pi_1^L(X) \oplus \pi_1^D(X) \oplus \pi_1^R(X)\) is injective. \label{2slice:thm:N0:2}
\end{enumerate}
Then \(X\) is a 2-slice.
Conversely, if a \(K\)-spectrum \(X\) is a 2-slice, then its only nontrivial homotopy Mackey functors of a \(K\)-spectrum \(X\) are \(\ul\pi_1(X)\) and \(\ul\pi_2(X)\) where \(\Si{2} H\ul\pi_2(X)\) is a 2-slice and \(\Si{1} H\ul\pi_1(X) \in [2,4]\), i.e., both \cref{2slice:thm:M} and \cref{2slice:thm:N0} hold.
\end{thm}
\begin{proof}
Let \(\ul\pi_2(X)=\ul M\) and \(\ul\pi_1(X)=\ul N\). If these are the only nontrivial homotopy Mackey functors of \(X\), we have a fiber sequence
\begin{align} \label{2slice:fibseq}
\Si{2} H\ul M\rightarrow X\rightarrow \Si{1} H\ul N.
\end{align}
By \cref{2slice:characterization:si1} and \cref{2slice:characterization:si2}, conditions \eqref{2slice:thm:M} - \eqref{2slice:thm:N0:2} show that \(\Si{2} H\ul M\) and \(\Si{1} H\ul N\) are 2-slices. Now if \(\Si{2} H\ul M\) and \(\Si{1} H\ul N\) are 2-slices, then \(X\) must be a 2-slice as well.
Conversely, suppose that \(X\) is a 2-slice. We then find that \(X\) has no homotopy above degree two and none below degree one; thus, we have the fiber sequence in \cref{2slice:fibseq}.
Because \(X\geq 2\) and \(\Si{2} H\ul M\geq 2\), it follows that \(\Si{1} H\ul N\geq 2\). So by \cref{thm:atleastn}, \(N_e = 0\). That \(\Si{1} H\ul N\leq 4\) follows from a similar argument as in \cref{2slice:characterization:si1}.
Rotating this fiber sequence gives \(H\ul N\rightarrow \Si{2} H\ul M \rightarrow X\). As \(H\ul N\) is a 0-slice and \(X\) is a 2-slice, we have \(\Si{2} H\ul M \in [0,2]\), so it must be that \(\Si{2} H\ul M = 2\). Consequently, by \cref{2slice:characterization:si2}, \cref{2slice:thm:M} holds.
\end{proof}
It is not necessary for condition \eqref{2slice:thm:N0:2} to hold for \(X\) to be a 2-slice as we show in the following example.
\begin{exmp}
Take \(X \simeq \Si{1+\beta} H\ul E\) where
\begin{equation*}
\ul E = \begin{tikzcd}[ampersand replacement=\&, column sep=scriptsize]
\& \mathbb{F}_2 \ar[d, color=black] \ar[dr,color=black] \& \\
0 \& \mathbb{F}_2 \& \mathbb{F}_2 \\
\& \mathbb{F}_2 \ar[u, color=black] \ar[ur, color=black] \&
\end{tikzcd}
\end{equation*}
Then \(X\) is a 2-slice with \(\ul\pi_2(X) = \ul f\) and \(\ul\pi_1(X) = \underline{g}\). But \(\Si{1} H\underline{g}\) is not a 2-slice.
\end{exmp}
\begin{proof}
We can construct \(\Si{1+\beta} H\ul E\) using the cofiber sequence
\begin{align*}
(K/L_+ \rightarrow S^0 \rightarrow S^\beta) \wedge \Si{1} H\ul E.
\end{align*}
The resulting long exact sequence in homotopy is
\begin{flushright}
\begin{tikzpicture}[scale=0.8, every node/.style={scale=0.75},xscale=0.9]
\node (pi2) at (-9,1.5) {$\ul \pi_2$};
%
\node (MD) at (0,1.5) {$\ul 0$};
%
\node (rMK) at (5.5,3) {$0$};
\node (rML) at (4,1.5) {$0$};
\node (rMD) at (5.5,1.5) {$0$};
\node (rMR) at (7,1.5) {$0$};
\node (rMe) at (5.5,0) {$\mathbb{F}_2$};
%
\draw[->] (MD) to (rML);
\end{tikzpicture}
\begin{tikzpicture}[scale=0.8, every node/.style={scale=0.75},xscale=0.9]
\node (pi1) at (-9,1.5) {$\ul \pi_1$};
%
\node (lMK) at (-5.5,3) {$0$};
\node (lML) at (-7,1.5) {$0$};
\node (lMD) at (-5.5,1.5) {$\mathbb{F}_2$};
\node (lMR) at (-4,1.5) {$\mathbb{F}_2$};
\node (lMe) at (-5.5,0) {$\mathbb{F}_2[K/L]$};
%
\node (MK) at (0,3) {$\mathbb{F}_2$};
\node (ML) at (-1.5,1.5) {$0$};
\node (MD) at (0,1.5) {$\mathbb{F}_2$};
\node (MR) at (1.5,1.5) {$\mathbb{F}_2$};
\node (Me) at (0,0) {$\mathbb{F}_2$};
%
\node (rMK) at (5.5,3) {$\mathbb{F}_2$};
\node (rML) at (4,1.5) {$0$};
\node (rMD) at (5.5,1.5) {$0$};
\node (rMR) at (7,1.5) {$0$};
\node (rMe) at (5.5,0) {$0$};
%
\arres{lMD}{lMe}{-4}{0}{$\Delta$}
\arres{lMR}{lMe}{0}{0}{$\Delta$}
\artr{lMe}{lMD}{0}{0}{$\nabla$}
\artr{lMe}{lMR}{0}{0}{$\nabla$}
\draw[->, black] (MK) to node[right,xshift={2pt}] {\scalebox{0.6}{$1$}} (MR);
\draw[->, black] (MK) to node[right,xshift={2pt}] {\scalebox{0.6}{$1$}} (MD);
\draw[->, black] (Me) to node[right,xshift={2pt}] {\scalebox{0.6}{$1$}} (MR);
\draw[->, black] (Me) to node[right,xshift={-2pt}] {\scalebox{0.6}{$1$}} (MD);
%
\draw[->] (lMK) to node[above, xshift={-1pt}] {} (MK);
\draw[->] (lMR) to node[above, xshift={-1pt}] {\scalebox{0.7}{$\begin{pmatrix}0 \\ 1 \\ 1 \end{pmatrix}$}} (ML);
\draw[->] (lMe) to node[above, xshift={-1pt}] {\scalebox{0.7}{$\nabla$}} (Me);
%
\draw[->>] (MK) to (rMK);
\draw[->>] (MR) to (rML);
\draw[->>] (Me) to (rMe);
\end{tikzpicture}
\end{flushright}
So \(\ul \pi_2(X) = \ul f\) and \(\ul \pi_1(X) = \underline{g}\). From this we see that \(X\geq 2\) by \cref{2slice:characterization}. Note that \(\Si{2} H\ul f\) is a 2-slice and \(\Si{1} H\underline{g}\) is a 4-slice.
To show that \(X\leq 2\), we need \([S^{k\rho_H+r}, \Si{1+\beta} H\ul E]^H=0\) for all \(k\geq \frac{3}{\abs{H}}\) and \(r\geq 0\). As \(i_H^*(X) \simeq \Si{2} H_{C_2} \ul f\) is a 2-slice, where \(H\) is an order two subgroup of \(K\), we only need to consider
\begin{align*}
[S^{k\rho_K+r}, \Si{1+\beta} H\ul E]^K = [S^{k+r-1}, \Si{-\alpha-\gamma-(k-1) \overline{\rho}_K} H\ul E]^K
\end{align*}
for all \(k\geq 1\) and \(r\geq 0\).
By \cref{EM:homotopy:reduction}, it is sufficient to examine
\([S^0, \Si{-\alpha-\gamma-(k-1)\overline{\rho}_K} H\ul E]^K\)
for all \(k\geq 1\).
From \cref{homotopy:-beta:M} we find that \([S^0, \Si{-\alpha-\gamma} H\ul E] = \ul 0\). Consequently, given any \(k\geq 1\), repeated application of \cref{Homotopy_Reduction} shows that
\begin{align*}
[S^0, \Si{-\alpha-\gamma-(k-1)\overline{\rho}_K} H\ul E] = \ul 0.
\end{align*}
That is, \(X\) is a 2-slice.
\end{proof}
\section{Cotowers for \texorpdfstring{\(\Si{-n} H\underline{\Z}\)}{HZ}} \label{sec:cotowers:K:Z}
We determine the slice towers of \(\Si{-n} H\underline{\Z}\) and \(\Si{-n} H\ul m^*\) for \(1\leq n\leq 5\).
\begin{exmp} \label{slices:K:m*:1-2}
By \cref{charac:slice:01-1} and \cref{m:mg:equiv}, \(\Si{-1} H\ul m^*\) is a \((-2)\)-slice and \(\Si{-2} H\ul m^*\) is a \((-4)\)-slice.
Alternatively, by \cref{m:mg:equiv}, \(\Si{\rho}\Si{-1} H\ul m^* \simeq \Si{1} H\ul {mg}\), which is a 2-slice by \cref{2slice:characterization:si1}. Consequently, \(\Si{-1} H\ul m^*\) is a \((-2)\)-slice.
\end{exmp}
\begin{exmp} \label{tower:-1:K:Z}
By \cite[Theorem 6-4]{U}, the cotower for \(\Si{-1} H\underline{\Z}\) is
\begin{equation*}
\begin{tikzcd}
\Si{-1} H\underline{\Z}^* \arrow[d] = P^{-1}_{-1} & \\
\Si{-1} H\underline{\Z}(2,1) \arrow[d] \arrow[r] & \Si{-1} H\ul m^* = P^{-2}_{-2} \\
\Si{-1} H\underline{\Z} \arrow[r] & \Si{-1} H\underline{g} = P^{-4}_{-4}
\end{tikzcd}
\end{equation*}
\end{exmp}
\begin{exmp} \label{tower:-2:K:Z}
We suspend the cotower in \cref{tower:-1:K:Z} by \(-1\) to get the cotower for \(\Si{-2} H\underline{\Z}\).
\begin{equation*}
\begin{tikzcd}
\Si{-2} H\underline{\Z}^* \arrow[d] = P^{-2}_{-2} & \\
\Si{-2} H\underline{\Z}(2,1) \arrow[d] \arrow[r] & \Si{-2} H\ul m^* = P^{-4}_{-4} \\
\Si{-2} H\underline{\Z} \arrow[r] & \Si{-2} H\underline{g} = P^{-8}_{-8}
\end{tikzcd}
\end{equation*}
\end{exmp}
\begin{exmp} \label{tower:-3:K:Z}
We suspend the cotower in \cref{tower:-2:K:Z} by \(-1\) to get a partial cotower for \(\Si{-3} H\underline{\Z}\).
\begin{equation*}
\begin{tikzcd}
\Si{-3} H\underline{\Z}^* \arrow[d] = P^{-3}_{-3} & \\
\Si{-3} H\underline{\Z}(2,1) \arrow[d] \arrow[r] & \Si{-3} H\ul m^* \\
\Si{-3} H\underline{\Z} \arrow[r] & \Si{-3} H\underline{g} = P^{-12}_{-12}
\end{tikzcd}
\end{equation*}
The issue here is that \(\Si{-3} H\ul m^*\) is not a slice.
However, the short exact sequence
\begin{align*}
\underline{g}^2 \rightarrow \phi^*_{LDR} (\underline{\F}^*) \rightarrow \ul m^*
\end{align*}
provides the tower
\begin{align*}
P^{-6}_{-6} = \Si{-3} H\phi^*_{LDR} (\underline{\F}^*) \rightarrow \Si{-3} H\ul m^* \rightarrow \Si{-2} H\underline{g}^2 = P^{-8}_{-8}.
\end{align*}
Consequently, the remaining slices of \(\Si{-3} H\underline{\Z}\) are \(P^{-6}_{-6} = \Si{-2p+1} H\phi^*_{LDR}( \underline{\F})\) and \(P^{-8}_{-8} = \Si{-2} H\underline{g}^2\).
\end{exmp}
\begin{exmp} \label{tower:-4:K:Z}
We suspend the cotower in \cref{tower:-3:K:Z} by \(-1\) to get a partial cotower for \(\Si{-4} H\underline{\Z}\).
\begin{equation*}
\begin{tikzcd}
\Si{-4} H\underline{\Z}^* \arrow[d] = P^{-4}_{-4} & \\
\Si{-4} H\underline{\Z}(2,1) \arrow[d] \arrow[r] & \Si{-4} H\ul m^* \\
\Si{-4} H\underline{\Z} \arrow[r] & \Si{-4} H\underline{g} = P^{-16}_{-16}
\end{tikzcd}
\end{equation*}
Again, \(\Si{-4} H\ul m^*\) is not a slice. But suspending the cotower for \(\Si{-3} H\ul m^*\) by \(-1\) provides the missing slices: \(P^{-8}_{-8} = \Si{-4} H\phi^*_{LDR}( \underline{\F}^*)\) and \(P^{-12}_{-12} = \Si{-3} H\underline{g}^2\).
\end{exmp}
\begin{exmp} \label{tower:-5:K:Z}
We suspend the cotower in \cref{tower:-4:K:Z} by \(-1\) and augment with the \(-\rho\) suspension of \cref{tower:-1:K:Z} to get a partial cotower for \(\Si{-5} H\underline{\Z}\).
\begin{equation*}
\begin{tikzcd}
\Si{-\rho-1} H\underline{\Z}^* \arrow[d] = P^{-5}_{-5} & \\
\Si{-\rho-1} H\underline{\Z}(2,1) \arrow[d] \arrow[r] & \Si{-\rho-1} H\ul m^* = P^{-6}_{-6} \\
\Si{-\rho-1} H\underline{\Z} \arrow[d] \arrow[r] & \Si{-2} H\underline{g} = P^{-8}_{-8} \\
\Si{-5} H\underline{\Z}(2,1) \arrow[d] \arrow[r] & \Si{-5} H\ul m^* \\
\Si{-5} H\underline{\Z} \arrow[r] & \Si{-5} H\underline{g} = P^{-20}_{-20}
\end{tikzcd}
\end{equation*}
This time, the cotower for \(\Si{-5} H\ul m^*\) is
\begin{equation*}
\begin{tikzcd}
\Si{-3\rho + 1} H\phi^*_{LDR}(\underline{\F}) = P^{-10}_{-10} \arrow[d] & \\
\Si{-5} H\phi^*_{LDR}(\underline{\F}^*) \arrow[d] \arrow[r] & \Si{-3} H\underline{g}^3 = P^{-12}_{-12} \\
\Si{-5} H\ul m^* \arrow[r] & \Si{-4} H\underline{g}^2 = P^{-16}_{-16}
\end{tikzcd}
\end{equation*}
The remaining slices of \(\Si{-5} H\underline{\Z}\) are then \(P^{-10}_{-10} = \Si{-3\rho+1} H\phi^*_{LDR}(\underline{\F})\), \newline
\(P^{-12}_{-12} = \Si{-3} H\underline{g}^3\), and \(P^{-16}_{-16} = \Si{-4} H\underline{g}^2\).
\end{exmp}
\begin{exmp} \label{tower:-7:K:Z}
The partial cotower for \(\Si{-7} H\underline{\Z}\) follows by suspending the partial cotower in \cref{tower:-5:K:Z} by \(-2\).
\begin{equation*}
\begin{tikzcd}
\Si{-\rho-3} H\underline{\Z}^* \arrow[d] = P^{-7}_{-7} & \\
\Si{-\rho-3} H\underline{\Z}(2,1) \arrow[d] \arrow[r] & \Si{-\rho-3} H\ul m^* \\
\Si{-\rho-3} H\underline{\Z} \arrow[d] \arrow[r] & \Si{-4} H\underline{g} = P^{-16}_{-16} \\
\Si{-7} H\underline{\Z}(2,1) \arrow[d] \arrow[r] & \Si{-7} H\ul m^* \\
\Si{-7} H\underline{\Z} \arrow[r] & \Si{-7} H\underline{g} = P^{-28}_{-28}
\end{tikzcd}
\end{equation*}
We now have a cotower for \(\Si{-\rho-3} H\ul m^*\).
\begin{align*}
P^{-10}_{-10} = \Si{-\rho-3} H\phi^*_{LDR} (\underline{\F}^*) \rightarrow \Si{-\rho-3} H\ul m^* \rightarrow \Si{-3} H\underline{g}^2 = P^{-12}_{-12},
\end{align*}
And a cotower for \(\Si{-7} H\ul m^*\).
\begin{equation*}
\begin{tikzcd}
\Si{-4\rho + 1} H\phi^*_{LDR}(\underline{\F}) = P^{-14}_{-14} \arrow[d] & \\
\Si{-\rho-5} H\phi^*_{LDR}(\underline{\F}^*) \arrow[d] \arrow[r] & \Si{-4} H\underline{g}^3 = P^{-16}_{-16} \\
\Si{-7} H\phi^*_{LDR}(\underline{\F}^*) \arrow[d] \arrow[r] & \Si{-5} H\underline{g}^3 = P^{-20}_{-20} \\
\Si{-7} H\ul m^* \arrow[r] & \Si{-6} H\underline{g}^2 = P^{-24}_{-24}
\end{tikzcd}
\end{equation*}
We now see interference from the cotower for \(\Si{-7} H\ul m^*\). Its \((-14)\)-slice appears below the \((-16)\)-slice in the partial cotower for \(\Si{-7} H\underline{\Z}\). Additionally, both of these (partial) cotowers have a \((-16)\)-slice.
\cref{-4kslices:K:Z} and \cref{-4k-2slices:K:Z} tell us that \(P^{-14}_{-14}(\Si{-7} H\underline{\Z})\) is indeed \(\Si{-4\rho + 1} H\phi^*_{LDR}(\underline{\F})\) and \(P^{-16}_{-16}(\Si{-7} H\underline{\Z})\) is \(\Si{-4} H\underline{g} \vee \Si{-4} H\underline{g}^3 \simeq \Si{-4} H\underline{g}^4\).
\end{exmp}
All partial cotowers for \(\Si{-n} H\underline{\Z}\) will follow this pattern of utilizing the homotopy equivalence \(\Si{-n} H\underline{\Z}^* \simeq \Si{-\rho-n+4} H\underline{\Z}\) to augment the bottom of the cotower for \(\Si{-n} H\underline{\Z}\) with the cotower for \(\Si{-\rho-n+4} H\underline{\Z}\).
\section{Slices of \texorpdfstring{\(\Si{-n} H_{K}\ul \mathbb{Z}\)}{HZ}} \label{sec:slices:K:Z}
Here we determine the slices of \(\Si{-n} H_{K}\underline{\Z}\).
\begin{prop} \label{slice:K:Z1-5:Z*1-4}
\(\Si{n} H\underline{\Z}\) is an \(n\)-slice for \(1\leq n\leq 5\) and \(\Si{-n} H\underline{\Z}^*\) is a \((-n)\)-slice for \(1\leq n\leq 4\).
\end{prop}
\begin{proof}
The \(K\)-spectra \(\Si{1} H\underline{\Z}\), \(\Si{1} H\underline{\Z}^*\), \(\Si{2} H\underline{\Z}\), and \(\Si{-1} H\underline{\Z}^*\) are 1-, 2-, and \((-1)\)-slices by \cref{charac:slice:01-1} and \cref{2slice:characterization}. The result then follows from \cref{equivalence:ZZ*} and the resulting equivalences
\begin{align*}
\Si{5} H \underline{\Z} \simeq \Si{\rho+1} H \ul \mathbb{Z}^*, \quad \Si{4} H \underline{\Z} \simeq \Si{\rho} H \underline{\Z}^*, \quad \text{ and }\quad \Si{3} H \ul\mathbb{Z} \simeq \Si{\rho-1} H\underline{\Z}^*.
\end{align*}
\end{proof}
\subsection{The \texorpdfstring{\((-n)\)}{(-n)}-slice} \label{sec:-nslice:K:Z}
We first establish a comparison of the slices of \(\Si{-n} H\underline{\Z}\) with those of \(\Si{-n+4} H\underline{\Z}\).
\begin{prop} \label{slice:reduction:ZZ^*}
Let \(n\geq 5\). Then
\begin{align*}
P^{-k}_{-k} ( \Si{-n} H\underline{\Z}) \simeq \Si{-\rho} P^{-k+4}_{-k+4} (\Si{-n+4} H\underline{\Z})
\end{align*}
for \(k\in[n,2n-1]\).
\end{prop}
\begin{proof}
By \cref{equivalence:ZZ*}, we have
\begin{align*}
\Si{-\rho} P^{-k+4}_{-k+4} (\Si{-n+4} H\underline{\Z}) \simeq \Si{\rho} \Si{-\rho} P^{-k}_{-k} (\Si{-n+4-\rho} H\underline{\Z}) \simeq P^{-k}_{-k} (\Si{-n} H\underline{\Z}^*).
\end{align*}
Thus, it is sufficient to compare the \((-k)\)-slices of \(\Si{-n} H\underline{\Z}\) and \(\Si{-n} H\underline{\Z}^*\).
Recall the injection \(\underline{\Z}^* \rightarrow \underline{\Z} \rightarrow \underline{\mathbb{M}}\) from \cref{sec:K:Mackey}. We wish to show that
\begin{align} \label{inj:ZZM:-j}
\Si{-n} H\underline{\Z}^* \rightarrow \Si{-n} H\underline{\Z} \xrightarrow{\mu}{} \Si{-n} \underline{\mathbb{M}}
\end{align}
induces an equivalence on slices strictly above level \(-2n\).
We take the Brown-Comenetz dual of \cref{inj:ZZM:-j} to find the fiber sequence
\begin{align*}
\Si{n} H\underline{\mathbb{M}} \xrightarrow{\iota}{} \Si{n} I_{\mathbb{Q}/\mathbb{Z}} H\underline{\Z} \rightarrow \Si{n} I_{\mathbb{Q}/\mathbb{Z}} H\underline{\Z}^*
\end{align*}
Now \(\Si{n} H\underline{\mathbb{M}}\) is \(n\)-connective, and as its underlying spectrum is contractible, when \(n\geq 1\), we know \(\Si{n} H\underline{\mathbb{M}} \geq 2n\) by \cref{thm:atleastn}. Then \cite[Lemma 4.28]{HHR} provides that \(\iota\) induces an equivalence on slices strictly below level \(2n\). It then follows from \cref{duality:prop} that \(\mu\) in \cref{inj:ZZM:-j} induces an equivalence on slices strictly above level \(-2n\).
\end{proof}
This is an analogous result to \cite[Proposition 5.3]{GY}. However, the injection \(\underline{\Z}^* \hookrightarrow \underline{\Z}\) allows us to simplify the argument.
\begin{prop} \label{-nslice:K:Z}
Let \(n\geq 1\) and take \(r\equiv n\pmod 4\) with \(1\leq r\leq 4\). Then the top slice of \(\Si{-n} H\underline{\Z}\) is
\begin{align*}
P^{-n}_{-n}( \Si{-n} H\underline{\Z}) = \Si{-\frac{n-r}{4}\rho -r} H\underline{\Z}^*
\end{align*}
\end{prop}
\begin{proof}
\cref{sec:cotowers:K:Z} gives the result for \(1\leq n\leq 4\), and \(n\geq 5\) follows from repeated application of \cref{slice:reduction:ZZ^*}.
\end{proof}
\subsection{The \texorpdfstring{\((-4k)\)}{(-4k)}-slices} \label{-4kslice:K:Z}
\phantom{a}
Here we determine the \((-4k)\)-slices of \(\Si{-n} H\underline{\Z}\).
\begin{prop} \label{-4kslices:K:Z}
For \(n\geq 1\),
\begin{align*}
P^{-4k}_{-4k} ( \Si{-n} H\underline{\Z} ) \simeq \left\lbrace \begin{array}{ll}
\Si{-k\rho} H\ul {mg} & 4k=n+2 \\
\Si{-k} H\underline{g}^{\frac{1}{2}(4k-n-1)} &
\begin{aligned}
4k\in [n+1,2n-2] \\ n\equiv 1\pmod 2
\end{aligned} \\
\Si{-k\rho} H(\underline{g}^{\frac{1}{2}(4k-n+2)-3} \oplus \phi^*_{LDR} \underline{\F} ) &
\begin{aligned}
4k\in[n+3,2n], \\ n\equiv 0\pmod 2
\end{aligned} \\
\Si{-k} H\underline{g}^{n-k+1} & 4k\in[2n+1,4n]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
We have the equivalence
\begin{align*}
P^{-4k}_{-4k}(\Si{-n} H\underline{\Z}) \simeq \Si{-k\rho} H\ul\pi_0( \Si{-n+k\rho} H\underline{\Z}).
\end{align*}
The restriction to each cyclic subgroup agrees with the slices found in \cref{tower:C2:Z:-n} and \cref{krho:fixedpts:K:Z} gives the fixed points. All that remains is to verify the result for \(4k\in[n+2,2n]\) with \(n\) even. For \mbox{\(4k\in[n+2,2n)\)}, this follows from \cref{tower:-1:K:Z}, \cref{tower:-2:K:Z}, \cref{tower:-3:K:Z}, \cref{tower:-4:K:Z}, and \cref{slice:reduction:ZZ^*}.
Now let \(4k=2n\). The transfer \(\tr{K}{L}\) fits in the cofiber sequence
\begin{equation*}
\begin{tikzcd}
(\Si{-n+k\rho} H\underline{\Z})^L \arrow[r, "\tr{K}{L}"] & (\Si{-n+k\rho} H\underline{\Z})^K \arrow[r] & (\Si{-n+k\rho+\beta} H\underline{\Z})^K.
\end{tikzcd}
\end{equation*}
This gives the exact sequence
\begin{equation} \label{transfer:fixedpts:proof:Z:-4k}
\begin{tikzcd}
\pi_n^L( \Si{k\rho} H\underline{\Z} ) \arrow[r, "\tr{K}{L}"] & \pi_n^K(\Si{k\rho} H\underline{\Z}) \arrow[r] & \pi_n^K (\Si{k\rho+\beta} H\underline{\Z}).
\end{tikzcd}
\end{equation}
We wish to show this transfer is trivial. By \cref{krho:fixedpts:K:Z} and \cref{krho+beta:fixedpoints:K:Z}, \cref{transfer:fixedpts:proof:Z:-4k} becomes
\begin{equation*}
\begin{tikzcd}
\mathbb{F}_2 \arrow[r, "\tr{K}{L}"] & \mathbb{F}_2^{\frac{1}{2}(n+2)} \arrow[r, two heads] & \mathbb{F}_2^{\frac{1}{2}(n+2)} \arrow[r] & 0 = \pi_{n-1}^L( \Si{k\rho} H\underline{\Z}).
\end{tikzcd}
\end{equation*}
Thus, \(\tr{K}{L}\) must be zero. A similar argument shows that \(\tr{K}{R}\) and \(\tr{K}{D}\) are trivial as well.
The restriction \(\res{K}{L}\) fits into the cofiber sequence
\begin{equation*}
\begin{tikzcd}
( \Si{-n+k\rho-\beta} H\underline{\Z})^K \arrow[r] & ( \Si{-n+k\rho} H\underline{\Z})^K \arrow[r, "\res{K}{L}"] & ( \Si{-n+k\rho} H\underline{\Z})^L
\end{tikzcd}
\end{equation*}
This gives the exact sequence
\begin{equation} \label{res:fixedpts:proof:Z:-4k}
\begin{tikzcd}
\pi_n^K( \Si{k\rho-\beta} H\underline{\Z} ) \arrow[r] & \pi_n^K(\Si{k\rho} H\underline{\Z}) \arrow[r, "\res{K}{L}"] & \pi_n^L(\Si{k\rho} H\underline{\Z}).
\end{tikzcd}
\end{equation}
We wish to show this restriction is surjective. By \cref{krho:fixedpts:K:Z} and \cref{krho-beta:fixedpoints:K:Z}, \cref{res:fixedpts:proof:Z:-4k} becomes
\begin{equation*}
\begin{tikzcd}
\pi_{n+1}^L( \Si{k\rho} H\underline{\Z} ) = 0 \arrow[r] & \mathbb{F}_2^{\frac{1}{2}(n)} \arrow[r, hook] & \mathbb{F}_2^{\frac{1}{2}(n)+1} \arrow[r, "\res{K}{L}"] & \mathbb{F}_2
\end{tikzcd}
\end{equation*}
Thus, \(\res{K}{L}\) must be surjective. A similar argument shows that \(\res{K}{R}\) and \(\res{K}{D}\) are surjective as well.
All that remains to be shown is that the restrictions have distinct kernels. Consider the cofiber sequence
\begin{equation*}
\begin{tikzcd}
\Si{-n+k\rho-\beta} H\underline{\Z} \arrow[r] & \Si{-n+k\rho} H\underline{\Z} \arrow[r] & K/L_+ \wedge \Si{-n+k\rho} H\underline{\Z}.
\end{tikzcd}
\end{equation*}
This results in the exact sequence
\begin{equation} \label{exactseq:-4kproof:K:Z}
\begin{tikzcd}
\ul\pi_{n} ( \Si{k\rho-\beta} H\underline{\Z} ) \arrow[r] & \ul\pi_{n} ( \Si{k\rho} H\underline{\Z}) \arrow[r] & \uparrow^K_L \downarrow^K_L \ul\pi_{n} (\Si{k\rho} H\underline{\Z} )
\end{tikzcd}
\end{equation}
which by \cref{ksigma+r:homotopy:C2:Z}, \cref{krho:fixedpts:K:Z}, and \cref{krho-beta:fixedpoints:K:Z} is
\begin{equation*}
\begin{tikzcd}[scale cd = 0.85,column sep = small]
&& \mathbb{F}_2^{\frac{n}{2}} \arrow[rrr, "\varphi"]
\arrow[d, black] \arrow[dr, black]
&&[+15pt]
& \mathbb{F}_2^{\frac{n}{2}+1} \arrow[rrr, "\res{K}{L}"]
\arrow[d, black] \arrow[dr, black] \arrow[dl, black]
&& [+15pt]
& \mathbb{F}_2
\arrow[dl, black, bend right = 10]
%
&& \\ [-5pt]
& 0
& \mathbb{F}_2
& \mathbb{F}_2 \arrow[r, "0 \text{ } 1 \text{ } 1"]
& \mathbb{F}_2
& \mathbb{F}_2
& \mathbb{F}_2 \arrow[r, "\Delta \text{ } 0 \text{ } 0"]
&
\mathbb{F}_2[K/L]
\arrow[ur, black, bend right = 10]
& 0
& 0
& \\ [-5pt]
&& 0 \arrow[rrr, "0"]
&&
& 0 \arrow[rrr, "0"]
&&
& 0
&&
\end{tikzcd}
\end{equation*}
Each restriction is surjective with kernel of rank \(\frac{n}{2}\). As
\begin{align*}
\dim \pi_{n}^K ( \Si{k\rho} H\underline{\Z}) = \frac{n}{2}+1,
\end{align*}
it is sufficient to show that the kernels are pairwise distinct.
Because the diagram on the left commutes, we find that
\begin{align*}
\text{im }\varphi \cap \ker(\res{K}{R}) = \{0\} \qquad \text{and} \qquad \text{im }\varphi \cap \ker(\res{K}{D}) = \{0\}.
\end{align*}
As \(\text{im }\varphi = \ker(\res{K}{L})\), we have that \(\ker(\res{K}{L})\) is distinct from \(\ker(\res{K}{R})\) and \(\ker(\res{K}{D})\). Replacing \(\beta\) by \(\alpha\) shows that \(\ker(\res{K}{R})\) and \(\ker(\res{K}{D})\) are distinct as well.
\end{proof}
\begin{prop} \label{krho:fixedpts:K:Z}
For \(G=K\) and \(k\geq 1\),
\begin{align*}
\pi_i^K( \Si{k\rho} H\underline{\Z}) &= \left\lbrace \begin{array}{ll}
\mathbb{Z} & i=4k \\
\mathbb{F}_2^{\frac{1}{2}(4k-1-i)} & i\in [2k+1, 4k-1] \text{ odd} \\
\mathbb{F}_2^{\frac{1}{2}(4k+2-i)} & i\in[2k+2, 4k-2] \text{ even} \\
\mathbb{F}_2^{i-k+1} & i\in[k,2k]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
Note that
\begin{align*}
P^{-4k}_{-4k}(\Si{-4k} H\underline{\Z}) &\simeq \Si{-k\rho} H\ul\pi_{4k}(\Si{k\rho} H\underline{\Z}).
\end{align*}
Thus, by \cref{-nslice:K:Z}, \(H\ul\pi_{4k}(\Si{k\rho} H\underline{\Z}) \simeq H\underline{\Z}\). So the result holds for \(i=4k\) and we only need consider \(i\leq 4k-1\).
We will use the resulting long exact sequences in homotopy resulting from the cofiber sequences
\begin{align}
(K/L_+ &\rightarrow S^0 \rightarrow S^\beta) \wedge \Si{k\rho+1} H\underline{\Z} \label{cofib:beta} \\
(K/R_+ &\rightarrow S^0 \rightarrow S^\alpha) \wedge \Si{k\rho+\beta+1} H\underline{\Z} \label{cofib:alpha} \\
(K/D_+ &\rightarrow S^0 \rightarrow S^\gamma) \wedge \Si{k\rho+\alpha+\beta+1} H\underline{\Z} \label{cofib:gamma} \\
\Si{k\rho} H\underline{\Z} &\xrightarrow{2}{} \Si{k\rho} H\underline{\Z} \rightarrow \Si{k\rho} H\underline{\F} \label{cofib:mult2}
\end{align}
where \cref{cofib:mult2} is induced by the short exact sequence of Mackey functors \(\underline{\Z} \xrightarrow{2}{} \underline{\Z} \rightarrow \underline{\F}\).
For \(k=0\), \cref{cofib:beta} - \cref{cofib:gamma} provide that
\begin{align*}
\ul\pi_n(\Si{\rho} H\underline{\Z}) &= \left\lbrace \begin{array}{cc}
\underline{\Z} & n=4 \\
\ul 0 & n=3 \\
\ul{mg} & n=2 \\
\underline{g} & n=1
\end{array}\right.
\end{align*}
Consequently, the result holds for \(k=1\). We now proceed by induction on \(k\). Assume the result holds for \(k\). By restriction to \(L\), \(D\), and \(R\), we find that \(\ul\pi_i(\Si{(k+1)\rho} H\underline{\Z})\) is a pullback over \(K\) for \(i\leq 4k+3\) odd, and consequently, 2-torsion as in \cite[Remark 2.13]{Z}.
In the long exact sequence of fixed points resulting from \cref{cofib:beta} - \cref{cofib:gamma}, we have at the \(K/K\) level,
\begin{equation*}
\begin{tikzcd}[scale cd = 0.9]
\pi_i^{C_2} (\Si{2k\rho_{C_2}+1} H_{C_2}\underline{\Z}) \arrow[r] &[-10pt] \pi_i^K( \Si{k\rho+1} H\underline{\Z}) \arrow[r] &[-10pt] \pi_{i}^K(\Si{k\rho+\beta + 1} H\underline{\Z}) \\
\pi_i^{C_2} (\Si{(2k+1)\rho_{C_2}} H_{C_2}\underline{\Z}) \arrow[r] & \pi_i^K( \Si{k\rho+\beta+1} H\underline{\Z}) \arrow[r] & \pi_{i}^K(\Si{k\rho+\alpha+\beta +1 } H\underline{\Z}) \\
\pi_i^{C_2} (\Si{(2k+2)\rho_{C_2}-1} H_{C_2}\underline{\Z}) \arrow[r] & \pi_i^K( \Si{k\rho+\alpha+\beta+ 1} H\underline{\Z}) \arrow[r] & \pi_{i}^K(\Si{(k+1)\rho} H\underline{\Z})
\end{tikzcd}
\end{equation*}
When \(2k+1\leq i \leq 4k+1\), by \cref{ksigma+r:homotopy:C2:Z},
\begin{align*}
\pi_i^{C_2} (\Si{2k\sigma + 2k+1} H_{C_2}\underline{\Z}) &= \left\lbrace \begin{array}{ll}
\mathbb{F}_2 & i \text{ odd} \\
0 & i \text{ even}
\end{array}\right.
\end{align*}
Now for \(i\) even, we have in our long exact sequence,
\begin{equation*}
\begin{tikzcd}
0 \ar[r] & \pi_i^K( \Si{k\rho+1} H\underline{\Z}) \ar[r, hook] & \pi_{i}^K(\Si{k\rho+\beta + 1} H\underline{\Z}) \ar[r] & \phantom{a} \\
\mathbb{F}_2 \ar[r] & \pi_{i-1}^K( \Si{k\rho+1} H\underline{\Z}) \ar[r, twoheadrightarrow] & \pi_{i-1}^K(\Si{k\rho+\beta + 1} H\underline{\Z}) \ar[r, "0", twoheadrightarrow] & \phantom{a}
\end{tikzcd}
\end{equation*}
Consequently, when \(i\) is even,
\begin{align*}
\text{2-rk } \pi_{i}^K(\Si{k\rho+\beta + 1} H\underline{\Z}) \leq \text{2-rk } \pi_i^K( \Si{k\rho+1} H\underline{\Z}) + 1,
\end{align*}
where equality occurs if \(\pi_i^K( \Si{k\rho+\beta + 1} H\underline{\Z}) = \pi_i^K( \Si{(k\rho+1} H\underline{\Z}) \oplus \mathbb{F}_2\).
And, when \(i\) is odd,
\begin{align*}
\text{2-rk } \pi_{i}^K(\Si{k\rho+\beta + 1} H\underline{\Z}) \leq \text{2-rk } \pi_i^K( \Si{k\rho+1} H\underline{\Z}).
\end{align*}
For \(i\leq 2k\) we have
\begin{equation*}
\begin{tikzcd}
0 \ar[r] & \pi_{i}^K( \Si{k\rho+1} H\underline{\Z}) \ar[r, hook, two heads] & \pi_{i}^K(\Si{k\rho+\beta + 1} H\underline{\Z}) \ar[r, "0", twoheadrightarrow] & \phantom{a}
\end{tikzcd}
\end{equation*}
Thus, for \(i\leq 2k\),
\begin{align*}
\pi_{i}^K(\Si{k\rho+\beta + 1} H\underline{\Z}) \cong \pi_{i}^K( \Si{k\rho+1} H\underline{\Z}).
\end{align*}
A similar statement holds for
\begin{align*}
\text{2-rk } \pi_{i}^K(\Si{ k\rho+\alpha+\beta +1} H\underline{\Z}) \qquad \text{ and } \qquad \text{2-rk } \pi_{i}^K(\Si{(k+1)\rho} H\underline{\Z}).
\end{align*}
From \cite[Corollary 7.2]{GY}, in the long exact sequence resulting from \cref{cofib:mult2}, we have
\begin{align*}
\mathbb{F}_2^l \cong \pi_{2i+1}^K(\Si{(i+1)\rho} H\underline{\Z}) &\xrightarrow{0} \pi_{2i+1}^K(\Si{(i+1)\rho} H\underline{\Z}) \hookrightarrow \pi_{2i+1}^K(\Si{(i+1)\rho} H\underline{\F}) \cong \mathbb{F}_2^{2i+1} \\
&\twoheadrightarrow \pi_{2i}^K(\Si{(i+1)\rho} H\underline{\Z}) \cong \pi_{2i-1}^K(\Si{i\rho} H\underline{\Z}) \cong \mathbb{F}_2^{i}.
\end{align*}
Consequently, \(l=i+1\) and \(\pi_{2i+1}^K(\Si{(i+1)\rho} H\underline{\Z})\cong \mathbb{F}_2^{i+1}\). We then have in our long exact sequence
\begin{align*}
\pi_{2i+2}^K(\Si{(i+1)\rho} H\underline{\Z}) \rightarrow \pi_{2i+2}^K(\Si{(i+1)\rho} H\underline{\F}) \cong \mathbb{F}_2^{2i+3} \twoheadrightarrow \mathbb{F}_2^{i+1} \cong \pi_{2i+1}^K(\Si{(i+1)\rho} H\underline{\Z}).
\end{align*}
Thus,
\begin{align*}
\text{2-rk } \pi_{2i+2}^K(\Si{(i+1)\rho} H\underline{\Z}) \geq i+2 = \text{2-rk } \pi_{2i+1}^K(\Si{i\rho} H\underline{\Z}) + 3.
\end{align*}
We achieve the maximum bound for \(\text{2-rk } \pi_{2i+2}^K(\Si{(i+1)\rho} H\underline{\Z})\); thus,
\begin{align*}
\pi_{2i+2}^K(\Si{(i+1)\rho} H\underline{\Z}) \cong \pi_{2i+1}^K(\Si{i\rho} H\underline{\Z}) \oplus \mathbb{F}_2^3 \cong \mathbb{F}_2^{i+2}.
\end{align*}
The rest of the result now follows from this long exact sequence in a similar manner.
\end{proof}
\begin{prop} \label{krho-beta:fixedpoints:K:Z}
For \(G=K\) and \(k\geq 1\),
\begin{align*}
\pi_i^K( \Si{k\rho-\beta} H\underline{\Z}) &= \left\lbrace \begin{array}{ll}
\mathbb{F}_2^{\frac{1}{2}(4k-i)} & i\in[2k, 4k-2] \text{ even} \\
\mathbb{F}_2^{\frac{1}{2}(4k-i-1)} & i\in[2k-1, 4k-3] \text{ odd} \\
\mathbb{F}_2^{i-k+1} & i\in[k,2k]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
This follows from a similar argument as in \cref{krho:fixedpts:K:Z}.
\end{proof}
\begin{prop} \label{krho+beta:fixedpoints:K:Z}
For \(G=K\) and \(k\geq 1\),
\begin{align*}
\pi_i^K( \Si{k\rho+\beta} H\underline{\Z}) &= \left\lbrace \begin{array}{ll}
\mathbb{F}_2^{\frac{1}{2}(4k-i)+1} & i\in[2k+2, 4k] \text{ even} \\
\mathbb{F}_2^{\frac{1}{2}(4k-i+1)} & i\in[2k+1, 4k-1] \text{ odd} \\
\mathbb{F}_2^{i-k+1} & i\in[k,2k]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
This follows from a similar argument as in \cref{krho:fixedpts:K:Z}.
\end{proof}
\subsection{The \texorpdfstring{\((-4k-2)\)}{(-4k-2)}-slices} \label{-4k-2slice:K:Z}
\phantom{a}
We now determine the \((-4k-2)\)-slices of \(\Si{-n} H\underline{\Z}\).
\begin{prop} \label{-4k-2slices:K:Z}
For even \(n\), \(\Si{-n} H\underline{\Z}\) has no \((-4k-2)\)-slices, except for possibly the \(-n\) slice.
For odd \(n\geq 1\),
\begin{align*}
P^{-4k-2}_{-4k-2} ( \Si{-n} H\underline{\Z} ) &\simeq \left\lbrace
\begin{array}{ll}
\Si{-k\rho-1} H\ul m^* & 4k+2=n+1 \\
\Si{-k\rho-1} H\phi^*_{LDR} \underline{\F}^* & 4k+2\in (n+1,2n]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
When \(n\) is even, the slices in \cref{-4kslices:K:Z} restrict to the all of the appropriate slices for \(\Si{-n} H_{C_2}\underline{\Z}\). If, then, \(\Si{-n} H_K\underline{\Z}\) has a \((-4k-2)\)-slice, it must be a pullback over \(K\). But this is a contradiction as such slices are \((-4k)\)-slices. Thus, for \(n\) even, \(\Si{-n} H_K\underline{\Z}\) has no \((-4k-2)\)-slices.
Now let \(n\) be odd. We first handle the case \(4k+2=2n\). By \cref{rho:susp:commutes},
\begin{align*}
P^{-4k-2}_{-4k-2}(\Si{-n} H_K\underline{\Z}) \simeq \Si{-(k+1)\rho} P^2_2( \Si{-n+(k+1)\rho} H_K\underline{\Z}).
\end{align*}
For clarity, let
\begin{align*}
X := P^2_2( \Si{-n+(k+1)\rho} H_K\underline{\Z}).
\end{align*}
Now from \cref{-2kslice:C2:HZ},
\begin{align*}
i_{H}^* X \simeq \Si{1} H_{C_2}\underline{g}.
\end{align*}
where \(H\) is \(L\), \(D\), or \(R\). Thus, by \cref{2slice:characterization},
\begin{align*}
\ul\pi_2( X) = \phi^*_{K} B \qquad \text{ and } \qquad \ul\pi_1( X ) = \ul A
\end{align*}
where \(B\) is some group and
\begin{equation*}
\ul A = \begin{tikzcd}[ampersand replacement=\&, column sep=scriptsize, bend right=15]
\& A_K \ar[dl] \ar[d] \ar[dr] \& \\
\mathbb{F}_2 \ar[ur] \& \mathbb{F}_2 \ar[u] \& \mathbb{F}_2 \ar[ul] \\
\& 0 \&
\end{tikzcd}
\end{equation*}
and \(A_K \rightarrow \mathbb{F}_2\oplus \mathbb{F}_2\oplus \mathbb{F}_2\) is injective. That is, \(A_K=\mathbb{F}_2^n\) with \(0\leq n\leq 3\).
If \(B\neq 0\), then \(X\) cannot be a 2-slice. Consequently, \(X\simeq \Si{1} H\ul A\).
Because \(\underline{\Z}\) is invariant under the automorphisms of \(K\), the spectrum \(\Si{-n} H\underline{\Z}\) is as well. Therefore, the slices of \(\Si{-n} H\underline{\Z}\) are also invariant under the automorphisms of \(K\).
Thus, \(\ul A\) must be one of the following:
\begin{align*}
\phi^*_{LDR} \ul f, \quad \ul m, \quad \ul{mg}, \quad \text{ or } \quad \phi^*_{LDR} \underline{\F}.
\end{align*}
Except for degree \(-k-1\), \(\Si{-(k+1)\rho +1} H\ul A\) has the same homotopy for each choice of \(A\).
For degree \(-k-1\), \(\ul\pi_{-k-1}(\Si{-(k+1)\rho +1} H\ul A)\) is a pullback over \(K\) of dimension 3, 2, 1, or 0.
From \cref{phiF:SSS} we find that \(\pi_{-k-1}^K(\Si{-(k+1)\rho +1} H\ul A)=0\). The only choice of \(\ul A\) that meets this requirement is \(\phi^*_{LDR}\underline{\F}\).
Now let \(4k+2\in[n+1,2n-1]\). The base cases are established in \cref{tower:-1:K:Z}, \cref{tower:-3:K:Z}, and \cref{tower:-5:K:Z}. The result then follows from \cref{slice:reduction:ZZ^*}.
\end{proof}
\section{Slices of \texorpdfstring{\(\Si{n} H\ul \mathbb{Z}\)}{HZ}} \label{sec:slices:n:K:Z}
Recall from \cref{slice:K:Z1-5:Z*1-4} that \(\Si{n} H\underline{\Z}\) is a slice for \(1\leq n\leq 5\).
\begin{prop} \label{slice:reduction:K:n:Z}
Let \(n\geq 6\). Then
\begin{align*}
P^k_k(\Si{n} H\underline{\Z}) &\simeq \Si{\rho} P^{k-4}_{k-4} (\Si{n-4} H\underline{\Z})
\end{align*}
for \(k\in[n,2n-7]\).
\end{prop}
\begin{proof}
We employ a similar argument as in \cref{slice:reduction:ZZ^*}. Note that
\begin{align*}
P^k_k(\Si{n} H\underline{\Z}) &\simeq \Si{\rho} P^{k-4}_{k-4}(\Si{n-\rho} H\underline{\Z}) \simeq \Si{\rho} P^{k-4}_{k-4}(\Si{n-4} H\underline{\Z}^*).
\end{align*}
Consequently, it is sufficient to compare the \((k-4)\)-slices of \(\Si{n-4} H\underline{\Z}\) and \(\Si{n-4} H\underline{\Z}^*\).
The exact sequence \(\underline{\Z}^* \rightarrow \underline{\Z} \rightarrow \underline{\mathbb{M}}\) provides the fiber sequence
\begin{align*}
\Si{j-1} H\underline{\mathbb{M}} \rightarrow \Si{j} H\underline{\Z}^* \xrightarrow{\iota}{} \Si{j} H\underline{\Z}.
\end{align*}
Then, because \(\Si{j-1} H\underline{\mathbb{M}}\geq 2j-2\), by \cite[Lemma 4.28]{HHR}, \(\iota\) induces an equivalence on slices strictly below level \(2j-2\). Taking \(j=n-4\) gives the result.
\end{proof}
\begin{exmp} \label{tower:6:K:Z}
The tower for \(\Si{6} H\underline{\Z}\) is
\begin{equation*}
\begin{tikzcd}
P^8_8 = \Si{2} H\underline{g} \arrow[r] & \Si{6} H\underline{\Z} \arrow[d] \\
& P^6_6 = \Si{\rho+2} H\underline{\Z}(2,1)^*
\end{tikzcd}
\end{equation*}
\end{exmp}
\begin{proof}
We have the short exact sequence \(\underline{\Z}^* \rightarrow \underline{\Z}(2,1)^* \rightarrow \underline{g}\). This leads to the cofiber sequence
\begin{equation*}
\begin{tikzcd}
P^8_8 = \Si{2} H\underline{g} \arrow[r] & \Si{6} H\underline{\Z} \simeq \Si{\rho+2} H\underline{\Z}^* \arrow[r] & \Si{\rho+2} H\underline{\Z}(2,1)^* = P^6_6.
\end{tikzcd}
\end{equation*}
\end{proof}
\begin{exmp} \label{tower:7:K:Z}
The tower for \(\Si{7} H\underline{\Z}\) is
\begin{equation*}
\begin{tikzcd}
P^{12}_{12} = \Si{3} H\underline{g} \arrow[r] & \Si{7} H\underline{\Z} \arrow[d] \\
P^{8}_8 = \Si{\rho+2} H\ul m \arrow[r] & \Si{\rho+3} H\underline{\Z}(2,1)^* \arrow[d] \\
& P^7_7 = \Si{\rho+3} H\underline{\Z}
\end{tikzcd}
\end{equation*}
\end{exmp}
\begin{proof}
We suspend the tower in \cref{tower:6:K:Z} by 1 and augment with the cofiber sequence \(\Si{\rho+2} H\ul m \rightarrow \Si{\rho+3} H\underline{\Z}(2,1)^* \rightarrow \Si{\rho+3} H\underline{\Z}\) which arises from the short exact sequence \(\underline{\Z}(2,1)^* \rightarrow \underline{\Z} \rightarrow \ul m\).
\end{proof}
We now determine the slices of \(\Si{n} H\underline{\Z}\).
\begin{thm} \label{duality:slices:K:Z}
Let \(n\geq 6\). For \(k\geq n+2\),
\begin{align*}
P^k_k (\Si{n} H\underline{\Z}) &\simeq \Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-k+4}_{-k+4} \, \Si{-n+5} H\underline{\Z}.
\end{align*}
\end{thm}
\begin{proof}
Take \(r\equiv n-5\pmod 4\) with \(1\leq r\leq 4\). We may map the top slice of \(\Si{-n+5} H\underline{\Z}\) into it to find the cofiber sequence
\begin{align}
P^{-n+5}_{-n+5} = \Si{-\frac{n-5-r}{4}\rho -r} H\underline{\Z}^* \rightarrow \Si{-n+5} H\underline{\Z} \rightarrow P^{-n+5-1} \, \Si{-n+5} H\underline{\Z}. \label{slices:cofiber}
\end{align}
Note that all slices of \(P^{-n+4} H\underline{\Z}\) are torsion, so then
\begin{align*}
I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+4} \, H\underline{\Z} = \Si{1} I_{\mathbb{Z}} \, P^{-n+4} \, \Si{-n+5} H\underline{\Z}.
\end{align*}
Apply \(I_\mathbb{Z}\) to \cref{slices:cofiber} and suspend by one to find
\begin{align*}
\Si{1} I_{\mathbb{Z}} \, P^{-n+4} \, \Si{-n+5} H\underline{\Z} \rightarrow \Si{n-4} H\underline{\Z}^* \rightarrow \Si{\frac{n-5-r}{4}\rho +r+1} H\underline{\Z}.
\end{align*}
We can rewrite this as
\begin{align*}
I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+4} \, \Si{-n+5} H\underline{\Z} \rightarrow \Si{n-\rho} H\underline{\Z} \rightarrow \Si{\frac{n-5-r}{4}\rho +r+1} H\underline{\Z}.
\end{align*}
Finally, suspend by \(\rho\) to obtain
\begin{align}
\Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+4} \,\Si{-n+5} H\underline{\Z} \rightarrow \Si{n} H\underline{\Z} \rightarrow \Si{\frac{n-1-r}{4}\rho +r+1} H\underline{\Z}. \label{slices:cofiber:1}
\end{align}
Note that \(\Si{\frac{n-1-r}{4}\rho +r+1} H\underline{\Z}\) is an \(n\)-slice and
\begin{align*}
\Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+4} \, \Si{-n+5} H\underline{\Z}\in [n,4(n-4)].
\end{align*}
Furthermore, if \(n\not\equiv 2\pmod 4\),
\begin{align*}
\Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+4} \, \Si{-n+5} H\underline{\Z}\in [n+1,4(n-4)].
\end{align*}
From \cref{duality:prop},
\begin{align*}
I_{\mathbb{Q}/\mathbb{Z}} \, P^{-k}_{-k} \, \Si{-n+5} H\underline{\Z} &\simeq P^{k}_{k} I_{\mathbb{Q}/\mathbb{Z}} \, \Si{-n+5} H\underline{\Z}.
\end{align*}
Consequently, \(\Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+4} \,\Si{-n+5} H\underline{\Z}\) provides all slices of \(\Si{n} H\underline{\Z}\).
Now suppose \(n\equiv 2\pmod 4\) so that \(r=1\). Then from \cref{duality:prop},
\begin{align*}
P^n_n ( \Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+4} \, \Si{-n+5} H\underline{\Z} ) &\simeq \Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n}_{-n} ( P^{-n+4} \, \Si{-n+5} H\underline{\Z} ) \\
&\simeq \Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, \Si{-(\frac{n-6}{4}+1)\rho+1} H\ul {mg} \\
&\simeq \Si{\frac{n-2}{4}\rho + 1} H\ul m
\end{align*}
Apply \(P^n_n(-)\) to \cref{slices:cofiber:1} to get the extension
\begin{align*}
\Si{\frac{n-2}{4}\rho + 1} H\ul m \rightarrow P^n_n \Si{n} H\underline{\Z} \rightarrow \Si{\frac{n-2}{4}\rho +2} H\underline{\Z}
\end{align*}
and the fiber sequence
\begin{align*}
\Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+5} H\underline{\Z} \rightarrow \Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+4} H\underline{\Z} \rightarrow P^n_n \Si{n} H\underline{\Z}.
\end{align*}
Now \(\Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+5} H\underline{\Z} \in [n+1,4(n-4)]\) and thus supplies the remaining slices of \(\Si{n} H\underline{\Z}\).
\end{proof}
\begin{prop} \label{slices:K:n:Z}
Let \(n\geq 6\) and set \(r\equiv n\pmod 4\) with \(2\leq r\leq 5\). The \(n\)-slice of \(\Si{n} H\underline{\Z}\) is
\begin{align*}
P^n_n \Si{n} H\underline{\Z} &\simeq \left\lbrace \begin{array}{ll}
\Si{\frac{n-2}{4}\rho + 2} H\underline{\Z}(2,1)^* & n\equiv 2\pmod 4 \\
\Si{\frac{n-r}{4}\rho + r} H\underline{\Z} & \text{otherwise}
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
For \(n\not\equiv 2\pmod 4\), this follows from \cref{slices:cofiber:1}. When \(n\equiv 2\pmod 4\) it follows from \cref{tower:6:K:Z} and repeated application of \cref{slice:reduction:K:n:Z}.
\end{proof}
\begin{prop} \label{slices:K:4k:Z}
Let \(n\geq 6\). The \(4k\)-slices of \(\Si{n} H\underline{\Z}\) are
\begin{align*}
P^{4k}_{4k} ( \Si{n} H\underline{\Z} ) \simeq \left\lbrace \begin{array}{ll}
\Si{k\rho} H\ul{mg}^* & 4k=n+1 \\
\Si{k} H\underline{g}^{\frac{1}{2}(4k-n)} & \begin{aligned}
4k\in [n+2,2n-8], \\ n\equiv 0\pmod 2
\end{aligned} \\
\Si{k\rho} H(\underline{g}^{\frac{1}{2}(4k-n+3)-3} \oplus \phi^*_{LDR} \underline{\F}^* ) &
\begin{aligned} 4k\in[n+2,2n-6], \\ n\equiv 1\pmod 2
\end{aligned} \\
\Si{k} H\underline{g}^{n-k-3} & 4k\in[2n-4,4(n-4)]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
This follows from \cref{-4kslices:K:Z} and \cref{duality:slices:K:Z}.
\end{proof}
\begin{prop} \label{slices:4k+2:K:Z}
For \(n\geq 6\), except for possibly the \(n\) slice,
the nontrivial \((4k+2)\)-slices of \(\Si{n} H\underline{\Z}\) are
\begin{align*}
P^{4k+2}_{4k+2} (\Si{n} H\underline{\Z}) &\simeq \left\lbrace \begin{array}{ll}
\Si{k\rho+1} H\phi^*_{LDR} \underline{\F} &
\begin{aligned} 4k+2\in[n+2,2n-6], \\ n\equiv 0\pmod 2
\end{aligned}
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
By \cref{2kslice:C2:HZ}, \(\Si{n} H_{C_2}\underline{\Z}\) has no \((4k+2)\)-slices except in the range \([n+2,2n-6]\) when \(n\) is even. So for \(4k+2\) not in this range, any such slice must be a pullback over \(K\). But then it is a \(4k\)-slice. For \(4k+2\in[n+2,2n-6]\), the result follows from \cref{-4k-2slices:K:Z} and \cref{duality:slices:K:Z}.
\end{proof}
\subsection{Comparison with the Slices of \texorpdfstring{\(\Si{n} H\underline{\F}\)}{HF}} \label{subsec:compGY}
\phantom{a}
This work is complementary to \cite{GY}, which calculates the slices of \(\Si{n} H\underline{\F}\) for \(n\geq 1\). One would hope that the exact sequence \(\underline{\Z} \xrightarrow{2}{} \underline{\Z} \rightarrow \underline{\F}\) could play a role in recovering the slices of \(\Si{n} H\underline{\Z}\) from the slices of \(\Si{n} H\underline{\F}\) or vice versa, but this is not always the case.
When \(G=C_2\), \cite[Theorem 3.18]{GY} shows that the slices of \(\Si{n} H_{C_2}\underline{\F}\) contain both even and odd suspensions of \(H_{C_2}\underline{g}\), whereas \cref{tower:C2:Z:n} shows that \(\Si{n} H_{C_2}\underline{\Z}\) has only even or odd suspensions of \(H_{C_2}\underline{g}\). This is illustrated in \cref{tab-C2slices}.
\begin{table}[h]
\caption[Comparison of {$C_2$}-slices]{Comparison of $C_2$-slices}
\label{tab-C2slices}
\begin{center}
{\renewcommand{\arraystretch}{1.5}\setlength{\tabcolsep}{15pt}
\begin{tabular}{|c|c|c|}
\hline
Slices of \(\Si{9} H_{C_2}\underline{\Z}\) & Slices of \(\Si{9} H_{C_2}\underline{\F}\) & Slices of \(\Si{10} H_{C_2}\underline{\Z}\) \\ \hline
&\(P^{14}_{14} = \Si{7} H_{C_2}\underline{g}\) & \(P^{14}_{14} = \Si{7} H_{C_2}\underline{g}\) \\
\(P^{12}_{12} = \Si{6} H_{C_2}\underline{g}\) &\(P^{12}_{12} = \Si{6} H_{C_2}\underline{g}\) & \\
&\(P^{10}_{10} = \Si{5} H_{C_2}\underline{g}\) & \(P^{10}_{10} = \Si{2\rho+2} H_{C_2}\underline{\Z}^*\) \\
\(P^{9}_{9} = \Si{2\rho+1} H_{C_2}\underline{\Z}^*\) & \(P^{9}_{9} = \Si{2\rho+1} H_{C_2}\underline{\F}^*\) & \\ \hline
\end{tabular} }
\end{center}
\end{table}
The \(2k\)-slices of \(\Si{n} H_{C_2}\underline{\Z}\) and \(\Si{n+1} H_{C_2}\underline{\Z}\) only combine to give the slices of \(\Si{n} H_{C_2}\underline{\F}\) when \(n\equiv 3,4\pmod 4\). When \(n\equiv 5,6\pmod 4\), the \(\Si{n} H_{C_2}\underline{\Z}\) and \(\Si{n+1} H_{C_2}\underline{\Z}\) slices miss the \((n+r)\)-slice of \(\Si{n} H_{C_2}\underline{\F}\), where \(r=1,2\), respectively. For example, neither \(\Si{9} H_{C_2}\underline{\Z}\) nor \(\Si{10} H_{C_2}\underline{\Z}\) has a slice equivalent to \(\Si{5} H_{C_2}\underline{g}\), but \(\Si{9} H_{C_2}\underline{\F}\) does.
We can recover the \((4k)\)-slices of \(\Si{n} H\underline{\F}\) from the \((4k)\)-slices of \(\Si{n} H_K\underline{\Z}\) and \(\Si{n+1} H_K\underline{\Z}\). As in \cref{krho:fixedpts:K:Z}, we use the sequence \(\underline{\Z} \xrightarrow{2}{} \underline{\Z} \rightarrow \underline{\F}\) to get the cofiber sequence \(\Si{-k\rho} H_K\underline{\Z} \rightarrow \Si{-k\rho} H_K\underline{\F} \rightarrow \Si{1-k\rho} H_K\underline{\Z}\). We then have the long exact sequence in homotopy
\begin{equation} \label{les:ZZF:4kslices}
\begin{tikzcd}[column sep=small, row sep=small]
\ul\pi_{-n}( \Si{-k\rho} H_K\underline{\Z}) \arrow[r, "2"] &
\ul\pi_{-n}( \Si{-k\rho} H_K\underline{\Z}) \arrow[r] & \phantom{\ul\pi_{-n}( \Si{-k\rho} H_K\underline{\Z})} \\
\ul\pi_{-n}( \Si{-k\rho} H_K\underline{\F}) \arrow[r] &
\ul\pi_{-n-1}( \Si{-k\rho} H_K\underline{\Z}) \arrow[r, "2"] &
\ul\pi_{-n-1}( \Si{-k\rho} H_K\underline{\Z}).
\end{tikzcd}
\end{equation}
When \(n\leq 4k-1\), all groups in \cref{les:ZZF:4kslices} are 2-torsion and the middle three terms become the exact sequence
\begin{equation} \label{ses:ZZF:4kslices}
\begin{tikzcd}[row sep=small]
\ul\pi_{-n}( \Si{-k\rho} H_K\underline{\Z}) \arrow[r, hook] &
\ul\pi_{-n}( \Si{-k\rho} H_K\underline{\F}) \arrow[r, twoheadrightarrow] &
\ul\pi_{-n-1}( \Si{-k\rho} H_K\underline{\Z}).
\end{tikzcd}
\end{equation}
When \(n=4k\), the left four terms of \cref{les:ZZF:4kslices} become the exact sequence
\begin{equation*}
\begin{tikzcd}[row sep=small]
\underline{\Z}^* \arrow[r, "2"] &
\underline{\Z}^* \arrow[r, twoheadrightarrow, "2"] &
\ul\pi_{-n-1}( \Si{-k\rho} H_K\underline{\Z}) \arrow[r] & \ul 0.
\end{tikzcd}
\end{equation*}
Consequently, the \((4k)\)-slices of \(\Si{n} H_K\underline{\F}\) are
\begin{align*}
P^{4k}_{4k}( \Si{n} H_K\underline{\F}) &\simeq \left\lbrace
\begin{array}{ll}
\Si{k\rho} H_K\underline{\F}^* & n=4k \\
\Si{k\rho} \ul E_{-n} & n\leq 4k-1
\end{array}\right.
\end{align*}
where \(\ul E_{-n}\) is the middle Mackey functor in \cref{ses:ZZF:4kslices}. This recovery is illustrated in \cref{tab-Kslices}.
\begin{table}[h]
\caption[Comparison of {$K$}-slices]{Comparison of $K$-slices}
\label{tab-Kslices}
\begin{center}
{\renewcommand{\arraystretch}{1.5}\setlength{\tabcolsep}{15pt}
\begin{tabular}{|c|c|c|}
\hline
Slices of \(\Si{7} H\underline{\Z}\) & Slices of \(\Si{7} H\underline{\F}\) & Slices of \(\Si{8} H\underline{\Z}\) \\ \hline
& \(P^{16}_{16} = \Si{4} H\underline{g}\) & \(P^{16}_{16} = \Si{4} H\underline{g}\) \\
\(P^{12}_{12} = \Si{3} H\underline{g}\) & \(P^{12}_{12} = \Si{3} H\underline{g}^3\) & \(P^{12}_{12} = \Si{3} H\underline{g}^2\) \\
&\(P^{10}_{10} = \Si{\rho+3} H\phi^*_{LDR} \underline{\F}\) & \(P^{10}_{10} = \Si{\rho+3} H\phi^*_{LDR} \underline{\F}\) \\
\(P^{8}_{8} = \Si{\rho+2} H\ul m\) & \(P^{8}_{8} = \Si{\rho+2} H\ul m\) & \(P^8_8 = \Si{\rho+4} H\underline{\Z}\) \\
\(P^7_7 = \Si{\rho+3} H\underline{\Z}\) & \(P^7_7 = \Si{\rho+3} H\underline{\F}\) & \\ \hline
\end{tabular} }
\end{center}
\end{table}
Except for the \(n\)-slice, all slices of \(\Si{7} H_K\underline{\F}\) are recovered from the slices of \(\Si{7} H_K\underline{\Z}\) and \(\Si{8} H_K\underline{\Z}\). It is not always the case, however, that the \((4k+2)\)-slices are recovered. For example, \(\Si{10} H_K\underline{\F}\) has a 14-slice (\cite[Example 6.14]{GY}), but neither \(\Si{10} H_K\underline{\Z}\) nor \(\Si{11} H_K\underline{\Z}\) have 14-slices.
\section{Homotopy Mackey Functor Computations} \label{sec:htpycomp}
Here we compute the homotopy Mackey functors of the slices of \(\Si{\pm n}H\underline{\Z}\).
\begin{prop} \label{htpy:krho:Z}
For \(k\geq 1\), the nontrivial homotopy Mackey functors of \(\Si{k\rho} H\underline{\Z}\) are
\begin{align*}
\ul\pi_n ( \Si{k\rho} H\underline{\Z} ) = \left\lbrace \begin{array}{ll}
\underline{\Z} & n=4k \\
\ul{mg} & n=4k-2 \\
\underline{g}^{\frac{1}{2}(4k-n-1)} & \begin{aligned}
n\in [2k,4k-3], \\ \text{ } n\equiv 1\pmod 2
\end{aligned} \\
\underline{g}^{\frac{1}{2}(4k-n+2)-3} \oplus \phi^*_{LDR} \underline{\F} & \begin{aligned}
n\in[2k,4k-3], \\ \text{ } n\equiv 0\pmod 2
\end{aligned} \\
\underline{g}^{n-k+1} & n\in[k,2k-1]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
For \(n\in[k,4k-2]\), this is a restatement of \cref{-4kslices:K:Z}. For \(n=4k\), the result follows from \cref{-nslice:K:Z} and repeated application of \cref{slice:reduction:ZZ^*}.
\end{proof}
\begin{prop} \label{htpy:-krho:Z}
For \(k\geq 1\), the nontrivial homotopy Mackey functors of \(\Si{-k\rho} H\underline{\Z}\) are
\begin{align*}
\ul\pi_{-n} ( \Si{-k\rho} H\underline{\Z} ) = \left\lbrace \begin{array}{ll}
\underline{\Z}^* & n=4k \\
\ul{mg}^* & n=4k-1 \\
\underline{g}^{\frac{1}{2}(4k-n)} & \begin{aligned}
n\in [2k+4,4k-3], \\ \text{ } n\equiv 0\pmod 2
\end{aligned} \\
\underline{g}^{\frac{1}{2}(4k-n+3)-3} \oplus \phi^*_{LDR} \underline{\F}^* & \begin{aligned}
n\in[2k+3,4k-2], \\ \text{ } n\equiv 1\pmod 2
\end{aligned} \\
\underline{g}^{n-k-3} & n\in[k+4,2k+2]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
For \(n\in[k+4,4k-1]\), this is a restatement of \cref{slices:K:4k:Z}. For \(n=4k\), the result follows from \cref{slices:K:n:Z} and repeated application of \cref{slice:reduction:K:n:Z}.
\end{proof}
\begin{prop} \label{htpy:krho:Z*}
For \(k\geq 1\), the nontrivial homotopy Mackey functors of \(\Si{k\rho} H\underline{\Z}^*\) are
\begin{align*}
\ul\pi_n ( \Si{k\rho} H\underline{\Z}^* ) = \left\lbrace \begin{array}{ll}
\underline{\Z} & n=4k \\
\ul{mg} & n=4k-2 \\
\underline{g}^{\frac{1}{2}(4k-n-1)} & \begin{aligned}
n\in [2k+2,4k-3], \\ \text{ } n\equiv 1\pmod 2
\end{aligned} \\
\underline{g}^{\frac{1}{2}(4k-n-2)-3} \oplus \phi^*_{LDR} \underline{\F} & \begin{aligned}
n\in[2k+2,4k-3], \\ \text{ } n\equiv 0\pmod 2
\end{aligned} \\
\underline{g}^{n-k+2} & n\in[k+3,2k+1]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
This follows from the equivalence \(\Si{k\rho} H\underline{\Z}^* \simeq \Si{(k-1)\rho+4} H\underline{\Z}\) and \cref{htpy:krho:Z}.
\end{proof}
\begin{prop} \label{htpy:krho:Z(2,1)*}
For \(k\geq 1\), the nontrivial homotopy Mackey functors of \(\Si{k\rho} H\underline{\Z}(2,1)^*\) are
\begin{align*}
\ul\pi_n ( \Si{k\rho} H\underline{\Z}(2,1)^* ) = \left\lbrace \begin{array}{ll}
\ul\pi_n(\Si{k\rho} H\underline{\Z}^*) & n\in[k+3,4k] \\
\underline{g} & n=k
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
The exact sequence \(\underline{\Z}^*\rightarrow \underline{\Z}(2,1)^* \rightarrow \underline{g}\) and corresponding cofiber sequence \(\Si{k\rho} H\underline{\Z}^* \rightarrow \Si{k\rho} H\underline{\Z}(2,1)^* \rightarrow \Si{k} H\underline{g}\) provide us with a long exact sequence in homotopy. We then have that the homotopy of \(\Si{k\rho} H\underline{\Z}(2,1)^*\) is the homotopy of \(\Si{k\rho} H\underline{\Z}^*\) with an additional \(\underline{g}\) in degree \(k\).
\end{proof}
\begin{prop} \label{htpy:krho:mg*}
For \(k\geq 2\), the nontrivial homotopy Mackey functors of \(\Si{k\rho} H\ul{mg}^*\) are
\begin{align*}
\ul\pi_n ( \Si{k\rho} H\ul{mg}^* ) = \left\lbrace \begin{array}{ll}
\phi^*_{LDR} \underline{\F} & n=2k\\
\underline{g}^{3} & n\in [k+2,2k-1] \\
\underline{g} & n=k+1
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
We have the equivalence \(\Si{k\rho} H\ul{mg}^* \simeq \Si{(k-1)\rho+2} H\ul m\). The result then follows from \cite[Proposition 7.3]{GY}.
\end{proof}
\begin{prop} \label{htpy:-krho:mg}
For \(k\geq 2\), the nontrivial homotopy Mackey functors of \(\Si{-k\rho} H\ul{mg}\) are
\begin{align*}
\ul\pi_{-n} ( \Si{-k\rho} H\ul{mg} ) = \left\lbrace \begin{array}{ll}
\phi^*_{LDR} \underline{\F}^* & n=2k \\
\underline{g}^{3} & n\in [k+2,2k-1] \\
\underline{g} & n=k+1
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
The result follows by taking the Brown-Comenetz dual of each Mackey functor in \cref{htpy:krho:mg*}.
\end{proof}
\begin{prop} \label{htpy:-krho:m}
For \(k\geq 2\), the nontrivial homotopy Mackey functors of \(\Si{-k\rho} H\ul m\) are
\begin{align*}
\ul\pi_{-n} ( \Si{-k\rho} H\ul{m} ) = \left\lbrace \begin{array}{ll}
\phi^*_{LDR} \underline{\F}^* & n=2k \\
\underline{g}^{3} & n\in [k+2,2k-1] \\
\underline{g}^2 & n=k+1
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
First, take the Brown-Comenetz dual of the Mackey functors in \cite[Proposition 7.4]{GY}. The result then follows from the equivalence \(\Si{-\rho} H\ul m \simeq \Si{-2} H\ul{mg}^*\).
\end{proof}
\begin{prop} \label{htpy:krho:phiF*}
We have the equivalences
\begin{align*}
\Si{k\rho} H\phi^*_{LDR}\underline{\F}^* \simeq \left\lbrace \begin{array}{ll}
\Si{2} H\phi^*_{LDR}\ul f & k=1 \\
\Si{4} H\phi^*_{LDR}\underline{\F} & k=2
\end{array}\right.
\end{align*}
Then for \(k\geq 3\), the nontrivial homotopy Mackey functors of \(\Si{k\rho} H\phi^*_{LDR} \underline{\F}^*\) are
\begin{align*}
\ul\pi_{n} ( \Si{k\rho} H\phi^*_{LDR} \underline{\F}^* ) = \left\lbrace \begin{array}{ll}
\phi^*_{LDR} \underline{\F} & n=2k\\
\underline{g}^{3} & n\in [k+2,2k-1]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
This is the pullback over \(L\), \(D\), and \(R\) of the Mackey functors in \cite[Proposition 3.6]{GY}.
\end{proof}
\begin{prop} \label{htpy:-krho:phiF}
We have the equivalences
\begin{align*}
\Si{-k\rho} H\phi^*_{LDR}\underline{\F} \simeq \left\lbrace \begin{array}{ll}
\Si{-2} H\phi^*_{LDR}\ul f & k=1 \\
\Si{-4} H\phi^*_{LDR}\underline{\F}^* & k=2
\end{array}\right.
\end{align*}
Then for \(k\geq 3\), the nontrivial homotopy Mackey functors of \(\Si{-k\rho} H\phi^*_{LDR} \underline{\F}\) are
\begin{align*}
\ul\pi_{-n} ( \Si{-k\rho} H\phi^*_{LDR} \underline{\F} ) = \left\lbrace \begin{array}{ll}
\phi^*_{LDR} \underline{\F}^* & n=2k\\
\underline{g}^{3} & n\in [k+2,2k-1]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
This is the pullback over \(L\), \(D\), and \(R\) of the Mackey functors in \cite[Proposition 3.7]{GY}.
\end{proof}
\section{Spectral Sequences} \label{sec:SSS}
\sseqset{
Fclass/.sseq style = {fill = black, diamond, draw, inner sep = 0.1ex},
classes= {fill, inner sep = 0pt, minimum size = 0.22em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 225, scale = 1.2 },
run off differentials = ->,
grid = go,
grid color = gray!30
Zclass/.sseq style={fill=black, rectangle, draw,inner sep=0.4ex},
Zstarclass/.sseq style={fill=none, rectangle, draw,inner sep=0.4ex},
Z4class/.sseq style={fill=none, circle, draw,inner sep=0.28ex},
Z4classSource/.sseq style={circle, inner sep = 0.28ex, draw, fill = red!70},
Z4classTarget/.sseq style={circle, inner sep = 0.28ex, draw, fill = black!40},
FstarLDRclass/.sseq style = {fill = none, regular polygon, sides=5, draw, inner sep = 0.4ex},
FLDRclass/.sseq style = {fill = black, regular polygon, regular polygon sides=5, draw, inner sep = 0.4ex},
mgstarclass/.sseq style = {fill = none, trapezium, draw, inner sep = 0.25ex},
mgclass/.sseq style = {fill = black, trapezium, draw, inner sep = 0.25ex},
mstarclass/.sseq style = {fill = none, regular polygon, regular polygon sides=3, draw, inner sep = 0.4ex},
}
\begin{sseqdata}[name=SI-1K,
y range={-10}{0},
x range={-10}{0},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zstarclass, "\phantom{*}" {inside,font=\tiny}](-1,-1)
\class[mstarclass](-1,-2)
\class[Z4class,"1" {inside,font=\tiny}](-1,-3)
\structline[dashed](-1,-3)(-1,-2)
\structline[dashed](-1,-2)(-1,-1)
\draw[fill=white](-9.9,-1.1) rectangle (-6.4,0.0);
\node at (-8.1,-0.6) {\Si{-1}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI-3K,
y range={-10}{0},
x range={-10}{0},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zstarclass, "\phantom{*}" {inside,font=\tiny}](-3,0)
\class[FstarLDRclass](-3,-3)
\class[Z4class,"1" {inside,font=\tiny}](-2,-6)
\class[Z4class,"2" {inside,font=\tiny}](-3,-9)
\d2(-2,-6)(-3,-3)
\structline[dashed](-3,-9)(-3,-3)
\structline[dashed](-3,-3)(-3,0)
\draw[fill=white](-9.9,-1.1) rectangle (-6.4,0.0);
\node at (-8.1,-0.6) {\Si{-3}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI-5K,
y range={-16}{0},
x range={-10}{0},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zstarclass, "\phantom{*}" {inside,font=\tiny}](-5,0)
\class[mgstarclass, "\phantom{*}" {inside,font=\tiny}](-4,-1)
\class[FstarLDRclass](-3,-3)
\class[FstarLDRclass](-5,-5)
\foreach \x in {3,...,1} {
\class[Z4class, "\x" {inside,font=\tiny}](-6+\x,-18+3*\x)
}
\class[Z4class,"1" {inside,font=\tiny}](-3,-2)
\class[Z4class,"1" {inside,font=\tiny}](-2,-4)
\class[Z4class,"1" {inside,font=\tiny}](-2,-6)
\class[Z4class,"3" {inside,font=\tiny}](-4,-6)
\d2(-2,-4)(-3,-2)
\d2(-3,-3)(-4,-1)
\replacesource
\d3(-2,-6)(-3,-3)
\d3(-3,-9)(-4,-6)
\d7(-4,-12)(-5,-5)
\structline[dashed](-5,-15)(-5,-5)
\structline[dashed](-5,-5)(-5,0)
\draw[fill=white](-9.9,-1.1) rectangle (-6.4,0.0);
\node at (-8.1,-0.6) {\Si{-5}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI-7K,
y range={-28}{0},
x range={-9}{0},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zstarclass, "\phantom{*}" {inside,font=\tiny}](-7,0)
\class[mgstarclass, "\phantom{*}" {inside,font=\tiny}](-6,-1)
\class[FstarLDRclass](-5,-5)
\class[FstarLDRclass](-7,-7)
\class[Z4class,"1" {inside,font=\tiny}](-5,-2)
\foreach \x in {1,...,4} {
\class[Z4class, "\x" {inside,font=\tiny}](-8+\x,-24+3*\x)
}
\foreach \x in {2,...,3} {
\class[Z4class, "\x" {inside,font=\tiny}](-1-\x,-15+3*\x)
}
\foreach \x in {0,...,1} {
\class[Z4class, "3" {inside,font=\tiny}](-6+\x,-8-\x)
}
\d3(-3,-9)(-4,-6)
\replacetarget
\d3(-4,-12)(-5,-9)
\replacesource
\d4(-5,-5)(-6,-1)
\replacesource
\d4(-4,-6)(-5,-2)
\d7(-4,-12)(-5,-5)
\d7(-5,-15)(-6,-8)
\d11(-6,-18)(-7,-7)
\structline[dashed](-7,-21)(-7,-7)
\structline[dashed](-7,-7)(-7,0)
\draw[fill=white](-3.9,-1.1) rectangle (-0.4,0.0);
\node at (-2.1,-0.6) {\Si{-7}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI-9K,
y range={-28}{0},
x range={-11}{0},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zstarclass, "\phantom{*}" {inside,font=\tiny}](-9,0)
\class[mgstarclass, "\phantom{*}" {inside,font=\tiny}](-8,-1)
\class[FstarLDRclass](-6,-3)
\foreach \x in {0,1,2} {
\class[FstarLDRclass](-9+2*\x,-9+2*\x)
}
\class[Z4class,"1" {inside,font=\tiny}](-3,-7)
\class[Z4class,"1" {inside,font=\tiny}](-3,-9)
\class[Z4class,"3" {inside,font=\tiny}](-4,-6)
\class[Z4class,"3" {inside,font=\tiny}](-4,-12)
\class[Z4class,"1" {inside,font=\tiny}](-4,-5)
\class[Z4class,"2" {inside,font=\tiny}](-5,-4)
\class[Z4class,"1" {inside,font=\tiny}](-7,-2)
\foreach \x in {1,...,5} {
\class[Z4class, "\x" {inside,font=\tiny}](-10+\x,-30+3*\x)
}
\foreach \x in {0,...,2} {
\class[Z4class, "3" {inside,font=\tiny}](-6-\x,-12+\x)
}
\foreach \x in {0,...,1} {
\class[Z4class, "3" {inside,font=\tiny}](-5-\x,-9+\x)
}
\d2(-4,-6)(-5,-4)
\replacesource
\d3(-4,-9-3*1)(-4-1,-6-3*1)
\d2(-3-2,-7+2)(-4-2,-5+2)
\d3(-3-2,-9-3*2)(-4-2,-6-3*2)
\replacesource
\foreach \x in {0} {
\d2(-3-\x,-7+\x)(-4-\x,-5+\x)
\d3(-3-\x,-9-3*\x)(-4-\x,-6-3*\x)
}
\foreach \x in {0,1} {
\d6(-6-\x,-8+\x)(-7-\x,-2+\x)
\replacesource
\d7(-5-\x,-15-3*\x)(-6-\x,-8-3*\x)
\replacesource
}
\d11(-6,-18)(-7,-7)
\d11(-7,-21)(-8,-10)
\d15(-8,-24)(-9,-9)
\structline[dashed](-9,-27)(-9,-9)
\structline[dashed](-9,-9)(-9,0)
\draw[fill=white](-3.9,-1.1) rectangle (-0.4,0.0);
\node at (-2.1,-0.6) {\Si{-9}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI6K,
y range={0}{10},
x range={0}{10},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zclass, "\phantom{*}" {inside,font=\tiny}](6,0)
\class[Z4class,"1" {inside,font=\tiny}](2,6)
\class[Z4class,"1" {inside,font=\tiny}](3,3)
\d3(3,3)(2,6)
\draw[fill=white](7.5,8.5) rectangle (10.3,9.6);
\node at (8.9,9.0) {\Si{6}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI7K,
y range={0}{10},
x range={0}{10},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zclass, "\phantom{*}" {inside,font=\tiny}](7,0)
\class[mgclass, "\phantom{*}" {inside,font=\tiny}](5,2)
\class[FLDRclass](4,4)
\class[Z4class,"1" {inside,font=\tiny}](4,3)
\class[Z4class,"1" {inside,font=\tiny}](3,5)
\class[Z4class,"1" {inside,font=\tiny}](3,9)
\d2(5,2)(4,4)
\replacetarget
\d2(4,3)(3,5)
\d5(4,4)(3,9)
\draw[fill=white](7.5,8.5) rectangle (10.3,9.6);
\node at (8.9,9) {\Si{7}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI8K,
y range={0}{13},
x range={0}{10},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zclass, "\phantom{*}" {inside,font=\tiny}](8,0)
\class[mgclass, "\phantom{*}" {inside,font=\tiny}](6,2)
\class[FLDRclass](5,5)
\class[Z4class,"3" {inside,font=\tiny}](4,6)
\class[Z4class,"2" {inside,font=\tiny}](3,9)
\class[Z4class,"1" {inside,font=\tiny}](5,3)
\d3(5,3)(4,6)
\d3(4,6)(3,9)
\class[Z4class,"1" {inside,font=\tiny}](4,12)
\d3(6,2)(5,5)
\replacetarget
\d7(5,5)(4,12)
\draw[fill=white](7.5,9.5) rectangle (10.3,10.6);
\node at (8.9,10) {\Si{8}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI10K,
y range={0}{19},
x range={0}{12},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zclass, "\phantom{*}" {inside,font=\tiny}](10,0)
\class[mgclass, "\phantom{*}" {inside,font=\tiny}](8,2)
\class[FLDRclass](7,7)
\class[Z4class,"1" {inside,font=\tiny}](6,18)
\class[Z4class,"2" {inside,font=\tiny}](5,15)
\class[Z4class,"1" {inside,font=\tiny}](3,9)
\class[Z4class,"3" {inside,font=\tiny}](6,8)
\class[Z4class,"3" {inside,font=\tiny}](4,12)
\class[Z4class,"3" {inside,font=\tiny}](5,9)
\class[Z4class,"1" {inside,font=\tiny}](4,6)
\class[Z4class,"1" {inside,font=\tiny}](7,3)
\d3(4,6)(3,9)
\d3(5,9)(4,12)
\d5(7,3)(6,8)
\replacetarget
\d7(6,8)(5,15)
\d5(8,2)(7,7)
\replacetarget
\d11(7,7)(6,18)
\draw[fill=white](9.5,18.5) rectangle (12.3,19.6);
\node at (10.9,19.0) {\Si{10}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI11K,
y range={0}{21},
x range={0}{13},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zclass, "\phantom{*}" {inside,font=\tiny}](11,0)
\class[mgclass, "\phantom{*}" {inside,font=\tiny}](9,2)
\class[FLDRclass](7,4)
\class[FLDRclass](8,8)
\class[FLDRclass](6,6)
\class[Z4class,"1" {inside,font=\tiny}](7,21)
\class[Z4class,"2" {inside,font=\tiny}](6,18)
\class[Z4class,"3" {inside,font=\tiny}](5,15)
\class[Z4class,"3" {inside,font=\tiny}](7,9)
\class[Z4class,"3" {inside,font=\tiny}](6,10)
\class[Z4class,"1" {inside,font=\tiny}](4,12)
\class[Z4class,"3" {inside,font=\tiny}](5,7)
\class[Z4class,"1" {inside,font=\tiny}](4,8)
\class[Z4class,"1" {inside,font=\tiny}](8,3)
\class[Z4class,"2" {inside,font=\tiny}](6,5)
\class[Z4class,"1" {inside,font=\tiny}](5,6)
\d2(5,6)(4,8)
\d2(7,4)(6,6)
\d2(6,5)(5,7)
\replacetarget
\d5(5,7)(4,12)
\d5(6,10)(5,15)
\d6(8,3)(7,9)
\replacetarget
\d6(9,2)(8,8)
\replacetarget
\d9(7,9)(6,18)
\d13(8,8)(7,21)
\draw[fill=white](10.5,20.5) rectangle (13.3,21.6);
\node at (11.9,21.0) {\Si{11}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI12K,
y range={0}{25},
x range={0}{12},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zclass, "\phantom{*}" {inside,font=\tiny}](12,0)
\class[mgclass, "\phantom{*}" {inside,font=\tiny}](10,2)
\class[FLDRclass](8,4)
\class[FLDRclass](9,9)
\class[FLDRclass](7,7)
\foreach \x in {1,...,4} {
\class[Z4class, "\x" {inside,font=\tiny}](9-\x,27-3*\x)
}
\foreach \x in {0,...,2} {
\class[Z4class, "3" {inside,font=\tiny}](8-\x,10+\x)
}
\class[Z4class,"1" {inside,font=\tiny}](9,3)
\class[Z4class,"2" {inside,font=\tiny}](7,5)
\class[Z4class,"1" {inside,font=\tiny}](6,6)
\class[Z4class,"2" {inside,font=\tiny}](4,12)
\foreach \x in {0,...,1} {
\class[Z4class, "3" {inside,font=\tiny}](6-\x,8+\x)
}
\d3(6,12)(5,15)
\replacetarget
\d6(7,11)(6,18)
\d7(9,3)(8,10)
\replacetarget
\d12(8,10)(7,21)
\d7(10,2)(9,9)
\replacetarget
\d15(9,9)(8,24)
\d3(6,6)(5,9)
\replacesource
\d3(5,9)(4,12)
\foreach \x in {1,...,2} {
\d3(6+\x,6-\x)(5+\x,9-\x)
\replacetarget
}
\d7(6,8)(5,15)
\draw[fill=white](1.5,23.5) rectangle (4.3,24.6);
\node at (2.9,24.0) {\Si{12}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI14K,
y range={0}{31},
x range={0}{14},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zclass, "\phantom{*}" {inside,font=\tiny}](14,0)
\class[mgclass, "\phantom{*}" {inside,font=\tiny}](12,2)
\class[FLDRclass](10,4)
\class[FLDRclass](11,11)
\class[FLDRclass](9,9)
\foreach \x in {1,...,5} {
\class[Z4class, "\x" {inside,font=\tiny}](11-\x,33-3*\x)
}
\foreach \x in {0,...,3} {
\class[Z4class, "3" {inside,font=\tiny}](10-\x,12+\x)
}
\class[Z4class,"1" {inside,font=\tiny}](11,3)
\class[Z4class,"2" {inside,font=\tiny}](9,5)
\class[Z4class,"1" {inside,font=\tiny}](8,6)
\class[Z4class,"1" {inside,font=\tiny}](4,12)
\class[Z4class,"3" {inside,font=\tiny}](5,15)
\class[Z4class,"1" {inside,font=\tiny}](5,9)
\foreach \x in {0,...,2} {
\class[Z4class, "3" {inside,font=\tiny}](8-\x,10+\x)
}
\d9(12,2)(11,11)
\replacetarget
\d19(11,11)(10,30)
\d9(11,3)(10,12)
\replacetarget
\d16(10,12)(9,27)
\foreach \x in {0,...,1} {
\d3\x(5+\x,9+3*\x)(4+\x,12+3*\x)
\replacetarget
}
\d3(7,15)(6,18)
\replacetarget
\d11(9,13)(8,24)
\d7(8,14)(7,21)
\replacetarget
\foreach \x in {2,...,2} {
\d5\x(8+\x,6-\x)(7+\x,11-\x)
\replacetarget
}
\d5(9,5)(8,10)
\replacetarget
\d5(8,6)(7,11)
\replacetarget
\d7(7,11)(6,18)
\d11(8,10)(7,21)
\draw[fill=white](1.5,29.5) rectangle (4.3,30.6);
\node at (2.9,30.0) {\Si{14}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI7KF,
y range={0}{13},
x range={0}{10},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Fclass, "\phantom{*}" {inside,font=\tiny}](7,0)
\class[mgclass, "\phantom{*}" {inside,font=\tiny}](6,1)
\class[FLDRclass](5,2)
\class[FLDRclass](4,4)
\class[FLDRclass](5,5)
\class[Z4class,"1" {inside,font=\tiny}](4,3)
\class[Z4class,"1" {inside,font=\tiny}](3,5)
\class[Z4class,"1" {inside,font=\tiny}](4,12)
\d2(5,2)(4,4)
\d2(4,3)(3,5)
\class[Z4class,"3" {inside,font=\tiny}](4,6)
\class[Z4class,"3" {inside,font=\tiny}](3,9)
\d3(4,6)(3,9)
\d4(6,1)(5,5)
\replacetarget
\d7(5,5)(4,12)
\draw[fill=white](7.5,9.5) rectangle (10.3,10.6);
\node at (8.9,10) {\Si{7}H_{K}\underline{\F}};
\end{sseqdata}
\begin{sseqdata}[name=SI10KF,
y range={0}{21},
x range={0}{12},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Fclass, "\phantom{*}" {inside,font=\tiny}](10,0)
\class[mgclass, "\phantom{*}" {inside,font=\tiny}](9,1)
\class[FLDRclass](8,2)
\foreach \x in {0,...,2} {
\class[FLDRclass](6+\x,6+\x)
}
\class[FLDRclass](7,3)
\d7(9,1)(8,8)
\replacetarget
\d5(8,2)(7,7)
\replacesource
\d3(7,3)(6,6)
\replacesource
\class[Z4class,"4" {inside,font=\tiny}](4,12)
\class[Z4class,"5" {inside,font=\tiny}](5,15)
\class[Z4class,"3" {inside,font=\tiny}](6,18)
\class[Z4class,"1" {inside,font=\tiny}](7,21)
\foreach \x in {0,...,2} {
\class[Z4class, "3" {inside,font=\tiny}](5+\x,7+\x)
}
\class[Z4class,"3" {inside,font=\tiny}](5,9)
\class[Z4class,"3" {inside,font=\tiny}](6,10)
\class[Z4class,"2" {inside,font=\tiny}](6,4)
\class[Z4class,"1" {inside,font=\tiny}](8,2)
\d7(8,2)(7,9)
\replacetarget
\d3(6,4)(5,7)
\d3(5,9)(4,12)
\d5(7,3)(6,8)
\d9(7,9)(6,18)
\d5(6,10)(5,15)
\d13(8,8)(7,21)
\draw[fill=white](9.5,20.5) rectangle (12.3,21.6);
\node at (10.9,21.0) {\Si{10}H_{K}\underline{\F}};
\end{sseqdata}
The slice spectral sequence for \(\Si{-n} H\underline{\Z}\) and \(\Si{n} H\underline{\Z}\) must recover the homotopy Mackey functors of each spectrum, that is, we must be left with \(\ul\pi_{-n}(\Si{-n}H\underline{\Z}) = \ul\pi_{n}(\Si{n}H\underline{\Z}) = \underline{\Z} \) and all other homotopy Mackey functors trivial. For most of the differentials, then, there is only one choice.
We use the indexing convention from \cite[Section 4.4.2]{HHR}. The Mackey functor \(\ul E_2^{t-n,t}\) is \(\ul\pi_n P^t_t(X)\). We also use the Adams convention, so that \(\ul\pi_n P^t_t(X)\) has coordinates \((n,n-t)\) and the differential,
\begin{align*}
d_r:\ul E_r^{s,t} \rightarrow \ul E_r^{s+r,t+r-1},
\end{align*}
points left one and up \(r\).
The symbols in \cref{SSS:mack} denote the Mackey functors in the slice spectral sequences shown.
\begin{table}[h]
\caption[Symbols for {$K$}-Mackey functors]{Symbols for $K$-Mackey functors}
\label{SSS:mack}
\begin{center}
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c|c|}
\hline
$\blacksquare=\underline{\Z}$
& $\bpent=\phi^*_{LDR} \underline{\F}$
& $\btrap=\ul{mg}$
\\\hline
$\square=\underline{\Z}^*$
& $\pent=\phi_{LDR}^*\underline{\F}^*$
& $\trap = \ul {mg}^*$
\\ \hline
$\underline{\F} = \bdiamond$ & $\gcirc=\underline{g}^n$ & $\triangle = \ul m^*$\\\hline
\end{tabular} }
\end{center}
\end{table}
\begin{exmp} \label{SSS:-1HZ}
The slices for \(\Si{-1} H\underline{\Z}\) are all a one-fold desuspension of
Eilenberg-MacLane spectra (\cref{tower:-1:K:Z}, \cref{fig:SSS:-1..-5:HZ}). Because each of these Mackey functors is in the same column, there are no differentials. Consequently, in the spectral sequence, we find a double extension:
\begin{equation} \label{double:extension}
\begin{tikzcd}[row sep=small]
\underline{\Z}^* \ar[dr] & & \\
& \underline{\Z}(2,1) \ar[dr] \ar[dl] & \\
\ul m^* & & \underline{\Z} \ar[dddll] \\
& & \\ \\
\underline{g} & &
\end{tikzcd}
\end{equation}
\end{exmp}
\begin{exmp} \label{SSS:-5HZ}
In the spectral sequence for \(\Si{-5} H\underline{\Z}\), \cref{fig:SSS:-1..-5:HZ}, because we can only be left with \(\ul\pi_{-5}(P^{-5}_{-5} \Si{-5} H\underline{\Z}) \cong \underline{\Z}\) and all differentials must go left one and up at least two, all differentials are forced. Once we have evaluated each differential, we once again find ourselves with the double extension in \cref{double:extension}.
\end{exmp}
\begin{exmp} \label{SSS:-9HZ}
In \cref{fig:SSS:-7..-9:HZ}, most of the differentials for \(\Si{-9} H\underline{\Z}\) are again forced by the fact that only \(\ul\pi_{-9}(P^{-9}_{-9} \Si{-9} H\underline{\Z}) \cong \underline{\Z}\) can survive the spectral sequence. For example, we have two choices for a differential from \(\pi_{-8}P^{-32}_{-32} \Si{-9} H\underline{\Z} \cong \underline{g}^2\). We find it must be
\begin{align*}
d_{15}: \underline{g}^2 \rightarrow \phi^*_{LDR}\underline{\F}^* \cong \ul\pi_{-9} (P^{-24}_{-24} \Si{-9} H\underline{\Z})
\end{align*}
so that we are left with the extension in \cref{double:extension}. Indeed, we will always be left with this extension once all differentials have been evaluated. Similarly, for \(n\equiv 1\pmod 4\), we will always have a
\begin{align*}
d_2:\phi^*_{LDR} \underline{\F}^* \xrightarrow{\cong}{} \phi^*_{LDR} \underline{\F}^*
\end{align*}
in the upper right corner.
\end{exmp}
In \cref{-4k-2slices:K:Z} we claim that for \(n\) odd,
\begin{align*}
P^{-2n}_{-2n} (\Si{-n} H\underline{\Z}) \simeq \Si{-(k+1)\rho +1} H\phi^*_{LDR} \underline{\F}.
\end{align*}
We now prove this claim.
\begin{prop} \label{phiF:SSS}
Let \(n\) be odd and take \(\ul A\) to be one of the Mackey functors listed in \cref{tab-SShtpydegrees}. The only choice of \(\ul A\) where the homotopy of \(P^{-2n}_{-2n} (\Si{-n} H\underline{\Z}) \simeq \Si{-(k+1)\rho+1} H\ul A\) fits into the spectral sequence for \(\Si{-n} H\underline{\Z}\) is \(\phi^*_{LDR} \underline{\F}\).
\end{prop}
\begin{proof}
Because \(\Si{-(k+1)\rho+1} H\ul A\) is a \((-4k-2)\)-slice, its \(\ul\pi_{-k-1}\) is located at \((-k-1,-3k-1)\). We argue that we cannot have a nonzero Mackey functor in this location.
\begin{table}[h]
\caption[Homotopy Comparison]{Homotopy Comparison}
\label{tab-SShtpydegrees}
\begin{center}
{\renewcommand{\arraystretch}{1.5}\setlength{\tabcolsep}{12pt}
\begin{tabular}{|c|c|} \hline
$\ul A$ & $\ul\pi _{-k-1}(\Si{-(k+1)\rho +1} H\ul A)$ \\ \hline
$\phi^*_{LDR} \ul f$ & $\underline{g}^3$ \\
$\ul m$ & $\underline{g}^2$ \\
$\ul{mg}$ & $\underline{g}$ \\
$\phi^*_{LDR} \underline{\F}$ & $\ul 0$ \rule[-2.4ex]{0pt}{2.2ex} \\ \hline
\end{tabular} }
\end{center}
\end{table}
In the spectral sequence for \(\Si{-n} H\underline{\Z}\), all slices below level \(-2n\) are \(\Si{-k} H\underline{g}^{n-k+1}\) where \(4k\in[2n+1,4n]\). These Mackey functors, for \(4k\leq 2n+1\), lie on the line \(y=-3k\). Thus, for the Mackey functors in \cref{tab-SShtpydegrees}, the source of a differential hitting it must be \((-k,-3k)\). This is not possible.
We now argue that the Mackey functor located at \((-k-1,-3k-1)\) cannot be the source of a differential.The first value of \(n\) for which we must determine the \((-2n)\)-slice is \(n=7\). The spectral sequence where \(\ul A=\phi^*_{LDR} \underline{\F}\) is shown in \cref{fig:SSS:-7..-9:HZ}. This spectral sequence leaves us with the appropriate homotopy for \(\Si{-7} H\underline{\Z}\).
Note there is a copy of \(\underline{g}^4\) located at \((-4,-12)\). For any other choice of \(\ul A\), we would have a nontrivial \(\ul\pi_{-4}\) located at \((-4,-10)\). For a differential originating from \((-4,-10)\), there are two possible targets: \(\phi^*_{LDR} \underline{\F}^*\) at \((-5,-5)\) and \(\underline{g}\) at \((-5,-2)\). However, these two Mackey functors must fit into the exact sequences
\begin{align*}
\underline{g}^2 \rightarrow \underline{g}^3\rightarrow \underline{g} \qquad \text{ and } \qquad
\underline{g}^4 \rightarrow \phi^*_{LDR} \underline{\F}^* \rightarrow \ul {mg}.
\end{align*}
Thus, we cannot have a nonzero Mackey functor at \((-4,-10)\).
We now consider the spectral sequence for \(\Si{-9} H\underline{\Z}\), located in \cref{fig:SSS:-7..-9:HZ}. We again use \(\Si{-(l+1)\rho + 1} H\phi^*_{LDR} \underline{\F}\) for the \((-2n)\)-slice. The resulting homotopy fits in the spectral sequence. For the other three choices of \(\ul A\) we would have a nontrivial \(\ul\pi_{-5}\) located at \((-5,-13)\). For a differential originating from \((-5,-13)\) there are two possible targets: \(\underline{g}^3\) at \((-6,-8)\) and \(\phi^*_{LDR}\underline{\F}^*\) at \((-6,-3)\).
However, we have a \(d_2:\phi^*_{LDR}\underline{\F}^* \twoheadrightarrow \phi^*_{LDR}\underline{\F}^*\) and a \(d_7:\underline{g}^5\twoheadrightarrow \underline{g}^3\). Thus, there is no target for a differential from \((-5,-13)\). Consequently, the only choice of \(\ul A\) that works is \(\phi^*_{LDR} \underline{\F}\).
There is a similar story for all odd \(n>9\). There will always be a \(d_2\) hitting the \(\phi^*_{LDR} \underline{\F}^*\) located at \((-k-2,-n+k+2)\). The only other possible targets for a differential from \((-k-1,-3k-1)\) are then the \(\underline{g}^3\)'s resulting from the homotopy of the other \((-4j-2)\)-slices. All of these will be hit by a differential from the \(\underline{g}^{n-k+1}\) located at \((-k-1,-3k-3)\). Thus, we cannot have a nonzero Mackey functor located at \((-k-1,-3k-1)\). The only choice of \(\ul A\) which satisfies this requirement is \(\phi^*_{LDR} \underline{\F}\).
\end{proof}
\begin{exmp} \label{SSS:6HZ}
For the positive, trivial suspensions of \(H\underline{\Z}\), we find that \(\Si{6} H\underline{\Z}\) has the first nontrivial slice tower. In \cref{fig:SSS:6..8:HZ}, we then see that there is only one possible differential. This \(d_3\) exists because we must be left with only \(\ul\pi_6 (\Si{6} H\underline{\Z}) \cong \underline{\Z}\).
\end{exmp}
\begin{exmp} \label{SSS:7HZ}
The spectral sequence for \(\Si{7} H\underline{\Z}\), in \cref{fig:SSS:6..8:HZ}, is more interesting. Here we find the differentials \(d_2:\phi^*_{LDR}F \rightarrow \ul{mg}\) and \(d_5:\ul{mg}\rightarrow \underline{g}\). Indeed, we will always see a \(d_{n-7}:\phi^*_{LDR}\underline{\F} \rightarrow \ul{mg}\) and \(d_{2n-9}:\ul{mg} \rightarrow \underline{g}\) on the right side of the spectral sequence for \(\Si{n} H_K\underline{\Z}\).
\end{exmp}
\begin{exmp} \label{SSS:7F}
Except for the homotopy of the \(n\)-slice of \(\Si{7} H_K\underline{\Z}, \quad \) \(\Si{7} H_K\underline{\F}\), and \(\Si{8} H_K\underline{\Z}\), the spectral sequences for \(\Si{7} H_K\underline{\Z}\) and \(\Si{8} H_K\underline{\Z}\) collapse to give the spectral sequence for \(\Si{7} H_K\underline{\F}\). We see in \cref{fig:SSS:6..8:HZ} the \(\underline{g}\) in \((3,9)\) and the \(\underline{g}^2\) in \((3,9)\) in the spectral sequences for \(\Si{7} H_K\underline{\Z}\) and \(\Si{8} H_K\underline{\Z}\), respectively, combine to give the \(\underline{g}^3\) in \((3,9)\) in the spectral sequence for \(\Si{7} H_K\underline{\F}\). Off the diagonal for the \(n\)-slice for \(\Si{7} H_K\underline{\Z}\) and \(\Si{8} H_K\underline{\Z}\) we have a single copy of \(\phi^*_{LDR}\underline{\F}\). These provide the two copies of \(\phi^*_{LDR}\underline{\F}\) off the diagonal for \(\Si{7} H_K\underline{\F}\).
\end{exmp}
\begin{exmp} \label{SSS:11HZ}
Now, in \cref{fig:SSS:10..11:HZ}, we have some choice of differentials in the spectral sequence for \(\Si{11} H\underline{\Z}\). Once we consider that only \(\ul\pi_{11}(\Si{11} H\underline{\Z}) \cong \underline{\Z}\) can be left, there is only one choice of each differential that provides the desired result. Analogously to the spectral sequence for \(\Si{-n} H\underline{\Z}\) where \(n\equiv 1\pmod 4\), we will always have a \(d_2:\phi^*_{LDR} \underline{\F} \xrightarrow{\cong}{} \phi^*_{LDR} \underline{\F}^*\) in the bottom left corner when \(n\equiv 3\pmod 4\).
\end{exmp}
\begin{exmp} \label{SSS:10F}
Again, with the exception of the homotopy of the \(n\)-slice, the spectral sequences for \(\Si{10} H_K\underline{\Z}\) and \(\Si{11} H_K\underline{\Z}\) collapse to give the spectral sequence for \(\Si{10} H_K\underline{\F}\) in \cref{fig:SSS:10..11:HZ}. As in \cref{SSS:7F}, the upper left diagonals in in the spectral sequences for \(\Si{10} H_K\underline{\Z}\) and \(\Si{11} H_K\underline{\Z}\) combine to even more copies of \(\underline{g}\) in the upper left diagonal in the spectral sequence for \(\Si{10} H_K\underline{\F}\). This will always be the case.
\end{exmp}
\clearpage
\begin{fig}\label{fig:SSS:-1..-5:HZ}
\begin{flushleft}
The slice spectral sequence over \(K\), \(n=-1,-3,-5\).
\end{flushleft}
\end{fig}
\begin{adjustbox}{minipage=[r][\textheight][b]{\paperwidth},scale={0.8}}
\printpage[name=SI-1K]
\printpage[name=SI-3K]\newline
\printpage[name=SI-5K]
\end{adjustbox}
\clearpage
\begin{fig}\label{fig:SSS:-7..-9:HZ}
\begin{flushleft}
The slice spectral sequence over \(K\), \(n=-7,-9\).
\end{flushleft}
\end{fig}
\begin{adjustbox}{scale=0.8}
\printpage[name=SI-7K]
\printpage[name=SI-9K]
\end{adjustbox}
\clearpage
\begin{fig}\label{fig:SSS:6..8:HZ}
\begin{flushleft}
The slice spectral sequence over \(K\), \(n=6,7,8\).
\end{flushleft}
\end{fig}
\begin{adjustbox}{minipage=[r][\textheight][b]{\paperwidth},scale={0.8}}
\printpage[name=SI6K]
\printpage[name=SI7K] \newline
\printpage[name=SI8K]
\printpage[name=SI7KF]
\end{adjustbox}
\vfill
\clearpage
\begin{fig}\label{fig:SSS:10..11:HZ}
\begin{flushleft}
The slice spectral sequence over \(K\), \(n=10,11\).
\end{flushleft}
\end{fig}
\begin{adjustbox}{scale={0.7}}
\printpage[name=SI10K]
\end{adjustbox} \newline
\begin{adjustbox}{scale=0.7}
\printpage[name=SI10KF]
\printpage[name=SI11K]
\end{adjustbox}
\clearpage
\begin{fig}\label{fig:SSS:12:HZ}
\begin{flushleft}
The slice spectral sequence over \(K\), \(n=12\).
\end{flushleft}
\end{fig}
\begin{adjustbox}{scale=0.9}
\printpage[name=SI12K]
\end{adjustbox}
\clearpage
\begin{fig}\label{fig:SSS:14:HZ}
\begin{flushleft}
The slice spectral sequence over \(K\), \(n=14\).
\end{flushleft}
\end{fig}
\begin{adjustbox}{scale=0.9}
\printpage[name=SI14K]
\end{adjustbox}
\begin{bibdiv}
\begin{biblist}
\bib{D}{article}{
author={Dugger, Daniel},
title={An Atiyah-Hirzebruch spectral sequence for $KR$-theory},
journal={$K$-Theory},
volume={35},
date={2005},
number={3-4},
pages={213--256 (2006)},
issn={0920-3036},
review={\MR{2240234}},
doi={10.1007/s10977-005-1552-9},
}
\bib{GM}{article}{
author={Greenlees, J. P. C.},
author={Meier, Lennart},
title={Gorenstein duality for real spectra},
journal={Algebr. Geom. Topol.},
volume={17},
date={2017},
number={6},
pages={3547--3619},
issn={1472-2747},
review={\MR{3709655}},
doi={10.2140/agt.2017.17.3547},
}
\bib{GY}{article}{
author={Guillou, B.},
author={Yarnall, C.},
title={The Klein four slices of $\Sigma^n H\underline{\F}_2$},
journal={Math. Z.},
volume={295},
date={2020},
number={3-4},
pages={1405--1441},
issn={0025-5874},
review={\MR{4125695}},
doi={10.1007/s00209-019-02433-3},
}
\bib{H}{article}{
author={Hill, Michael A.},
title={The equivariant slice filtration: a primer},
journal={Homology Homotopy Appl.},
volume={14},
date={2012},
number={2},
pages={143--166},
issn={1532-0073},
review={\MR{3007090}},
doi={10.4310/HHA.2012.v14.n2.a9},
}
\bib{HHR}{article}{
author={Hill, M. A.},
author={Hopkins, M. J.},
author={Ravenel, D. C.},
title={On the nonexistence of elements of Kervaire invariant one},
journal={Ann. of Math. (2)},
volume={184},
date={2016},
number={1},
pages={1--262},
issn={0003-486X},
review={\MR{3505179}},
doi={10.4007/annals.2016.184.1.1},
}
\bib{HHR2}{article}{
author={Hill, Michael A.},
author={Hopkins, M. J.},
author={Ravenel, D. C.},
title={The slice spectral sequence for certain $RO(C_{p^n})$-graded
suspensions of $H\underline{\bf Z}$},
journal={Bol. Soc. Mat. Mex. (3)},
volume={23},
date={2017},
number={1},
pages={289--317},
issn={1405-213X},
review={\MR{3633137}},
doi={10.1007/s40590-016-0129-3},
}
\bib{HS}{article}{
author={Heard, Drew},
author={Stojanoska, Vesna},
title={$K$-theory, reality, and duality},
journal={J. K-Theory},
volume={14},
date={2014},
number={3},
pages={526--555},
issn={1865-2433},
review={\MR{3349325}},
doi={10.1017/is014007001jkt275},
}
\bib{HY}{article}{
author={Hill, Michael A.},
author={Yarnall, Carolyn},
title={A new formulation of the equivariant slice filtration with
applications to $C_p$-slices},
journal={Proc. Amer. Math. Soc.},
volume={146},
date={2018},
number={8},
pages={3605--3614},
issn={0002-9939},
review={\MR{3803684}},
doi={10.1090/proc/13906},
}
\bib{U}{article}{
author={Ullman, John},
title={On the slice spectral sequence},
journal={Algebr. Geom. Topol.},
volume={13},
date={2013},
number={3},
pages={1743--1755},
issn={1472-2747},
review={\MR{3071141}},
doi={10.2140/agt.2013.13.1743},
}
\bib{U2}{book}{
author={Ullman, John Richard},
title={On the Regular Slice Spectral Sequence},
note={Thesis (Ph.D.)--Massachusetts Institute of Technology},
publisher={ProQuest LLC, Ann Arbor, MI},
date={2013},
pages={(no paging)},
review={\MR{3211466}},
}
\bib{V}{article}{
author={Voevodsky, Vladimir},
title={Open problems in the motivic stable homotopy theory. I},
conference={
title={Motives, polylogarithms and Hodge theory, Part I},
address={Irvine, CA},
date={1998},
},
book={
series={Int. Press Lect. Ser.},
volume={3},
publisher={Int. Press, Somerville, MA},
},
date={2002},
pages={3--34},
review={\MR{1977582}},
}
\bib{Y}{article}{
author={Yarnall, Carolyn},
title={The slices of $S^n\wedge H\underline{\mathbb{Z}}$ for cyclic
$p$-groups},
journal={Homology Homotopy Appl.},
volume={19},
date={2017},
number={1},
pages={1--22},
issn={1532-0073},
review={\MR{3628673}},
doi={10.4310/HHA.2017.v19.n1.a1},
}
\bib{Z}{article}{
AUTHOR = {Zeng, Mingcong},
TITLE = {Eilenberg-Mac Lane spectra in cyclic \(p\)-groups},
eprint = {https://arxiv.org/abs/1710.01769},
year = {2018},
}
\bib{Z2}{thesis}{
author={Zou, Yan},
title={\(RO(D_{2p})\)-graded Slice Spectral Sequence of \(H\underline{\Z}\)},
note={Thesis (Ph.D.)--University of Rochester},
date={2018},
}
\end{biblist}
\end{bibdiv}
\end{document}
\section{Introduction} \label{sec:intro}
The slice filtration, a filtration of genuine \(G\)-spectra, was developed by Hill, Hopkins, and Ravenel in their solution to the Kervaire invariant-one problem \cite{HHR} and is a generalization of Dugger's filtration \cite{D}. It was modeled after Voevodsky's motivic slice filtration \cite{V} and is an equivariant generalization of the Postnikov tower. Rather than decomposing a \(G\)-spectrum into Eilenberg-Mac Lane specra, as does the Postnikov tower, instead the slice filtration decomposes a \(G\)-spectrum into ``\(n\)-slices''.
There is a complete characterization of all \(n\)-slices where \(-1\leq n\leq 1\), listed in \cref{charac:slice:01-1}. This, combined with \cref{rho:susp:commutes}, characterizes all slices in degrees congruent to \(-1\), 0, or 1, modulo the order of \(G\). For \(G=C_2\times C_2\), we are then only missing the \((4k+2)\)-slices. In \cref{sec:2slice} we finish this characterization with the following result.
\textit{\cref{2slice:characterization}: Suppose the only nontrivial homotopy Mackey functors of a \((C_2\times C_2)\)-spectrum \(X\) are \(\ul\pi_1(X)\) and \(\ul\pi_2(X)\) where certain maps in each Mackey functor are injective. Then \(X\) is a 2-slice. Conversely, if \(X\) is a 2-slice, then its only nontrivial homotopy Mackey functors are \(\ul\pi_1(X)\) and \(\ul\pi_2(X)\) where \(\Si{2} H\ul\pi_2(X)\) is a 2-slice and \(\Si{1} H\ul\pi_1(X) \in [2,4]\).}
Much work has been done computing the slices of certain \(RO(G)\)-graded suspensions of \(H_G\underline{\Z}\) including \(G=C_{p^n}\) by \cite{HHR2} and \cite{Y}, and \mbox{\(G=D_{2p}\)} by \cite{Z2}. \cite{GY} computes the slices of \(\Si{n} H_K \underline{\F}\) where \(K=C_2\times C_2\). Most of these slices are \(RO(K)\)-graded suspensions of \(H\ul\pi_{-i}(\Si{-k\rho_K} H_K\underline{\F})\) for \(i\) in the range \([k+3,4k]\). We primarily focus on the slices of \(\Si{\pm n} H_K\underline{\Z}\). Although we have cofiber sequences relating \(\Si{n} H_K\underline{\Z}\) to \(\Si{n} H_K\underline{\F}\), we can only recover some information about the former from the latter.
As for the slices of \(\Si{\pm n} H_K\underline{\Z}\), the main result can be summarized as follows:
\textit{Main Result: For \(n<0\), all nontrivial slices of \(\Si{n} H_K\underline{\Z}\) are given by:
\begin{align*}
P^{i}_{i}(\Si{n} H_K\underline{\Z}) \simeq \Si{-V} H_K\ul M
\end{align*}
where \(i\) is in the range \([4n,n]\).
For \(0\leq n\leq 5\), \(\Si{n} H\underline{\Z}\) is an \(n\)-slice. Finally, for \(n>5\),
\begin{align*}
P^i_i(\Si{n} H_K\underline{\Z}) \simeq \Si{V} H_K\ul M
\end{align*}
where \(i\) is in the range \([n,4(n-4)]\).
The representations \(V\) and Mackey functors \(\ul M\) are given in \cref{-nslice:K:Z}, \cref{-4kslices:K:Z}, \cref{-4k-2slices:K:Z}, \cref{slices:K:n:Z}, \cref{slices:K:4k:Z}, and \cref{slices:4k+2:K:Z}.}
The paper is organized as follows. In \cref{sec:background}, we review the slice filtration and relevant dualities. The story for \(K\) must restrict to the corresponding results for \(C_2\), so we review these results in \cref{sec:C2}. \cref{sec:K:Mackey} provides us with the main Mackey functors for \(K\) and some pertinent results for \cref{sec:2slice}, in which we characterize all 2-slices over \(K\). We provide some slice towers in \cref{sec:cotowers:K:Z} and describe the slices of \(\Si{-n} H_K\underline{\Z}\) in
\cref{sec:-nslice:K:Z}. In \cref{sec:slices:n:K:Z}, we use Brown-Comenetz duality and the slices of \(\Si{-n} H\underline{\Z}\) to obtain the slices of \(\Si{n} H_K\underline{\Z}\). We then compute the homotopy Mackey functors of the slices of \(\Si{\pm n} H_K\underline{\Z}\) in \cref{sec:htpycomp}. Finally, we provide some examples of the slice spectral sequence for \(\Si{\pm n} H_K\underline{\Z}\) in \cref{sec:SSS}.
The author is grateful for the guidance of Bert Guillou and some helpful conversations with Vigleik Angelveit. Figures \ref{fig:SSS:-1..-5:HZ}, \ref{fig:SSS:-7..-9:HZ}, \ref{fig:SSS:6..8:HZ}, \ref{fig:SSS:10..11:HZ}, \ref{fig:SSS:12:HZ}, and \ref{fig:SSS:14:HZ} were created using Hood Chatham's \texttt{spectralsequences} package.
\section{Background} \label{sec:background}
In this section we give background for the slice filtration as well as Brown-Comenetz and Anderson duality, and \(K\)-representations. Here, except for \cref{subsec:signrep}, \(G\) is any finite group.
\subsection{The Slice Filtration} \label{slice_filt_intro}
\phantom{a}
We start with a brief review of the equivariant slice filtration. For more details see \cite[Section 4]{HHR}.
\begin{defn} \label{def:tau_n}
Let \(\text{Sp}^G\) be the category of genuine \(G\)-spectra. Let \(\tau^G_{\geq n} \subseteq \text{Sp}^G\) be the localizing subcategory generated by \(G\)-spectra of the form \(\Si{\infty}_G G_+ \wedge_H S^{k\rho_H}\) where \(H\leq G\), \(\rho_H\) is the regular representation of \(H\), and \(k\abs{H} \geq n\). We write \(X\geq n\) to mean that \(X\in \tau^G_{\geq n}\).
\end{defn}
\begin{defn} \label{def:lessthan}
We say that \(X<n\) if
\begin{align*}
[S^{k\rho_H+r}, X]^H = 0
\end{align*}
for all \(r\geq 0\) and all subgroups \(H\leq G\) such that \(k\abs{H}\geq n\).
\end{defn}
\begin{thm}{\cite[Corollary 2.9, Theorem 2.10]{HY}} \label{thm:atleastn}
Let \(n\geq 0\). Then \(X\geq n\) if and only if
\begin{align*}
\pi_k(X^H) = 0 \quad \text{for } \quad k<\frac{n}{\abs{H}}.
\end{align*}
\end{thm}
\begin{prop} \label{charac:slice:01-1}
\phantom{a}
\begin{enumerate}
\item \cite[Proposition 4.50]{HHR} \(X\) is a 0-slice if and only if \(X\simeq H\ul M\) for some \(\ul M\in \text{Mack}(G)\).
\item \cite[Proposition 4.50]{HHR} \(X\) is a 1-slice if and only if \(X\simeq \Si{1} H\ul M\) for some \(\ul M\in \text{Mack}(G)\) with injective restrictions.
\item \cite[Theorem 6-4]{U} \(\Si{-1} H\ul M\) is a \((-n)\)-slice iff \(\ul M\) has surjective transfers for \(\abs{H}\geq n\) and \(M(G/H) = 0\) for \(\abs{H}<n\).
\end{enumerate}
\end{prop}
It is important to note that \cite{HHR} uses the original slice filtration whereas we employ the regular slice filtration from \cite{U}. Except for an indexing difference of one, the results are the same.
\begin{prop}{\cite[Corollary 4.25]{HHR}} \label{rho:susp:commutes}
For any \(k\in\mathbb{Z}\),
\begin{align*}
P^{k+\abs{G}}_{k+\abs{G}} (\Si{\rho} X) &\simeq \Si{\rho} P^k_k(X).
\end{align*}
\end{prop}
That is, suspension by the regular representation commutes with the slice filtration.
Given some surjection of groups \(\phi_N:G\rightarrow G/N\) where \(N\unlhd G\), there is a geometric pullback functor \(\phi^*_N : \text{Sp}^{G/N} \rightarrow \text{Sp}^G\) \cite[Definition 4.1]{H}.
\begin{prop}{\cite[Conjecture 4.11]{H}, \cite[Corollary 4-5]{U}} \label{pullback:slice}
Let \(N\unlhd G\). If the \((G/N)\)-spectrum \(X\) is a \(k\)-slice over \(G/N\), then \(\phi^*_N X\) is a \(k[G:N]\)-slice over \(G\).
\end{prop}
\subsection{Brown-Comenetz and Anderson Duality} \label{BCandA:duality}
\phantom{A}
As in \cite{HS}, we write \(I_{\mathbb{Q}/\mathbb{Z}}\) to indicate the representing \(G\)-spectrum of the cohomology theory \(X\mapsto \text{Hom}(\pi_{-\ast}^G X, \mathbb{Q}/\mathbb{Z})\). The Brown-Comenetz dual of \(X\) is then defined to be the function \(G\)-spectrum \(F(X, I_{\mathbb{Q}/\mathbb{Z}})\). Similarly, \(I_{\mathbb{Q}}\) represents \(X\mapsto \text{Hom}(\pi_{-\ast} X, \mathbb{Q})\) and \(I_\mathbb{Q} X = F(X, I_{\mathbb{Q}})\). Finally, the Anderson dual of \(X\) is \(I_{\mathbb{Z}} X = F(X,I_\mathbb{Z})\), where \(I_\mathbb{Z}\) is the fiber of the natural map \(I_{\mathbb{Q}} \rightarrow I_{\mathbb{Q}/\mathbb{Z}}\).
\begin{exmp}{\cite[Section 3A]{GM}, \cite{HS}}
For a torsion Eilenberg-MacLane spectrum \(H\ul M\),
\begin{align*}
I_\mathbb{Z} H\ul M &\simeq \Si{-1} H\ul M^*
\end{align*}
and
\begin{align*}
I_{\mathbb{Q}/\mathbb{Z}} H\ul M &\simeq H\ul M^*.
\end{align*}
One should note that \cite{HS} deals with non-equivariant spectra and \cite[Section 3A]{GM} refers specifically to \(H\ul M\) as an \(\mathbb{F}_2\)-torsion spectrum. This example, however, follows easily from their work and \cite[Corollary I.7.3]{U2}. See \cite[Section 3A, Section 3B]{GM} for a more detailed discussion of equivariant Anderson duality.
\end{exmp}
\begin{prop}(\cite[Theorem I.7.7, Theorem I.7.8]{U2}) \label{duality:prop}
For a spectrum \(X\),
\begin{align*}
X\geq n &\Leftrightarrow I_{\mathbb{Q}/\mathbb{Z}} X\leq -n.
\end{align*}
In particular,
\begin{align*}
P^n_k I_{\mathbb{Q}/\mathbb{Z}} X &\cong I_{\mathbb{Q}/\mathbb{Z}} P^{-k}_{-n} X.
\end{align*}
\end{prop}
That is, the Brown-Comenetz dualization functor dualizes slice status.
\subsection{\texorpdfstring{\(K\)}{K4}-Representations} \label{subsec:signrep}
\phantom{a}
Recall that \(\sigma\) is the one-dimensional sign representation of \(C_2\). The sign representations of \(K\) are then the pullbacks \(p_1^*\sigma\), \(m^*\sigma\), and \(p_2^*\sigma\) where \(p_1\), \(m\), and \(p_2\) are
\begin{equation*}
\begin{tikzcd}[ampersand replacement=\&]
\& C_2 \arrow[dr, "\sigma"] \\
C_2\times C_2 \arrow[ur, "p_1"] \arrow[r, "m"] \arrow[dr, "p_2"] \& C_2 \arrow[r, "\sigma"] \& \mathbb{R} \\
\& C_2 \arrow[ur, "\sigma"] \&
\end{tikzcd}
\end{equation*}
As in \cite{GY}, we will write
\begin{align*}
\alpha=p_1^*\sigma, \qquad \beta=p_2^*\sigma, \qquad \text{and } \qquad \gamma=m^*\sigma.
\end{align*}
The regular representation of \(K\) is then \(\rho_{K} = 1 + \alpha + \beta + \gamma\).
\section{Review of \texorpdfstring{\(C_2\)}{C2}} \label{sec:C2}
The Lewis diagram for a \(C_2\)-Mackey functor takes the form
\begin{equation*}
\begin{tikzcd}
\ul M(C_2) \arrow[dd, color = black, bend right = 20] \\ \\
\ul M(e) \arrow[uu, color = black, bend right = 20]
\end{tikzcd}
\end{equation*}
where \(\ul M(e)\) has a \(C_2/e\) action. Arrows going down the diagram are called restrictions and arrows going up are called transfers.
For the sake of clarity in large diagrams, in a general Mackey functor \(\ul M\) we will henceforward denote \(\ul M(H)\) by \(M_H\).
\begin{prop} \label{C2:2slice}
A \(C_2\)-spectrum \(Y\) is a 2-slice over \(C_2\) if and only if the only nontrivial homotopy Mackey functors are \(\ul\pi_1(Y)\) and \(\ul\pi_2(Y)\), where
\begin{enumerate}
\item The restriction map \(\res{C_2}{e}: \pi_2^{C_2}(Y) \rightarrow \pi_2^e(Y)\) is injective \label{2slice:C2:M}
\item \(\pi_1^e(Y) = 0\). \label{2slice:C2:N}
\end{enumerate}
That is, both \(\Si{2} H\ul\pi_2(Y)\) and \(\Si{1} \ul\pi_1(Y)\) are 2-slices.
\end{prop}
\begin{proof}
Let \(\ul M=\ul\pi_2(Y)\) and \(\ul N=\ul\pi_1(Y)\) and suppose that they are the only nontrivial homotopy Mackey functors of \(Y\). We show that \(\Si{2} H\ul M\) and \(\Si{1} H\ul N\) are 2-slices if and only if \cref{2slice:C2:M} and \cref{2slice:C2:N} hold.
First, consider \(\Si{2} H\ul M\). We immediately see that this spectrum is at least two. As for \(\Si{2} H\ul M\leq 2\), the cofiber sequence
\begin{align*}
(S^{-\sigma} \rightarrow S^0 \rightarrow C_2/e_+) \wedge \Si{2} H\ul M
\end{align*}
provides the homotopy
\begin{center}
\begin{tikzpicture}
\node at (-4.25,1) {\(\ul\pi_2 (\Si{-\sigma} H\ul M) \quad =\)};
\node (lMC) at (-1.5,2) {\(\ker(\res{C_2}{e})\)};
\node (lMe) at (-1.5,0) {\(0\)};
%
\node at (0,1) {and};
%
\node at (2.5,1) {$\ul\pi_1 (\Si{-\sigma} H\ul M) \quad =$};
\node (MC) at (5,2) {$M_e / \text{im } \res{C_2}{e}$};
\node (Me) at (5,0) {$\mathbb{Z}_\sigma \otimes M_e$};
%
\arres{MC}{Me}{0}{0}{}
\artr{Me}{MC}{0}{0}{}
\end{tikzpicture}
\end{center}
Then the cofiber sequences
\begin{align*}
(S^{-k\sigma} \rightarrow S^{-(k-1)\sigma} \rightarrow C_2/e_{+}\wedge S^{-(k-1)\sigma})\wedge \Si{2} H\ul M
\end{align*}
reveal that \(\ul\pi_2^{C_2}(\Si{-k\sigma} \Si{2} H\ul M) = \ker(\res{C_2}{e})\) for all \(k\geq 2\).
By definition, \(\Si{2} H\ul M\leq 2\) if and only if
\(\ul\pi_2^{C_2}(\Si{-k\sigma} \Si{2} H\ul M) = 0\) for all \(k\geq 2\). Consequently, \(\Si{2} H\ul M\) is a 2-slice if and only if \(\res{C_2}{e}\) is injective .
As for \(\Si{1} H\ul N\), by \cref{thm:atleastn}, we find that \(Y\geq 2\) if and only if \(N_{e} = 0\). But then \(\Si{1} H\ul N\) is the pullback \(\phi^*_K \Si{1} HN_{e}\) and consequently, a 2-slice by \cref{pullback:slice}.
Conversely, assume \(Y\) is a 2-slice. By \cref{thm:atleastn}, we know \(\ul\pi_1^e(Y) = 0\) and \(\ul \pi_k(Y) = \ul 0\) for \(k\leq 0\). For \(k\geq 3\), the slice status of \(Y\) and \cref{def:lessthan} dictate that \(\ul \pi_k(Y) = \ul 0\). Consequently, the only nontrivial homotopy Mackey functors of \(Y\) and \(\ul\pi_1(Y)=\ul N\) and \(\ul\pi_2(Y)=\ul M\). Thus, we have a fiber sequence \(\Si{2} H\ul M\rightarrow Y\rightarrow \Si{1} H\ul N\).
If either \(\Si{2} H\ul M\) or \(\Si{1} H\ul N\) is trivial, the result follows from above. So assume that both are nontrivial. Because \(N_e = 0\), \(\Si{1} H\ul N\) is a 2-slice. In particular, \(\Si{1} H\ul N\leq 2\), and since \(Y\leq 2\), we have that \(\Si{2} H\ul M\leq 2\). Consequently, as \(\Si{2} H\ul M\geq 2\), we have that \(\Si{2} H\ul M\) is a 2-slice.
\end{proof}
From \cite{GY}, we will see the \(C_2\)-Mackey functors in \cref{tab-oldC2Mackey}.
\begin{table}[h]
\caption[Familiar {$C_2$}-Mackey functors.]{Familiar $C_2$-Mackey functors.}
\label{tab-oldC2Mackey}
\begin{center}
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c|c|}
\hline
$\underline{\Z}$
& $\underline{\Z}^*$
& ${\underline{\widehat{f}}}$
\\ \hline
$\MackC{\mathbb{Z}}{\mathbb{Z}}{blue}{orange}{1}{2}$
& $\MackC{\mathbb{Z}}{\mathbb{Z}}{orange}{blue}{2}{1}$
& $\MackC{0}{\mathbb{Z}_\sigma}{white}{white}{}{}$
\\
\hline
$\underline{\F}$
& $\underline{\F}^*$
& $\ul f$
\\ \hline
$\MackC{\mathbb{F}_2}{\mathbb{F}_2}{black}{white}{1}{}$
& $\MackC{\mathbb{F}_2}{\mathbb{F}_2}{white}{black}{}{1}$
& $\MackC{0}{\mathbb{F}_2}{white}{white}{}{}$
\\
\hline
$\underline{g}$ && \\
$\MackC{\mathbb{F}_2}{0}{white}{white}{}{}$ && \\ \hline
\end{tabular} }
\end{center}
\end{table}
\begin{prop}[\cite{GY}] \label{C2:equiv}
There are equivalences
\begin{enumerate}
\item \(\Si{4} H_{C_2}\underline{\Z} \simeq \Si{2\rho} H_{C_2}\underline{\Z}^*\)
\item \(\Si{2} H_{C_2}\ul f \simeq \Si{\rho} H_{C_2}\underline{\F}^*\)
\item \(\Si{1} H_{C_2}\underline{g} \simeq \Si{\rho} H_{C_2}\underline{g}\) \label{C2:g:2nslice}
\end{enumerate}
\end{prop}
Note, in particular, that \cref{C2:g:2nslice} makes \(\Si{n} H_{C_2}\underline{g}\) a \(2n\)-slice for any \(n\in \mathbb{Z}\).
\begin{prop} \label{ksigma+r:homotopy:C2:Z}
For \(k,r\geq 0\),
\begin{align*}
\ul \pi_i (\Si{k\sigma + r} H_{C_2}\underline{\Z}) &= \left\lbrace \begin{array}{ll}
\underline{\Z} & i=k+r, \text{ } k \text{ even} \\
{\underline{\widehat{f}}} & i=k+r, \text{ } k \text{ odd} \\
\underline{g} & i\in[r,k+r-1], \text{ } i\equiv r\pmod 2
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
We calculate \(\ul \pi_i (\Si{k\sigma} H_{C_2}\underline{\Z})\) and then shift the degrees by \(r\). The result follows by induction on \(j\geq 1\) using the cofiber sequence
\begin{align*}
\Si{(j-1)\sigma+2} H\underline{\Z} \simeq \Si{j\sigma} H\underline{\Z}^* \rightarrow \Si{j\sigma} H\underline{\Z} \rightarrow \Si{j\sigma} H\underline{g}\simeq H\underline{g}.
\end{align*}
\end{proof}
\subsection{Slices of \texorpdfstring{\(\Si{\pm n} H_{C_2}\underline{\Z}\)}{HZ}} \label{towers:C2:Z:V} \hspace{2in}
Because the slices for \(\Si{\pm n} H_{K}\underline{\Z}\) over \(K\) restrict to the corresponding slices over \(C_2\), we must know these slices over \(C_2\). They are as follows.
\begin{prop} \label{C2:Z:slice:0-6}
\(\Si{n} H_{C_2}\underline{\Z}\) is an \(n\)-slice for \(0\leq n\leq 6\).
\end{prop}
\begin{proof}
For \(0\leq n\leq 2\), this follows from \cref{charac:slice:01-1} and \cref{C2:2slice}. By the same results, \(\Si{n} H\underline{\Z}^*\) is an \(n\)-slice for \(0\leq n\leq 2\). Furthermore, \(\Si{-1} H\underline{\Z}^*\) is a \((-1)\)-slice by \cref{charac:slice:01-1}.
Then, by \cref{C2:2slice} and \cref{rho:susp:commutes},
\(\Si{n} H\underline{\Z} \simeq \Si{n-4+2\rho} H\underline{\Z}^*\) is a \(n\)-slice for \(3\leq n\leq 6\).
\end{proof}
\begin{prop} \label{tower:C2:Z:n}
Let \(n\geq 7\) and take \(r\equiv n\pmod 4\) with \(3\leq r\leq 6\). The slice tower of \(\Si{n} H_{C_2}\underline{\Z}\) is
\begin{equation*}
\begin{tikzcd}
\arrow[r] P^{2n-6}_{2n-6} = \Si{n-3} H\underline{g} &
\arrow[d] \Si{n} H\underline{\Z} \\
\arrow[r] P^{2n-10}_{2n-10} = \Si{n-5} H\underline{g} &
\arrow[d] \Si{2\rho + (n-4)} H\underline{\Z} \\
& \arrow[d] \vdots \\
\arrow[r] P^{n+r-2}_{n+r-2} = \Si{\frac{n+r}{2}-1} H\underline{g} & \arrow[d] \Si{\frac{n-r-4}{2}\rho+(r+4)} H\underline{\Z} \\
& P^n_n = \Si{\frac{n-r}{2}\rho + r} H\underline{\Z}
\end{tikzcd}
\end{equation*}
\end{prop}
\begin{proof}
The exact sequence \(\underline{\Z}^* \rightarrow \underline{\Z} \rightarrow \underline{g}\) and the homotopy equivalence \(\Si{n} H\underline{\Z} \simeq \Si{n-4+2\rho} H\underline{\Z}^*\) provide the fiber sequence
\begin{align*}
\Si{n-3} H\underline{g} \rightarrow \Si{n} H\underline{\Z} \rightarrow \Si{2\rho + (n-4)} H\underline{\Z}.
\end{align*}
We then augment this sequence with its appropriate \(2\rho\) suspensions until \(\Si{\frac{n-r}{2}\rho + r} H\underline{\Z}\), a slice, is reached.
\end{proof}
\begin{cor} \label{2kslice:C2:HZ}
Let \(n\geq 7\) and take \(r\equiv n\pmod 4\) with \(3\leq r\leq 6\). The \((2k)\)-slices of \(\Si{n} H_{C_2}\underline{\Z}\) are
\begin{align*}
P^{2k}_{2k}( \Si{n} H_{C_2}\underline{\Z} ) &\simeq \Si{k} H_{C_2}\underline{g}
\end{align*}
for \(k\equiv n+1\pmod 2\) and \(k\in \left[\frac{n+r}{2}-1,\ldots, n-3\right]\).
\end{cor}
\begin{prop} \label{tower:C2:Z:-n}
Let \(n\geq 1\) and take \(r\equiv n\pmod 4\) with \(1\leq r\leq 4\). The slice cotower of \(\Si{-n} H_{C_2}\underline{\Z}\) is
\begin{equation*}
\begin{tikzcd}
\arrow[d] \Si{-\frac{n-r}{2}\rho-r} H\underline{\Z}^* = P^{-n}_{-n} & \\
\arrow[d] \arrow[r] \Si{-\frac{n-r}{2}\rho -r} H\underline{\Z} & \Si{-\frac{n+r}{2}} H\underline{g} = P^{-n-r}_{-n-r} \\
\arrow[d] \vdots & \\
\arrow[d] \arrow[r] \Si{-2\rho -n+4} H\underline{\Z} & \Si{-n+2} H\underline{g} = P^{-2n+4}_{-2n+4} \\
\arrow[r] \Si{-n} H\underline{\Z} & \Si{-n} H\underline{g} = P^{-2n}_{-2n}
\end{tikzcd}
\end{equation*}
\end{prop}
\begin{proof}
The exact sequence \(\underline{\Z}^* \rightarrow \underline{\Z} \rightarrow \underline{g}\) and the homotopy equivalence \(\Si{-n} H\underline{\Z} \simeq \Si{2\rho-n-4} H\underline{\Z}\) provide the cofiber sequence
\begin{align*}
\Si{-2\rho-n+4} H\underline{\Z} \rightarrow \Si{-n} H\underline{\Z} \rightarrow \Si{-n} H\underline{g}.
\end{align*}
We then augment this sequence with its appropriate \(-2\rho\) suspensions until \(\Si{-\frac{n-r}{2}\rho-r} H\underline{\Z}\), a slice, is reached.
\end{proof}
\begin{cor} \label{-2kslice:C2:HZ}
Let \(n\geq 1\) and take \(r\equiv n\pmod 4\) with \(1\leq r\leq 4\). The \((-2k)\)-slices of \(\Si{-n} H_{C_2}\underline{\Z}\) are
\begin{align*}
P^{-2k}_{-2k}( \Si{-n} H_{C_2}\underline{\Z} ) &\simeq \Si{-k} H_{C_2}\underline{g}
\end{align*}
for \(k\equiv n\pmod 2\) and \(k\in \left[\frac{n+r}{2},n\right]\).
\end{cor}
\section{\texorpdfstring{\(K\)}{K}-Mackey Functors \label{sec:K:Mackey} \hspace{2in}}
Recall that \(K=C_2\times C_2\) and let \(L\), \(D\), and \(R\) be the left, diagonal, and right subgroups of \(K\), respectively. The Lewis diagram for a \(K\)-Mackey functor takes the form
\begin{center}
\begin{tikzpicture}
\node (MK) at (0,4) {$M_K$};
\node (ML) at (-2,2) {$M_L$};
\node (MD) at (0,2) {$M_D$};
\node (MR) at (2,2) {$M_R$};
\node (Me) at (0,0) {$M_e$};
\arres{MK}{ML}{0}{0}{}
\arres{MK}{MD}{0}{0}{}
\arres{MK}{MR}{0}{0}{}
\artr{ML}{MK}{0}{0}{}
\artr{MD}{MK}{0}{0}{}
\artr{MR}{MK}{0}{0}{}
\arres{ML}{Me}{0}{0}{}
\arres{MD}{Me}{0}{0}{}
\arres{MR}{Me}{0}{0}{}
\artr{Me}{ML}{0}{0}{}
\artr{Me}{MD}{0}{0}{}
\artr{Me}{MR}{0}{0}{}
\end{tikzpicture}
\end{center}
with a Weyl group action \(W_G(H) \circlearrowright M_H\) at each level. If we do not indicate the Weyl group actions for a specific Mackey functor, then they are trivial.
A number of Mackey functors from \cite{GY} will make their appearance.
\begin{table}[h]
\caption[Familiar {$K$}-Mackey functors.]{Familiar $K$-Mackey functors.}
\label{tab-oldKMackey}
\begin{center}
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c|}
\hline
$\ul {mg}$
& $\ul {m}$
\\ \hline
$\MackKupperR{\mathbb{F}_2^2}{\mathbb{F}_2}{black}{p_1}{\nabla}{p_2}$
& $\MackKupperR{\mathbb{F}_2}{\mathbb{F}_2}{black}{1}{1}{1}$
\\
\hline
$\phi_{LDR}^*\underline{\F}$
& $\underline{g}^n$
\\ \hline
$\MackKupperR{\mathbb{F}_2^3}{\mathbb{F}_2}{black}{p_1}{p_2}{p_3}$
& $\MackKgn{\mathbb{F}_2^n}$
\\
\hline
\end{tabular} }
\end{center}
\end{table}
We will also see the duals of the Mackey functors in \cref{tab-oldKMackey}.
\begin{prop}(\cite[Proposition 4.8]{GY}) \label{m:mg:equiv}
There are equivalences
\begin{align}
\Si{-\rho} H_K\ul m &\simeq \Si{-2} H_K\ul{mg}^* \\
\Si{-2} H_K\ul m^* &\simeq \Si{-\rho} H_K\ul{mg}
\end{align}
\end{prop}
We will also see the new Mackey functors in \cref{tab-newKMackey}.
\begin{table}[h]
\caption[{New $K$}-Mackey functors.]{New $K$-Mackey functors.}
\label{tab-newKMackey}
\begin{center}
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c|}
\hline
$\underline{\Z}$
& $\underline{\Z}^*$
\\ \hline
$\MackKall{\mathbb{Z}}{blue}{blue}{orange}{orange}$
& $\MackKall{\mathbb{Z}}{orange}{orange}{blue}{blue}$
\\ \hline
$\underline{\Z}(2,1)$
& $\underline{\Z}(2,1)^*$
\\ \hline
$\MackKall{\mathbb{Z}}{orange}{blue}{blue}{orange}$
& $\MackKall{\mathbb{Z}}{blue}{orange}{orange}{blue}$
\\ \hline
$\underline{\mathbb{M}}$ & \\ \hline
${\begin{tikzcd}[ampersand replacement=\&, bend right=2.5ex, column sep=scriptsize]
\& \mathbb{Z}/4 \ar[dl, color=blue] \ar[d, color=blue] \ar[dr,color=blue] \& \\
\mathbb{Z}/2 \ar[ur, color=orange] \& \mathbb{Z}/2 \ar[u, color=orange] \& \mathbb{Z}/2 \ar[ul, color=orange] \\
\& 0 \&
\end{tikzcd}}$
& \\ \hline
\end{tabular} }
\end{center}
\end{table}
In \cref{tab-newKMackey}, \(\underline{\Z}^*\) is the dual to the constant Mackey functor \(\underline{\Z}\) and \(\underline{\Z}(2,1)^*\) is the dual to \(\underline{\Z}(2,1)\), so named for consistency with \cite{Y}. The Mackey functor \(\underline{\mathbb{M}}\) results from the injection \(\underline{\Z}^*\hookrightarrow \underline{\Z} \twoheadrightarrow \underline{\mathbb{M}}\). In each Mackey functor, the blue arrows are multiplication by one and the orange arrows are multiplication by two.
\begin{prop} \label{equivalence:ZZ*}
We have the equivalence
\begin{align*}
\Si{-\rho} H_K\underline{\Z} &\simeq \Si{-4} H_K\underline{\Z}^*.
\end{align*}
\end{prop}
\begin{proof}
This follows from \cref{homotopy:-rho:M}.
\end{proof}
\subsection{Induced Mackey Functors} \label{Induction_Formula}
\phantom{a}
We will now give an explicit description of \(K\)-Mackey functors induced from the cyclic subgroups.
We recall the following standard result.
\begin{lem}[Shearing isomorphism] \label{shearingiso}
Let \(M\) be a \(\mathbb{Z}[G]\)-module and \(H\) a subgroup of \(G\). Then we have an isomorphism of \(\mathbb{Z}[G]\)-modules,
\begin{center}
\begin{tikzpicture}
%
\node (L) at (-1.5,0) {$\mathbb{Z}[G]\otimes_{H} M$};
\node (M) at (0,0) {$\cong$};
\node (R) at (1.5,0) {$\mathbb{Z}[G/H]\otimes M$};
\node (loop1) at (-.8,0) {$\phantom{A}$};
\node (loop2) at (2.3,0) {$\phantom{A}$};
%
\draw[->,out=240,in=300,loop, looseness=4] (loop1) to
node[below] {\scalebox{0.6}{$H$}} (loop1);
\draw[->,out=240,in=300,loop, looseness=4] (loop2) to node[below] {\scalebox{0.6}{$G$}} (loop2);
\end{tikzpicture}
\end{center}
where \(G\) acts on the left of \(\mathbb{Z}[G]\) and diagonally on \(\mathbb{Z}[G/H]\otimes M\) via the map \(g\otimes m \mapsto gH \otimes g\cdot m\).
\end{lem}
Let \(\ul N\) be a \(C_2\)-Mackey functor. Take \(h\inK\) and set \(\Delta_h:N_e \rightarrow \mathbb{Z}[K] \otimes_{\mathbb{Z}[L]} N_e\) and \(\nabla_h:\mathbb{Z}[K] \otimes_{\mathbb{Z}[L]} N_e \rightarrow N_e\) to be \(n\mapsto (e+h)\otimes n\) and \(e\otimes n,h\otimes n\mapsto n\), respectively.
Then \(\uparrow^K_L \ul N\) is
\begin{equation*}
\raisebox{12ex}{$\uparrow^K_L \ul N \quad \cong$} \quad
\begin{tikzpicture}
\node (MK) at (0,4) {$N_{C_2}$};
\node (ML) at (-2,2) {$\mathbb{Z}[K/L] \otimes N_{C_2}$};
\node (MD) at (0,2) {$N_e$};
\node (MR) at (2,2) {$N_e$};
\node (Me) at (0,0) {$\mathbb{Z}[K] \otimes_{\mathbb{Z}[L]} N_e$};
\arres{MK}{ML}{0}{0}{$\Delta$}
\arres{MK}{MD}{-6}{-16}{$\res{C_2}{e}$}
\arres{MK}{MR}{-3}{-15}{$\res{C_2}{e}$}
\artr{ML}{MK}{0}{0}{$\nabla$}
\artr{MD}{MK}{0}{0}{$\tr{C_2}{e}$}
\artr{MR}{MK}{0}{0}{$\tr{C_2}{e}$}
\draw[->,black,bend right=2ex] (ML) to node[sloped, anchor=center, xshift={0pt},yshift={-5pt}] {\scalebox{0.5}{$\text{id}\otimes \res{C_2}{e}$}} (Me);
\arres{MD}{Me}{-5}{0}{$\Delta_d$}
\arres{MR}{Me}{0}{0}{$\Delta_r$}
\draw[->,black,bend right=2ex] (Me) to node[sloped, anchor=center, xshift={0pt},yshift={-5pt}] {\scalebox{0.5}{$\text{id} \otimes \tr{C_2}{e}$}} (ML);
\artr{Me}{MD}{0}{0}{$\nabla_d$}
\artr{Me}{MR}{0}{0}{$\nabla_r$}
\end{tikzpicture}
\end{equation*}
The Weyl group \(W_K(D)\) acts on \(\uparrow_L^K \ul{N}(D)= N_e\) via the isomorphism \(W_K(D)\cong C_2\) and the given action of \(C_2 = W_{C_2}(e)\) on \(N_e\). Similarly for the action of \(W_{K}(R)\) on \(N_e\). As for \(L\), \(W_K(L)\) acts only on \(\mathbb{Z}[K/L]\). Finally, \(W_K(e) = K\) acts on the \(\mathbb{Z}[K]\) factor of \(\mathbb{Z}[K]\otimes_{\mathbb{Z}[L]} N_e\).
We are now going to define the unit map \(\ul{M}\rightarrow \uparrow^K_{L} \downarrow^K_L \ul M\). The pullback along \(K\twoheadrightarrow K/L\cong C_2\) of \(\mathbb{Z}\hookrightarrow \mathbb{Z}[C_2]\) is an inclusion \(\mathbb{Z}\hookrightarrow \mathbb{Z}[K/L]\). Tensoring with \(M_e\) gives
\begin{equation*}
\begin{tikzcd}
M_e \arrow[r, hook, "i_L"] & \mathbb{Z}[K/L]\otimes M_e.
\end{tikzcd}
\end{equation*}
We will also use \(i_L\) to denote the inclusion of \(K/L\) fixed points
\begin{equation*}
\begin{tikzcd}
M_L \arrow[r, hook, "i_L"] & \mathbb{Z}[K/L]\otimes M_L.
\end{tikzcd}
\end{equation*}
Consider
\begin{align*}
\nabla^r: \mathbb{Z}[K/L]\otimes M_e \rightarrow M_e \qquad \text{and} \qquad \Delta^r:M_e \rightarrow \mathbb{Z}[K/L]\otimes M_e.
\end{align*}
Let \(\nabla^r\) be the action map and define \(\Delta^r\) by \(m\mapsto e\otimes m + r\otimes r\cdot m\). Then, for \(\ul M\in \text{Mack}(K)\), the map \(\ul M \rightarrow \uparrow^K_{L} \downarrow^K_L \ul M\) is
\begin{equation*}
\begin{tikzpicture}
\node (MK) at (0,4) {$M_K$};
\node (ML) at (-2,2) {$M_L$};
\node (MD) at (0,2) {$M_D$};
\node (MR) at (2,2) {$M_R$};
\node (Me) at (0,0) {$M_e$};
%
\node (rMK) at (7,4) {$M_L$};
\node (rML) at (5,2) {$\mathbb{Z}[K/L] \otimes M_L$};
\node (rMD) at (7,2) {$M_e$};
\node (rMR) at (9,2) {$M_e$};
\node (rMe) at (7,0) {$\mathbb{Z}[K/L] \otimes M_e$};
\arres{MK}{ML}{0}{0}{}
\arres{MK}{MD}{0}{0}{}
\arres{MK}{MR}{0}{0}{}
\artr{ML}{MK}{0}{0}{}
\artr{MD}{MK}{0}{0}{}
\artr{MR}{MK}{0}{0}{}
\arres{ML}{Me}{0}{0}{}
\arres{MD}{Me}{0}{0}{}
\arres{MR}{Me}{0}{0}{}
\artr{Me}{ML}{0}{0}{}
\artr{Me}{MD}{0}{0}{}
\artr{Me}{MR}{0}{0}{}
\arres{rMK}{rML}{0}{0}{$\Delta^r$}
\arres{rMK}{rMD}{0}{0}{}
\arres{rMK}{rMR}{0}{0}{}
\artr{rML}{rMK}{0}{0}{$\nabla^r$}
\artr{rMD}{rMK}{0}{0}{}
\artr{rMR}{rMK}{0}{0}{}
\draw[->,black,bend right=2ex] (rML) to node[sloped, anchor=center, xshift={0pt},yshift={-5pt}] {\scalebox{0.7}{$\text{id}\otimes r\cdot \res{L}{e}$}} (rMe);
\arres{rMD}{rMe}{-5}{2}{$\Delta^d$}
\arres{rMR}{rMe}{0}{0}{$\Delta^r$}
\draw[->,black,bend right=2ex] (rMe) to node[sloped, anchor=center, xshift={0pt},yshift={5pt}] {\scalebox{0.7}{$\text{id}\otimes r\cdot \tr{L}{e}$}} (rML);
\artr{rMe}{rMD}{0}{0}{$\nabla^d$}
\artr{rMe}{rMR}{0}{0}{$\nabla^r$}
%
\draw[->] (MK) to node[above, xshift={-1pt}] {\scalebox{0.7}{$\res{K}{L}$}} (rMK);
\draw[->] (MR) to node[above, xshift={-1pt}] {\scalebox{0.7}{$\begin{pmatrix}i_L \\ \res{D}{e} \\ \res{R}{e} \end{pmatrix}$}} (rML);
\draw[->] (Me) to node[above, xshift={-1pt}] {\scalebox{0.7}{$i_L$}} (rMe);
\end{tikzpicture}
\end{equation*}
Again, \(W_K(D)\) acts on \(M_e\) via the \(C_2\)-action \(W_{C_2}(e)\circlearrowright M_e\). The same applies for \(W_K(R)\circlearrowright M_e\). Note we have used \cref{shearingiso} to rewrite the bottom group in \(\uparrow^K_L \downarrow^K_L \ul M\). We now have a diagonal action \(K/L\) on \(\mathbb{Z}[K/L]\otimes M_L\). Similarly, \(K/e\) acts diagonally on \(\mathbb{Z}[K/L]\otimes M_e\).
\begin{exmp}
For \(\ul M=\underline{\Z}\), $\uparrow_L^K \underline{\Z}$ and $\underline{\Z} \rightarrow \uparrow_L^K \underline{\Z}$ are as follows.
\begin{equation*}
\begin{tikzpicture}
\node (MK) at (0,4) {$\mathbb{Z}$};
\node (ML) at (-2,2) {$\mathbb{Z}$};
\node (MD) at (0,2) {$\mathbb{Z}$};
\node (MR) at (2,2) {$\mathbb{Z}$};
\node (Me) at (0,0) {$\mathbb{Z}$};
%
\node (rMK) at (7,4) {$\mathbb{Z}$};
\node (rML) at (5,2) {$\mathbb{Z}[K/L] \otimes \mathbb{Z}$};
\node (rMD) at (7,2) {$\mathbb{Z}$};
\node (rMR) at (9,2) {$\mathbb{Z}$};
\node (rMe) at (7,0) {$\mathbb{Z}[K/L] \otimes \mathbb{Z}$};
\draw[->,blue,bend right=2ex] (MK) to (ML);
\draw[->,blue,bend right=2ex] (MK) to (MD);
\draw[->,blue,bend right=2ex] (MK) to (MR);
\draw[->,orange,bend right=2ex] (ML) to (MK);
\draw[->,orange,bend right=2ex] (MD) to (MK);
\draw[->,orange,bend right=2ex] (MR) to (MK);
\draw[->,blue,bend right=2ex] (ML) to (Me);
\draw[->,blue,bend right=2ex] (MD) to (Me);
\draw[->,blue,bend right=2ex] (MR) to (Me);
\draw[->,orange,bend right=2ex] (Me) to (ML);
\draw[->,orange,bend right=2ex] (Me) to (MD);
\draw[->,orange,bend right=2ex] (Me) to (MR);
\arres{rMK}{rML}{0}{0}{$\Delta^r$}
\draw[->,blue,bend right=2ex] (rMK) to (rMD);
\draw[->,blue,bend right=2ex] (rMK) to (rMR);
\artr{rML}{rMK}{0}{0}{$\nabla^r$}
\draw[->,orange,bend right=2ex] (rMD) to (rMK);
\draw[->,orange,bend right=2ex] (rMR) to (rMK);
\draw[->,blue,bend right=2ex] (rML) to node[sloped, anchor=center, xshift={0pt},yshift={-5pt}] {\scalebox{0.7}{$\text{id}\otimes r\cdot 1$}} (rMe);
\arres{rMD}{rMe}{-5}{2}{$\Delta^d$}
\arres{rMR}{rMe}{0}{0}{$\Delta^r$}
\draw[->,orange,bend right=2ex] (rMe) to node[sloped, anchor=center, xshift={0pt},yshift={5pt}] {\scalebox{0.7}{$\text{id}\otimes r\cdot 2$}} (rML);
\artr{rMe}{rMD}{0}{0}{$\nabla^d$}
\artr{rMe}{rMR}{0}{0}{$\nabla^r$}
%
\draw[->] (MK) to node[above, xshift={-1pt}] {\scalebox{0.7}{$1$}} (rMK);
\draw[->] (MR) to node[above, xshift={-1pt}] {\scalebox{0.7}{$\begin{pmatrix}i_L \\ 1 \\ 1 \end{pmatrix}$}} (rML);
\draw[->] (Me) to node[above, xshift={-1pt}] {\scalebox{0.7}{$i_L$}} (rMe);
\end{tikzpicture}
\end{equation*}
Again, blue denotes multiplication by one and orange multiplication by two.
\end{exmp}
\section{2-slice Characterization \hspace{2in}} \label{sec:2slice}
Consider a \(K\)-spectrum \(X\). Then by \cref{rho:susp:commutes},
\begin{align*}
P^n_n X &\simeq \Si{\frac{n-r}{4}\rho} P^r_r \left( \Si{-\frac{n-r}{4}\rho} X \right)
\end{align*}
where \(n\equiv r\pmod 4\) and \(0\leq r\leq 3\). Thus, to know the slices of \(X\), we need only know the 0-, 1-, 2-, and 3-slices of certain suspensions of \(X\). \cref{charac:slice:01-1} and \cref{rho:susp:commutes} characterize the 0-, 1-, and 3-slices. We now characterize the 2-slices.
\begin{prop} \label{Homotopy_Reduction}
Let \(G\) be a finite group, \(H\) an index two subgroup of \(G\), and \(\sigma^H\) the sign representation from \(G\rightarrow G/H\).
For a \(G\)-spectrum \(X\), if \(\downarrow^G_H \ul \pi_{n+1}(X) = \ul 0 = \downarrow^G_H \ul \pi_n(X)\), the natural map \(\Si{-\sigma^H} X \rightarrow X\) induces an isomorphism on \(\ul\pi_{n}\). In particular, if \(\pi_{n+1}^H(X) = 0 = \pi_n^G(X)\), then \(\pi_n^G(\Si{-\sigma^H} X) = 0\).
\end{prop}
\begin{proof}
Because \(\sigma^H\) is one-dimensional we may construct \(\Si{-\sigma^H} X\) with the cofiber sequence \((S^{-\sigma^H} \rightarrow S^0 \rightarrow G/H_+) \wedge X\). The resulting long exact sequence in homotopy is
\begin{align} \label{les:reduction}
\uparrow^G_H \downarrow^G_H \ul\pi_{n+1}(X) \rightarrow \ul\pi_n(\Si{-\sigma^H} X) \rightarrow \ul\pi_n(X) \rightarrow \uparrow^G_H \downarrow^G_H \ul\pi_{n}(X).
\end{align}
As \(\uparrow^G_H \downarrow^G_H \ul\pi_{n+1}(X) = \ul 0 = \uparrow^G_H \downarrow^G_H \ul \pi_n(X)\),
we find that \(\ul\pi_n(\Si{-\sigma^H} X) \cong \ul\pi_n(X)\).
Now suppose that \(\pi_{n+1}^H(X) = 0 = \pi_n^G(X)\). Because \(\pi^H_{n+1}(X)\) is the value of \(\uparrow^G_H \downarrow^G_H \ul\pi_{n+1}(X)\) at the orbit \(G/H\), the left three terms of \cref{les:reduction} prove the exact sequence
\begin{align*}
0 \rightarrow \pi_n^G(\Si{-\sigma^H} X) \rightarrow 0.
\end{align*}
Consequently, \(\pi_n^G(\Si{-\sigma^H} X) = 0\).
\end{proof}
\begin{cor} \label{EM:homotopy:reduction}
Let \(H\ul M\) be an Eilenberg-MacLane \(G\)-spectrum and \(V \cong \mathbf{s} \oplus \bigoplus_{i=1}^r \sigma^{H_i}\) be a real representation with \(s\) copies of the trivial representation and each \(\sigma^{H_i}\) the sign representation from \(G\rightarrow G/H_i\), where \(H_i\) is an index two subgroup of \(G\). Then \(\Si{-V} H\ul M\) does not have nontrivial homotopy above degree \(-s\).
\end{cor}
\begin{proof}
First apply \cref{les:reduction} to \(X_1 := \Si{-\sigma^{H_1}} H\ul M\) with \(n\geq 1\). Then repeat for \mbox{\(X_i:= \Si{-\sigma^{H_i}} X_{i-1}\)} where \(2\leq i\leq r\). Then \(\Si{-V} H\ul M \simeq \Si{-s} X_{r}\) has no homotopy above degree \(-s\).
\end{proof}
Recall from \cref{subsec:signrep} that \(\alpha\), \(\beta\), and \(\gamma\) are the sign representations of \(K\).
\begin{lem} \label{homotopy:-beta:M}
Let \(\ul M\) be a \(K\)-Mackey functor. The nontrivial homotopy of \(\Si{-\beta} H\ul M\) is
\begin{center}
\begin{tikzpicture}[scale=0.8, every node/.style={scale=0.75}, bend right=2ex]
\node (pi0) at (-8,2) {$\ul \pi_0 \quad =$};
\node (lMK) at (-5,4) {$\ker \res{K}{L}$};
\node (lML) at (-7,2) {$0$};
\node (lMD) at (-5,2) {$\ker \res{D}{e}$};
\node (lMR) at (-3,2) {$\ker \res{R}{e}$};
\node (lMe) at (-5,0) {$0$};
%
%
\begin{scope}[xshift=-3cm]
\node (pi-1) at (3,2) {$\ul \pi_{-1} \quad =$};
\node (MK) at (7,4) {$M_L / \text{im } \res{K}{L}$};
\node (ML) at (5,2) {$\mathbb{Z}_\sigma^L\otimes M_L$};
\node (MD) at (7,2) {$M_e / \text{im } \res{D}{e}$};
\node (MR) at (9,2) {$M_e / \text{im } \res{R}{e}$};
\node (Me) at (7,0) {$\mathbb{Z}_\sigma^L\otimes M_e$};
\end{scope}
\draw[->,black] (lMK) to (lMR);
\draw[->,black] (lMR) to (lMK);
\draw[->,black] (lMK) to (lMD);
\draw[->,black] (lMD) to (lMK);
\arres{MK}{ML}{-2}{0}{$\varphi_r^{K L}$}
\draw[->,black] (MK) to (MD);
\draw[->,black] (MK) to (MR);
\artr{ML}{MK}{0}{0}{$\pi$}
\draw[->,black] (MD) to (MK);
\draw[->,black] (MR) to (MK);
\draw[->,black] (ML) to node[sloped,anchor=center,xshift={0pt},yshift={-5pt}] {\scalebox{0.7}{$1\otimes r\cdot \res{L}{e}$}} (Me);
\draw[->,black] (Me) to node[sloped,anchor=center,xshift={0pt},yshift={5pt}] {\scalebox{0.7}{$1\otimes r\cdot \tr{L}{e}$}} (ML);
\arres{MD}{Me}{-6}{0}{$\varphi_r^{De}$}
\arres{MR}{Me}{1}{2}{$\varphi_r^{Re}$}
\artr{Me}{MD}{0}{0}{$\pi$}
\artr{Me}{MR}{0}{0}{$\pi$}
\end{tikzpicture}
\end{center}
Here, \(\varphi_r^{K L}\) is induced by \(\Delta^r\) in the square
\begin{equation*}
\begin{tikzcd}
M_L \arrow[r, "\pi"] \arrow[d, "\Delta^r"] & M_L/ \text{im } \res{K}{L} \arrow[d, "\varphi_r^{K L}"] \\
\mathbb{Z}[K/L] \otimes M_L \arrow[r, "q_L"] & \mathbb{Z}_\sigma^L\otimes M_L
\end{tikzcd}
\end{equation*}
and is given by \(\varphi_r^{KL}(m) = m-r\cdot m\). The maps \(\varphi_r^{De}:M_e/\text{im } \res{D}{e}\rightarrow \mathbb{Z}_\sigma^L\otimes M_e\) and \(\varphi_r^{Re}:M_e/\text{im } \res{R}{e}\rightarrow \mathbb{Z}_\sigma^L\otimes M_e\) are defined similarly.
\end{lem}
\begin{proof}
We have the cofiber sequence \((S^{-\beta} \rightarrow S^0 \rightarrow K/L_+)\wedge H\ul M\). The result then follows from the description of the map \(\ul M \rightarrow \uparrow^K_L\downarrow^K_L \ul M\) given in \cref{Induction_Formula}.
\end{proof}
\begin{prop} \label{homotopy:-rho:M}
Let \(\ul M\) be a \(K\)-Mackey functor. The nontrivial homotopy of \(\Si{-\alpha-\beta-\gamma} H\ul M\) is
\begin{tikzpicture}[scale=0.8, every node/.style={scale=0.75}]
\node (pi0) at (-9,2) {$\ul \pi_0 \quad =$};
%
\node (lMK) at (-5,4) {$\ker (\res{K}{L} + \res{K}{D} + \res{K}{R})$};
\node (lML) at (-7,2) {$0$};
\node (lMD) at (-5,2) {$0$};
\node (lMR) at (-3,2) {$0$};
\node (lMe) at (-5,0) {$0$};
%
%
%
\begin{scope}[xshift=-3cm]
\node (pi-1) at (3,2) {$\ul \pi_{-1} \quad =$};
%
\node (MK) at (7,4) {$E_2$};
\node (ML) at (5,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^L &\otimes \\ \ker & \res{L}{e}
\end{aligned}$}};
\node (MD) at (7,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^D &\otimes \\ \ker & \res{D}{e}
\end{aligned}$}};
\node (MR) at (9,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^R &\otimes \\ \ker & \res{R}{e}
\end{aligned}$}};
\node (Me) at (7,0) {$0$};
\end{scope}
%
\arres{MK}{ML}{0}{0}{}
\arres{MK}{MD}{0}{0}{}
\arres{MK}{MR}{0}{0}{}
\artr{ML}{MK}{0}{0}{}
\artr{MD}{MK}{0}{0}{}
\artr{MR}{MK}{0}{0}{}
\end{tikzpicture}
\begin{tikzpicture}[scale=0.8, every node/.style={scale=0.75}]
\node (pi-2) at (-9,2) {$\ul \pi_{-2} \quad =$};
%
\node (lMK) at (-5,4) {$E_3$};
\node (lML) at (-7,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^L \otimes& \\ \ker\varphi_r^{Le}
\end{aligned}$}};
\node (lMD) at (-5,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^D \otimes& \\ \ker \varphi_r^{De}
\end{aligned}$}};
\node (lMR) at (-3,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^R \otimes& \\ \ker\varphi_r^{Re}
\end{aligned}$}};
\node (lMe) at (-5,0) {$0$};
%
\arres{lMK}{lML}{0}{0}{}
\arres{lMK}{lMD}{0}{0}{}
\arres{lMK}{lMR}{0}{0}{}
\artr{lML}{lMK}{0}{0}{}
\artr{lMD}{lMK}{0}{0}{}
\artr{lMR}{lMK}{0}{0}{}
%
%
\begin{scope}[xshift=-3cm]
\node (pi-3) at (3,2) {$\ul \pi_{-3} \quad =$};
%
\node (MK) at (7,4) {$(M_e / \text{im } \varphi_r^{De}) / \text{im } \varphi_r^{R,L}$};
\node (ML) at (5,2) {$M_e / \text{im } \varphi_d^{Le}$};
\node (MD) at (7,2) {$M_e / \text{im } \varphi_r^{De}$};
\node (MR) at (9,2) {$M_e / \text{im } \varphi_r^{Re}$};
\node (Me) at (7,0) {$M_e$};
\end{scope}
%
\arres{MK}{ML}{0}{0}{$\varphi_d$}
\arres{MK}{MD}{-5}{-10}{$\varphi_l$}
\arres{MK}{MR}{0}{-14}{$\varphi_d$}
\artr{ML}{MK}{0}{0}{$\pi$}
\artr{MD}{MK}{0}{0}{$\pi$}
\artr{MR}{MK}{0}{0}{$\pi$}
\arres{ML}{Me}{-15}{-5}{$\varphi_{l}^{dLe}$}
\arres{MD}{Me}{-10}{5}{$\varphi_{d}^{rDe}$}
\arres{MR}{Me}{0}{0}{$\varphi_{l}^{rRe}$}
\artr{Me}{ML}{0}{0}{$\pi$}
\artr{Me}{MD}{0}{0}{$\pi$}
\artr{Me}{MR}{0}{0}{$\pi$}
\end{tikzpicture}
where \(E_1\), \(E_2\), and \(E_3\) are extensions
\begin{equation*}
\begin{tikzcd}[row sep=small]
M_e/\text{im } \varphi_r^{De} \arrow[r] & E_3 \arrow[r] & (M_e/\text{im } \res{R}{e})/\text{im } \res{L}{e} \\
\ker\res{D}{e} \arrow[r] & E_2 \arrow[r] & E_1 \\
\ker\res{R}{e} \arrow[r] & E_1 \arrow[r] & M_L/\text{im } \res{K}{L}
\end{tikzcd}
\end{equation*}
Let \(\varphi_h^{*}:\overline{M_e} \rightarrow \overline{M_e}'\) be one of the maps shown. Then \(\varphi_h^{*}(m) = m-h\cdot m\). Additionally, \((M_e/\text{im } \varphi_r^{De})/\text{im } \varphi_r^{R,L}\) is the cokernel of the map
\begin{equation*}
\begin{tikzcd}
(M_e/\text{im } \res{R}{e})/\text{im } \res{L}{e} \arrow[r, "\varphi_r^{R,L}"] & M_e/\text{im } \varphi_r^{De}.
\end{tikzcd}
\end{equation*}
\end{prop}
\begin{proof}
\cref{homotopy:-beta:M} supplies the homotopy for \(\Si{-\beta} H\ul M\). We continue constructing \(\Si{-\alpha-\beta-\gamma} H\ul M\) iteratively. The cofiber sequence
\begin{align*}
(S^{-\alpha} \rightarrow S^0 \rightarrow K/R_+) \wedge \Si{-\beta} H\ul M
\end{align*}
results in the homotopy of \(\Si{-\alpha-\beta} H\ul M\)
\begin{tikzpicture}[scale=0.8, every node/.style={scale=0.75}]
\node (pi0) at (-9,2) {$\ul \pi_0 \quad =$};
%
\node (lMK) at (-5,4) {$\ker (\res{K}{L} + \res{K}{R})$};
\node (lML) at (-7,2) {$0$};
\node (lMD) at (-5,2) {$\ker \res{K}{D}$};
\node (lMR) at (-3,2) {$0$};
\node (lMe) at (-5,0) {$0$};
%
\arres{lMK}{lMD}{0}{0}{}
\artr{lMD}{lMK}{0}{0}{}
%
%
\begin{scope}[xshift=-3cm]
\node (pi-1) at (3,2) {$\ul \pi_{-1} \quad =$};
\node (MK) at (7,4) {$E_1$};
\node (ML) at (5,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^L &\otimes \\ \ker & \res{L}{e}
\end{aligned}$}};
\node (MD) at (7,2) {\scalebox{0.9}{$\ker\varphi_r^{De}$}};
\node (MR) at (9,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^R &\otimes \\ \ker & \res{R}{e}
\end{aligned}$}};
\node (Me) at (7,0) {$0$};
\end{scope}
%
\arres{MK}{ML}{0}{0}{}
\arres{MK}{MD}{0}{0}{}
\arres{MK}{MR}{0}{0}{}
\artr{ML}{MK}{0}{0}{}
\artr{MD}{MK}{0}{0}{}
\artr{MR}{MK}{0}{0}{}
\end{tikzpicture}
\begin{tikzpicture}[scale=0.8, every node/.style={scale=0.75}]
\node (pi-2) at (-9,2) {$\ul \pi_{-2} \quad =$};
%
\node (lMK) at (-5,4) {$(M_e/\text{im } \res{D}{e})/ \text{im } \res{L}{e}$};
\node (lML) at (-7,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^L &\otimes \\ M_e / &\text{im } \res{L}{e}
\end{aligned}$}};
\node (lMD) at (-5,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^D &\otimes \\ M_e / &\text{im } \varphi_r^{De}
\end{aligned}$}};
\node (lMR) at (-3,2) {\scalebox{0.9}{$\begin{aligned}
\mathbb{Z}_\sigma^R &\otimes \\ M_e / &\text{im } \res{R}{e}
\end{aligned}$}};
\node (lMe) at (-5,0) {$\mathbb{Z}_\sigma^D \otimes M_e$};
%
\arres{lMK}{lML}{0}{0}{$\varphi_r^{D,L}$}
\arres{lMK}{lMD}{-8}{-17}{$\varphi_r^{D,L}$}
\arres{lMK}{lMR}{0}{-23}{$\varphi_d^{D,L}$}
\artr{lML}{lMK}{-11}{0}{$\pi$}
\artr{lMD}{lMK}{0}{0}{$\pi$}
\artr{lMR}{lMK}{0}{0}{$\pi$}
\arres{lML}{lMe}{-10}{-8}{$\varphi_d^{Le}$}
\arres{lMD}{lMe}{-6}{3}{$\varphi_d^{rDe}$}
\draw[->,black,bend right=2ex] (lMR) to node[sloped, anchor=center, xshift={0pt},yshift={-5pt}] {\scalebox{0.7}{$1\otimes \varphi_r^{Re}$}} (lMe);
\artr{lMe}{lML}{0}{0}{$\pi$}
\artr{lMe}{lMD}{-2}{0}{$\pi$}
\draw[->,black,bend right=2ex] (lMe) to node[sloped, anchor=center, xshift={0pt},yshift={-5pt}] {\scalebox{0.7}{$1\otimes \pi$}} (lMR);
\end{tikzpicture}
Finally, the cofiber sequence \((S^{-\gamma} \rightarrow S^0 \rightarrow K/D_+) \wedge \Si{-\alpha-\beta} H\ul M\) provides the result.
\end{proof}
\begin{lem} \label{Kernel_Equivalences}
Let \(\phi:G\rightarrow N\) and \(\psi:G\rightarrow M\) be group homomorphisms. Then
\begin{align*}
\ker(\phi, \psi) = \ker(\phi) \cap \ker(\psi) = \ker( \phi\vert_{\ker(\psi)})
\end{align*}
where \((\phi,\psi):G\rightarrow N\oplus M\) is defined by \(g\mapsto \phi(g)\oplus \psi(g)\).
\end{lem}
\begin{prop} \label{2slice:equivalences:si2}
Let \(\ul M\) be a \(K\) Mackey functor where the restrictions \(\res{L}{e}\), \(\res{D}{e}\), and \(\res{R}{e}\) are injective. The following are equivalent.
\begin{enumerate}[(A)]
\item The sequence
\begin{align*}
M_K \xrightarrow{\res{K}{L}+\res{K}{D}+\res{K}{R}} M_L \oplus M_D \oplus M_R \xrightarrow{\left(\begin{smallmatrix} -\res{L}{e} & \res{L}{e} & 0 \\ \res{D}{e} & 0 & -\res{D}{e} \\ 0 & -\res{R}{e} & \res{R}{e} \end{smallmatrix}\right)} M_e^3
\end{align*}
is exact. \label{2slice:si2:exact1}
\item \(\text{im } \res{K}{H_1} = (\res{H_1}{e})^{-1}(\text{im } \res{H_2}{e}) \cap (\res{H_1}{e})^{-1}(\text{im } \res{H_3}{e})\) where \(H_1\), \(H_2\), \(H_3\) are distinct order two subgroups of \(K\). \label{2slice:res:injective}
This equality is represented by the diagram below.
\begin{equation*}
\begin{tikzcd}
\text{im } \res{K}{H_1} \arrow[d, hook] \ar[r] & \text{im } \res{H_2}{e} \cap \text{im } \res{H_3}{e} \ar[d, hook] \\
M_{H_1} \ar[r, "\res{H_1}{e}"] & M_e
\end{tikzcd}
\end{equation*}
\end{enumerate}
\end{prop}
\begin{proof}
Without loss of generality, let \(H_1=L\), \(H_2=D\), and \(H_3=R\). For convenience, set \(I = (\res{L}{e})^{-1}(\text{im } \res{D}{e}) \cap (\res{L}{e})^{-1}(\text{im } \res{R}{e})\).
\cref{2slice:si2:exact1} \(\Rightarrow\) \cref{2slice:res:injective}:
Let \(x\in \text{im } \res{K}{L}\). Then we have \(k\in M_K\) such that \(\res{K}{L}(k) = x\). Additionally, \(\res{L}{e}(x) = \res{D}{e} \res{K}{D} (k) = \res{R}{e} \res{K}{R} (k)\). It follows that \(x\in I\).
Now suppose \(x\in I\). We then have \(y\in D\) and \(z\in R\) such that
\begin{align*}
\res{L}{e}(x) = \res{D}{e}(y) = \res{R}{e}(z).
\end{align*}
Thus, \((x,y,z) \in \ker \left(\begin{smallmatrix} -\res{L}{e} & \res{L}{e} & 0 \\ \res{D}{e} & 0 & -\res{D}{e} \\ 0 & -\res{R}{e} & \res{R}{e} \end{smallmatrix}\right)\). Consequently, we have \(x = \res{K}{L}(k)\) for some \(k\in M_K\); that is, \(x\in \text{im } \res{K}{L}\).
\cref{2slice:res:injective} \(\Rightarrow\) \cref{2slice:si2:exact1}:
Let \((x,y,z)\in \ker \left(\begin{smallmatrix} -\res{L}{e} & \res{L}{e} & 0 \\ \res{D}{e} & 0 & -\res{D}{e} \\ 0 & -\res{R}{e} & \res{R}{e} \end{smallmatrix}\right)\). Then \(\res{L}{e}(x) = \res{D}{e}(y) = \res{R}{e}(z)\), so \(x\in \text{im } \res{K}{L}\). So there is some \(k\in M_K\) such that \(x = \res{K}{L}(k)\). Because \(\res{D}{e}\) and \(\res{R}{e}\) are injective, it must be that \(y = (\res{D}{e})^{-1}(\res{L}{e}(x))\) and \(z = (\res{R}{e})^{-1}(\res{L}{e}(x))\). Hence, \cref{2slice:si2:exact1} is exact.
\end{proof}
\begin{prop} \label{2slice:characterization:si1}
The \(K\)-spectrum \(\Si{1} H\ul N\) is a 2-slice if and only if \(N_e = 0\) and \(N_K \rightarrow N_L \oplus N_D \oplus N_R\) is injective.
\end{prop}
\begin{proof}
By \cref{thm:atleastn}, we find that \(\Si{1} H\ul N\geq 2\) if and only if \(N_e = 0\). Consequently, going forward we may assume \(N_e = 0\).
Now \(\Si{1} H\ul N\leq 2\) if and only if \([S^{k\rho_H+r}, \Si{2} H\ul N]^H=0\) for all \(k\geq \frac{3}{\abs{H}}\) and \(r\geq 0\).
Because \(N_e = 0\), \cref{C2:2slice} implies that \(i_H^*(\Si{1} H\ul N)\) is a 2-slice, where \(H\) is an order two subgroup of \(K\). Thus, to finish the equivalence, we only need consider
\begin{align*}
[S^{k\rho_K+r}, \Si{1} H\ul N]^K = [S^{k+r-1}, \Si{-k\overline{\rho}_K} H\ul N]^K
\end{align*}
for all \(k\geq 1\) and \(r\geq 0\).
From \cref{EM:homotopy:reduction}, we need only concern ourselves with \([S^0, \Si{-k\overline{\rho}_K} H\ul N]\). From \cref{homotopy:-rho:M},
\begin{align*}
\ul \pi_0(\Si{-\overline\rho} H\ul N) &= \phi^*_K( \ker \res{K}{L} \cap \ker \res{K}{D} \cap \ker \res{K}{R}).
\end{align*}
Therefore \cref{Homotopy_Reduction} shows that \(\ul \pi_0(\Si{-\overline\rho} H \ul N)\) vanishes if and only if \(\ul \pi_0(\Si{-k\overline\rho} H \ul N)\) vanishes for all \(k\geq 1\). Hence, \(\Si{1} H\ul N\leq 2\) if and only if \(\ker \res{K}{L} \cap \ker \res{K}{D} \cap \ker \res{K}{R} = \{0\}\). By \cref{Kernel_Equivalences} this is equivalent to \(N_K \rightarrow N_L \oplus N_D \oplus N_R\) being injective.
\end{proof}
\begin{prop} \label{2slice:characterization:si2}
The spectrum \(\Si{2} H\ul M\) is a 2-slice if and only if all restrictions of \(\ul M\) are injective and the sequence
\begin{align*}
M_K \xrightarrow{\res{K}{L}+\res{K}{D}+\res{K}{R}} M_L \oplus M_D \oplus M_R \xrightarrow{\left(\begin{smallmatrix} -\res{L}{e} & \res{L}{e} & 0 \\ \res{D}{e} & 0 & -\res{D}{e} \\ 0 & -\res{R}{e} & \res{R}{e} \end{smallmatrix}\right)} M_e^3
\end{align*}
is exact.
\end{prop}
\begin{proof}
Note that \(\Si{2} H\ul M \geq 2\), so we just need to show that \(\Si{2} H\ul M\leq 2\). Now \(\Si{2} H\ul M\leq 2\) if and only if \([S^{k\rho_H+r}, \Si{2} H\ul M]^H=0\) for all \(k\geq \frac{3}{\abs{H}}\) and \(r\geq 0\). For \(H\) an order two subgroup of \(K\), \(i_H^* \Si{2} H\ul M\leq 2\) if and only if \(\res{H}{e}\)
is injective. Consequently, going forward we may assume \(\res{L}{e}\), \(\res{D}{e}\), and \(\res{R}{e}\) are injective.
To finish the equivalence, we only need consider
\begin{align*}
[S^{k\rho_K+r}, \Si{2} H\ul M]^K = [S^{k+r-2}, \Si{-k\overline{\rho}_K} H\ul M]^K
\end{align*}
for all \(k\geq 1\) and \(r\geq 0\). By \cref{EM:homotopy:reduction}, it is enough to examine \([S^0, \Si{-k\overline{\rho}_K} H\ul M]\) and \([S^{-1}, \Si{-k\overline{\rho}_K} H\ul M]\). From \cref{homotopy:-rho:M},
\begin{align*}
\ul \pi_0(\Si{-\overline\rho} H\ul M) &= \phi^*_K( \ker \res{K}{L})
\end{align*}
and
\begin{align*}
\ul \pi_{-1}(\Si{-\overline\rho} H\ul M) &= \phi^*_K \left( \frac{(\res{L}{e})^{-1}(\text{im }\res{D}{e}) \cap (\res{L}{e})^{-1}(\text{im } \res{R}{e})}{\text{im } \res{K}{L}} \right).
\end{align*}
Note that \cref{homotopy:-rho:M} states
\begin{align*}
\ul \pi_0(\Si{-\overline\rho} H\ul M) = \phi^*_K( \ker \res{K}{L} \cap \ker\res{K}{D} \cap\ker\res{K}{R}),
\end{align*}
but because the lower restrictions are injective, these kernels coincide.
\cref{Homotopy_Reduction} then yields that
\begin{align*}
\ul\pi_0^G(\Si{-k\overline\rho} H\ul M) &\cong \ul\pi_0^G(\Si{-\overline\rho} H\ul M) = \ker \res{K}{L} \\
\ul\pi_{-1}^G(\Si{-k\overline\rho} H\ul M) &\cong \ul\pi_{-1}^G(\Si{-\overline\rho} H\ul M) = \frac{(\res{L}{e})^{-1}(\text{im }\res{D}{e}) \cap (\res{L}{e})^{-1}(\text{im } \res{R}{e})}{\text{im } \res{K}{L}}
\end{align*}
for all \(k\geq 2\). By \cref{def:lessthan}, we find that \(\Si{2} H\ul M\leq 2\) if and only if these two homotopy groups vanish. Hence, \(\Si{2} H\ul M\leq 2\) if and only if \(\ker \res{K}{L}=\{0\}\) and \(\text{im } \res{K}{L} = (\res{L}{e})^{-1}(\text{im }\res{D}{e}) \cap (\res{L}{e})^{-1}(\text{im } \res{R}{e})\). Because the homotopy of \(\Si{-\alpha-\beta-\gamma} H\ul M\) is invariant under the order of construction -- that is, whether \(H\ul M\) is suspended by say \(-\alpha\) or \(-\beta\) first -- we find that all upper restrictions must be injective and that \cref{2slice:res:injective} must hold.
\end{proof}
\begin{thm} \label{2slice:characterization}
Suppose the only nontrivial homotopy Mackey functors of a \(K\)-spectrum \(X\) are \(\ul\pi_1(X)\) and \(\ul\pi_2(X)\) where
\begin{enumerate}
\item All restrictions of \(\ul\pi_2(X)\) are injective and the sequence
\begin{align*}
\pi_2^K(X) \xrightarrow{\res{K}{L}+\res{K}{D}+\res{K}{R}} \pi_2^L(X) \oplus \pi_2^D(X) \oplus \pi_2^R(X) \xrightarrow{\left(\begin{smallmatrix} -\res{L}{e} & \res{L}{e} & 0 \\ \res{D}{e} & 0 & -\res{D}{e} \\ 0 & -\res{R}{e} & \res{R}{e} \end{smallmatrix}\right)} \pi_2^e(X)^3
\end{align*}
is exact. \label{2slice:thm:M}
\item \(\pi_1^e(X) = 0\). \label{2slice:thm:N0}
\item \(\pi_1^K(X) \rightarrow \pi_1^L(X) \oplus \pi_1^D(X) \oplus \pi_1^R(X)\) is injective. \label{2slice:thm:N0:2}
\end{enumerate}
Then \(X\) is a 2-slice.
Conversely, if a \(K\)-spectrum \(X\) is a 2-slice, then its only nontrivial homotopy Mackey functors of a \(K\)-spectrum \(X\) are \(\ul\pi_1(X)\) and \(\ul\pi_2(X)\) where \(\Si{2} H\ul\pi_2(X)\) is a 2-slice and \(\Si{1} H\ul\pi_1(X) \in [2,4]\), i.e., both \cref{2slice:thm:M} and \cref{2slice:thm:N0} hold.
\end{thm}
\begin{proof}
Let \(\ul\pi_2(X)=\ul M\) and \(\ul\pi_1(X)=\ul N\). If these are the only nontrivial homotopy Mackey functors of \(X\), we have a fiber sequence
\begin{align} \label{2slice:fibseq}
\Si{2} H\ul M\rightarrow X\rightarrow \Si{1} H\ul N.
\end{align}
By \cref{2slice:characterization:si1} and \cref{2slice:characterization:si2}, conditions \eqref{2slice:thm:M} - \eqref{2slice:thm:N0:2} show that \(\Si{2} H\ul M\) and \(\Si{1} H\ul N\) are 2-slices. Now if \(\Si{2} H\ul M\) and \(\Si{1} H\ul N\) are 2-slices, then \(X\) must be a 2-slice as well.
Conversely, suppose that \(X\) is a 2-slice. We then find that \(X\) has no homotopy above degree two and none below degree one; thus, we have the fiber sequence in \cref{2slice:fibseq}.
Because \(X\geq 2\) and \(\Si{2} H\ul M\geq 2\), it follows that \(\Si{1} H\ul N\geq 2\). So by \cref{thm:atleastn}, \(N_e = 0\). That \(\Si{1} H\ul N\leq 4\) follows from a similar argument as in \cref{2slice:characterization:si1}.
Rotating this fiber sequence gives \(H\ul N\rightarrow \Si{2} H\ul M \rightarrow X\). As \(H\ul N\) is a 0-slice and \(X\) is a 2-slice, we have \(\Si{2} H\ul M \in [0,2]\), so it must be that \(\Si{2} H\ul M = 2\). Consequently, by \cref{2slice:characterization:si2}, \cref{2slice:thm:M} holds.
\end{proof}
It is not necessary for condition \eqref{2slice:thm:N0:2} to hold for \(X\) to be a 2-slice as we show in the following example.
\begin{exmp}
Take \(X \simeq \Si{1+\beta} H\ul E\) where
\begin{equation*}
\ul E = \begin{tikzcd}[ampersand replacement=\&, column sep=scriptsize]
\& \mathbb{F}_2 \ar[d, color=black] \ar[dr,color=black] \& \\
0 \& \mathbb{F}_2 \& \mathbb{F}_2 \\
\& \mathbb{F}_2 \ar[u, color=black] \ar[ur, color=black] \&
\end{tikzcd}
\end{equation*}
Then \(X\) is a 2-slice with \(\ul\pi_2(X) = \ul f\) and \(\ul\pi_1(X) = \underline{g}\). But \(\Si{1} H\underline{g}\) is not a 2-slice.
\end{exmp}
\begin{proof}
We can construct \(\Si{1+\beta} H\ul E\) using the cofiber sequence
\begin{align*}
(K/L_+ \rightarrow S^0 \rightarrow S^\beta) \wedge \Si{1} H\ul E.
\end{align*}
The resulting long exact sequence in homotopy is
\begin{flushright}
\begin{tikzpicture}[scale=0.8, every node/.style={scale=0.75},xscale=0.9]
\node (pi2) at (-9,1.5) {$\ul \pi_2$};
%
\node (MD) at (0,1.5) {$\ul 0$};
%
\node (rMK) at (5.5,3) {$0$};
\node (rML) at (4,1.5) {$0$};
\node (rMD) at (5.5,1.5) {$0$};
\node (rMR) at (7,1.5) {$0$};
\node (rMe) at (5.5,0) {$\mathbb{F}_2$};
%
\draw[->] (MD) to (rML);
\end{tikzpicture}
\begin{tikzpicture}[scale=0.8, every node/.style={scale=0.75},xscale=0.9]
\node (pi1) at (-9,1.5) {$\ul \pi_1$};
%
\node (lMK) at (-5.5,3) {$0$};
\node (lML) at (-7,1.5) {$0$};
\node (lMD) at (-5.5,1.5) {$\mathbb{F}_2$};
\node (lMR) at (-4,1.5) {$\mathbb{F}_2$};
\node (lMe) at (-5.5,0) {$\mathbb{F}_2[K/L]$};
%
\node (MK) at (0,3) {$\mathbb{F}_2$};
\node (ML) at (-1.5,1.5) {$0$};
\node (MD) at (0,1.5) {$\mathbb{F}_2$};
\node (MR) at (1.5,1.5) {$\mathbb{F}_2$};
\node (Me) at (0,0) {$\mathbb{F}_2$};
%
\node (rMK) at (5.5,3) {$\mathbb{F}_2$};
\node (rML) at (4,1.5) {$0$};
\node (rMD) at (5.5,1.5) {$0$};
\node (rMR) at (7,1.5) {$0$};
\node (rMe) at (5.5,0) {$0$};
%
\arres{lMD}{lMe}{-4}{0}{$\Delta$}
\arres{lMR}{lMe}{0}{0}{$\Delta$}
\artr{lMe}{lMD}{0}{0}{$\nabla$}
\artr{lMe}{lMR}{0}{0}{$\nabla$}
\draw[->, black] (MK) to node[right,xshift={2pt}] {\scalebox{0.6}{$1$}} (MR);
\draw[->, black] (MK) to node[right,xshift={2pt}] {\scalebox{0.6}{$1$}} (MD);
\draw[->, black] (Me) to node[right,xshift={2pt}] {\scalebox{0.6}{$1$}} (MR);
\draw[->, black] (Me) to node[right,xshift={-2pt}] {\scalebox{0.6}{$1$}} (MD);
%
\draw[->] (lMK) to node[above, xshift={-1pt}] {} (MK);
\draw[->] (lMR) to node[above, xshift={-1pt}] {\scalebox{0.7}{$\begin{pmatrix}0 \\ 1 \\ 1 \end{pmatrix}$}} (ML);
\draw[->] (lMe) to node[above, xshift={-1pt}] {\scalebox{0.7}{$\nabla$}} (Me);
%
\draw[->>] (MK) to (rMK);
\draw[->>] (MR) to (rML);
\draw[->>] (Me) to (rMe);
\end{tikzpicture}
\end{flushright}
So \(\ul \pi_2(X) = \ul f\) and \(\ul \pi_1(X) = \underline{g}\). From this we see that \(X\geq 2\) by \cref{2slice:characterization}. Note that \(\Si{2} H\ul f\) is a 2-slice and \(\Si{1} H\underline{g}\) is a 4-slice.
To show that \(X\leq 2\), we need \([S^{k\rho_H+r}, \Si{1+\beta} H\ul E]^H=0\) for all \(k\geq \frac{3}{\abs{H}}\) and \(r\geq 0\). As \(i_H^*(X) \simeq \Si{2} H_{C_2} \ul f\) is a 2-slice, where \(H\) is an order two subgroup of \(K\), we only need to consider
\begin{align*}
[S^{k\rho_K+r}, \Si{1+\beta} H\ul E]^K = [S^{k+r-1}, \Si{-\alpha-\gamma-(k-1) \overline{\rho}_K} H\ul E]^K
\end{align*}
for all \(k\geq 1\) and \(r\geq 0\).
By \cref{EM:homotopy:reduction}, it is sufficient to examine
\([S^0, \Si{-\alpha-\gamma-(k-1)\overline{\rho}_K} H\ul E]^K\)
for all \(k\geq 1\).
From \cref{homotopy:-beta:M} we find that \([S^0, \Si{-\alpha-\gamma} H\ul E] = \ul 0\). Consequently, given any \(k\geq 1\), repeated application of \cref{Homotopy_Reduction} shows that
\begin{align*}
[S^0, \Si{-\alpha-\gamma-(k-1)\overline{\rho}_K} H\ul E] = \ul 0.
\end{align*}
That is, \(X\) is a 2-slice.
\end{proof}
\section{Cotowers for \texorpdfstring{\(\Si{-n} H\underline{\Z}\)}{HZ}} \label{sec:cotowers:K:Z}
We determine the slice towers of \(\Si{-n} H\underline{\Z}\) and \(\Si{-n} H\ul m^*\) for \(1\leq n\leq 5\).
\begin{exmp} \label{slices:K:m*:1-2}
By \cref{charac:slice:01-1} and \cref{m:mg:equiv}, \(\Si{-1} H\ul m^*\) is a \((-2)\)-slice and \(\Si{-2} H\ul m^*\) is a \((-4)\)-slice.
Alternatively, by \cref{m:mg:equiv}, \(\Si{\rho}\Si{-1} H\ul m^* \simeq \Si{1} H\ul {mg}\), which is a 2-slice by \cref{2slice:characterization:si1}. Consequently, \(\Si{-1} H\ul m^*\) is a \((-2)\)-slice.
\end{exmp}
\begin{exmp} \label{tower:-1:K:Z}
By \cite[Theorem 6-4]{U}, the cotower for \(\Si{-1} H\underline{\Z}\) is
\begin{equation*}
\begin{tikzcd}
\Si{-1} H\underline{\Z}^* \arrow[d] = P^{-1}_{-1} & \\
\Si{-1} H\underline{\Z}(2,1) \arrow[d] \arrow[r] & \Si{-1} H\ul m^* = P^{-2}_{-2} \\
\Si{-1} H\underline{\Z} \arrow[r] & \Si{-1} H\underline{g} = P^{-4}_{-4}
\end{tikzcd}
\end{equation*}
\end{exmp}
\begin{exmp} \label{tower:-2:K:Z}
We suspend the cotower in \cref{tower:-1:K:Z} by \(-1\) to get the cotower for \(\Si{-2} H\underline{\Z}\).
\begin{equation*}
\begin{tikzcd}
\Si{-2} H\underline{\Z}^* \arrow[d] = P^{-2}_{-2} & \\
\Si{-2} H\underline{\Z}(2,1) \arrow[d] \arrow[r] & \Si{-2} H\ul m^* = P^{-4}_{-4} \\
\Si{-2} H\underline{\Z} \arrow[r] & \Si{-2} H\underline{g} = P^{-8}_{-8}
\end{tikzcd}
\end{equation*}
\end{exmp}
\begin{exmp} \label{tower:-3:K:Z}
We suspend the cotower in \cref{tower:-2:K:Z} by \(-1\) to get a partial cotower for \(\Si{-3} H\underline{\Z}\).
\begin{equation*}
\begin{tikzcd}
\Si{-3} H\underline{\Z}^* \arrow[d] = P^{-3}_{-3} & \\
\Si{-3} H\underline{\Z}(2,1) \arrow[d] \arrow[r] & \Si{-3} H\ul m^* \\
\Si{-3} H\underline{\Z} \arrow[r] & \Si{-3} H\underline{g} = P^{-12}_{-12}
\end{tikzcd}
\end{equation*}
The issue here is that \(\Si{-3} H\ul m^*\) is not a slice.
However, the short exact sequence
\begin{align*}
\underline{g}^2 \rightarrow \phi^*_{LDR} (\underline{\F}^*) \rightarrow \ul m^*
\end{align*}
provides the tower
\begin{align*}
P^{-6}_{-6} = \Si{-3} H\phi^*_{LDR} (\underline{\F}^*) \rightarrow \Si{-3} H\ul m^* \rightarrow \Si{-2} H\underline{g}^2 = P^{-8}_{-8}.
\end{align*}
Consequently, the remaining slices of \(\Si{-3} H\underline{\Z}\) are \(P^{-6}_{-6} = \Si{-2p+1} H\phi^*_{LDR}( \underline{\F})\) and \(P^{-8}_{-8} = \Si{-2} H\underline{g}^2\).
\end{exmp}
\begin{exmp} \label{tower:-4:K:Z}
We suspend the cotower in \cref{tower:-3:K:Z} by \(-1\) to get a partial cotower for \(\Si{-4} H\underline{\Z}\).
\begin{equation*}
\begin{tikzcd}
\Si{-4} H\underline{\Z}^* \arrow[d] = P^{-4}_{-4} & \\
\Si{-4} H\underline{\Z}(2,1) \arrow[d] \arrow[r] & \Si{-4} H\ul m^* \\
\Si{-4} H\underline{\Z} \arrow[r] & \Si{-4} H\underline{g} = P^{-16}_{-16}
\end{tikzcd}
\end{equation*}
Again, \(\Si{-4} H\ul m^*\) is not a slice. But suspending the cotower for \(\Si{-3} H\ul m^*\) by \(-1\) provides the missing slices: \(P^{-8}_{-8} = \Si{-4} H\phi^*_{LDR}( \underline{\F}^*)\) and \(P^{-12}_{-12} = \Si{-3} H\underline{g}^2\).
\end{exmp}
\begin{exmp} \label{tower:-5:K:Z}
We suspend the cotower in \cref{tower:-4:K:Z} by \(-1\) and augment with the \(-\rho\) suspension of \cref{tower:-1:K:Z} to get a partial cotower for \(\Si{-5} H\underline{\Z}\).
\begin{equation*}
\begin{tikzcd}
\Si{-\rho-1} H\underline{\Z}^* \arrow[d] = P^{-5}_{-5} & \\
\Si{-\rho-1} H\underline{\Z}(2,1) \arrow[d] \arrow[r] & \Si{-\rho-1} H\ul m^* = P^{-6}_{-6} \\
\Si{-\rho-1} H\underline{\Z} \arrow[d] \arrow[r] & \Si{-2} H\underline{g} = P^{-8}_{-8} \\
\Si{-5} H\underline{\Z}(2,1) \arrow[d] \arrow[r] & \Si{-5} H\ul m^* \\
\Si{-5} H\underline{\Z} \arrow[r] & \Si{-5} H\underline{g} = P^{-20}_{-20}
\end{tikzcd}
\end{equation*}
This time, the cotower for \(\Si{-5} H\ul m^*\) is
\begin{equation*}
\begin{tikzcd}
\Si{-3\rho + 1} H\phi^*_{LDR}(\underline{\F}) = P^{-10}_{-10} \arrow[d] & \\
\Si{-5} H\phi^*_{LDR}(\underline{\F}^*) \arrow[d] \arrow[r] & \Si{-3} H\underline{g}^3 = P^{-12}_{-12} \\
\Si{-5} H\ul m^* \arrow[r] & \Si{-4} H\underline{g}^2 = P^{-16}_{-16}
\end{tikzcd}
\end{equation*}
The remaining slices of \(\Si{-5} H\underline{\Z}\) are then \(P^{-10}_{-10} = \Si{-3\rho+1} H\phi^*_{LDR}(\underline{\F})\), \newline
\(P^{-12}_{-12} = \Si{-3} H\underline{g}^3\), and \(P^{-16}_{-16} = \Si{-4} H\underline{g}^2\).
\end{exmp}
\begin{exmp} \label{tower:-7:K:Z}
The partial cotower for \(\Si{-7} H\underline{\Z}\) follows by suspending the partial cotower in \cref{tower:-5:K:Z} by \(-2\).
\begin{equation*}
\begin{tikzcd}
\Si{-\rho-3} H\underline{\Z}^* \arrow[d] = P^{-7}_{-7} & \\
\Si{-\rho-3} H\underline{\Z}(2,1) \arrow[d] \arrow[r] & \Si{-\rho-3} H\ul m^* \\
\Si{-\rho-3} H\underline{\Z} \arrow[d] \arrow[r] & \Si{-4} H\underline{g} = P^{-16}_{-16} \\
\Si{-7} H\underline{\Z}(2,1) \arrow[d] \arrow[r] & \Si{-7} H\ul m^* \\
\Si{-7} H\underline{\Z} \arrow[r] & \Si{-7} H\underline{g} = P^{-28}_{-28}
\end{tikzcd}
\end{equation*}
We now have a cotower for \(\Si{-\rho-3} H\ul m^*\).
\begin{align*}
P^{-10}_{-10} = \Si{-\rho-3} H\phi^*_{LDR} (\underline{\F}^*) \rightarrow \Si{-\rho-3} H\ul m^* \rightarrow \Si{-3} H\underline{g}^2 = P^{-12}_{-12},
\end{align*}
And a cotower for \(\Si{-7} H\ul m^*\).
\begin{equation*}
\begin{tikzcd}
\Si{-4\rho + 1} H\phi^*_{LDR}(\underline{\F}) = P^{-14}_{-14} \arrow[d] & \\
\Si{-\rho-5} H\phi^*_{LDR}(\underline{\F}^*) \arrow[d] \arrow[r] & \Si{-4} H\underline{g}^3 = P^{-16}_{-16} \\
\Si{-7} H\phi^*_{LDR}(\underline{\F}^*) \arrow[d] \arrow[r] & \Si{-5} H\underline{g}^3 = P^{-20}_{-20} \\
\Si{-7} H\ul m^* \arrow[r] & \Si{-6} H\underline{g}^2 = P^{-24}_{-24}
\end{tikzcd}
\end{equation*}
We now see interference from the cotower for \(\Si{-7} H\ul m^*\). Its \((-14)\)-slice appears below the \((-16)\)-slice in the partial cotower for \(\Si{-7} H\underline{\Z}\). Additionally, both of these (partial) cotowers have a \((-16)\)-slice.
\cref{-4kslices:K:Z} and \cref{-4k-2slices:K:Z} tell us that \(P^{-14}_{-14}(\Si{-7} H\underline{\Z})\) is indeed \(\Si{-4\rho + 1} H\phi^*_{LDR}(\underline{\F})\) and \(P^{-16}_{-16}(\Si{-7} H\underline{\Z})\) is \(\Si{-4} H\underline{g} \vee \Si{-4} H\underline{g}^3 \simeq \Si{-4} H\underline{g}^4\).
\end{exmp}
All partial cotowers for \(\Si{-n} H\underline{\Z}\) will follow this pattern of utilizing the homotopy equivalence \(\Si{-n} H\underline{\Z}^* \simeq \Si{-\rho-n+4} H\underline{\Z}\) to augment the bottom of the cotower for \(\Si{-n} H\underline{\Z}\) with the cotower for \(\Si{-\rho-n+4} H\underline{\Z}\).
\section{Slices of \texorpdfstring{\(\Si{-n} H_{K}\ul \mathbb{Z}\)}{HZ}} \label{sec:slices:K:Z}
Here we determine the slices of \(\Si{-n} H_{K}\underline{\Z}\).
\begin{prop} \label{slice:K:Z1-5:Z*1-4}
\(\Si{n} H\underline{\Z}\) is an \(n\)-slice for \(1\leq n\leq 5\) and \(\Si{-n} H\underline{\Z}^*\) is a \((-n)\)-slice for \(1\leq n\leq 4\).
\end{prop}
\begin{proof}
The \(K\)-spectra \(\Si{1} H\underline{\Z}\), \(\Si{1} H\underline{\Z}^*\), \(\Si{2} H\underline{\Z}\), and \(\Si{-1} H\underline{\Z}^*\) are 1-, 2-, and \((-1)\)-slices by \cref{charac:slice:01-1} and \cref{2slice:characterization}. The result then follows from \cref{equivalence:ZZ*} and the resulting equivalences
\begin{align*}
\Si{5} H \underline{\Z} \simeq \Si{\rho+1} H \ul \mathbb{Z}^*, \quad \Si{4} H \underline{\Z} \simeq \Si{\rho} H \underline{\Z}^*, \quad \text{ and }\quad \Si{3} H \ul\mathbb{Z} \simeq \Si{\rho-1} H\underline{\Z}^*.
\end{align*}
\end{proof}
\subsection{The \texorpdfstring{\((-n)\)}{(-n)}-slice} \label{sec:-nslice:K:Z}
We first establish a comparison of the slices of \(\Si{-n} H\underline{\Z}\) with those of \(\Si{-n+4} H\underline{\Z}\).
\begin{prop} \label{slice:reduction:ZZ^*}
Let \(n\geq 5\). Then
\begin{align*}
P^{-k}_{-k} ( \Si{-n} H\underline{\Z}) \simeq \Si{-\rho} P^{-k+4}_{-k+4} (\Si{-n+4} H\underline{\Z})
\end{align*}
for \(k\in[n,2n-1]\).
\end{prop}
\begin{proof}
By \cref{equivalence:ZZ*}, we have
\begin{align*}
\Si{-\rho} P^{-k+4}_{-k+4} (\Si{-n+4} H\underline{\Z}) \simeq \Si{\rho} \Si{-\rho} P^{-k}_{-k} (\Si{-n+4-\rho} H\underline{\Z}) \simeq P^{-k}_{-k} (\Si{-n} H\underline{\Z}^*).
\end{align*}
Thus, it is sufficient to compare the \((-k)\)-slices of \(\Si{-n} H\underline{\Z}\) and \(\Si{-n} H\underline{\Z}^*\).
Recall the injection \(\underline{\Z}^* \rightarrow \underline{\Z} \rightarrow \underline{\mathbb{M}}\) from \cref{sec:K:Mackey}. We wish to show that
\begin{align} \label{inj:ZZM:-j}
\Si{-n} H\underline{\Z}^* \rightarrow \Si{-n} H\underline{\Z} \xrightarrow{\mu}{} \Si{-n} \underline{\mathbb{M}}
\end{align}
induces an equivalence on slices strictly above level \(-2n\).
We take the Brown-Comenetz dual of \cref{inj:ZZM:-j} to find the fiber sequence
\begin{align*}
\Si{n} H\underline{\mathbb{M}} \xrightarrow{\iota}{} \Si{n} I_{\mathbb{Q}/\mathbb{Z}} H\underline{\Z} \rightarrow \Si{n} I_{\mathbb{Q}/\mathbb{Z}} H\underline{\Z}^*
\end{align*}
Now \(\Si{n} H\underline{\mathbb{M}}\) is \(n\)-connective, and as its underlying spectrum is contractible, when \(n\geq 1\), we know \(\Si{n} H\underline{\mathbb{M}} \geq 2n\) by \cref{thm:atleastn}. Then \cite[Lemma 4.28]{HHR} provides that \(\iota\) induces an equivalence on slices strictly below level \(2n\). It then follows from \cref{duality:prop} that \(\mu\) in \cref{inj:ZZM:-j} induces an equivalence on slices strictly above level \(-2n\).
\end{proof}
This is an analogous result to \cite[Proposition 5.3]{GY}. However, the injection \(\underline{\Z}^* \hookrightarrow \underline{\Z}\) allows us to simplify the argument.
\begin{prop} \label{-nslice:K:Z}
Let \(n\geq 1\) and take \(r\equiv n\pmod 4\) with \(1\leq r\leq 4\). Then the top slice of \(\Si{-n} H\underline{\Z}\) is
\begin{align*}
P^{-n}_{-n}( \Si{-n} H\underline{\Z}) = \Si{-\frac{n-r}{4}\rho -r} H\underline{\Z}^*
\end{align*}
\end{prop}
\begin{proof}
\cref{sec:cotowers:K:Z} gives the result for \(1\leq n\leq 4\), and \(n\geq 5\) follows from repeated application of \cref{slice:reduction:ZZ^*}.
\end{proof}
\subsection{The \texorpdfstring{\((-4k)\)}{(-4k)}-slices} \label{-4kslice:K:Z}
\phantom{a}
Here we determine the \((-4k)\)-slices of \(\Si{-n} H\underline{\Z}\).
\begin{prop} \label{-4kslices:K:Z}
For \(n\geq 1\),
\begin{align*}
P^{-4k}_{-4k} ( \Si{-n} H\underline{\Z} ) \simeq \left\lbrace \begin{array}{ll}
\Si{-k\rho} H\ul {mg} & 4k=n+2 \\
\Si{-k} H\underline{g}^{\frac{1}{2}(4k-n-1)} &
\begin{aligned}
4k\in [n+1,2n-2] \\ n\equiv 1\pmod 2
\end{aligned} \\
\Si{-k\rho} H(\underline{g}^{\frac{1}{2}(4k-n+2)-3} \oplus \phi^*_{LDR} \underline{\F} ) &
\begin{aligned}
4k\in[n+3,2n], \\ n\equiv 0\pmod 2
\end{aligned} \\
\Si{-k} H\underline{g}^{n-k+1} & 4k\in[2n+1,4n]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
We have the equivalence
\begin{align*}
P^{-4k}_{-4k}(\Si{-n} H\underline{\Z}) \simeq \Si{-k\rho} H\ul\pi_0( \Si{-n+k\rho} H\underline{\Z}).
\end{align*}
The restriction to each cyclic subgroup agrees with the slices found in \cref{tower:C2:Z:-n} and \cref{krho:fixedpts:K:Z} gives the fixed points. All that remains is to verify the result for \(4k\in[n+2,2n]\) with \(n\) even. For \mbox{\(4k\in[n+2,2n)\)}, this follows from \cref{tower:-1:K:Z}, \cref{tower:-2:K:Z}, \cref{tower:-3:K:Z}, \cref{tower:-4:K:Z}, and \cref{slice:reduction:ZZ^*}.
Now let \(4k=2n\). The transfer \(\tr{K}{L}\) fits in the cofiber sequence
\begin{equation*}
\begin{tikzcd}
(\Si{-n+k\rho} H\underline{\Z})^L \arrow[r, "\tr{K}{L}"] & (\Si{-n+k\rho} H\underline{\Z})^K \arrow[r] & (\Si{-n+k\rho+\beta} H\underline{\Z})^K.
\end{tikzcd}
\end{equation*}
This gives the exact sequence
\begin{equation} \label{transfer:fixedpts:proof:Z:-4k}
\begin{tikzcd}
\pi_n^L( \Si{k\rho} H\underline{\Z} ) \arrow[r, "\tr{K}{L}"] & \pi_n^K(\Si{k\rho} H\underline{\Z}) \arrow[r] & \pi_n^K (\Si{k\rho+\beta} H\underline{\Z}).
\end{tikzcd}
\end{equation}
We wish to show this transfer is trivial. By \cref{krho:fixedpts:K:Z} and \cref{krho+beta:fixedpoints:K:Z}, \cref{transfer:fixedpts:proof:Z:-4k} becomes
\begin{equation*}
\begin{tikzcd}
\mathbb{F}_2 \arrow[r, "\tr{K}{L}"] & \mathbb{F}_2^{\frac{1}{2}(n+2)} \arrow[r, two heads] & \mathbb{F}_2^{\frac{1}{2}(n+2)} \arrow[r] & 0 = \pi_{n-1}^L( \Si{k\rho} H\underline{\Z}).
\end{tikzcd}
\end{equation*}
Thus, \(\tr{K}{L}\) must be zero. A similar argument shows that \(\tr{K}{R}\) and \(\tr{K}{D}\) are trivial as well.
The restriction \(\res{K}{L}\) fits into the cofiber sequence
\begin{equation*}
\begin{tikzcd}
( \Si{-n+k\rho-\beta} H\underline{\Z})^K \arrow[r] & ( \Si{-n+k\rho} H\underline{\Z})^K \arrow[r, "\res{K}{L}"] & ( \Si{-n+k\rho} H\underline{\Z})^L
\end{tikzcd}
\end{equation*}
This gives the exact sequence
\begin{equation} \label{res:fixedpts:proof:Z:-4k}
\begin{tikzcd}
\pi_n^K( \Si{k\rho-\beta} H\underline{\Z} ) \arrow[r] & \pi_n^K(\Si{k\rho} H\underline{\Z}) \arrow[r, "\res{K}{L}"] & \pi_n^L(\Si{k\rho} H\underline{\Z}).
\end{tikzcd}
\end{equation}
We wish to show this restriction is surjective. By \cref{krho:fixedpts:K:Z} and \cref{krho-beta:fixedpoints:K:Z}, \cref{res:fixedpts:proof:Z:-4k} becomes
\begin{equation*}
\begin{tikzcd}
\pi_{n+1}^L( \Si{k\rho} H\underline{\Z} ) = 0 \arrow[r] & \mathbb{F}_2^{\frac{1}{2}(n)} \arrow[r, hook] & \mathbb{F}_2^{\frac{1}{2}(n)+1} \arrow[r, "\res{K}{L}"] & \mathbb{F}_2
\end{tikzcd}
\end{equation*}
Thus, \(\res{K}{L}\) must be surjective. A similar argument shows that \(\res{K}{R}\) and \(\res{K}{D}\) are surjective as well.
All that remains to be shown is that the restrictions have distinct kernels. Consider the cofiber sequence
\begin{equation*}
\begin{tikzcd}
\Si{-n+k\rho-\beta} H\underline{\Z} \arrow[r] & \Si{-n+k\rho} H\underline{\Z} \arrow[r] & K/L_+ \wedge \Si{-n+k\rho} H\underline{\Z}.
\end{tikzcd}
\end{equation*}
This results in the exact sequence
\begin{equation} \label{exactseq:-4kproof:K:Z}
\begin{tikzcd}
\ul\pi_{n} ( \Si{k\rho-\beta} H\underline{\Z} ) \arrow[r] & \ul\pi_{n} ( \Si{k\rho} H\underline{\Z}) \arrow[r] & \uparrow^K_L \downarrow^K_L \ul\pi_{n} (\Si{k\rho} H\underline{\Z} )
\end{tikzcd}
\end{equation}
which by \cref{ksigma+r:homotopy:C2:Z}, \cref{krho:fixedpts:K:Z}, and \cref{krho-beta:fixedpoints:K:Z} is
\begin{equation*}
\begin{tikzcd}[scale cd = 0.85,column sep = small]
&& \mathbb{F}_2^{\frac{n}{2}} \arrow[rrr, "\varphi"]
\arrow[d, black] \arrow[dr, black]
&&[+15pt]
& \mathbb{F}_2^{\frac{n}{2}+1} \arrow[rrr, "\res{K}{L}"]
\arrow[d, black] \arrow[dr, black] \arrow[dl, black]
&& [+15pt]
& \mathbb{F}_2
\arrow[dl, black, bend right = 10]
%
&& \\ [-5pt]
& 0
& \mathbb{F}_2
& \mathbb{F}_2 \arrow[r, "0 \text{ } 1 \text{ } 1"]
& \mathbb{F}_2
& \mathbb{F}_2
& \mathbb{F}_2 \arrow[r, "\Delta \text{ } 0 \text{ } 0"]
&
\mathbb{F}_2[K/L]
\arrow[ur, black, bend right = 10]
& 0
& 0
& \\ [-5pt]
&& 0 \arrow[rrr, "0"]
&&
& 0 \arrow[rrr, "0"]
&&
& 0
&&
\end{tikzcd}
\end{equation*}
Each restriction is surjective with kernel of rank \(\frac{n}{2}\). As
\begin{align*}
\dim \pi_{n}^K ( \Si{k\rho} H\underline{\Z}) = \frac{n}{2}+1,
\end{align*}
it is sufficient to show that the kernels are pairwise distinct.
Because the diagram on the left commutes, we find that
\begin{align*}
\text{im }\varphi \cap \ker(\res{K}{R}) = \{0\} \qquad \text{and} \qquad \text{im }\varphi \cap \ker(\res{K}{D}) = \{0\}.
\end{align*}
As \(\text{im }\varphi = \ker(\res{K}{L})\), we have that \(\ker(\res{K}{L})\) is distinct from \(\ker(\res{K}{R})\) and \(\ker(\res{K}{D})\). Replacing \(\beta\) by \(\alpha\) shows that \(\ker(\res{K}{R})\) and \(\ker(\res{K}{D})\) are distinct as well.
\end{proof}
\begin{prop} \label{krho:fixedpts:K:Z}
For \(G=K\) and \(k\geq 1\),
\begin{align*}
\pi_i^K( \Si{k\rho} H\underline{\Z}) &= \left\lbrace \begin{array}{ll}
\mathbb{Z} & i=4k \\
\mathbb{F}_2^{\frac{1}{2}(4k-1-i)} & i\in [2k+1, 4k-1] \text{ odd} \\
\mathbb{F}_2^{\frac{1}{2}(4k+2-i)} & i\in[2k+2, 4k-2] \text{ even} \\
\mathbb{F}_2^{i-k+1} & i\in[k,2k]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
Note that
\begin{align*}
P^{-4k}_{-4k}(\Si{-4k} H\underline{\Z}) &\simeq \Si{-k\rho} H\ul\pi_{4k}(\Si{k\rho} H\underline{\Z}).
\end{align*}
Thus, by \cref{-nslice:K:Z}, \(H\ul\pi_{4k}(\Si{k\rho} H\underline{\Z}) \simeq H\underline{\Z}\). So the result holds for \(i=4k\) and we only need consider \(i\leq 4k-1\).
We will use the resulting long exact sequences in homotopy resulting from the cofiber sequences
\begin{align}
(K/L_+ &\rightarrow S^0 \rightarrow S^\beta) \wedge \Si{k\rho+1} H\underline{\Z} \label{cofib:beta} \\
(K/R_+ &\rightarrow S^0 \rightarrow S^\alpha) \wedge \Si{k\rho+\beta+1} H\underline{\Z} \label{cofib:alpha} \\
(K/D_+ &\rightarrow S^0 \rightarrow S^\gamma) \wedge \Si{k\rho+\alpha+\beta+1} H\underline{\Z} \label{cofib:gamma} \\
\Si{k\rho} H\underline{\Z} &\xrightarrow{2}{} \Si{k\rho} H\underline{\Z} \rightarrow \Si{k\rho} H\underline{\F} \label{cofib:mult2}
\end{align}
where \cref{cofib:mult2} is induced by the short exact sequence of Mackey functors \(\underline{\Z} \xrightarrow{2}{} \underline{\Z} \rightarrow \underline{\F}\).
For \(k=0\), \cref{cofib:beta} - \cref{cofib:gamma} provide that
\begin{align*}
\ul\pi_n(\Si{\rho} H\underline{\Z}) &= \left\lbrace \begin{array}{cc}
\underline{\Z} & n=4 \\
\ul 0 & n=3 \\
\ul{mg} & n=2 \\
\underline{g} & n=1
\end{array}\right.
\end{align*}
Consequently, the result holds for \(k=1\). We now proceed by induction on \(k\). Assume the result holds for \(k\). By restriction to \(L\), \(D\), and \(R\), we find that \(\ul\pi_i(\Si{(k+1)\rho} H\underline{\Z})\) is a pullback over \(K\) for \(i\leq 4k+3\) odd, and consequently, 2-torsion as in \cite[Remark 2.13]{Z}.
In the long exact sequence of fixed points resulting from \cref{cofib:beta} - \cref{cofib:gamma}, we have at the \(K/K\) level,
\begin{equation*}
\begin{tikzcd}[scale cd = 0.9]
\pi_i^{C_2} (\Si{2k\rho_{C_2}+1} H_{C_2}\underline{\Z}) \arrow[r] &[-10pt] \pi_i^K( \Si{k\rho+1} H\underline{\Z}) \arrow[r] &[-10pt] \pi_{i}^K(\Si{k\rho+\beta + 1} H\underline{\Z}) \\
\pi_i^{C_2} (\Si{(2k+1)\rho_{C_2}} H_{C_2}\underline{\Z}) \arrow[r] & \pi_i^K( \Si{k\rho+\beta+1} H\underline{\Z}) \arrow[r] & \pi_{i}^K(\Si{k\rho+\alpha+\beta +1 } H\underline{\Z}) \\
\pi_i^{C_2} (\Si{(2k+2)\rho_{C_2}-1} H_{C_2}\underline{\Z}) \arrow[r] & \pi_i^K( \Si{k\rho+\alpha+\beta+ 1} H\underline{\Z}) \arrow[r] & \pi_{i}^K(\Si{(k+1)\rho} H\underline{\Z})
\end{tikzcd}
\end{equation*}
When \(2k+1\leq i \leq 4k+1\), by \cref{ksigma+r:homotopy:C2:Z},
\begin{align*}
\pi_i^{C_2} (\Si{2k\sigma + 2k+1} H_{C_2}\underline{\Z}) &= \left\lbrace \begin{array}{ll}
\mathbb{F}_2 & i \text{ odd} \\
0 & i \text{ even}
\end{array}\right.
\end{align*}
Now for \(i\) even, we have in our long exact sequence,
\begin{equation*}
\begin{tikzcd}
0 \ar[r] & \pi_i^K( \Si{k\rho+1} H\underline{\Z}) \ar[r, hook] & \pi_{i}^K(\Si{k\rho+\beta + 1} H\underline{\Z}) \ar[r] & \phantom{a} \\
\mathbb{F}_2 \ar[r] & \pi_{i-1}^K( \Si{k\rho+1} H\underline{\Z}) \ar[r, twoheadrightarrow] & \pi_{i-1}^K(\Si{k\rho+\beta + 1} H\underline{\Z}) \ar[r, "0", twoheadrightarrow] & \phantom{a}
\end{tikzcd}
\end{equation*}
Consequently, when \(i\) is even,
\begin{align*}
\text{2-rk } \pi_{i}^K(\Si{k\rho+\beta + 1} H\underline{\Z}) \leq \text{2-rk } \pi_i^K( \Si{k\rho+1} H\underline{\Z}) + 1,
\end{align*}
where equality occurs if \(\pi_i^K( \Si{k\rho+\beta + 1} H\underline{\Z}) = \pi_i^K( \Si{(k\rho+1} H\underline{\Z}) \oplus \mathbb{F}_2\).
And, when \(i\) is odd,
\begin{align*}
\text{2-rk } \pi_{i}^K(\Si{k\rho+\beta + 1} H\underline{\Z}) \leq \text{2-rk } \pi_i^K( \Si{k\rho+1} H\underline{\Z}).
\end{align*}
For \(i\leq 2k\) we have
\begin{equation*}
\begin{tikzcd}
0 \ar[r] & \pi_{i}^K( \Si{k\rho+1} H\underline{\Z}) \ar[r, hook, two heads] & \pi_{i}^K(\Si{k\rho+\beta + 1} H\underline{\Z}) \ar[r, "0", twoheadrightarrow] & \phantom{a}
\end{tikzcd}
\end{equation*}
Thus, for \(i\leq 2k\),
\begin{align*}
\pi_{i}^K(\Si{k\rho+\beta + 1} H\underline{\Z}) \cong \pi_{i}^K( \Si{k\rho+1} H\underline{\Z}).
\end{align*}
A similar statement holds for
\begin{align*}
\text{2-rk } \pi_{i}^K(\Si{ k\rho+\alpha+\beta +1} H\underline{\Z}) \qquad \text{ and } \qquad \text{2-rk } \pi_{i}^K(\Si{(k+1)\rho} H\underline{\Z}).
\end{align*}
From \cite[Corollary 7.2]{GY}, in the long exact sequence resulting from \cref{cofib:mult2}, we have
\begin{align*}
\mathbb{F}_2^l \cong \pi_{2i+1}^K(\Si{(i+1)\rho} H\underline{\Z}) &\xrightarrow{0} \pi_{2i+1}^K(\Si{(i+1)\rho} H\underline{\Z}) \hookrightarrow \pi_{2i+1}^K(\Si{(i+1)\rho} H\underline{\F}) \cong \mathbb{F}_2^{2i+1} \\
&\twoheadrightarrow \pi_{2i}^K(\Si{(i+1)\rho} H\underline{\Z}) \cong \pi_{2i-1}^K(\Si{i\rho} H\underline{\Z}) \cong \mathbb{F}_2^{i}.
\end{align*}
Consequently, \(l=i+1\) and \(\pi_{2i+1}^K(\Si{(i+1)\rho} H\underline{\Z})\cong \mathbb{F}_2^{i+1}\). We then have in our long exact sequence
\begin{align*}
\pi_{2i+2}^K(\Si{(i+1)\rho} H\underline{\Z}) \rightarrow \pi_{2i+2}^K(\Si{(i+1)\rho} H\underline{\F}) \cong \mathbb{F}_2^{2i+3} \twoheadrightarrow \mathbb{F}_2^{i+1} \cong \pi_{2i+1}^K(\Si{(i+1)\rho} H\underline{\Z}).
\end{align*}
Thus,
\begin{align*}
\text{2-rk } \pi_{2i+2}^K(\Si{(i+1)\rho} H\underline{\Z}) \geq i+2 = \text{2-rk } \pi_{2i+1}^K(\Si{i\rho} H\underline{\Z}) + 3.
\end{align*}
We achieve the maximum bound for \(\text{2-rk } \pi_{2i+2}^K(\Si{(i+1)\rho} H\underline{\Z})\); thus,
\begin{align*}
\pi_{2i+2}^K(\Si{(i+1)\rho} H\underline{\Z}) \cong \pi_{2i+1}^K(\Si{i\rho} H\underline{\Z}) \oplus \mathbb{F}_2^3 \cong \mathbb{F}_2^{i+2}.
\end{align*}
The rest of the result now follows from this long exact sequence in a similar manner.
\end{proof}
\begin{prop} \label{krho-beta:fixedpoints:K:Z}
For \(G=K\) and \(k\geq 1\),
\begin{align*}
\pi_i^K( \Si{k\rho-\beta} H\underline{\Z}) &= \left\lbrace \begin{array}{ll}
\mathbb{F}_2^{\frac{1}{2}(4k-i)} & i\in[2k, 4k-2] \text{ even} \\
\mathbb{F}_2^{\frac{1}{2}(4k-i-1)} & i\in[2k-1, 4k-3] \text{ odd} \\
\mathbb{F}_2^{i-k+1} & i\in[k,2k]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
This follows from a similar argument as in \cref{krho:fixedpts:K:Z}.
\end{proof}
\begin{prop} \label{krho+beta:fixedpoints:K:Z}
For \(G=K\) and \(k\geq 1\),
\begin{align*}
\pi_i^K( \Si{k\rho+\beta} H\underline{\Z}) &= \left\lbrace \begin{array}{ll}
\mathbb{F}_2^{\frac{1}{2}(4k-i)+1} & i\in[2k+2, 4k] \text{ even} \\
\mathbb{F}_2^{\frac{1}{2}(4k-i+1)} & i\in[2k+1, 4k-1] \text{ odd} \\
\mathbb{F}_2^{i-k+1} & i\in[k,2k]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
This follows from a similar argument as in \cref{krho:fixedpts:K:Z}.
\end{proof}
\subsection{The \texorpdfstring{\((-4k-2)\)}{(-4k-2)}-slices} \label{-4k-2slice:K:Z}
\phantom{a}
We now determine the \((-4k-2)\)-slices of \(\Si{-n} H\underline{\Z}\).
\begin{prop} \label{-4k-2slices:K:Z}
For even \(n\), \(\Si{-n} H\underline{\Z}\) has no \((-4k-2)\)-slices, except for possibly the \(-n\) slice.
For odd \(n\geq 1\),
\begin{align*}
P^{-4k-2}_{-4k-2} ( \Si{-n} H\underline{\Z} ) &\simeq \left\lbrace
\begin{array}{ll}
\Si{-k\rho-1} H\ul m^* & 4k+2=n+1 \\
\Si{-k\rho-1} H\phi^*_{LDR} \underline{\F}^* & 4k+2\in (n+1,2n]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
When \(n\) is even, the slices in \cref{-4kslices:K:Z} restrict to the all of the appropriate slices for \(\Si{-n} H_{C_2}\underline{\Z}\). If, then, \(\Si{-n} H_K\underline{\Z}\) has a \((-4k-2)\)-slice, it must be a pullback over \(K\). But this is a contradiction as such slices are \((-4k)\)-slices. Thus, for \(n\) even, \(\Si{-n} H_K\underline{\Z}\) has no \((-4k-2)\)-slices.
Now let \(n\) be odd. We first handle the case \(4k+2=2n\). By \cref{rho:susp:commutes},
\begin{align*}
P^{-4k-2}_{-4k-2}(\Si{-n} H_K\underline{\Z}) \simeq \Si{-(k+1)\rho} P^2_2( \Si{-n+(k+1)\rho} H_K\underline{\Z}).
\end{align*}
For clarity, let
\begin{align*}
X := P^2_2( \Si{-n+(k+1)\rho} H_K\underline{\Z}).
\end{align*}
Now from \cref{-2kslice:C2:HZ},
\begin{align*}
i_{H}^* X \simeq \Si{1} H_{C_2}\underline{g}.
\end{align*}
where \(H\) is \(L\), \(D\), or \(R\). Thus, by \cref{2slice:characterization},
\begin{align*}
\ul\pi_2( X) = \phi^*_{K} B \qquad \text{ and } \qquad \ul\pi_1( X ) = \ul A
\end{align*}
where \(B\) is some group and
\begin{equation*}
\ul A = \begin{tikzcd}[ampersand replacement=\&, column sep=scriptsize, bend right=15]
\& A_K \ar[dl] \ar[d] \ar[dr] \& \\
\mathbb{F}_2 \ar[ur] \& \mathbb{F}_2 \ar[u] \& \mathbb{F}_2 \ar[ul] \\
\& 0 \&
\end{tikzcd}
\end{equation*}
and \(A_K \rightarrow \mathbb{F}_2\oplus \mathbb{F}_2\oplus \mathbb{F}_2\) is injective. That is, \(A_K=\mathbb{F}_2^n\) with \(0\leq n\leq 3\).
If \(B\neq 0\), then \(X\) cannot be a 2-slice. Consequently, \(X\simeq \Si{1} H\ul A\).
Because \(\underline{\Z}\) is invariant under the automorphisms of \(K\), the spectrum \(\Si{-n} H\underline{\Z}\) is as well. Therefore, the slices of \(\Si{-n} H\underline{\Z}\) are also invariant under the automorphisms of \(K\).
Thus, \(\ul A\) must be one of the following:
\begin{align*}
\phi^*_{LDR} \ul f, \quad \ul m, \quad \ul{mg}, \quad \text{ or } \quad \phi^*_{LDR} \underline{\F}.
\end{align*}
Except for degree \(-k-1\), \(\Si{-(k+1)\rho +1} H\ul A\) has the same homotopy for each choice of \(A\).
For degree \(-k-1\), \(\ul\pi_{-k-1}(\Si{-(k+1)\rho +1} H\ul A)\) is a pullback over \(K\) of dimension 3, 2, 1, or 0.
From \cref{phiF:SSS} we find that \(\pi_{-k-1}^K(\Si{-(k+1)\rho +1} H\ul A)=0\). The only choice of \(\ul A\) that meets this requirement is \(\phi^*_{LDR}\underline{\F}\).
Now let \(4k+2\in[n+1,2n-1]\). The base cases are established in \cref{tower:-1:K:Z}, \cref{tower:-3:K:Z}, and \cref{tower:-5:K:Z}. The result then follows from \cref{slice:reduction:ZZ^*}.
\end{proof}
\section{Slices of \texorpdfstring{\(\Si{n} H\ul \mathbb{Z}\)}{HZ}} \label{sec:slices:n:K:Z}
Recall from \cref{slice:K:Z1-5:Z*1-4} that \(\Si{n} H\underline{\Z}\) is a slice for \(1\leq n\leq 5\).
\begin{prop} \label{slice:reduction:K:n:Z}
Let \(n\geq 6\). Then
\begin{align*}
P^k_k(\Si{n} H\underline{\Z}) &\simeq \Si{\rho} P^{k-4}_{k-4} (\Si{n-4} H\underline{\Z})
\end{align*}
for \(k\in[n,2n-7]\).
\end{prop}
\begin{proof}
We employ a similar argument as in \cref{slice:reduction:ZZ^*}. Note that
\begin{align*}
P^k_k(\Si{n} H\underline{\Z}) &\simeq \Si{\rho} P^{k-4}_{k-4}(\Si{n-\rho} H\underline{\Z}) \simeq \Si{\rho} P^{k-4}_{k-4}(\Si{n-4} H\underline{\Z}^*).
\end{align*}
Consequently, it is sufficient to compare the \((k-4)\)-slices of \(\Si{n-4} H\underline{\Z}\) and \(\Si{n-4} H\underline{\Z}^*\).
The exact sequence \(\underline{\Z}^* \rightarrow \underline{\Z} \rightarrow \underline{\mathbb{M}}\) provides the fiber sequence
\begin{align*}
\Si{j-1} H\underline{\mathbb{M}} \rightarrow \Si{j} H\underline{\Z}^* \xrightarrow{\iota}{} \Si{j} H\underline{\Z}.
\end{align*}
Then, because \(\Si{j-1} H\underline{\mathbb{M}}\geq 2j-2\), by \cite[Lemma 4.28]{HHR}, \(\iota\) induces an equivalence on slices strictly below level \(2j-2\). Taking \(j=n-4\) gives the result.
\end{proof}
\begin{exmp} \label{tower:6:K:Z}
The tower for \(\Si{6} H\underline{\Z}\) is
\begin{equation*}
\begin{tikzcd}
P^8_8 = \Si{2} H\underline{g} \arrow[r] & \Si{6} H\underline{\Z} \arrow[d] \\
& P^6_6 = \Si{\rho+2} H\underline{\Z}(2,1)^*
\end{tikzcd}
\end{equation*}
\end{exmp}
\begin{proof}
We have the short exact sequence \(\underline{\Z}^* \rightarrow \underline{\Z}(2,1)^* \rightarrow \underline{g}\). This leads to the cofiber sequence
\begin{equation*}
\begin{tikzcd}
P^8_8 = \Si{2} H\underline{g} \arrow[r] & \Si{6} H\underline{\Z} \simeq \Si{\rho+2} H\underline{\Z}^* \arrow[r] & \Si{\rho+2} H\underline{\Z}(2,1)^* = P^6_6.
\end{tikzcd}
\end{equation*}
\end{proof}
\begin{exmp} \label{tower:7:K:Z}
The tower for \(\Si{7} H\underline{\Z}\) is
\begin{equation*}
\begin{tikzcd}
P^{12}_{12} = \Si{3} H\underline{g} \arrow[r] & \Si{7} H\underline{\Z} \arrow[d] \\
P^{8}_8 = \Si{\rho+2} H\ul m \arrow[r] & \Si{\rho+3} H\underline{\Z}(2,1)^* \arrow[d] \\
& P^7_7 = \Si{\rho+3} H\underline{\Z}
\end{tikzcd}
\end{equation*}
\end{exmp}
\begin{proof}
We suspend the tower in \cref{tower:6:K:Z} by 1 and augment with the cofiber sequence \(\Si{\rho+2} H\ul m \rightarrow \Si{\rho+3} H\underline{\Z}(2,1)^* \rightarrow \Si{\rho+3} H\underline{\Z}\) which arises from the short exact sequence \(\underline{\Z}(2,1)^* \rightarrow \underline{\Z} \rightarrow \ul m\).
\end{proof}
We now determine the slices of \(\Si{n} H\underline{\Z}\).
\begin{thm} \label{duality:slices:K:Z}
Let \(n\geq 6\). For \(k\geq n+2\),
\begin{align*}
P^k_k (\Si{n} H\underline{\Z}) &\simeq \Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-k+4}_{-k+4} \, \Si{-n+5} H\underline{\Z}.
\end{align*}
\end{thm}
\begin{proof}
Take \(r\equiv n-5\pmod 4\) with \(1\leq r\leq 4\). We may map the top slice of \(\Si{-n+5} H\underline{\Z}\) into it to find the cofiber sequence
\begin{align}
P^{-n+5}_{-n+5} = \Si{-\frac{n-5-r}{4}\rho -r} H\underline{\Z}^* \rightarrow \Si{-n+5} H\underline{\Z} \rightarrow P^{-n+5-1} \, \Si{-n+5} H\underline{\Z}. \label{slices:cofiber}
\end{align}
Note that all slices of \(P^{-n+4} H\underline{\Z}\) are torsion, so then
\begin{align*}
I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+4} \, H\underline{\Z} = \Si{1} I_{\mathbb{Z}} \, P^{-n+4} \, \Si{-n+5} H\underline{\Z}.
\end{align*}
Apply \(I_\mathbb{Z}\) to \cref{slices:cofiber} and suspend by one to find
\begin{align*}
\Si{1} I_{\mathbb{Z}} \, P^{-n+4} \, \Si{-n+5} H\underline{\Z} \rightarrow \Si{n-4} H\underline{\Z}^* \rightarrow \Si{\frac{n-5-r}{4}\rho +r+1} H\underline{\Z}.
\end{align*}
We can rewrite this as
\begin{align*}
I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+4} \, \Si{-n+5} H\underline{\Z} \rightarrow \Si{n-\rho} H\underline{\Z} \rightarrow \Si{\frac{n-5-r}{4}\rho +r+1} H\underline{\Z}.
\end{align*}
Finally, suspend by \(\rho\) to obtain
\begin{align}
\Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+4} \,\Si{-n+5} H\underline{\Z} \rightarrow \Si{n} H\underline{\Z} \rightarrow \Si{\frac{n-1-r}{4}\rho +r+1} H\underline{\Z}. \label{slices:cofiber:1}
\end{align}
Note that \(\Si{\frac{n-1-r}{4}\rho +r+1} H\underline{\Z}\) is an \(n\)-slice and
\begin{align*}
\Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+4} \, \Si{-n+5} H\underline{\Z}\in [n,4(n-4)].
\end{align*}
Furthermore, if \(n\not\equiv 2\pmod 4\),
\begin{align*}
\Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+4} \, \Si{-n+5} H\underline{\Z}\in [n+1,4(n-4)].
\end{align*}
From \cref{duality:prop},
\begin{align*}
I_{\mathbb{Q}/\mathbb{Z}} \, P^{-k}_{-k} \, \Si{-n+5} H\underline{\Z} &\simeq P^{k}_{k} I_{\mathbb{Q}/\mathbb{Z}} \, \Si{-n+5} H\underline{\Z}.
\end{align*}
Consequently, \(\Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+4} \,\Si{-n+5} H\underline{\Z}\) provides all slices of \(\Si{n} H\underline{\Z}\).
Now suppose \(n\equiv 2\pmod 4\) so that \(r=1\). Then from \cref{duality:prop},
\begin{align*}
P^n_n ( \Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+4} \, \Si{-n+5} H\underline{\Z} ) &\simeq \Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n}_{-n} ( P^{-n+4} \, \Si{-n+5} H\underline{\Z} ) \\
&\simeq \Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, \Si{-(\frac{n-6}{4}+1)\rho+1} H\ul {mg} \\
&\simeq \Si{\frac{n-2}{4}\rho + 1} H\ul m
\end{align*}
Apply \(P^n_n(-)\) to \cref{slices:cofiber:1} to get the extension
\begin{align*}
\Si{\frac{n-2}{4}\rho + 1} H\ul m \rightarrow P^n_n \Si{n} H\underline{\Z} \rightarrow \Si{\frac{n-2}{4}\rho +2} H\underline{\Z}
\end{align*}
and the fiber sequence
\begin{align*}
\Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+5} H\underline{\Z} \rightarrow \Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+4} H\underline{\Z} \rightarrow P^n_n \Si{n} H\underline{\Z}.
\end{align*}
Now \(\Si{\rho} I_{\mathbb{Q}/\mathbb{Z}} \, P^{-n+5} H\underline{\Z} \in [n+1,4(n-4)]\) and thus supplies the remaining slices of \(\Si{n} H\underline{\Z}\).
\end{proof}
\begin{prop} \label{slices:K:n:Z}
Let \(n\geq 6\) and set \(r\equiv n\pmod 4\) with \(2\leq r\leq 5\). The \(n\)-slice of \(\Si{n} H\underline{\Z}\) is
\begin{align*}
P^n_n \Si{n} H\underline{\Z} &\simeq \left\lbrace \begin{array}{ll}
\Si{\frac{n-2}{4}\rho + 2} H\underline{\Z}(2,1)^* & n\equiv 2\pmod 4 \\
\Si{\frac{n-r}{4}\rho + r} H\underline{\Z} & \text{otherwise}
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
For \(n\not\equiv 2\pmod 4\), this follows from \cref{slices:cofiber:1}. When \(n\equiv 2\pmod 4\) it follows from \cref{tower:6:K:Z} and repeated application of \cref{slice:reduction:K:n:Z}.
\end{proof}
\begin{prop} \label{slices:K:4k:Z}
Let \(n\geq 6\). The \(4k\)-slices of \(\Si{n} H\underline{\Z}\) are
\begin{align*}
P^{4k}_{4k} ( \Si{n} H\underline{\Z} ) \simeq \left\lbrace \begin{array}{ll}
\Si{k\rho} H\ul{mg}^* & 4k=n+1 \\
\Si{k} H\underline{g}^{\frac{1}{2}(4k-n)} & \begin{aligned}
4k\in [n+2,2n-8], \\ n\equiv 0\pmod 2
\end{aligned} \\
\Si{k\rho} H(\underline{g}^{\frac{1}{2}(4k-n+3)-3} \oplus \phi^*_{LDR} \underline{\F}^* ) &
\begin{aligned} 4k\in[n+2,2n-6], \\ n\equiv 1\pmod 2
\end{aligned} \\
\Si{k} H\underline{g}^{n-k-3} & 4k\in[2n-4,4(n-4)]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
This follows from \cref{-4kslices:K:Z} and \cref{duality:slices:K:Z}.
\end{proof}
\begin{prop} \label{slices:4k+2:K:Z}
For \(n\geq 6\), except for possibly the \(n\) slice,
the nontrivial \((4k+2)\)-slices of \(\Si{n} H\underline{\Z}\) are
\begin{align*}
P^{4k+2}_{4k+2} (\Si{n} H\underline{\Z}) &\simeq \left\lbrace \begin{array}{ll}
\Si{k\rho+1} H\phi^*_{LDR} \underline{\F} &
\begin{aligned} 4k+2\in[n+2,2n-6], \\ n\equiv 0\pmod 2
\end{aligned}
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
By \cref{2kslice:C2:HZ}, \(\Si{n} H_{C_2}\underline{\Z}\) has no \((4k+2)\)-slices except in the range \([n+2,2n-6]\) when \(n\) is even. So for \(4k+2\) not in this range, any such slice must be a pullback over \(K\). But then it is a \(4k\)-slice. For \(4k+2\in[n+2,2n-6]\), the result follows from \cref{-4k-2slices:K:Z} and \cref{duality:slices:K:Z}.
\end{proof}
\subsection{Comparison with the Slices of \texorpdfstring{\(\Si{n} H\underline{\F}\)}{HF}} \label{subsec:compGY}
\phantom{a}
This work is complementary to \cite{GY}, which calculates the slices of \(\Si{n} H\underline{\F}\) for \(n\geq 1\). One would hope that the exact sequence \(\underline{\Z} \xrightarrow{2}{} \underline{\Z} \rightarrow \underline{\F}\) could play a role in recovering the slices of \(\Si{n} H\underline{\Z}\) from the slices of \(\Si{n} H\underline{\F}\) or vice versa, but this is not always the case.
When \(G=C_2\), \cite[Theorem 3.18]{GY} shows that the slices of \(\Si{n} H_{C_2}\underline{\F}\) contain both even and odd suspensions of \(H_{C_2}\underline{g}\), whereas \cref{tower:C2:Z:n} shows that \(\Si{n} H_{C_2}\underline{\Z}\) has only even or odd suspensions of \(H_{C_2}\underline{g}\). This is illustrated in \cref{tab-C2slices}.
\begin{table}[h]
\caption[Comparison of {$C_2$}-slices]{Comparison of $C_2$-slices}
\label{tab-C2slices}
\begin{center}
{\renewcommand{\arraystretch}{1.5}\setlength{\tabcolsep}{15pt}
\begin{tabular}{|c|c|c|}
\hline
Slices of \(\Si{9} H_{C_2}\underline{\Z}\) & Slices of \(\Si{9} H_{C_2}\underline{\F}\) & Slices of \(\Si{10} H_{C_2}\underline{\Z}\) \\ \hline
&\(P^{14}_{14} = \Si{7} H_{C_2}\underline{g}\) & \(P^{14}_{14} = \Si{7} H_{C_2}\underline{g}\) \\
\(P^{12}_{12} = \Si{6} H_{C_2}\underline{g}\) &\(P^{12}_{12} = \Si{6} H_{C_2}\underline{g}\) & \\
&\(P^{10}_{10} = \Si{5} H_{C_2}\underline{g}\) & \(P^{10}_{10} = \Si{2\rho+2} H_{C_2}\underline{\Z}^*\) \\
\(P^{9}_{9} = \Si{2\rho+1} H_{C_2}\underline{\Z}^*\) & \(P^{9}_{9} = \Si{2\rho+1} H_{C_2}\underline{\F}^*\) & \\ \hline
\end{tabular} }
\end{center}
\end{table}
The \(2k\)-slices of \(\Si{n} H_{C_2}\underline{\Z}\) and \(\Si{n+1} H_{C_2}\underline{\Z}\) only combine to give the slices of \(\Si{n} H_{C_2}\underline{\F}\) when \(n\equiv 3,4\pmod 4\). When \(n\equiv 5,6\pmod 4\), the \(\Si{n} H_{C_2}\underline{\Z}\) and \(\Si{n+1} H_{C_2}\underline{\Z}\) slices miss the \((n+r)\)-slice of \(\Si{n} H_{C_2}\underline{\F}\), where \(r=1,2\), respectively. For example, neither \(\Si{9} H_{C_2}\underline{\Z}\) nor \(\Si{10} H_{C_2}\underline{\Z}\) has a slice equivalent to \(\Si{5} H_{C_2}\underline{g}\), but \(\Si{9} H_{C_2}\underline{\F}\) does.
We can recover the \((4k)\)-slices of \(\Si{n} H\underline{\F}\) from the \((4k)\)-slices of \(\Si{n} H_K\underline{\Z}\) and \(\Si{n+1} H_K\underline{\Z}\). As in \cref{krho:fixedpts:K:Z}, we use the sequence \(\underline{\Z} \xrightarrow{2}{} \underline{\Z} \rightarrow \underline{\F}\) to get the cofiber sequence \(\Si{-k\rho} H_K\underline{\Z} \rightarrow \Si{-k\rho} H_K\underline{\F} \rightarrow \Si{1-k\rho} H_K\underline{\Z}\). We then have the long exact sequence in homotopy
\begin{equation} \label{les:ZZF:4kslices}
\begin{tikzcd}[column sep=small, row sep=small]
\ul\pi_{-n}( \Si{-k\rho} H_K\underline{\Z}) \arrow[r, "2"] &
\ul\pi_{-n}( \Si{-k\rho} H_K\underline{\Z}) \arrow[r] & \phantom{\ul\pi_{-n}( \Si{-k\rho} H_K\underline{\Z})} \\
\ul\pi_{-n}( \Si{-k\rho} H_K\underline{\F}) \arrow[r] &
\ul\pi_{-n-1}( \Si{-k\rho} H_K\underline{\Z}) \arrow[r, "2"] &
\ul\pi_{-n-1}( \Si{-k\rho} H_K\underline{\Z}).
\end{tikzcd}
\end{equation}
When \(n\leq 4k-1\), all groups in \cref{les:ZZF:4kslices} are 2-torsion and the middle three terms become the exact sequence
\begin{equation} \label{ses:ZZF:4kslices}
\begin{tikzcd}[row sep=small]
\ul\pi_{-n}( \Si{-k\rho} H_K\underline{\Z}) \arrow[r, hook] &
\ul\pi_{-n}( \Si{-k\rho} H_K\underline{\F}) \arrow[r, twoheadrightarrow] &
\ul\pi_{-n-1}( \Si{-k\rho} H_K\underline{\Z}).
\end{tikzcd}
\end{equation}
When \(n=4k\), the left four terms of \cref{les:ZZF:4kslices} become the exact sequence
\begin{equation*}
\begin{tikzcd}[row sep=small]
\underline{\Z}^* \arrow[r, "2"] &
\underline{\Z}^* \arrow[r, twoheadrightarrow, "2"] &
\ul\pi_{-n-1}( \Si{-k\rho} H_K\underline{\Z}) \arrow[r] & \ul 0.
\end{tikzcd}
\end{equation*}
Consequently, the \((4k)\)-slices of \(\Si{n} H_K\underline{\F}\) are
\begin{align*}
P^{4k}_{4k}( \Si{n} H_K\underline{\F}) &\simeq \left\lbrace
\begin{array}{ll}
\Si{k\rho} H_K\underline{\F}^* & n=4k \\
\Si{k\rho} \ul E_{-n} & n\leq 4k-1
\end{array}\right.
\end{align*}
where \(\ul E_{-n}\) is the middle Mackey functor in \cref{ses:ZZF:4kslices}. This recovery is illustrated in \cref{tab-Kslices}.
\begin{table}[h]
\caption[Comparison of {$K$}-slices]{Comparison of $K$-slices}
\label{tab-Kslices}
\begin{center}
{\renewcommand{\arraystretch}{1.5}\setlength{\tabcolsep}{15pt}
\begin{tabular}{|c|c|c|}
\hline
Slices of \(\Si{7} H\underline{\Z}\) & Slices of \(\Si{7} H\underline{\F}\) & Slices of \(\Si{8} H\underline{\Z}\) \\ \hline
& \(P^{16}_{16} = \Si{4} H\underline{g}\) & \(P^{16}_{16} = \Si{4} H\underline{g}\) \\
\(P^{12}_{12} = \Si{3} H\underline{g}\) & \(P^{12}_{12} = \Si{3} H\underline{g}^3\) & \(P^{12}_{12} = \Si{3} H\underline{g}^2\) \\
&\(P^{10}_{10} = \Si{\rho+3} H\phi^*_{LDR} \underline{\F}\) & \(P^{10}_{10} = \Si{\rho+3} H\phi^*_{LDR} \underline{\F}\) \\
\(P^{8}_{8} = \Si{\rho+2} H\ul m\) & \(P^{8}_{8} = \Si{\rho+2} H\ul m\) & \(P^8_8 = \Si{\rho+4} H\underline{\Z}\) \\
\(P^7_7 = \Si{\rho+3} H\underline{\Z}\) & \(P^7_7 = \Si{\rho+3} H\underline{\F}\) & \\ \hline
\end{tabular} }
\end{center}
\end{table}
Except for the \(n\)-slice, all slices of \(\Si{7} H_K\underline{\F}\) are recovered from the slices of \(\Si{7} H_K\underline{\Z}\) and \(\Si{8} H_K\underline{\Z}\). It is not always the case, however, that the \((4k+2)\)-slices are recovered. For example, \(\Si{10} H_K\underline{\F}\) has a 14-slice (\cite[Example 6.14]{GY}), but neither \(\Si{10} H_K\underline{\Z}\) nor \(\Si{11} H_K\underline{\Z}\) have 14-slices.
\section{Homotopy Mackey Functor Computations} \label{sec:htpycomp}
Here we compute the homotopy Mackey functors of the slices of \(\Si{\pm n}H\underline{\Z}\).
\begin{prop} \label{htpy:krho:Z}
For \(k\geq 1\), the nontrivial homotopy Mackey functors of \(\Si{k\rho} H\underline{\Z}\) are
\begin{align*}
\ul\pi_n ( \Si{k\rho} H\underline{\Z} ) = \left\lbrace \begin{array}{ll}
\underline{\Z} & n=4k \\
\ul{mg} & n=4k-2 \\
\underline{g}^{\frac{1}{2}(4k-n-1)} & \begin{aligned}
n\in [2k,4k-3], \\ \text{ } n\equiv 1\pmod 2
\end{aligned} \\
\underline{g}^{\frac{1}{2}(4k-n+2)-3} \oplus \phi^*_{LDR} \underline{\F} & \begin{aligned}
n\in[2k,4k-3], \\ \text{ } n\equiv 0\pmod 2
\end{aligned} \\
\underline{g}^{n-k+1} & n\in[k,2k-1]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
For \(n\in[k,4k-2]\), this is a restatement of \cref{-4kslices:K:Z}. For \(n=4k\), the result follows from \cref{-nslice:K:Z} and repeated application of \cref{slice:reduction:ZZ^*}.
\end{proof}
\begin{prop} \label{htpy:-krho:Z}
For \(k\geq 1\), the nontrivial homotopy Mackey functors of \(\Si{-k\rho} H\underline{\Z}\) are
\begin{align*}
\ul\pi_{-n} ( \Si{-k\rho} H\underline{\Z} ) = \left\lbrace \begin{array}{ll}
\underline{\Z}^* & n=4k \\
\ul{mg}^* & n=4k-1 \\
\underline{g}^{\frac{1}{2}(4k-n)} & \begin{aligned}
n\in [2k+4,4k-3], \\ \text{ } n\equiv 0\pmod 2
\end{aligned} \\
\underline{g}^{\frac{1}{2}(4k-n+3)-3} \oplus \phi^*_{LDR} \underline{\F}^* & \begin{aligned}
n\in[2k+3,4k-2], \\ \text{ } n\equiv 1\pmod 2
\end{aligned} \\
\underline{g}^{n-k-3} & n\in[k+4,2k+2]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
For \(n\in[k+4,4k-1]\), this is a restatement of \cref{slices:K:4k:Z}. For \(n=4k\), the result follows from \cref{slices:K:n:Z} and repeated application of \cref{slice:reduction:K:n:Z}.
\end{proof}
\begin{prop} \label{htpy:krho:Z*}
For \(k\geq 1\), the nontrivial homotopy Mackey functors of \(\Si{k\rho} H\underline{\Z}^*\) are
\begin{align*}
\ul\pi_n ( \Si{k\rho} H\underline{\Z}^* ) = \left\lbrace \begin{array}{ll}
\underline{\Z} & n=4k \\
\ul{mg} & n=4k-2 \\
\underline{g}^{\frac{1}{2}(4k-n-1)} & \begin{aligned}
n\in [2k+2,4k-3], \\ \text{ } n\equiv 1\pmod 2
\end{aligned} \\
\underline{g}^{\frac{1}{2}(4k-n-2)-3} \oplus \phi^*_{LDR} \underline{\F} & \begin{aligned}
n\in[2k+2,4k-3], \\ \text{ } n\equiv 0\pmod 2
\end{aligned} \\
\underline{g}^{n-k+2} & n\in[k+3,2k+1]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
This follows from the equivalence \(\Si{k\rho} H\underline{\Z}^* \simeq \Si{(k-1)\rho+4} H\underline{\Z}\) and \cref{htpy:krho:Z}.
\end{proof}
\begin{prop} \label{htpy:krho:Z(2,1)*}
For \(k\geq 1\), the nontrivial homotopy Mackey functors of \(\Si{k\rho} H\underline{\Z}(2,1)^*\) are
\begin{align*}
\ul\pi_n ( \Si{k\rho} H\underline{\Z}(2,1)^* ) = \left\lbrace \begin{array}{ll}
\ul\pi_n(\Si{k\rho} H\underline{\Z}^*) & n\in[k+3,4k] \\
\underline{g} & n=k
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
The exact sequence \(\underline{\Z}^*\rightarrow \underline{\Z}(2,1)^* \rightarrow \underline{g}\) and corresponding cofiber sequence \(\Si{k\rho} H\underline{\Z}^* \rightarrow \Si{k\rho} H\underline{\Z}(2,1)^* \rightarrow \Si{k} H\underline{g}\) provide us with a long exact sequence in homotopy. We then have that the homotopy of \(\Si{k\rho} H\underline{\Z}(2,1)^*\) is the homotopy of \(\Si{k\rho} H\underline{\Z}^*\) with an additional \(\underline{g}\) in degree \(k\).
\end{proof}
\begin{prop} \label{htpy:krho:mg*}
For \(k\geq 2\), the nontrivial homotopy Mackey functors of \(\Si{k\rho} H\ul{mg}^*\) are
\begin{align*}
\ul\pi_n ( \Si{k\rho} H\ul{mg}^* ) = \left\lbrace \begin{array}{ll}
\phi^*_{LDR} \underline{\F} & n=2k\\
\underline{g}^{3} & n\in [k+2,2k-1] \\
\underline{g} & n=k+1
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
We have the equivalence \(\Si{k\rho} H\ul{mg}^* \simeq \Si{(k-1)\rho+2} H\ul m\). The result then follows from \cite[Proposition 7.3]{GY}.
\end{proof}
\begin{prop} \label{htpy:-krho:mg}
For \(k\geq 2\), the nontrivial homotopy Mackey functors of \(\Si{-k\rho} H\ul{mg}\) are
\begin{align*}
\ul\pi_{-n} ( \Si{-k\rho} H\ul{mg} ) = \left\lbrace \begin{array}{ll}
\phi^*_{LDR} \underline{\F}^* & n=2k \\
\underline{g}^{3} & n\in [k+2,2k-1] \\
\underline{g} & n=k+1
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
The result follows by taking the Brown-Comenetz dual of each Mackey functor in \cref{htpy:krho:mg*}.
\end{proof}
\begin{prop} \label{htpy:-krho:m}
For \(k\geq 2\), the nontrivial homotopy Mackey functors of \(\Si{-k\rho} H\ul m\) are
\begin{align*}
\ul\pi_{-n} ( \Si{-k\rho} H\ul{m} ) = \left\lbrace \begin{array}{ll}
\phi^*_{LDR} \underline{\F}^* & n=2k \\
\underline{g}^{3} & n\in [k+2,2k-1] \\
\underline{g}^2 & n=k+1
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
First, take the Brown-Comenetz dual of the Mackey functors in \cite[Proposition 7.4]{GY}. The result then follows from the equivalence \(\Si{-\rho} H\ul m \simeq \Si{-2} H\ul{mg}^*\).
\end{proof}
\begin{prop} \label{htpy:krho:phiF*}
We have the equivalences
\begin{align*}
\Si{k\rho} H\phi^*_{LDR}\underline{\F}^* \simeq \left\lbrace \begin{array}{ll}
\Si{2} H\phi^*_{LDR}\ul f & k=1 \\
\Si{4} H\phi^*_{LDR}\underline{\F} & k=2
\end{array}\right.
\end{align*}
Then for \(k\geq 3\), the nontrivial homotopy Mackey functors of \(\Si{k\rho} H\phi^*_{LDR} \underline{\F}^*\) are
\begin{align*}
\ul\pi_{n} ( \Si{k\rho} H\phi^*_{LDR} \underline{\F}^* ) = \left\lbrace \begin{array}{ll}
\phi^*_{LDR} \underline{\F} & n=2k\\
\underline{g}^{3} & n\in [k+2,2k-1]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
This is the pullback over \(L\), \(D\), and \(R\) of the Mackey functors in \cite[Proposition 3.6]{GY}.
\end{proof}
\begin{prop} \label{htpy:-krho:phiF}
We have the equivalences
\begin{align*}
\Si{-k\rho} H\phi^*_{LDR}\underline{\F} \simeq \left\lbrace \begin{array}{ll}
\Si{-2} H\phi^*_{LDR}\ul f & k=1 \\
\Si{-4} H\phi^*_{LDR}\underline{\F}^* & k=2
\end{array}\right.
\end{align*}
Then for \(k\geq 3\), the nontrivial homotopy Mackey functors of \(\Si{-k\rho} H\phi^*_{LDR} \underline{\F}\) are
\begin{align*}
\ul\pi_{-n} ( \Si{-k\rho} H\phi^*_{LDR} \underline{\F} ) = \left\lbrace \begin{array}{ll}
\phi^*_{LDR} \underline{\F}^* & n=2k\\
\underline{g}^{3} & n\in [k+2,2k-1]
\end{array}\right.
\end{align*}
\end{prop}
\begin{proof}
This is the pullback over \(L\), \(D\), and \(R\) of the Mackey functors in \cite[Proposition 3.7]{GY}.
\end{proof}
\section{Spectral Sequences} \label{sec:SSS}
\sseqset{
Fclass/.sseq style = {fill = black, diamond, draw, inner sep = 0.1ex},
classes= {fill, inner sep = 0pt, minimum size = 0.22em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 225, scale = 1.2 },
run off differentials = ->,
grid = go,
grid color = gray!30
Zclass/.sseq style={fill=black, rectangle, draw,inner sep=0.4ex},
Zstarclass/.sseq style={fill=none, rectangle, draw,inner sep=0.4ex},
Z4class/.sseq style={fill=none, circle, draw,inner sep=0.28ex},
Z4classSource/.sseq style={circle, inner sep = 0.28ex, draw, fill = red!70},
Z4classTarget/.sseq style={circle, inner sep = 0.28ex, draw, fill = black!40},
FstarLDRclass/.sseq style = {fill = none, regular polygon, sides=5, draw, inner sep = 0.4ex},
FLDRclass/.sseq style = {fill = black, regular polygon, regular polygon sides=5, draw, inner sep = 0.4ex},
mgstarclass/.sseq style = {fill = none, trapezium, draw, inner sep = 0.25ex},
mgclass/.sseq style = {fill = black, trapezium, draw, inner sep = 0.25ex},
mstarclass/.sseq style = {fill = none, regular polygon, regular polygon sides=3, draw, inner sep = 0.4ex},
}
\begin{sseqdata}[name=SI-1K,
y range={-10}{0},
x range={-10}{0},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zstarclass, "\phantom{*}" {inside,font=\tiny}](-1,-1)
\class[mstarclass](-1,-2)
\class[Z4class,"1" {inside,font=\tiny}](-1,-3)
\structline[dashed](-1,-3)(-1,-2)
\structline[dashed](-1,-2)(-1,-1)
\draw[fill=white](-9.9,-1.1) rectangle (-6.4,0.0);
\node at (-8.1,-0.6) {\Si{-1}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI-3K,
y range={-10}{0},
x range={-10}{0},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zstarclass, "\phantom{*}" {inside,font=\tiny}](-3,0)
\class[FstarLDRclass](-3,-3)
\class[Z4class,"1" {inside,font=\tiny}](-2,-6)
\class[Z4class,"2" {inside,font=\tiny}](-3,-9)
\d2(-2,-6)(-3,-3)
\structline[dashed](-3,-9)(-3,-3)
\structline[dashed](-3,-3)(-3,0)
\draw[fill=white](-9.9,-1.1) rectangle (-6.4,0.0);
\node at (-8.1,-0.6) {\Si{-3}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI-5K,
y range={-16}{0},
x range={-10}{0},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zstarclass, "\phantom{*}" {inside,font=\tiny}](-5,0)
\class[mgstarclass, "\phantom{*}" {inside,font=\tiny}](-4,-1)
\class[FstarLDRclass](-3,-3)
\class[FstarLDRclass](-5,-5)
\foreach \x in {3,...,1} {
\class[Z4class, "\x" {inside,font=\tiny}](-6+\x,-18+3*\x)
}
\class[Z4class,"1" {inside,font=\tiny}](-3,-2)
\class[Z4class,"1" {inside,font=\tiny}](-2,-4)
\class[Z4class,"1" {inside,font=\tiny}](-2,-6)
\class[Z4class,"3" {inside,font=\tiny}](-4,-6)
\d2(-2,-4)(-3,-2)
\d2(-3,-3)(-4,-1)
\replacesource
\d3(-2,-6)(-3,-3)
\d3(-3,-9)(-4,-6)
\d7(-4,-12)(-5,-5)
\structline[dashed](-5,-15)(-5,-5)
\structline[dashed](-5,-5)(-5,0)
\draw[fill=white](-9.9,-1.1) rectangle (-6.4,0.0);
\node at (-8.1,-0.6) {\Si{-5}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI-7K,
y range={-28}{0},
x range={-9}{0},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zstarclass, "\phantom{*}" {inside,font=\tiny}](-7,0)
\class[mgstarclass, "\phantom{*}" {inside,font=\tiny}](-6,-1)
\class[FstarLDRclass](-5,-5)
\class[FstarLDRclass](-7,-7)
\class[Z4class,"1" {inside,font=\tiny}](-5,-2)
\foreach \x in {1,...,4} {
\class[Z4class, "\x" {inside,font=\tiny}](-8+\x,-24+3*\x)
}
\foreach \x in {2,...,3} {
\class[Z4class, "\x" {inside,font=\tiny}](-1-\x,-15+3*\x)
}
\foreach \x in {0,...,1} {
\class[Z4class, "3" {inside,font=\tiny}](-6+\x,-8-\x)
}
\d3(-3,-9)(-4,-6)
\replacetarget
\d3(-4,-12)(-5,-9)
\replacesource
\d4(-5,-5)(-6,-1)
\replacesource
\d4(-4,-6)(-5,-2)
\d7(-4,-12)(-5,-5)
\d7(-5,-15)(-6,-8)
\d11(-6,-18)(-7,-7)
\structline[dashed](-7,-21)(-7,-7)
\structline[dashed](-7,-7)(-7,0)
\draw[fill=white](-3.9,-1.1) rectangle (-0.4,0.0);
\node at (-2.1,-0.6) {\Si{-7}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI-9K,
y range={-28}{0},
x range={-11}{0},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zstarclass, "\phantom{*}" {inside,font=\tiny}](-9,0)
\class[mgstarclass, "\phantom{*}" {inside,font=\tiny}](-8,-1)
\class[FstarLDRclass](-6,-3)
\foreach \x in {0,1,2} {
\class[FstarLDRclass](-9+2*\x,-9+2*\x)
}
\class[Z4class,"1" {inside,font=\tiny}](-3,-7)
\class[Z4class,"1" {inside,font=\tiny}](-3,-9)
\class[Z4class,"3" {inside,font=\tiny}](-4,-6)
\class[Z4class,"3" {inside,font=\tiny}](-4,-12)
\class[Z4class,"1" {inside,font=\tiny}](-4,-5)
\class[Z4class,"2" {inside,font=\tiny}](-5,-4)
\class[Z4class,"1" {inside,font=\tiny}](-7,-2)
\foreach \x in {1,...,5} {
\class[Z4class, "\x" {inside,font=\tiny}](-10+\x,-30+3*\x)
}
\foreach \x in {0,...,2} {
\class[Z4class, "3" {inside,font=\tiny}](-6-\x,-12+\x)
}
\foreach \x in {0,...,1} {
\class[Z4class, "3" {inside,font=\tiny}](-5-\x,-9+\x)
}
\d2(-4,-6)(-5,-4)
\replacesource
\d3(-4,-9-3*1)(-4-1,-6-3*1)
\d2(-3-2,-7+2)(-4-2,-5+2)
\d3(-3-2,-9-3*2)(-4-2,-6-3*2)
\replacesource
\foreach \x in {0} {
\d2(-3-\x,-7+\x)(-4-\x,-5+\x)
\d3(-3-\x,-9-3*\x)(-4-\x,-6-3*\x)
}
\foreach \x in {0,1} {
\d6(-6-\x,-8+\x)(-7-\x,-2+\x)
\replacesource
\d7(-5-\x,-15-3*\x)(-6-\x,-8-3*\x)
\replacesource
}
\d11(-6,-18)(-7,-7)
\d11(-7,-21)(-8,-10)
\d15(-8,-24)(-9,-9)
\structline[dashed](-9,-27)(-9,-9)
\structline[dashed](-9,-9)(-9,0)
\draw[fill=white](-3.9,-1.1) rectangle (-0.4,0.0);
\node at (-2.1,-0.6) {\Si{-9}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI6K,
y range={0}{10},
x range={0}{10},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zclass, "\phantom{*}" {inside,font=\tiny}](6,0)
\class[Z4class,"1" {inside,font=\tiny}](2,6)
\class[Z4class,"1" {inside,font=\tiny}](3,3)
\d3(3,3)(2,6)
\draw[fill=white](7.5,8.5) rectangle (10.3,9.6);
\node at (8.9,9.0) {\Si{6}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI7K,
y range={0}{10},
x range={0}{10},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zclass, "\phantom{*}" {inside,font=\tiny}](7,0)
\class[mgclass, "\phantom{*}" {inside,font=\tiny}](5,2)
\class[FLDRclass](4,4)
\class[Z4class,"1" {inside,font=\tiny}](4,3)
\class[Z4class,"1" {inside,font=\tiny}](3,5)
\class[Z4class,"1" {inside,font=\tiny}](3,9)
\d2(5,2)(4,4)
\replacetarget
\d2(4,3)(3,5)
\d5(4,4)(3,9)
\draw[fill=white](7.5,8.5) rectangle (10.3,9.6);
\node at (8.9,9) {\Si{7}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI8K,
y range={0}{13},
x range={0}{10},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zclass, "\phantom{*}" {inside,font=\tiny}](8,0)
\class[mgclass, "\phantom{*}" {inside,font=\tiny}](6,2)
\class[FLDRclass](5,5)
\class[Z4class,"3" {inside,font=\tiny}](4,6)
\class[Z4class,"2" {inside,font=\tiny}](3,9)
\class[Z4class,"1" {inside,font=\tiny}](5,3)
\d3(5,3)(4,6)
\d3(4,6)(3,9)
\class[Z4class,"1" {inside,font=\tiny}](4,12)
\d3(6,2)(5,5)
\replacetarget
\d7(5,5)(4,12)
\draw[fill=white](7.5,9.5) rectangle (10.3,10.6);
\node at (8.9,10) {\Si{8}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI10K,
y range={0}{19},
x range={0}{12},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zclass, "\phantom{*}" {inside,font=\tiny}](10,0)
\class[mgclass, "\phantom{*}" {inside,font=\tiny}](8,2)
\class[FLDRclass](7,7)
\class[Z4class,"1" {inside,font=\tiny}](6,18)
\class[Z4class,"2" {inside,font=\tiny}](5,15)
\class[Z4class,"1" {inside,font=\tiny}](3,9)
\class[Z4class,"3" {inside,font=\tiny}](6,8)
\class[Z4class,"3" {inside,font=\tiny}](4,12)
\class[Z4class,"3" {inside,font=\tiny}](5,9)
\class[Z4class,"1" {inside,font=\tiny}](4,6)
\class[Z4class,"1" {inside,font=\tiny}](7,3)
\d3(4,6)(3,9)
\d3(5,9)(4,12)
\d5(7,3)(6,8)
\replacetarget
\d7(6,8)(5,15)
\d5(8,2)(7,7)
\replacetarget
\d11(7,7)(6,18)
\draw[fill=white](9.5,18.5) rectangle (12.3,19.6);
\node at (10.9,19.0) {\Si{10}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI11K,
y range={0}{21},
x range={0}{13},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zclass, "\phantom{*}" {inside,font=\tiny}](11,0)
\class[mgclass, "\phantom{*}" {inside,font=\tiny}](9,2)
\class[FLDRclass](7,4)
\class[FLDRclass](8,8)
\class[FLDRclass](6,6)
\class[Z4class,"1" {inside,font=\tiny}](7,21)
\class[Z4class,"2" {inside,font=\tiny}](6,18)
\class[Z4class,"3" {inside,font=\tiny}](5,15)
\class[Z4class,"3" {inside,font=\tiny}](7,9)
\class[Z4class,"3" {inside,font=\tiny}](6,10)
\class[Z4class,"1" {inside,font=\tiny}](4,12)
\class[Z4class,"3" {inside,font=\tiny}](5,7)
\class[Z4class,"1" {inside,font=\tiny}](4,8)
\class[Z4class,"1" {inside,font=\tiny}](8,3)
\class[Z4class,"2" {inside,font=\tiny}](6,5)
\class[Z4class,"1" {inside,font=\tiny}](5,6)
\d2(5,6)(4,8)
\d2(7,4)(6,6)
\d2(6,5)(5,7)
\replacetarget
\d5(5,7)(4,12)
\d5(6,10)(5,15)
\d6(8,3)(7,9)
\replacetarget
\d6(9,2)(8,8)
\replacetarget
\d9(7,9)(6,18)
\d13(8,8)(7,21)
\draw[fill=white](10.5,20.5) rectangle (13.3,21.6);
\node at (11.9,21.0) {\Si{11}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI12K,
y range={0}{25},
x range={0}{12},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zclass, "\phantom{*}" {inside,font=\tiny}](12,0)
\class[mgclass, "\phantom{*}" {inside,font=\tiny}](10,2)
\class[FLDRclass](8,4)
\class[FLDRclass](9,9)
\class[FLDRclass](7,7)
\foreach \x in {1,...,4} {
\class[Z4class, "\x" {inside,font=\tiny}](9-\x,27-3*\x)
}
\foreach \x in {0,...,2} {
\class[Z4class, "3" {inside,font=\tiny}](8-\x,10+\x)
}
\class[Z4class,"1" {inside,font=\tiny}](9,3)
\class[Z4class,"2" {inside,font=\tiny}](7,5)
\class[Z4class,"1" {inside,font=\tiny}](6,6)
\class[Z4class,"2" {inside,font=\tiny}](4,12)
\foreach \x in {0,...,1} {
\class[Z4class, "3" {inside,font=\tiny}](6-\x,8+\x)
}
\d3(6,12)(5,15)
\replacetarget
\d6(7,11)(6,18)
\d7(9,3)(8,10)
\replacetarget
\d12(8,10)(7,21)
\d7(10,2)(9,9)
\replacetarget
\d15(9,9)(8,24)
\d3(6,6)(5,9)
\replacesource
\d3(5,9)(4,12)
\foreach \x in {1,...,2} {
\d3(6+\x,6-\x)(5+\x,9-\x)
\replacetarget
}
\d7(6,8)(5,15)
\draw[fill=white](1.5,23.5) rectangle (4.3,24.6);
\node at (2.9,24.0) {\Si{12}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI14K,
y range={0}{31},
x range={0}{14},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Zclass, "\phantom{*}" {inside,font=\tiny}](14,0)
\class[mgclass, "\phantom{*}" {inside,font=\tiny}](12,2)
\class[FLDRclass](10,4)
\class[FLDRclass](11,11)
\class[FLDRclass](9,9)
\foreach \x in {1,...,5} {
\class[Z4class, "\x" {inside,font=\tiny}](11-\x,33-3*\x)
}
\foreach \x in {0,...,3} {
\class[Z4class, "3" {inside,font=\tiny}](10-\x,12+\x)
}
\class[Z4class,"1" {inside,font=\tiny}](11,3)
\class[Z4class,"2" {inside,font=\tiny}](9,5)
\class[Z4class,"1" {inside,font=\tiny}](8,6)
\class[Z4class,"1" {inside,font=\tiny}](4,12)
\class[Z4class,"3" {inside,font=\tiny}](5,15)
\class[Z4class,"1" {inside,font=\tiny}](5,9)
\foreach \x in {0,...,2} {
\class[Z4class, "3" {inside,font=\tiny}](8-\x,10+\x)
}
\d9(12,2)(11,11)
\replacetarget
\d19(11,11)(10,30)
\d9(11,3)(10,12)
\replacetarget
\d16(10,12)(9,27)
\foreach \x in {0,...,1} {
\d3\x(5+\x,9+3*\x)(4+\x,12+3*\x)
\replacetarget
}
\d3(7,15)(6,18)
\replacetarget
\d11(9,13)(8,24)
\d7(8,14)(7,21)
\replacetarget
\foreach \x in {2,...,2} {
\d5\x(8+\x,6-\x)(7+\x,11-\x)
\replacetarget
}
\d5(9,5)(8,10)
\replacetarget
\d5(8,6)(7,11)
\replacetarget
\d7(7,11)(6,18)
\d11(8,10)(7,21)
\draw[fill=white](1.5,29.5) rectangle (4.3,30.6);
\node at (2.9,30.0) {\Si{14}H_{K}\underline{\Z}};
\end{sseqdata}
\begin{sseqdata}[name=SI7KF,
y range={0}{13},
x range={0}{10},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Fclass, "\phantom{*}" {inside,font=\tiny}](7,0)
\class[mgclass, "\phantom{*}" {inside,font=\tiny}](6,1)
\class[FLDRclass](5,2)
\class[FLDRclass](4,4)
\class[FLDRclass](5,5)
\class[Z4class,"1" {inside,font=\tiny}](4,3)
\class[Z4class,"1" {inside,font=\tiny}](3,5)
\class[Z4class,"1" {inside,font=\tiny}](4,12)
\d2(5,2)(4,4)
\d2(4,3)(3,5)
\class[Z4class,"3" {inside,font=\tiny}](4,6)
\class[Z4class,"3" {inside,font=\tiny}](3,9)
\d3(4,6)(3,9)
\d4(6,1)(5,5)
\replacetarget
\d7(5,5)(4,12)
\draw[fill=white](7.5,9.5) rectangle (10.3,10.6);
\node at (8.9,10) {\Si{7}H_{K}\underline{\F}};
\end{sseqdata}
\begin{sseqdata}[name=SI10KF,
y range={0}{21},
x range={0}{12},
y tick step = 2,
x tick step = 2,
struct lines = blue,
classes= {fill, inner sep = 0pt, minimum size = 0.2em},
class labels={below=2pt},
differentials=black,
class pattern=linear,
class placement transform = { rotate = 135, scale = 2 },
xscale=0.6,
yscale = 0.6,
run off differentials = ->,
grid = go,
grid color = gray!3
]
\class[Fclass, "\phantom{*}" {inside,font=\tiny}](10,0)
\class[mgclass, "\phantom{*}" {inside,font=\tiny}](9,1)
\class[FLDRclass](8,2)
\foreach \x in {0,...,2} {
\class[FLDRclass](6+\x,6+\x)
}
\class[FLDRclass](7,3)
\d7(9,1)(8,8)
\replacetarget
\d5(8,2)(7,7)
\replacesource
\d3(7,3)(6,6)
\replacesource
\class[Z4class,"4" {inside,font=\tiny}](4,12)
\class[Z4class,"5" {inside,font=\tiny}](5,15)
\class[Z4class,"3" {inside,font=\tiny}](6,18)
\class[Z4class,"1" {inside,font=\tiny}](7,21)
\foreach \x in {0,...,2} {
\class[Z4class, "3" {inside,font=\tiny}](5+\x,7+\x)
}
\class[Z4class,"3" {inside,font=\tiny}](5,9)
\class[Z4class,"3" {inside,font=\tiny}](6,10)
\class[Z4class,"2" {inside,font=\tiny}](6,4)
\class[Z4class,"1" {inside,font=\tiny}](8,2)
\d7(8,2)(7,9)
\replacetarget
\d3(6,4)(5,7)
\d3(5,9)(4,12)
\d5(7,3)(6,8)
\d9(7,9)(6,18)
\d5(6,10)(5,15)
\d13(8,8)(7,21)
\draw[fill=white](9.5,20.5) rectangle (12.3,21.6);
\node at (10.9,21.0) {\Si{10}H_{K}\underline{\F}};
\end{sseqdata}
The slice spectral sequence for \(\Si{-n} H\underline{\Z}\) and \(\Si{n} H\underline{\Z}\) must recover the homotopy Mackey functors of each spectrum, that is, we must be left with \(\ul\pi_{-n}(\Si{-n}H\underline{\Z}) = \ul\pi_{n}(\Si{n}H\underline{\Z}) = \underline{\Z} \) and all other homotopy Mackey functors trivial. For most of the differentials, then, there is only one choice.
We use the indexing convention from \cite[Section 4.4.2]{HHR}. The Mackey functor \(\ul E_2^{t-n,t}\) is \(\ul\pi_n P^t_t(X)\). We also use the Adams convention, so that \(\ul\pi_n P^t_t(X)\) has coordinates \((n,n-t)\) and the differential,
\begin{align*}
d_r:\ul E_r^{s,t} \rightarrow \ul E_r^{s+r,t+r-1},
\end{align*}
points left one and up \(r\).
The symbols in \cref{SSS:mack} denote the Mackey functors in the slice spectral sequences shown.
\begin{table}[h]
\caption[Symbols for {$K$}-Mackey functors]{Symbols for $K$-Mackey functors}
\label{SSS:mack}
\begin{center}
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{|c|c|c|}
\hline
$\blacksquare=\underline{\Z}$
& $\bpent=\phi^*_{LDR} \underline{\F}$
& $\btrap=\ul{mg}$
\\\hline
$\square=\underline{\Z}^*$
& $\pent=\phi_{LDR}^*\underline{\F}^*$
& $\trap = \ul {mg}^*$
\\ \hline
$\underline{\F} = \bdiamond$ & $\gcirc=\underline{g}^n$ & $\triangle = \ul m^*$\\\hline
\end{tabular} }
\end{center}
\end{table}
\begin{exmp} \label{SSS:-1HZ}
The slices for \(\Si{-1} H\underline{\Z}\) are all a one-fold desuspension of
Eilenberg-MacLane spectra (\cref{tower:-1:K:Z}, \cref{fig:SSS:-1..-5:HZ}). Because each of these Mackey functors is in the same column, there are no differentials. Consequently, in the spectral sequence, we find a double extension:
\begin{equation} \label{double:extension}
\begin{tikzcd}[row sep=small]
\underline{\Z}^* \ar[dr] & & \\
& \underline{\Z}(2,1) \ar[dr] \ar[dl] & \\
\ul m^* & & \underline{\Z} \ar[dddll] \\
& & \\ \\
\underline{g} & &
\end{tikzcd}
\end{equation}
\end{exmp}
\begin{exmp} \label{SSS:-5HZ}
In the spectral sequence for \(\Si{-5} H\underline{\Z}\), \cref{fig:SSS:-1..-5:HZ}, because we can only be left with \(\ul\pi_{-5}(P^{-5}_{-5} \Si{-5} H\underline{\Z}) \cong \underline{\Z}\) and all differentials must go left one and up at least two, all differentials are forced. Once we have evaluated each differential, we once again find ourselves with the double extension in \cref{double:extension}.
\end{exmp}
\begin{exmp} \label{SSS:-9HZ}
In \cref{fig:SSS:-7..-9:HZ}, most of the differentials for \(\Si{-9} H\underline{\Z}\) are again forced by the fact that only \(\ul\pi_{-9}(P^{-9}_{-9} \Si{-9} H\underline{\Z}) \cong \underline{\Z}\) can survive the spectral sequence. For example, we have two choices for a differential from \(\pi_{-8}P^{-32}_{-32} \Si{-9} H\underline{\Z} \cong \underline{g}^2\). We find it must be
\begin{align*}
d_{15}: \underline{g}^2 \rightarrow \phi^*_{LDR}\underline{\F}^* \cong \ul\pi_{-9} (P^{-24}_{-24} \Si{-9} H\underline{\Z})
\end{align*}
so that we are left with the extension in \cref{double:extension}. Indeed, we will always be left with this extension once all differentials have been evaluated. Similarly, for \(n\equiv 1\pmod 4\), we will always have a
\begin{align*}
d_2:\phi^*_{LDR} \underline{\F}^* \xrightarrow{\cong}{} \phi^*_{LDR} \underline{\F}^*
\end{align*}
in the upper right corner.
\end{exmp}
In \cref{-4k-2slices:K:Z} we claim that for \(n\) odd,
\begin{align*}
P^{-2n}_{-2n} (\Si{-n} H\underline{\Z}) \simeq \Si{-(k+1)\rho +1} H\phi^*_{LDR} \underline{\F}.
\end{align*}
We now prove this claim.
\begin{prop} \label{phiF:SSS}
Let \(n\) be odd and take \(\ul A\) to be one of the Mackey functors listed in \cref{tab-SShtpydegrees}. The only choice of \(\ul A\) where the homotopy of \(P^{-2n}_{-2n} (\Si{-n} H\underline{\Z}) \simeq \Si{-(k+1)\rho+1} H\ul A\) fits into the spectral sequence for \(\Si{-n} H\underline{\Z}\) is \(\phi^*_{LDR} \underline{\F}\).
\end{prop}
\begin{proof}
Because \(\Si{-(k+1)\rho+1} H\ul A\) is a \((-4k-2)\)-slice, its \(\ul\pi_{-k-1}\) is located at \((-k-1,-3k-1)\). We argue that we cannot have a nonzero Mackey functor in this location.
\begin{table}[h]
\caption[Homotopy Comparison]{Homotopy Comparison}
\label{tab-SShtpydegrees}
\begin{center}
{\renewcommand{\arraystretch}{1.5}\setlength{\tabcolsep}{12pt}
\begin{tabular}{|c|c|} \hline
$\ul A$ & $\ul\pi _{-k-1}(\Si{-(k+1)\rho +1} H\ul A)$ \\ \hline
$\phi^*_{LDR} \ul f$ & $\underline{g}^3$ \\
$\ul m$ & $\underline{g}^2$ \\
$\ul{mg}$ & $\underline{g}$ \\
$\phi^*_{LDR} \underline{\F}$ & $\ul 0$ \rule[-2.4ex]{0pt}{2.2ex} \\ \hline
\end{tabular} }
\end{center}
\end{table}
In the spectral sequence for \(\Si{-n} H\underline{\Z}\), all slices below level \(-2n\) are \(\Si{-k} H\underline{g}^{n-k+1}\) where \(4k\in[2n+1,4n]\). These Mackey functors, for \(4k\leq 2n+1\), lie on the line \(y=-3k\). Thus, for the Mackey functors in \cref{tab-SShtpydegrees}, the source of a differential hitting it must be \((-k,-3k)\). This is not possible.
We now argue that the Mackey functor located at \((-k-1,-3k-1)\) cannot be the source of a differential.The first value of \(n\) for which we must determine the \((-2n)\)-slice is \(n=7\). The spectral sequence where \(\ul A=\phi^*_{LDR} \underline{\F}\) is shown in \cref{fig:SSS:-7..-9:HZ}. This spectral sequence leaves us with the appropriate homotopy for \(\Si{-7} H\underline{\Z}\).
Note there is a copy of \(\underline{g}^4\) located at \((-4,-12)\). For any other choice of \(\ul A\), we would have a nontrivial \(\ul\pi_{-4}\) located at \((-4,-10)\). For a differential originating from \((-4,-10)\), there are two possible targets: \(\phi^*_{LDR} \underline{\F}^*\) at \((-5,-5)\) and \(\underline{g}\) at \((-5,-2)\). However, these two Mackey functors must fit into the exact sequences
\begin{align*}
\underline{g}^2 \rightarrow \underline{g}^3\rightarrow \underline{g} \qquad \text{ and } \qquad
\underline{g}^4 \rightarrow \phi^*_{LDR} \underline{\F}^* \rightarrow \ul {mg}.
\end{align*}
Thus, we cannot have a nonzero Mackey functor at \((-4,-10)\).
We now consider the spectral sequence for \(\Si{-9} H\underline{\Z}\), located in \cref{fig:SSS:-7..-9:HZ}. We again use \(\Si{-(l+1)\rho + 1} H\phi^*_{LDR} \underline{\F}\) for the \((-2n)\)-slice. The resulting homotopy fits in the spectral sequence. For the other three choices of \(\ul A\) we would have a nontrivial \(\ul\pi_{-5}\) located at \((-5,-13)\). For a differential originating from \((-5,-13)\) there are two possible targets: \(\underline{g}^3\) at \((-6,-8)\) and \(\phi^*_{LDR}\underline{\F}^*\) at \((-6,-3)\).
However, we have a \(d_2:\phi^*_{LDR}\underline{\F}^* \twoheadrightarrow \phi^*_{LDR}\underline{\F}^*\) and a \(d_7:\underline{g}^5\twoheadrightarrow \underline{g}^3\). Thus, there is no target for a differential from \((-5,-13)\). Consequently, the only choice of \(\ul A\) that works is \(\phi^*_{LDR} \underline{\F}\).
There is a similar story for all odd \(n>9\). There will always be a \(d_2\) hitting the \(\phi^*_{LDR} \underline{\F}^*\) located at \((-k-2,-n+k+2)\). The only other possible targets for a differential from \((-k-1,-3k-1)\) are then the \(\underline{g}^3\)'s resulting from the homotopy of the other \((-4j-2)\)-slices. All of these will be hit by a differential from the \(\underline{g}^{n-k+1}\) located at \((-k-1,-3k-3)\). Thus, we cannot have a nonzero Mackey functor located at \((-k-1,-3k-1)\). The only choice of \(\ul A\) which satisfies this requirement is \(\phi^*_{LDR} \underline{\F}\).
\end{proof}
\begin{exmp} \label{SSS:6HZ}
For the positive, trivial suspensions of \(H\underline{\Z}\), we find that \(\Si{6} H\underline{\Z}\) has the first nontrivial slice tower. In \cref{fig:SSS:6..8:HZ}, we then see that there is only one possible differential. This \(d_3\) exists because we must be left with only \(\ul\pi_6 (\Si{6} H\underline{\Z}) \cong \underline{\Z}\).
\end{exmp}
\begin{exmp} \label{SSS:7HZ}
The spectral sequence for \(\Si{7} H\underline{\Z}\), in \cref{fig:SSS:6..8:HZ}, is more interesting. Here we find the differentials \(d_2:\phi^*_{LDR}F \rightarrow \ul{mg}\) and \(d_5:\ul{mg}\rightarrow \underline{g}\). Indeed, we will always see a \(d_{n-7}:\phi^*_{LDR}\underline{\F} \rightarrow \ul{mg}\) and \(d_{2n-9}:\ul{mg} \rightarrow \underline{g}\) on the right side of the spectral sequence for \(\Si{n} H_K\underline{\Z}\).
\end{exmp}
\begin{exmp} \label{SSS:7F}
Except for the homotopy of the \(n\)-slice of \(\Si{7} H_K\underline{\Z}, \quad \) \(\Si{7} H_K\underline{\F}\), and \(\Si{8} H_K\underline{\Z}\), the spectral sequences for \(\Si{7} H_K\underline{\Z}\) and \(\Si{8} H_K\underline{\Z}\) collapse to give the spectral sequence for \(\Si{7} H_K\underline{\F}\). We see in \cref{fig:SSS:6..8:HZ} the \(\underline{g}\) in \((3,9)\) and the \(\underline{g}^2\) in \((3,9)\) in the spectral sequences for \(\Si{7} H_K\underline{\Z}\) and \(\Si{8} H_K\underline{\Z}\), respectively, combine to give the \(\underline{g}^3\) in \((3,9)\) in the spectral sequence for \(\Si{7} H_K\underline{\F}\). Off the diagonal for the \(n\)-slice for \(\Si{7} H_K\underline{\Z}\) and \(\Si{8} H_K\underline{\Z}\) we have a single copy of \(\phi^*_{LDR}\underline{\F}\). These provide the two copies of \(\phi^*_{LDR}\underline{\F}\) off the diagonal for \(\Si{7} H_K\underline{\F}\).
\end{exmp}
\begin{exmp} \label{SSS:11HZ}
Now, in \cref{fig:SSS:10..11:HZ}, we have some choice of differentials in the spectral sequence for \(\Si{11} H\underline{\Z}\). Once we consider that only \(\ul\pi_{11}(\Si{11} H\underline{\Z}) \cong \underline{\Z}\) can be left, there is only one choice of each differential that provides the desired result. Analogously to the spectral sequence for \(\Si{-n} H\underline{\Z}\) where \(n\equiv 1\pmod 4\), we will always have a \(d_2:\phi^*_{LDR} \underline{\F} \xrightarrow{\cong}{} \phi^*_{LDR} \underline{\F}^*\) in the bottom left corner when \(n\equiv 3\pmod 4\).
\end{exmp}
\begin{exmp} \label{SSS:10F}
Again, with the exception of the homotopy of the \(n\)-slice, the spectral sequences for \(\Si{10} H_K\underline{\Z}\) and \(\Si{11} H_K\underline{\Z}\) collapse to give the spectral sequence for \(\Si{10} H_K\underline{\F}\) in \cref{fig:SSS:10..11:HZ}. As in \cref{SSS:7F}, the upper left diagonals in in the spectral sequences for \(\Si{10} H_K\underline{\Z}\) and \(\Si{11} H_K\underline{\Z}\) combine to even more copies of \(\underline{g}\) in the upper left diagonal in the spectral sequence for \(\Si{10} H_K\underline{\F}\). This will always be the case.
\end{exmp}
\clearpage
\begin{fig}\label{fig:SSS:-1..-5:HZ}
\begin{flushleft}
The slice spectral sequence over \(K\), \(n=-1,-3,-5\).
\end{flushleft}
\end{fig}
\begin{adjustbox}{minipage=[r][\textheight][b]{\paperwidth},scale={0.8}}
\printpage[name=SI-1K]
\printpage[name=SI-3K]\newline
\printpage[name=SI-5K]
\end{adjustbox}
\clearpage
\begin{fig}\label{fig:SSS:-7..-9:HZ}
\begin{flushleft}
The slice spectral sequence over \(K\), \(n=-7,-9\).
\end{flushleft}
\end{fig}
\begin{adjustbox}{scale=0.8}
\printpage[name=SI-7K]
\printpage[name=SI-9K]
\end{adjustbox}
\clearpage
\begin{fig}\label{fig:SSS:6..8:HZ}
\begin{flushleft}
The slice spectral sequence over \(K\), \(n=6,7,8\).
\end{flushleft}
\end{fig}
\begin{adjustbox}{minipage=[r][\textheight][b]{\paperwidth},scale={0.8}}
\printpage[name=SI6K]
\printpage[name=SI7K] \newline
\printpage[name=SI8K]
\printpage[name=SI7KF]
\end{adjustbox}
\vfill
\clearpage
\begin{fig}\label{fig:SSS:10..11:HZ}
\begin{flushleft}
The slice spectral sequence over \(K\), \(n=10,11\).
\end{flushleft}
\end{fig}
\begin{adjustbox}{scale={0.7}}
\printpage[name=SI10K]
\end{adjustbox} \newline
\begin{adjustbox}{scale=0.7}
\printpage[name=SI10KF]
\printpage[name=SI11K]
\end{adjustbox}
\clearpage
\begin{fig}\label{fig:SSS:12:HZ}
\begin{flushleft}
The slice spectral sequence over \(K\), \(n=12\).
\end{flushleft}
\end{fig}
\begin{adjustbox}{scale=0.9}
\printpage[name=SI12K]
\end{adjustbox}
\clearpage
\begin{fig}\label{fig:SSS:14:HZ}
\begin{flushleft}
The slice spectral sequence over \(K\), \(n=14\).
\end{flushleft}
\end{fig}
\begin{adjustbox}{scale=0.9}
\printpage[name=SI14K]
\end{adjustbox}
\begin{bibdiv}
\begin{biblist}
\bib{D}{article}{
author={Dugger, Daniel},
title={An Atiyah-Hirzebruch spectral sequence for $KR$-theory},
journal={$K$-Theory},
volume={35},
date={2005},
number={3-4},
pages={213--256 (2006)},
issn={0920-3036},
review={\MR{2240234}},
doi={10.1007/s10977-005-1552-9},
}
\bib{GM}{article}{
author={Greenlees, J. P. C.},
author={Meier, Lennart},
title={Gorenstein duality for real spectra},
journal={Algebr. Geom. Topol.},
volume={17},
date={2017},
number={6},
pages={3547--3619},
issn={1472-2747},
review={\MR{3709655}},
doi={10.2140/agt.2017.17.3547},
}
\bib{GY}{article}{
author={Guillou, B.},
author={Yarnall, C.},
title={The Klein four slices of $\Sigma^n H\underline{\F}_2$},
journal={Math. Z.},
volume={295},
date={2020},
number={3-4},
pages={1405--1441},
issn={0025-5874},
review={\MR{4125695}},
doi={10.1007/s00209-019-02433-3},
}
\bib{H}{article}{
author={Hill, Michael A.},
title={The equivariant slice filtration: a primer},
journal={Homology Homotopy Appl.},
volume={14},
date={2012},
number={2},
pages={143--166},
issn={1532-0073},
review={\MR{3007090}},
doi={10.4310/HHA.2012.v14.n2.a9},
}
\bib{HHR}{article}{
author={Hill, M. A.},
author={Hopkins, M. J.},
author={Ravenel, D. C.},
title={On the nonexistence of elements of Kervaire invariant one},
journal={Ann. of Math. (2)},
volume={184},
date={2016},
number={1},
pages={1--262},
issn={0003-486X},
review={\MR{3505179}},
doi={10.4007/annals.2016.184.1.1},
}
\bib{HHR2}{article}{
author={Hill, Michael A.},
author={Hopkins, M. J.},
author={Ravenel, D. C.},
title={The slice spectral sequence for certain $RO(C_{p^n})$-graded
suspensions of $H\underline{\bf Z}$},
journal={Bol. Soc. Mat. Mex. (3)},
volume={23},
date={2017},
number={1},
pages={289--317},
issn={1405-213X},
review={\MR{3633137}},
doi={10.1007/s40590-016-0129-3},
}
\bib{HS}{article}{
author={Heard, Drew},
author={Stojanoska, Vesna},
title={$K$-theory, reality, and duality},
journal={J. K-Theory},
volume={14},
date={2014},
number={3},
pages={526--555},
issn={1865-2433},
review={\MR{3349325}},
doi={10.1017/is014007001jkt275},
}
\bib{HY}{article}{
author={Hill, Michael A.},
author={Yarnall, Carolyn},
title={A new formulation of the equivariant slice filtration with
applications to $C_p$-slices},
journal={Proc. Amer. Math. Soc.},
volume={146},
date={2018},
number={8},
pages={3605--3614},
issn={0002-9939},
review={\MR{3803684}},
doi={10.1090/proc/13906},
}
\bib{U}{article}{
author={Ullman, John},
title={On the slice spectral sequence},
journal={Algebr. Geom. Topol.},
volume={13},
date={2013},
number={3},
pages={1743--1755},
issn={1472-2747},
review={\MR{3071141}},
doi={10.2140/agt.2013.13.1743},
}
\bib{U2}{book}{
author={Ullman, John Richard},
title={On the Regular Slice Spectral Sequence},
note={Thesis (Ph.D.)--Massachusetts Institute of Technology},
publisher={ProQuest LLC, Ann Arbor, MI},
date={2013},
pages={(no paging)},
review={\MR{3211466}},
}
\bib{V}{article}{
author={Voevodsky, Vladimir},
title={Open problems in the motivic stable homotopy theory. I},
conference={
title={Motives, polylogarithms and Hodge theory, Part I},
address={Irvine, CA},
date={1998},
},
book={
series={Int. Press Lect. Ser.},
volume={3},
publisher={Int. Press, Somerville, MA},
},
date={2002},
pages={3--34},
review={\MR{1977582}},
}
\bib{Y}{article}{
author={Yarnall, Carolyn},
title={The slices of $S^n\wedge H\underline{\mathbb{Z}}$ for cyclic
$p$-groups},
journal={Homology Homotopy Appl.},
volume={19},
date={2017},
number={1},
pages={1--22},
issn={1532-0073},
review={\MR{3628673}},
doi={10.4310/HHA.2017.v19.n1.a1},
}
\bib{Z}{article}{
AUTHOR = {Zeng, Mingcong},
TITLE = {Eilenberg-Mac Lane spectra in cyclic \(p\)-groups},
eprint = {https://arxiv.org/abs/1710.01769},
year = {2018},
}
\bib{Z2}{thesis}{
author={Zou, Yan},
title={\(RO(D_{2p})\)-graded Slice Spectral Sequence of \(H\underline{\Z}\)},
note={Thesis (Ph.D.)--University of Rochester},
date={2018},
}
\end{biblist}
\end{bibdiv}
\end{document} |
\section{Dataset Statistics}
Here we list the statistics of 7 public dataset in Table \ref{data}.
\begin{table}[h]
\centering
\begin{tabular}{ccccc}
\toprule
& Nodes & Edges & Classes & Features \\ \midrule
Cora & 2,708 & 5,429 & 7 & 1,433 \\
Citeseer & 3,327 & 4,732 & 6 & 3,703 \\
Polblogs & 1,490 & 19,025 & 2 & - \\
USA & 1,190 & 13,599 & 4 & - \\
Brazil & 131 & 1,038 & 4 & - \\
AIDS & 31,385 & 64,780 & 38 & 4 \\
ENZYMES & 19,580 & 74,564 & 3 & 18 \\
\bottomrule
\end{tabular}
\caption{Dataset statistics}
\label{data}
\end{table}
\section{Details of Random Sampling}
\begin{algorithm}[bh]
\caption{Random sampling from probabilistic vector to binary adjacency matrix}
\label{random sampling}
\textbf{Input}: Probabilistic vector $\mathbf{a}$, number of trials $K$\\
\textbf{Parameter}: Edge density $\rho$\\
\textbf{Output}: Binary matrix $A$;
\begin{algorithmic}[1]
\STATE Normalize probabilistic vector: $\hat{\mathbf{a}}$ = $\mathbf{a}/\|\mathbf{a}\|_1$
\FOR{k = 1,2 $\cdots$ K}
Draw binary vector $\mathbf{a}^{(k)}$ by sampling $\lfloor \rho n \rfloor$ edges according to probabilistic vector $\hat{\mathbf{a}}$.
\ENDFOR
\STATE Choose a vector $\mathbf{a}^*$ from $\{\mathbf{a}^{(k)}\}$ which yields the smallest loss $\mathcal{L}_{attack}$.
\STATE Convert $\mathbf{a}^*$ to binary adjacency matrix $A$
\STATE \textbf{return} $A$
\end{algorithmic}
\end{algorithm}
\section{Proof of Edge Influence}
\begin{theorem}
The adversary advantage is greater for edges with greater influence.
\end{theorem}
The proof is based on the lemma \ref{lemma} from \cite{wu2016methodology}.
\begin{Lemma}
\label{lemma}
Suppose the target model $f$ is trained on data distribution $p(\mathcal{X},\mathcal{Y})$. $\mathcal{X}= (x_s, x_{ns})$, where $x_s$ and $x_{ns}$ denote the sensitive and non-sensitive part of feature respectively. The optimal adversary advantage is
\[P_{\mathcal{X} \sim p(\mathcal{X},\mathcal{Y})}[f(x_s=1, x_{ns}) \neq f(x_s=0, x_{ns})] .\]
\end{Lemma}
\noindent
\begin{proof}
In our graph setting, $x_{ns}$ refers to feature matrix $X$ and $x_s$ refers to edges.
Set $P[f^i_{\theta^*}(A, X)= y_i]$ = $P_{f_{\theta^*}}(y_i~ |~ A, X) = p$; $P[f^i_{\theta^*}(A_{-e}, X)= y_i] = P_{f_{\theta^*}}(y_i ~|~ A_{-e}, X) = q$.
Without loss of generality, the prediction accuracy is higher with edge $e$.
So, we set $q\le p$.
The adversary advantage is $Adv =[p(1-q)+q(1-p)]$.
Through variable substitution ($x=p-q, y=p+q$), we have $Adv= (y+\frac{x^2-y^2}{2}) $ and $\frac{\partial Adv}{\partial \mathcal{I}(e)} = \frac{\partial Adv}{\partial x}= p-q \ge 0$.
\end{proof}
\bibliographystyle{named}
\section{Summary Review}
The paper studies the problem of model inversion attack of GNNs, i.e., inferring the edges from GNNs. The reviewers generally agree the studied problem is interesting/novel and the use of projected gradient descent is feasible. The author rebuttal also clarified several concerns raised by reviewers, e.g., the generated graph is not an adversarial example and does not require to be unnoticeable. However, the novelty of the proposed attack is kind of limited from the methodological perspective. The authors are also encouraged to improve the paper presentation, provide comparisons with more state-of-the-art baselines, add convergence analysis, and provide justifications of the chosen datasets and assumptions of underlying attacks (e.g., the attacker's knowledge and capability).
\section{Changes of Resubmission Paper}
The authors are grataful for the valuable suggestions offered by reviewers. We have conducted major revisions to address the reviewers' concerns and improve this paper. Firstly, the authors provide comparisons with state-of-the-art baselines on more datasets. Secondly, convergence analysis, parameter settings, dataset separation and assumptions of the attacks are demonstrated clearly in the new paper. Thirdly, we have modified our model inversion method which takes the intrinsic properties of graph into consideration. Last but not least, we improved the presentation of the whole paper to make it easier for reviewers and readers to understand and follow.
\section{Reviews}
\subsection{Reviewer 1}
1. {Summary} Please summarize the main claims/contributions of the paper in your own words.\\
This paper proposes an interesting problem: Graph Model Inversion, that infers the edges from the known GNN models.\\
2. {Novelty} How novel is the paper?\\
Paper make non-trivial advances over past work\\
3. {Soundness} Is the paper technically sound?\\
I have not checked all details, but the paper appears to be technically sound\\
4. {Impact} How important is the paper likely to be, considering both methodological contributions and impact on application areas?\\
The paper will impact a moderate number of researchers\\
5. {Clarity} Is the paper well-organized and clearly written?\\
Good: paper is well organized but language can be improved\\
6. {Evaluation} Are claims well supported by experimental results?\\
Moderate: Experimental results are weak: important baselines are missing, or improvements are not significant\\
7. {Resources} How impactful will this work be via sharing datasets, code and/or other resources?\\
Not applicable: no shared resources\\
8. (Reproducibility) Would the experiments in the paper be easy to reproduce? (It may help to consult the paper’s reproducibility checklist.)\\
Good: e.g., code/data available, but some details of experimental settings are missing/unclear\\
\\
9. {Reasons to Accept} Please describe the paper’s key strengths.\\
1. Novel problem and intersting setting.\\
2. Using PGD to solve this problem\\
\\
10. {Reasons to Reject} Please describe the paper’s key weaknesses.\\
1. The motiveation to adopt PGD should be elaborated.\\
2. Comparsion to the other inversion attack is needed.\\
3. The explanation of the algorithm is not clear.\\
4. Experimental setting is node clear.\\
\\
11. {Detailed Comments} Please provide other detailed comments and constructive feedback.\\
In general, I found this work interesting, since the research on the problem of model inversion attack targeting graph neural networks is quite few and novel. However, the current manuscript can be further improved; see my questions/comments as below.\\
\\
1. I agree with the authors that projected gradient decent is used given the discrete nature of graphs as in [1]. However, could the authors provide a theoretical or empirical analysis on how the error would be if pure gradient decent is used under graph setting other than projected gradient decent.\\
\\
2. Further, though the authors mention that none of the existing model inversion attack methods can be directly used under the graph setting, it would be interesting to see how these existing model inversion attack would perform with the same setting. This could be implemented by such as naively adding threshold to make the recovered features in these methods discrete.\\
\\
3. In step 4 of Algorithm 1, the statement “if needed” is quite vague and should be explained more in details.\\
\\
4. Details upon how the baselines (embedding similarity and feature similarity) were constructed and tuned in your experimental settings are missed, such as how to select the threshold, since the embedding similarity also achieves comparable performance.\\
\\
------ Post Rebuttal -------\\
I've read the rebuttal from authors and most of my points has been addressed. Moreover, I think the novelty of the problem setting should be considered. Despite the negative opiontion from some reviewers, I would be positive to accept this paper.
\\
[1] Xu, K., Chen, H., Liu, S., Chen, P. Y., Weng, T. W., Hong, M., & Lin, X. (2019, August). Topology attack and defense for graph neural networks: an optimization perspective. In Proceedings of the 28th International Joint Conference on Artificial Intelligence (pp. 3961-3967). AAAI Press.\\
\\
12. {QUESTIONS FOR THE AUTHORS} Please provide questions for authors to address during the author feedback period. (Please number them)\\
See the detailed comments.\\
13. {Ethical Considerations} Please highlight any ethical considerations that must be considered before a final decision (it may help to consult the paper’s ethics statement, if provided)\\
None\\
\\
14. (OVERALL SCORE)\\
6 - Above threshold of acceptance\\
\\
19. I acknowledge that I have read the author's rebuttal and made whatever changes to my review where necessary.\\
Agreement accepted\\
\subsection{Reviewer 2}
1. {Summary} Please summarize the main claims/contributions of the paper in your own words. (Do not provide any review in this box)\\
The paper (1). designs a gradient based, white-box strategy, Graph Model Inversion (GMI), to generate the graph adversarial samples which are used to attack the SOTA graph models (e.g. GCN). (2). It also introduces the differential privacy method into graph model to help the graph model to defend the attack from the adversarial samples generated in (1).\\
2. {Novelty} How novel is the paper?\\
Paper contributes some new ideas\\
3. {Soundness} Is the paper technically sound?\\
I have not checked all details, but the paper appears to be technically sound\\
4. {Impact} How important is the paper likely to be, considering both\\ methodological contributions and impact on application areas?
The paper will impact a moderate number of researchers\\
5. {Clarity} Is the paper well-organized and clearly written?\\
Good: paper is well organized but language can be improved\\
6. {Evaluation} Are claims well supported by experimental results?\\
Moderate: Experimental results are weak: important baselines are missing, or improvements are not significant\\
7. {Resources} How impactful will this work be via sharing datasets, code and/or other resources? (It may help to consult the paper’s reproducibility checklist.)\\
Fair: some may find shared resources useful in future work\\
8. (Reproducibility) Would the experiments in the paper be easy to reproduce? (It may help to consult the paper’s reproducibility checklist.)
Meets Minimum Standard: e.g., code/data unavailable, but paper is clear enough that an expert could confidently reproduce\\
9. {Reasons to Accept} Please describe the paper’s key strengths.\\
(1). Besides proposing the attacking strategy, the paper also proposed a defense strategy to protect the graph model from the attack by the adversarial samples.\\
(2.) The paper analyzed the correlation between edge importance and inversion risk in a way.\\
10. {Reasons to Reject} Please describe the paper’s key weaknesses.\\
1. The proposed gradient based method for attacking GNN is not very novel.\\
2. The discussion of Experiment part is not very clear.\\
3. Many previous related works are not considered in the paper.\\
11. {Detailed Comments} Please provide other detailed comments and constructive feedback.\\
(1). One of an important thing in adversarial attack on graph model is that the perturbation on generated adversarial graph should be unnoticeable. Therefore, the paper should show some evidences (e.g. node degree distribution) to show the similarity between the original graphs and their adversarial samples.\\
\\
(2). In the Method part, I doubt that max L(a) may cause the asymmetry of adjacency matrix A.\\
\\
(3). As for the related work, the author ignores a lot of previous works. Many adversarial attack works on GNN have been proposed including meta-gradient on adjacency (discrete relaxation), black-box attack via reinforcement learning:\\
For example:\\
Adversarial attacks on neural networks for graph data, Daniel Zugner\\
12. {QUESTIONS FOR THE AUTHORS} Please provide questions for authors to address during the author feedback period. (Please number them)\\
(1). A few discussions in Experiment part are confused. For example: (1), The RQ1 want to show the effect of the designed attack strategy. However, the author used results in table 1 to support RQ1 which seems not valid. I think it may be solid to use the Figure 5 (a-b) to support the RQ 1. (2), when reading the Fig. 5 (a-b), why there are two “Projection lower bound”? One of them may be “upper bound”?\\
\\
(2). One of an important thing in adversarial attack on graph model is that the perturbation on generated adversarial graph should be unnoticeable. Therefore, the paper should show some evidences (e.g. node degree distribution) to show the similarity between the original graphs and their adversarial samples.\\
\\
14. (OVERALL SCORE)\\
5 - Below threshold of acceptance\\
\\
19. I acknowledge that I have read the author's rebuttal and made whatever changes to my review where necessary.\\
Agreement accepted
\subsection{Reviewer 3}
1. {Summary} Please summarize the main claims/contributions of the paper in your own words. (Do not provide any review in this box)\\
A model inversion attack on GNN based on the GNN model, and features and labels from a subset of vertices is proposed. The attack is based on solving a constrained optimization problem using PGD. A defense based on adding edges with high feature similarity is also proposed.\\
2. {Novelty} How novel is the paper?\\
Main ideas of the paper are known or incremental advances over past work\\
3. {Soundness} Is the paper technically sound?\\
I have not checked all details, but the paper appears to be technically sound\\
4. {Impact} How important is the paper likely to be, considering both methodological contributions and impact on application areas?\\
The paper will have low overall impact\\
5. {Clarity} Is the paper well-organized and clearly written?\\
Good: paper is well organized but language can be improved\\
6. {Evaluation} Are claims well supported by experimental results?\\
Poor: The experimental design is flawed and unconvincing\\
7. {Resources} How impactful will this work be via sharing datasets, code and/or other resources? (It may help to consult the paper’s reproducibility checklist.)\\
Fair: some may find shared resources useful in future work\\
8. (Reproducibility) Would the experiments in the paper be easy to reproduce? (It may help to consult the paper’s reproducibility checklist.)\\
Meets Minimum Standard: e.g., code/data unavailable, but paper is clear enough that an expert could confidently reproduce\\
9. {Reasons to Accept} Please describe the paper’s key strengths.\\
The paper addresses a timely and important topic on privacy of the connectivity graph in GNN models.\\
10. {Reasons to Reject} Please describe the paper’s key weaknesses.\\
The experiment results are unconvincing due to the datasets used. Unclear what is the performance in bigger and more diverse datasets.\\
11. {Detailed Comments} Please provide other detailed comments and constructive feedback.\\
1. Please state the dimensions of each parameter like $W^k$ carefully. $M_k$ is used in (1) and also as the dimension of the feature of node v in layer k. The equation (3) does not seem to be properly written.\\
\\
2. Please proofread the paper. There are various grammatical mistakes. E.g., "one optimization problem" should be "an optimization problem", "will severely undermines", "the rest missing links", etc. There are also formatting mistakes like writing "if" in math font instead of text font.\\
\\
12. {QUESTIONS FOR THE AUTHORS} Please provide questions for authors to address during the author feedback period. (Please number them)\\
1. How large is $V_GMI$? There should be investigations to show the impact of the subset size and choice on the attack and defense performance.\\
\\
2. How important are labels in the attack performance? Why are labels not considered in the defense mechanism? [My point was why the authors did not consider using labels in the defense mechanism and why they designed their defense as such? They have not answered this question satisfactorily.]\\
\\
3. It is well known that the inference performance of the datasets Cora and Citeseer are highly sensitive to the underlying graph topology (after all, these are citation networks based on word documents). Hence there is no surprise in Fig. 4. To test the efficacy of the proposed attacks, the authors should consider other datasets where node features play a more important role. [No satisfactory answer from authors.]\\
\\
4. Unclear if the paper is performing transductive learning of the GNN model. Are the GCN, GAT and GraphSAGE pre-trained models or the authors retrain them? How are the testing nodes chosen?\\
13. {Ethical Considerations} Please highlight any ethical considerations that must be considered before a final decision (it may help to consult the paper’s ethics statement, if provided)
\\
Nil\\
14. (OVERALL SCORE)\\
3 - Clear reject\\
19. I acknowledge that I have read the author's rebuttal and made whatever changes to my review where necessary.\\
Agreement accepted
\subsection{Reviewer 4}
Questions
1. {Summary} Please summarize the main claims/contributions of the paper in your own words. (Do not provide any review in this box)\\
This paper studies an important problem, i.e., model inversion attack of Graph Neural Networks (GNN) models. Due to the discreteness of graph data, the authors resort to projected gradient descent with convex relaxation and design a denoising module to improve the performance of model inversion attack on GNNs. The authors also provide studies on the connection between model inversion risk and edge importance and find edges with higher importance are more likely to be recovered, and propose a simple defense method based on graph preprocessing. Furthermore, the authors also conduct experiments to evaluate the effectiveness of the proposed methods.\\
2. {Novelty} How novel is the paper?\\
Paper contributes some new ideas\\
3. {Soundness} Is the paper technically sound?\\
The paper has minor technical flaws that are easily fixable\\
4. {Impact} How important is the paper likely to be, considering both methodological contributions and impact on application areas?\\
The paper will impact a moderate number of researchers\\
5. {Clarity} Is the paper well-organized and clearly written?\\
Good: paper is well organized but language can be improved\\
6. {Evaluation} Are claims well supported by experimental results?\\
Good: Experimental results are sufficient, though more analysis would significantly add support to the claims\\
7. {Resources} How impactful will this work be via sharing datasets, code and/or other resources? (It may help to consult the paper’s reproducibility checklist.)\\
Fair: some may find shared resources useful in future work\\
8. (Reproducibility) Would the experiments in the paper be easy to reproduce? (It may help to consult the paper’s reproducibility checklist.)\\
Good: e.g., code/data available, but some details of experimental settings are missing/unclear\\
9. {Reasons to Accept} Please describe the paper’s key strengths.\\
$[+]$ The authors propose graph model inversion attack which applies projected gradient descent on graphs via convex relaxation for edge reconstruction.\\
\\
$[+]$ The authors evaluate the relation between edge importance and model inversion risk and find edges with higher importance are more likely to be reconstructed.\\
\\
$[+]$ The authors propose one defense method.\\
10. {Reasons to Reject} Please describe the paper’s key weaknesses.\\
$[-]$ It would be better if the authors empirically and theoretically analyze the convergence of the proposed method.\\
\\
$[-]$ The authors fail to cite state-of the-art works. Currently, different defense mechanisms against model inversion attacks are proposed, e.g., [1,2,3]. It would be better if the authors give more discussion and conduct comparative experiments.\\
\\
$[-]$ The assumptions about attacker’s knowledge and capability that the authors make are very strong.\\
\\
$[-]$ More details are needed. The authors relax $a$ into convex space. It would be better if the authors analyze the approximation errors before and after the relaxation of the vector $a$. Additionally, it would be better if the authors provide more details on how to select the optimal dropout rates, which are applied to adjacency vector $a$.\\
\\
$[1]$ “Improving Robustness to Model Inversion Attacks via Mutual Information Regularization”, 2020.\\
$[2]$ “Defending Model Inversion and Membership Inference Attacks via Prediction Purification”, 2020.\\
$[3]$ “Model inversion attacks that exploit confidence information and basic countermeasures”, 2015.\\
\\
11. {Detailed Comments} Please provide other detailed comments and constructive feedback.\\
This paper studies an important problem, i.e., model inversion attack of Graph Neural Networks (GNN) models. Due to the discreteness of graph data, the authors resort to projected gradient descent with convex relaxation and design a denoising module to improve the performance of model inversion attack on GNNs. The authors also provide studies on the connection between model inversion risk and edge importance and find edges with higher importance are more likely to be recovered, and propose a simple defense method based on graph preprocessing. Furthermore, the authors also conduct experiments to evaluate the effectiveness of the proposed methods. However, I still have the following concerns.
\\
$[-]$ It would be better if the authors empirically and theoretically analyze the convergence of the proposed method.\\
\\
$[-]$ The authors fail to cite state-of the-art works. Currently, different defense mechanisms against model inversion attacks are proposed, e.g., [1,2,3]. It would be better if the authors give more discussion and conduct comparative experiments.\\
\\
$[-]$ The assumptions about attacker’s knowledge and capability that the authors make are very strong.\\
\\
$[-]$ More details are needed. The authors relax $a$ into convex space. It would be better if the authors analyze the approximation errors before and after the relaxation of vector $a$. Additionally, it would be better if the authors provide more details on how to select the optimal dropout rates, which are applied to adjacency vector $a$.\\
\\
$[1]$ “Improving Robustness to Model Inversion Attacks via Mutual Information Regularization”, 2020.\\
$[2]$ “Defending Model Inversion and Membership Inference Attacks via Prediction Purification”, 2020.\\
$[3]$ “Model inversion attacks that exploit confidence information and basic countermeasures”, 2015.\\
\\
12. {QUESTIONS FOR THE AUTHORS} Please provide questions for authors to address during the author feedback period. (Please number them)
Please see the above detailed comments.\\
14. (OVERALL SCORE)\\
5 - Below threshold of acceptance\\
19. I acknowledge that I have read the author's rebuttal and made whatever changes to my review where necessary.\\
Agreement accepted\\
\end{document}
\section{Defense for GMI attack}
One of the most popular countermeasure against privacy attack is differential privacy (DP). Its definition is based on the idea that if two databases differ only by one record and are used by the same algorithm, the output of that algorithm should be similar. More formally:
\begin{myDef}
(($\epsilon,\delta$)-differential privacy) A randomized mechanism $\mathcal{M}$ with domain $\mathcal{R}$ and output $\mathcal{S}$ satisfies ($\epsilon,\delta$)-differential privacy if for any neighboring datasets $D, D' \in \mathcal{R}$, which differ by at most one record and for any subsets of outputs $S$ it holds that:
\begin{equation}
Pr[\mathcal{M}(D)\in S]\le e^{\epsilon}Pr[\mathcal{M}(D')\in S]+\delta,
\end{equation}
\end{myDef}
\noindent where $\epsilon$ is the privacy budget and $\delta$ is the failure probability.\par
\begin{figure}[!t]
\centering
\includegraphics[width=0.47\textwidth]{figs/illustration2.pdf}
\caption{Defense based on graph preprocessing}
\label{illustration2}
\end{figure}
Differential privacy is one general approach for protecting privacy. However, it is not designed for graph, especially for graph model inversion attack. Here we propose one simple defense method based on graph preprocessing (summarized in algorithm 2). Instead of adding noise in GNN model training, our method adds pseudo edges in the training graph to "hide" real edges from model inversion. Specifically, to preserve utility of trained GNN model, we add pseudo edges with high feature similarity (Figure~\ref{illustration2}). Compared with convolutional neural networks for image data, graph neural network generally has fewer layers and parameters. GNN models essentially aggregate features according to graph structure. Previous work on graph adversarial learning~\cite{wu2019adversarial} suggests that adding edges with low feature similarity will severely undermines the utility of GNN model in node classification. Compared with adding edges randomly in the training graph, adding edges with high feature similarity helps preserve utility.
\begin{algorithm}[h]
\caption{Defense for GMI attack based on graph preprocessing}
\begin{algorithmic}[1]
\Require
Training graph adjacency matrix $A$;
Feature vectors $x_i$;
Edge selection threshold $t$;
\Ensure
Trained GNN model, $f$;
\State Normalize feature vectors $x_i$
\State Calculate similarity matrix $S$
\For{all entries in $A$}
\If {$S_{ij}\textgreater t$ and $A_{ij}=0$}
\ThenApply Set $A_{ij}=1$
\EndIf
\EndFor
\State Train GNN model on $A$
\label{code1}
\end{algorithmic}
\end{algorithm}
\section{Introduction}
This short example shows a contrived example on how to format the authors' information for {\it IJCAI--21 Proceedings}.
\section{Author names}
Each author name must be followed by:
\begin{itemize}
\item A newline {\tt \textbackslash{}\textbackslash{}} command for the last author.
\item An {\tt \textbackslash{}And} command for the second to last author.
\item An {\tt \textbackslash{}and} command for the other authors.
\end{itemize}
\section{Affiliations}
After all authors, start the affiliations section by using the {\tt \textbackslash{}affiliations} command.
Each affiliation must be terminated by a newline {\tt \textbackslash{}\textbackslash{}} command. Make sure that you include the newline on the last affiliation too.
\section{Mapping authors to affiliations}
If some scenarios, the affiliation of each author is clear without any further indication (\emph{e.g.}, all authors share the same affiliation, all authors have a single and different affiliation). In these situations you don't need to do anything special.
In more complex scenarios you will have to clearly indicate the affiliation(s) for each author. This is done by using numeric math superscripts {\tt \$\{\^{}$i,j, \ldots$\}\$}. You must use numbers, not symbols, because those are reserved for footnotes in this section (should you need them). Check the authors definition in this example for reference.
\section{Emails}
This section is optional, and can be omitted entirely if you prefer. If you want to include e-mails, you should either include all authors' e-mails or just the contact author(s)' ones.
Start the e-mails section with the {\tt \textbackslash{}emails} command. After that, write all emails you want to include separated by a comma and a space, following the same order used for the authors (\emph{i.e.}, the first e-mail should correspond to the first author, the second e-mail to the second author and so on).
You may ``contract" consecutive e-mails on the same domain as shown in this example (write the users' part within curly brackets, followed by the domain name). Only e-mails of the exact same domain may be contracted. For instance, you cannot contract <EMAIL>" and <EMAIL>" because the domains are different.
\end{document}
\section{Introduction}
Machine learning (ML) algorithms based on deep neural networks have achieved remarkable success in a range of domains such as computer vision~\cite{zhang2020secret}, natural language processing~\cite{ijcai2020-525}, and graph analysis~\cite{wang2019mcne}.
Meanwhile, the impact of machine learning techniques on privacy is receiving more and more attention because many machine learning applications involve processing sensitive user data (e.g., purchase records)~\cite{ijcai2020-481}.
Attackers may exploit the output (i.e., black-box attack) or the parameters (i.e., white-box attack) of machine learning models to potentially reveal sensitive information in training data.
\par
According to the attacker's goal, privacy attacks can be categorized into several types, such as membership inference attack \cite{shokri2017membership}, model extraction attack \cite{tramer2016stealing}, and model inversion attack \cite{fredrikson2015model}.
Of particular interest to this paper is model inversion attack which aims to extract sensitive features of training data given output labels and partial knowledge of non-sensitive features.
Model inversion attack was firstly introduced by \cite{fredrikson2014privacy}, where an attacker, given a linear regression model for personalized medicine and some demographic information about a patient, could predict the patient’s genetic markers.
Generally, model inversion relies on the correlation between features and output labels and try to maximize a posteriori (MAP) or likelihood estimation (MLE) to recover sensitive features.
Recently, efforts have been made to extend model inversion to attack other machine learning models, in particular Convolutional Neural Networks (CNN) ~\cite{fredrikson2015model,aivodji2019gamin}. Thus far, most model inversion attacks are investigated in the grid-like domain (e.g., images), leaving its effect on the non-grid domain (e.g., graph structured data) an open problem.
Graph Neural Network (GNN) as one of the state-of-the-art graph analysis tools shows excellent results in various applications on graph-structured data \cite{kipf2016semi,velivckovic2017graph}.
However, the fact that many GNN-based applications such as recommendation systems \cite{wu2019session} and social relationship analysis \cite{wang2019mcne} rely on processing sensitive graph data raises great privacy concerns.
Studying model inversion attack on GNNs helps us understand the vulnerability of GNN models and enable us to avoid privacy risks in advance.
\begin{figure}[!t]
\centering
\includegraphics[width=0.48\textwidth]{figs/illustration.png}
\caption{One motivation scenario in social networks}
\label{motivation}
\end{figure}
\textbf{Motivation scenario:}
Figure \ref{motivation} shows one concrete motivation scenario of model inversion attack on GNN models. Users' friendships are sensitive relational data, and users want to keep them private. Sometimes social network data is collected with user permission to train GNN models for better service. For example, these trained GNN models are used to classify friends or recommend advertisements. Then, the trained models are broadcast to customers or local clients. If the attacker can obtain the trained GNN model from malicious clients, with some auxiliary knowledge crawled from the internet, model inversion attack can be performed to reconstruct friendships among users.
In this paper, we draw attention to model inversion attacks extracting private graph data from GNN models. We focus on the following adversarial scenario. Given the trained GNN model and some auxiliary knowledge (node labels and attributes), the adversary aims to reconstruct all the edges among nodes in the training dataset. However, model inversion attack on graphs brings unique challenges. Firstly, existing model inversion attack methods can barely be applied to the graph setting due to the discrete nature of graphs. Different from the continuous image data, gradient computation and optimization on binary edges of the graph are difficult.
Secondly, current model inversion methods fail to exploit the intrinsic properties of graph such as sparsity and feature smoothness. In addition, existing model inversion attack methods cannot fully leverage the information of node attributes and GNN models. For example, node pairs with similar attributes or embeddings are more likely to have edges.
To address the aforementioned challenges, we propose \textbf{Graph} \textbf{M}odel \textbf{I}nversion attack (GraphMI) for edge reconstruction.
GraphMI is designed with two important modules: the projected gradient module and the graph auto-encoder module.
The projected gradient module is able to tackle the edge discreteness via convex relaxation while preserving graph sparsity and feature smoothness.
The graph auto-encoder module is designed to take all the information of node attributes, graph topology and target model parameters into consideration for graph reconstruction.
Based on GraphMI, we investigate the relation between edge influence and model inversion risk and find that edges with greater influence are more likely to be reconstructed.
Furthermore, we show that differential privacy, in its canonical form, is of little avail to defend against GraphMI.
Experimental results on several public datasets show the effectiveness of GraphMI \footnote{https://github.com/zaixizhang/GraphMI}.
\section{Related Work}
Based on the attacker's goal, privacy attacks can be categorized into several types such as membership inference attack \cite{shokri2017membership}, model extraction attack \cite{tramer2016stealing} and model inversion attack \cite{fredrikson2015model}. Membership inference attack tries to determine whether one sample was used to train the machine learning model; Model extraction attack is one black-box privacy attack. It tries to extract information of model parameters and reconstruct one substitute model that behaves similarly to the target model. Model inversion attack, which is the focus of this paper, aims to reconstruct sensitive features corresponding to labels of target machine learning models.\par
Model inversion attack was firstly presented in \cite{fredrikson2014privacy} for linear regression models. \cite{fredrikson2015model} extended model inversion attack to extract faces from shallow neural networks. They cast the model inversion as an optimization problem and solve the problem by gradient descent with modifications to the images. Furthermore, several model inversion attacks in the black-box setting or assisted with generative methods are proposed~\cite{aivodji2019gamin,zhang2020secret} in the image domain.
Thus far, no existing model inversion attack has focused on the graph domain.
\par
\section{Problem Formulation}
\subsection{Preliminaries on GNNs}
One task that GNN models are commonly used for is semi-supervised node classification~\cite{kipf2016semi}. Given a single network topology with node attributes and a known subset of node labels, GNNs are efficient to infer the classes of unlabeled nodes. Before defining GNN, we firstly introduce the following notations of graph. Let $\mathcal{G}=(\mathcal{V,E})$ be an undirected and unweighted graph, where $\mathcal{V}$ is the vertex (i.e. node) set with size $|\mathcal{V}|=N$, and $\mathcal{E}$ is the edge set. Denote $\mathcal{A}\in\{0,1\}^{N\times N}$ as an adjacent matrix containing information of network topology and $X \in \mathds{R}^{ N\times l}$ as a feature matrix with dimension $l$. In a GNN model, each node $i$ is associated with a feature vector $\textbf{x}_i \in \mathds{R}^l$ and a scalar label $y_i$. GNN is used to predict the classes of unlabeled nodes under the adjacency matrix A and the labeled node data $\{(\textbf{x}_i,y_i)\}^{N_{train}}_{i=1}$. GNN uses all nodes' input features but only $N_{train}< N$ labeled nodes in the training phase.\par
Formally, the k-th layer of a GNN model obeys the message passing rule and can be modeled by one message passing phase and one readout update phase:
\begin{equation}
m_v^{k+1}= \mathop{\sum}_{u\in \mathcal{N}(v)}\mathbf{M}_k(h_v^k, h_u^k, e_{uv}),
\end{equation}
\begin{equation}
h_v^{k+1}= \mathbf{U}_k(h_v^k, m_v^{k+1}),
\end{equation}
where $\mathbf{M}_k$ denotes the message passing function and $\mathbf{U}_k$ is the vertex update function. $\mathcal{N}(v)$ is the neighbors of $v$ in graph $\mathcal{G}$. $h_v^k $ is the feature vector of node $v$ at layer $k$ and $ e_{uv}$ denotes the edge feature. $h^0_v = \textbf{x}_v $ is the input feature vector of node $v$. \par
Specifically, Graph Convolutional Network (GCN)~\cite{kipf2016semi}, a well-established method for semi-supervised node classification, obeys the following rule to aggregate neighboring features:
\begin{equation}
H^{k+1}= \sigma \big (\hat D^{-\frac{1}{2}} \hat A \hat D^{-\frac{1}{2}} H^k W^k \big),
\end{equation}
where $\hat A = A + I_N$ is the adjacency matrix of the graph $\mathcal{G}$ with self connections added and $\hat D$ is a diagonal matrix with $\hat D_{ii} = \sum_j \hat A_{ij}$. $\sigma(\cdot)$ is the ReLU function. $H^k$ and $W^k$ are the feature matrix and the trainable weight matrix of the k-th layer respectively. $H^0 = X$ is the input feature matrix. Note that in most of this paper, we focus on two-layer GCN for the node classification. Later, we show that our graph model inversion attack can be also performed on other types of GNNs, including GAT~\cite{velivckovic2017graph} and GraphSAGE~\cite{hamilton2017inductive}.
\subsection{Problem Definition}
\begin{figure}[!t]
\centering
\includegraphics[width=0.48\textwidth]{figs/flow.png}
\caption{Overview of GraphMI}
\label{illustration}
\end{figure}
We refer to the trained model subjected to model inversion attack as the target model. In this paper, we will firstly train a GNN for node classification task from scratch as the target model. We assume a threat model similar to the existing model inversion attacks \cite{fredrikson2015model}.
\paragraph{Attacker's Knowledge and Capability:}
We will focus on the white-box setting. The attacker is assumed to have access to the target model $f$ and can employ some inference technique to discover the adjacency matrix $A$ of the training graph. In most of the paper, we assume the attacker has labels of all the nodes. In addition to the target model $f$ and node labels, the attacker may have other auxiliary knowledge to facilitate model inversion such as node attributes, node IDs or edge density. We will discuss the impact of auxiliary knowledge and the number of node labels on attack performance in the following sections.
\paragraph{Model Inversion of Graph Neural Networks:}
Let $\theta$ be the model parameters of target model $f$. During the training phase, $f$ is trained to minimize the loss function $\mathcal{L}(\theta, X, A, Y)$:
\begin{equation}
\theta^*={\textup{arg }}\mathop{\textup{min}}\limits_{\theta}\mathcal{L}(\theta, X, A, Y),
\end{equation}
where $Y$ is the vector of node labels and $X$ is the feature matrix.
Given the trained model and its parameters, graph model inversion aims to find the adjacency matrix $A^*$ that maximizes the posterior possibility:
\begin{equation}
A^*={\textup{arg }}\mathop{\textup{max}}\limits_{A}P(A | X, Y, \theta^*).
\end{equation}
\begin{algorithm}[tb]
\caption{GraphMI}
\label{code}
\textbf{Input}: Target GNN model $f_{\theta*}$;
Node label vector $Y$;
Node feature matrix $X$;
Learning rate, $\eta_t$;
Iterations $T$;\\
\textbf{Output}: Reconstructed $A$
\begin{algorithmic}[1]
\STATE $\mathbf{a}^{(0)}$ is set to zeros
\STATE Let $t=0$
\WHILE{t \textless T}
\STATE Gradient descent: $\mathbf{a}^{(t)}= \mathbf{a}^{(t-1)}- \eta_t \nabla \mathcal{L}_{attack}(\mathbf{a})$;
\STATE Call Projection operation in (\ref{projection})
\ENDWHILE
\STATE Call Graph auto-encoder module in (\ref{GAE})
\STATE Call Random sampling module.
\STATE \textbf{return} $A$
\end{algorithmic}
\end{algorithm}
\section{Proposed Algorithm}
Next we introduce GraphMI, our proposed model inversion attack on GNN models.
\subsection{Attack Overview}
Figure \ref{illustration} shows the overview of GraphMI. Generally, GraphMI is one optimization-based attack method, which firstly employs projected gradient descent on the graph to find the ``optimal'' network topology for node labels. Then the adjacency matrix and feature matrix will be sent to the graph auto-encoder module of which parameters are transfered from the target model. Finally, we can interpret the optimized graph as the edge probability matrix and sample a binary adjacency matrix. We summarize GraphMI in Algo \ref{code}.
\subsection{Details of Modules}
\paragraph{Projected Gradient Descent Module:} We treat model inversion on GNNs as one optimization problem: given node features or node IDs, we want to minimize the cross-entropy loss between true labels $y_i$ and predicted labels $\hat{y_i}$ from the target GNN model $f_{\theta*}$. The intuition is that the reconstructed adjacency matrix will be similar to the original adjacency matrix if the loss between true labels and predicted labels is minimized. The attack loss on node $i$ is denoted by $\ell_i(A, f_{\theta^*}, \textbf{x}_i, y_i)$ where $A$ is the reconstructed adjacency matrix, $\theta^*$ is the model parameter of the target model $f$ and $\textbf{x}_i$ is the node feature vector of node $i$. The objective function can be formulated as:
\begin{equation}
\begin{split}
\mathop{\textup{min}}\limits_{A \in \{0,1\}^{N\times N}} \mathcal{L}_{GNN}(A) &=\frac{1}{N}\sum_{i=1} ^N \ell_i(A, f_{\theta^*}, \textbf{x}_i, {y_i}) \\s.t.\quad A &= A^\top.
\end{split}
\label{equ1}
\end{equation}
\par
In many real-world graphs, such as social networks, citation networks, and web pages, connected nodes are likely to have similar features \cite{wu2019adversarial}. Based on this observation, we need to ensure the feature smoothness in the optimized graph. The feature smoothness can be captured by the following loss term $\mathcal{L}_s$:
\begin{equation}
\begin{split}
\mathcal{L}_{s} = \frac{1}{2}\sum_{i, j =1} ^N A_{i, j}(\textbf{x}_i-\textbf{x}_j)^2,
\end{split}
\end{equation}
where $A_{i,j}$ indicates the connection between node $v_i$ and $v_j$ in the optimized graph and $(\textbf{x}_i-\textbf{x}_j)^2$ measures the feature difference between $v_i$ and $v_j$. $\mathcal{L}_s$ can also be represented as:
\begin{equation}
\begin{split}
\mathcal{L}_s = tr(X^\top L X),
\end{split}
\end{equation}
where $L=D-A$ is the laplacian matrix of $A$ and $D$ is the diagonal matrix of $A$. In this paper, to make feature smoothness
independent of node degrees, we use the normalized lapacian matrix $\hat{L} = D^{-1/2}LD^{-1/2}$ instead:
\begin{equation}
\begin{split}
\mathcal{L}_s = tr(X^\top \hat{L} X)=\frac{1}{2}\sum_{i, j =1} ^N A_{i, j}(\frac{\textbf{x}_i}{\sqrt{d_i}}-\frac{\textbf{x}_j}{\sqrt{d_j}})^2,
\end{split}
\end{equation}
where $d_i$ and $d_j$ denote the degree of node $v_i$ and $v_j$. To encourage the sparsity of graph structure, F norm of adjacency matrix $A$ is also added to the loss function. The final objective function is:
\begin{equation}
\begin{split}
{\textup{arg }}\mathop{\textup{min}}\limits_{A \in \{0,1\}^{N\times N}}\mathcal{L}_{attack} &=\mathcal{L}_{GNN}+\alpha \mathcal{L}_s+\beta \|A\|_F
\\s.t. \quad A &= A^\top,\label{equ3}
\end{split}
\end{equation}
where $\alpha$ and $\beta$ are hyper-parameters that control the contribution of feature smoothing and graph sparsity.
Solving equation (\ref{equ3}) is a combinatorial optimization problem due to edge discreteness. For ease of gradient computation and update, we firstly replace the symmetric reconstructed adjacency matrix $A$ with its vector form $\mathbf{a}$ that consists of $n:=N(N-1)/2$ unique variables in $A$. Adjacency matrix $A$ and vector $\mathbf{a}$ can be converted to each other easily, which ensures the optimized adjacency matrix is symmetric. Then we relax $\mathbf{a} \in \{0,1\}^n$ into convex space $\mathbf{a} \in [0,1]^n$. We can perform model inversion attack by firstly solving the following optimization problem:
\begin{equation}
{\textup{arg }}\mathop{\textup{min}}\limits_{\mathbf{a} \in [0,1]^n}\mathcal{L}_{attack} =\mathcal{L}_{GNN}+\alpha \mathcal{L}_s+\beta \|\mathbf{a}\|_2.
\label{equ4}
\end{equation}
\par The continuous optimization problem \ref{equ4} is solved by projected gradient descent (PGD):
\begin{equation}
\mathbf{a}^{t+1} = P_{[0,1]} \big[\mathbf{a}^t -\eta_t g_t \big],
\end{equation}
where $t$ is the iteration index of PGD, $\eta_t$ is the learning rate, $g_t$ is the gradients of loss $\mathcal{L}_{attack}$ in \ref{equ3} evaluated at $\textbf{a}^t$, and
\begin{equation}
P_{[0,1]}[x]=\left\{
\begin{array}{rcl}
0 & & x < 0\\
1 & & x > 1\\
x & & otherwise
\end{array} \right.
\label{projection}
\end{equation}
is the projection operator.
\paragraph{Graph Auto-encoder Module:} In GraphMI, we propose to use graph auto-encoder (GAE) \cite{kipf2016variational} to post-process the optimized adjacency matrix $A$. GAE is composed of two components: encoder and decoder. We transfer part of the parameters from the target model $f_{\theta^*}$ to the encoder. Specifically, feature matrix and adjacency vector $\mathbf{a}$ are sent to the $f_{\theta^*}$ and the node embedding matrix $Z$ is generated by taking the penultimate layer of the target model $f_{\theta^*}$, which is denoted as $H_{\theta^*}(\mathbf{a}, X)$. Then the decoder will reconstruct adjacency matrix $A$ by applying logistic sigmoid function to the inner product of $Z$:
\begin{equation}
A={\rm sigmoid}(ZZ^\top), {\rm with}~ Z=H_{\theta^*}(\mathbf{a}, X).\label{GAE}
\end{equation}
\par The node embeddings generated by the graph auto-encoder module encode the information from node attributes, graph topology, and the target GNN model. Intuitively, node pairs with close embeddings are more likely to form edges.
\paragraph{Random Sampling Module:} After the optimization problem is solved, the solution $A$ can be interpreted as a probabilistic matrix, which represents the possibility of each edge. We could use random sampling to recover the binary adjacency matrix; see details in the appendix.\\
\subsection{Analysis on Correlation between Edge Influence and Inversion Risk}
In previous work \cite{wu2016methodology}, researchers found feature influence to be an essential factor in incurring privacy risk. In our context of graph model inversion attack, sensitive features are edges. Here we want to characterize the correlation between edge influence and inversion risk. Given label vector $Y$, adjacency matrix $A$ and feature matrix $X$, the performance of target model $f_{\theta^*}$ for the prediction can be measured by prediction accuracy:
\begin{equation}
ACC(f_{\theta^*}, A, X)=\frac{1}{N}\sum_{i=1}^N\mathds{1}(f_{\theta^*}^i (A, X)=y_i),
\end{equation}
where, $f_{\theta^*}^i (A, X)$ is the predicted label for node i. The influence of edge $e$ can be defined as:
\begin{equation}
{\mathcal I}(e)=ACC(f_{\theta^*}, A, X) - ACC(f_{\theta^*}, A_{-e}, X),
\label{edge}
\end{equation}
where $A_{-e}$ denotes removing the edge $e$ from the adjacency matrix $A$. \cite{wu2016methodology} proposed to use adversary advantage to characterize model inversion risk of features. The model inversion advantage of adversary $\mathcal{A}$ is defined to be $P[\mathcal{A}(X, f_{\theta^*})=e]- 1/2$, where $P[\mathcal{A}(X, f_{\theta^*})= e]$ is the probability that adversary $\mathcal{A}$ correctly infer the existence of edge $e$. Next, we introduce our theorem.
\begin{theorem}
The adversary advantage is greater for edges with greater influence.
\end{theorem}
We defer the proof to the appendix. Intuitively, edges with greater influence are more likely to be recovered by GraphMI because these edges have a greater correlation with the model output. In the following section, we will validate our theorem with experiments.
\section{Experiments}
\begin{table*}[!t]
\centering
\begin{tabular}{lp{0.7cm}<{\centering}p{0.7cm}<{\centering}p{0.7cm}<{\centering}p{0.7cm}<{\centering}p{0.7cm}<{\centering}p{0.7cm}<{\centering}p{0.7cm}<{\centering}p{0.7cm}<{\centering}p{0.7cm}<{\centering}p{0.7cm}<{\centering}p{0.7cm}<{\centering}p{0.7cm}<{\centering}p{0.7cm}<{\centering}p{0.7cm}<{\centering}}
\toprule
\multirow{2}{*}{\textbf{Method}}& \multicolumn{2}{c}{\textbf{Cora}} & \multicolumn{2}{c}{\textbf{Citeseer}} & \multicolumn{2}{c}{\textbf{Polblogs}}& \multicolumn{2}{c}{\textbf{USA}} &\multicolumn{2}{c}{\textbf{Brazil}} &\multicolumn{2}{c}{\textbf{AIDS}} &\multicolumn{2}{c}{\textbf{ENZYMES}} \\
\cmidrule(r){2-3} \cmidrule(r){4-5} \cmidrule(r){6-7} \cmidrule(r){8-9}\cmidrule(r){10-11}\cmidrule(r){12-13}\cmidrule(r){14-15}
& AUC & AP & AUC & AP & AUC & AP & AUC &AP&AUC&AP&AUC &AP&AUC&AP \\ \midrule
Attr. Sim. & 0.803 & 0.808 & \textbf{0.889} & \textbf{0.891}&-&-&-&-&-&- & 0.731 & 0.727 & 0.564 & 0.567 \\
MAP & 0.747 & 0.708 & 0.693 & 0.755 & 0.688 & 0.751 & 0.594 & 0.601 & 0.638 &0.661&0.642&0.653&0.617&0.643 \\
GraphMI & \textbf{0.868} & \textbf{0.883} & 0.878 & 0.885 & \textbf{0.793} & \textbf{0.797} & \textbf{0.806}& \textbf{0.813} & \textbf{0.866} &\textbf{0.888}&\textbf{0.802} &\textbf{0.809}&\textbf{0.678}&\textbf{0.684} \\ \bottomrule
\end{tabular}
\caption{Results of model inversion attack on Graph Neural Networks}
\label{result}
\end{table*}
In this section, we present the experimental results to show the effectiveness of GraphMI. Specifically, our experiments are designed to answer the following research questions:
\begin{itemize}
\item \textbf{RQ1}: How effective is GraphMI?
\item \textbf{RQ2}: Which edges are more likely to be
\item \textbf{RQ3}: Is differential privacy an effective countermeasure against model inversion attacks on GNN?
\end{itemize}
\subsection{Experimental Settings}
\textbf{Datasets: }Our graph model inversion attack method is evaluated on 7 public datasets from 4 categories. The detailed statistics of them are listed in the appendix.
\begin{itemize}
\item \textit{Citation Networks}: We use Cora and Citeseer \cite{sen2008collective}. Here, nodes are documents with corresponding bag-of-words features and edges denote citations among nodes. Class labels denote the subfield of research that
the papers belong to.
\item \textit{Social Networks}: Polblogs \cite{adamic2005political} is the network of political blogs whose nodes do not have features.
\item \textit{Air-Traffic Networks}: The air-traffic networks are based on flight records from USA and Brazil. Each node is an airport and
an edge indicates a commercial airline route
between airports. Labels denote the level of activity
in terms of people and flights passing through an
airport \cite{ribeiro2017struc2vec}.
\item \textit{Chemical Networks}: AIDS \cite{riesen2008iam} and ENZYMES \cite{borgwardt2005protein} are chemical datasets that contain many molecure graphs, each node is an atom and each link represents chemical bonds.
\end{itemize}
\paragraph{Target Models:}In our evaluation, we use 3 state-of-the-art GNN models: GCN \cite{kipf2016semi}, GAT \cite{velivckovic2017graph} and GraphSAGE \cite{hamilton2017inductive}.
The parameters of the models are the same as those set in the original papers. To train a target model, 10\% randomly sampled nodes are used as the training set. All GNN models are trained for 200 epochs with an early stopping strategy based on convergence behavior and accuracy on a validation set containing 20\% randomly sampled nodes. In GraphMI attack experiments, attackers have labels of all the nodes and feature vectors. All the experiments are conducted on Tesla V100 GPUs.
\paragraph{Parameter Settings:}In experiments, we set $\alpha=0.001$, $\beta=0.0001$, $\eta_t=0.1$ and $T=100$ as the default setting. We show how to find optimal values for hyper-parameters in the following section.
\paragraph{Metrics:}Since our attack is unsupervised, the attacker cannot find a threshold to make a concrete prediction through the algorithm. To evaluate our attack, we use AUC (area under the ROC curve) and AP (average precision) as our metrics, which is consistent with previous works \cite{kipf2016variational}. In experiments, we use all the edges from the training graph and the same number of randomly sampled pairs of unconnected nodes (non-edges) to evaluate AUC and AP.
\begin{figure}[t]
\centering
\includegraphics[width=0.7\linewidth]{figs/gnn.png}
\caption{Attack performance of GraphMI on different Graph Neural Networks.}
\label{gnn type}
\end{figure}
\subsection{Results and Discussions}
\paragraph{Attack Performance.}Results for model inversion attacks on GCN are summarized in table \ref{result}. There are two baseline methods, attribute similarity (abbreviated as Attr. Sim.) and MAP. Attribute similarity is measured by cosine distance among node attributes, which is commonly used in previous works \cite{he2020stealing}. We adapt the model inversion method from \cite{fredrikson2015model}, MAP to the graph neural network setting as the other baseline. Note that some datasets such as Polblogs dataset do not have node attributes, so that we assign one-hot vectors as their attributes. They are not applicable for the attack based on attribute similarity. As can be observed in table \ref{result}, GraphMI achieves the best performance across nearly all the datasets, which demonstrates the effectiveness of GraphMI. One exception is Citesser where the attack performance of GraphMI is relatively lower than attribute similarity, which could be explained by more abundant node attribute information of Citeseer compared with other datasets. Thus using node attribute similarity alone could achieve good performance in the Citesser dataset.
In figure \ref{gnn type}, we show the attack performance of GraphMI on three GNNs. We observe that GraphMI has better attack performance on GAT model. This may be explained by the fact that GAT model is more powerful and is able to build a stronger correlation between graph topology and node labels. GraphMI can take advantage of such a stronger correlation and achieve better attack performance. \par
In figure \ref{rq1}, we present the influence of node label proportion on attack performance. As can be observed from the plot, with fewer node labels, the attack performance will drop gradually. But GraphMI can still achieve over 80 $\%$ AUC and AP when only 20$\%$ node labels are available, which again verifies the effectiveness of GraphMI.
\par We also explore the sensitivity of hyper-parameters $\alpha$ and $\beta$ for GraphMI. In the experiments, we alter the value of
$\alpha$ and $\beta$ to see how they affect the performance of GraphMI. Specifically, we vary $\alpha$ from 0.00025 to 0.008 and $\beta$ from 0.00005 to 0.0016 in a log scale of base 2. The attack performance change of GraphMI is illustrated in Fig~\ref{param analysis}. As we can observe, the attack performance of GraphMI can be boosted when choosing proper values for all the hyper-parameters.
\begin{figure}[t]
\centering
\includegraphics[width=0.7\linewidth]{figs/label.png}
\caption{Impact of node label proportion.}
\label{rq1}
\end{figure}
\begin{figure}[t]
\centering
\subfigure[]{\includegraphics[width=0.48\linewidth]{figs/alpha.png}}
\subfigure[]{\includegraphics[width=0.48\linewidth]{figs/beta.png}}
\caption{ Results of parameter analysis on Cora dataset}
\label{param analysis}
\end{figure}
\paragraph{Edge Influence.}We do experiments to verify our claim that edges with greater influence are more likely to be inferred successfully through model inversion attack. Note that it will be very time-consuming to measure the influence of each edge exactly. According to equation (\ref{edge}), removing edges with greater influence will cause greater drop of prediction accuracy. To select edges with great influence, we apply the state-of-the-art topology attack \cite{xu2019topology} on graphs by removing edges. In Figure \ref{edge influence}, we show that for edges with top $5\%$ influence GraphMI achieves the attack AUC of nearly 1.00 in Cora dataset. This implies that the privacy leakage will be more severe if sensitive edges are those with greater influence.
\begin{figure}[t]
\centering
\includegraphics[width=0.7\linewidth]{figs/influence.png}
\caption{Impact of edge influence on the performance of the GraphMI attack.}
\label{edge influence}
\end{figure}
\begin{table}[h]
\centering
\begin{tabular}{lrc}
\toprule
Method & ACC & GraphMI AUC\\ \midrule
$\epsilon=1.0 $ & 0.48 & 0.60 \\
$\epsilon=5.0 $ & 0.65 & 0.72 \\
$\epsilon=10.0 $ & 0.78 & 0.84 \\
no DP & 0.80 & 0.87 \\ \bottomrule
\end{tabular}
\caption{The performance of the GraphMI attack against GCN trained with differential privacy on Cora dataset}
\label{DP}
\end{table}
\paragraph{Defense Performance of Differentail Privacy.}
Differential privacy (DP) is one general approach for protecting privacy.
Here, we investigate the impact of differential privacy on GraphMI attacks. $(\epsilon,\delta)$ -
DP is ensured by adding Gaussian noise to clipped gradients in each training iteration \cite{abadi2016deep}. In experiments, $\delta$ is set to $10^{-5}$ and the noise scale is varied to obtain target GNN models with different $\epsilon$ from 1.0 to 10.0.
The GraphMI attack performance and their model utility are presented in Table \ref{DP}. As the privacy budget $\epsilon$ drops, the performance of GraphMI attack deteriorates at the price of a huge utility drop. Generally, enforcing DP on target models cannot prevent GraphMI attack.
\section{Conclusion}
In this paper, we presented GraphMI, a model inversion attack method against Graph Neural Networks.
Our method was specifically designed and optimized for extracting private graph-structured data from GNNs.
Extensive experimental results showed its effectiveness on several state-of-the-art graph neural networks.
We also explored and evaluated the impact of node label proportion and edge influence on the attack performance.
Finally, we showed that imposing differential privacy on graph neural networks can hardly protect privacy while preserving decent utility.
This paper provided potential tools for investigating the privacy risks of deep learning models on graph-structured data.
Interesting future directions include: 1) Extending the current work to a black-box setting.2) Design countermeasures with a better trade-off between utility and privacy.
\section*{Acknowledgments}
This research was partially supported by grants from the National Natural Science Foundation of China (Grants No.61922073 and U20A20229), and the Fundamental Research Funds for the Central Universities (Grant No.WK2150110021).
\bibliographystyle{named}
|
\section{Introduction}
The geometry of weight manifolds and functional spaces represented by artificial neural networks is an important window to ‘understanding’ machine learning. Many open questions in machine learning, when viewed through the lens of geometry, can be related to finding points or paths of equivalent function in the weight and functional spaces \cite{cohen2020separability, jia2019geometric, hauser2018principles, anderson2017high}. Although geometric structure plays a key role in determining the properties of neural network training, application of methods from differential geometry to machine learning is complicated by the dependence on millions of network parameters, the non-convex structure of loss functions, and the lack of computationally efficient algorithms that harness the geometric structure to traverse weight or functional space \cite{auer1996exponentially, choromanska2015loss}.
In this paper, we propose a mathematical framework grounded in differential geometry for constructing path-connected sets of deep neural networks that have equivalent functional performance on a task. The key contribution of our paper is that we view the networks' weights space as a pseudo-Riemannian manifold equipped with a distance metric that represents task performance while simultaneously capturing task-independent network properties, like network sparseness. We formalize the ``search'' for a suitable network (based on the application of interest) as a dynamic movement on the curved pseudo-Riemannian manifold \cite{amari1982differential}. Further, we demonstrate that geodesics, minimum length paths, on the network weights space provide high performance paths that the network can traverse to maintain performance while `searching-out' for other networks that satisfy additional objectives. Specifically we develop a procedure based on the geodesic equation to find sets of path connected networks that achieve high performance while also satisfying a second objective like sparsification or mitigating catastrophic interference.
We demonstrate that our framework can be applied to solve three (seemingly unrelated) major problems in machine learning: (i) Discovering sparse counterparts of dense neural networks and high-accuracy paths that connect the two (dense and sparse networks), (ii) Enabling continual learning by mitigating catastrophic forgetting and (iii) Finding high-accuracy paths connecting two trained deep networks (mode-connectivity) in a non-convex loss landscape. These applications, when viewed through the lens of differential geometry can be solved by finding points or paths of equivalent function in the functional space of deep networks. Broadly, our paper demonstrates that differential geometry can provide a mathematical framework and novel algorithms to unify open problems in machine learning in a common geometric language.
\section{Related work}
The parameters of a neural network encodes the function that maps a set of inputs to outputs. Although the function mapping input/output is crucial for a large number of ML applications, the intractability of the function-space has veered researchers away and instead focus on techniques and analyses that concern the parameters of the neural network. The introduction of information geometry by Amari \cite{amari1982differential} pioneered the analysis of neural networks from the lens of their function and output spaces. Natural gradient descent (NGD) \cite{amari1998natural} utilized the KL distance between functions to efficiently train neural networks, by evaluating the Fisher-information matrix for scaling gradient updates based on parameters informativeness. Since then, many distance metrics\cite{neyshabur2015path}, like Deep relative trust \cite{bernstein2020distance}, the L$^2$ distance between neural network functions in Hilbert space \cite{benjamin2018measuring} have been developed for computing the functional distance between neural networks. In addition, distance metrics and algorithms have been developed to navigate the object manifolds in order to learn the best transformations for pattern recognition applications \cite{simard1998transformation}.
\section{Geodesics: Mathematical Framework}
We develop a mathematical framework grounded in differential geometry for navigating the space of neural networks to discover novel networks that have high task-performance while satisfying additional constraints on task-independent network properties, like sparseness.
\textbf{Preliminaries}:
We represent a feed-forward neural network (NN) as a smooth, $\mathbb{C}^\infty$function $f(\mathbf{x};\mathbf{w})$, that maps an input vector, $\mathbf{x} \in \mathbb{R}^{\text{k}}$, to an output vector, $f(\mathbf{x};\mathbf{w}) = \mathbf{y}\in \mathbb{R}^{\text{m}}$. The function, $f(\mathbf{x};\mathbf{w})$, is parameterized by a vector of weights, $\mathbf{w} \in \mathbb{R}^\text{n}$, that are typically set in training to solve a specific task. We refer to $W= \mathbb{R}^n$ as the \textit{weight space} ($W$) of the network, and we refer to $\mathcal{Y} = \mathbb{R}^m$ as the \textit{functional space} \cite{mache2006trends}. We also define a loss function, $L: \mathbb{R}^\text{m} \times\mathbb{R} \rightarrow \mathbb{R}$, that provides a scalar measure of network performance for a given task (Figure 1). Please note that the functional space $\mathcal{Y}$ and the loss space $L$ are \textbf{task-dependent} spaces, while the weights space $W$ is \textbf{task-independent} and encodes network properties, like fraction of non-zero weights.
We construct a metric tensor ($\mathbf{g}$) to evaluate how infinitesimal movements in the weights space $W$ impacts movement in the functional space ($\mathcal{Y}$), effectively measuring the functional-similarity of networks before and after weight perturbation. The metric tensor can be applied at any point in $W$ to measure the functional impact of an arbitrary network weights perturbation.
To construct a metric mathematically, we fix the input, $\mathbf{x}$, into a network and ask how the output of the network, $f(\mathbf{x},\mathbf{w})$, moves on the functional space, $\mathcal{Y}$, given an infinitesimal weight perturbation, $\mathbf{du}$, in $W$ where $\mathbf{w_p} = \mathbf{w_t} + \mathbf{du}$. For an infinitesimal perturbation $\mathbf{du}$,
\begin{align}
f(\mathbf{x},\mathbf{w_t} +\mathbf{du}) \approx f(\mathbf{x},\mathbf{w_t}) +\mathbf{ J_{w_t}} \ \mathbf{du},
\vspace{-2mm}
\end{align}
where $\mathbf{ J_{w_t}}$ is the Jacobian of $f(\mathbf{x},\mathbf{w_t})$ for a fixed $\mathbf{x}$, $J_{i,j} = \frac{\partial f_i}{\partial w^j}$, evaluated at $\mathbf{w_t}$. We measure the change in functional performance given weight perturbation $\mathbf{du}$ as:
\begin{align}
\label{local metric}
d(\mathbf{w_t},\mathbf{w_p}) = |f(\mathbf{x},\mathbf{w_t})-f(\mathbf{x},\mathbf{w_p})|^2 = \mathbf{du}^T \ (\mathbf{J_{w_t}(\mathbf{x})}^T \ \mathbf{J_{w_t}(\mathbf{x})}) \ \mathbf{du} =
\mathbf{du}^T \ \mathbf{g_{w_t}(\mathbf{x})} \ \mathbf{du}
\end{align}
where $\mathbf{g_{w_t}(\mathbf{x})}$ = $\mathbf{J_{w_t}}$($\mathbf{x})^T$ $\mathbf{J_{w_t}(\mathbf{x})}$ is the metric tensor evaluated at the point $\mathbf{w_t} \in W$ for a single datapoint ($\mathbf{x}$). The metric tensor is an $n \times n$ symmetric matrix that defines an inner product and local distance metric, $\langle \mathbf{du}, \mathbf{du} \rangle_{\mathbf{w}} = \mathbf{du^T} \ \mathbf{g_{w}(\mathbf{x})} \ \mathbf{du}$, on the tangent space of the manifold, $T_w(W)$ at each $\mathbf{w} \in W$. Explicitly,
\begin{equation}
g_{ij}(\mathbf{x}) = \sum_{k=1}^{m}\frac{\partial f_k(\textbf{x},\mathbf{w})}{\partial \mathbf{w^i}}\frac{\partial f_k(\textbf{x},\mathbf{w})}{\partial \mathbf{w^j}},
\end{equation}where the partial derivatives $\frac{\partial f_k(\textbf{x},\mathbf{w})}{\partial \mathbf{w^i}}$ measure change in functional output of a network given a change in weight. In the appendix, we extend the metric formulation to cases where we consider a set of N training data points, $\mathbf{X}$, and view $\mathbf{g}$ as the average of metrics derived from individual training examples. $ \mathbf{g_w} = \mathbf{g_w(X)} = \Sigma_{i=1}^N\mathbf{g_w(x_i)}/N$. The metric, $\mathbf{g}$, provides a local measure of \textit{functional distance} on the pseudo-Riemmanian manifold $(W,\mathbf{g})$. At each point in weight space, the metric defines the length, $\langle \mathbf{du}, \mathbf{du} \rangle_{\mathbf{w}}$, of a local perturbation by its impact on the functional output of the network (Figure 1B).
\begin{figure}[t]
\centering
\centerline{ \includegraphics[width=0.8\linewidth]{Figures/geodesic_framework.png}}
\caption{\textbf{Geometric framework for constructing paths in functional space} (A) Three networks ($\mathbf{w_1},\mathbf{w_2},\mathbf{w_3}$) in weights space $W$ and their relative distance in functional space and loss space. High-performance path is determined by asking how movement in weight space changes functional performance and loss through introduction of a pullback metric $\mathbf{g}$. (B) The metric tensor ($\mathbf{g}$) is evaluated by analyzing the effect of infinitesimal perturbation in the tangent space of a network.(C) Paths between pairs of networks are modeled as long range movement of network weights along a path, $\gamma(t)$, in weight space. }
\label{fig:schematics}
\vspace{-4mm}
\end{figure}
Our framework discovers a network that satisfies multiple objectives by constructing a geodesic between two locations in the networks weights space ($W$): one defined by a network that maximizes task performance, $\mathbf{w_t}$, and another defined by a network that satisfies task-independent constraints, the secondary goal, $\mathbf{w_a}$. $\mathbf{w_a}$ could be a single network (if known) or a subspace of networks that satisfy the constraints (if the network is yet to be discovered).
The global path is constructed in order to simultaneously minimize the movement of the network on the functional space while moving in weights space towards a second point to achieve a secondary goal. We use the metric tensor to determine the functional change across a path-connected set of networks in the networks weights space leading to the second point $\mathbf{w_a}$. Mathematically, the metric changes as we move in $W$ due to the curvature of the ambient space that reflects changes in the vulnerability of a network to weight perturbation (Figure 1C).
As a network moves along a path $\gamma(t) \in W$ from start network $\mathbf{w_t}$ to the second point encoding the secondary goal $\mathbf{w_a}$, we can analyze the integrated impact on the network performance by using the metric ($\mathbf{g}$) to calculate the length of the path $\gamma(t)$ as:
\begin{equation}
S(\gamma) = \int^1 _0 {\langle \frac{d\gamma(t)}{dt}, \frac{d\gamma(t)}{dt}\rangle_{\gamma(t)} } \ dt ,
\label{eq:geodesicLength}
\end{equation}
where $\langle \frac{d\gamma(t)}{dt}, \frac{d\gamma(t)}{dt}\rangle_{\gamma(t)} = \frac{d\gamma(t)}{dt}^T \mathbf{g}_{\gamma(t)} \frac{d\gamma(t)}{dt}$ is the infinitesimal functional change accrued while traversing path $\gamma(t) \in W$. As the shortest path in functional space - min($S(\gamma$)) is desirable to ensure that the path connected networks are functionally similar, we evaluate the geodesic from $\mathbf{w_t}$ to $\mathbf{w_a}$.
Minimizing $S(\gamma)$ is equivalent to solving the geodesics on $W$ equipped with metric tensor $\mathbf{g}$.
\begin{align}
\frac{d^2 w^\eta}{dt^2} + \Gamma_{\mu\nu}^\eta\frac{dw^\mu}{dt} \frac{dw^\nu}{dt} = 0
\label{eq:geodesicMotion}
\end{align}
where, $w^j$ defines the j'th basis vector of the weights space $W$, $\Gamma_{\mu\nu}^\eta$ specifies the Christoffel symbols $( \Gamma_{\mu \nu}^\eta = \sum_r \frac{1}{2}g_{\eta r}^{-1}(\frac{\partial g_{r \mu}}{\partial x^\nu} + \frac{\partial g_{r \nu}}{\partial x^\mu} - \frac{\partial g_{\mu \nu}}{\partial x^r}))$ on the manifold. The Christoffel symbols record infinitesimal changes in the metric tensor ($\mathbf{g}$) along a set of directions on the manifold (Appendix). Since the computation and memory for evaluating Christoffel symbols scales as third order polynomial of network parameters ($\mathcal{O}(n^3)$), we propose an optimization algorithm for evaluating `approximate' geodesics in the manifold.
\textbf{Optimization procedure for approximating geodesics :}
Inspired by the local formulation of the geodesics equation, we introduce an optimization procedure: Geo($\mathbf{w_t}$, $\mathbf{g}$, $\mathbf{w_a}$, $s$) to construct a global path to find networks that have high performance as well as satisfy additional task-independent constraints. The inputs to the procedure Geo($\mathbf{w_t}$, $\mathbf{g}$, $\mathbf{w_a}$, $s$) are: (i) Start network that maximizes task performance ($\mathbf{w_t}$), (ii) Metric to measure change in task-performance when network moves on weights space ($\mathbf{g}$), (iii) Second network that encodes the secondary goal ($\mathbf{w_a}$) and (iv) User-defined number of steps taken along the path (s). The output of the optimization procedure is the path $\gamma(t)$ beginning from $\gamma(0)$ = $\mathbf{w_t}$ and ending at $\gamma(1) = \mathbf{w_c}$. $\mathbf{w_c}$ could be $\mathbf{w_a}$ for a complete traversal, or could be a different network if the stopping criterion terminates the optimization before the complete traversal.
Starting at $\mathbf{w_t}$, we iteratively solve for $\theta(\mathbf{w})$ using equation-6,7 to traverse the path from $\mathbf{w_t}$ to $\mathbf{w_a}$ in the networks weights space ($W$). $\theta(\mathbf{w})$ is a vector at point $\mathbf{w} \in W$ whose length measures a linear estimate of the change in performance of the network incurred by moving an infinitesimal distance in a given direction in weight space. The procedure finds a direction that (i) minimizes the functional change between networks on a task before and after an update $\{$min: $\langle \theta(\mathbf{w}), \theta(\mathbf{w})\rangle_\mathbf{w}$ = $\theta(\mathbf{w})^T\mathbf{g_w}\theta(\mathbf{w})$ $\}$ while (ii) moving towards the target network ($\mathbf{w_a}$), achieved by maximizing the dot-product of the tangent vector and vector pointing towards $\mathbf{w_a}$: $\{$ max: $\theta(\mathbf{w})^T$ $(\mathbf{w_a}-\mathbf{w})$ $\}$. Having evaluated a suitable $\theta(\mathbf{w})$, we update the networks weights via equation-7, where $\eta$ is the step-size of the update.
We fix the metric tensor ($\mathbf{g}$) as a representation of the performance on the task as it measures the functional difference before and after every update step, while the direction ($\mathbf{w_a} - \mathbf{w}$) towards $\mathbf{w_a}$ encodes the secondary goal.
\vspace{-3mm}
\begin{equation}
\text{argmin}_{\theta(\mathbf{w})} \ \langle \theta(\mathbf{w}), \theta(\mathbf{w}) \rangle_\mathbf{w} - \beta \ \theta(\mathbf{w})^T (\mathbf{w_a}-\mathbf{w}) \\ \text{ subject to: } \theta(\mathbf{w})^T \theta(\mathbf{w}) \leq 0.01. \\
\end{equation}
\vspace{-4mm}
\begin{align}
\Delta \mathbf{w} = \eta \frac{\theta(\mathbf{w})}{||\theta(\mathbf{w})||}
\end{align}
The stopping criterion for Geo($\mathbf{w_t}$, $\mathbf{g}$, $\mathbf{w_a}$, $s$) are as follows: (i) Network traverses the entire path from $\mathbf{w_t}$ to $\mathbf{w_a}$, (ii) Number of steps taken reaches user-defined $s$ fed as an input to the procedure and (iii) Network makes small oscillations (moving to and away from $\mathbf{w_a}$). Please note that if stopping criterion (ii) or (iii) is reached, the output of the procedure is the path that terminates at a network different from $\mathbf{w_a}$.
Our optimization procedure is a quadratic program that trades off, through the hyper-parameter $\beta$, motion towards the target network that encodes the secondary goal ($\mathbf{w_a}$) and the maximization of the functional performance of the intermediate networks along the path (elaborated in the appendix). The strategy discovers multiple paths from the trained network $\mathbf{w_{t}}$ to $\mathbf{w_a}$ (encoding secondary goal) where networks maintain high functional performance during traversal. Of the many paths obtained, we can select the path with the shortest total length (with respect to the metric $\mathbf{g}$) as the best approximation to the geodesic in the manifold.
\section{Geodesic framework applied to three distinct ML problems }
In the sections that follow, we recast three distinct open questions in ML through the lens of geometry and find solutions by constructing approximate geodesics. The three applications are: (i) Sparsifying networks by traversing geodesics, (ii) Alleviating catastrophic forgetting via geodesics and (iii) Connecting modes of deep neural networks by constructing geodesic paths.
\subsection{Sparsifying networks by traversing geodesics}
Network sparsification has gained importance in the recent years. Although deep networks are powerful systems, they require lots of computation and memory making their deployment in resource-constrained environments like mobile phones and smart-devices challenging \cite{blalock2020state}.
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{Figures/schematic_sparseNet2.png}
\caption{\textbf{Geodesic framework for discovering sparse networks:} (A) Dense network ($\mathbf{w_t}$) and 4 networks ($\mathbf{w_2}, \mathbf{w_3}, \mathbf{w_4}, \mathbf{w_5}$) on the p-sparse hyperplane in the Weight space (left) and their relative distances on the functional manifold (right). (B) $\mathbf{w_t}$, $\mathbf{w_2}$ to $\mathbf{w_5}$ represented on the loss surface. }
\label{fig:sparseSchematic}
\vspace{-4mm}
\end{figure}
\begin{wrapfigure}{r}{0.4\textwidth}
\includegraphics[width=1\linewidth]{Figures/lotteryTicket_comparison.png}
\caption{Difference in Top-1 $(\Delta_1)$ test accuracy between the discovered sparse networks and their reference dense network. A comparison with the lottery ticket hypothesis is presented.}
\label{fig:sparse_results2}
\vspace{-3mm}
\end{wrapfigure}
Here, we aim to discover a network that simultaneously has (i) high performance on a particular task (eg. CIFAR-10 classification) and (ii) satisfies sparsity constraints, by setting a fraction of its weights to zero. Our optimization procedure (Geo($\mathbf{w_t}$, $\mathbf{g}$, $\mathbf{w_a}$, $s$)) addresses both the objectives by (i) setting the start point $\mathbf{w_t}$ to a dense trained network on the task, followed by computing the metric tensor ($\mathbf{g}$) for the task (eg CIFAR-10 classification) to evaluate change in task-performance while (ii) moving towards a target network $\mathbf{w_a}$:= p-sparse network, which has p$\%$ of its networks' weights set to zero (encodes the additional sparsity constraint).
Here, the target sparse network is yet to be discovered, so we designate $\mathbf{w_a}$ as a member of the subspace of networks that satisfies the sparsity constraint. So, we define a p-sparse subspace ($H_p$) in the networks' weights space as a collection of all networks with p$\%$ of their weights set to zero. $H_p = \{ \mathbf{w} \in \mathcal{R}^n : ||\mathbf{w}||_0 = \frac{np}{100} \}$. We choose $\mathbf{w_a}$ to be the projection of the dense network on $H_p$; $\mathbf{w_a}$ = proj($\mathbf{w_t}, H_p$). We constantly update our target network to proj($\mathbf{w}$, $H_p$) every $n_s$ steps taken by the network.
\begin{algorithm}
\caption{Discovering sparse networks via geodesics}
\begin{algorithmic}[1]
\While{$\mathbf{w_t} \not\in H_p$} \Comment{If start network isn't on p-sparse hyperplane}
\State $\mathbf{w_a} \leftarrow proj(\mathbf{w_t}, H_p$) \Comment{Project network on p-sparse hyperplane}
\State $\gamma(t) \leftarrow Geo(\mathbf{w_t}, \mathbf{g}, \mathbf{w_a}, n_s$)
\State $\mathbf{w_c} \leftarrow \gamma(1)$ \Comment{End point of geodesic after taking $n_s$ steps }
\State $\mathbf{w_t} \leftarrow \mathbf{w_c}$ \Comment{New start network to evaluate geodesic}
\EndWhile
\end{algorithmic}
\end{algorithm}
Figure-\ref{fig:sparseSchematic}A illustrates our adaptation of the geodesic framework for finding functionally similar p-sparse networks by constructing a geodesic to the p-sparse subspace ($H_p$). Here, $\mathbf{w_t}$ is the trained dense network, and $\mathbf{w_2}$, $\mathbf{w_3}$, $\mathbf{w_4}$ and $\mathbf{w_5}$ are p-sparse networks on $H_p$. Their corresponding positions in the functional space highlight the functional closeness of some p-sparse networks to the trained network over others. In figure-\ref{fig:sparseSchematic}B, p-sparse network - $\mathbf{w_2}$ is functionally closest to $\mathbf{w_t}$.
We demonstrate our geodesic strategy for sparsification on: (i) Multilayer perceptron (LeNet) trained on MNIST and (ii) VGG-11 trained on CIFAR-10.
In figure-\ref{fig:sparse_results2}, we show that the geodesic strategy discovers sparse networks that perform at test accuracies comparable to the reference dense trained network $\mathbf{w_t}$. Our results (figure-\ref{fig:sparse_results}A supersede the existing benchmarks for LeNet-300-100 compression reported in literature for extreme levels of sparsity \cite{manessi2018automated, babaeizadeh2016noiseout, frankle2018lottery, alford2019training}. A comparison with the lottery ticket hypothesis \cite{frankle2018lottery} is presented in fig-\ref{fig:sparse_results2}.
In addition to finding the p-sparse network (on $H_p$), we obtain a high-performance path connecting the dense-MLP trained on MNIST to the discovered sparse network on $H_p$. In Figure-\ref{fig:sparse_results}B, we show that
path-connected networks from the dense network to $H_{50}$, $H_{75}$, $H_{90}$ and $H_{95}$ also perform at an accuracy $\geq$97$\%$.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{Figures/SparsifyNetworks2.png}
\caption{\textbf{Traversing geodesics to sparsify networks} (A) Test performance of sparse LeNet architectures on MNIST discovered by traversing the geodesic from the dense-MLP to the p-sparse hyperplane. The geodesic locates sparse networks that perform $\geq$97.5$\%$ for sparsities ranging from from 50 to 98.1$\%$ and finds a sparse network performing at $\sim$97$\%$ for 99$\%$ sparsity and $\sim$96.8$\%$ for 99.1$\%$ sparse network. (B) The paths traversed from the trained dense-MLP to p-sparse hyperplane (p$\in$[50,75,90,95]) are high-performing as they perform at an accuracy $\geq$97.4$\%$. (C) Test accuracy and (D) number of network update epochs for geodesic recovery (orange) vs fine-tuning (blue) of VGG11 trained on CIFAR-10, while 50 (out of 64) conv-filters are deleted from layer1 in VGG11. (D) Geodesic recovery requires $\leq$30 total update epochs, while fine-tuning requires upto 120 epochs. }
\label{fig:sparse_results}
\vspace{-7mm}
\end{figure}
Our geodesic strategy discovers structured sparse counterparts of VGG-11 trained on CIFAR-10. In Figure-\ref{fig:sparse_results}C , our geodesic approach yields high-performance paths that connect the dense-VGG11 network to its sparser counterpart wherein 50 (out of 64) conv filters from the first layer are zeroed out. We compared our strategy to traditional heuristic fine-tuning to demonstrate that our approach is both rationale and computationally efficient. Specifically, an iterative prune-train cycle achieved through structured pruning of a single node at a time, coupled with SGD re-training requires upto 120 training epochs to identify a sparsification path. However, our geodesic strategy finds paths that quantitatively out-perform the iterative prune-train procedure and obtains these paths with only 30 training epochs (Figure-\ref{fig:sparse_results}D).
\subsection{Alleviating Catastrophic forgetting by traversing geodesics}
Neural networks succumb to catastrophic forgetting (CF) during sequential training of tasks because training on sequential tasks alters the weights between nodes in the neural network which are locations of "stored knowledge", resulting in the abrupt loss of "memory" of all information from previous tasks \cite{mccloskey1989catastrophic, ratcliff1990connectionist}. Previous attempts to solve the problem of CF faced by deep networks was accomplished by meticulous tuning of network hyperparameters \cite{bengio2013empirical, srivastava2013compete} accompanied by standard regularization methods. Addressing earlier limitations, Kirkpatrick et al \cite{kirkpatrick2017overcoming} proposed elastic weight consolidation, wherein they evaluate a Fisher information matrix to guide retraining of network on a new task.
\begin{figure}[H]
\centering
\includegraphics[width=1\linewidth]{Figures/CF_all.png}
\caption{\textbf{Sequential Permuted-MNIST tasks: } (A) Two LeNet's are trained on permute-MNIST task-1,2 ($\mathbf{w_1}, \mathbf{w_2}$). The geodesic between $\mathbf{w_2}$ and $\mathbf{w_1}$ using metric tensor on task-2 ($\mathbf{g^2}$) discovers a network that performs at 98$\%$ test-accuracy on both tasks. The best network along the linear path performs at 80$\%$ on both tasks. (B) PCA projections of the network weights show geodesic obtained from Geo($\mathbf{w_2}$, $\mathbf{g^2}$, $\mathbf{w_1}$, $N_s$) (magenta) and linear path (blue). Both paths begin from $\mathbf{w_t} = \mathbf{w_2}$ net trained on task-2 (red circle) and move to the target $\mathbf{w_a} = \mathbf{w_1}$ net trained on task-1 (red triangle). (C) A third network pre-trained on Task-3 ($\mathbf{w_3}$) moves to target network $\mathbf{w_a}$ trained on task-1,2 obtained from (A). The geodesic finds networks that performs at [97$\%$, 97$\%$, 98$\%$] while the best network along linear path performs at [70$\%$, 60$\%$, 60$\%$] on Tasks-1,2,3 respectively. (D) A fifth network pretrained on Task-5, moves to the target network $\mathbf{w_a}$ trained on task-1,2,3,4. The geodesic path finds networks that perform at [94,95,96,96,98$\%$] on Tasks 1 to 5 respectively while the best network along the linear path performs at $\sim$60$\%$ on all 5 tasks. (E) The 'circle' and 'triangle' correspond to CNN's trained on MNIST and Fashion-MNIST datasets respectively. The x,y axes are tSNE projections of the network weights and z axes corresponds to the mean performance of the network on both, MNIST and F-MNIST. The linear path (blue) between the networks hosts networks that perform on average 40$\%$ on both tasks, while the geodesic approach discovers a curved path that performs at 94$\%$ on MNIST and 82$\%$ on F-MNIST. (F) A comparison of the linear path and geodesic approach reveals that the curved path converges at a network that simultaneously performs at 94$\%$ on MNIST and 82$\%$ on F-MNIST, while the linear path finds a network that performs at 40$\%$ on MNIST and F-MNIST. }
\label{fig:CF_MNISTpermute}
\vspace{-4mm}
\end{figure}
We apply our geodesic framework in a novel fashion to mitigate CF while training networks on sequential tasks. To alleviate CF while learning $k$ sequential tasks, our goal is to discover a network that achieves a high performance on all $k$ tasks, given access to only one task at a time. Here, task is synonymous to dataset. For instance, training a network sequentially on MNIST, followed by Fashion-MNIST constitutes two sequential tasks.
We evaluate $k$ metric tensors ($\mathbf{g^1}$, $\mathbf{g^2}$, ..., $\mathbf{g^k}$) corresponding to $k$ sequential tasks (or datasets - ($X_1$,$X_2$, ..., $X_k$)). The metric, $\mathbf{g^i}$ provides a local measure of the functional distance on the pseudo-Riemannian manifold (W,$\mathbf{g^i}$), ie it measures the change in performance on task-i as the network moves on the weights space. Our geodesic approach discovers networks that perform well on sequential tasks by constructing geodesics between two locations in the weights space, one defined by network trained on the most recent task (task-i), while the other location is defined by the network trained on all previous tasks (task-1,2,..,i-1). The metric $\mathbf{g^i}$ measures the change in performance on task-i as the network moves towards a location in the weights space defined by another network trained on all previous tasks (task-1,2, ..., i-1). Therefore, our optimization strategy finds a set of path-connected networks beginning at network trained on task-i, and moving to another network trained on all previous tasks (1,2,..,i-1) with the objective of minimizing the change in performance on task-i alone. The procedure converges\footnote{One of the stopping criterion mentioned in section-2 is reached - where the network makes small oscillations towards and away from the target network.} at a network that performs well on all tasks (including most recent task-i) seen until then (task-1,2,...., i).
\begin{wrapfigure}{r}{0.5\textwidth}
\includegraphics[width=1\linewidth]{Figures/CF_schema2.png}
\caption{\textbf{Framework for mitigating catastrophic interference via geodesics} (A) Pretrained network on task-1 ($\mathbf{w_1}$) (B) Geodesic evaluated using metric on task-2 ($\mathbf{g^2}$) connecting $\mathbf{w_2}$ (pretrained on task-2) and $\mathbf{w_1}$ converges at $w_{12}$.}
\label{fig:CF_schema}
\vspace{-3mm}
\end{wrapfigure}
In figure-\ref{fig:CF_schema}, we illustrate our framework in action for learning two sequential tasks. We begin by training two different networks on the sequential tasks one after the other to get $\mathbf{w_1}$ and $\mathbf{w_2}$ trained on task-1,2 respectively. Subsequently, we compute the geodesic on the networks weights space, beginning from $\mathbf{w_t}$ = $\mathbf{w_2}$, network trained on task-2 by traversing the manifold (W, $\mathbf{g^2}$) towards the target network $\mathbf{w_a}$ = $\mathbf{w_1}$, network trained on task-1. The metric $\mathbf{g^2}$ measures the change in performance on task-2 as the network moves towards a location in the weights space defined by the network trained on the first task. The procedure converges at a network $\mathbf{w_{1,2}}$ that performs well on both tasks.
\vspace{-3mm}
\begin{algorithm}
\caption{Alleviating CF: Learning $k$ sequential tasks via geodesics}
\begin{algorithmic}[1]
\State Train net on Task-1 ($\mathbf{w_1}$) \Comment{Train random network on Task-1}
\State Set $i$ = 2 \Comment{Update to Task-2}
\While{$i$ $\leq$ k} \Comment{Iterate procedure for $k$ tasks}
\State Train net on Task:i ($\mathbf{w_i}$) \Comment{Train random network on latest task}
\State $\mathbf{w_p}$ = $\mathbf{w_{1:i-1}}$ \Comment{$\mathbf{w_p}$ := network that performs well on all previous tasks: task:(1,2,..,i-1)}
\State $\mathbf{g} = \mathbf{g^i}$ \Comment{Compute metric tensor for most recent task (task-i)}
\State $\gamma(t) \leftarrow Geo(\mathbf{w_i}, \mathbf{g}, \mathbf{w_p}, N_s) $ \Comment{Constructing geodesic path}
\State $\mathbf{w_p} \leftarrow \gamma(1)$ \Comment{Terminal net of geodesic performs well on all tasks}
\State Set $i$ = $i$ + 1 \Comment{Moving on to the next task}
\EndWhile
\end{algorithmic}
\end{algorithm}
\vspace{-3mm}
We apply our geodesic framework to the classic permuted-MNIST task \cite{le2015simple}, wherein every new task corresponds to a (fixed) unique permutation of the input pixels of the MNIST dataset. For $k$ tasks, we have $k$ permuted MNIST datasets. We use the Lenet-MLP (784-300-100-10) architecture to test our framework on sequential learning of permute-MNIST tasks.
To highlight the performance of our optimization strategy, we compare the geodesic path with the linear path between two locations in weights space: one defined by the network trained on the latest task, while the other trained on all previous tasks - obtained by iterative application of the optimization strategy. Figure-\ref{fig:CF_MNISTpermute}A shows that the network discovered along the geodesic path performs well on both tasks, and is much better than the ones uncovered by the linear path.
We also demonstrate that our strategy can be scaled up to an arbitrary number of tasks. Fig-\ref{fig:CF_MNISTpermute} C,D, shows 5 permuted-MNIST tasks learnt sequentially without facing CF.
Our results show improved performance over strategies like SGD (appendix) with dropout regularization and SGD with L2 regularization proposed earlier and is comparable to EWC\footnote{Note that EWC uses a large neural network with upto 1000 nodes for their 3 task analysis, while our networks use only 400 for 5 tasks.
}. \cite{kirkpatrick2017overcoming}. We extend our analysis to CNN's\footnote{CNN architecture used is detailed in the supplementary} trained on 2 different datasets (tasks): MNIST and Fashion-MNIST (figure-\ref{fig:CF_MNISTpermute}E,F).
\subsection{Achieving mode connectivity via geodesics}
We apply our geodesics framework to discover high-performance paths for connecting different instances of trained DNN's (modes). As the loss landscape of DNNs are non-convex, can rely on millions of parameters and are studded with a large number of local optima and saddle points \cite{auer1996exponentially, dauphin2014identifying}, it makes the search for discovering high performance paths between the two modes of DNN's challenging. It is also observed that the linear path connecting the modes incurs high loss implying the presence of isolated local optima in the loss landscape \cite{keskar2016large, goodfellow2014qualitatively, garipov2018loss}.
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{Figures/modeConnect3.png}
\caption{\textbf{Discovering high performance paths connecting network modes: } (A) The two modes of VGG16 trained on CIFAR-10 perform at 94.2$\%$ and 95.2$\%$ respectively. The linear segment (blue) connecting the modes incurs very high loss ($\sim$10$\%$) while the geodesic procedure (orange) finds a curved path with test accuracy $\geq$88$\%$. (B) The two modes of ResNet164 trained on CIFAR-10 perform at 92.44$\%$ and 92.12$\%$ respectively. The linear segment (blue) connecting the modes incurs very high loss ($\sim$20$\%$) while the geodesic procedure (orange) finds a curved path with test accuracy $\geq$85$\%$. (C) The two modes of VGG16 trained on CIFAR-100 perform at accuracy 74.4$\%$ and 75.21$\%$ respectively. The linear segment (blue) connecting the modes incurs very high loss ($\sim$1$\%$) while the geodesic procedure (orange) finds a curved path with test accuracy $\geq$55$\%$. }
\label{fig:modeConnect}
\vspace{-4mm}
\end{figure}
Our experiments are performed on VGG16 trained on CIFAR-10, CIFAR-100 and 164-layer ResNet bottleneck trained on CIFAR-10. These models and datasets were chosen to demonstrate that our algorithm works on a wide range of network architectures and datasets. For each model and dataset chosen, we train two networks with different random initializations to find two mode, corresponding to two optima in the loss landscape ($\mathbf{w_1}$, $\mathbf{w_2}$). Subsequently, Geo($\mathbf{w_1}$, $\mathbf{g}$, $\mathbf{w_2}$) constructs a geodesic starting from the first mode $\mathbf{w_1}$ to the target network (second mode) $\mathbf{w_2}$, while minimizing functional difference of networks along the path on the task (CIFAR-10/100 classification), using metric tensor $\mathbf{g}$ computed on the same task. In figure-\ref{fig:modeConnect} we contrast the high-performance path obtained from the optimization strategy from the linear path that connects the two modes of (A) VGG-16 on CIFAR-10, (B) ResNet-164 on CIFAR-10 and (C) VGG-16 on CIFAR-100.
\begin{algorithm}
\caption{Discovering high performance paths connecting modes of deep networks}
\begin{algorithmic}[1]
\State Train deep network 1 ($\mathbf{w_1}$) \Comment{Training deep network with random seed 1}
\State Train deep network 2 ($\mathbf{w_2}$) \Comment{Training deep network with random seed 2}
\State $\gamma(t) \leftarrow Geo(\mathbf{w_1}, \mathbf{g}, \mathbf{w_2})$ \Comment{Geodesic from $\mathbf{w_1}$ to $\mathbf{w_2}$}
\end{algorithmic}
\end{algorithm}
\section{Discussion}
We have established a mathematical framework to construct global paths for navigating the space of neural networks to discover novel networks that have high task-performance while satisfying additional constraints on task-independent network properties. We have shown our framework in action for solving 3 major problems in ML: (i) Network sparsification, (ii) Mitigation of catastrophic interference when learning sequential tasks and (iii) Finding high-accuracy paths to connect modes of deep networks. With AI being built into many critical applications, the need for real-time processing and continuous learning on personal devices is on the rise. Our algorithm can be used for catering to this need. In addition, we note that local processing on personal devices increases data security for the user as information remains local, without having to be streamed to the cloud. However we acknowledge that increasing accessibility of powerful hand-held AI systems in the society could have a negative impact, especially if its present in the wrong hands. Therefore, we believe widespread adoption of personal learning systems in the society should accompany educational programs for developers and subscribers regarding ethical use of AI.
\printbibliography
\section*{Checklist}
\begin{enumerate}
\item For all authors...
\begin{enumerate}
\item Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
\answerYes{}
\item Did you describe the limitations of your work?
\answerYes{Mentioned in the supplementary}
\item Did you discuss any potential negative societal impacts of your work?
\answerYes{}
\item Have you read the ethics review guidelines and ensured that your paper conforms to them?
\answerYes{}
\end{enumerate}
\item If you are including theoretical results...
\begin{enumerate}
\item Did you state the full set of assumptions of all theoretical results?
\answerNA{}
\item Did you include complete proofs of all theoretical results?
\answerNA{}
\end{enumerate}
\item If you ran experiments...
\begin{enumerate}
\item Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)?
\answerYes{Code attached in supplementary}
\item Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)?
\answerYes{In the supplementary}
\item Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)?
\answerYes{wherever applicable}
\item Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)?
\answerYes{in the SI}
\end{enumerate}
\item If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
\begin{enumerate}
\item If your work uses existing assets, did you cite the creators?
\answerYes{}
\item Did you mention the license of the assets?
\answerNA{}
\item Did you include any new assets either in the supplemental material or as a URL?
\answerNA{}
\item Did you discuss whether and how consent was obtained from people whose data you're using/curating?
\answerNA{all datasets are traditional ML datasets(mnist, cifar, etc)}
\item Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content?
\answerNA{}
\end{enumerate}
\item If you used crowdsourcing or conducted research with human subjects...
\begin{enumerate}
\item Did you include the full text of instructions given to participants and screenshots, if applicable?
\answerNA{}
\item Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable?
\answerNA{}
\item Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation?
\answerNA{}
\end{enumerate}
\end{enumerate}
\end{document}
|
\subsection{Motivation}
Deep learning (DL) provides efficient solutions for a number of problems that were difficult to solve with traditional
computing techniques; \hbox{e.g.}\xspace, automatic driving, speech recognition and machine translation.
They are widely supported by DL libraries, such as TensorFlow~\cite{TensorFlow}, Keras~\cite{Keras}, Torch~\cite{torch}, Cafe~\cite{Cafe} and Theano~\cite{Theano}; these libraries facilitate programmers to develop, train, and run models efficiently.
While DL libraries make it easy to incorporate DL into software systems, engineers still fail in using them correctly because of their unique semantic and data requirements~\cite{wanmachine}.
Tensor shape faults are the main type of DL faults; they can occur in all stages of a DL pipeline~\cite{Islam2019ACS}.
DL programs build models on the basis of computational graphs: each node in a computational graph represents an operation (OP), and a tensor (i.e., an input or output of an operation) corresponds to an edge of the graph. Each tensor has its shape properties, such as the number of dimensions and the size of each dimension. When tensors flow and change in computational graphs, their shapes also flow and change.
An OP manipulates tensors and usually restricts their shapes.
When a restriction condition of OPs is not met, a \emph{tensor shape fault} is produced, which usually leads to a program crash.
For example, a misuse of a DL API in a two-dimensional tensor multiplication may cause a tensor shape fault, because the multiplication requires the size of the second dimension of the first matrix to be equal to the size of the first dimension of the second matrix. Fig. \ref{example} shows a more complex example (\S 3).
A recent study reveals that tensor shape faults are frequent, covering 45\% of failures in TensorFlow programs~\cite{Verma2020ShapeFlowDS}. The high prevalence of tensor shape faults and their serious consequences
inspires research on shape faults.
Previous efforts focus on detecting tensor shape faults.
One solution to this is \emph{static shape fault detection}, which analyzes the source code of DL programs against detection rules.
Ariadne~\cite{Dolby2018AriadneAF} tracks and analyzes tensor usages, checking whether each tensor is provided with the desired shape.
Pythia~\cite{Lagouvardos2020StaticAO} extends Ariadne, which translates Python programs into intermediate representations and designs rules to detect shape faults w.r.t. TensorFlow's OPs.
All the static techniques require detection rules to be well prepared.
Another mainstream is \emph{dynamic shape fault detection}, which analyzes the programs' execution information and detects tensor shape faults.
The only previous work is ShapeFlow~\cite{Verma2020ShapeFlowDS}, an abstract interpreter for TensorFlow programs.
It rewrites TensorFlow's APIs and runs programs, extracting the tensor shapes and then constructing shape computational graphs for tracking shape changes.
Up to now, there exist few researches on automatically repairing these faults.
To detect and repair tensor shape faults efficiently,
a dataset of shape faults should be constructed from real projects, and analyzed to obtain insights into tensor shape faults.
Zhang \hbox{et al.}\xspace~\cite{Zhang2018AnES} build a dataset of buggy TensorFlow programs in their empirical study, while the dataset contains only 14 tensor shape faults collected from StackOverflow and 9 ones from Github. Islam \hbox{et al.}\xspace construct a dataset from StackOverflow~\cite{Islam2019ACS}, containing 32 and 37 tensor shape faults in TensorFlow and Keras programs, respectively. Meanwhile, Islam's dataset does not contain buggy programs and their patches, since it is not designed for program analysis and bug repair.
In this paper, we empirically study crashing tensor shape faults in StackOverflow. We build \SFData, a dataset of crashing tensor shape faults that contains 146 buggy programs and their crash messages, patches and test data.
By analyzing these data, we reveal four types of faults, helping researchers understand these faults and as well engineers detect and repair them.
We have released our whole data suite of \SFData and detailed study results on Github~\footnote{https://github.com/tensfa/tensfa}.
The rest of the paper is organized as follows. Section II provides
a background and related work of tensor shape faults.
Section III gives a motivating example.
Section IV describes the construction process of \SFData.
Section V presents the results of data analysis on \SFData,
and Section VI concludes.
\subsection{Preliminaries}
\begin{figure}[t]
\centering
\includegraphics[width=0.4\textwidth]{faults-eps-converted-to.pdf}
\caption{Relationship between tensor shape faults and crashing faults. Here \ding {172}, \ding {173}, and \ding {174} represent non-crashing tensor shape faults, crashing tensor shape faults, and the other crashing faults, respectively. In our study, we mainly focus on crashing tensor shape faults.}
\label{faults}
\end{figure}
First defines the notions used in this paper.
\begin{definition}
\textit{(Operation)}. An operation $OP$ is a function that accepts a set of parameters, each of which can either be an input tensor or a variable value, and produces a set of output tensors and may as well updates the parameters' values:
$$
(ot_{0}, ot_{1}, \dots, ot_{m}~|~ov_{0}, ov_{1}, \dots, ov_{n}) $$
$$= OP (it_{0}, it_{1}, \dots, it_{p}~|~iv_{0}, iv_{1}, \dots, iv_{q}),$$
where $ot_{0 \leq i \leq m}$ and $it_{0 \leq i \leq n}$ are an output tensor and an input one, respectively; $ov_{0 \leq i \leq p}$ and $iv_{0 \leq i \leq q}$ is a left value and a right value of the operation.
\end{definition}
\smallskip
\begin{definition}
\textit{(Shape Restriction)}. Let $t$ be a tensor and $shape(t)$ be the dimensionality of $t$. A shape restriction is a set of conditions a tensor's shape needs to meet:
$$
restrict(t)=\{cond_1,cond_2, \dots,cond_n\};$$
$OP$'s shape restriction is a set of the restrictions on all of its tensors, i.e.,
$$restrict(OP)=(\cup_{0\leq k \leq m} restrict(ot_k)) $$
$$\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\cup~(\cup_{0\leq k \leq p} restrict(it_k)).$$
\end{definition}
\smallskip
\begin{definition}
\textit{(Tensor Shape Fault)}.
A tensor shape fault in a DL program is a violation of shape restrictions, i.e.,
$$tsf(OP) \implies (\exists t \in \{it_0, \dots, it_p, ot_0, \dots, ot_m\}, $$
$$\exists cond \in restrict(t)|\bullet ~cond(shape(t))= false).$$
\end{definition}
\medskip
As Fig. \ref{faults} shows, a tensor shape fault may or may not cause a crash. On the other hand, a crash may or may not be raised by a tensor shape fault. Correspondingly, we define \textit{crashing tensor shape fault} as follows.
\smallskip
\begin{definition}
\textit{(Crashing Tensor Shape Fault)}. A \emph{crashing fault} is a fault that can cause a program to crash. A \emph{crashing tensor shape fault} is a tensor shape fault that can cause a DL program to crash.
\end{definition}
\subsection{Related Work}
There has been a number of researches~\cite{Zhang2018AnES, Islam2019ACS, Islam2020RepairingDN, HumbatovaJBR0T20, jia2020empirical} on studying DL faults from open source repositories.
Humbatova \hbox{et al.}\xspace~\cite{HumbatovaJBR0T20} introduce a taxonomy of faults in DL systems.
Li \hbox{et al.}\xspace~\cite{jia2020empirical} find that many TensorFlow bugs are in its interfaces.
Zhang \hbox{et al.}\xspace~\cite{Zhang2018AnES} collect data from GitHub and StackOverflow to study faults in TensorFlow programs, examining root causes and symptoms of the faults as well as methods used for detecting and locating faults. Their dataset contains 23 TensorFlow programs with shape faults.
Islam \hbox{et al.}\xspace~\cite{Islam2019ACS} examine programs related to five popular deep learning libraries
to understand types of errors, root causes, impacts, and anti-patterns. They further study fault repairs on GitHub and StackOverflow and summarize the repair patterns and challenges~\cite{Islam2020RepairingDN}. Islam's dataset contains 69 TensorFlow/Keras programs with shape faults.
Compared with the above studies, our empirical study and dataset focus on crashing tensor shape faults, while previous approaches are targeting general DL faults.
Furthermore, \SFData contains the shape faults and their buggy programs, patches and tests, which are critical to reproduce and repair faults. Besides, \SFData is larger than existing datasets.
\subsection{Data collection}
On StackOverflow, we search for posts using keywords ``\texttt{[tensorflow] or [keras] ValueError shape answers:1..}'', where
\begin{itemize}
\item \texttt{TensorFlow} and \texttt{Keras} programs are chosen because they are two popular, and as well closely-related deep learning libraries. They are open-sourced at GitHub, where TensorFlow is of 154K stars and 84.3K forks, and Keras is of 50.9K stars and 18.7K forks, as of Apr. 2021;
\item the keyword \texttt{ValueError} indicates that posts related with ValueError are acquired, since an empirical study reveals that 71.43\% of tensor shape faults in TensorFlow and Keras programs are reported as value errors \cite{Islam2019ACS};
\item and the keyword \texttt{answers:1..} indicates that we search for posts with at least one answer, aiming at improving the quality of the obtained posts.
\end{itemize}
\subsection{Data processing}
We manually check the retrieved posts, comments and answers to determine whether the corresponding programs contain shape faults and eliminate the duplicate ones. Four strategies are then taken
for processing data, allowing a high-quality dataset to be constructed.
\textit{S1: Extracting information from posts.} For each post, we extract the buggy program, the crash message, the patch and the tests if they are available; we also extract the post descriptions, comments, and answers.
\textit{S2: Reusing code in Github.} We search for open source projects on Github to retrieve the code, bug reports, patches, and/or test data if a fault is incompletely described (e.g., the patch is missing).
\textit{S3: Generating random tests.} For the programs without tests, we generate random tests as their inputs; the random strategy is frequently used in answering issues on StackOverflow, such as the post \#58277932.
\textit{S4: Producing crash messages.} For each buggy program, we run it and its patch against the tests, obtaining the crash message and validating the patch.
All of the posts, buggy programs, faults, and patches are carefully checked. In this study, we check the data manually, but discuss and confirm the results by taking a rigorous code/fault inspection process.
\subsection{A summary of the \SFData dataset}
As Table \ref{table2} shows,
\SFData consists of 146 buggy programs with crashing tensor shape faults. Among them, 59 are TensorFlow programs and 87 are Keras ones. Each record has the attributes: \texttt{[ProgramID, crash message, \{fault type\}, postID in StackOverflow, original code, patch, \{test data\}]}; for each tensor shape fault, its program, tests and patch are provided such that the fault can be reproduced.
Note that the average number of faults in each program is 1.08 (158/146), because a buggy program can contain one or more shape faults.
Besides, we provide other crashing faults as negative samples; they are collected from StackOverflow and Zhang's empirical study~\cite{Zhang2018AnES}.
\medskip\noindent\textbf{\SFData \hbox{vs.}\xspace IslamData.} IslamData is a dataset whose buggy programs with tensor shape faults are extracted from Islam \hbox{et al.}\xspace's empirical study~\cite{Islam2019ACS}. It can be used to measure the fault detection techniques. Comparatively, \SFData contains the source code of buggy programs, crash messages, (expected) patches and tests; it is also designed for measuring fault repair techniques, letting the patches be the ground truths of the repair techniques.
\subsection{Shape Fault Analysis}
After an analysis of the \SFData dataset, we categorize all of the crashing tensor shape faults into four types, as Table \ref{table1} shows. Let a computational graph be constructed for deep learning. A shape fault may occur during either its construction or its execution stage.
\begin{itemize}
\item
The \emph{construction-stage faults} can be divided into {PRI} and {ALI}. A PRI fault occurs when the shapes of parameters of a general OP violate the shape restrictions (\hbox{e.g.}\xspace, StackOverflow \#54452974). An ALI fault occurs when the output shape of the previous layer is incompatible with the input shape of the next layer (\hbox{e.g.}\xspace, StackOverflow \#47842931).
\item The \emph{execution-stage faults} can be divided into {FII} and {LOI}. An FII fault refers to an incompatibility between a feature data shape and an input tensor's shape. An LOI fault refers to an incompatibility between a label data shape and an output tensor's shape (\hbox{e.g.}\xspace, StackOverflow \#42821125).
\end{itemize}
Statistically, the execution-stage faults account for 65.82\% of crashing tensor shape faults. The construction-stage faults account for 23.16\% and 50.79\% of crashing tensor shape faults in Keras and TensorFlow programs, respectively. One main reason for this is that Keras's APIs are well encapsulated and easy to use, reducing tensor shape faults during the graphs' construction stage.
\begin{framed}
\noindent\emph{Finding 1}:
Shape faults may occur during the construction and the execution of computational graphs, while the execution-stage faults are more than the construction-stage ones. The better a DL library encapsulates its APIs, the less possible its programs suffer from tensor shape faults during the construction stage.
\end{framed}
\subsection{Analysis of Fault Detection Capability}
We conduct further analysis on \SFData by employing existing fault detection techniques.
We observe that Ariadne~\cite{Dolby2018AriadneAF}, Pythia~\cite{Lagouvardos2020StaticAO} and ShapeFlow~\cite{Verma2020ShapeFlowDS} require a set of well-designed checking rules for detecting tensor shape faults.
However, there are 3016 OPs in TensorFlow and 535 OPs in Keras, and each OP may have trivial restrictions on parameters.
Correspondingly, all of these approaches are less effective, as they define only a small number of restriction rules; restrictions may also change over time, as the DL libraries can update frequently. Furthermore, the rule-based approaches are facing the overfitting problem, leading to a low recall in shape fault detection.
Notably, although restrictions are diverse, the crash messages raised by tensor shape faults are similar to each other.
For example, 91.78\% of crashing tensor shape faults produce shape-related crash messages (i.e., messages containing the word ``shape''). Meanwhile, only 11.80\% of other crashing faults can produce such messages.
It is promising to leverage ML techniques to learn from crash messages for distinguishing tensor shape faults from other crashing faults.
\begin{framed}
\noindent\emph{Finding 2}:
Rule-based approaches suffer from the overfitting problem, making their detection capabilities be constrained by the sufficiency of the rules; existing techniques also face challenges in adapting to the evolution of DL libraries. Correspondingly, detection techniques may benefit from machine learning methods in classifying tensor shape faults and exploring fault detecting rules.
\end{framed}
\section{Introduction}
\input{1_introduction}
\section{Background}
\input{2_background}
\section{A Motivating Example}
\input{3_a_motivating_example}
\section{Dataset}
\input{4_dataset}
\section{Analysis and Results}
\input{5_fault_analysis}
\section{Conclusion}
\input{6_conclusion}
\balance
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\begin{figure}[ht]
\centering
\begin{tabular}{p{3.65cm}<{\centering} p{3.65cm}<{\centering}}
w/o FAA & w/ FAA
\\
\end{tabular}
\centering
\begin{minipage}[h]{0.49\linewidth}
\centering\includegraphics[width=.99\linewidth]{figures/figure_1_a.pdf}
\end{minipage}
\begin{minipage}[h]{0.49\linewidth}
\centering\includegraphics[width=.99\linewidth]{figures/figure_1_a_FAA.pdf}
\end{minipage}
\begin{minipage}[h]{0.49\linewidth}
\centering\includegraphics[width=.99\linewidth]{figures/figure_1_b.pdf}
\end{minipage}
\begin{minipage}[h]{0.49\linewidth}
\centering\includegraphics[width=.99\linewidth]{figures/figure_1_b_FAA.pdf}
\end{minipage}
\begin{minipage}[h]{0.49\linewidth}
\centering\includegraphics[width=.99\linewidth]{figures/figure_1_c.pdf}
\end{minipage}
\begin{minipage}[h]{0.49\linewidth}
\centering\includegraphics[width=.99\linewidth]{figures/figure_1_c_FAA.pdf}
\end{minipage}
\caption{
Our robust domain adaptation alleviates overfitting effectively: Both supervised learning with source data (row 1) and unsupervised learning with target data (in rows 2 and 3 for adversarial learning and self-training) in unsupervised domain adaptation suffer from clear overfitting as illustrated by decreased training losses (blue curves) vs increased target test losses (red curves) in column 1. Our Fourier adversarial attacking (FAA) generates novel adversarial samples, which regulates the minimization of training losses and alleviates overfitting effectively with decreased target test loss as shown in column 2 (Best viewed in color).}
\label{fig:intro}
\end{figure}
Deep convolutional neural networks (CNNs)~\cite{krizhevsky2012imagenet,simonyan2014very,he2016deep} have defined new state of the arts in various computer vision tasks~\cite{chen2017deeplab,long2015fully,ren2015faster,he2017mask,krizhevsky2012imagenet,simonyan2014very,he2016deep}, but their trained models often over-fit to the training data and experience clear performance drops for data from different sources due to the existence of \textit{domain gaps}. Unsupervised domain adaptation (UDA) has been investigated to address the \textit{domain gaps} by leveraging unlabeled target data. To this end, most existing UDA works~\cite{kang2019contrastive,kang2018deep,tsai2018learning,tzeng2017adversarial, luo2019taking, vu2019advent, Chen_2018_CVPR,chen2018domain} involve supervised losses on source data and unsupervised losses on target data for learning a model that performs well in target domains. However, as illustrated in Fig.~\ref{fig:intro}, these methods often face more severe overfitting (as compared with the classical supervised learning) as supervised source losses in UDA has an extra \textit{domain gap} (for test data in target domains) and unsupervised target losses are often noisy due to the lack of annotations.
Overfitting exists in almost all deep network training, which is undesired and often degrades the generalization of the trained deep network models while applied to new data. One way of identifying whether overfitting is happening is to check whether the generalization gap, \ie, the difference between the test loss and the training loss, is increasing or not~\cite{goodfellow2016deep} as shown in Fig.~\ref{fig:intro}. Various strategies have been investigated to alleviate overfitting through weight regularization~\cite{hanson1988comparing}, dropout~\cite{srivastava2014dropout}, mixup~\cite{zhang2017mixup}, label smoothing~\cite{szegedy2016rethinking}, batch normalization~\cite{ioffe2015batch}, etc. However, all these strategies were designed for supervised and semi-supervised learning where training data and test data usually have very similar distributions. For domain adaptive learning, they do not fit in well due to the negligence of domain gaps that widely exist between data of different domains.
We design a robust domain adaptation technique that introduces a novel Fourier adversarial attacking (FAA) technique to mitigate the overfitting in unsupervised domain adaptation. FAA mitigates overfitting by generating adversarial samples that prevent over-minimization of supervised and unsupervised UDA losses as illustrated in Fig.~\ref{fig:intro}. Specifically, FAA decomposes training images into multiple frequency components (FCs) and only perturbs FCs that capture little semantic information. Unlike traditional attacking that restricts the magnitude of perturbation noises to keep image semantics intact, FAA allows large magnitude of perturbation in its generated adversarial samples but has minimal modification of image semantics. This feature is critical to unsupervised domain adaptation which usually involves clear domain gaps and so requires adversarial sample with large perturbations. By introducing the FAA-generated adversarial samples in training, networks can continue the ``random walk" and avoid over-fitting and drift into an area with a flat loss landscape~\cite{chaudhari2019entropy,keskar2016large,li2017visualizing}, leading to more robust domain adaptation.
The contributions of this work can be summarized in three aspects. \textit{First}, we identify the overfitting issue in unsupervised domain adaptation and introduce adversarial attacking to mitigate overfitting by preventing training objectives from over-minimization. \textit{Second}, we design an innovative Fourier Adversarial Attacking (FAA) technique to generate novel adversarial samples to mitigate overfitting in domain adaptation. FAA is generic which can work for both supervised source loss and unsupervised target losses effectively. \textit{Third}, we conducted extensive experiments over multiple computer vision tasks in semantic segmentation, object detection and image classification. All experiments show that our method mitigates overfitting and improve domain adaption consistently.
\section{Related Works}
\textbf{Domain Adaptation:} Domain adaptation has been studied extensively for mitigating data annotation constraints. Most existing works can be broadly classified into three categories. The first category is \textit{adversarial training} based which employs a discriminator to align source and target domains in the feature, output or latent space \cite{long2016unsupervised,tzeng2017adversarial,luo2019taking,tsai2018learning,saito2018maximum,huang2021mlan,vu2019advent,tsai2019domain,huang2021category,yang2020part,zhang2021detr,guan2021uncertainty,huang2021semi}. The second category is \textit{image translation} based which adapt image appearance to mitigate domain gaps \cite{li2019bidirectional,zhan2019ga,yang2020fda,zhang2021spectral}. The third category is \textit{self-training} based which predict pseudo labels or minimize entropy to guide iterative learning from target samples \cite{zou2018unsupervised, zou2019confidence,luo2021unsupervised,guan2021scale,huang2021cross,guan2021domain}.
Domain adaptation involves two typical training losses, namely, supervised loss over labeled source data and unsupervised loss over unlabeled target data. State-of-the-art methods tend to over-minimize the two types of losses which directly leads to deviated models with suboptimal adaptation as illustrated in Fig.~\ref{fig:intro}. We design a robust domain adaptation technique that addresses this issue by preventing loss over-minimization.
\textbf{Overfitting in Network Training:} Overfitting is a common phenomenon in deep network training which has been widely studied in the deep learning and computer vision community~\cite{belkin2018overfitting,caruana2001overfitting,ng1997preventing,roelofs2019meta,werpachowski2019detecting,wang2021embracing}. Most existing works address overfitting by certain regularization strategies such as weight decay~\cite{hanson1988comparing}, dropout~\cite{srivastava2014dropout}, $l_{1}$ regularization~\cite{tibshirani1996regression}, mix-up~\cite{zhang2017mixup}, label smoothing~\cite{szegedy2016rethinking}, batch normalization~\cite{ioffe2015batch}, virtual adversarial training~\cite{miyato2018virtual}, flooding~\cite{ishida2020we}, etc. However, these strategies were mostly designed for supervised or semi-supervised learning which do not fit in well for domain adaptive learning that usually involves domain gaps and unsupervised losses. We design a adversarial attacking technique that mitigates the overfitting in domain adaptive learning effectively.
\textbf{Adversarial Attacking:} Adversarial attacking has been studied in various security problems. For example, \cite{szegedy2013intriguing} shows that adversarial samples
can easily confuse CNN models. The following works improved adversarial attacking from different aspects via fast gradient signs \cite{goodfellow2014explaining}, minimal adversarial perturbation \cite{moosavi2016deepfool}, universal adversarial perturbations \cite{moosavi2017universal}, gradient-free attacking \cite{baluja2017adversarial}, transferable adversarial sample generation\cite{liu2016delving}, etc. Adversarial attacking has also been applied to other tasks, e.g., \cite{zheng2016improving,miyato2018virtual} employed adversarial samples to mitigate over-fitting in supervised and semi-supervised learning, \cite{volpi2018generalizing} generated adversarial samples for data augmentation, and \cite{liu2019transferable, yang2020adversarial} augments transferable features for domain divergence minimization.
Most existing adversarial attacking methods commonly constrain the magnitude of perturbation noises for minimal modification of image semantics. However, such generated adversarial samples cannot tackle overfitting in domain adaptive learning well which usually involves a domain gap of fair magnitude. We design an innovative Fourier adversarial attacking technique that allows to generate adversarial samples without magnitude constraint yet with minimal modification of image semantics, more details to be described in the ensuing subsections.
\section{Method}
We achieve robust domain adaptation via Fourier adversarial attacking as illustrated in Fig.~\ref{fig:stru}. The training consists of two phases, namely, an \textit{Attacking Phase} and a \textit{Defending Phase}. Given a training image, the attacking phase learns to identify the right FCs with limited semantic information that allow perturbation noises of large magnitude. It also learns to generate adversarial samples (with perturbable FCs) with minimal modification of image semantics. During the defending phase, the generated adversarial sample is applied to mitigate overfitting by preventing over-minimization of training losses, more details to be described in the ensuing subsections.
\subsection{Task Definition}
We focus on the problem of unsupervised domain adaptation (UDA). Given labeled source data \{$X_{s}$, $Y_{s}$\} and unlabeled target data $X_{t}$, our goal is to learn a task model $F$ that performs well on $X_{t}$. The \textit{baseline} model is trained with the labeled source data only:
\begin{equation}
\begin{split}
\mathcal{L}_{sup}(X_{s}, Y_{s}; F) = l(F(X_{s}), Y_{s}),
\end{split}
\label{eq_baseline}
\end{equation}
where $l(\cdot)$ denotes an accuracy-related loss, \eg, the standard cross-entropy loss.
\begin{figure*}[ht]
\centering
\subfigure{\includegraphics[width=0.98\linewidth]{figures/figure_3.pdf}}
\caption{
Overview of our proposed robust domain adaptaion (RDA) via Fourier Adversarial Attacking (FAA). RDA has a bidirectional training process which consists of an \textit{Attacking Phase} and a \textit{Defending Phase}. Given a training image, the attacking phase learns to generate adversarial samples with minimal modification of image semantics. During the defending phase, the generated adversarial samples are applied to mitigate overfitting by preventing over-minimization of training losses. Given an input image, FAA decomposes it into multiple FCs representation, identifies perturbable FCs, and then perturbs them through weighted sum with the corresponding FCs from a randomly picked reference image. The objective of FFA is to maximize task losses while ensuring minimal modification of image semantics.
}
\label{fig:stru}
\end{figure*}
\subsection{Fourier Adversarial Attacking}
Our proposed Fourier adversarial attacking (FAA) generates adversarial samples to attack the training loss to mitigate overfitting in domain adaptation, as shown in Fig.~\ref{fig:stru}. In adversarial sample generation, it first employs Fourier transformation to decompose input images into multiple frequency components (FCs) and then inject perturbation to non-semantic FCs which allows perturbation of large magnitude but with minimal modification of image semantics.
\textbf{Fourier Decomposition:} Inspired by JPEG \cite{wallace1992jpeg,pennebaker1992jpeg} and frequency-domain learning \cite{huang2021fsdr,jiang2020focal}, we transform an image $x$ into frequency space and decompose it into multiple FCs which allows explicit manipulation of each FC and more controllable perturbations. We employ Fourier transformation to convert $x$ into frequency space and further decompose it into multiple FCs of equivalent bandwidth:
\begin{equation}
\begin{split}
z = \mathcal{F}(x),\nonumber\\
z^{fc} = \mathcal{D}(z;N),\nonumber
\end{split}
\end{equation}
where $\mathcal{F}(\cdot)$ stands for Fourier transformation \cite{bracewell1986fourier}; $z$ denotes the frequency-space representation of $x$; $\mathcal{D}(z;N)$ denotes a function that decomposes $z$ into $N$ FCs $z^{fc} = \{z^{1}, z^{2}, ... , z^{N-1}, z^{N}\}$ of equivalent bandwidth.
We consider a gray-scale image $x \in \mathbb{R}^{H \times H}$ for defining $\mathcal{D}(z;N)$ formally. We thus have $z \in \mathbb{C}^{H \times H}$, where $\mathbb{C}$ denotes complex numbers. We use $z(i,j)$ to index $z$ at $(i,j)$, and $c_{i}$ and $c_{j}$ to denote the image centroid (\ie, image center at $(H/2, H/2)$). The equation $\{z^{1}, z^{2}, ... , z^{N-1}, z^{N}\}= \mathcal{D}(z;N)$ can be formally defined by:
\begin{equation}
z^{n}(i, j) =
\left\{
\begin{array}{ll}
z(i, j), &\text{if} \ \frac{n-1}{N} < d((i,j),(c_{i},c_{j})) < \frac{n}{N},\\
0, &\text{otherwise},\\
\end{array}
\right.\nonumber
\end{equation}
where $d(\cdot,\cdot)$ denotes Euclidean distance, $N$ denotes how many FCs the input is supposed to be decomposed into and $n$ denotes the FC index.
To get $z^{fc} = \{z^{1}, z^{2}, ... , z^{N-1}, z^{N}\}= \mathcal{D}(z;N)$, we perform the equation $N$ times by changing $n$ from $1$ to $N$. If $x$ does not have a square size, we first up-sample the short side to be same as the long side before this process and down-sample it back to the original size after processing. If $x$ has more than one channel ($e.g.$, RGB image $x \subset \mathbb{R}^{n \times n \times 3}$), we process each channel independently.
\textbf{Adversarial Attacking:} With the decomposed FCs $z^{fc} = \{z^{1}, z^{2}, ... , z^{N-1}, z^{N}\}$, we attack domain adaption losses by perturbing partial FCs without magnitude constraint. Specifically, we employ a learnable gate module \cite{jang2016categorical,tucker2017rebar, xu2020learning,maddison2016concrete} to select certain FCs for perturbation. This gate $G$ assigns a binary score to each FC, where `1' indicates this FC is selected for perturbation (multiplied by `1') while `0' shows this FC is discarded (reset to all zero values). $G$ works with Gumbel-Softmax, a differentiable sampling mechanism for categorical variables that can be trained via standard back-propagation. Please refer to \cite{jang2016categorical, tucker2017rebar, xu2020learning, maddison2016concrete} for details.
Given a reference image (an randomly selected target-domain image) in FC representation $z^{fc}_{ref} = \{z^{1}_{ref}, z^{2}_{ref}, ... , z^{N-1}_{ref}, z^{N}_{ref}\}= \mathcal{D}(\mathcal{F}(x_{ref});N)$, we employ this gate module to block some FCs (all reset to zero), and apply the selected FCs to perturb the corresponding FCs of the input image $x$:
\begin{equation}
\hat{z}^{fc} = (1 - G(z^{fc}))z^{fc} + G(z^{fc})z^{fc}_{ref},
\label{eq_gating}
\end{equation}
where the learnable gate module $G$ enables binary channel attention that selects which FCs to perturb.
Specifically, for the identified FCs of the input image $x$, we extract the corresponding FCs of a reference image $X_{ref}$ and use them as perturbation noises. We generate adversarial samples in this manner because the perturbation noises (\ie, non-semantic FCs of $z^{fc}_{ref}$) from target natural images are more reasonable and meaningful as compared to a random noises/signals as in many existing adversarial sample generation methods. In addition, the use of non-semantic FCs of target samples mitigates inter-domain gaps which helps to improve target-domain performance in domain adaptation.
With the perturbed FCs $\hat{z}^{fc}$, we convert them all back (via inverse Fourier transformation) to get the adversarial sample $x^{FAA}$:
\begin{equation}
x^{FAA} = \mathcal{F}^{-1}( \mathcal{C}(\hat{z}^{fc})),
\end{equation}
where $\mathcal{C}(\cdot)$ denotes an inverse process of $\mathcal{D}(\cdot;N)$ that `compose’ the decomposed $z^{fc}$ back to the full representation by summing all the elements over frequency channels.
The \textit{Fourier Decomposition} and \textit{Adversarial Attacking} can be combined to form the FAA as follows:
\begin{equation}
x^{FAA} = \mathcal{A}(x),
\label{eq_attacker}
\end{equation}
where $\mathcal{A}$ takes an image $x$ and outputs an adversarial sample $x^{FAA}$ via the FAA as described. $\mathcal{A}$ has two sub-modules with learnable parameters, $i.e.$, the gate module $G$ and the single-layer neural network for spatial weighting map $\mathcal{S}$ generation. The rest operations in $\mathcal{A}$ are deterministic such as Fourier transformation and its inverse ($\mathcal{F}(\cdot)$ and $\mathcal{F}^{-1}(\cdot)$), decomposition and re-composition ($\mathcal{D}(\cdot;N)$ and $ \mathcal{C}(\cdot)$)
\subsection{FAA Training}
The proposed FAA involves three types of losses including task loss (\ie, UDA training loss in this work) that is to be attacked, the gate related loss that constrains the proportion of perturbable FCs, and the reconstruction loss that aims to minimize the attacking effects over image semantics.
Given input images $X \subset \mathbb{R}^{H \times W \times 3}$, a task model $F$ and the Fourier attacker $\mathcal{A}$, we denote the task loss by $\mathcal{L}(X;\mathcal{A},F)$. Note the task loss refers to UDA training loss in this work but it can be other types of losses such as supervised or unsupervised training losses. The gate related loss and the reconstruction loss losses are defined by:
\begin{equation}
\begin{split}
\mathcal{L}_{gat}(X;\mathcal{A}) =
\left\{
\begin{array}{ll}
\sum(G(z^{fc})), &\text{if} \ \sum(G(z^{fc}))>N*p,\\
0, &\text{otherwise},\\
\end{array}
\right.\\\nonumber
\mathcal{L}_{rec}(X;\mathcal{A}) = |R(X)- R(\mathcal{A}(X))|,\nonumber
\end{split}
\label{eq_gate_recon}
\end{equation}
where $G(z^{fc})$ is a gating process in $\mathcal{A}$ as described in Eq.~\ref{eq_gating}. $R$ is a band-pass filter to get the mid-frequency content \cite{huang2021fsdr} that captures semantic information ($e.g.$, structures and shapes) and thus the consistency loss can ensure that the selected FCs contains minimal semantic information. $p$ is a hyper-parameter that constrains the maximum number of perturbable FCs by $N*p$.
The overall training objective of FAA is formulated by:
\begin{equation}
\max_{\mathcal{A}} \mathcal{L}(X;\mathcal{A},F) - \mathcal{L}_{gat}(X;\mathcal{A}) - \mathcal{L}_{rec}(X;\mathcal{A})
\label{eq_attacking}
\end{equation}
\subsection{Robust Domain Adaptation}
Sections 3.2 and 3.3 describe the ``Attacking Phase" that generates adversarial samples via FAA.
This section presents the proposed robust domain adaptation technique where FAA is used to mitigate the overfitting in UDA in a ``Defending Phase”. Specifically, we apply the FAA-generated adversarial samples to prevent the over-minimization of UDA training losses in domain adaptation.
Given the task model $F$ and the Fourier attacker $\mathcal{A}$, the training objective of the task model $F$ is formulated by:
\begin{equation}
\min_{F} \mathcal{L}(X;\mathcal{A},F)
\label{eq_defending}
\end{equation}
The optimization functions of both attacking (\ie, Eq.~\ref{eq_attacking}) and defending (\ie, Eq.~\ref{eq_defending}) are generic and applicable to various \textit{tasks} and \textit{data}. Specifically, the model $F$ could be for semantic segmentation, object detection or image classification task. The data could be labeled source data with supervised source losses or unlabeled target data with unsupervised target losses such as adversarial loss~\cite{tsai2018learning, luo2019taking, vu2019advent, yang2020adversarial, Pan_2020_CVPR, liu2019transferable, huang2020contextual, saito2019strong}, self-training loss~\cite{zou2018unsupervised, zou2019confidence, li2019bidirectional, wang2020differential, kim2020learning,yang2020fda} or entropy loss~\cite{vu2019advent, grandvalet2005semi}, etc. Below are a supervised loss and a self-training-based unsupervised loss here for reference:
\begin{equation}
\begin{split}
\mathcal{L}_{s}^{sup}(X_{s};\mathcal{A}, F) = l(F(\mathcal{A}(X_{s})), Y_{s}), \\\nonumber
\mathcal{L}_{t}^{un}(X_{t};\mathcal{A}, F) = l(F(\mathcal{A}(X_{t})), \hat{Y}_{t}), \nonumber
\end{split}
\label{eq_gate_recon}
\end{equation}
where $l(\cdot)$ denotes an accuracy-related loss, \eg, cross entropy loss; $\mathcal{A}(X_{s})$ and $\mathcal{A}(X_{t})$ are inputs perturbed by FAA, and $\hat{Y}_{t}$ is the pseudo label of unlabeled target data.
\begin{algorithm}[t]
\caption{The proposed Robust Domain Adaptation via Fourier adversarial attacking (FAA).
}\label{algorithm_FAA_RDA}
\begin{algorithmic}[1]
\REQUIRE Input data $X$; task model $F$;
Fourier attacker $\mathcal{A}$
\ENSURE Learnt parameters $\theta$ of task model $F$
\FOR{$iter = 1$ \textbf{to} $Max\_Iter$}
\STATE Sample a batch of $X$ uniformly from the dataset
\STATE \textbf{\emph{Defending phase (minimize loss):}}
\STATE Update $F$ via Eq.~\ref{eq_defending}
\STATE \textbf{\emph{Attacking phase (maximize loss):}}
\STATE Update $\mathcal{A}$ via Eq.~\ref{eq_attacking}
\ENDFOR
\RETURN $F$
\end{algorithmic}
\end{algorithm}
In summary, our robust domain adaptation has a bidirectional training framework including an \textit{Attacking Phase} and a \textit{Defending Phase} as shown in Fig.~\ref{fig:stru}. In the \textit{Attacking Phase}, the task model $F$ is fixed and the attacker $\mathcal{A}$ generates adversarial samples to increase the training loss. In the \textit{Defending Phase}, $\mathcal{A}$ is fixed and $F$ is updated to reduce the training loss. These two phases are conducted in an alternative way way and form a bidirectional training framework (adversarial training). Please refer to Algorithm~\ref{algorithm_FAA_RDA} for details.
\begin{table*}[!t]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{c|ccccccccccccccccccc|c}
\hline
Method & Road & SW & Build & Wall & Fence & Pole & TL & TS & Veg. & Terrain & Sky & PR & Rider & Car & Truck & Bus & Train & Motor & Bike & mIoU\\
\hline
Baseline~\cite{he2016deep} &75.8 &16.8 &77.2 &12.5 &21.0 &25.5 &30.1 &20.1 &81.3 &24.6 &70.3 &53.8 &26.4 &49.9 &17.2 &25.9 &6.5 &25.3 &36.0 &36.6\\
\textbf{+FAA-S} &89.8 &39.1 &81.7 &27.6 &19.9 &34.2 &35.9 &23.3 &82.1 &29.5 &76.6 &58.3 &26.0 &82.1 &32.5 &45.2 &15.3 &26.9 &33.5 &\textbf{45.2}\\\hline
AdaptSeg~\cite{tsai2018learning} &86.5 &36.0 &79.9 &23.4 &23.3 &23.9 &35.2 &14.8 &83.4 &33.3 &75.6 &58.5 &27.6 &73.7 &32.5 &35.4 &3.9 &30.1 &28.1 &42.4\\
\textbf{+FAA-S} &89.1 &38.7 &82.0 &28.5 &23.6 &33.2 &34.9 &22.2 &83.7 &34.0 &77.5 &59.0 &26.6 &83.5 &33.8 &45.8 &15.8 &28.3 &33.9 &46.0\\
\textbf{+FAA-T} &91.0 &44.3 &82.3 &28.1 &21.7 &35.6 &36.8 &18.7 &84.2 &38.3 &73.7 &61.4 &28.5 &85.9 &38.6 &45.7 &14.9 &32.4 &33.3 &47.1\\
\textbf{+FAA} &91.2 &44.7 &82.5 &27.7 &24.7 &36.7 &36.5 &24.7 &84.9 &38.1 &78.4 &62.2 &27.4 &84.9 &39.3 &46.7 &12.9 &31.9 &36.7 &\textbf{48.0}\\\hline
ST~\cite{zou2018unsupervised} &88.5 &26.3 &82.0 &24.5 &19.9 &33.6 &37.6 &19.8 &82.7 &26.2 &76.8 &60.0 &26.6 &82.9 &28.9 &31.0 &6.3 &24.5 &31.2 &42.6\\
\textbf{+FAA-S} &91.4 &52.1 &81.6 &32.6 &26.4 &35.3 &36.2 &22.5 &84.6 &35.8 &77.3 &60.5 &27.9 &84.1 &37.0 &43.8 &17.2 &25.6 &35.9 &47.8\\
\textbf{+FAA-T} &91.6 &49.0 &83.2 &32.7 &26.5 &38.9 &46.5 &23.6 &83.2 &35.3 &77.9 &60.1 &28.5 &85.5 &38.1 &41.4 &16.9 &25.6 &36.4 &48.5\\
\textbf{+FAA} &92.0 &52.6 &82.5 &33.9 &28.5 &39.4 &45.7 &33.6 &84.3 &38.4 &80.5 &60.4 &29.1 &85.4 &39.9 &44.4 &16.2 &29.6 &35.1 &\textbf{50.1}\\
\hline
CLAN~\cite{luo2019taking} &87.0 &27.1 &79.6 &27.3 &23.3 &28.3 &35.5 &24.2 &83.6 &27.4 &74.2 &58.6 &28.0 &76.2 &33.1 &36.7 &{6.7} &{31.9} &31.4 &43.2\\
\textbf{+FAA} &92.0 &53.4 &82.8 &32.5 &29.8 &36.8 &36.6 &28.5 &83.3 &35.8 &77.1 &61.2 &30.1 &83.8 &36.8 &46.2 &11.7 &28.6 &35.6 &\textbf{48.6}\\\hline
AdvEnt~\cite{vu2019advent} &89.4 &33.1 &81.0 &26.6 &26.8 &27.2 &33.5 &24.7 &{83.9} &{36.7} &78.8 &58.7 &30.5 &{84.8} &38.5 &44.5 &1.7 &31.6 &32.4 &45.5\\
\textbf{+FAA} &92.1 &49.0 &82.5 &31.2 &28.6 &38.3 &37.1 &27.1 &84.8 &39.4 &79.1 &61.3 &29.1 &85.0 &38.2 &46.6 &14.3 &32.4 &36.9 &\textbf{49.1}\\\hline
IDA~\cite{pan2020unsupervised} &90.6 &37.1 &82.6 &30.1 &19.1 &29.5 &32.4 &20.6 &85.7 &40.5 &79.7 &58.7 &31.1 &86.3 &31.5 &48.3 &0.0 &30.2 &35.8 &46.3\\
\textbf{+FAA} &91.4 &50.3 &83.4 &33.0 &27.2 &37.7 &38.4 &27.0 &84.4 &41.8 &79.9 &59.7 &30.6 &85.1 &37.7 &47.5 &14.2 &31.2 &35.3 &\textbf{49.3}\\\hline
CRST~\cite{zou2019confidence} &91.0 &55.4 &80.0 &33.7 &21.4 &37.3 &32.9 &24.5 &85.0 &34.1 &80.8 &57.7 &24.6 &84.1 &27.8 &30.1 &26.9 &26.0 &42.3 &47.1\\
\textbf{+FAA} &92.7 &56.4 &84.4 &34.3 &29.2 &38.2 &48.8 &47.1 &85.5 &42.4 &86.0 &62.3 &32.9 &85.3 &40.8 &49.8 &22.7 &23.0 &36.8 &\textbf{52.6}\\\hline
BDL~\cite{li2019bidirectional} &91.0 &44.7 &84.2 &34.6 &27.6 &30.2 &36.0 &36.0 &85.0 &43.6 &83.0 &58.6 &31.6 &83.3 &35.3 &49.7 &3.3 &28.8 &35.6 &48.5\\
\textbf{+FAA} &92.8 &54.1 &85.3 &33.8 &28.1 &41.0 &46.1 &47.6 &84.8 &42.7 &82.8 &63.5 &32.5 &84.1 &37.0 &50.3 &15.5 &23.2 &36.5 &\textbf{51.7}\\\hline
CrCDA~\cite{huang2020contextual} &92.4 &55.3 &82.3 &31.2 &29.1 &32.5 &33.2 &35.6 &83.5 &34.8 &84.2 &58.9 &32.2 &84.7 &40.6 &46.1 &2.1 &31.1 &32.7 &48.6 \\
\textbf{+FAA} &92.9 &55.2 &83.4 &31.8 &31.1 &39.4 &47.8 &46.4 &83.4 &38.4 &85.7 &61.2 &31.0 &85.7 &39.1 &46.3 &12.0 &33.4 &38.9 &\textbf{51.7}\\\hline
SIM~\cite{wang2020differential} &90.6 &44.7 &84.8 &34.3 &28.7 &31.6 &35.0 &37.6 &84.7 &43.3 &85.3 &57.0 &31.5 &83.8 &42.6 &48.5 &1.9 &30.4 &39.0 &49.2\\
\textbf{+FAA} &92.2 &53.7 &84.6 &34.2 &29.1 &38.0 &47.0 &45.3 &85.0 &43.8 &84.9 &60.3 &29.2 &85.2 &39.5 &47.3 &12.4 &32.3 &44.3 &\textbf{52.0}\\\hline
TIR~\cite{kim2020learning} &92.9 &55.0 &85.3 &34.2 &31.1 &34.9 &40.7 &34.0 &85.2 &40.1 &87.1 &61.0 &31.1 &82.5 &32.3 &42.9 &0.3 &36.4 &46.1 &50.2\\
\textbf{+FAA} &93.0 &55.5 &84.0 &33.0 &28.2 &38.1 &46.6 &45.8 &84.9 &41.7 &86.1 &61.2 &33.7 &84.2 &38.3 &46.5 &15.3 &34.9 &44.6 &\textbf{52.4}\\\hline
FDA~\cite{yang2020fda} &92.5 &53.3 &82.4 &26.5 &27.6 &36.4 &40.6 &38.9 &82.3 &39.8 &78.0 &62.6 &34.4 &84.9 &34.1 &53.1 &16.9 &27.7 &46.4 &50.5\\
\textbf{+FAA} &93.1 &55.0 &84.7 &33.1 &29.5 &38.7 &49.3 &44.9 &84.8 &41.6 &80.2 &62.3 &33.2 &85.6 &37.3 &51.3 &18.5 &34.6 &45.3 &\textbf{52.8}\\\hline
\end{tabular}
}
\caption{
Domain adaptive semantic segmentation experiments over task GTA5 $\rightarrow$ Cityscapes (\textit{+FAA-S}, \textit{+FAA-T}, and \textit{+FAA} mean to apply our FAA to attack supervised source losses, unsupervised target losses, and both types of losses, respectively).}
\label{table:gta2city}
\end{table*}
\begin{table*}[h]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{c|cccccccccccccccc|c|c}
\hline
Method & Road & SW & Build & Wall\textsuperscript{*} & Fence\textsuperscript{*} & Pole\textsuperscript{*} & TL & TS & Veg. & Sky & PR & Rider & Car & Bus & Motor & Bike & mIoU & mIoU\textsuperscript{*}\\
\hline
Baseline~\cite{he2016deep} &55.6 &23.8 &74.6 &9.2 &0.2 &24.4 &6.1 &12.1 &74.8 &79.0 &55.3 &19.1 &39.6 &23.3 &13.7 &25.0 &33.5 &38.6\\
PatAlign~\cite{tsai2019domain} &82.4 &38.0 &78.6 &8.7 &0.6 &26.0 &3.9 &11.1 &75.5 &84.6 &53.5 &21.6 &71.4 &32.6 &19.3 &31.7 &40.0 &46.5\\
AdaptSeg~\cite{tsai2018learning} &84.3 &42.7 &77.5 &- &- &- &4.7 &7.0 &77.9 &82.5 &54.3 &21.0 &72.3 &32.2 &18.9 &32.3 &- &46.7\\
CLAN~\cite{luo2019taking} &81.3 &37.0 &{80.1} &- &- &- &{16.1} &{13.7} &78.2 &81.5 &53.4 &21.2 &73.0 &32.9 &{22.6} &30.7 &- &47.8\\
AdvEnt~\cite{vu2019advent} &85.6 &42.2 &79.7 &{8.7} &0.4 &25.9 &5.4 &8.1 &{80.4} &84.1 &{57.9} &23.8 &73.3 &36.4 &14.2 &{33.0} &41.2 &48.0\\
IDA~\cite{pan2020unsupervised} &84.3 &37.7 &79.5 &5.3 &0.4 &24.9 &9.2 &8.4 &80.0 &84.1 &57.2 &23.0 &78.0 &38.1 &20.3 &36.5 &41.7 &48.9\\
CrCDA~\cite{huang2020contextual} &86.2 &44.9 &79.5 &8.3 &0.7 &27.8 &9.4 &11.8 &78.6 &86.5 &57.2 &26.1 &76.8 &39.9 &21.5 &32.1 &42.9 &50.0\\\hline
TIR~\cite{kim2020learning} &92.6 &53.2 &79.2 &- &- &- &1.6 &7.5 &78.6 &84.4 &52.6 &20.0 &82.1 &34.8 &14.6 &39.4 &- &49.3\\
\textbf{+FAA} &89.1 &51.3 &78.7 &8.9 &0.1 &27.1 &19.1 &17.0 &82.4 &83.9 &55.8 &27.7 &84.5 &39.6 &26.9 &38.3 &45.7 &\textbf{53.4}\\\hline
CRST~\cite{zou2019confidence} &67.7 &32.2 &73.9 &10.7 &1.6 &37.4 &22.2 &31.2 &80.8 &80.5 &60.8 &29.1 &82.8 &25.0 &19.4 &45.3 &43.8 &50.1\\
\textbf{+FAA} &89.3 &49.6 &79.3 &12.9 &0.1 &37.2 &20.6 &29.8 &83.7 &84.0 &62.6 &28.0 &84.9 &37.1 &30.1 &48.2 &48.6 &\textbf{55.9}\\\hline
BDL~\cite{li2019bidirectional} &86.0 &46.7 &80.3&-&-&- &14.1 &11.6 &79.2 &81.3 &54.1 &27.9 &73.7 &42.2 &25.7 &45.3 &- &51.4\\
\textbf{+FAA} &89.7 &47.5 &82.7 &6.2 &0.1 &31.8 &25.0 &22.7 &82.3 &84.8 &59.0 &26.2 &83.4 &39.2 &31.5 &37.9 &46.9 &\textbf{54.8}\\\hline
SIM~\cite{wang2020differential} &83.0 &44.0 &80.3 &- &- &- &17.1 &15.8 &80.5 &81.8 &59.9 &33.1 &70.2 &37.3 &28.5 &45.8 &- &52.1\\
\textbf{+FAA} &88.3 &47.8 &82.8 &9.3 &1.1 &36.7 &24.5 &26.8 &79.4 &83.7 &62.5 &32.1 &81.7 &36.3 &29.2 &46.4 &48.0 &\textbf{55.5}\\\hline
FDA~\cite{yang2020fda} &79.3 &35.0 &73.2 &- &- &- &19.9 &24.0 &61.7 &82.6 &61.4 &31.1 &83.9 &40.8 &38.4 &51.1 &- &52.5\\
\textbf{+FAA} &89.4 &39.0 &79.8 &8.1 &1.2 &33.0 &22.6 &28.1 &81.8 &82.0 &60.9 &30.1 &82.7 &41.4 &37.5 &48.9 &47.9 &\textbf{55.7}\\\hline
\end{tabular}
}
\caption{
Domain adaptive semantic segmentation experiments over task SYNTHIA $\rightarrow$ Cityscapes (\textit{+FAA} means to include FAA to attack supervised source loss and unsupervised target loss in domain adaptation).}
\label{table:synthia2city}
\end{table*}
This bidirectional training framework uses FAA to prevent over-minimization of UDA losses by forcing them oscillating around a small value. In another word, it ensures that the task model $F$ can continue the “random walk” and drift into an area with a flat loss landscape~\cite{chaudhari2019entropy,keskar2016large,li2017visualizing}, leading to robust domain adaptation.
\begin{table*}[t]
\centering
\footnotesize
\begin{tabular}{l|p{1cm}<{\centering}p{1cm}<{\centering}p{1cm}<{\centering}p{1cm}<{\centering}p{1cm}<{\centering}
p{1cm}<{\centering}p{1cm}<{\centering}p{1cm}<{\centering}|c}
\hline
Method & person & rider & car & truck & bus & train & mcycle & bicycle & mAP \\
\hline
Baseline~\cite{ren2015faster} & 24.4 & 30.5 & 32.6 & 10.8 & 25.4 & 9.1 & 15.2 & 28.3 & 22.0 \\
MAF~\cite{he2019multi} & 28.4 & 39.5 & 43.9 & 23.8 & 39.9 & 33.3 & 29.2 & 33.9 & 34.0 \\
SCDA~\cite{zhu2019adapting} & 33.5 & 38.0 & 48.5 & 26.5 & 39.0 & 23.3 & 28.0 & 33.6 & 33.8 \\
DA~\cite{chen2018domain} &25.0 &31.0 &40.5 &22.1 &35.3 &20.2 &20.0 &27.1 &27.6\\
MLDA~\cite{Xie_2019_ICCV} &33.2 &44.2 &44.8 &28.2 &41.8 &28.7 &30.5 &36.5 &36.0\\
DMA~\cite{kim2019diversify} &30.8 &40.5 &44.3 &27.2 &38.4 &34.5 &28.4 &32.2 &34.6\\
CAFA~\cite{hsu2020every} &41.9 &38.7 &56.7 &22.6 &41.5 &26.8 &24.6 &35.5 &36.0 \\
\hline
SWDA~\cite{saito2019strong} &36.2 &35.3 &43.5 &30.0 &29.9 &42.3 &32.6 &24.5 &34.3\\
\textbf{+FAA} &39.5 &41.3 &47.0 &34.5 &39.3 &44.0 &31.9 &28.4 &\textbf{38.3}\\\hline
CRDA~\cite{xu2020exploring} & 32.9 & 43.8 & 49.2 & 27.2 & 45.1 & 36.4 & 30.3 & 34.6 & 37.4 \\
\textbf{+FAA} &37.4 &46.4 &48.3 &32.6 &46.5 &39.3 &32.3 &35.3 &\textbf{39.8}\\
\hline
\end{tabular}
\caption{Experimental results of domain adaptive object detection over the adaptation task Cityscapes $\rightarrow$ Foggy Cityscapes (\textit{+FAA} means to include FAA to attack supervised source loss and unsupervised target loss in domain adaptation).}
\label{table:det_city2fog}
\end{table*}
\section{Experiments}
This section presents experiments including datasets and implementation details, domain adaptation for semantic segmentation (with ablation studies), object detection, and image classification tasks, and discussion, more details to be described in the ensuing subsections.
\subsection{Datasets}
\textbf{Adaptation for semantic segmentation:} We consider two synthesized-to-real segmentation tasks: 1) GTA5 \cite{richter2016playing} $\rightarrow$ Cityscapes \cite{cordts2016cityscapes} and 2) SYNTHIA \cite{ros2016synthia} $\rightarrow$ Cityscapes. GTA5 contains $24,966$ synthetic images and shares $19$ categories with Cityscapes. For SYNTHIA, we use `SYNTHIA-RAND-CITYSCAPES' which contains $9,400$ synthetic images and shares 16 categories with Cityscapes. Cityscapes contains $2975$/$500$ training/validation images. Following \cite{tsai2018learning, zou2018unsupervised}, we adapt towards the Cityscapes training set and evaluate on the Cityscapes validation set.
\textbf{Adaptation for object detection:} We consider two adaptation tasks: 1) Cityscapes $\rightarrow$ Foggy Cityscapes~\cite{sakaridis2018semantic} and 2) Cityscapes $\rightarrow$ BDD100k~\cite{yu2018bdd100k}. For Cityscapes, we convert instance segmentation annotations to bounding boxes in experiments. Foggy Cityscapes was created by applying synthetic fog on Cityscapes images. BDD100k contains $100k$ images including $70k$ for training and $10k$ for validation. Following \cite{xu2020exploring,saito2019strong,chen2018domain}, we use a BDD100k subset \textit{daytime} in experiments, which includes $36,728$ training images and $5,258$ validation images.
\begin{table*}[h]
\centering
\footnotesize
\begin{tabular}{l|p{1cm}<{\centering}p{1cm}<{\centering}p{1cm}<{\centering}p{1cm}<{\centering}p{1cm}<{\centering}p{1cm}<{\centering}p{1cm}<{\centering}|c}
\hline
Method & person & rider & car & truck & bus & mcycle & bicycle & mAP \\
\hline
Baseline~\cite{ren2015faster} & 26.9 & 22.1 & 44.7 & 17.4 & 16.7 & 17.1 & 18.8 & 23.4 \\
DA~\cite{chen2018domain} & 29.4 & 26.5 & 44.6 & 14.3 & 16.8 & 15.8 & 20.6 & 24.0 \\
\hline
SWDA~\cite{saito2019strong} & 30.2 & 29.5 & 45.7 & 15.2 & 18.4 & 17.1 & 21.2 & 25.3 \\
\textbf{+FAA} &32.0 &32.2 &49.7 &19.8 &23.9 &18.5 &24.0 &\textbf{28.6}\\\hline
CRDA~\cite{xu2020exploring} & 31.4 & 31.3 & 46.3 & 19.5 & 18.9 & 17.3 & 23.8 & 26.9 \\
\textbf{+FAA} &32.9 &33.1 &51.4 &21.8 &24.5 &19.1 &26.2 &\textbf{29.9}\\
\hline
\end{tabular}
\caption{
Experimental results of domain adaptive object detection over the adaptation task Cityscapes $\rightarrow$ BDD100k (\textit{+FAA} means to include FAA to attack supervised source loss and unsupervised target loss in domain adaptation).}
\label{table:det_city2BDD}
\end{table*}
\textbf{Adaptation for image classification:} We adopt two adaptation benchmarks VisDA17~\cite{peng2018visda} and Office-31~\cite{saenko2010adapting}. VisDA17 includes $152,409$ synthetic images of $12$ categories as source and $55,400$ real images as target. Office-31 contains images of $31$ classes from Amazon (A), Webcam (W) and DSLR (D) that have $2817$, $795$ and $498$ images, respectively. We evaluate on six tasks A$\rightarrow$W, D$\rightarrow$W, W$\rightarrow$D, A$\rightarrow$D, D$\rightarrow$A, and W$\rightarrow$A as in~\cite{zou2019confidence, saenko2010adapting, sankaranarayanan2018generate}.
\subsection{Implementation Details}
\textbf{Semantic segmentation:} We use DeepLab-V2~\cite{chen2017deeplab} with ResNet101 \cite{he2016deep} as the segmentation network as in~\cite{tsai2018learning,zou2018unsupervised}. We use SGD optimizer \cite{bottou2010large} with a momentum $0.9$ and a weight decay $1e-4$. The initial learning rate is $2.5e-4$ and decayed by a polynomial policy of power $0.9$ \cite{chen2017deeplab}.
\textbf{Object detection:} As in~\cite{xu2020exploring, saito2019strong, chen2018domain}, we use Faster R-CNN~\cite{chen2017deeplab} with VGG-16~\cite{simonyan2014very} as the detection network. We use SGD optimizer \cite{bottou2010large} with a momentum $0.9$ and a weight decay $5e-4$. The initial learning rate is $1e-3$ for $50k$ iterations and reduced to $1e-4$ for another $20k$ iterations \cite{xu2020exploring,saito2019strong,chen2018domain}. In all experiments, we set image shorter side to 600 and employ RoIAlign~\cite{he2017mask} for feature extraction.
\textbf{Image classification:} For fair comparisons, we follow ~\cite{zou2019confidence, saenko2010adapting,sankaranarayanan2018generate} and use ResNet-101/ResNet-50~\cite{he2016deep} (pre-trained with ImageNet \cite{deng2009imagenet}) as backbones. We use SGD optimizer \cite{bottou2010large} with a momentum $0.9$ and a weight decay $5e-4$. The learning rate is $1e-3$ and the batch size is $32$ \cite{zou2019confidence}.
We set the parameter $p$ and the number of FCs $N$ at $0.1$ and $96$.
The band-pass filter $R$ follows \cite{huang2021fsdr} with mid-pass and low-/high-rejected designs to get the mid-frequency content that captures semantic information (e.g., structures and shapes).
\subsection{Domain Adaptive Semantic Segmentation}
Table~\ref{table:gta2city} shows experimental results over semantic segmentation task GTA5 $\rightarrow$ Cityscapes. It can be seen that FAA is generic and can be applied to attack both \textit{Baseline} (for preventing overfitting in supervised source loss) and state-of-the-art UDA methods (for preventing overfitting in both supervised source loss and unsupervised target loss). In addition, incorporating FAA improves both \textit{Baseline} and UDA methods clearly and consistently.
We perform ablation studies over two representative UDA methods using adversarial alignment~\cite{tsai2018learning} and self-training~\cite{zou2018unsupervised}, where \textit{+FAA-S}, \textit{+FAA-T} and \textit{+FAA} address the overfitting of supervised source losses, unsupervised target losses and both losses, respectively. It can be seen that \textit{+FAA-S} and \textit{+FAA-T} both improve domain adaptation by large margins. This shows that both supervised source objective and unsupervised target objective introduce clear overfitting and FAA mitigates the overfitting effectively. In addition, \textit{+FAA} performs clearly the best. This shows that preventing the two learning objectives from over-minimization is complementary as overftting in the two learning objectives affects generalization in different manners. Specifically, supervised source loss has domain gap and over-minimizing it guides the model to over-memorize source data whereas unsupervised target loss is noisy and over-minimizing it leads to deviated solutions with accumulated errors.
\begin{table*}[!t]
\centering
\centering
\footnotesize
\begin{tabular}{c|cccccccccccc|c}
\hline
Method & Aero & Bike & Bus & Car & Horse & Knife & Motor & Person & Plant & Skateboard & Train & Truck & Mean\\
\hline
Res-101 \cite{saito2018adversarial} & 55.1 & 53.3 & 61.9 & 59.1 & 80.6 & 17.9 & 79.7 & 31.2 & 81.0 & 26.5 & 73.5 & 8.5 & 52.4\\
MMD \cite{long2015learning} & 87.1 & 63.0 & 76.5 & 42.0 & 90.3 & 42.9 & 85.9 & 53.1 & 49.7 & 36.3 & 85.8 & 20.7 & 61.1\\
DANN \cite{ganin2016domain} & 81.9 & 77.7 & 82.8 & 44.3 & 81.2 & 29.5 & 65.1 & 28.6 & 51.9 & 54.6 & 82.8 & 7.8 & 57.4\\
ENT \cite{grandvalet2005semi} & 80.3 & 75.5 & 75.8 & 48.3 & 77.9 & 27.3 & 69.7 & 40.2 & 46.5 & 46.6 & 79.3 & 16.0 & 57.0\\
MCD \cite{saito2018maximum} & 87.0 & 60.9 & {83.7} & 64.0 & 88.9 & 79.6 & 84.7 & {76.9} & {88.6} & 40.3 & 83.0 & 25.8 & 71.9\\
ADR \cite{saito2018adversarial} & 87.8 & 79.5 & {83.7} & 65.3 & {92.3} & 61.8 & {88.9} & 73.2 & 87.8 & 60.0 & {85.5} & {32.3} & 74.8\\
SimNet-Res152 \cite{pinheiro2018unsupervised} & {94.3} & 82.3 & 73.5 & 47.2 & 87.9 & 49.2 & 75.1 & 79.7 & 85.3 & 68.5 & 81.1 & 50.3 & 72.9\\
GTA-Res152 \cite{sankaranarayanan2018generate} & - & - & - & - & - & - & - & - & - & - & - & - & 77.1\\
\hline
CBST~\cite{zou2018unsupervised} &87.2 & 78.8 & 56.5 & 55.4 & 85.1 & 79.2 & 83.8 & 77.7 & 82.8 & 88.8 & 69.0 & 72.0 & 76.4\\
\textbf{CBST+FAA} &90.6 &80.3 &79.6 &67.1 &86.7 &80.2 &86.2 &77.1 &86.2 &87.1 &80.6 &71.8 &\textbf{81.1}\\\hline
CRST~\cite{zou2019confidence} & 88.0 & 79.2 & 61.0 & 60.0 & 87.5 & 81.4 & 86.3 & 78.8 & 85.6 & 86.6 & 73.9 & 68.8 &78.1\\
\textbf{CRST+FAA} &91.6 &80.5 &81.5 &70.7 &89.6 &81.0 &87.5 &79.9 &87.1 &86.4 &81.0 &75.1 &\textbf{82.7}\\
\hline
\end{tabular}
\caption{
Experimental results of domain adaptive image classification task on VisDA17 (\textit{+FAA} means to include FAA to attack supervised source loss and unsupervised target loss in domain adaptation).}
\label{table:visda17}
\end{table*}
\begin{table}[t]
\centering
\resizebox{0.95\linewidth}{!}{
\begin{tabular}{c|cccccc|c}
\hline
Method & A$\rightarrow$W & D$\rightarrow$W & W$\rightarrow$D & A$\rightarrow$D & D$\rightarrow$A & W$\rightarrow$A & Mean\\
\hline
ResNet-50 \cite{he2016deep} & 68.4 & 96.7 & 99.3 & 68.9 & 62.5 & 60.7 & 76.1\\
DAN \cite{long2015learning} & 80.5 & 97.1 & 99.6 & 78.6 & 63.6 & 62.8 & 80.4\\
RTN \cite{long2016unsupervised} & 84.5 & 96.8 & 99.4 & 77.5 & 66.2 & 64.8 & 81.6\\
DANN \cite{ganin2016domain} & 82.0 & 96.9 & 99.1 & 79.7 & 68.2 & 67.4 & 82.2\\
ADDA \cite{tzeng2017adversarial} & 86.2 & 96.2 & 98.4 & 77.8 & 69.5 & 68.9 & 82.9\\
JAN \cite{long2017deep} & 85.4 & 97.4 & 99.8 & 84.7 & 68.6 & 70.0 & 84.3\\
GTA \cite{sankaranarayanan2018generate} & {89.5} & 97.9 & 99.8 & 87.7 & 72.8 & 71.4 & 86.5\\
\hline
CBST~\cite{zou2018unsupervised} & 87.8 & 98.5 & {100} & 86.5 & 71.2 & 70.9 & 85.8\\
\textbf{CBST+FAA} &90.9 &98.9 &99.8 &92.5 &76.7 &76.2 &\textbf{89.1}\\\hline
TAT~\cite{liu2019transferable} & 92.5 & 99.3 &100.0 & 93.2 & 73.1 & 72.1 & 88.4\\
\textbf{TAT+FAA} &92.4 &99.8 &99.4 &94.7 &78.2 &77.1 &\textbf{90.3}\\\hline
CRST~\cite{zou2019confidence} & 89.4 & {98.9} & {100} & 88.7 & 72.6 & 70.9 & 86.8\\
\textbf{CRST+FAA} &92.3 &99.2 &99.7 &94.4 &80.5 &78.7 &\textbf{90.8}\\
\hline
\end{tabular}
}
\caption{
Experimental results of domain adaptive image classification over Office-31 (\textit{+FAA} means to include FAA to attack supervised source loss and unsupervised target loss in domain adaptation).}
\label{table:office}
\end{table}
Table~\ref{table:synthia2city} shows experimental results over semantic segmentation task SYNTHIA $\rightarrow$ Cityscapes. We can observe that FAA improves state-of-the-art UDA methods in the similar manner as in Table~\ref{table:gta2city}. Note we applied FAA to a few representative UDA methods only due to space limit.
\subsection{Domain Adaptive Object Detection}
Table~\ref{table:det_city2fog} shows domain adaptive object detection over the task Cityscapes $\rightarrow$ Foggy Cityscapes. It can be seen that FAA boosts mAP by over $+2.5\%$ for both SWDA~\cite{saito2019strong} and CRDA~\cite{xu2020exploring}. Note that we did not apply FAA to other listed UDA methods due to space limit.
Table~\ref{table:det_city2BDD} shows domain adaptive object detection over the task Cityscapes $\rightarrow$ BDD100k. It can be seen that including FAA outperforms state-of-the-art UDA methods consistently as in Table~\ref{table:det_city2fog}. Note that we did not apply FAA to other listed UDA methods due to space limit.
\subsection{Domain Adaptive Image Classification}
We presents experimental results on VisDA17 in Table~\ref{table:visda17} in per-class accuracy. It can be seen that incorporating FAA outperforms state-of-the-art UDA methods consistently. This applies to UDA methods that employ stronger backbones ResNet-152~\cite{pinheiro2018unsupervised,sankaranarayanan2018generate}.
Table~\ref{table:office} shows domain adaptive image classification experiments over Office-31 (all using the same backbone ResNet-50). We can see that incorporating FAA leads to robust domain adaptation and improve the image classification consistently by large margins.
\subsection{Discussion}
\begin{figure*}[t]
\begin{tabular}{p{3cm}<{\centering} p{3cm}<{\centering} p{3cm}<{\centering} p{3cm}<{\centering} p{3cm}<{\centering}}
\raisebox{-0.5\height}{\includegraphics[width=1.1\linewidth,height=0.55\linewidth]{figures/appendix/figure_seg_al/frankfurt_000000_010351_leftImg8bit.png}}
& \raisebox{-0.5\height}{\includegraphics[width=1.1\linewidth,height=0.55\linewidth]{figures/appendix/figure_seg_al/frankfurt_000000_010351_leftImg8bit_color_source_only.png}}
& \raisebox{-0.5\height}{\includegraphics[width=1.1\linewidth,height=0.55\linewidth]{figures/appendix/figure_seg_al/frankfurt_000000_010351_leftImg8bit_color_AL.png}}
& \raisebox{-0.5\height}{\includegraphics[width=1.1\linewidth,height=0.55\linewidth]{figures/appendix/figure_seg_al/frankfurt_000000_010351_leftImg8bit_color.png}}
& \raisebox{-0.5\height}{\includegraphics[width=1.1\linewidth,height=0.55\linewidth]{figures/appendix/figure_seg_al/frankfurt_000000_010351_gtFine_color.png}}
\vspace{-2.5 pt}
\\
\raisebox{-0.5\height}{Target Image}
& \raisebox{-0.5\height}{Baseline}
& \raisebox{-0.5\height}{AdaptSeg~\cite{tsai2018learning}}
& \raisebox{-0.5\height}{\textbf{Ours(AdaptSeg+FAA)}}
& \raisebox{-0.5\height}{Ground Truth}
\\
\end{tabular}
\vspace{2.5 pt}
\caption{
Qualitative illustration of domain adaptive semantic segmentation task GTA5 to Cityscapes task. Our robust domain adaption (RDA) employs Fourier adversarial attacking to mitigate the overfitting in domain adaptive learning effectively, which produces better semantic segmentation by mitigating over-fitting in domain adaptive learning. The differences are clearer for less-frequent categories such as traffic-sign, pole and bus, etc.
}
\label{fig:results_seg_al}
\end{figure*}
\textbf{Qualitative illustration of domain adaptation in segmentation}
We compare our robust domain adaptation (RDA) with the recent domain adaptation method~\cite{tsai2018learning} qualitatively over semantic segmentation task. As Fig.~\ref{fig:results_seg_al} shows, the proposed RDA outperforms the baselines clearly.
\textbf{Overfitting Mitigation:} We compared FAA with existing overfitting mitigation methods. Most existing methods address overfitting through certain network regularization by noise injection to hidden units (\textit{Dropout}), label-dropout (\textit{Label smooth}), gradient ascent (\textit{Flooding}), data and label mixing (\textit{Mixup}), gradient based adversarial attacking (\textit{FGSM}), virtual-label based adversarial attacking (\textit{VAT}), etc. Table~\ref{tab:comp_overfitting} shows experimental results over the task GTA$\rightarrow$Cityscapes. It can be seen that existing regularization does not perform well in the domain adaptation task. The major reason is that existing methods were designed for supervised and semi-supervised learning where training and test data usually have little domain gap. The proposed FAA mitigates overfitting with clear performance gains as it allows large magnitude of perturbation noises
which is critical to the effectiveness of its generated adversarial samples due to the existence of `domain gaps’ in UDA.
Due to the space limit, we provide more visualization of the qualitative segmentation examples (including their comparisons with the recent domain adaptation method~\cite{tsai2018learning}) in the supplementary material.
\begin{table}[t]
\centering
\begin{footnotesize}
\begin{tabular}{c|ccc}
\hline
Method &mIoU & gain
\\\hline
ST~\cite{zou2018unsupervised} &42.6 &N.A.
\\\hline
+Dropout~\cite{srivastava2014dropout} &43.1 &+0.5\\
+Label smooth~\cite{szegedy2016rethinking} &43.4 &+0.8\\
+Mixup~\cite{zhang2017mixup} &43.6 &+1.0\\
+FGSM~\cite{goodfellow2014explaining} &43.9 &+1.3\\
+VAT~\cite{miyato2018virtual} &44.3 &+1.7\\
+Flooding~\cite{ishida2020we} &44.1 &+1.5\\
\textbf{+FAA} &\textbf{50.1} &\textbf{+7.5}\\
\hline
\end{tabular}
\end{footnotesize}
\caption{
Comparison with existing overfitting migitation methods: For the semantic segmentation task GTA $\rightarrow$ Cityscapes, FAA performs the best consistently by large margins.
}
\label{tab:comp_overfitting}
\end{table}
\section{Conclusion}
In this work, we presented RDA, a robust domain adaptation technique that mitigates overfitting in UDA via a novel Fourier adversarial attacking (FAA).
We achieve robust domain adaptation by a novel Fourier adversarial attacking (FAA) method that allows large magnitude of perturbation noises but has minimal modification of image semantics.
With FAA-generated adversarial samples, the training can continue the `random walk’ and drift into an area with a flat loss landscape, leading to more robust domain adaptation.
Extensive experiments over multiple domain adaptation tasks show that RDA can work with different computer vision tasks (\ie, segmentation, detection and classification) with superior performance. We will explore disentanglement-based adversarial attacking and its applications to other computer vision tasks. We will also study how FAA could mitigate over-fitting in classical supervised learning and semi-supervised learning as well as the recent contrast-based unsupervised representation learning.
\section*{Acknowledgement}
This study is supported under the RIE2020 Industry Alignment Fund – Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from Singapore Telecommunications Limited (Singtel), through Singtel Cognitive and Artificial Intelligence Lab for Enterprises (SCALE@NTU).
{\small
\bibliographystyle{ieee_fullname}
|
\section{\label{Introduction}Introduction}
Many systems in the real world can be simplified as networks, where each node (vertex) represents an individual and an edge exists between two nodes if the two corresponding individuals interact in some way. Examples include friendships in social networks of interactions among people \citep{HE202134,West18355}, molecular bindings in biological networks of molecules \citep{SantoliniE6375,2020Robustness,Gro2019Representing}, and web hyperlinks in the World Wide Web \citep{barabasi2000scale}. In the past, most studies have only treated networks as unadorned sets of nodes and their links \citep{girvan2002community,5fortunato2009community}. In recent years, most network data, however, are accompanied by the contents that describe the properties of nodes. For example, a user on Twitter and users whom he/she follow represents relationships while Twitter lists and tweets they post describe the profile of the user. Such networks represented by the semantic contents combining with links (also called attribute information and topology information respectively) are referred as attributed networks or attributed graphs \citep{li2018community,2020Network,ren2021block}.
As a network structure sheds light on the behavior of a system in a way, a large number of studies have been devoted to the detection of community structures in networks \citep{5fortunato2009community,chunaev2020community,2009Communities}. However, most of these studies have focused on methods of discovering traditional communities, i.e., groupings of nodes with dense internal connections and sparse external ones. In fact, with the emergence of various complex networks in different fields, including social, information, biological and physical sciences, different types of network structures are discovered and studied. Examples include bipartite structure, core-periphery structure, and their mixture structure, etc. Here traditional communities and other types of communities are referred to as generalized structures\citep{matias2014modeling,2015Generalized}.
In recent years, some methods have been proposed for the exploration of structures contained in node-attribute networks \citep{chunaev2020community,2015Clustering}, which roughly fall into two categories according to how to use attributes: the methods that use full attribute space \citep{bojchevski2018bayesian,chang2019generative,jin2019detecting,chen2016network,chai2013combining} and the methods that explore subspaces of attributes \citep{gunnemann2011db,gunnemann2013spectral,perozzi2014focused,wu2018mining,gunnemann2013efficient,huang2017joint,chen2020attributed}. The former fuses structure and all available attributes to improves community detection quality, the latter believes that part attributes are related to obtaining good-quality communities. One of the former subclasses is the probabilistic model-based methods, which generate links and node attributes through a joint probability function and the model parameters. The generative models can be further classified into two categories in terms of types of network structure: one mainly detects traditional communities, and the other detects generalized communities, including traditional communities. For example, Yang et al. \citep{yang2009combining} combined a popularity-based conditional link Model PCL with a discriminative content (DC) model for community detection (termed PCL\_DC). By introducing node productivity, Yang et al. \citep{yang2010directed} further developed a popularity and productivity link model PPL, and the corresponding united model was called PPL\_DC. However, both PCL\_DC and PPL\_DC only detect traditional network structure since they assume that nodes in the same community have more opportunities to link each other. On the contrary, due to Newman’s mixture model (NMM) \citep{newman2007mixture}, the method BNPA proposed by Chen et al. \citep{chen2016network} can detect generalized network structure. Thanks to the block structure assumption of SBM \citep{holland1983stochastic}, the models of Chai et al. \citep{chai2013combining}, He et al. \citep{he2017joint}, and Chang et al. \citep{chang2019generative} can discover generalized network structure. However, Chai et al. \citep{chen2016network} also used a DC model for attributes so that the proposed model PPSB\_DC was not good at semantic interpretability because the model only learned a weight vector of attributes for communities. He et al. \citep{he2017joint} developed a model NEMBP that had good semantic interpretability because the relationship between a community and its corresponding attributes was characterized. However, NEMBP needed to specify the number of both topics and communities in advance. PSB\_PG \citep{chang2019generative} also had good semantic interpretability but had nonlinear complexity in a naive EM algorithm, which means the method works well for networks of moderate size.
Here we propose a principle generative model to address the above problems. Firstly, based on the connectivity behavior of nodes, a model that can generate topology information is proposed. Secondly, based on the homogeneity of attributes, a model that can generate attribute information is developed. Finally, the two models are united together by sharing the latent locations of nodes and a generative node-attribute network model (named GNAN) is formed naturally. The main contributions of the proposed method are summarized as follows.
\begin {itemize}
\item The proposed model that combines topological information with attribute information can detect communities more accurately than the model that only uses topology information, which means the node attributes are effectively utilized and can complement the network structure.
\item The new model GNAN can classify the nodes of a network into groups such that the link patterns of each group are similar in some sense. Therefore, the model can detect not only traditional communities, but also a range of other types of structures in networks, such as bipartite structure, core-periphery structure, and their mixture structure.
\item The dependency between attributes and communities can be automatically learned by our model and thus we can ignore the attributes that do not contain useful information. A case study is provided to show the ability of our model in the semantic interpretability of communities. And experiments on both synthetic and real-world networks show that the new model is competitive with other state-of-the-art models.
\end {itemize}
The rest of the paper is organized as follows. In Sect.II, a generative node-attribute network model is described. In Sect.III, the model parameters are inferred and the corresponding algorithm is designed. In Sect.IV, the new algorithm is evaluated and compared with some related methods on both synthetic and real-world networks. Finally, In Sect.V, we derive the conclusions.
\section{\label{GNAN}GNAN: A generative node-attribute network model}
Let $ G(V, E, X) $ be a mathematical formalization for a network, where $ V = \left\{ {1,2, \cdots ,N} \right\} $ is the set of nodes (vertices), $ E = \left\{ {e_1 ,e_2 , \cdots ,e_M } \right\} $ is the set of links (edges), $ N $ and $ M $ are the number of nodes and links, respectively. $ X = \left( {x_{ik} } \right)_{N \times K} $ is a node-attribute matrix of a network and represents the attribute information contained in a network data set, $ K $ is the dimension of the node attributes, $ x_{ik} = 1 $ means node $ i $ has the $ k $th attribute, or 0 otherwise. The topology information of a network is represented by an adjacent matrix $ A = \left( {a_{ij} } \right)_{N \times N} $, where $ a_{ij} = 1 $ if a link between the pair of nodes $ (i, j) \in E$, or 0 otherwise. Suppose $ V_1, V_2, \cdots , V_C $ are the communities embedded in a network, and $ \bigcup\limits_{r = 1}^C {V_r } = V $.
\textbf{Modeling the links of a network} In order to enable the topology structure generated by the new model to form a wide range of network structures, we consider introducing a group of parameters that characterize the connectivity behavior of nodes. A community is a set of nodes that share the same connectivity behavior, which is in line with our intuition (See a toy example in FIG.\ref{fig0} \citep{matias2014modeling}).
\begin{figure}[htbp]
\centering
\includegraphics[width=14.5cm]{linkpattern.pdf}
\caption{A toy network with 2 communities represented by colors (blue or red). The left panel means that a community is a set of nodes that share a large number of connections, i.e., a traditional community. The right panel means that a community is a set of nodes that share the same connectivity behavior, i.e., a generalized community. Of course, the situation of the left panel can also be regarded as an example based on the connectivity behavior of nodes.}\label{fig0}
\end{figure}
In this paper, we adopt a parameter matrix $ \Theta = \left( {\theta _{rj} } \right)_{C \times N} $ to measure the connectivity behavior of nodes, where the entry $ \theta_{rj} $ represents the probability that any particular node in community $ V_r $ sends an edge to node $ j $, and $ \sum\limits_{j = 1}^N {\theta _{rj} } = 1 $, which is also used in NMM \citep{newman2007mixture}. Thus, $ \theta_{rj} $ characterizes the preference of nodes in community $ V_r $ about which other nodes they like to link to. These preferences form the communities in which nodes have similar patterns of connection to others. To generate an expected link between a pair of nodes $ (i, j) $, another parameter matrix $ {\rm T} = \left( {\tau _{ir} } \right)_{N \times C} $ is introduced to our model, where the entry $ \tau _{ir} $ represents the probability that a node $ i $ falls into community $ V_r $, and $ \sum\limits_{r = 1}^C {\tau _{ir} } = 1 $. Based on both parameters $ \tau _{ir} $ and $ \theta_{rj} $, an expected link between node pairs $ (i, j) $ through $ V_r $ is $ \hat a_{ij,r} = \tau _{ir} \theta _{rj} $. Summing over communities $ V_r $, the expected number of links between a pair of nodes $ (i, j) $ is $ \hat a_{ij} = \sum\limits_{r = 1}^C {\tau _{ir} \theta _{rj} } $. Suppose directed links are placed independently between node pairs with probabilities that are Poisson distribution, the likelihood of topology information is
\[
\Pr \left( {A|{\rm T},\Theta } \right) = \prod\limits_{i,j = 1}^N {\frac{{\hat a_{ij}^{a_{ij} } }}{{a_{ij} !}}\exp \left\{ { - \hat a_{ij} } \right\}} = \prod\limits_{i,j = 1}^N {\frac{{\left( {\sum\limits_{r = 1}^C {\tau _{ir} \theta _{rj} } } \right)^{a_{ij} } }}{{a_{ij} !}}\exp \left\{ { - \sum\limits_{r = 1}^C {\tau _{ir} \theta _{rj} } } \right\}}. \eqno(1)
\]
\textbf{Modeling the attributes of nodes in a network} In order to automatically learn the dependency between attributes and communities, a parameter matrix $ \Phi = \left( {\phi _{rk} } \right)_{C \times K} $ is introduced to the model of generating node attributes, where the entry $ \phi _{rk} $ represents the probability that a community $ V_r $ has the $ k $th attribute, and $ \sum\limits_{k = 1}^K {\phi _{rk} } = 1 $. Therefore, a node $i$ in community $V_r$ possessing $k$th attribute can be represented as $ \hat x_{ik,r} = \tau_{ir} \phi _{rk} $. Summing over all communities $V_r$, the expected propensity of a node $i$ possessing $k$th attribute is $ \hat x_{ik} = \sum\limits_r {\tau_{ir} \phi _{rk} } $. Suppose $x_{ik} (k=1,2,\dots,K)$ is independent and identically distributed, we have
\[
\Pr \left( {X|{\rm T},\Phi } \right) = \prod\limits_{i = 1}^N {\prod\limits_{k = 1}^K {\frac{{\hat x_{ik}^{x_{ik} } }}{{x_{ik} !}}\exp \left\{ { - \hat x_{ik} } \right\}} } = \prod\limits_{i = 1}^N {\prod\limits_{k = 1}^K {\frac{{\left( {\sum\limits_{r = 1}^C {\tau _{ir} \phi _{rk} } } \right)^{x_{ik} } }}{{x_{ik} !}}\exp \left\{ { - \sum\limits_{r = 1}^C {\tau _{ir} \phi _{rk} } } \right\}} }. \eqno(2)
\]
In effect, $ \phi _{rk} $ characterizes the preference of which attributes nodes in community $ V_r $ possess, which is similar to the parameter $ \theta_{rj} $ in some sense. The preferences define communities in which nodes have the same attributes (named attribute communities here to discriminate the communities formed by $ \theta_{rj} $). We can order the attributes learned by $ \theta_{rj} $ and easily choose some important attributes for a certain community. These selected attributes may come from the same topic or different ones if the attributes have topics, but they are shared simultaneously by nodes in community $ V_r $, we call such property homogeneity of attributes. The shared attributes naturally build up the semantic interpretation for each community. Note that the constraints are added to the columns of the parameter $ \Phi $, which means that an attribute can maintain a close relationship with multiple communities at the same time. Therefore, there may not be a one-to-one relationship between communities and attribute topics if the attributes have topics.
\textbf{GNAN} By sharing the latent position of nodes, the joint likelihood function for generating node-attribute network can be described as follows:
\[
\begin{array}{l}
\Pr \left( {A{\rm{,}}X|{\rm T},\Theta ,\Phi } \right) = \Pr \left( {A|{\rm T},\Theta } \right)\Pr \left( {X|{\rm T},\Phi } \right) \\
\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; = \prod\limits_{i,j = 1}^N {\frac{{\left( {\sum\limits_{r = 1}^C {\tau _{ir} \theta _{rj} } } \right)^{a_{ij} } }}{{a_{ij} !}}\exp \left\{ { - \sum\limits_{r = 1}^C {\tau _{ir} \theta _{rj} } } \right\}} \\
\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \times \prod\limits_{i = 1}^N {\prod\limits_{k = 1}^K {\frac{{\left( {\sum\limits_{r = 1}^C {\tau _{ir} \phi _{rk} } } \right)^{x_{ik} } }}{{x_{ik} !}}\exp \left\{ { - \sum\limits_{r = 1}^C {\tau _{ir} \phi _{rk} } } \right\}} }, \\
\end{array} \eqno(3)
\]
where `` $ \times $ " represents multiplication. Under the assumption of the sparsity of links and attributes, a unified Poisson distribution likelihood function makes the model both reasonable and easy to calculate.
\section{A proposed algorithm for detecting generalized structure with GNAN}
In this section, firstly, the parameters $ {\rm T} $ and $ \Phi $ in model Eq.(3) are inferred by the EM algorithm. The learned parameter $ {\rm T} $ can help us to derive the network structures embedded in a network. The inferred parameter $ \Phi $ represents the dependency of communities and attributes, helping to explain why these nodes come together. Then, based on the inferred parameters, an algorithm for GNAN is designed.
Because the parameters in model Eq.(3) are related to the potential position of nodes that cannot be observed (i.e., a latent variable), it is difficult to directly estimate them. The EM algorithm can conveniently handle this type of parameter estimation problem with latent variables. Considering the logarithm of the model Eq.(3), ignoring constants and terms independent of parameters and latent variables, we have
\[
L\left( {{\rm T},\Theta ,\Phi } \right) = \sum\limits_{i.j = 1}^N {\left[ {a_{ij} \log \left( {\sum\limits_{r = 1}^C {\tau _{ir} \theta _{rj} } } \right) - \sum\limits_{r = 1}^C {\tau _{ir} \theta _{rj} } } \right]} + \sum\limits_{i = 1}^N {\sum\limits_{k = 1}^K {\left[ {x_{ik} \log \left( {\sum\limits_{r = 1}^C {\tau _{ir} \phi _{rk} } } \right) - \sum\limits_{r = 1}^C {\tau _{ir} \phi _{rk} } } \right]} }. \eqno(4)
\]
From Jensen's inequality, the lower bound of the log-likelihood Eq.(4) is as follows:
\[
\bar L\left( {{\rm T},\Theta ,\Phi } \right) = \sum\limits_{ijr} {a_{ij} q_{ij,r} \log \frac{{\tau _{ir} \theta _{rj} }}{{q_{ij,r} }}} - \sum\limits_{ijr} {\tau _{ir} \theta _{rj} } + \sum\limits_{ikr} {x_{ik} h_{ik,r} \log \frac{{\tau _{ir} \phi _{rk} }}{{h_{ik,r} }}} - \sum\limits_{ikr} {\tau _{ir} \phi _{rk} }, \eqno(5)
\]
where
\[
q_{ij,r} = \frac{{\tau _{ir} \theta _{rj} }}{{\sum\limits_{r = 1}^C {\tau _{ir} \theta _{rj} } }},\;h_{ik,r} = \frac{{\tau _{ir} \phi _{rk} }}{{\sum\limits_{r = 1}^C {\tau _{ir} \phi _{rk} } }}. \eqno(6)
\]
are the expected probabilities of a node pair $ (i, j) $ in community $ V_r $ to be linked and those of nodes $i ( \in V_r ) $ possessing $k$th attribute, respectively. By using the Lagrange multiplicator method, we can obtain the estimates of $ {\rm T} $, $\Theta$, and $\Phi$ that maximize the lower bound $ \bar L\left( {{\rm T},\Theta ,\Phi } \right) $ in Eq.(5) in the following.
\[
\tau _{ir} = \frac{{\sum\limits_j {a_{ij} q_{ij,r} } + \sum\limits_k {x_{ik} h_{ik,r} } }}{{\sum\limits_{jr} {a_{ij} q_{ij,r} } + \sum\limits_{kr} {x_{ik} h_{ik,r} } }},\;\theta _{rj} = \frac{{\sum\limits_i {a_{ij} q_{ij,r} } }}{{\sum\limits_{ij} {a_{ij} q_{ij,r} } }},\;\phi _{rk} = \frac{{\sum\limits_i {x_{ik} h_{ik,r} } }}{{\sum\limits_{ik} {x_{ik} h_{ik,r} } }}. \eqno(7)
\]
See the appendix for a detailed derivation. Eqs.(6,7) build up our EM algorithm for GNAN, which will converge \citep{wu1983convergence}.
\begin{algorithm}[H]
\caption{EM algorithm for GNAN}
\renewcommand{\algorithmicrequire}{ \textbf{Input:}}
\renewcommand{\algorithmicensure}{ \textbf{Output:}}
\label{Al:01}
\begin{algorithmic}[1]
\REQUIRE~~\\
the adjacency matrix $A$, the attribute matrix $ X $\\
the number of communities $C$, \\
the maximum iteration $I_{T}$, and the threshold $\epsilon$ \\
\ENSURE ~~\\
the inferred parameters $ {\rm T},\Theta, \Phi $
\vspace{2mm}
\STATE Initialize $ {\rm T^{(0)}},\Theta^{(0)}, \Phi^{(0)} $ by Eq.(7).
\STATE Compute lower bound $ \bar L\left( {{\rm T^{(0)}},\Theta^{(0)} ,\Phi^{(0)} } \right) $ by Eq.(5).
\FOR{$t=1:I_{T}$}
\vspace{1mm}
\STATE E-step: Compute $ q_{ij,r},\; h_{ik,r} $ by Eq.(6).
\vspace{1mm}
\STATE M-step: Compute $ {\rm T^{(t)}},\Theta^{(t)}, \Phi^{(t)} $ by Eq.(7).
\vspace{1mm}
\STATE Compute lower bound $ \bar L\left( {\rm T^{(t)}},\Theta^{(t)}, \Phi^{(t)} \right) $ by Eq.(5).
\vspace{1mm}
\IF{ $ \left| \bar L\left( {\rm T^{(t)}},\Theta^{(t)}, \Phi^{(t)} \right)- \bar L\left( {\rm T^{(t-1)}},\Theta^{(t-1)}, \Phi^{(t-1)} \right) \right| < \epsilon $ or $ t=I_{T} $}
\vspace{1mm}
\STATE $ \hat{\rm T} = {\rm T^{(t)}} ,\; \hat{\Theta} = \Theta ^{(t)} ,\; \hat{\Phi} = \Phi ^{(t)} $; STOP;
\ENDIF
\ENDFOR
\end{algorithmic}
\end{algorithm}
Algorithm 1 (called GNAN for simplicity) will converge to a local optimum of the likelihood. Therefore, there usually have different solutions from different starting points. For the robustness of our results, the initialization about parameters $ {\rm T},\Theta, \Phi $ is generated from a uniform distribution over [0.5-$ \zeta $,0.5+$ \zeta $], where $ \zeta $ is a random perturbation. In the following experiments, the maximum iteration $I_{T}=500$, and the threshold $\epsilon=10^{-6}$.
The time complexity of the new algorithm is dominated by updating $ q_{ij, r}, h_{ik, r} $ in step 4 and $ \tau _{ir}, \theta _{rj}, \phi _{rk} $ in step 5. Updating them for all nodes takes $ O(MC) $ and $ O(NKC) $ operations, which is linear in the number of communities $C$, the number of links $M$, the dimension of each node attribute $ K $, and the number of nodes $N$. Therefore, the total time complexity of the algorithm is $O((MC+NKC)I_T)$, where $I_T$ is the number of iterations. The memory of the method outlined above is dominated by updating $ q_{ij, r} $ and $ h_{ik, r} $ in step 4. The space required to store $ q_{ij, r} $ is $ O(MC) $ while the $ h_{ik, r} $ is $ O(NKC) $. Therefore, the whole memory use of the algorithm is $ O(MC+NKC) $, which is linear in $ N, M, C $, and $ K $. Obviously, the complexity of the GNAN is lower than that of the PSB\_PG ($ O(MC^2+NKC^2) $).
\section{Experiments}
Firstly, the ability of the new method to discover community and node-attribute information was shown on artificial networks. Then, we applied our method to a real friendship network to show the semantic interpretation of communities in a case study. Finally, our new model GNAN was evaluated on synthetic and real-world networks with a range of known network structures in comparison with 4 state-of-the-art methods. The metric NMI (Normalized Mutual Information) \citep{danon2005comparing} was used here to evaluate an algorithm running on a network with ground-truth communities.
\textbf{NMI.} Suppose $V = \left( {V_{1} ,V_{2} , \cdots ,V_{C} } \right)$ are true communities in a network, $V^{*} = \left( {V_{1}^{*} ,V_{2}^{*} , \cdots ,V_{C}^{*} } \right)$ are inferred communities. NMI is defined as follows
\[
NMI(V ,V^{*} ) = \frac{\displaystyle {-2 \sum\limits_{i = 1}^C {\sum\limits_{j = 1}^C {N_{ij} \log \frac{{NN_{ij} }}{{N_i^{V } N_j^{V^{*} } }}} } }}{{\sqrt { \displaystyle \left( {\sum\limits_{i = 1}^C {N_i^{V} \log \frac{{N_i^{V} }}{N}} } \right)\left( {\sum\limits_{j = 1}^C {N_j^{V^{*} } \log \frac{{N_j^{V^{*} } }}{N}} } \right)} }}, \eqno(8)
\]
where $C$ is the number of communities in a network; $N$ is the number of nodes; $N_{ij}$ is the number of nodes in the true community $V_i$ that are assigned to the inferred community $V_{j}^{*}$; ${N_i^{V} }$ is the number of nodes in the true community $V_i$; ${N_j^{V^{*} } }$ is the number of nodes in the inferred community $V_{j}^{*}$. A larger $NMI$ means a better partition.
\subsection{Performance on synthetic networks}
In this section, the performance of the GNAN algorithm was tested on artificial networks that were generated by the standard stochastic blockmodel (SBM) \citep{holland1983stochastic}. In fact, the SBM can produce flexible and challenging synthetic networks with a wide variety of network structures. And the strength of network structures is easily controlled. After generating artificial networks, the attributes related to communities were produced by 0-1 distribution $ Bin(1, p) $, where $ p $ measures how well a community matches attributes. The larger the values of $ p $, the stronger the dependency between attributes and communities. Here, assume that each community has a strong dependency with 10 attributes ($ p > 0.1 $), but has little relationship with the remaining attributes ($ p=0.1 $).
\textbf{Community structure.} Suppose the parameter generating network structures in SBM is $ \omega $ which has the following particular form
\[
\omega ^{{\rm{planted}}} = \left( \begin{array}{l}
\omega \;\;\lambda \;\;\lambda \;\;\lambda \\
\lambda \;\;\omega \;\;\lambda \;\;\lambda \\
\lambda \;\;\lambda \;\;\omega \;\;\lambda \\
\lambda \;\;\lambda \;\;\lambda \;\;\omega \\
\end{array} \right), \eqno(9)
\]
where $ \omega \geq \lambda $. The smaller the difference between $ \omega $ and $ \lambda $, the vaguer the network structures. $ \omega = \lambda $ means a fully random network with no group structure. The 10 attributes that have a strong relationship with each community were produced by $ Bin(1, p) $ with $ p=0.3, 0.5, 0.7 $, or 0.9. The results were shown in FIG.\ref{fig2}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.9\textwidth]{nmi4onarti.pdf}
\caption{The performance of GNAN on networks with traditional communities. In this group of studies, $ \lambda \equiv 0.02 $, $ \omega = 0.02, 0.04, 0.06, 0.08, 0.10$, the corresponding network structures are gradually clear, the sizes of 4 communities are 80, 100, 120, and 200, respectively. (a) A tested network with community structure ($ \lambda = 0.02,\;\omega = 0.10 $) \textemdash a set of communities with dense internal connections and sparse external ones, ``1, 2, 3", and ``4" are 4 communities, respectively. (b) An attribute matrix used in this group of tests ($ p=0.7 $), 10 attributes are matched to each community. (c) Both links and attributes were used (Link+Attr), only links were used (Link). (d) Only attributes were used (Attr). ``p=0.3, 0.5, 0.7, 0.9" measure the strength that a community matches attributes.}\label{fig2}
\end{figure}
From FIG.\ref{fig2} (c), we can easily see that (1) community detection results with both topology and attributes are better than ones without attributes; (2) the clearer the network structure, the better the detection effect; (3) the closer the relationship between the community and the attribute, the better the detection accuracy. These conclusions are consistent with our intuition, which shows the node-attribute information can help to improve community detection. Especially when the network has no community structure ($ \omega = \lambda =0.02 $), the results of community detection are entirely determined by attributes (i.e., $ \phi_{rk} $). Similar conclusions are shown in FIG.\ref{fig2} (d), where links are not considered, community detection is controlled by attributes ($ \phi_{rk} $), and the larger the $ p $, the better the community detection.
\textbf{Disassortative structure.} A group of five networks including disassortative structure was generated by SBM with parameter
\[
\omega ^{{\rm{planted}}} = \left( {\begin{array}{*{20}c}
{0.05} & {\lambda _1 } & {\lambda _1 + 0.1} \\
{\lambda _1 } & {0.03} & {\lambda _1 + 0.05} \\
{\lambda _1 + 0.1} & {\lambda _1 + 0.05} & {0.02} \\
\end{array}} \right) \eqno(10)
\]
where $ \lambda_1 > 0.05 $. The smaller the $ \lambda_1 $, the vaguer the network structures. The results were shown in FIG. \ref{fig3}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.9\textwidth]{nmi4ondi.pdf}
\caption{The performance of GNAN on networks with disassortative communities. In this group of studies, $ \lambda_1 = 0.10, 0.15, 0.20, 0.25, 0.30$, the corresponding network structures are gradually clear, the sizes of 3 communities are 100, 150, and 250, respectively. (a) An illustration with disassortative structure ($ \lambda_1 = 0.20 $), in which nodes have most of their connections outside their community. (b) An attribute matrix ($ p=0.7 $), 10 attributes are matched to each community. (c) Both links and attributes were used (Link+Attr), only links were used (Link). (d) Only attributes were used (Attr). Other symbols are the same as the ones in FIG. \ref{fig2}.}\label{fig3}
\end{figure}
As in FIG. \ref{fig3} (c-d), the same conclusion as in FIG. \ref{fig2} can be derived. The combination of attribute information and topological information improves the accuracy of community discovery. These conclusions are in line with expectations.
\textbf{Mixture structure.} A group of five networks including bipartite structure, community structure and core-periphery structure was generated by SBM with parameter
\[
\omega ^{{\rm{planted}}} = \left( \begin{array}{l}
0\;\;\omega _1 \;\;\lambda \;\;\lambda \;\;\lambda \\
\omega _1 \;\;0\;\;\;\lambda \;\;\lambda \;\;\lambda \\
\lambda \;\;\lambda \;\;\omega _2 \;\;\lambda \;\;\lambda \\
\lambda \;\;\lambda \;\;\lambda \;\;\omega _3 \;\;\omega _4 \\
\lambda \;\;\lambda \;\;\lambda \;\;\omega _4 \;\;0 \\
\end{array} \right) \eqno(11)
\]
to evaluate the performance of our new model. The 10 attributes that have a strong relationship with each community were produced by $ Bin(1, p) $ with $ p=0.3, 0.5, 0.7 $, or 0.9. The results were shown in FIG. \ref{fig4}.
\begin{figure}[htb]
\centering
\includegraphics[width=0.9\textwidth]{nmim4onarti.pdf}
\caption{The performance of GNAN on networks with mixture structures (bipartite structure, community structure, and core-periphery structure). The parameters corresponding to the 5 networks are as follows: $ \lambda \equiv 0.02 $, and network 1: $ \omega_1 = \omega_2 = \omega_3 = \omega_4 = 0.05$; network 2: $ \omega_1 = \omega_2 = \omega_3 = \omega_4 = 0.1$; network 3: $ \omega_1 = 0.1, \omega_2 = 0.2, \omega_3 = 0.4, \omega_4 = 0.1$; network 4: $ \omega_1 = 0.1, \omega_2 = 0.3, \omega_3 = 0.4, \omega_4 = 0.1$; network 5: $ \omega_1 = 0.1, \omega_2 = 0.5, \omega_3 = 0.4, \omega_4 = 0.1$. (a) An illustration with a mixture structure (network 3). Other symbols are the same as the ones in FIG.\ref{fig2}.} \label{fig4}
\end{figure}
From FIG. \ref{fig4} (c-d), on the networks with mixture structure, the performance of GNAN is almost the same as the performance on the network with community structure (FIG. \ref{fig2}) and disassortative structure (FIG. \ref{fig3}). These phenomena in both FIG. \ref{fig2} and FIG. \ref{fig3} mean that our new model GNAN can effectively use the node-attribute information to improve the community detection on networks with generalized structures.
The above experiments were mainly designed to test community detection. In order to evaluate whether our model GNAN can automatically discover important node attributes and ignore attributes without important information, we designed the following group of experiments.
\textbf{The ability to discover attribute information.} The network used here was the one in FIG. \ref{fig2} ($ \lambda = 0.02, \omega = 0.10 $). The attributes corresponding to each community were designed as follows.
\begin{table}[H]
\centering
\caption{The dependency between communities and attributes. ``-" represents that the 10 attributes were generated by $ Bin(1, p) $ with $ p=0.1 $, which means that these 10 attributes are noisy for the corresponding community. Note that attributes 31-40 are noisy for all communities.}
\resizebox{0.7\textwidth}{!}{
\begin{tabular}{c c c c c}
\hline
Attributes & 1-10 & 11-20 & 21-30 & 31-40 \\
\hline
Community\_1& Strong ($ p=0.9 $) & Strong ($ p=0.9 $) & - &- \\
Community\_2 & Strong ($ p=0.9 $) & Strong ($ p=0.9 $) & - & - \\
Community\_3 & - & - & Strong ($ p=0.7 $) & - \\
Community\_4 & - & - & Strong ($ p=0.7 $) & - \\
\hline
\end{tabular}%
}
\label{tab1}%
\end{table}%
Using this node-attribute network, the inferred dependencies by our model GNAN were shown in FIG. \ref{fig5}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.9\textwidth]{phionarti.pdf}
\caption{The inferred $ \phi_{rk} $ which characters the ability to find attribute information. ``Ignored attributes" means that the 10 attributes are noisy for all communities such that they can be ignored when considering the importance of attributes.}\label{fig5}
\end{figure}
As can be seen in FIG. \ref{fig5}, for community\_1, the values of $ \phi_{rk} $ from 1 to 20 are significantly larger than the rest, which means the corresponding 20 attributes are discovered by the GNAN. This conclusion is in line with the truth in TAB. \ref{tab1}. For the other three communities, the same conclusions are derived. Note that attributes 31-40 are noisy for all communities in TAB. \ref{tab1}, the corresponding values of $ \phi_{rk} $ are always small compared to the ones of strongly dependent attributes. These phenomena show that the model GNAN has the ability to find important attributes, which is beneficial to the semantic interpretation of communities in practice. We will see this conclusion in the following case study.
\subsection{Semantic interpretation of communities: an example about the friendship between Lazega’s lawyers}
In this section, through a small real-world network, we revealed what the main characteristics of each community are and tried to explain why they became friends. This data set includes 71 attorneys and 575 links and comes from a network study of corporate law partnership that was carried out in a Northeastern US corporate law firm, referred to as SG \& R, 1988-1991 in New England. The dataset has various members' attributes as follows:
\begin{table}[htbp]
\centering
\resizebox{0.8\textwidth}{!}{
\begin{tabular}{l l}
\textbullet $ \; $seniority (range: 1-71) & $ \; $ \textbullet $ \; $status (1=partner; 2=associate) \\
\textbullet $ \; $gender (1=man; 2=woman) & $ \; $ \textbullet $ \; $office (1=Boston; 2=Hartford; 3=Providence) \\
\textbullet $ \; $years with the firm (range: 1-32) & $ \; $ \textbullet $ \; $age (range: 26-67) \\
\textbullet $ \; $practice (1=litigation; 2=corporate) & $ \; $ \textbullet $ \; $law school (1: harvard, yale; 2: ucon; 3: other) \\
\end{tabular}%
}
\label{tab0}%
\end{table}%
where office means the office in which they work and other names of items are self-explanatory. Because the number of communities is not given in advance in this friendship network, by maximizing the modularity measure (Q) over all possible partitions (Q$ _{max} $=0.4088), 4 ground-truth communities were obtained, where 2 isolated points (NO.44 and NO.47) were removed. The communities were shown in FIG. \ref{fig5}, where a color represents a community. Here, the attribute variables ``age" and ``years with the firm" were discretized as shown in TAB. \ref{tab2}.
\begin{table}[H]
\centering
\caption{The inferred $ \phi_{rk} $ for each community, and the ones that greater than 0.1 were in bold.}
\begin{tabular}{c|ccccc}
\hline
\multicolumn{2}{c}{Attributes} & Comminity\_1 & Comminity\_2 & Comminity\_3 & Comminity\_4 \\
\hline
\multirow{2}[2]{*}{status } & partner & \textbf{0.1479 } & 0.0990 & 0.0000 & 0.0443 \\
& associate & 0.0000 & 0.0338 & \textbf{0.1609 } & 0.0614 \\
\hline
\multirow{2}[2]{*}{gender} & man & \textbf{0.1361 } & \textbf{0.1334 } & 0.0817 & 0.0682 \\
& woman & 0.0000 & 0.0171 & 0.0560 & 0.0906 \\
\hline
\multirow{3}[2]{*}{office} & Boston & \textbf{0.1232 } & 0.0000 & \textbf{0.1321 } & \textbf{0.1065 } \\
& Hartford & 0.0000 & \textbf{0.2002 } & 0.0000 & 0.0000 \\
& Providence & 0.0063 & 0.0000 & 0.0000 & 0.0138 \\
\hline
\multirow{3}[2]{*}{age} & $ \leq $35 & 0.0000 & 0.0684 & \textbf{0.1111 } & 0.0000 \\
& 36-45 & 0.0000 & 0.0411 & 0.0366 & \textbf{0.1704 } \\
& $ \geq $ 46 & \textbf{0.1453 } & 0.0000 & 0.0000 & 0.0000 \\
\hline
\multirow{3}[2]{*}{years with the firm} & 1-4 & 0.0000 & 0.0499 & \textbf{0.1414 } & 0.0000 \\
& 5-9 & 0.0000 & 0.0301 & 0.0000 & \textbf{0.1540 } \\
& $ \geq $10 & \textbf{0.1554 } & 0.0359 & 0.0000 & 0.0000 \\
\hline
\multirow{2}[2]{*}{practice} & litigation & 0.0724 & 0.0771 & 0.0923 & 0.0924 \\
& corporate & 0.0659 & 0.0798 & 0.0434 & 0.0576 \\
\hline
\multirow{3}[2]{*}{law school} & harvard or yale & 0.0781 & 0.0000 & 0.0000 & 0.0227 \\
& ucon & 0.0291 & 0.0757 & 0.0836 & 0.0421 \\
& other & 0.0404 & 0.0584 & 0.0608 & 0.0759 \\
\hline
\end{tabular}%
\label{tab2}%
\end{table}%
\begin{figure}[htb]
\centering
\includegraphics[width=0.8\textwidth]{lazegafdandattr.pdf}
\caption{A real-world network with 71 attorneys and 575 links between them. Each node has 18 attributes that belong to 6 classes as shown in TAB. \ref{tab2}. By maximizing the modularity measure (Q) over all possible partitions (Q$ _{max} $=0.4088), 4 ground-truth communities were obtained, where 2 isolated points (NO.44 and NO.47) were removed. Different colors represent different communities. The semantic interpretation inferred by the new method was shown near the corresponding community.}\label{fig6}
\end{figure}
As can be seen in TAB. \ref{tab2}, taking 0.1 as a boundary, Community\_1 has the most attributes (5 attributes), while Community\_2 has the least (2 attributes), which shows that there may be many reasons why people become friends, or they may not need many reasons. From the perspective of attributes, the attribute variables ``practice " and ``law school" are not important to every community, on the contrary, ``office" is important to each community, which means that distance (or geographic location) is one of the important factors that affect people's friendships, while the specific works they are engaged in and the schools they used to attend are not so important for friendships in this network. These conclusions show the ability of our model GNAN to learn attributes automatically, which helps the semantic interpretation of each community. A visual semantic explanation was shown in FIG. \ref{fig5}.
From TAB. \ref{tab2} and FIG. \ref{fig5}, we can try to explain why people build their friendships. From the workplace, each community seems to be formed like this: firstly, people in office Hartford formed a community (Community\_2). Then, the rest people (mainly office Boston) were divided into three communities: Community\_4 contains middle-aged persons (36-45 years old) who have worked in the law firm for between 5 and 10 years; Community\_3 is mainly composed of young people (less than 35 years old) who have worked for no more than 5 years; people in Community\_1 seem to be friends because they have been in the company for a long time ($ \geq 10 $), have almost the same status (partner), are old ($ \geq 46 $), and work in the same office (Boston).
\subsection{Comparison of our model GNAN with other state-of-the-art models}
We showed our model GNAN for both community detection and the ability to find important attributes in the above experiments. Next, we would compare the new model GNAN with other 4 state-of-the-art models: PSB\_PG, NEMBP, BNPA, and PPSB\_DC, which are all probabilistic generative models and can detect generalized structure. The results were shown in the following TAB. \ref{tab3} and TAB. \ref{tab4}.
\begin{table}[htbp]
\centering
\caption{Comparison results (mean $ \pm $ error) on artificial networks with a range of network structures. The best results were in bold. The networks $ \omega $0.06 and $ \omega $0.04 are the ones $ \omega $=0.06 and $ \omega $=0.04 in FIG. \ref{fig2}. ``m3" and ``m4" mean the third and fourth networks in FIG. \ref{fig4}. ``d0.1" means the network $ \lambda_1=0.1 $ in FIG. \ref{fig3}. ``cp" means a network with core-periphery structure in FIG. \ref{fig4}, and parameter $ \omega_3=0.12, \omega_4=0.1$. ``$ \omega $0.1\_noisyAttr" is the network with noisy attributes in TAB. \ref{tab1}.}
\resizebox{\textwidth}{!}{
\begin{tabular}{llccccc}
\hline
Network & Structure & GANA & PSB\_PG & NEMBP & BNPA & PPSB\_DC \\
\hline
SBM\_$ \omega $0.06\_Attr0.9 & Community & \textbf{1.0000$ \pm $0.0000} & \textbf{1.0000$ \pm $0.0000} & 0.9960$ \pm $0.0000 & \textbf{1.0000$ \pm $0.0000} & 0.9485$ \pm $0.0728 \\
SBM\_$ \omega $0.04\_Attr0.9 & Community & \textbf{0.9922$ \pm $0.0048} & 0.9844$ \pm $0.0000 & 0.9127$ \pm $0.0947 & 0.9901$ \pm $0.0206 & 0.9382$ \pm $0.0564 \\
SBM\_m3\_Attr0.5 & Mixture & \textbf{0.9908$ \pm $0.0035} & 0.9378$ \pm $0.0782 & 0.9527$ \pm $0.0883 & 0.7569$ \pm $0.0175 & 0.5526$ \pm $0.0771 \\
SBM\_m4\_Attr0.5 & Mixture & 0.9798$ \pm $0.0459 & \textbf{0.9963$ \pm $0.0102} & 0.9777$ \pm $0.0629 & 0.86854$ \pm $0.0000 & 0.7577$ \pm $0.0077 \\
SBM\_d0.1\_Attr0.5 & Disassortative & \textbf{0.9416$ \pm $0.0909} & 0.9285$ \pm $0.1085 & 0.8046$ \pm $0.0000 & 0.8020$ \pm $0.0000 & 0.6428$ \pm $0.0160 \\
SBM\_d0.1\_Attr0.3 & Disassortative & \textbf{0.8236$ \pm $0.0855} & 0.7746$ \pm $0.1567 & 0.8052$ \pm $0.0000 & 0.8026$ \pm $0.0000 & 0.3791$ \pm $0.0000 \\
SBM\_cp\_Attr0.5 & Core-periphery & 0.9393$ \pm $0.0000 & 0.9519$ \pm $0.0000 & \textbf{1.0000$ \pm $0.0000} & 0.9426$ \pm $0.0027 & 0.6271$ \pm $0.0106 \\
SBM\_cp\_Attr0.4 & Core-periphery & \textbf{0.9162$ \pm $0.0000} & 0.9024$ \pm $0.0049 & 0.8006$ \pm $0.0210 & 0.9144$ \pm $0.0017 & 0.3726$ \pm $0.0062 \\
SBM\_$ \omega $0.1\_noisyAttr & Community & \textbf{0.9443$ \pm $0.0750} & 0.8295$ \pm $0.0684 & 0.9348$ \pm $0.0816 & 0.8970$ \pm $0.0000 & 0.6608$ \pm $0.0000 \\
\hline
\end{tabular}%
}
\label{tab3}%
\end{table}%
As in TAB. \ref{tab3}, on the networks with traditional structure, the performance of these five methods is good, especially the methods GANA, PSB\_PG, and BNPA. On the networks with mixture structure, the methods GANA, PSB\_PG, and BNPA are superior to the rest 2 methods. On the networks with disassortative structure and core-periphery structure, all these methods except PPSB\_DC are highly efficacious. Simply speaking, on most of the tested networks, the community detection quality of the new method is the best in all methods. On the contrary, the performance of the method PPSB\_DC is the worst.
\begin{table}[htbp]
\centering
\caption{Comparison results (mean $ \pm $ error) on real-world networks. The best results were in bold.}
\resizebox{\textwidth}{!}{
\begin{tabular}{llccccccccc}
\hline
Network & N & M & K & C & Structure & GANA & PSB\_PG & NEMBP & BNPA & PPSB\_DC \\
\hline
Cornell & 195 & 304 & 1703 & 5 & Community & \textbf{0.3505$ \pm $0.0533} & 0.3115$ \pm $0.0576 & 0.1890$ \pm $0.0416 & 0.2211$ \pm $0.0077 & 0.1257$ \pm $0.0110 \\
Texas & 187 & 328 & 1703 & 5 & Community & \textbf{0.3226$ \pm $0.0280} & 0.3072$ \pm $0.0362 & 0.3093$ \pm $0.0222 & 0.1922$ \pm $0.0296 & 0.1576$ \pm $0.0101 \\
Washington & 230 & 446 & 1703 & 5 & Community & \textbf{0.3433$ \pm $0.0414} & 0.3013$ \pm $0.0323 & 0.2085$ \pm $0.0407 & 0.1697$ \pm $0.0158 & 0.2818$ \pm $0.0444 \\
Wisconsin & 265 & 530 & 1703 & 5 & Community & \textbf{0.4200$ \pm $0.0258} & 0.3729$ \pm $0.0279 & 0.3004$ \pm $0.0427 & 0.2696$ \pm $0.0126 & 0.2272$ \pm $0.0407 \\
Cora & 2708 & 5429 & 1433 & 7 & Mixture & 0.3594$ \pm $0.0373 & 0.3442$ \pm $0.0382 & 0.4188$ \pm $0.0255 & \textbf{0.4780$ \pm $0.0303} & 0.4659$ \pm $0.0090 \\
Citeseer & 3312 & 4723 & 3703 & 6 & Mixture & 0.2606$ \pm $0.0296 & 0.2543$ \pm $0.0364 & 0.2325$ \pm $0.0192 & 0.2958$ \pm $0.0321 & \textbf{0.3753$ \pm $0.0382} \\
\hline
\end{tabular}%
}
\label{tab4}%
\end{table}%
From TAB. \ref{tab4}, on the first four networks with community structure, the methods GANA and PSB\_PG are superior to the other 2 methods. Instead, on networks with mixture structure (Cora and Citeseer), the methods BNPA and PPSB\_DC are superior to the rest four methods. The performance of various methods on the real and artificial networks is not completely consistent, which means that there is a gap between the computer-generated network structure and the real network structure. However, whether on synthetic networks or real networks, the community detection quality of the new method GANA is competitive with the other state-of-the-art methods.
In addition, the complexity of the PSB\_PG mentioned above is higher than that of our new method. We showed the comparison results on real networks (see FIG. \ref{fig7}). From FIG. \ref{fig7}, the new algorithm GNAN is superior to the algorithm PSB\_PG, especially on the latter two networks.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\textwidth]{complex.pdf}
\caption{Comparison of the complexity. In the two algorithms compared, the running parameters are the same as the original. The time and memory here were obtained when the corresponding algorithm reaches its stop condition. }\label{fig7}
\end{figure}
\section{Conclusions}
In summary, based on the connectivity behavior of nodes and homogeneity of attributes, we have developed a generative node-attribute network model GNAN that combines topological information and attribute information. The major contributions: (1) The new model GNAN can detect a range of network structures, by experiments in Sect.IV, we have shown this feature. (2) The node attributes that match the true community assignments of nodes can be automatically learned by our model. We have designed an experiment to test this ability of the model GNAN. By using this ability, a case study has been provided to show the semantic interpretability of communities. (3) The new model detects communities more accurately than the model that only uses topology information. Experiments on both synthetic and real-world networks have shown that the new model is competitive with other state-of-the-art models.
\section*{Acknowledgements}
This work is supported in part by the National Natural Science Foundation of China (granted No. 61876016), National Key R \& D Program of China (No. 2018AAA0100302), the Higher Education Innovation Ability Improvement Project in Gansu Province (No. 2020-098A), and the Research Project at Tianshui Normal University (No. CXJ2020-28). The authors thank the anonymous reviewers for their constructive comments.
\section*{Appendix A}
Note that the constraint on $ \tau_{ir} $: $ \sum\limits_{r = 1}^C {\tau _{ir} } = 1 $, and ignoring the constants not related to T, we have
\[
\tilde L\left( {\rm T} \right) = \sum\limits_{ijr} {a_{ij} q_{ij,r} \log \tau _{ir} } - \sum\limits_{ijr} {\tau _{ir} \theta _{rj} } + \sum\limits_{ikr} {x_{ik} h_{ik,r} \log \tau _{ir} } - \sum\limits_{ikr} {\tau _{ir} \varphi _{rk} } + \sum\limits_i {\lambda _i \left( {\sum\limits_r {\tau _{ir} } - 1} \right)}. \eqno(A1)
\]
Taking the first-order partial derivative of the Lagrangian $\tilde L\left( {\rm T} \right)$ with respect to $\tau_{ir}$, we have
\[
\frac{{\partial \tilde L\left( {\rm T} \right)}}{{\partial \tau _{ir} }} = \frac{{\sum\limits_j {a_{ij} q_{ij,r} } }}{{\tau _{ir} }} - \sum\limits_j {\theta _{rj} } + \frac{{\sum\limits_k {x_{ik} h_{ik,r} } }}{{\tau _{ir} }} - \sum\limits_k {\phi _{rk} } + \lambda _i. \eqno(A2)
\]
Set $ \frac{{\partial \tilde L\left( {\rm T} \right)}}{{\partial \tau _{ir} }} = 0 $, we have
\[
\frac{{\sum\limits_j {a_{ij} q_{ij,r} } + \sum\limits_k {x_{ik} h_{ik,r} } }}{{\tau _{ir} }} - 2 + \lambda _i {\rm{ = 0}}, \eqno(A3)
\]
\[
\sum\limits_j {a_{ij} q_{ij,r} } + \sum\limits_k {x_{ik} h_{ik,r} } - 2\tau _{ir} + \lambda _i \tau _{ir} {\rm{ = 0}}, \eqno(A4)
\]
\[
\sum\limits_{jr} {a_{ij} q_{ij,r} } + \sum\limits_{kr} {x_{ik} h_{ik,r} } - 2\sum\limits_r {\tau _{ir} } + \lambda _i \sum\limits_r {\tau _{ir} } {\rm{ = 0}}. \eqno(A5)
\]
By $ (A3) $ and $ (A5) $, we have
\[
\tau _{ir} = \frac{{\sum\limits_j {a_{ij} q_{ij,r} } + \sum\limits_k {x_{ik} h_{ik,r} } }}{{\sum\limits_{jr} {a_{ij} q_{ij,r} } + \sum\limits_{kr} {x_{ik} h_{ik,r} } }}. \eqno(A6)
\]
Similarly, by Eq.(5) and the constraint $ \sum\limits_{j = 1}^N {\theta _{rj} } = 1 $, we have
\[
\tilde L\left( \Theta \right) = \sum\limits_{ijr} {a_{ij} q_{ij,r} \log \theta _{rj} - } \sum\limits_{ijr} {\tau _{ir} \theta _{rj} } + \sum\limits_r {\eta _r \left( {\sum\limits_j {\theta _{rj} } - 1} \right)}. \eqno(A7)
\]
Taking the first order partial derivative of the Lagrangian $\tilde L\left( \Theta \right)$ with respect to $\theta _{rj}$ and set it to be zero, we have
\[
\frac{{\sum\limits_i {a_{ij} q_{ij,r} } }}{{\theta _{rj} }} - \sum\limits_i {\tau _{ir} } + \eta _r = 0. \eqno(A8)
\]
\[
\sum\limits_{ij} {a_{ij} q_{ij,r} } - \sum\limits_j {\theta _{rj} } \sum\limits_i {\tau _{ir} } + \sum\limits_j {\eta _r \theta _{rj} } = 0. \eqno(A9)
\]
By $ (A8) $ and $ (A9) $, we have
\[
\theta _{rj} = \frac{{\sum\limits_i {a_{ij} q_{ij,r} } }}{{\sum\limits_{ij} {a_{ij} q_{ij,r} } }}. \eqno(A10)
\]
Finally, using the condition $ \sum\limits_{k = 1}^K {\phi _{rk} } = 1 $ and Eq.(5), we have
\[
\tilde L\left( \Phi \right) = \sum\limits_{ikr} {x_{ik} h_{ik,r} \log \phi _{rk} } - \sum\limits_{ikr} {\tau _{ir} \phi _{rk} } + \sum\limits_r {\rho _r \left( {\sum\limits_k {\phi _{rk} } - 1} \right)}. \eqno(A11)
\]
Taking the first derivative of the Lagrangian $\tilde L\left( \Phi \right)$ with respect to $\phi _{rk}$ and set it to be zero, we have
\[
\frac{{\sum\limits_i {x_{ik} h_{ik,r} } }}{{\phi _{rk} }} - \sum\limits_i {\tau _{ir} } + \rho _r = 0. \eqno(A12)
\]
\[
\sum\limits_{ik} {x_{ik} h_{ik,r} } - \sum\limits_{ik} {\tau _{ir} \phi _{rk} } + \sum\limits_k {\eta _r \phi _{rk} } = 0. \eqno(A13)
\]
By $ (A12) $ and $ (A13) $, we have
\[
\phi _{rk} = \frac{{\sum\limits_i {x_{ik} h_{ik,r} } }}{{\sum\limits_{ik} {x_{ik} h_{ik,r} } }}. \eqno(A14)
\]
|
\section{Introduction}
A further understanding of causality beyond observational data is critical across many domains including statistics, computer science, education, public policy, economics, and health care. Although randomized controlled trials (RCT) are usually considered as the gold standard for causal inference, estimating causal effects from observational data has received growing attention owing to the increasing availability of data and the low costs compared to RCT.
When estimating treatment effects from observational data, we face two major issues, i.e., missing counterfactual outcomes and treatment selection bias. The foremost challenge for solving these two issues is the existence of confounders, which are the variables that affect both treatment assignment and outcome. Unlike RCT, treatments are typically not assigned at random in observational data. Due to the confounders, subjects would have a preference for a certain treatment option, which leads to a bias of the distribution for the confounders among different treatment options. This phenomenon exacerbates the difficulty of counterfactual outcome estimation. For most of existing methods~\cite{hill2011bayesian,chu2020matching, li2016matching, li2017matching, shalit2017estimating, wager2018estimation, yao2018representation, alaa2017bayesian, yao2019estimation}, the \emph{strong ignorability} assumption is the most important prerequisite. It assumes given covariates, the treatment assignment is independent of the potential outcomes and for any value of covariates, treatment assignment is not deterministic. Strong ignorability is also known as the \emph{no unmeasured confounders} assumption. This assumption requires that all the confounders be observed and sufficient to characterize the treatment assignment mechanism. Moreover, strong ignorability is a sufficient condition for the individual treatment effect (ITE) function to be identifiable~\cite{imbens2009recent}.
However, due to the fact that identifying all of the confounders is impossible in practice, the strong ignorability assumption is usually untenable. By leveraging big data, it becomes possible to find a proxy for the hidden confounders. Network information, which serves as an efficient structured representation of non-regular data, is ubiquitous in the real world. Advanced by the powerful representation capabilities of various graph neural networks, networked data has recently received increasing attention~\cite{kipf2016semi,velivckovic2017graph,velickovic2019deep,jiang2019censnet}. Besides, it can be used to help recognize the patterns of hidden confounders. A network deconfounder~\cite{guo2019learning} is proposed to recognize hidden confounders by combining the graph convolutional networks~\cite{kipf2016semi} and counterfactual regression~\cite{shalit2017estimating}.
\begin{figure}[t]
\centering
\includegraphics[width=0.3\textwidth]{samples/example1.PNG}
\vspace{-2mm}
\caption{Example of the imbalance of network structure.}
\label{exmaple}
\vspace{-2mm}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=0.85\textwidth]{samples/imbalance.PNG}
\vspace{-2mm}
\caption{Under the assumption that each node has the same possibility to be connected with another node regardless of node's treatment assignment, for $n$ nodes, there should be $\frac{n^2}{4}-\frac{n}{2}$ homogeneous edges (that link the nodes in the same group, i.e., treatment-treatment or control-control) and $\frac{n^2}{4}$ heterogeneous edges (that link the nodes in different groups, i.e., treatment-control). The number of heterogeneous edges should be greater than that of homogeneous edges. However, in the benchmarks of causal inference with networked data (\textit{BlogCatalog} and \textit{Flickr}), the homogeneous edges are consistently greater than heterogeneous edges for both datasets. Besides, as the selection bias increases, the difference between homogeneous and heterogeneous edges gets larger. This result totally agrees with our expectation that, in the causal inference problem, the network structure is imbalanced. The relationship is more likely to appear among people who are in the same group.}
\label{imbalance}
\end{figure*}
The networked observational data consists of two components, node features and network structures. Due to the confounding bias in causal inference problem, the imbalance not only exists in distributions of feature variables in treatment and control groups but also in network structures. For example, in social networks, the links are more likely to appear among more similar people, so the subjects are more likely to follow other subjects in the same group as shown in Fig.~\ref{exmaple}, which will aggravate the imbalance in the representation space learned by graph neural networks. Fig.~\ref{imbalance} shows the existence of imbalanced network structures in the benchmarks of causal inference with networked data (\textit{BlogCatalog} and \textit{Flickr}). Unlike the networked data in traditional graph learning tasks, such as node classification and link detection, the networked data under the causal inference problem has its particularity, i.e., imbalanced network structure. For most existing work on networked observational data, they did not consider this peculiarity of graph structure under causal inference settings. Directly applying graph neural networks designed for traditional graph learning tasks cannot capture all of the information from imbalanced networked data.
To fully exploit the information in the networked data with the imbalanced network structure, we propose a Graph Infomax Adversarial Learning method (GIAL) to estimate the treatment effects from networked observational data. In our model, structure mutual information is maximized to help graph neural networks to extract a representation space, which best represents observed and hidden confounders from the networked data with the imbalanced structure. Also, adversarial learning is applied to balance the learned representation distributions of treatment and control groups and to generate the potential outcomes for each unit across two groups. Overall, GIAL can make full use of network structure to recognize patterns of hidden confounders, which has been validated by extensive experiments on benchmark datasets.
\section{Background}
Suppose that the observational data contain $n$ units and each unit received one of two or more treatments. Let $t_i$ denote the treatment assignment for unit $i$; $i=1,...,n$. For binary treatments, $t_i = 1$ is for the treatment group, and $t_i=0$ for the control group. The outcome for unit $i$ is denoted by $Y_{t}^i$ when treatment $t$ is applied to unit $i$; that is, $Y_{1}^i$ is the potential outcome of unit $i$ in the treatment group and $Y_0^i$ is the potential outcome of unit $i$ in the control group. For observational data, only one of the potential outcomes is observed according to the actual treatment assignment of unit $i$. The observed outcome is called the factual outcome, and the remaining unobserved potential outcomes are called counterfactual outcomes. Let $X \in \mathbb{R}^d$ denote all observed variables of a unit.
Let $\mathcal{G}(\mathcal{V},\mathcal{E})$ denote an undirected graph, where $\mathcal{V}$ represents $n$ nodes in $\mathcal{G}$ and $\mathcal{E}$ is a set of edges between nodes. According to the adjacency relationships in $\mathcal{E}$, the corresponding adjacent matrix $A \in \mathbb{R}^{n \times n}$ of the graph $\mathcal{G}$ can be defined as follows. If $(v_i,v_j) \in \mathcal{E}$, $A_{ij} = 1$, otherwise $A_{ij}=0$. When edges have different weights, $A_{ij}$ can be assigned to a real value.
In this paper, we explore the observational data as networks. In particular, the graph $\mathcal{G}$ is the networked observational data. Every node in $\mathcal{V}$ is one unit in observational data, an edge in $\mathcal{V}$ describes the relationship between a pair of units, and adjacent matrix $A$ represents the whole network structure. Therefore, the observational data can be denoted as $(\{x_i, t_i, y_i\}_{i=1}^n,A)$. We follow the potential outcome framework for estimating treatment effects ~\cite{rubin1974estimating}. The individual treatment effect (ITE) for unit $i$ is the difference between the potential treated and control outcomes, which is defined as: $ \text{ITE}_i = Y_1^i - Y_0^i, \quad (i=1,...,n).$
The average treatment effect (ATE) is the difference between the mean potential treated and control outcomes, which is defined as $\text{ATE}=\frac{1}{n}\sum_{i=1}^{n}(Y_1^i - Y_0^i), \quad (i=1,...,n).$ The success of the potential outcome framework is based on the strong ignorability assumption, which ensures that the treatment effect can be identified~\cite{imbens2015causal, yao2020survey}.
\begin{assumption} \textbf{Strong Ignorability}:
Given covariates $X$, treatment assignment $T$ is independent of the potential outcomes, i.e., $(Y_1, Y_0) \protect\mathpalette{\protect\independenT}{\perp} T | X$ and for any value of $\,X$, treatment assignment is not deterministic, i.e.,$P(T = t | X = x) > 0$, for all $t$ and $x$.
\end{assumption}
In our model, we relax the strong ignorability and allow the existence of hidden confounders. We aim to use network structure information to recognize the hidden confounders and then estimate treatment effects based on the learned confounder representations.
\section{The Proposed Framework}
\subsection{Motivation}
The foremost challenge of causal inference from observational data is how to recognize hidden confounders. Recently, leveraging the powerful representation capabilities of various graph neural networks, network structures can be utilized to help recognize the patterns of hidden confounders in networked observational data.
Due to the particularity of the causal inference problem, the networked data in causal inference is different from that in traditional graph learning tasks such as node classification and link detection. As network information is incorporated into the model, we face a new imbalance issue,i.e., imbalance of network structure in addition to the imbalance of observed covariate distributions. A link has a larger probability of appearing between two more similar people. It implies that one unit is more likely to be connected to other units in the same group. Therefore, directly applying traditional graph learning methods to learn the representation of networked data could not fully exploit the useful information for causal inference.
It is essential to design a new method that can capture the representation of hidden confounders implied from the imbalanced network structure and observed confounders that exist in the covariates simultaneously. To solve this problem, we propose the Graph Infomax Adversarial Learning method (GIAL) to estimate the treatment effects from the networked observational data, which can recognize patterns of hidden confounders from imbalanced network structure.
\begin{figure*}[h]
\centering
\includegraphics[width=0.8\textwidth]{samples/framework.PNG}
\vspace{-2mm}
\caption{Framework of our Graph Infomax Adversarial Learning method (GIAL). Graph neural networks and structure mutual information are utilized to learn the representations of hidden confounders and observed confounders. Then the potential outcome generator is applied to infer the potential outcomes of units across treatment and control groups based on the learned representation space and treatment assignment. At the same time, the counterfactual outcome discriminator is incorporated to remove the imbalance in the learned representations of treatment and control groups.}
\vspace{-2mm}
\label{framework}
\end{figure*}
\subsection{Model Architecture}
As shown in Fig.~\ref{framework}, our GIAL consists of four main components, i.e., confounder representation learning, structure mutual information maximization, potential outcome generator, and counterfactual outcome discriminator. Firstly, we utilize the graph neural network and structure mutual information to learn the representations of hidden confounders and observed confounders, by mapping the feature covariates and network structure simultaneously into a representation space. Then the potential outcome generator is applied to infer the potential outcomes of units across treatment and control groups based on the learned representation space and treatment assignment. At the same time, the counterfactual outcome discriminator is incorporated to remove the imbalance in the learned representations of treatment and control groups, and thus it improves the prediction accuracy of potential outcomes inferred in the outcome generator by playing a minimax game. In the following, we present the details of each component.
\textbf{Confounder Representation Learning.}
Based on the graph $\mathcal{G}(\mathcal{V},\mathcal{E})$, our goal is to learn the representation of confounders by a function $g : X \times A \rightarrow R, R \in \mathbb{R}^d$, which is parameterized by a graph neural network. To better capture information resided in the networked data, we separately adopt two powerful graph neural network methods, i.e., the graph convolutional network (GCN) ~\cite{kipf2016semi} and graph attention network layers (GAT) ~\cite{velivckovic2017graph}, to learn the representation space. For these two models, their effectiveness of the learned representations has been verified in various graph learning tasks. The major difference between GCN and GAT is how the information from the one-hop neighborhood is aggregated. For GCN, a graph convolution operation is used to produce the normalized sum of the node features of neighbors. GAT introduces the attention mechanism to better quantify the importance of each edge. Here, we want to find out which model is better to unravel patterns of hidden confounders from the networked data with imbalanced covariate and imbalanced network structure.
For the graph convolutional network (GCN) model, the representation learning function $g : X \times A \rightarrow R$ is parameterized with the following layer-wise propagation rule:
\begin{equation}
r^{(l+1)}=\sigma (\Tilde{D}^{-\frac{1}{2}}\Tilde{A}\Tilde{D}^{-\frac{1}{2}}r^{(l)}W^{(l)}),
\end{equation}
where $\Tilde{A}=A+I_n$ is the adjacency matrix of graph $\mathcal{G}(\mathcal{V},\mathcal{E})$ with inserted self-loops, i.e., the identity matrix $I_n$. $\Tilde{D}$ is its corresponding degree matrix, i.e., $\Tilde{D}_{ii} = \sum_{j}\Tilde{A}_{ij}$ and $W^{(l)}$ is a layer-specific trainable weight matrix. $\sigma(\cdot)$ denotes an activation function and here we apply the parametric ReLU (PReLU) function ~\cite{he2015delving}. A number of GCN layers can be stacked to approximate the function $g : X \times A \rightarrow R$.
For the graph attention network (GAT) model, the representation of confounder for the $i$-th node is a function of its covariates and receptive field. Here, we define the $i$-th node $v_i$ and its one-hop neighbor nodes as the receptive field $\mathcal{N}(v_i)$. The representation learning function $g : X \times A \rightarrow R$ is parameterized with the following equation:
\begin{equation}
r^{(l+1)}_i=\sigma\left(\sum_{j\in \mathcal{N}(i)} {\alpha^{(l)}_{ij} W^{(l)}r^{(l)}_j}\right),
\label{eq-gat}
\end{equation}
where $W^{(l)}$ is the learnable weight matrix and $W^{(l)}r^{(l)}_j$ is a linear transformation of the lower layer representation $r^{(l)}_j$. $\sigma(\cdot)$ is the activation function for nonlinearity. In Eq.~\eqref{eq-gat}, the representation of the $i$-th node and its neighbors are aggregated together, scaled by the normalized attention scores $\alpha^{(l)}_{ij}$.
\begin{equation}
\alpha^{(l)}_{ij}=\frac{\exp(\text{LeakyReLU}( {a^{(l)}}^T(W^{(l)}r^{(l)}_i||W^{(l)}r^{(l)}_j)))}{\sum_{k\in \mathcal{N}(i)}^{}\exp(\text{LeakyReLU}( {a^{(l)}}^T(W^{(l)}r^{(l)}_i||W^{(l)}r^{(l)}_k)))},
\end{equation}
where softmax is used to normalize the attention scores on each node's incoming edges. The pair-wise attention score between two neighbors is calculated by $\text{LeakyReLU}(a^{(l)^T}(W^{(l)}r^{(l)}_i||W^{(l)}r^{(l)}_j))$. Here, it first concatenates the linear transformation of the lower layer representations for two nodes, i.e., $W^{(l)}r^{(l)}_i||W^{(l)}r^{(l)}_j$, where $||$ denotes concatenation, and then it takes a dot product of itself and a learnable weight vector $a^{(l)}$. Finally, the LeakyReLU function is applied.
To stabilize the learning process, a multi-head attention mechanism is employed. We compute multiple different attention maps and finally aggregate all the learned representations. In particular, $K$ independent attention mechanisms execute the transformation of Eq.~\eqref{eq-gat}, and then their outputs are merged in two ways:
\begin{equation}
\text{concatenation}: r^{(l+1)}_{i} =||_{k=1}^{K}\sigma\left(\sum_{j\in \mathcal{N}(i)}\alpha_{ij}^{k}W^{k}r^{(l)}_{j}\right)
\end{equation}
or
\begin{equation}
\text{average}: h_{i}^{(l+1)}=\sigma\left(\frac{1}{K}\sum_{k=1}^{K}\sum_{j\in\mathcal{N}(i)}\alpha_{ij}^{k}W^{k}h^{(l)}_{j}\right)
\end{equation}
When performing the multi-head attention on the final layer of the network, concatenation is no longer sensible. Thus, we use the concatenation for intermediary layers and the average for the final layer. An arbitrary number of GAT layers can be stacked to approximate the function $g: X \times A \rightarrow R$.
\textbf{Structure Mutual Information Maximization.}
Inspired by a recent successful unsupervised graph learning method ~\cite{velickovic2019deep}, we maximize structure mutual information to capture the imbalanced graph structure with respect to treatment and control nodes in the networked observational data. We aim to learn representations that can capture the imbalanced structure of the entire graph. Specifically, we utilize a structure summary function, $f : R \rightarrow S, S \in \mathbb{R}^d$, to summarize the learned representation into an entire graph structure representation, i.e., $S=f(g(X,A))$. From the observations in empirical evaluations, the structure summary function could be defined as $s = \sigma(\frac{1}{n}\sum_{i=1}^{n}r_i)$ to best capture the entire graph structure, where $\sigma$ is the logistic sigmoid activation function.
Here, our purpose is to learn a representation vector, which can capture the entire graph structure encoded by the graph structure summary vector $s$ and also reflect the abnormal imbalance in the graph structure. Therefore, we aim at maximizing the mutual information between the learned representation vector $r_i$ and the structure summary vector $s$.
Mutual information is a fundamental quantity for measuring the relationship between random variables. For example, the dependence of two random variables $W$ and $Z$ is quantified by mutual information as~\cite{belghazi2018mutual}:
\begin{equation}
I(W;Z)=\int_{\mathcal{W}\times\mathcal{Z}}{\text{log}\frac{d\mathbb{P}_{WZ}}{d\mathbb{P}_W\otimes\mathbb{P}_Z}d\mathbb{P}_{WZ}},
\end{equation}
where $\mathbb{P}_{WZ}$ is the joint probability distribution, and $\mathbb{P}_W=\int_{\mathcal{W}}d\mathbb{P}_{WZ}$ and $\mathbb{P}_Z=\int_{\mathcal{Z}}d\mathbb{P}_{WZ}$ are the marginals. However, mutual information has historically been difficult to compute. From the viewpoint of Shannon
information theory, mutual information can be estimated as Kullback-Leibler divergence:
\begin{equation}
I(W;Z)=H(W)-H(W|Z)=D_{KL}(\mathbb{P}_{WZ} || \mathbb{P}_W\otimes\mathbb{P}_Z).
\end{equation}
Actually, in our model, it is unnecessary to use the exact KL-based formulation of MI, as we only want to maximize the mutual information between representation vector $r_i$ and structure summary vector $s$. A simple and stable alternative based on the Jensen-Shannon divergence (JSD) can be utilized. Thus, we follow the intuitions from deep infomax~\cite{hjelm2018learning} and deep graph infomax ~\cite{velickovic2019deep} to maximize the mutual information.
To act as an agent for maximizing the mutual information, one discriminator $d: R \times S \rightarrow P, P \in \mathbb{R}$ is employed. The discriminator is formulated by a simple bilinear scoring function with nonlinear activation: $d(r_i,s)=\sigma({r_i}^TWs)$, which estimates the probability of the $i$-th node representation contained within the graph structure summary $s$. $W$ is a learnable scoring matrix.
To implement the discriminator, we also need to create the negative samples compared with original samples and then use the discriminator to distinguish which one is from positive samples (original networked data) and which one is from the negative samples (created fake networked data), such that the original graph structure information could be correctly captured. The choice of the negative sampling procedure will govern the specific kinds of structural information that is desirable to be captured ~\cite{velickovic2019deep}. Here, we focus on the imbalance between the edges that link nodes in the same group and those that link nodes in the different groups, i.e., treatment unit to treatment unit, treatment unit to control unit, and control unit to control unit. Therefore, our discriminator is designed to force the representations to capture this imbalanced structure by creating negative samples where the original adjacency matrix $A$ is preserved, whereas the negative samples $\Tilde{X}$ are obtained by the row-wise shuffling of $X$. That is, the created fake networked data consists of the same nodes as the original graph, but they are located in different places in the same structure. Thus, the nodes at both ends of the edges may change the treatment choices, e.g., from treatment to control, from control to treatment, or remain unchanged. Then we also conduct the confounder representation learning for the created fake networked data $(\Tilde{X}, A)$ to get the $\Tilde{r}_i$. With the proposed discriminator, we could have $d(r_i,s)$ and $d(\Tilde{r}_i,s)$, which indicate the probabilities of containing the representations of the $i$-th positive sample and negative sample in the graph structure summary, respectively.
We optimize the discriminator to maximize mutual information between $r_i$ and $s$ based on the Jensen Shannon divergence via a noise-contrastive type objective with a standard binary cross-entropy (BCE) loss~\cite{velickovic2019deep,hjelm2018learning}:
\begin{equation}
\mathcal{L}_m = \frac{1}{2n}\Big(\sum_{i=1}^n \mathbb{E}_{(X,A)}[\text{log} \, d(r_i,s)]+\sum_{j=1}^n \mathbb{E}_{(\Tilde{X},A)}[\text{log} \, (1-d(\Tilde{r}_i,s))]\Big).
\label{eq: mutual information}
\end{equation}
\textbf{Potential Outcome Generator.}
So far, we have learned the representation space of confounders from networked data with the imbalanced network structure and imbalanced covariates. The function $\Psi: R \times T \rightarrow Y$ maps the representation of hidden confounders and observed confounders as well as a treatment to the corresponding potential outcome, which is parameterized by a feed-forward deep neural network with multiple hidden layers and non-linear activation functions. The function $\Psi: R \times T \rightarrow Y$ uses representations and treatment options as inputs to predict potential outcomes. The output of $\Psi$ estimates potential outcomes across treatment and control groups, including the estimated factual outcome $\hat{y}^f$ and the estimated counterfactual outcomes $\hat{y}^{cf}$. The factual outcomes $y^f$ are used to minimize the loss of prediction $\hat{y}^f$. We aim to minimize the mean squared error in predicting factual outcomes:
\begin{equation}
\mathcal{L}_\Psi = \frac{1}{n}\sum_{i=1}^{N}(\hat{y}^f_i-y^f_i)^2,
\label{eq: factual}
\end{equation}
where $\hat{y}_i=\Psi(r_i,t_i)$ denotes the inferred observed outcome of unit $i$ corresponding to the factual treatment $t_i$.
\textbf{Counterfactual Outcome Discriminator.} The counterfactual outcome discriminator is intended to remove the imbalance of confounder representations between treatment and control groups, and thus it could improve the prediction accuracy of potential outcomes inferred by the outcome generator. We define the counterfactual outcome discriminator as $\Phi: R \times T \times (Y^f \text{or} \ \hat{Y}^{cf}) \rightarrow P$, where $P$ is the discriminator's judgement, i.e., probability that this outcome for unit $i$ given $R$ and $T$ is factual outcome. $P$ is defined as:
\begin{equation}
P=
\begin{cases}
P(\text{judges} \ y^f \text{as factual} |x,t) \, \text{if} \ t \ \text{is factual treatment choice} \\
P(\text{judges} \ \hat{y}^{cf} \text{as factual} |x,t)\, \text{if} \ t \ \text{is not factual treatment choice}. \\
\end{cases}
\label{Eqn: p}
\end{equation}
To improve the accuracy of prediction and avoid risk of losing the influence of treatment $t$ and potential outcomes $(y^f\text{or} \ \hat{y}^{cf})$ due to high dimensional representation vector, we adopt separate head networks for treatment and control groups ~\cite{shalit2017estimating}. Besides, to improve the influence of $(y^f,\hat{y}^{cf})$ in the discriminator, we add $(y^f \text{or} \ \hat{y}^{cf})$ into each layer of the neural network, repetitively.
The discriminator deals with a binary classification task, which assigns one label (i.e., factual outcome or counterfactual outcome) to the vector concatenating the representation vector $r$ and potential outcome $(y^f \text{or} \ \hat{y}^{cf})$ under the treatment head network and control head network, respectively. Thus, the loss of discrimination is measured by the cross-entropy with truth probability, where $P^{\text{truth}}=1$ if $y^f$ is input, and $P^{\text{truth}}=0$ if $\hat{y}^{cf}$ is input. In each iteration of training, we make sure to input the same number of units in the treatment and control groups to ensure that there exist the same number of factual outcomes as counterfactual outcomes in each head network to overcome the imbalanced classification. The inputs of discriminator are generated by the outcome generator $\Psi(R,T)$, and then the cross entropy loss of the counterfactual outcome discriminator is defined as:
\begin{equation}
\begin{split}
\mathcal{L}_{\Phi,\Psi} = &-\frac{1}{2n}\sum_{t=0}^{1}\sum_{i=1}^{n}(p^{\text{truth}}_{ti}\log(p_{ti})+(1-p^{\text{truth}}_{ti})\log(1-p_{ti})),
\label{Eqn: observed and potential outcome}
\end{split}
\end{equation}
where $p^{\text{truth}}_{ti}$ is the indicator that this input outcome for unit $i$ under treatment option $t$ is the observed factual outcome or inferred outcome from generator module, i.e., $p^{\text{truth}}_{ti}$ equals 1 or 0, separately. $P_{ti}$ is the probability judged by discriminator that how likely this input outcome for unit $i$ under treatment option $t$ is a factual outcome.
Thus far, we have introduced the outcome generator to estimate potential outcomes for each unit across treatment and control groups, and the discriminator to determine if the potential outcome is factual, given a unit's confounder representation under treatment or control group. In the initial iterations of the model training, the outcome generator may generate potential outcomes that are very different from factual outcomes as determined by the discriminator. As the model is trained further, the discriminator may no longer be able to distinguish the generated counterfactual outcome and the factual outcome. At this point, we have attained all potential outcomes for each unit under treatment and control groups. For the training procedure of optimizing the outcome generator and discriminator, the minimax game is adopted. Putting all of the above together, the objective function of our Graph Infomax Adversarial Learning (GIAL) method is:
\begin{equation}
\begin{split}
\text{min}_{\Psi} \text{max}_{\Phi,m} \ ( \mathcal{L}_\Psi + \alpha \mathcal{L}_m -\beta \mathcal{L}_{\Phi,\Psi}),
\label{Eqn: minimax}
\end{split}
\end{equation}
where $\alpha$ and $\beta$ are the hyper-parameters controlling the trade-off among the outcome generator, mutual information, and discriminator.
\subsection{Overview of GIAL}
The proposed Graph Infomax Adversarial Learning method (GIAL) can estimate the treatment effects from networked observational data, which utilizes the graph neural network (GCN or GAT) and structure mutual information to learn the representations of hidden confounders and observed confounders, by mapping the feature covariates and network structure simultaneously into a representation space. Adversarial learning is also employed to mitigate the representation imbalance between treatment and control groups and to predict the counterfactual outcomes. After obtaining the counterfactual outcomes, GIAL can estimate the treatment effects.
We summarize the procedures of GIAL as follows:
\begin{enumerate}
\item Create the negative samples $(\Tilde{X}, A)$ by the row-wise shuffling of $X$ and keeping the original adjacency matrix $A$.
\item Learn the representation space $R$ for the positive samples $(X,A)$ by function $g : X \times A \rightarrow R$ by a graph neural network.
\item Learn the representation space $\Tilde{R}$ for the negative samples $(\Tilde{X},A)$ by function $g : \Tilde{X} \times A \rightarrow \Tilde{R}$ by the same graph neural network as Step 2.
\item Utilize a structure summary function $f : R^{n\times d} \rightarrow S$ to summarize the learned representation into a graph-level structure representation, i.e., $s=f(g(X,A))$.
\item Employ a discriminator $d: R \times S \rightarrow P$ to obtain $d(r_i,s)$ and $d(\Tilde{r}_i,s)$, which are the probabilities that the representations of $i$-th positive and negative samples are contained within the original graph structure summary $s$.
\item Utilize functions $g$, $f$ and $d$ to maximize mutual information between $R$ and $S$.
\item Use potential outcome generator $\Psi: R \times T \rightarrow Y$ to estimate the potential outcomes.
\item Apply counterfactual discriminator $\Phi: R \times T \times (Y^f \text{or} \ \hat{Y}^{cf}) \rightarrow P$ to remove imbalance of confounder representations between treatment and control group.
\item Here, Steps 6, 7, and 8 in the procedure are jointly trained together by optimizing minimax rule Eq.~(\ref{Eqn: minimax}) about $\mathcal{L}_m$, $\mathcal{L}_\Psi$, and $\mathcal{L}_{\Phi,\Psi}$ to update parameters in $g$, $f$, $d$, $\Phi$, and $\Psi$.
\end{enumerate}
\section{Experiments}
In this section, we conduct experiments on two semi-synthetic networked datasets, including the BlogCatalog and Flickr, to evaluate the following aspects: (1) Our proposed method can improve treatment effect estimation with respect to average treatment effect and individualized treatment effect compared to the state-of-the-art methods. (2) The structure mutual information can help representations capture more hidden confounder information, and thus increase the predictive accuracy for counterfactual outcomes. (3) The proposed method is robust to the hyperparameters.
\subsection{Dataset}
\textbf{BlogCatalog.} BlogCatalog is a social blog directory that manages the bloggers and their blogs. In this dataset, each unit is a blogger and each edge represents the social relationship between two bloggers. The features are bag-of-words representations of keywords in bloggers' descriptions. We follow the assumptions and procedures of synthesizing the outcomes and treatment assignments in ~\cite{guo2019learning}. In this semi-synthetic networked dataset, the outcomes are the opinions of readers on each blogger and the treatment options are mobile devices or desktops on which blogs are read more. If the blogger's blogs are read more on mobile devices, the blogger is in the treatment group; if they are read more on desktops, the blogger is in the control group. We also assume that the topics of bloggers with the social relationship can causally affect their treatment assignment and readers' opinions on them. To model readers' preference on reading some topics from mobile devices and others from desktops, one LDA topic model ~\cite{guo2019learning} is trained. Three settings of datasets are created with $k=0.5, 1,\text{and}\, 2$ that represent the magnitude of the confounding bias in the dataset. $k=0$ means the treatment assignment is random and there is no selection bias, and greater $k$ means larger selection bias.
\textbf{Flickr.} Flickr is a popular photo-sharing and hosting service, and it supports an active community where people can share each other's photos. In the Flickr dataset, each unit is a user and each edge represents the social relationship between two users. The features of each user represent a list of tags of interest. The same settings and simulation procedures as BlogCatalog dataset are adopted here. Table~\ref{tab:freq} presents an overview of these two datasets.
\begin{table}
\caption{Properties of BlogCatalog and Flickr datasets.}
\label{tab:freq}
\scalebox{0.9}{
\begin{tabular}{lll}
\toprule
Datasets & \textbf{BlogCatalog} & \textbf{Flickr}\\
\midrule
Nodes & 5,196 & 7,575\\
Features & 8,189 & 12,047\\
Edges & 171,743 & 239,738\\
Treatments & 2 & 2\\
\bottomrule
\end{tabular}}
\end{table}
\subsection{Baseline Methods}
We compare the proposed GIAL with the following baseline methods. \textbf{Network Deconfounder (ND)} ~\cite{guo2019learning} utilizes the GCN and integral probability metric to learn balanced representations to recognize patterns of hidden confounders from the network dataset. \textbf{Counterfactual Regression (CFRNET)} ~\cite{shalit2017estimating} maps the original features into a balanced representation space by minimizing integral probability metric between treatment and control representation spaces. \textbf{Treatment-agnostic Representation Networks (TARNet)} ~\cite{shalit2017estimating} is a variant of counterfactual regression without balance regularization. \textbf{Causal Effect Variational Autoencoder (CEVAE)} ~\cite{louizos2017causal} is based on Variational Autoencoder (VAE), which simultaneously estimates the unknown latent space summarizing the confounders and the causal effect. \textbf{Causal Forests (CF)} ~\cite{wager2018estimation} is a nonparametric forest-based method for estimating heterogeneous treatment effects by extending Breiman's random forest algorithm. \textbf{Bayesian Additive Regression Trees (BART)}~\cite{chipman2010bart} is a nonparametric Bayesian regression model, which uses dimensionally adaptive random basis elements.
\subsection{Descriptive Data Analysis}
\begin{figure}[h!]
\centering
\includegraphics[width=0.2\textwidth]{samples/example.PNG}
\vspace{-2mm}
\caption{Example of complete graph. The solid line represents heterogeneous edge and the dashed line means homogeneous edge.}
\vspace{-2mm}
\label{complete}
\end{figure}
\begin{table}
\caption{Summary of homogeneous edges and heterogeneous edges for the BlogCatalog datasets and Flickr datasets.}
\vspace{-2mm}
\label{homogeneous}
\centering
\scalebox{0.9}{
\begin{tabular}{llll}
\toprule
\multicolumn{1}{c}{Dataset} & \multicolumn{1}{c}{k} & \multicolumn{1}{c}{Homogeneous} & \multicolumn{1}{c}{Heterogeneous} \\
\midrule
& 0.5 & \textbf{94524.5} & 77218.5\\
BlogCatalog & 1 & \textbf{101102.8} & 70640.2\\
& 2 & \textbf{116031.8} & 55711.2\\
\midrule
& 0.5 & \textbf{124320.9} & 115417.1\\
Flickr & 1 & \textbf{130978.5} & 108759.5\\
& 2 & \textbf{141957.3} & 97780.7\\
\bottomrule
\end{tabular}}
\vspace{-2mm}
\end{table}
Before estimating the treatment effects from these two networked datasets, we provide the descriptive data analysis to demonstrate the existence of network structural imbalance in the networked data for causal inference problems.
\begin{table*}[t]
\caption{Performance comparison on BlogCatalog and Flickr datasets with different $k\in{0.5, 1, 2}$. We present the mean value of $\sqrt{\epsilon_\text{PEHE}}$ and $\epsilon_\text{ATE}$ on the test sets. Results of baseline methods on the same datasets are reported in~\cite{guo2019learning}.}
\vspace{-2mm}
\label{result}
\centering
\scalebox{0.9}{
\begin{tabular}{lllllllllllll}
\toprule
& \multicolumn{6}{c}{BlogCatalog} & \multicolumn{6}{c}{Flickr} \\
\cmidrule(lr){2-7} \cmidrule(lr){8-13}
& \multicolumn{2}{c}{k=0.5} & \multicolumn{2}{c}{k=1} & \multicolumn{2}{c}{k=2} & \multicolumn{2}{c}{k=0.5} & \multicolumn{2}{c}{k=1} & \multicolumn{2}{c}{k=2} \\
\cmidrule(lr){2-3} \cmidrule(lr){4-5} \cmidrule(lr){6-7} \cmidrule(lr){8-9} \cmidrule(lr){10-11} \cmidrule(lr){12-13}
Method & $\sqrt{\epsilon_\text{PEHE}}$ & $\epsilon_\text{ATE}$ & $\sqrt{\epsilon_\text{PEHE}}$ & $\epsilon_\text{ATE}$ &
$\sqrt{\epsilon_\text{PEHE}}$ & $\epsilon_\text{ATE}$ & $\sqrt{\epsilon_\text{PEHE}}$ & $\epsilon_\text{ATE}$ & $\sqrt{\epsilon_\text{PEHE}}$ & $\epsilon_\text{ATE}$ &
$\sqrt{\epsilon_\text{PEHE}}$ & $\epsilon_\text{ATE}$\\
\midrule
BART~\cite{chipman2010bart} & 4.808&2.680 &5.770 &2.278 &11.608 &6.418 & 4.907 & 2.323 & 9.517 & 6.548 & 13.155 & 9.643\\
CF~\cite{wager2018estimation} &7.456 &1.261 &7.805 &1.763 &19.271 &4.050 & 8.104& 1.359 & 14.636& 3.545 & 26.702 & 4.324\\
CEVAE~\cite{louizos2017causal} & 7.481& 1.279&10.387 &1.998 & 24.215&5.566 & 12.099 &1.732 & 22.496& 4.415&42.985 & 5.393\\
TARNet~\cite{shalit2017estimating} & 11.570& 4.228& 13.561& 8.170& 34.420&13.122 & 14.329 &3.389 & 28.466& 5.978& 55.066& 13.105\\
$\text{CFRNET}_\text{MMD}$~\cite{shalit2017estimating} &11.536 &4.127 & 12.332&5.345 & 34.654& 13.785 &13.539 &3.350 & 27.679& 5.416& 53.863& 12.115\\
$\text{CFRNET}_\text{Wass}$~\cite{shalit2017estimating} & 10.904& 4.257& 11.644&5.107 & 34.848&13.053 & 13.846&3.507 &27.514 & 5.192& 53.454& 13.269\\
ND~\cite{guo2019learning} & 4.532 &0.979 & 4.597&0.984 & 9.532&2.130 &4.286 & 0.805&5.789 & 1.359& 9.817& 2.700\\
\midrule
$\text{GIAL}_\text{GAT}$ (Ours) &4.215 &0.912 &4.258 &0.937 &9.119 &1.982 &4.015 &0.773 &5.432 & 1.2312&9.428&2.586\\
$\text{GIAL}_\text{GCN}$ (Ours) & \textbf{4.023} &\textbf{0.841} &\textbf{4.091} &\textbf{0.883} &\textbf{8.927} &\textbf{1.780} &\textbf{3.938} &\textbf{0.682}& \textbf{5.317}&\textbf{1.194} &\textbf{9.275} &\textbf{2.245}\\
\bottomrule
\end{tabular}}
\vspace{-2mm}
\end{table*}
\begin{table}[t]
\vspace{-2mm}
\caption{Summary of results in ablation studies.}
\vspace{-2mm}
\label{ablation}
\centering
\scalebox{0.9}{
\begin{tabular}{lllllll}
\toprule
& \multicolumn{2}{c}{k=0.5} & \multicolumn{2}{c}{k=1} & \multicolumn{2}{c}{k=2} \\
\cmidrule(lr){2-3} \cmidrule(lr){4-5} \cmidrule(lr){6-7}
& $\sqrt{\epsilon_\text{PEHE}}$ & $\epsilon_\text{ATE}$ & $\sqrt{\epsilon_\text{PEHE}}$ & $\epsilon_\text{ATE}$ &
$\sqrt{\epsilon_\text{PEHE}}$ & $\epsilon_\text{ATE}$ \\
\midrule
\textbf{BlogCatalog} & \multicolumn{6}{c}{} \\
GIAL &4.023 &0.841 &4.091 &0.883 &8.927 &1.780 \\
GIAL (w/o SMI) &4.422&0.982 &4.481 &0.981 &9.315 &2.142\\
GIAL (w/o CD) &4.482 & 0.987&4.951 &1.023 &13.598 &3.215\\
\midrule
\textbf{Flickr} & \multicolumn{6}{c}{} \\
GIAL &3.938 &0.682& 5.317&1.194 &9.275 &2.245\\
GIAL (w/o SMI) &4.158 &0.792 &5.694 &1.375 & 9.673&2.661 \\
GIAL (w/o CD) &4.284 &0.812 &6.127 &1.435 &11.524 &3.564 \\
\bottomrule
\end{tabular}
}
\vspace{-2mm}
\end{table}
According to graph theory, in the complete graph which is a simple undirected graph where every pair of distinct nodes is connected by a unique edge, there are $\frac{n\times(n-1)}{2}$ edges for $n$ nodes. We assume that the $n$ nodes are evenly divided into treatment group and control group with the same $\frac{n}{2}$ nodes in each group, and also each node has the same possibility to have an edge (relationship) with another node regardless of the node's treatment assignment. Then, this graph is still a complete graph with $\frac{n\times(n-1)}{2}$ edges. Now the edges in this graph are put into two categories: (a) the homogeneous group including the edges that link the nodes in the same group (treatment-treatment or control-control); (b) the heterogeneous group including the edges that link the nodes in different groups (treatment-control). Under the assumption that each node has the same possibility to be connected with another node regardless of the node's treatment assignment, we can find that in the homogeneous group, there are $\frac{n^2}{4}-\frac{n}{2}$ edges and in the heterogeneous group, there are $\frac{n^2}{4}$ edges. The number of edges in the heterogeneous group should be greater than that in the homogeneous group. For example, as shown in Fig.~\ref{complete}, there is one complete graph with 6 nodes including 3 treatment nodes and 3 control nodes. The heterogeneous group has 9 edges, while the homogeneous group has 6 edges.
We separately calculate the average numbers of homogeneous edges and heterogeneous edges for the BlogCatalog datasets and Flickr datasets, then report them in Table~\ref{homogeneous}. We can observe that the homogeneous edges are consistently greater than the heterogeneous edges for both datasets with different $k$. This result totally agrees with our expectation that, in the causal inference problem, the network structure is imbalanced. Therefore, the relationship is more likely to appear among people who are in the same group. This is the major difference between traditional graph learning tasks and the causal inference task on networked data, which is also the motivation of our proposed model.
\subsection{Experimental Settings}
In the following experiments, we randomly sample $60\%$ and $20\%$ of the units as the training set and validation set, and use the remaining $20\%$ units to form the test set. For each dataset with a different imbalance $k$, the simulation procedures are repeated 10 times and we report the average mean.
\textbf{GIAL.} By using different graph neural networks to learn the representation space from the networked dataset, the proposed GIAL method has two variants denoted as GIAL$_\text{GCN}$ and GIAL$_\text{GAT}$, which adopt the original implementation of graph convolutional network ~\cite{kipf2016semi} and graph attention network (GAT)~\cite{velickovic2019deep}, respectively. Besides, a squared $l_2$ norm regularization with hyperparameter $10^{-4}$ is added into our model to mitigate the overfitting issue. The hyperparameters of our method are chosen based on performance on the validation dataset, and the searching range is shown in Table~\ref{hyperparameter}. The Adam SGD optimizer ~\cite{kingma2014adam} is used to train the final objective function Eq.~(\ref{Eqn: minimax}) with an initial learning rate of 0.001 and an early stopping strategy with patience of 100 epochs.
\begin{table}[th!]
\vspace{-2mm}
\caption{Hyperparameters and ranges.}
\vspace{-2mm}
\label{hyperparameter}
\centering
\scalebox{0.9}{
\begin{tabular}{ll}
\toprule
\multicolumn{1}{c}{Hyperparameter} & \multicolumn{1}{c}{Range} \\
\midrule
$\alpha$, $\beta$ & 0, $10^{-4}$,$10^{-3}$ ,$10^{-2}$ ,$10^{-1}$ \\
\midrule
Dim. of confounder representation & 50, 100, 150, 200\\
\midrule
No. of GCN and GAT layers & 1, 2, 3\\
\midrule
No. of attention heads in GAT & 1, 2, 3, 4\\
\midrule
No. of outcome generator layer & 1, 2, 3, 4\\
\bottomrule
\end{tabular}
}
\end{table}
\textbf{Baseline Methods.} BART, CF, CEVAE, TARNet, and CFRNET are not originally designed for the networked observational data, so they cannot directly utilize the network information. To be fair, we concatenate the corresponding row of adjacency matrix to the original features, but this strategy cannot effectively improve the performance of baselines due to the curse of dimensionality. Besides, we adopt their default hyperparameter settings~\cite{guo2019learning}.
\subsection{Results}
For the BlogCatalog and Flickr datasets, we adopt two commonly used evaluation metrics to evaluate the performance of our method and baselines. The first one is the error of ATE estimation, which is defined as $\epsilon_\text{ATE} = |\text{ATE} - \widehat{\text{ATE}}|$, where \text{ATE} is the true value and $\widehat{\text{ATE}}$ is an estimated \text{ATE}. The second one is the error of expected precision in estimation of heterogeneous effect (PEHE)~\cite{hill2011bayesian}, which is defined as $\epsilon_\text{PEHE} = \frac{1}{n}\sum_{i=1}^{n}(\text{ITE}_i-\widehat{\text{ITE}}_i)^2$, where $\text{ITE}_i$ is the true \text{ITE} for unit $i$ and $\widehat{\text{ITE}}_i$ is an estimated \text{ITE} for unit $i$.
Table~\ref{result} shows the performance of our method and baseline methods on the BlogCatalog and Flickr datasets over 10 realizations. We report the average results of $\sqrt{\epsilon_\text{PEHE}}$ and $\epsilon_\text{ATE}$ on the test sets. $\text{GIAL}_\text{GCN}$ achieves the best performance with respect to $\sqrt{\epsilon_\text{PEHE}}$ and $\epsilon_\text{ATE}$ in all cases of both datasets. Although the $\text{GIAL}_\text{GAT}$ also has obvious improvements compared to baseline methods, it is outperformed by $\text{GIAL}_\text{GCN}$. GCN demonstrates clear superiority over GAT when recognizing patterns of hidden confounders from imbalanced network structure. Because $k=0.5, 1,\text{and}\, 2$ is used to represent the magnitude of the confounding bias in both datasets, results show that GIAL consistently outperforms the baseline methods under different levels of divergence, and our method is robust to a high level of confounding bias. Compared to baseline methods (e.g., CFRNET) only relying on observed confounders but without utilizing the network information, our model is capable of recognizing the patterns of hidden confounders from the network structure. Compared to baseline methods with learning network information (e.g., ND), our model has significant performance advantages, which demonstrates our model can capture more information from an imbalanced network structure. The reason is that our method maximizes the structure mutual information, instead of directly adopting the graph learning method without considering the specificity of networked data in the causal inference problem.
\subsection{Model Evaluation}
Experimental results on both datasets show that GIAL obtains a more accurate estimation of the ATE and ITE than the state-of-the-art methods. We further evaluate the performance of GIAL from two perspectives, including the effectiveness of each component, and its robustness to hyper-parameters.
We perform two ablation studies of $\text{GIAL}_\text{GCN}$ on both datasets. The first one is GIAL (w/o SMI) where the structure mutual information maximizing module is removed. We directly adopt graph neural networks to learn the representation space without considering the structural imbalance of networked data. The second ablation study is GIAL (w/o CD) where the counterfactual outcome discriminator is removed and there is not any restriction on the divergence between the representation distributions of treatment and control groups.
As shown in Table~\ref{ablation}, the performance becomes poor after removing either the structure mutual information or counterfactual outcome discriminator, compared to the original GIAL. More specifically, after removing the structure mutual information, $\sqrt{\epsilon_\text{PEHE}}$ and $\epsilon_\text{ATE}$ increase dramatically and have similar performance to other baseline methods. Besides, as the bias ($k$) increases, the difference between the performance of GIAL (w/o CD) and the original GIAL increases further. Therefore, the structure mutual information and counterfactual outcome discriminator are essential components of our model.
\begin{figure*}[t]
\centering
\includegraphics[width=1\linewidth]{samples/gial_parameter.PNG}
\caption{Sensitivity analysis for $\alpha$ and $\beta$ of structure mutual information and counterfactual outcome discriminator.}
\label{sensitivity}
\end{figure*}
\begin{comment}
\begin{table*}[h!]
\caption{Summary of results in ablation studies.}
\label{ablation}
\centering
\begin{tabular}{lllllllllllll}
\toprule
& \multicolumn{6}{c}{BlogCatalog} & \multicolumn{6}{c}{Flickr} \\
\cmidrule(lr){2-7} \cmidrule(lr){8-13}
& \multicolumn{2}{c}{k=0.5} & \multicolumn{2}{c}{k=1} & \multicolumn{2}{c}{k=2} & \multicolumn{2}{c}{k=0.5} & \multicolumn{2}{c}{k=1} & \multicolumn{2}{c}{k=2} \\
\cmidrule(lr){2-3} \cmidrule(lr){4-5} \cmidrule(lr){6-7} \cmidrule(lr){8-9} \cmidrule(lr){10-11} \cmidrule(lr){12-13}
Method & $\sqrt{\epsilon_\text{PEHE}}$ & $\epsilon_\text{ATE}$ & $\sqrt{\epsilon_\text{PEHE}}$ & $\epsilon_\text{ATE}$ &
$\sqrt{\epsilon_\text{PEHE}}$ & $\epsilon_\text{ATE}$ & $\sqrt{\epsilon_\text{PEHE}}$ & $\epsilon_\text{ATE}$ & $\sqrt{\epsilon_\text{PEHE}}$ & $\epsilon_\text{ATE}$ &
$\sqrt{\epsilon_\text{PEHE}}$ & $\epsilon_\text{ATE}$\\
\midrule
GIAL &4.023 &0.841 &4.091 &0.883 &8.927 &1.780 &3.938 &0.682& 5.317&1.194 &9.275 &2.245\\
GIAL (w/o SMI) &4.422&0.982 &4.481 &0.981 &9.315 &2.142 &4.158 &0.792 &5.694 &1.375 & 9.673&2.661 \\
GIAL (w/o CD) &4.482 & 0.987&4.951 &1.023 &13.598 &3.215&4.284 &0.812 &6.127 &1.435 &11.524 &3.564 \\
\bottomrule
\end{tabular}
\end{table*}
\end{comment}
Next, we explore the model's sensitivity to the most important parameters $\alpha$ and $\beta$, which control the ability to capture the graph structure and handle the confounding bias when estimating the potential outcomes. We show the results of $\sqrt{\epsilon_\text{PEHE}}$ and $\epsilon_\text{ATE}$ on BlogCatalog dataset with different $k$ in Fig.~\ref{sensitivity}. We observe that the performance is stable over a large parameter range. It confirms the effectiveness and robustness of structure mutual information and counterfactual outcome discriminator in GIAL, which is consistent with our ablation studies, i.e., GIAL (w/o SMI) and GIAL (w/o CD).
\section{Related Work}
The related work is presented along with two directions: learning causal effects from observational data and graph neural networks.
Various causal effect estimation methods for observational data have sprung up. For most existing methods, the strong ignorability assumption is the most important prerequisite. However, this assumption might be untenable in practice. A series of methods have been proposed to relax the strong ignorability assumption. A latent variable is inferred as a substitute for unobserved confounders~\cite{wang2019blessings}. Variational Autoencoder has been used to infer the relationships between the observed confounders based on the assumption joint distribution of the latent confounders and the observed confounders can be approximately recovered solely from the observations ~\cite{louizos2017causal}. Recently, some work aims to relax the strong ignorability assumption via network knowledge, where the network connecting the units is a proxy is for the unobserved confounding. The network deconfounder~\cite{guo2019learning} learns representations of confounders from network data by adopting the graph convolutional networks. Another work utilizes graph attention networks to learn representations and mitigates confounding bias by representation balancing and treatment prediction, simultaneously ~\cite{guo2020ignite}. Causal network embedding (CNE)~\cite{veitch2019using} is proposed to learn node embeddings from network data to represent confounders by reducing the causal estimation problem to a semi-supervised prediction of both the treatments and outcomes. For the existing methods about networked data, they do not dig deeply on what is the essential difference between the networked data under the causal inference problem and the networked data for traditional graph learning tasks such as node classification, link detection, etc. This is the reason why we propose this GIAL model, instead of directly adopting the GCN or GAT to learn the representation from the networked data.
Graph learning is increasingly becoming fascinating as more and more real-world data can be modeled as networked data. Graph convolutional network~\cite{kipf2016semi} is an effective approach for semi-supervised learning on networked data, via a localized first-order approximation of spectral graph convolutions. Graph attention network (GAT) ~\cite{velivckovic2017graph} is an attention-based architecture leveraging masked self-attentional layers where nodes are able to attend over their neighborhoods' features. Deep graph infomax (DGI) ~\cite{velickovic2019deep} is one approach for learning node representations within networked data in an unsupervised manner, which relies on maximizing mutual information between patch representations and high-level summaries of graphs. In our model, we extend the idea in DGI originally aimed for unsupervised learning to representation learning under the causal inference setting. Utilizing the structure mutual information can help representations capture the imbalanced structure that is specific to the causal inference problem.
\section{Conclusion}
In this paper, we propose the Graph Infomax Adversarial Learning method (GIAL) to capture the hidden confounders and estimate the treatment effects from networked observational data. GIAL makes full use of the network structure to capture more information by recognizing the imbalance in the network structure. Our work clarifies the greatest particularity of networked data under the causal inference problem compared with traditional graph learning tasks, that is, the structural imbalance due to confounding bias between treatment and control groups. Extensive experiments show the effectiveness and advantages of the proposed GIAL method.
\section*{ACKNOWLEDGMENTS}
We would like to thank the anonymous reviewers for their insightful comments. This research is supported in part by the U.S. Army Research Office Award under Grant Number W911NF-21-1-0109.
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}\label{sec:introduction}
Image deblurring is an important problem in the field of image restoration and has attracted wide attention with a long history \cite{josa1972_richardson_bayesian_deblurring, siggraph2006_fergus_removing_shake, tip1998_chan1998_tv_deblur}. The goal of deblurring is to recover the underlying clean image from an observed blurry one. Such a blurry image can be resulted by different reasons, such as defocusing and camera shaking, which can in general be modeled as (assuming the blur process is uniform and spatially invariant)
\begin{equation}\label{blurmodel}
\bm{y} = \bm{h} \ast \bm{z} + \bm{n},
\end{equation}
where $\bm{y}$ is the blurry image, $\bm{z}$ is the corresponding clean one, $\bm{h}$ is the blur kernel (point spread function), $\bm{n}$ is the noise, often modeled as additive white Gaussian noise, and $\ast$ denotes the 2D convolution operator. When the blur kernel $\bm{h}$ is unknown, we need to simultaneously estimate $\bm{h}$ and $\bm{z}$ solely from the given observation $\bm{y}$, which is commonly referred as blind image deblurring or blind deconvolution. It is a highly ill-posed inverse problem, since there can be different pairs of $\bm{z}$ and $\bm{h}$ explaining one same blurred image $\bm{y}$.
Many different methods have been proposed to address this problem, which can be mainly categorized into two groups, i.e., optimization-based methods and deep learning ones. Traditional optimization-based methods, from the probabilistic perspective, generally formulate the image deblurring task as a maximum-a-posteriori (MAP) estimation problem \cite{cvpr2011_krishnan_normalized_sparsity_prior, tog2008_shan_high_quality, cvpr2013_xu_l0norm, cvpr2014_pan_l0text, cvpr2016_pan_dark_channel, eccv2014_michaeli_2014_patch_prior, nips2009_krishnan_hyper_laplacian} or Bayesian posterior inference (specifically variational inference (VI)) one \cite{siggraph2006_fergus_removing_shake, tip2015_zhou_variational_dirichlet_deblur, book2017_bishop_review}, by properly designing priors to both the to-be-estimated image \cite{cvpr2013_xu_l0norm, cvpr2014_pan_l0text, cvpr2016_pan_dark_channel, nips2009_krishnan_hyper_laplacian, siggraph2006_fergus_removing_shake, eccv2014_michaeli_2014_patch_prior} and blur kernel \cite{icassp1997_molina_dirichlet_deblur, tip2015_zhou_variational_dirichlet_deblur}. However, the manually designed priors are still subjective and may not precisely characterize the intrinsic distributions of natural images and blur kernels in real world, limiting the performance of a deblurring algorithm. Besides, solving the complex optimization usually causes heavy computational cost, which is always unaffordable in real applications. Recently, motivated by their great success in computer vision, especially for other image restoration problems, \cite{pami2015_dong_cnn_sr, tip2017_zhang_dncnn}, deep learning methods have also attracted increasing attention in image deblurring research.
Early attempts along this line used deep neural networks (DNNs) as a component of traditional methods, e.g., estimating the blur kernel \cite{cvpr15_sun_learning_kernel, cvpr17_gong_motion_flow, pami15_schuler_learning_deblur, eccv16_chakrabarti_deblur}. Later, learning an explicit mapping from a blurred image to its deblurred one in an end-to-end fashion
has become a dominant methodology \cite{bmvc15_hradis_cnn_deblur, cvpr17_nah_deepdeblur, cvpr18_tao_srn}. This approach is especially superior in extracting underlying knowledge from large datasets and fast deblurring for test images.
Existing deep deblurring methods mainly focus on designing fancy network architectures, while pay less attention to make full use of the information delivered by physical blur model \eqref{blurmodel}, which limits further performance improvement for them. In contrast, \eqref{blurmodel} plays a crucial role for achieving a promising performance in optimization based methods, especially Bayesian ones \cite{siggraph2006_fergus_removing_shake, cvpr2009_levin_understanding_deblurring, pami2020_pan_physics_gan}. Based on these observations, and also inspired by recent advances in deep Bayesian modeling for image restoration \cite{nips2019_yue_vdn}, we propose a deep variational Bayesian framework for training deblurring neural networks in the blind setting, aiming at simultaneously estimating the clean image and blur kernel. Specifically, under the generative model of blurry image, we construct the variational distributions parameterized by DNNs to approximate the true posteriors of both the latent clean image and blur kernel, and learn the network parameters by optimizing the evidence lower bound (ELBO). After training, the inference networks can then be directly used to approximate the corresponding posteriors of clean image and blur kernel in an amortized way, efficiently from an observed blurry image.
The main advantages of the proposed method can be summarized as follows. First, by virtue of the Bayesian generative model, the physical blur model \eqref{blurmodel} can be naturally encoded as the likelihood term in the ELBO to guide learning the inference networks. Therefore, the inference networks for the clean image and blur kernel can both be jointly trained under the supervision of data driven priors, and interact with each other due to the likelihood. Second, the proposed framework provides, in principle, a natural unification of previous methodologies that consider kernel information for training deblurring DNNs, in a concise Bayesian way, and thus is expected to achieve a better performance. Third, our framework can also be straightforwardly incorporated with any existing end-to-end DNNs to enhance their performance. All of these advantages have been substantiated by our experiments.
The remainder of this paper is organized as follows: Section \ref{sec:related_work} introduces the related studies. Section \ref{sec:method} presents the proposed deep Bayesian framework, including Bayesian generative model, deep VI procedure and network architectures. Section \ref{sec:experiments} shows the experimental results and analyses. Final conclusion is made in Section \ref{sec:conclusion}.
\section{Related Work}\label{sec:related_work}
\textbf{Optimization based methods}
The main focus along this research line is to design priors that can precisely characterize the intrinsic structures of clean images and blur kernels, in order to make the ill-posed blind deblurring problem solvable.
A series of priors have been developed for clean images, including total variation \cite{tip1998_chan1998_tv_deblur}, sparsity of gradients \cite{siggraph2006_fergus_removing_shake, tog2008_shan_high_quality}, hyper-Laplacian \cite{nips2009_krishnan_hyper_laplacian}, $L_0$-norm prior \cite{cvpr2013_xu_l0norm, cvpr2014_pan_l0text}, internal patch prior \cite{eccv2014_michaeli_2014_patch_prior} and dark channel prior \cite{cvpr2016_pan_dark_channel}. For blur kernels, the aforementioned priors can still be applied, while there are also some specifically designed priors, e.g., Dirichlet prior \cite{icassp1997_molina_dirichlet_deblur, tip2015_zhou_variational_dirichlet_deblur} and piecewise-linear prior {\cite{tip2014_oh_piecewise_prior}}.
After specifying the priors, such a blind image deblurring problem can be formulated as a MAP estimation\cite{tog2008_shan_high_quality, cvpr2013_xu_l0norm, cvpr2014_pan_l0text, cvpr2016_pan_dark_channel, eccv2014_michaeli_2014_patch_prior, nips2009_krishnan_hyper_laplacian} or Bayesian posterior inference (more specifically VI) \cite{siggraph2006_fergus_removing_shake, tip2015_zhou_variational_dirichlet_deblur} problem, and solved by off-the-shelf optimization tools.
\textbf{Deep learning based methods}
Having witnessed the great success in other fields, deep learning has also been investigated in dealing with the image deblurring task. Early methods treat DNNs as a part of traditional optimization based ones, by virtue of their prediction ability. For example, Schuler \textit{et al.} \cite{pami15_schuler_learning_deblur} designed a convolutional neural network (CNN) to simulate iterative optimization for image deblurring. Sun \textit{et al.} \cite{cvpr15_sun_learning_kernel} used a CNN to classify the blur kernel in patch-level. Gong \textit{et al.} \cite{cvpr17_gong_motion_flow} went further to predict the motion flow in a pixel-level. Chakrabarti \cite{eccv16_chakrabarti_deblur} used CNN to predict the kernel in the frequency domain. It is then gradually popular to directly parameterize the mapping from a blurry image and its clean one with CNNs in a kernel free paradigm, and then learn the parameters by a large collection of blurry-clean image pairs. Hradiš \textit{et al.} \cite{bmvc15_hradis_cnn_deblur} directly predicted latent clean images from observed blurred images by a fully CNN. Nah \textit{et al.} \cite{cvpr17_nah_deepdeblur} introduced a multi-scale CNN to directly recover the latent clean image.
Tao \textit{et al.} \cite{cvpr18_tao_srn} improved the architecture by using a scale-recurrent network and a ConvLSTM \cite{nips2015_shi_convlstm} layer. The generative adversarial network (GAN) \cite{nips2014_goodfellow_gan} is also applied to image deblurring, e.g., Kupyn \textit{et al.} \cite{cvpr2018_kupyn_deblurgan} utilized the Wasserstein GAN \cite{icml2017_arjovsky_wgan} to learn a generator that restores the latent image. Recently, there are some studies that borrow fundamental principles from traditional methods. Pan \textit{et al.} \cite{pami2020_pan_physics_gan} introduce the consistency loss motivated from the likelihood in traditional methods. Cai \textit{et al.} \cite{tip2020_cai_dcp_nn} embed the dark and bright channel prior to a deep multi-scale neural network. Ren \textit{et al.} \cite{cvpr2020_ren_dip_deblur} tried to directly solve the traditional blind deblurring optimization, while constrain the solution space by deep image priors \cite{cvpr2018_ulyanov_dip}.
\section{Deep Variational Bayesian Framework for Blind Image Deblurring}\label{sec:method}
Let $\mathcal{D}=\{ \bm{x}^{(j)}, \bm{y}^{(j)}, \bm{k}^{(j)} \}_{j=1}^N$ be the training dataset, where $\bm{x}^{(j)}$, $\bm{y}^{(j)}$ and $\bm{k}^{(j)}$ denote the clean image, blurry image and blur kernel, respectively, in the $j$th training tuple, and $N$ be the number of training tuples. Our goal is to construct parametric forms with DNNs to approximate the posteriors of both clean image and blur kernel, given only the observed blurry image, under the Bayesian framework. With these parametric variational posteriors trained on $\mathcal{D}$, we can directly infer the posterior distributions of both clean image and blur kernel from a test blurry image. Note that in practical training deblurring datasets \cite{cvpr17_nah_deepdeblur}, for a blurry image, the corresponding ``clean'' image is often obtained by complicated acquisition process, which might not be the exact latent clean image. Besides, the groundtruth blur kernels are also unavailable. Nevertheless, these two issues can be naturally addressed within our Bayesian formulation.
\subsection{Bayesian Model of Blind Image Deblurring with Data Driven Priors}\label{sec:model}
Let $(\bm{x}, \bm{y}, \bm{k})$, with $\bm{x} \in \mathbb{R}^{h \times w}$, $\bm{y} \in \mathbb{R}^{h \times w}$, and $\bm{k} \in \mathbb{R}^{h' \times w'} $, be any training tuple in $\mathcal{D}$, where $h \times w$ and $h' \times w'$ are respectively the sizes of image and kernel. As aforementioned, the ``clean'' image $\bm{x}$ and blur kernel $\bm{k}$ might be inaccurate or unavailable, and thus we introduce notations $\bm{z}$ and $\bm{h}$ for the latent clean image and blur kernel, respectively. With these notations, we can build our Bayesian model for blind image deblurring as follows.
\textbf{Likelihood:} Considering the physical blur model \eqref{blurmodel}, we can have the following likelihood for the observed blurry image $\bm{y}$, given the latent clean image $\bm{z}$ and blur kernel $\bm{h}$ \cite{siggraph2006_fergus_removing_shake, tip2016_molina_vi_deblur}:
\begin{equation}\label{blur_likelihood}
p(\bm{y} | \bm{z}, \bm{h}) = \prod\nolimits_{i=1}^d \mathcal{N}(y_i ; (\bm{h} \ast \bm{z})_i, \sigma^2),
\end{equation}
where $\mathcal{N}(\cdot ; \mu,\sigma^2)$ denotes the Gaussian distribution with mean $\mu$ and variance $\sigma^2$, subscript $i$ denotes the $i$th pixel within the image (by taking the image matrix as a vector) and $d=hw$ for notation convenience. It should be mentioned that, in this work, we treat the variance $\sigma^2$ as a hyper-parameter to reduce the complexity of overall model. Nevertheless, we can also treat it as a to-be-estimated variable and learn its posterior distribution from data as in \cite{nips2019_yue_vdn}.
To complete the full Bayesian model, we then need introduce priors to $\bm{z}$ and $\bm{h}$. Different from that in traditional optimization based methods, where the priors were specified in a subjective way, we place these priors in a data driven fashion with the aid of training set $\mathcal{D}$, which is detailed in the following.
\textbf{Data Driven Prior for Clean Image:} As mentioned before, $\bm{x}$ in the training set might not exactly be the latent clean image $\bm{z}$. Nevertheless, it can be expected that $\bm{x}$ is close to $\bm{z}$, and thus provide strong prior information to it. Thus, we can naturally assume the prior distribution of $\bm{z}$ as
\begin{equation}\label{piror_z}
p(\bm{z}) = \prod\nolimits_{i=1}^{d} \mathcal{N}(z_i ; x_i, \varepsilon_0^2)
\end{equation}
where $\varepsilon_0^2$ is a hyper-parameter, which measures the uncertainty of $\bm{z}$. Since typical deblurring datasets are carefully collected, it is expected that $\bm{x}$ is close enough to $\bm{z}$, and we can correspondingly assume a small $\varepsilon_0^2$. In the extreme case that $\varepsilon_0^2$ tends to 0, the prior degenerates to the Dirac distribution with all mass centered at $\bm{x}$. We will empirically analyze the effect of $\varepsilon_0^2$ in our experiments.
\textbf{Data Driven Prior for Blur Kernel:} Specifying a proper prior distribution for the blur kernel is more difficult, since there is a constraint that the sum of all elements in a blur kernel should equal to 1. Fortunately, Dirichlet distribution defined on a simplex well meet this requirement. Its probability density function, with parameter $\bm{\alpha}\!=\!(\alpha _{1}, \dots, \alpha_{M}),\alpha_m\!>\!0,m\!=\!1,\dots,M$ has the following form:
\begin{equation}
\text{Dir}(\bm{\pi};\bm{\alpha})=\frac{1}{Z(\bm{\alpha})}\prod\nolimits_{m=1}^{M} \pi_{m}^{\alpha _{m}-1},\quad 0<\pi_m<1,\quad m=1,\dots,M,
\end{equation}
where $Z(\bm{\alpha})=\prod_{m=1}^{M}\Gamma(\alpha_m)/\Gamma\big(\sum_{m=1}^{M}\alpha_m\big)$ is the normalizing factor, with $\Gamma(\cdot)$ being the gamma function. Besides, the mean and variance are given by, denoting $\alpha_{0}=\sum_{m=1}^{M}\alpha_m$,
\begin{equation}\label{dir_mean_variance}
{\mathbb{E}}[\pi_{m}] = \alpha _{m}/\alpha _{0}, \quad
{\text{Var}}[\pi_{m}] = \big(\alpha _{m}(\alpha _{0} - \alpha _{m})\big)/\big(\alpha _{0}^{2}(\alpha _{0} + 1)\big).
\end{equation}
With the above knowledge, we can naturally place a Dirichlet prior to the latent blur kernel $\bm{h}$, and the problem is how to set its hyper-parameter. Since $\bm{k}$, if available, in the training set should provide strong information to $\bm{h}$, it is reasonable to assume $\mathbb{E}[\bm{h}]=\bm{k}$. Also, as can be seen from \eqref{dir_mean_variance}, the mean of a Dirichlet distribution remains the same under a scaling transformation of parameters. Therefore, we can introduce the following prior distribution to $\bm{h}$:
\begin{equation}
p(\bm{h})=\text{Dir}(\bm{h};c\bm{k})\propto\prod\nolimits_{i'=1}^{d'} h_{i'}^{ck_{i'}-1},
\label{prior_h}
\end{equation}
where $d'=h'w'$ for convenience, and $c>0$ is a scaling factor to be tuned. Although hyper-parameter $c$ does not affect the mean provided $\bm{k}$ being specified, it does determine the variance according to \eqref{dir_mean_variance}, and correspondingly the prior uncertainty of $\bm{h}$. Thus, if $\bm{k}$ is given in the training set, we may set $c$ to a relatively larger value, which results in a smaller prior uncertainty for $\bm{h}$. In contrast, if $\bm{k}$ is unavailable, we can only roughly estimate it from the given $\bm{y}$ and $\bm{x}$, and it is better to set $c$ small to avoid the inaccurate prior information being too strong.
\textbf{Posterior Inference Goal:} Giving both the likelihood and priors, the goal then turns to inferring the posterior distribution of the latent variables $\bm{z}$ and $\bm{h}$ from the observed blurry image $\bm{y}$, which, by the Bayes' rule, can be theoretically formulated as
\begin{equation} \label{true_posterior}
p(\bm{z}, \bm{h} | \bm{y}) \propto p(\bm{z}, \bm{h}, \bm{y}) =
p(\bm{y} | \bm{z}, \bm{h}) p(\bm{z}) p(\bm{h}).
\end{equation}
Exactly computing this posterior is intractable, especially for a test image out of the training set. Therefore, we turn to constructing the approximate posterior in an amortized way \cite{iclr2014_kingma_vae, icml2014_rezende_2014_vae} with DNNs, and then train the inference networks on the training set. After that, the trained networks can be used to directly infer the posterior of $\bm{z}$ and $\bm{h}$ by input a test blurry image $\bm{y}$.
\subsection{Variational Formulations of Posterior}
In order to achieve the goal mentioned in the previous subsection, we first need to construct the variational form of the posterior, denoted as $q(\bm{z}, \bm{h} | \bm{y})$, to approximate the intractable true posterior in \eqref{true_posterior}. Since $p(\bm{z}, \bm{h} | \bm{y})$ is a joint distribution of $\bm{z}$ and $\bm{h}$, there are multiple ways to formulate the variational posterior, and we discuss two possible ones according to the way of inferring $\bm{z}$.
\textbf{Kernel-free Inference Structure for $\bm{z}$:} The basic mean-field assumption for $q(\bm{z}, \bm{h} | \bm{y})$ is
\begin{equation}\label{KF_form}
q(\bm{z}, \bm{h} | \bm{y}) = q(\bm{z} | \bm{y}) q(\bm{h} | \bm{y}),
\end{equation}
which assumes the conditional independence between $\bm{z}$ and $\bm{h}$ given $\bm{y}$. Then we further specify $q(\bm{z} | \bm{y})$ and $q(\bm{h} | \bm{y})$ with the same families as $q(\bm{z})$ and $q(\bm{h})$, respectively:
\begin{equation}\label{KF_vpost}
q_{\Psi}(\bm{z} | \bm{y}) =
\prod\nolimits_{i=1}^d \mathcal{N}\big(z_i; \mu_i(\bm{y};\Psi), m_i^2(\bm{y};\Psi)\big),\quad q_{\Phi}(\bm{h} | \bm{y}) = \text{Dir}\big(\bm{h}; \bm{\xi}(\bm{y};\Phi)\big),
\end{equation}
where $\Psi$ denotes the parameters of the inference network for $\bm{z}$, which takes $\bm{y}$ as its input and outputs the mean $\mu_i(\bm{y};\Psi)$ and variance $m_i^2(\bm{y};\Psi)$, and thus can be recognized as \emph{Deblur-Net}; similarly, $\Phi$ parameterizes the inference network for $\bm{h}$, which also takes $\bm{y}$ as input while outputs the parameters $\bm{\xi}(\bm{y};\Phi)=\big(\xi_{1}(\bm{y};\Phi),\dots,\xi_{d'}(\bm{y};\Phi)\big)$, and thus can be referred to as \emph{Kernel-Net}.
\textbf{Kernel-dependent Inference Structure for $\bm{z}$:} Another formulation for variational posterior $q(\bm{z}, \bm{h} | \bm{y})$ is obtained by conditional probability rule:
\begin{equation}\label{KD_form}
q(\bm{z}, \bm{h} | \bm{y}) = q(\bm{z} |\bm{h}, \bm{y}) q(\bm{h} | \bm{y}),
\end{equation}
where the posterior of $\bm{z}$ is also conditioned on $\bm{h}$ in addition to $\bm{y}$. This dependence can be easily achieved by taking both $\bm{h}$ and $\bm{y}$ as the inputs of \emph{Deblur-Net}, leading to following variational form:
\begin{equation}\label{KD_vpost}
q_{\Psi}(\bm{z} |\bm{h}, \bm{y}) =
\prod\nolimits_{i=1}^d \mathcal{N}\big(z_i; \mu_i(\bm{h},\bm{y};\Psi), m_i^2(\bm{h},\bm{y};\Psi)\big).
\end{equation}
This parameterization is rational, since the blur kernel, if accurately estimated, can provide useful information for deblurring an image. The variational form of $q_{\Phi}(\bm{h} | \bm{y})$ remains the same as in \eqref{KF_vpost}.
\emph{Remark:} Mathematically, it is also hold that $q(\bm{z}, \bm{h} | \bm{y}) = q(\bm{h} |\bm{z}, \bm{y})q(\bm{z} | \bm{y})$, which leads to another valid variational posterior formulation. However, considering that our ultimate goal is to deblur $\bm{y}$, this formulation is less attractive, since it requires estimating $\bm{z}$ first before inferring $\bm{h}$.
\subsection{Network Learning with Evidence Lower Bound}
After the variational posterior parameterized by DNNs having be constructed, we need to derive the evidence lower bound (ELBO), which acts as the objective function for learning the network parameters $\Psi$ and $\Phi$. For notation convenience, we simplify $\mu_i(\bm{y};\Psi)$ and $\mu_i(\bm{h},\bm{y};\Psi)$ to $\mu_i$, $m_i^2(\bm{y};\Psi)$ and $m_i^2(\bm{h},\bm{y};\Psi)$ to $m_i^2$, and $\xi_{i'}(\bm{y};\Phi)$ to $\xi_{i'}$ in the following.
Based on conventional decomposition of marginal likelihood in VI, we can get
\begin{equation}
\log p(\bm{y})=\mathcal{L}(\bm{y};\Psi,\Phi)+D_{KL}\big(q(\bm{z},\bm{h}|\bm{y})\|p(\bm{z},\bm{h}|\bm{y})\big),
\end{equation}
where $D_{KL}(\cdot|\cdot)$ is the KL divergence between two distributions, and $\mathcal{L}(\bm{y};\Psi,\Phi)$ is the ELBO we want to maximize. According to the two variational formulations \eqref{KF_form} and \eqref{KD_form}, we correspondingly have the following two ELBOs (derivations are provided in Appendix \ref{app:elbo_derive}):
\small
\begin{equation}\label{KF_elbo}
\mathcal{L}_{\text{KF}}(\bm{y};\Psi,\!\Phi)=\mathbb{E}_{q( \bm{z} | \bm{y}) q( \bm{h} | \bm{y})} [ \log p( \bm{y} | \bm{h}, \bm{z} )]- D_{KL}( q(\bm{z} | \bm{y} ) || p( \bm{z}) )
- D_{KL}(q( \bm{h} | \bm{y}) || p(\bm{h})),
\end{equation}
\normalsize
and
\small
\begin{equation}\label{KD_elbo}
\mathcal{L}_{\text{KD}}(\bm{y};\Psi,\!\Phi)= \mathbb{E}_{q( \bm{z} | \bm{h}, \bm{y}) q( \bm{h} | \bm{y} )} [ \log p( \bm{y} | \bm{h}, \bm{z})] - \mathbb{E}_{q( \bm{h} | \bm{y})} [ D_{KL}( q( \bm{z} | \bm{h}, \bm{y} ) || p(\bm{z}))] - D_{KL}(q(\bm{h} | \bm{y}) || p( \bm{h})),
\end{equation}
\normalsize
where we omit the subscripts in $q_{\Psi}(\bm{z} |\bm{h}, \bm{y})$, $q_{\Psi}(\bm{z} |\bm{h}, \bm{y})$ and $q_{\Phi}(\bm{h} |\bm{y})$. The KL divergence terms in the above two equations can then be calculated analytically as follows:
\small
\begin{equation}\label{kl_gaussian}
D_{KL}(q(\bm{z} | \bm{y}) || p(\bm{z}))=D_{KL}(q(\bm{z} | \bm{h}, \bm{y}) || p(\bm{z})) =
\sum_{i=1}^d \left\{\frac{\left(\mu_{i} - x_{i}\right)^{2}}{2 \varepsilon_{0}^{2}}
+ \frac{1}{2}\left[\frac{m_{i}^{2}}{\varepsilon_{0}^{2}}-\log\frac{m_{i}^{2}}{\varepsilon_{0}^{2}}-1\right]\right\},
\end{equation}
\begin{equation}
\begin{aligned}
D_{KL}(q(\bm{h} | \bm{y}) || p(\bm{h})) = &~\log \Gamma\Big(\sum\nolimits_{i'=1}^{d'}\xi_{i'}\Big) -\log \Gamma\Big(\sum\nolimits_{i'=1}^{d'}ck_{i'}\Big) - \sum\nolimits_{i'=1}^{d'} \log \Gamma(\xi_k) \\
&+ \sum\nolimits_{i'=1}^{d'} \log \Gamma(ck_{i'}) + \sum\nolimits_{i'=1}^{d'} (\xi_{i'} - ck_{i'}) \Big(\psi_0(\xi_{i'}) - \psi_0\Big(\sum\nolimits_{i'=1}^{d'}\xi_{i'}\Big)\Big),
\end{aligned}
\end{equation}
\normalsize
where $\psi_0(\cdot)$ is the digamma function. Note that $D_{KL}(q(\bm{z} | \bm{y}) || p(\bm{z}))$ and $D_{KL}(q(\bm{z} | \bm{y}, \bm{h}) || p(\bm{z}))$ have the same form, since $q(\bm{z} | \bm{y})$ and $q(\bm{z} |\bm{h}, \bm{y})$ both follow Gaussian distributions parameterized by DNNs, while the difference lies in the parametric forms of $\mu_i$ and $m_i^2$ as can be seen in \eqref{KF_vpost} and \eqref{KD_vpost}. The first term in either \eqref{KF_elbo} or \eqref{KD_elbo} can not be analytically calculated due to the intractable integral. Nevertheless, reparameterization trick \cite{iclr2014_kingma_vae, icml2014_rezende_2014_vae, nips2018_figurnov_implicit_trick} can be applied to obtaining an unbiased gradient estimator for it, which has been integrated into modern deep learning frameworks like PyTorch \cite{nips2019_pytorch}.
Given a training dataset $\mathcal{D}$, we now can optimize the network parameters $\Psi$ and $\Phi$ with respect to the ELBO \eqref{KF_elbo} or \eqref{KD_elbo}, over all data in $\mathcal{D}$, which respectively corresponds to the following objectives:
\begin{equation}\label{final_objective}
\max_{\Psi, \Phi} \sum\nolimits_{j=1}^N\mathcal{L}_{\text{KF}}(\bm{y}^{(j)};\Psi,\Phi),\quad \max_{\Psi, \Phi} \sum\nolimits_{j=1}^N\mathcal{L}_{\text{KD}}(\bm{y}^{(j)};\Psi,\Phi).
\end{equation}
\emph{Remark:} The three terms involved in the ELBO, either $\mathcal{L}_{\text{KF}}$ or $\mathcal{L}_{\text{KD}}$, play important roles. Firstly, the second term plays the role of supervising \emph{Deblur-Net} by the clean training image, which acts similarly as the loss functions of most end-to-end trained deblurring DNNs. Secondly, the third term provides supervision information to \emph{Kernel-Net}, which is often ignored in many existing methods. Thirdly, the first term (likelihood), attributed to the physical blur model, includes both \emph{Deblur-Net} and \emph{Kernel-Net}, making them interact with each other, which is also less considered in previous methods. To summarize, with the derived ELBO, \emph{Deblur-Net} and \emph{Kernel-Net} can both be supervised by data-driven priors and interact with each other during the training process.
\begin{figure}[t]
\centering\vspace{-1mm}
\includegraphics[width=0.98\textwidth]{main_img.pdf}
\vspace{-2mm}
\caption{\small{The architecture of the proposed framework with kernel-dependent inference structure.
}}
\label{fig_network}
\vspace{-5mm}
\end{figure}
\subsection{Network Architecture}\label{sec:net_arch}
\textbf{Design of \emph{Deblur-Net}:} For the kernel-free inference structure for $\bm{z}$, we design a relative simple yet effective baseline \emph{Deblur-Net} for evaluation. We adopt the widely used U-Net \cite{miccai2015_unet}, with four scales in the encoder and decoder, to capture multi-scale features of an image, and make some modifications to it. Specifically, we first use the modified ResBlock in \cite{cvpr17_nah_deepdeblur} but replace the ReLU activation with LeakyReLU one and use a small kernel size of 3. Then we stack a series of ResBlocks (the number of ResBlocks varies in our experiments depending on different purposes) to construct the encoding and decoding blocks. The number of channels at all layers is set to 64. The downsample is done by a convolution layer with filter size 3 and stride 2, while the upsample by a transposed convolution layer with filter size 5 and stride 2.
For the kernel-dependent inference structure for $\bm{z}$, the \emph{Deblur-Net} should also take the blur kernel, in addition to the blurry image, as its input. To do so, we adopt the SFT layer in \cite{cvpr2019_gu_ikc} to the encoding and decoding blocks. We also use a sub-network to map the blur kernel from matrix to vector.
\textbf{Design of \emph{Kernel-Net}:}
We use a ResNet \cite{cvpr2016_he_resnet} as feature extractor for \emph{Kernel-Net}, followed by a linear transformation layer. Then we apply a softmax layer after that and rescale it to make the final output non-negative and normalized. It should be mentioned that accurately estimating the blur kernel is not a easy task, and there is still a large room to improve the this baseline \emph{Kernel-Net}.
The overall network architecture of the proposed framework with kernel-dependent inference is illustrated in Fig. \ref{fig_network}. For the kernel-free inference, the architecture is almost the same, except that the output of \emph{Kernel-Net} does not be input to \emph{Deblur-Net}.
\subsection{Further Discussions}\label{sec:discussion}
\textbf{Non-uniform Blur:} Our deblurring framework is constructed based on the physical blur model \eqref{blurmodel}, which assumes uniform blur kernel across the spatial domain. For non-uniform blurring, we can follow previous studies \cite{cvpr2016_pan_softseg_deblurring, pami2020_pan_physics_gan}, and treat the blur kernel locally uniform within a relatively small image patch.
Since in the training phase, we indeed train the inference networks with patches cropped from images in the training set, the local uniformity of the blur kernel approximately holds, and therefore the proposed framework is also applicable. In the testing phase, the kernel-free inference can be done without any specific modification, while the kernel-dependent inference can be implemented by first performing deblurring on overlapping image patches and then aggregating them to obtain the final result, just like many traditional image restoration methods \cite{ksvd_image, wnnm_j}. However, the above strategy is only a rough approximation, and the application to non-uniform deblurring is still a major limitation of the proposed framework, which should be further investigated in future.
\textbf{Relationship to Kernel-involved Methodologies:} In previous studies, there are mainly two different methodologies in deep learning based deblurring methods that take the kernel information into consideration. The first is represented by Pan \textit{et al.} \cite{pami2020_pan_physics_gan}, who introduced a loss term motivated by the physical blur model \eqref{blurmodel}. This loss term plays the similar role as the likelihood
in ELBOs \eqref{KF_elbo} and \eqref{KD_elbo} under our framework. The second is by Kaufman \textit{et al.} \cite{cvpr2020_kaufman_asnet}, who designed a network to estimate the blur kernel from the blurry image, and input the estimated kernel to the main deblurring network as auxiliary information. This strategy is consistent with the kernel-dependent inference structure for $\bm{z}$ under our framework.
Compared with the two existing kernel-involved methodologies, our framework provides a more complete and concise unification, naturally by Bayesian inference principle without any tricky designing, which can not only explain the effectiveness of the existing method from Bayesian perspective, but also facilitate to network training as demonstrated in experiments.
\textbf{Incorporating with Existing Networks:} Though we have designed simple baseline networks in Section \ref{sec:net_arch}, our proposed framework is indeed very general, and can be incorporated with any existing DNNs designed for image deblurring task, as training guidance. As verified in Section \ref{sec:enhance}, the performance can be improved by virtue of the our Bayesian modeling and learning framework.
\section{Experimental Results}\label{sec:experiments}
In this section, we present the experimental results of the proposed blind image deblurring framework, in comparison with current state-of-the-art methods. We call our method \emph{Variational Bayesian Deblurring Network}, denoted as VBDeblurNet. More results are provided in Appendix \ref{app:more_results}.
\subsection{Experiment Settings}\label{sec:exp_settings}
\textbf{Datasets:} We evaluate the proposed method mainly on three datasets that are widely used for evaluating the deep learning based image deblurring methods, including \emph{Text}
by Hradiš \textit{et al.} \cite{bmvc15_hradis_cnn_deblur}, \emph{RealBlur}
by Rim \textit{et al.} \cite{eccv2020_rim_realblur} and \emph{GoPro}
by Nah \textit{et al.} \cite{cvpr17_nah_deepdeblur}. The three datasets represent typical situations we may encounter when applying a deep learning based method. Specifically, the \emph{Text} dataset is composed of blurry-clean-kernel image tuples with uniform blur, 60,000 for training and 1,000 for testing;
the \emph{RealBlur}\footnote{We use the \emph{RealBlur-J} subset in experiments.} dataset contains 3,758 blurry-clean image pairs for training and 980 for testing, while the blur kernels are approximately uniform; and the \emph{GoPro} dataset consists of 2,103 blurry-clean training pairs from 22 sequences and 1,111 testing pairs from 11 sequences, with non-uniform blur kernels. As mentioned in Section \ref{sec:model}, on the latter two datasets, the prior information of kernels can be estimated from training patches.
\textbf{Implementation:} The weights of networks are initialized according to \cite{cvpr2015_he_init}. In each iteration, we randomly crop the entire images to a size of $256 \times 256$ patch for training. The Adam optimizer \cite{iclr2015_kingma_adam} using default parameters is adopted for training, with initial learning rate $1e\text{-}4$, and being reduced to $1e\text{-}5$ as loss becoming stable and finally to $1e\text{-}6$. Without explicitly specified, we set the hyper-parameter $\sigma^2$ in (\ref{blur_likelihood}) to $1e\text{-}5$, $\varepsilon_0^2$ in (\ref{piror_z}) to $1e\text{-}6$, and $c$ in \eqref{prior_h} to $2e4$, respectively. Analyses for the effects of these hyper-parameters are provided in Appendix \ref{app:hyperparam}.
\subsection{Model Verification and Ablation Study}
We first construct experiments to verify the effectiveness of the proposed framework with ablation study. Since the \emph{Text} dataset contains the complete and accurate information of blur kernels, we can use it to fully verify the effectiveness of the proposed framework. For a illustration purpose, we use a relatively small \emph{Deblur-Net} with 4 ResBlocks in each encoding or decoding block of the U-Net. We consider four strategies, with different loss functions and inference structures, for obtaining a deblurring network. The first two are training under the supervision of naive mean squared error (MSE), with kernel-free and kernel-dependent inference structures, which are respectively denoted as MSE$_{\text{KF}}$ and MSE$_{\text{KD}}$. The latter two are the proposed framework with kernel-free and kernel-dependent inference structures, denoted as VBDeblurNet$_{\text{KF}}$ and VBDeblurNet$_{\text{KD}}$, respectively.
The deblurring performance of each strategy, in terms of \emph{peak signal-to-noise-ratio} (PSNR) and \emph{structure similarity} (SSIM) \cite{ssim}, is summarized in Table \ref{ablation_study_table_1}. As can be seen from the table, with the same inference structure, the performance can be largely improved by virtue of the proposed VBDeblurNet framework. Besides, even with the simpler kernel-free inference structure, the proposed VBDeblurNet$_{\text{KF}}$ can outperform MSE$_{\text{KD}}$ that uses kernel information in inference.
These results evidently verify the effectiveness of the proposed framework.
\begin{table}[!ht]
\caption{Deblurring results on \emph{Text} dataset, with different loss functions and inference structures. The best and the second best results are highlighted in bold and Italic bold, respectively.}
\label{ablation_study_table_1}
\vspace{-1mm}
\centering
\begin{adjustbox}{width=0.6\columnwidth,center}
\begin{tabular}{c|cccc}
\toprule
Method & MSE$_{\text{KF}}$ & VBDeblurNet$_{\text{KF}}$ & MSE$_{\text{KD}}$ & VBDeblurNet$_{\text{KD}}$ \\
\hline
PSNR & 28.70 & \textit{\textbf{29.07}} & 29.06 & \textbf{29.58} \\
SSIM & 0.9845 & \textbf{0.9860} & 0.9846 & \textit{\textbf{0.9855}} \\
\bottomrule
\end{tabular}
\end{adjustbox}
\vspace{-4mm}
\end{table}
\begin{table}[!t]
\caption{Comparison of competing methods in terms of PSNR and SSIM on \emph{Text} dataset.}
\label{text_result}
\vspace{-2mm}
\centering
\begin{adjustbox}{width=\columnwidth,center}
\begin{tabular}{c|cc|cccc|cc}
\toprule
Method & Xu \textit{et al.} \cite{cvpr2013_xu_l0norm} & Pan \textit{et al.} \cite{cvpr2016_pan_dark_channel} & Hradi{\v{s}} \textit{et al.} \cite{bmvc15_hradis_cnn_deblur}& Nah \textit{et al.} \cite{cvpr17_nah_deepdeblur} & Pan \textit{et al.} \cite{pami2020_pan_physics_gan} & Kaufman \textit{et al.} \cite{cvpr2020_kaufman_asnet} & VBDeblurNet$_{\text{KF}}$ & VBDeblurNet$_{\text{KD}}$ \\
\hline
PSNR & 17.52 & 18.47 & 26.53 & 26.81 &28.80 & 27.90 & \emph{\textbf{29.70}} & \textbf{30.38} \\
SSIM & 0.4186 & 0.6127 & 0.9422 & 0.9743 &0.9744 & 0.9604 & \emph{\textbf{0.9862}} & \textbf{0.9872} \\
\bottomrule
\end{tabular}
\end{adjustbox}
\vspace{-2mm}
\end{table}
\begin{figure}[!t]
\centerin
\includegraphics[width=0.98\textwidth]{show_text.pdf}
\vspace{-2mm}
\caption{Visual deblurring results of competing methods on \emph{Text} dataset, together with the estimated blur kernels if available.}
\label{fig_text}
\vspace{-2mm}
\end{figure}
\begin{table}[!t]
\caption{Comparison of competing methods in terms of PSNR and SSIM on \emph{RealBlur-J} dataset.}
\label{realblur-j_result}
\vspace{-1mm}
\centering
\begin{adjustbox}{width=\columnwidth,center}
\begin{tabular}{c|ccc|ccc|cc}
\toprule
Method & Hu \textit{et al.} \cite{cvpr2014_hu_low_light_deblurring} & Xu \textit{et al.} \cite{cvpr2013_xu_l0norm} & Pan \textit{et al.} \cite{cvpr2016_pan_dark_channel}& DeblurGAN-v2 \cite{iccv2019_Kupyn_deblurgan_v2} & SRN \cite{cvpr18_tao_srn} & MPRNet \cite{cvpr2021_zamir_mprnet} & VBDeblurNet$_{\text{KF}}$ & VBDeblurNet$_{\text{KD}}$ \\
\hline
PSNR & 26.41 & 27.14 & 27.22 & 29.69 & 31.38 & \emph{\textbf{31.76}} &\textbf{31.85} & 31.73\\
SSIM & 0.8028 & 0.8303 & 0.7901 & 0.8703 & 0.9091 & \textbf{0.9220} &\emph{\textbf{0.9132}} & 0.9106\\
\bottomrule
\end{tabular}
\end{adjustbox}
\vspace{-5mm}
\end{table}
\subsection{Comparison with State-of-the-art Methods}
In this subsection, we compared the proposed method with state-of-the-art methods, including traditional optimization-based methods \cite{cvpr2013_xu_l0norm, cvpr2014_pan_l0text, cvpr2016_pan_dark_channel, cvpr2014_hu_low_light_deblurring} and deep learning-based methods \cite{bmvc15_hradis_cnn_deblur, cvpr17_nah_deepdeblur, pami2020_pan_physics_gan, cvpr2020_ren_dip_deblur, cvpr2018_kupyn_deblurgan, iccv2019_Kupyn_deblurgan_v2, cvpr2021_zamir_mprnet, cvpr18_zhang_dynamic_rnn, cvpr2019_zhang_patch_deblurring}, on the three datasets introduced in Section \ref{sec:exp_settings}. The \emph{Deblur-Net} used here is with 6 Resblocks in each encoding or decoding block, in order to achieve a good performance. We show the quantitative results in terms of two commonly used metrics, PSNR and SSIM, in Tables \ref{text_result}, \ref{realblur-j_result} and \ref{gopro_result}. Note that the competing methods on different datasets are also different. This is because that we want to make sure that each deep learning based method has been trained and tested on the same dataset, by either citing the results from corresponding papers or retraining the models by ourselves.
It can be seen from Tables \ref{text_result} that, both versions of the proposed framework significantly improve the previous state-of-the-art results on \emph{Text} dataset. This is not surprising, since the structure of the training image tuples in this dataset can be faithfully characterized by our proposed Bayesian generative model, and thus the performance can be naturally benefited by our deep variational inference framework.
On the other two datasets, although the groundtruth blur kernels are unavailable, the benefits of the proposed framework can also be clearly observed from Tables \ref{realblur-j_result} and \ref{gopro_result}. Specifically, the proposed method performs at least comparable with the current state-of-the-art one, i.e., MPRNet, in terms of both PSNR and SSIM, especially that it outperforms MPRNet on \emph{RealBlur-J} dataset in terms of PSNR. The promising performance of MPRNet is not surprising, since authors have made a lot of efforts in network designing. Comparatively, our DNN architecture is much simpler, mainly built on classic structures, such as U-Net and ResBlock, and its performance is mostly attributed to the proposed VBDeblurNet framework. From this perspective, these quantitative results indeed substantiate the effectiveness of the proposed framework. Besides, we can also straightforwardly apply our framework to MPRNet, in order to further pursue better performance.
For a more intuitive illustration, we draw the example visual results in Figs. \ref{fig_text}, \ref{fig_realblur} and \ref{fig_gopro} on the three datasets, and more results are provided in Appendix \ref{app:more_results}. Fig. \ref{fig_text} shows visual deblurring results an image in \emph{Text} dataset by different deep learning methods contained in Table \ref{text_result}. As can be seen, the deblurred images by our method have better visual quality, especially for producing sharper edges and digits. Compared the two inference structures of our method, kernel-dependent one unsurprisingly performs better in more accurately recovering digit ``5'', since it makes full use of the kernel information. Besides, the estimated kernels by our method are also close to the ground truth. For the deblurring example on \emph{RealBlur-J} dataset, shown in Fig. \ref{fig_realblur}, our method produces slightly better results than that of MPRNet both in terms of PSNR and visual quality. Visual results on \emph{GoPro} dataset shown in Fig. \ref{fig_gopro} are also very interesting. Specifically, although MPRNet achieves higher PSNR, its visual quality is obviously not the best, especially that it fails to recover the number plate on the car. As a comparison, our VBDeblurNet$_{\text{KD}}$ method can relatively better recover this number plate, especially for digit ``4'' within it.
\begin{figure}[!t]
\centering
\includegraphics[width=\textwidth]{show_realblur_re.pdf}
\vspace{-6mm}
\caption{Visual deblurring results of competing methods on \emph{RealBlur-J} dataset.}
\label{fig_realblur}
\vspace{-2mm}
\end{figure}
\begin{table}[!t]
\caption{Comparison of competing methods in terms of PSNR and SSIM on \emph{GoPro} dataset.}
\label{gopro_result}
\vspace{-2mm}
\centering
\begin{adjustbox}{width=\columnwidth,center}
\begin{tabular}{c|ccccccc|cc}
\toprule
Method & Nah \textit{et al.} \cite{cvpr17_nah_deepdeblur} & DeblurGAN \cite{cvpr2018_kupyn_deblurgan} & SRN \cite{cvpr18_tao_srn}& Gao \textit{et al.} \cite{cvpr2019_gao_sharing_deblurring} & DBGAN \cite{cvpr2020_zhang_double_gan_deblurring} & Zhang \textit{et al.} \cite{cvpr2019_zhang_patch_deblurring} & MPRNet \cite{cvpr2021_zamir_mprnet} & VBDeblurNet$_{\text{KF}}$ & VBDeblurNet$_{\text{KD}}$ \\
\hline
PSNR & 29.08 & 28.70 & 30.21 & 30.92 & 31.10 & 31.20 & \textbf{32.66} & 31.87 & \emph{\textbf{32.03}} \\
SSIM & 0.9135 & 0.858 & 0.9352 & 0.9421 & 0.9424 & 0.940 & \textbf{0.9589} & 0.9519 & \emph{\textbf{0.9531}} \\
\bottomrule
\end{tabular}
\end{adjustbox}
\vspace{-2mm}
\end{table}
\begin{figure}[!t]
\centerin
\includegraphics[width=0.98\textwidth]{show_gopro.pdf}
\vspace{-2mm}
\caption{Visual deblurring results of competing methods on \emph{GoPro} dataset.}
\label{fig_gopro}
\end{figure}
\subsection{Enhancing Existing Networks by Proposed Framework}\label{sec:enhance}
As discussed in Section \ref{sec:discussion}, the proposed framework can be incorporated with existing DNNs, to improve their deblurring performances. To verify this, we train on the \emph{Text} dataset two existing deblurring networks, i.e., the ones proposed by Nah \textit{et al.} \cite{cvpr17_nah_deepdeblur}, and Kaufman and Fattal \cite{cvpr2020_kaufman_asnet}, under MSE loss and with our framework, respectively. The former network uses kernel-free inference structure and the latter adopt the kernel-dependent one. The results in terms of PSNR and SSIM are summarized in Table \ref{ablation_study_table_2}. As can be seen from the Table that, the performance of existing networks can be evidently improved by the proposed framework, no matter which types of inference structures they are built on. These improvements can be attributed to two advantages led to by our Bayesian modeling and learning framework. Firstly, the kernel information is naturally included in the training objective as auxiliary supervision. More importantly, \emph{Deblur-Net} and \emph{Kernel-Net} can interact with each other, under the guidance of the likelihood term led to by the physical blur model, to promote the overall training.
\begin{table}[!t]
\vspace{-2mm}
\caption{Comparison of existing networks trained with MSE and VBDeblurNet on \emph{Text} dataset.}
\label{ablation_study_table_2}
\vspace{-2mm}
\centering
\begin{adjustbox}{width=0.75\columnwidth,center}
\begin{tabular}{c|ccc|ccc}
\toprule
\multirow{2}{*}{Method} & \multicolumn{3}{c|}{Nah \textit{et al.} \cite{cvpr17_nah_deepdeblur}} & \multicolumn{3}{c}{Kaufman \textit{et al.} \cite{cvpr2020_kaufman_asnet}} \\
\cline{2-7}
& MSE & VBDeblurNet & $\Delta~\uparrow$ & MSE & VBDeblurNet & $\Delta~\uparrow$\\
\hline
PSNR & 26.81 & 27.03 & 0.22 & 27.90 & 28.10 & 0.20 \\
SSIM & 0.9743 & 0.9756 & 0.0013 & 0.9607 & 0.9636 & 0.0029 \\
\bottomrule
\end{tabular}
\end{adjustbox}
\vspace{-4mm}
\end{table}
\section{Conclusion}\label{sec:conclusion}
We have proposed a new deep variational Bayesian framework, i.e., VBDeblurNet, for blind image deblurring, under which, the approximate posterior of the latent clean image and blur kernel, parameterized by DNNs, can be jointly learned in an amortized inference fashion. By virtue of the fully Bayesian formulation, the inference networks for clean image and blur kernel can both be supervised by data-driven priors, and interact with each other by the likelihood attributed to the physical blur model during training, which thus leads to promising deblurring performance. Comprehensive experiments have verified the effectiveness of the proposed framework, showing that it can not only achieve a promising results with relatively simple networks, but also be able to enhance the deblurring performance of existing DNNs. In the future, we will devote to improving the applicability of VBDeblurNet in non-uniform deblurring.
\bibliographystyle{plain}
|
\section{Introduction}
In this paper, we study the stochastic reaction-diffusion equation on the whole line $ {\mathbb R}$ driven by multiplicative space-time white noise given as follows:
\begin{numcases}{}
\mathrm{d}u(t,x) = \frac{1}{2}\Delta u(t,x) \,\mathrm{d}t+ b(u(t,x)) \,\mathrm{d}t \nonumber\\
~~~~~~~~~~~~~ + \sigma(u(t,x)) \,W(\mathrm{d}t,\mathrm{d}x), \ t>0, x\in \mathbb{R} , \nonumber\\
\label{1.a} u(0,x)=u_0(x), \quad x\in \mathbb{R} .
\end{numcases}
The coefficients $b, \sigma$ are two deterministic measurable functions from $\mathbb{R}$ to from $\mathbb{R}$, $W$ is a space-time white noise on $\mathbb{R}_{+}\times \mathbb{R}$ defined on some filtrated probability space $(\Omega, {\cal F}, \{{\cal F}_t\}_{t\geq 0}, \mathbb{P})$.
\vskip 0.3cm
There exist numerous work in the literature on stochastic reaction-diffusion equations driven by space-time white noise covering a wide range of topics.
We refer the reader to \cite{DPZ}, \cite{cerrai-1}, \cite{K} and references therein. The majority of the work are focused on stochastic reaction-diffusion equations defined on finite intervals (i.e., the space variable belongs to a fixed finite interval) instead of the whole real line $\mathbb{R}$, partly due to the essential difficulties brought by the non-compactness of the whole space. We like to mention some relevant existing work on the well-posedness of the stochastic reaction-diffusion equations on the real line. In the early paper \cite{S}, the author obtained the existence and uniqueness of solutions of stochastic reaction-diffusion equations on the real line under the Lipschitz conditions of the coefficients. Pathwise uniqueness were established in \cite{MP} and \cite{MPS} for stochastic reaction-diffusion equations on the real line with H\"older continuous coefficients.
\vskip 0.4cm
It is well known that the equation (\ref{1.a}) admits a unique global solution when the coefficients fulfill the usual Lipschitz condition, in particular, being of linear growth. We are concerned here with the well-posedness of the stochastic reaction-diffusion equation (\ref{1.a}) with superlinear drift. Several papers in the literature discuss stochastic partial differential equations with locally Lipschitz coefficients that have polynomial growth and/or satisfy certain monotonicity conditions (see \cite{cerrai-1,DMP,LR}, for instance). The typical example of such a coefficient is $b(u) = -u^3$, which has the effect of ``pulling the solution back toward the origin.'' In the joint paper \cite{DKZ} with Dalang and Khoshnevisan by the second named author , stochastic reaction-diffusion equations(SRDEs) on finite intervals were considered and it was proved that
if the coefficients are locally Lipschitz and of $(|z|\log|z|)$-growth, then the SRDEs is globally well-posed. Unfortunately, the methods in \cite{DKZ} are not valid for SRDEs on the whole line $\mathbb{R}$ because typically the supremum norm of the solution explodes, i.e., $|u(t)|_{\infty}=\sup_{x\in {\mathbb R}}|u(t,x)|=\infty$ . The global well-posedness on the whole line $\mathbb{R}$ under the logarithmic nonlinearity was left open.
\vskip 0.4cm
The goal of this article is to fill in this gap. More precisely, we prove that
if the drift $b$ is locally Log-Lipschitz and if $|b(z)|=O(|z|\log|z|)$, then the stochastic reaction-diffusion equation \eqref{1.a} is globally well-posed. The precise statements are given in the next section. Because of the nature of the nonlinearity, we are forced to work with the first order moment of the solutions on the space $C_{tem}({\mathbb R})$ with a specially designed norm
$$\sup_{t\leq T, x\in\mathbb{R}}\left(|u(t,x)|e^{-\lambda |x|e^{\beta t}}\right),$$
where, unlike the usual norm in $C_{tem}({\mathbb R})$, the exponent also depends on time $t$ in a particular way. We need to establish some new, precise (lower order) moment estimates of stochastic convolution on the real line and hence obtain some a priori estimates of the solution. We like to stress that it is harder to get precise lower order moment estimate than high order for stochastic convolutions. To obtain the pathwise uniqueness, one of the difficulties is that we are not able to apply the usual localization procedure as in the literature because the usual uniform norm of the solution on the real line explodes. To overcome the difficulty, we provide a new type of Gronwall's inequalities, which is of independent interest.
\vskip 0.4cm
Now we describe the content of the paper in more details. In Section 2, we present the framework for the stochastic reaction-diffusion equations driven by space-time white noise and state our main results. In Section 3, we will prove two Gronwall-type inequalities and obtain some estimates associated with the heat kernel of the Laplacian operator. In Section 4, we establish new lower order moment estimates of the stochastic convolution with respect to the space-time white noise and obtain a priori estimates of the solutions of the stochastic reaction-diffusion equations. In Section 5, we approximate the coefficients $b$ and $\sigma$ by smooth functions and establish the tightness of the laws of the solutions of the corresponding approximating equations. As a consequence, we obtain the existence of weak solution ( in the probabilistic sense). Section 6 is devoted to the proof of the pathwise uniqueness of the stochastic reaction-diffusion equation under the local Log-Lipschitz conditions of the coefficients.
\section{Statement of main results}\label{S:2}
\setcounter{equation}{0}
Let us recall the following definition.
\begin{definition}
A random field solution to equation (\ref{1.a}) is a jointly measurable and adapted space-time process $u:=\{u(t,x): (t,x)\in\mathbb{R}_{+}\times\mathbb{R}\}$ such that for every $(t,x)\in \mathbb{R}_{+}\times \mathbb{R}$,
\begin{align}\label{definition solution}
u(t,x)= & P_t u_0(x)+\int_0^t\int_{\mathbb{R}} p_{t-s}(x,y) b(u(s,y)) \,\mathrm{d}s\mathrm{d}y \nonumber\\
& + \int_0^t\int_{\mathbb{R}} p_{t-s}\sigma(u(s,y)) \,W(\mathrm{d}s,\mathrm{d}y), \quad \mathbb{P}-a.s.,
\end{align}
where $p_{t}(x,y):=\frac{1}{\sqrt{2\pi t}} e^{-\frac{(x-y)^2}{2t}}$, and $\{P_t\}_{t\geq 0}$ is the corresponding heat semigroup on $\mathbb{R}$.
\end{definition}
\begin{remark}
The above mild form is equivalent to the weak (in the sense of partial differential equations) formulation of the stochastic reaction-diffusion equations. We refer readers to \cite{WA} for details.
\end{remark}
We also recall the so-called $C_{tem}$ space defined by
\begin{align*}
C_{tem}:=\left\{f\in C(\mathbb{R}): \sup_{x\in\mathbb{R}}|f(x)|e^{-\lambda |x|}<\infty \text{ for any } \lambda>0 \right\},
\end{align*}
and endow it with the metric defined by
\begin{align*}
d(f,g):=\sum_{n=1}^{\infty}\frac{1}{2^n}\min\left\{1, \sup_{x\in\mathbb{R}}|f(x)-g(x)|e^{-\frac{1}{n} |x|}\right\} ,
\end{align*}
for any $f,g\in C_{tem}$.
Then $f_n\rightarrow f$ in $C_{tem}$ iff $\sup_{x\in\mathbb{R}}|f_n(x)-f(x)|e^{-\lambda |x|}\rightarrow 0$ as $n\rightarrow\infty$ for any $\lambda>0$, and $(C_{tem}, d)$ is a Polish space.
\vskip 0.3cm
Next we introduce the following conditions of nonlinear term $b$. Set $\log_{+}(u):=\log_{+}(1\vee u)$ for any $u\geq 0$.
\begin{itemize}
\item [(H1)] $b$ is continuous, and there exist two nonnegative constants $c_1$ and $c_2$ such that for any $u\in\mathbb{R}$,
\begin{align}\label{210124.2000}
|b(u)|\leq c_1 |u|\log_{+}|u| + c_2.
\end{align}
\item [(H2)] There exist nonnegative constants $c_3, c_4, c_5$, such that for any $u,v\in\mathbb{R}$,
\begin{align}
|b(u)-b(v)|\leq c_3|u-v|\log_{+}\frac{1}{|u-v|}+ c_4 \log_{+}(|u|\vee|v|)|u-v| + c_5|u-v|.
\end{align}
\end{itemize}
Note that condition (H2) implies condition (H1). A typical example of function $b$ that satisfies (H2) is given below.
\vskip 0.4cm
\begin{example}\label{A.2}
The function $x\mapsto x\log|x|$ satisfies the local log-Lipschitz condition (H2), more precisely, for any $x,y\in\mathbb{R}$,
\begin{align}\label{05161157}
\left|x\log|x|-y\log|y|\right|\leq |x-y|\log\frac{1}{|x-y|} + [\log_{+}(|x|\vee|y|) +1 +\log 2]|x-y|.
\end{align}
\end{example}
\noindent {\bf Proof}.
Without loss of generality, we may assume $|y|\leq |x|$. We divide the proof into two cases.
Case 1: $|x| \leq 1$. We have
\begin{align}
\left|x\log|x|-y\log|y|\right|\leq & |x-y|\left|\log|x|\right|+ |y|\left|\log|x|-\log|y|\right| \nonumber\\
\leq & |x-y|\log\frac{1}{|x|} + |x-y| \nonumber\\
\leq & |x-y|\left(\log\frac{1}{|x-y|} +\log 2\right) +|x-y| ,
\end{align}
since $|x-y|\leq 2|x|$ leads to
\begin{align}
\log\frac{1}{|x|}\leq \log\frac{1}{|x-y|} + \log 2.
\end{align}
Case 2. $|x|\geq 1$. We have
\begin{align}
\left|x\log|x|-y\log|y|\right|\leq & |x-y|\left|\log|x|\right|+ |y|\left|\log|x|-\log|y|\right| \nonumber\\
\leq & |x-y|\times \log_{+}(|x|\vee|y|) +|x-y| .
\end{align}
Combining these two cases together yields (\ref{05161157}).
$\blacksquare$
\vskip 0.5cm
Now we can state the main results of this paper.
\begin{theorem}\label{thm1}
Assume $u_0 \in C_{tem}$ and that (H1) is satisfied. If $\sigma$ is bounded and continuous, then there exists a weak ( in the probabilistic sense) solution to the stochastic reaction-diffusion equation (\ref{1.a}) with sample paths a.s. in $C(\mathbb{R}_{+}, C_{tem})$.
\end{theorem}
\vskip 0.6cm
\begin{theorem}\label{thm3}
Assume $u_0 \in C_{tem}$ and that (H2) is satisfied. If $\sigma$ is bounded and Lipchitz, then the pathwise uniqueness holds for solutions of (\ref{1.a}) in $C(\mathbb{R}_{+}, C_{tem})$. Hence there exists a unique strong solution to (\ref{1.a}) in $C(\mathbb{R}_{+}, C_{tem})$.
\end{theorem}
\section{Preliminaries}
\setcounter{equation}{0}
In this section, we will provide two Gronwall-type inequalities which play an important role in this paper. Moreover, we also present some estimates associated with the heat kernel of the Laplacian operator which will be used in our analysis later.
\vskip 0.3cm
Lemma \ref{A.1} is a slight modification of Theorem 3.1 in \cite{W}, and is proved in Lemma 7.2 of \cite{SZ2}. We give a short proof here for completeness. Set $\log_{+}(r):= \log(r\vee 1)$.
\begin{lemma}\label{A.1}
Let $X, a, c_1, c_2$ be nonnegative functions on $\mathbb{R}_{+}$, $M$ an increasing function with $M(0)\geq 1$. Moreover, suppose that $c_1, c_2$ be integrable on finite time intervals. Assume that for any $t\geq 0$,
\begin{align}\label{5.3}
X(t)+a(t)\leq M(t)+\int_0^t c_1(s)X(s)\,\mathrm{d}s +\int_0^t c_2(s) X(s)\log_{+} X(s) \,\mathrm{d}s ,
\end{align}
and the above integral is finite. Then for any $t\geq 0$,
\begin{align}\label{5.4}
X(t)+a(t)\leq M(t)^{\exp(C_2(t))}\exp\left(\exp(C_2(t))\int_0^t c_1(s)\exp(-C_2(s))\,\mathrm{d}s\right) ,
\end{align}
where $C_2(t):=\int_0^t c_2(s)\,\mathrm{d}s $.
\end{lemma}
\noindent {\bf Proof}. Fix any $T>0$. Let
\[
Y(t):= M(T)+\int_0^t c_1(s)X(s)\,\mathrm{d}s +\int_0^t c_2(s) X(s)\log_{+} X(s) \,\mathrm{d}s, \quad t\in[0,T] .
\]
We see that $Y$ is almost surely differentiable on $[0,T]$, $Y(t)\geq 1$ and
\[
X(t)+a(t)\leq Y(t), \quad \forall\, t\in [0,T] .
\]
This leads to
\begin{align}\label{5.5}
Y^{\prime}(t)= & c_1(t)X(t)+ c_2(t) X(t)\log_{+} X(t) \nonumber\\
\leq & c_1(t)Y(t)+ c_2(t) Y(t)\log_{+} Y(t) \nonumber\\
= & c_1(t)Y(t)+ c_2(t) Y(t)\log Y(t) .
\end{align}
Thus,
\begin{align}\label{5.6}
\left(\log Y\right)^{\prime}(t)\leq c_1(t)+ c_2(t)\log Y(t) .
\end{align}
Solving this ordinary differential inequality, we get for any $t\in [0,T]$,
\begin{align}\label{5.7}
\log Y(t) \leq \exp(C_2(t))\left[\log M(T) + \int_0^t c_1(s)\exp(-C_2(s))\,\mathrm{d}s\right] .
\end{align}
Therefore, we obtain
\begin{align}\label{5.8}
& X(T)+a(T) \leq Y(T) \nonumber\\
\leq & M(T)^{\exp(C_2(T))}\exp\left(\exp(C_2(T))\int_0^T c_1(s)\exp(-C_2(s))\,\mathrm{d}s\right) .
\end{align}
By the arbitrariness of $T$, (\ref{5.4}) is deduced.
$\blacksquare$
\begin{lemma}\label{A.3}
Let $Y(t)$ be a nonnegative function on $\mathbb{R}_+$. Let $c_1$ and $c_2$ be non-negative, increasing functions on $\mathbb{R}_+$. Let $\varepsilon\in [0,1)$ be a constant and $c_3: \mathbb{R}_{+}\times(\varepsilon,1)\longmapsto\mathbb{R}_{+}$ be a function that is increasing with respect the first variable. Suppose that for any $\theta\in (\varepsilon,1)$, the following integral inequality holds
\begin{align}\label{210217.2130}
Y(t)\leq c_1(t)\int_0^t Y(s)\,\mathrm{d}s + c_2(t)\int_0^t Y(s)\log_{+}\frac{1}{Y(s)}\,\mathrm{d}s + c_3(t,\theta)\int_0^t Y(s)^{\theta} \,\mathrm{d}s ,\quad \forall\, t\geq 0 .
\end{align}
If for any $t>0$,
\begin{align}\label{210219.1709}
\limsup_{\theta\rightarrow 1-} \, (1-\theta) c_3(t,\theta) <\infty ,
\end{align}
then $Y(t)= 0$ for any $t\geq 0$. In particular, if $c_3(t,\theta) \leq \frac{c(t)}{1-\theta}$ and $c$ is an increasing function with respect to $t$, then (\ref{210219.1709}) holds.
\end{lemma}
\noindent {\bf Proof}.
It suffices to prove that for any $T>0$, $Y(\cdot)=0$ on $[0,T]$. In order to prove this, let
\begin{align}
\delta_T := \limsup_{\theta\rightarrow 1-} \, (1-\theta) c_3(T,\theta), \nonumber\\
T^* : = \min \left\{ T, \frac{1}{3\delta_T}, \frac{e}{3 c_2(T)} \right\} .
\end{align}
Step 1. We first prove $Y(t)=0$ for any $t\in[0,T^*]$. Since
\begin{align}
\sup_{x\geq 0}\left(x\log_{+}\frac{1}{x}\right)=\frac{1}{e} ,
\end{align}
we have
\begin{align}
Y(t)\leq & c_1(t)\int_0^t Y(s)\,\mathrm{d}s + \frac{c_2(t)}{1-\theta}\int_0^t Y(s)^{\theta}Y(s)^{1-\theta}\log_{+}\frac{1}{Y(s)^{1-\theta}}\,\mathrm{d}s \nonumber\\
& + c_3(t,\theta)\int_0^t Y(s)^{\theta} \,\mathrm{d}s \nonumber\\
\leq & c_1(t)\int_0^t Y(s)\,\mathrm{d}s + \left[ \frac{c_2(t)}{e(1-\theta)} + c_3(t,\theta)\right] \int_0^t Y(s)^{\theta}\,\mathrm{d}s ,\quad \forall\, t\geq 0 .
\end{align}
For $t\in[0,T]$, let
\begin{align}
\Phi(t) := c_1(T) \int_0^t Y(s) \,\mathrm{d}s + \left[ \frac{c_2(T)}{e(1-\theta)} + c_3(T,\theta)\right] \int_0^t Y(s)^{\theta}\,\mathrm{d}s .
\end{align}
Then $Y(t)\leq \Phi(t)$ for any $t\in[0,T]$. Thus,
\begin{align}
\frac{\mathrm{d}}{\mathrm{d}t}\Phi(t) = & c_1(T) Y(t) + \left[ \frac{c_2(T)}{e(1-\theta)} + c_3(T,\theta)\right] Y(t)^{\theta} \nonumber\\
\leq & c_1(T)\Phi(t) + \left[\frac{c_2(T)}{e(1-\theta)} + c_3(T,\theta)\right] \Phi(t)^{\theta} .
\end{align}
Without loss of generality, we can assume that $\Phi(t)>0$ for any $t\in(0,T]$, otherwise we can take the zero time to be $\min\{t: \Phi(t)>0\}$.
Multiplying $(1-\theta)\Phi(t)^{-\theta}$ on both sides of the above inequality yields
\begin{align}
\frac{\mathrm{d}}{\mathrm{d}t}\left(\Phi(t)^{1-\theta}\right) \leq (1-\theta) c_1(T)\Phi(t)^{1-\theta} + \left[ \frac{c_2(T)}{e} + c_3(T,\theta)(1-\theta)\right].
\end{align}
Solving the above inequality, we obtain
\begin{align}
\Phi(t)^{1-\theta}\leq \left[ \frac{c_2(T)}{e} + c_3(T,\theta)(1-\theta)\right] \int_0^t e^{(1-\theta)c_1(T)(t-s)} \,\mathrm{d}s .
\end{align}
Hence
\begin{align}
Y(t)\leq & \Phi(t) \leq \left\{ \left[\frac{c_2(T)T^*}{e} + (1-\theta) c_3(T,\theta)T^* \right] e^{(1-\theta)c_1(T) T^*} \right\}^{\frac{1}{1-\theta}} \nonumber\\
\leq & e^{c_1(T)T^*} \left\{ \frac{c_2(T)T^*}{e} + (1-\theta) c_3(T,\theta)T^* \right\}^{\frac{1}{1-\theta}} ,
\end{align}
for any $t\in[0,T^*]$. Letting $\theta\rightarrow 1$ in the above inequality and in view of the definition of $T^*$, we can see that
\begin{align}
Y(t)=0,\quad \forall\, t\in[0,T^*].
\end{align}
Step 2.
From (\ref{210217.2130}) it follows that
\begin{align}
Y(t+T^*)\leq & c_1(T)\int_0^{t+T^*} Y(s)\,\mathrm{d}s + c_2(T)\int_0^{t+T^*} Y(s)\log_{+}\frac{1}{Y(s)}\,\mathrm{d}s \nonumber\\
& + c_3(T,\theta)\int_0^{t+T^*} Y(s)^{\theta} \,\mathrm{d}s \nonumber\\
\leq & c_1(T)\int_{T^*}^{t+T^*} Y(s)\,\mathrm{d}s + c_2(T)\int_{T^*}^{t+T^*} Y(s)\log_{+}\frac{1}{Y(s)}\,\mathrm{d}s \nonumber\\
& + c_3(T,\theta)\int_{T^*}^{t+T^*} Y(s)^{\theta} \,\mathrm{d}s \nonumber\\
\leq & c_1(T)\int_{0}^{t} Y(s+T^*)\,\mathrm{d}s + c_2(T)\int_{0}^{t} Y(s+T^*)\log_{+}\frac{1}{Y(s)}\,\mathrm{d}s \nonumber\\
& + c_3(T,\theta)\int_{0}^{t} Y(s+T^*)^{\theta} \,\mathrm{d}s ,\quad t\in [0, T-T^*] .
\end{align}
By Step 1, we see that
\begin{align}
Y(t+T^*) = 0, \quad t\in[0, T^*\wedge(T-T^*)] ,
\end{align}
that is $Y(t)=0$ for any $t\in[0, T\wedge 2T^* ]$. Repeating this argument, we see that $Y(t)=0$ for any $t\in [0,T]$. The arbitrariness of $T$ leads $Y(t)=0$ for any $t>0$.
$\blacksquare$
\vskip 0.3cm
Recall that $p_{t}(x,y):=\frac{1}{\sqrt{2\pi t}} e^{-\frac{(x-y)^2}{2t}}$. Now we state some estimates of the heat kernel $p_t(x,y)$ used in this paper.
For any $x\in\mathbb{R}$ and $t>0$,
\begin{align}
\label{05132049.1} \int_{\mathbb{R}} p_{t}(x,y)e^{\eta |y|} \,\mathrm{d}y \leq & 2e^{\frac{\eta^2 t}{2}}e^{\eta |x|}, \quad \forall\ \eta\in\mathbb{R} , \\
\label{2102121.2100} \int_{\mathbb{R}} p_{t}(x,y)^2 e^{\eta |y|} \,\mathrm{d}y \leq & \frac{1}{\sqrt{\pi t}}e^{\frac{\eta^2 t}{4}}e^{\eta |x|}, \quad \forall\ \eta\in\mathbb{R}, \\
\label{05161718} \int_{\mathbb{R}}p_{t}(x,y) e^{\eta |y|} \eta |y| \,\mathrm{d}y \leq & e^{\frac{\eta^2 t}{2}}e^{\eta |x|}\eta|x| + 2e^{\frac{\eta^2 t}{2}}\left(\eta^2 t+ \eta\sqrt{\frac{t}{2\pi}}\right) e^{\eta |x|} ,
\quad \forall\ \eta >0 .
\end{align}
The above three estimates can be obtained through straightforward calculations. Moreover, the following
lemma is needed.
\begin{lemma}\label{A.4}
The following estimates of the heat kernel $p_{t}(x,y)$ hold.
\begin{itemize}
\item [(i)] For any $x,y\in\mathbb{R}$, $\theta\in[0,1]$, $0< s\leq t$,
\begin{align}
|p_t(x,y)-p_s(x,y)|\leq \frac{(2\sqrt{2})^\theta |t-s|^{\theta}}{s^{\theta}}\big(p_s(x,y)+p_t(x,y)+p_{2t}(x,y)\big) .
\end{align}
\item [(ii)] For any $x,y\in\mathbb{R}$ and $t>0$,
\begin{align}
\int_{\mathbb{R}}|p_t(x,z)-p_t(y,z)|\,\mathrm{d}z\leq \sqrt{\frac{2}{\pi}}\times \frac{|x-y|}{\sqrt{t}}.
\end{align}
\item [(iii)] For any $x,y\in\mathbb{R}$ and $\eta, t>0$,
\begin{align}
\int_{\mathbb{R}}|p_t(x,z)-p_t(y,z)|e^{\eta |z|}\,\mathrm{d}z\leq 2\sqrt{2}\times \frac{|x-y|}{\sqrt{t}}\times e^{\eta^2 t}\times e^{\eta (|x|+|x-y|)}.
\end{align}
\item [(iv)] For any $x,y\in\mathbb{R}$ and $\eta, t>0$,
\begin{align}
& \int_{\mathbb{R}}|p_t(x,z)-p_t(y,z)|e^{\eta |z|}\eta |z| \,\mathrm{d}z \nonumber\\
\leq & \frac{\sqrt{2}|x-y|}{\sqrt{t}} \times \bigg[ e^{\eta^2 t}\times e^{\eta (|x|+|x-y|)}\eta (|x|+|x-y|) \nonumber\\
& + 2e^{ \eta^2 t }\bigg(2 \eta^2 t+ \eta\sqrt{\frac{t}{\pi}}\bigg) e^{\eta (|x|+|x-y|)}\bigg].
\end{align}
\item [(v)] For any $x,y\in\mathbb{R}$ and $0<s\leq t$,
\begin{align}\label{210222.1639}
\int_0^{s} \int_{\mathbb{R}} |p_{t-r}(x,z) -p_{s-r}(y,z)|^2 \,\mathrm{d}r\mathrm{d}z \leq \frac{\sqrt{2}-1}{\sqrt{\pi}} |t-s|^{\frac{1}{2}} + \frac{2}{\sqrt{\pi}}|x-y| .
\end{align}
\end{itemize}
\end{lemma}
\noindent {\bf Proof}.
Proof of (i). On the one hand, by the mean value theorem, there exists some $\xi\in[s,t]$ such that
\begin{align}\label{05161649.1}
& |p_t(x,y)-p_s(x,y)| \nonumber\\
\leq & \frac{1}{\sqrt{2\pi t}}\left|e^{-\frac{(x-y)^2}{2t}}-e^{-\frac{(x-y)^2}{2s}}\right| + \left|\frac{1}{\sqrt{2\pi t}}-\frac{1}{\sqrt{2\pi s}}\right|e^{-\frac{(x-y)^2}{2s}} \nonumber\\
\leq & \frac{1}{\sqrt{2\pi t}} e^{-\frac{(x-y)^2}{2\xi}}\times \frac{(x-y)^2}{2\xi^2}\times|t-s|+ \frac{1}{\sqrt{2\pi}}\times \frac{|t-s|}{2s^{3/2}}e^{-\frac{(x-y)^2}{2s}} \nonumber\\
\leq & \frac{2}{\sqrt{2\pi t}} e^{-\frac{(x-y)^2}{2\xi}}\times e^{\frac{(x-y)^2}{4\xi}}\times \frac{|t-s|}{s} + \frac{|t-s|}{2s}p_s(x,y) \nonumber\\
\leq & 2\sqrt{2}\times \frac{|t-s|}{s}p_{2t}(x,y) + \frac{|t-s|}{2s}p_s(x,y) \nonumber\\
\leq & \frac{2\sqrt{2}|t-s|}{s}\big(p_s(x,y)+p_t(x,y)+p_{2t}(x,y)\big) ,
\end{align}
where we have used $z\leq e^z$ for any $z\geq 0$. On the other hand,
\begin{align}\label{05161649.2}
|p_t(x,y)-p_s(x,y)|\leq & p_s(x,y)+ p_t(x,y) + p_{2t}(x,y) .
\end{align}
Combining (\ref{05161649.1}) and (\ref{05161649.2}) together yields (i).
Proof of (ii). Obviously,
\begin{align}\label{05161703}
|p_t(x,z)-p_t(y,z)| = & \frac{1}{\sqrt{2\pi t}}\bigg|\int_0^1 d e^{-\frac{|x-z+\rho (y-x)|^2}{2t}}\bigg| \nonumber\\
\leq & \frac{|x-y|}{\sqrt{2\pi t}} \times \int_0^1 e^{-\frac{|x-z+\rho (y-x)|^2}{2t}} \times \frac{|x-z+\rho (y-x)|}{t} \,\mathrm{d}\rho .
\end{align}
Due to the Fubini theorem,
\begin{align}
& \int_{\mathbb{R}}|p_t(x,z)-p_t(y,z)|\,\mathrm{d}z \nonumber\\
\leq & \int_0^1 \frac{|x-y|}{t} \int_{\mathbb{R}}\frac{1}{\sqrt{2\pi t}} |x-z+\rho (y-x)| e^{-\frac{|x-z+\rho (y-x)|^2}{2t}} \,\mathrm{d}z \mathrm{d}\rho \nonumber\\
\leq & \int_0^1 \frac{|x-y|}{t} \sqrt{\frac{2t}{\pi}} \,\mathrm{d}\rho = \sqrt{\frac{2}{\pi t}}|x-y| .
\end{align}
Proof of (iii).
Note that $z\leq e^{\frac{z^2}{4}}$ for any $z\geq 0$. By (\ref{05161703}), we have
\begin{align}\label{05161703.1}
|p_t(x,z)-p_t(y,z)| \leq \sqrt{\frac{2}{t}}|x-y|\times \int_0^1 \frac{1}{\sqrt{2\pi\times 2t}} e^{-\frac{|x+\rho (y-x)-z|^2}{4t}} \,\mathrm{d}\rho .
\end{align}
The Fubini theorem and (\ref{05132049.1}) leads to
\begin{align}
& \int_{\mathbb{R}}|p_t(x,z)-p_t(y,z)|e^{\eta |z|}\,\mathrm{d}z \nonumber\\
\leq & \sqrt{\frac{2}{t}}|x-y| \int_0^1 \int_{\mathbb{R}} \frac{1}{\sqrt{2\pi\times 2t}} e^{-\frac{|x+\rho (y-x)-z|^2}{4t}} e^{\eta|z|} \,\mathrm{d}z \mathrm{d}\rho \nonumber\\
\leq & \sqrt{\frac{2}{t}}|x-y| \times \int_0^1 2 e^{\eta^2 t}\times e^{\eta |x+\rho (y-x)|} \,\mathrm{d}\rho \nonumber\\
\leq & 2\sqrt{2}\times \frac{|x-y|}{\sqrt{t}}\times e^{\eta^2 t}\times e^{\eta (|x|+|x-y|)} .
\end{align}
Proof of (iv). By (\ref{05161703.1}), the Fubini theorem and (\ref{05161718}), we have
{\allowdisplaybreaks \begin{align}
& \int_{\mathbb{R}}|p_t(x,z)-p_t(y,z)|e^{\eta |z|}\eta |z| \,\mathrm{d}z \nonumber\\
\leq & \sqrt{\frac{2}{t}}|x-y| \int_0^1 \int_{\mathbb{R}} \frac{1}{\sqrt{2\pi\times 2t}} e^{-\frac{|x+\rho (y-x)-z|^2}{4t}} e^{\eta|z|}\eta |z| \,\mathrm{d}z \mathrm{d}\rho \nonumber\\
\leq & \sqrt{\frac{2}{t}}|x-y| \int_0^1 \bigg[e^{\eta^2 t}\times e^{\eta |x+\rho (y-x)|}\eta|x+\rho(y-x)| \nonumber\\
& + 2e^{\eta^2 t}\bigg(2\eta^2 t+ \eta\sqrt{\frac{t}{\pi}}\bigg) e^{\eta |x+\rho (y-x)|}\bigg] \,\mathrm{d}\rho \nonumber\\
\leq & \frac{\sqrt{2}|x-y|}{\sqrt{t}} \times \bigg[ e^{\eta^2 t}\times e^{\eta (|x|+|x-y|)}\eta (|x|+|x-y|) \nonumber\\
& + 2e^{ \eta^2 t }\bigg(2\eta^2 t+ \eta\sqrt{\frac{t}{\pi}}\bigg)e^{\eta (|x|+|x-y|)}\bigg].
\end{align} }
Proof of (iv). This inequality can be found in Lemma 6.2 of \cite{S}, here we just give the explicit constant by straightforward calculations.
This completes the proof of Lemma \ref{A.4}.
$\blacksquare$
\section{Moment estimates}
\setcounter{equation}{0}
In this section, we will establish estimates for moments of stochastic convolutions, and obtain some a priori estimates for solutions of equation (\ref{1.a}).
We begin with the estimates of high order moments of stochastic convolutions.We stress that the precise lower order moment estimates are harder to get.
\begin{lemma}\label{210121.1036}
Let $h: \mathbb{R}_{+} \longmapsto \mathbb{R}_{+}$ be an increasing function.
Let $\{\sigma(s,y): (s,y)\in\mathbb{R}_+\times [0,1]\}$ be a random field such that the following stochastic convolution with respect to space time white noise is well defined.
Let $\tau$ be a stopping time.
Then for any $p>10$ and $T>0$, there exists a constant $C_{p, h(T), T}>0$ such that
\begin{align}\label{210119.2121}
& \mathbb{E} \sup_{(t,x)\in [0,T\wedge\tau]\times \mathbb{R}}\Bigg\{\left|\int_0^t\int_{\mathbb{R}}p_{t-s}(x,y)\sigma(s,y)\,W(\mathrm{d}s,\mathrm{d}y)\right| e^{-h(t)|x|}\Bigg\}^p \nonumber\\
\leq & C_{p, h(T), T}\,{\mathbb{E}}\int_0^{T\wedge\tau} \int_{\mathbb{R}}|\sigma(t,x)|^p e^{-p h(t) |x|} \, \mathrm{d}x \mathrm{d}t .
\end{align}
In particular, if $\sigma$ is bounded and $h$ is a positive constant, then the left hand side of (\ref{210119.2121}) is finite.
\end{lemma}
\noindent {\bf Proof}.
We employ the factorization method (see e.g. \cite{DPZ}). The proof here is inspired by \cite{SZ1}. Choose $\frac{3}{2p}<\alpha<\frac{1}{4}-\frac{1}{p}$. This is possible because we assume $p>10$. Let
\begin{align}
(J_{\alpha}\sigma)(s,y):&= \int_0^s\int_{\mathbb{R}} (s-r)^{-\alpha}p_{s-r}(y,z)\sigma(r,z)\,W(\mathrm{d}r,\mathrm{d}z), \\
(J^{\alpha-1}f)(t,x):&= \frac{\sin\pi\alpha}{\pi}\int_0^t\int_{\mathbb{R}} (t-s)^{\alpha-1}p_{t-s}(x,y)f(s,y)\,\mathrm{d}s\mathrm{d}y.
\end{align}
\noindent From the stochastic Fubini theorem (see Theorem 2.6 in \cite{W}), it follows that for any $(t,x)\in\mathbb{R}_+\times \mathbb{R}$,
\begin{align}\label{103.1}
\int_0^t\int_{\mathbb{R}} p_{t-s}(x,y)\sigma(s,y)\,W(\mathrm{d}s,\mathrm{d}y)=J^{\alpha-1}(J_{\alpha}\sigma)(t,x).
\end{align}
\noindent By H\"{o}lder's inequality and (\ref{2102121.2100}), we have
{\allowdisplaybreaks\begin{align}\label{104.1}
& {\mathbb{E}}\sup_{(t,x)\in[0,T\wedge\tau]\times\mathbb{R}}\bigg\{\left|\int_0^t\int_{\mathbb{R}} p_{t-s}(x,y)\sigma(s,y)\,W(\mathrm{d}s,\mathrm{d}y)\right| e^{-h(t) |x|} \bigg\}^p \nonumber\\
=& {\mathbb{E}}\sup_{(t,x)\in[0,T\wedge\tau]\times\mathbb{R}}\bigg\{\left|\frac{\sin\pi\alpha}{\pi} \int_0^t\int_{\mathbb{R}} (t-s)^{\alpha-1} p_{t-s}(x,y)J_{\alpha}\sigma(s,y)\,\mathrm{d}s\mathrm{d}y\right| e^{-h(t) |x|}\bigg\}^p \nonumber\\
\leq & \left|\frac{\sin\pi\alpha}{\pi}\right|^p {\mathbb{E}}\sup_{(t,x)\in[0,T\wedge\tau]\times\mathbb{R}}\bigg\{\bigg[\int_0^t (t-s)^{\alpha-1} \nonumber\\
&~~~~~~~~~~~~~\times\left(\int_{\mathbb{R}} p_{t-s}(x,y)|J_{\alpha}\sigma(s,y)|\,\mathrm{d}y\right)\,\mathrm{d}s\bigg]^p e^{-ph(t) |x|}\bigg\} \nonumber\\
\leq & \left|\frac{\sin\pi\alpha}{\pi}\right|^p {\mathbb{E}}\sup_{(t,x)\in[0,T\wedge\tau]\times \mathbb{R}}\Bigg\{\bigg[\int_0^t (t-s)^{\alpha-1} \nonumber\\
&~~~~~~~~~~~~~\times\left(\int_{\mathbb{R}} p_{t-s}(x,y)e^{\frac{ph(t)}{2}|y|}e^{-\frac{ph(t)}{2}|y|}|J_{\alpha}\sigma(s,y)|^{\frac{p}{2}}\,\mathrm{d}y\right)^{\frac{2}{p}}\,\mathrm{d}s\bigg]^p e^{-p h(t) |x|}\Bigg\}\nonumber\\
\leq & \left|\frac{\sin\pi\alpha}{\pi}\right|^p {\mathbb{E}}\sup_{(t,x)\in[0,T\wedge\tau]\times \mathbb{R}}\Bigg\{\bigg[\int_0^t (t-s)^{\alpha-1} \left(\int_{\mathbb{R}} p_{t-s}(x,y)^2 e^{ph(t) |y|}\,\mathrm{d}y\right)^{\frac{1}{2}\times\frac{2}{p}} \nonumber\\
&~~~~~~~~~~~~~\times\left(\int_{\mathbb{R}}|J_{\alpha}\sigma(s,y)|^p e^{-ph(t) |y|}\,\mathrm{d}y\right)^{\frac{1}{2}\times\frac{2}{p}}\,\mathrm{d}s\bigg]^p e^{-ph(t) |x|}\Bigg\} \nonumber\\
\leq & \left|\frac{\sin\pi\alpha}{\pi}\right|^p \frac{1}{\sqrt{\pi}} e^{\frac{p^2h(T)^2}{4}T} \times {\mathbb{E}}\sup_{t\in[0,T\wedge\tau]} \Bigg[\int_0^t (t-s)^{\alpha-1-\frac{1}{2p}} \nonumber\\
& ~~~~~~~~~~~~~~\times \left(\int_{\mathbb{R}}|J_{\alpha}\sigma(s,y)|^p e^{-ph(t) |y|}\,\mathrm{d}y\right)^{\frac{1}{p}}\,\mathrm{d}s\Bigg]^p \nonumber\\
\leq & \left|\frac{\sin\pi\alpha}{\pi}\right|^p \frac{1}{\sqrt{\pi}} e^{\frac{p^2h(T)^2}{4}T} \times {\mathbb{E}}\sup_{t\in[0,T\wedge\tau]}\Bigg\{ \left(\int_0^t (t-s)^{(\alpha-1-\frac{1}{2p})\frac{p}{p-1}}\,\mathrm{d}s\right)^{\frac{p-1}{p}\times p} \nonumber\\
&~~~~~~~~~~~~~\times\left(\int_0^t\int_{\mathbb{R}}|J_{\alpha}\sigma(s,y)|^p e^{-ph(t) |y|}\,\mathrm{d}y\mathrm{d}s\right)^{\frac{1}{p}\times p} \Bigg\} \nonumber\\
\leq & \left|\frac{\sin\pi\alpha}{\pi}\right|^p \frac{1}{\sqrt{\pi}} e^{\frac{p^2h(T)^2}{4}T} \times \left(\int_0^T s^{(\alpha-1-\frac{1}{2p})\frac{p}{p-1}}\,\mathrm{d}s\right)^{p-1} \nonumber\\
& ~~~~~~~~~~~~~\times\int_0^T\int_{\mathbb{R}} {\mathbb{E}}\left[|J_{\alpha}\sigma(s,y)|^p \mathbbm{1}_{[0, \tau]}(s) \right] e^{-ph(s) |y|}\,\mathrm{d}y\mathrm{d}s \nonumber\\
\leq & C_{p, h(T), T, \alpha}^{\prime} \int_0^T\int_{\mathbb{R}} {\mathbb{E}}\Big[ \left|\int_0^s\int_{\mathbb{R}} (s-r)^{-\alpha}p_{s-r}(y,z)\sigma(r,z)\,W(\mathrm{d}r,\mathrm{d}z)\right|^p \nonumber\\
&~~~~~~~~~~~~~~~~~~~~~~~~~~~~ \times\mathbbm{1}_{[0, \tau]}(s) \Big] e^{-ph(s) |y|}\,\mathrm{d}y\mathrm{d}s ,
\end{align}}
\noindent where we have used the condition $\alpha >\frac{3}{2p}$ in the last inequality, so that
\begin{align}\label{210217.1615}
C_{p, h(T), T, \alpha}^{\prime} =& \left|\frac{\sin\pi\alpha}{\pi}\right|^p \frac{1}{\sqrt{\pi}} e^{\frac{p^2h(T)^2}{4}T} \times \left(\int_0^T s^{(\alpha-1-\frac{1}{2p})\frac{p}{p-1}}\,\mathrm{d}s\right)^{p-1} \nonumber\\
=& \left|\frac{\sin\pi\alpha}{\pi}\right|^p \frac{1}{\sqrt{\pi}} e^{\frac{p^2h(T)^2}{4}T} \times \left(\frac{p-1}{\alpha p-\frac{3}{2}}\right)^{p-1} T^{\alpha p -\frac{3}{2}} .
\end{align}
For any fixed $(s,y)\in[0,T]\times\mathbb{R}$, let
\begin{align}
M_t:= \int_0^t\int_{\mathbb{R}} (s-r)^{-\alpha}p_{s-r}(y,z)\sigma(r,z) \mathbbm{1}_{[0, \tau]}(r) \,W(\mathrm{d}r,\mathrm{d}z), \quad t\in [0,s] .
\end{align}
Then it is easy to see that $\{M_t\}_{t\in [0,s]}$ is a martingale. Applying the Bukrholder-Davis-Gundy inequality (see Proposition 4.4 in \cite{K} and also \cite{W}), we have for $t\in [0,s]$,
\begin{align}\label{210217.1647}
\mathbb{E}|M_t|^p \leq & (4p)^{\frac{p}{2}} \,\mathbb{E}\langle M\rangle_t^{\frac{p}{2}} \nonumber\\
=&(4p)^{\frac{p}{2}}\,\mathbb{E}\left(\int_0^{t\wedge\tau}\int_{\mathbb{R}} (s-r)^{-2\alpha}p_{s-r}(y,z)^2\sigma(r,z)^2\,\mathrm{d}r\mathrm{d}z\right)^{\frac{p}{2}}.
\end{align}
Hence by the local property of the stochastic integral (see Lemma A.1 in Appendix of \cite{SZ1}), we get
\begin{align}\label{210217.1647.1}
& {\mathbb{E}} \left[\left|\int_0^s\int_{\mathbb{R}} (s-r)^{-\alpha}p_{s-r}(y,z)\sigma(r,z)\,W(\mathrm{d}r,\mathrm{d}z)\right|^p \mathbbm{1}_{[0, \tau]}(s) \right]\nonumber\\
= & {\mathbb{E}} \left[\left|\int_0^s\int_{\mathbb{R}} (s-r)^{-\alpha}p_{s-r}(y,z)\sigma(r,z)\mathbbm{1}_{[0, \tau]}(r) \,W(\mathrm{d}r,\mathrm{d}z)\right|^p \mathbbm{1}_{[0, \tau]}(s) \right]\nonumber\\
\leq & {\mathbb{E}} \left|\int_0^s\int_{\mathbb{R}} (s-r)^{-\alpha}p_{s-r}(y,z)\sigma(r,z)\mathbbm{1}_{[0, \tau]}(r) \,W(\mathrm{d}r,\mathrm{d}z)\right|^p \nonumber\\
\leq & (4p)^{\frac{p}{2}}\,\mathbb{E}\left(\int_0^{s\wedge\tau}\int_{\mathbb{R}} (s-r)^{-2\alpha}p_{s-r}(y,z)^2\sigma(r,z)^2\,\mathrm{d}r\mathrm{d}z\right)^{\frac{p}{2}} .
\end{align}
Note that $p_t(x,y)\leq(2\pi t)^{-\frac{1}{2}}$ for any $t>0$ and $x,y\in\mathbb{R}$. Using (\ref{210217.1647.1}), H\"{o}lder's inequality and the Fubini theorem, we get
\begin{align}\label{210217.1612}
& \int_0^T\int_{\mathbb{R}} {\mathbb{E}}\left[ \left|\int_0^s\int_{\mathbb{R}} (s-r)^{-\alpha}p_{s-r}(y,z)\sigma(r,z)\,W(\mathrm{d}r,\mathrm{d}z)\right|^p \mathbbm{1}_{[0, \tau]}(s) \right] e^{-ph(s) |y|}\,\mathrm{d}y\mathrm{d}s \nonumber\\
\leq & (4p)^{\frac{p}{2}}\int_0^T\int_{\mathbb{R}}{\mathbb{E}}\left\{\int_0^{s\wedge\tau}\int_{\mathbb{R}} (s-r)^{-2\alpha}p_{s-r}^2(y,z)\sigma^2(r,z)\, \mathrm{d}r \mathrm{d}z \right\}^{\frac{p}{2}} e^{-ph(s) |y|}\,\mathrm{d}y\mathrm{d}s \nonumber\\
\leq & \left(\frac{4p}{\sqrt{2\pi}}\right)^{\frac{p}{2}}\int_0^T\int_{\mathbb{R}}{\mathbb{E}}\Bigg\{\left[\int_0^{s\wedge\tau}\int_{\mathbb{R}} \bigg|(s-r)^{-2\alpha -\frac{1}{2}}p_{s-r}^{1-\frac{2}{p}}(y,z)\bigg|^{\frac{p}{p-2}}\, \mathrm{d}r \mathrm{d}z\right]^{\frac{p-2}{p}} \nonumber\\
&~~~~~~~~~~~\times \left[\int_0^{s\wedge\tau}\int_{\mathbb{R}} \bigg|p_{s-r}^{\frac{2}{p}}(y,z)\sigma^2(r,z)\bigg|^{\frac{p}{2}}\, \mathrm{d}r \mathrm{d}z\right]^{\frac{2}{p}}
\Bigg\}^{\frac{p}{2}} e^{-ph(s) |y|}\,\mathrm{d}y\mathrm{d}s \nonumber\\
\leq & \left(\frac{4p}{\sqrt{2\pi}}\right)^{\frac{p}{2}} \int_0^T \left[\int_0^s (s-r)^{-(2\alpha +\frac{1}{2}) \frac{p}{p-2} }\, \mathrm{d}r \right]^{\frac{p-2}{2}} \nonumber\\
&~~~~~~~~~~~\times {\mathbb{E}} \int_0^{s\wedge\tau}\int_{\mathbb{R}} \left(\int_{\mathbb{R}}p_{s-r}(y,z)e^{-ph(s) |y|}\,\mathrm{d}y \right) |\sigma(r,z)|^p \,\mathrm{d}z \mathrm{d}r\mathrm{d}s \nonumber\\
\leq & C_{p, h(T), T, \alpha}^{\prime\prime} \,{\mathbb{E}} \int_0^{T\wedge\tau}\int_{\mathbb{R}} |\sigma(r,z)|^p e^{-ph(r)|z|} \, \mathrm{d}z \mathrm{d}r ,
\end{align}
where (\ref{05132049.1}) and the condition $\alpha<\frac{1}{4}-\frac{1}{p}$ are used to see that
\begin{align}\label{210217.1618}
C_{p, h(T), T, \alpha}^{\prime\prime} = & \left(\frac{4p}{\sqrt{2\pi}}\right)^{\frac{p}{2}} \times T \times \left(\int_0^T r^{-(2\alpha +\frac{1}{2}) \frac{p}{p-2} }\, \mathrm{d}r \right)^{\frac{p-2}{2}} \times 2e^{\frac{p^2h(T)^2}{2}T}\nonumber\\
= & \left(\frac{4p}{\sqrt{2\pi}}\right)^{\frac{p}{2}} \times \left(\frac{p-2}{\frac{p}{2}-2-2\alpha p}\right)^{\frac{p-2}{2}} T^{\frac{p}{4}-\alpha p}\times 2e^{\frac{p^2h(T)^2}{2}T} .
\end{align}
Combining (\ref{104.1}) with (\ref{210217.1612}), we obtain
\begin{align}
& \mathbb{E} \sup_{(t,x)\in[0,T\wedge\tau]\times \mathbb{R}}\Bigg\{\left|\int_0^t\int_{\mathbb{R}}p_{t-s}(x,y)\sigma(s,y)W(\mathrm{d}s,\mathrm{d}y)\right| e^{-h(t)|x|}\Bigg\}^p \nonumber\\
\leq & C_{p, h(T), T} {\mathbb{E}} \int_0^{T\wedge\tau}\int_{\mathbb{R}} |\sigma(r,z)|^p e^{-ph(r)|z|} \, \mathrm{d}z \mathrm{d}r ,
\end{align}
where
\begin{align}\label{C_{T,p}}
C_{p, h(T), T}= \min_{\frac{3}{2p}<\alpha<\frac{1}{4}-\frac{1}{p}}C^{\prime}_{p, h(T), T, \alpha}\times C^{\prime\prime}_{p, h(T), T, \alpha} .
\end{align}
In view of (\ref{210217.1615}) and (\ref{210217.1618}), a straightforward calculation leads to
\begin{align}\label{210217.2153}
C_{p, h(T), T} < 2 \sqrt{2}\, p^{\frac{p}{2}} \left(\frac{2}{\pi}\right)^p \left(\frac{1}{\sqrt{2\pi}}\right)^{\frac{p}{2}+1} \left(\frac{6p-8}{p-10}\right)^{\frac{3p}{2}-2} T^{\frac{p}{4}-\frac{3}{2}} e^{\frac{3 p^2h(T)^2}{4}T}.
\end{align}
$\blacksquare$
\begin{proposition}\label{estimates 003}
Let $h: \mathbb{R}_{+} \longmapsto \mathbb{R}_{+}$ be an increasing function.
Let $\{\sigma(s,y): (s,y)\in\mathbb{R}_+\times [0,1]\}$ be a random field such that the following stochastic convolution with respect to the space time white noise is well defined.
Let $\tau$ be a stopping time.
Then for any $\epsilon, T>0$, and $0<p\leq 10$, there exists a constant $C_{\epsilon, p, h(T), T}$ such that
\begin{align}\label{101.2}
& \mathbb{E} \sup_{(t,x)\in [0,T\wedge\tau]\times \mathbb{R}}\Bigg\{\left|\int_0^t\int_{\mathbb{R}}p_{t-s}(x,y)\sigma(s,y)\,W(\mathrm{d}s,\mathrm{d}y)\right| e^{-h(t)|x|}\Bigg\}^p \nonumber\\
\leq & \epsilon \,\mathbb{E} \sup_{(t,x)\in[0,T\wedge\tau]\times\mathbb{R}}\left(|\sigma(t,x)| e^{-h(t)|x|} \right)^p \nonumber\\
& + C_{\epsilon,p,h(T),T} \,\mathbb{E}\int_0^{T\wedge\tau}\int_{\mathbb{R}} \left|\sigma(t,x)\right|^p e^{-ph(t)|x|}\,\mathrm{d}x\mathrm{d}t .
\end{align}
\end{proposition}
\begin{remark}
The constant $C_{\epsilon,p,h(T),T}$ is increasing with respect to $T$ and $C_{\epsilon,p,h(0),0}=0$ .
\end{remark}
\noindent {\bf Proof}. The following proof is inspired by \cite{SZ1}. The proof is divided into two steps.
Step 1. We first show that for any $\rho, T>0$ and $q>10$,
\begin{align}\label{102.1}
& \mathbb{P}\left(\sup_{(t,x)\in[0,T\wedge\tau]\times\mathbb{R}}\left[\left|\int_0^t\int_{\mathbb{R}} p_{t-s}(x,y)\sigma(s,y)\,W(\mathrm{d}s,\mathrm{d}y) \right|e^{-h(t)|x|}\right]>\rho\right) \nonumber\\
\leq & \mathbb{P}\left(\int_0^{T\wedge\tau}\int_{\mathbb{R}}\left|\sigma(s,y)\right|^q e^{-qh(s)|y|}\,\mathrm{d}y\mathrm{d}s >\rho^q\right) \nonumber\\
& + \frac{C_{q,h(T),T}}{\rho^q}\mathbb{E}\min\left\{\rho^q, \int_0^{T\wedge\tau}\int_{\mathbb{R}}\left|\sigma(s,y)\right|^q e^{-qh(t)|y|}\,\mathrm{d}y\mathrm{d}s\right\}.
\end{align}
Here the constant $C_{q,h(T), T}$ is the constant $C_{p,h(T),T}$ in (\ref{210119.2121}) with $p$ replaced by $q$.
To prove (\ref{102.1}), we set
\begin{align}
\Omega_{\rho}:=\left\{\omega\in\Omega: \int_0^{T\wedge\tau} \int_{\mathbb{R}}|\sigma(s,y)|^q e^{-qh(s)|y|}\,\mathrm{d}y\mathrm{d}s\leq \rho^q\right\} .
\end{align}
By Chebyshev's inequality, we have
\begin{align}\label{105.1}
& \mathbb{P}\left(\sup_{(t,x)\in[0,T\wedge\tau]\times\mathbb{R}}\left[\left|\int_0^t\int_{\mathbb{R}} p_{t-s}(x,y)\sigma(s,y)\,W(\mathrm{d}s,\mathrm{d}y) \right|e^{-h(t)|x|}\right]>\rho\right) \nonumber\\
\leq & \mathbb{P}(\Omega\backslash\Omega_{\rho}) + \mathbb{P}\left(\sup_{(t,x)\in[0,T\wedge\tau]\times\mathbb{R}}\left[\left|\int_0^t\int_{\mathbb{R}} p_{t-s}(x,y)\sigma(s,y)\,W(\mathrm{d}s,\mathrm{d}y) \right| \mathbbm{1}_{\Omega_{\rho}} e^{-h(t)|x|} \right] >\rho\right) \nonumber\\
\leq & \mathbb{P}\left(\int_0^{T\wedge\tau}\int_{\mathbb{R}}\left|\sigma(s,y)\right|^q e^{-qh(s)|y|}\,\mathrm{d}y\mathrm{d}s >\rho^q\right) \nonumber\\
& + \frac{1}{\rho^q} \mathbb{E}\sup_{(t,x)\in[0,T\wedge\tau]\times\mathbb{R}} \left\{\left|\mathbbm{1}_{\Omega_{\rho}}\int_0^t\int_{\mathbb{R}} p_{t-s}(x,y)\sigma(s,y)\,W(\mathrm{d}s,\mathrm{d}y) \right|^q e^{-qh(t)|x|}\right\} .
\end{align}
Now, we introduce the random field
\begin{align}
\widetilde{\sigma}(s,y):= \sigma(s,y)\mathbbm{1}_{\left\{\omega\in\Omega: \ \int_0^{s\wedge\tau}\int_{\mathbb{R}}|\sigma(r,y)|^q e^{-qh(r)|y|} \,\mathrm{d}y\mathrm{d}r \leq\rho^q\right\}}.
\end{align}
Note that the stochastic integral of $\widetilde{\sigma}(\cdot, \cdot)$ with respect to the space time white noise $W$ is well defined.
By the local property of the stochastic integral (see Lemma A.1 in Appendix of \cite{SZ1}),
\begin{align}
& \mathbbm{1}_{\Omega_{\rho}}\int_0^t\int_{\mathbb{R}} p_{t-s}(x,y)\sigma(s,y)\,W(\mathrm{d}s,\mathrm{d}y) \nonumber \\
= & \mathbbm{1}_{\Omega_{\rho}}\int_0^t\int_{\mathbb{R}} p_{t-s}(x,y)\widetilde{\sigma}(s,y)\,W(\mathrm{d}s,\mathrm{d}y), \quad \mathbb{P}-a.s..
\end{align}
Hence using the bound (\ref{210119.2121}), we get
\begin{align}\label{105.2}
& \mathbb{E}\sup_{(t,x)\in[0,T\wedge\tau]\times\mathbb{R}} \left\{\left|\mathbbm{1}_{\Omega_{\rho}}\int_0^t\int_{\mathbb{R}} p_{t-s}(x,y)\sigma(s,y)\,W(\mathrm{d}s,\mathrm{d}y) \right|^q e^{-qh(t)|x|}\right\} \nonumber\\
\leq & \mathbb{E}\sup_{(t,x)\in[0,T\wedge\tau]\times\mathbb{R}} \left\{\left|\int_0^t\int_{\mathbb{R}} p_{t-s}(x,y)\widetilde{\sigma}(s,y)\,W(\mathrm{d}s,\mathrm{d}y) \right|^q e^{-qh(t)|x|}\right\} \nonumber\\
\leq & C_{q,h(T),T} \,\mathbb{E}\int_0^{T\wedge\tau} \int_{\mathbb{R}} \left|\widetilde{\sigma}(s,y)\right|^q e^{-qh(s)|y|}\,\mathrm{d}s \nonumber\\
\leq & C_{q,h(T),T} \,\mathbb{E}\min\left\{\rho^q, \int_0^{T\wedge\tau}\int_{\mathbb{R}}\left|\sigma(s,y)\right|^q e^{-qh(s)|y|}\,\mathrm{d}y\mathrm{d}s\right\} .
\end{align}
Combining (\ref{105.1}) with (\ref{105.2}), we obtain (\ref{102.1}).
\vskip 0.3cm
Step 2.
Let now $0<p\leq 10$. From (\ref{102.1}) and Lemma A.2 in Appendix of \cite{SZ1}, it follows that
{\allowdisplaybreaks
\begin{align}\label{210217.2157}
& \mathbb{E} \sup_{(t,x)\in[0,T\wedge\tau]\times\mathbb{R}}\left\{\left|\int_0^t\int_{\mathbb{R}} p_{t-s}(x,y)\sigma(s,y)\,W(\mathrm{d}s,\mathrm{d}y) \right| e^{-h(t)|x|} \right\}^p \nonumber\\
= & \int_0^{\infty} p\rho^{p-1} \mathbb{P}\left(\sup_{(t,x)\in[0,T\wedge\tau]\times\mathbb{R}}\left[\left|\int_0^t\int_{\mathbb{R}} p_{t-s}(x,y)\sigma(s,y)\,W(\mathrm{d}s,\mathrm{d}y) \right| e^{-h(t)|x|}\right]>\rho\right)\,\mathrm{d}\rho \nonumber\\
\leq & \int_0^{\infty} p\rho^{p-1} \mathbb{P}\left(\int_0^{T\wedge\tau} \int_{\mathbb{R}}|\sigma(s,y)|^q e^{-qh(s)|y|}\,\mathrm{d}y\mathrm{d}s>\rho^q\right)\,\mathrm{d}\rho \nonumber\\
& + C_{q,h(T),T}\int_0^{\infty} p\rho^{p-1-q}\,\mathbb{E}\min\left\{\rho^q, \int_0^{T\wedge\tau} \int_{\mathbb{R}}|\sigma(s,y)|^q e^{-qh(s)|y|}\,\mathrm{d}y\mathrm{d}s\right\}\,\mathrm{d}\rho \nonumber\\
= & C_{p,q,h(T),T}\,\mathbb{E} \left[\int_0^{T\wedge\tau}\int_{\mathbb{R}} |\sigma(s,y)|^q e^{-qh(s)|y|}\,\mathrm{d}y\mathrm{d}s\right]^{\frac{p}{q}} \nonumber\\
\leq & C_{p,q,h(T),T} \,\mathbb{E}\Bigg[\sup_{(s,y)\in[0,T\wedge\tau]\times\mathbb{R}}\left(|\sigma(s,y)|e^{-h(s)|y|}\right)^{\frac{(q-p)p}{q}} \nonumber\\
& ~~~~~~~~~~~~~~~~\times \left(\int_0^{T\wedge\tau}\int_{\mathbb{R}}\left|\sigma(s,y)\right|^p e^{-ph(s)|y|}\,\mathrm{d}y \mathrm{d}s \right)^{\frac{p}{q}} \Bigg] \nonumber\\
\leq & \epsilon \,\mathbb{E} \sup_{(s,y)\in[0,T\wedge\tau]\times \mathbb{R}}\left(|\sigma(s,y)|e^{-h(s)|y|}\right)^p \nonumber\\
& + C_{p,q,h(T),T}\times C_{\epsilon,p,q,h(T),T}\, \mathbb{E}\int_0^{T\wedge\tau}\int_{\mathbb{R}}\left|\sigma(s,y)\right|^p e^{-ph(s)|y|}\,\mathrm{d}y \mathrm{d}s ,
\end{align}}
\noindent where
\begin{align}\label{C_Tpq}
C_{p,q,h(T),T}:=1+ \frac{q}{q-p} C_{q,h(T),T} ,
\end{align}
and we have used the following Young inequality
\begin{align}\label{C_Tpqe}
ab\leq & \frac{\epsilon}{C_{p,q,h(T),T}}\,a^{\frac{q}{q-p}} + C_{\epsilon,p,q,h(T),T}\,b^{\frac{q}{p}}, \quad \forall\, a,b>0, \nonumber\\ C_{\epsilon,p,q,h(T),T}:= & p\left(\frac{q-p}{\epsilon/C_{p,q,h(T),T}}\right)^{\frac{q-p}{p}} q^{-\frac{q}{p}} .
\end{align}
Set
\begin{align}\label{C_{T,p,epsilon}}
C_{\epsilon,p,h(T),T}:=\inf_{q>10} C_{p,q,h(T),T}\times C_{\epsilon,p,q,h(T),T} .
\end{align}
\noindent Combining (\ref{C_Tpq}) and (\ref{C_Tpqe}) gives
\begin{align}\label{C_Tpe}
C_{\epsilon,p,h(T),T}= \inf_{q>10} \left\{\frac{p}{q-p} q^{-\frac{q}{p}} \epsilon^{1-\frac{q}{p}} \left(q-p+qC_{q,h(T),T}\right)^{\frac{q}{p}}\right\} ,
\end{align}
where the constant $C_{q,h(T),T}$ is bounded by the right hand side of (\ref{210217.2153}) with $p$ replaced by $q$.
Now, (\ref{101.2}) follows from (\ref{210217.2157}) with the constant $C_{\epsilon,p,h(T),T}$ defined above.
$\blacksquare$
\vskip 0.3cm
Next, we will establish an a priori estimate of solutions to (\ref{1.a}). Throughout this paper, we will use the following notations. For $\lambda, \kappa >0$, set
\begin{align}
\beta(\lambda, \kappa):= & \max\left\{\frac{\lambda^2}{2}, 4\kappa \right\}, \\
T^*(\lambda, \kappa):= & \frac{1}{2\beta(\lambda,\kappa)}\left[1+\log\left(\frac{4\beta(\lambda,\kappa)}{\lambda^2}\log\frac{\beta(\lambda,\kappa)}{2\kappa}\right)\right] .
\end{align}
It is easy to see that for any $\kappa>0$, $T^*(\lambda,\kappa)\rightarrow\infty$ as $\lambda\rightarrow 0$.
\begin{lemma}\label{lemma 3.1}
Assume that (H1) is satisfied and $\sigma$ is bounded.
Let $u$ be a solution of (\ref{1.a}).
Set also
\begin{gather}
\label{2.2} V(t,x):=\int_0^t\int_{\mathbb{R}} p_{t-s}(x,y)\sigma(u(s,y)) \,W(\mathrm{d}s,\mathrm{d}y) .
\end{gather}
Then for any $\lambda>0$ and $T\leq T^*(\lambda, c_1)$, there exists a constant $C_{\lambda,c_1,T}$ such that the following a priori estimate holds for $\mathbb{P}$-a.s.,
\begin{align}\label{2.3}
& \sup_{t\leq T, x\in\mathbb{R}}\left(|u(t,x)|e^{-\lambda |x|e^{\beta t}}\right) \nonumber\\
\leq & C_{\lambda, c_1, T}\times \bigg\{1+ 2c_2 T +4e^{\frac{\lambda^2 T}{2}}\sup_{x\in\mathbb{R}}\left(|u_0(x)|e^{-\lambda |x|}\right) \nonumber\\
& + 2 \sup_{(t,x)\in[0,T]\times\mathbb{R}}\left(|V(t,x)|e^{-\lambda |x|}\right)\bigg\}^{e^{4c_1 T e^{\frac{\lambda^2}{4\beta} e^{2\beta T -1}}}} ,
\end{align}
where we write $\beta$ instead of $\beta(\lambda, c_1)$ for simplicity, and the constant $c_1$ is same as that in condition (H1).
\end{lemma}
\begin{remark}
Lemma \ref{lemma 3.1} actually implies that the solutions of (\ref{1.a}) don't blow up in the space $C_{tem}$, since we can take sufficiently small $\lambda>0$ such that $T^*(\lambda,c_1)$ can be larger than any given number.
\end{remark}
\vskip 0.3cm
\noindent {\bf Proof}.
Set
\[U(T):= \sup_{(t,x)\in [0,T]\times\mathbb{R}}\left(|u(t,x)|e^{-\lambda |x|e^{\beta t}}\right) .\]
From (\ref{definition solution}), we have
\begin{align}\label{200404.1559}
U(T) \leq & \sup_{t\leq T, x\in\mathbb{R}}\left(|P_t u_0(x)|e^{-\lambda |x|e^{\beta t}}\right) + \sup_{t\leq T, x\in\mathbb{R}}\left(|V(t,x)|e^{-\lambda |x|}\right) \nonumber\\
& + \sup_{t\leq T, x\in\mathbb{R}}\left\{\left|\int_0^t\int_{\mathbb{R}}p_{t-s}(x,y)b(u(s,y)) \,\mathrm{d}s\mathrm{d}y\right|e^{-\lambda |x|e^{\beta t}}\right\} .
\end{align}
\noindent Now we estimate the three terms on the right hand side of the above inequality.
\begin{align}\label{05161111}
& \sup_{t\leq T, x\in\mathbb{R}}\left(|P_t u_0(x)|e^{-\lambda |x|e^{\beta t}}\right) \nonumber\\
\leq & \sup_{t\leq T, x\in\mathbb{R}}\left\{\left|\int_{\mathbb{R}}p_t(x,y)u_0(y)\,\mathrm{d}y\right|e^{-\lambda |x|}\right\} \nonumber\\
\leq & \sup_{y\in\mathbb{R}}\left(|u_0(y)|e^{-\lambda |y|}\right)\times \sup_{t\leq T, x\in\mathbb{R}}\left\{\int_{\mathbb{R}}p_t(x,y) e^{\lambda |y|}\,\mathrm{d}y \times e^{-\lambda |x|}\right\} \nonumber\\
\leq & 2e^{\frac{\lambda^2 T}{2}} \sup_{y\in\mathbb{R}}\left(|u_0(y)|e^{-\lambda |y|}\right) ,
\end{align}
where we have used (\ref{05132049.1}).
Applying Lemma \ref{210121.1036} and using the boundedness of $\sigma$, we get that
for any $p, q, T>0$,
\begin{align}\label{05142145}
\mathbb{E}\sup_{t\leq T, x\in\mathbb{R}}\left(|V(t,x)|^p e^{-q |x|}\right)<\infty .
\end{align}
In particular,
\begin{align}\label{200404.1600}
\sup_{t\leq T, x\in\mathbb{R}}\left(|V(t,x)|e^{-\lambda |x|}\right)<+\infty , \quad \mathbb{P}-a.s..
\end{align}
\noindent The nonlinear term can be estimated as follows. By (\ref{210124.2000}) and $\log_{+}(ab)\leq \log_{+}a +\log_{+}b$ for any $a,b>0$, we have
\begin{align}\label{05161114}
&
\sup_{t\leq T, x\in\mathbb{R}}\left\{\left|\int_0^t\int_{\mathbb{R}}p_{t-s}(x,y)b(u(s,y))\,\mathrm{d}s\mathrm{d}y\right|e^{-\lambda |x|e^{\beta t}}\right\} \nonumber\\
\leq &
\sup_{t\leq T, x\in\mathbb{R}}\left\{\int_0^t\int_{\mathbb{R}}p_{t-s}(x,y)\left( c_1|u(s,y)|\log_{+} |u(s,y)| + c_2\right)\,\mathrm{d}s\mathrm{d}y \times e^{-\lambda |x|e^{\beta t}}\right\} \nonumber\\
\leq &
c_2 T + c_1\sup_{t\leq T, x\in\mathbb{R}}\bigg\{\int_0^t\int_{\mathbb{R}}p_{t-s}(x,y) e^{\lambda |y| e^{\beta s}}\times\left(|u(s,y)|e^{-\lambda |y|e^{\beta s}}\right) \nonumber\\
&~~~~~~~~~~~~~~~~~\times\log_{+}\left[\left(|u(s,y)|e^{-\lambda |y| e^{\beta s}}\right)\times e^{\lambda |y| e^{\beta s}}\right] \,\mathrm{d}s\mathrm{d}y \times e^{-\lambda |x|e^{\beta t}}\bigg\} \nonumber\\
\leq &
c_2 T
+ c_1\sup_{t\leq T, x\in\mathbb{R}}\bigg\{\int_0^t \sup_{y\in\mathbb{R}} \left[\left(|u(s,y)|e^{-\lambda |y|e^{\beta s}} \right)\times \log_{+}\left(|u(s,y)|e^{-\lambda |y| e^{\beta s}}\right)\right] \nonumber\\
&~~~~~~~~~~~~~~~~\times \int_{\mathbb{R}}p_{t-s}(x,y) e^{\lambda |y| e^{\beta s}} \,\mathrm{d}y\mathrm{d}s \times e^{-\lambda |x|e^{\beta t}}\bigg\} \nonumber\\
&
+ c_1\sup_{t\leq T, x\in\mathbb{R}}\bigg\{\int_0^t \sup_{y\in\mathbb{R}}\left(|u(s,y)|e^{-\lambda |y|e^{\beta s}}\right) \nonumber\\
&~~~~~~~~~~~~~~~\times\int_{\mathbb{R}}p_{t-s}(x,y) e^{\lambda |y| e^{\beta s}} \lambda |y| e^{\beta s} \,\mathrm{d}y\mathrm{d}s \times e^{-\lambda |x|e^{\beta t}}\bigg\} \nonumber\\
=& : c_2 T + I + II.
\end{align}
Note that the function $x\mapsto x\log_{+}x$ is increasing on $[0,\infty)$, so we have
\begin{align}\label{05161112}
I \leq &
c_1\sup_{t\leq T, x\in\mathbb{R}}\bigg\{\int_0^t \sup_{y\in\mathbb{R},r\leq s} \left[\left(|u(r,y)|e^{-\lambda |y|e^{\beta r}} \right)\times \log_{+}\left(|u(r,y)|e^{-\lambda |y| e^{\beta r}}\right)\right] \nonumber\\
& ~~~~~~~~~~~~~~\times 2 e^{\frac{\lambda^2 (t-s) e^{2\beta s}}{2}} e^{\lambda |x|e^{\beta s}} \,\mathrm{d}s \times e^{-\lambda |x|e^{\beta t}}\bigg\} \nonumber\\
\leq & 2c_1 \sup_{t\leq T}\bigg\{ \sup_{s\leq t}\left(e^{\frac{\lambda^2 (t-s) e^{2\beta s}}{2}}\right) \int_0^t U(s)\log_{+} U(s) \,\mathrm{d}s \bigg\} \nonumber\\
\leq & 2c_1 e^{\frac{\lambda^2}{4\beta}e^{2\beta T-1}} \int_0^T U(s)\log_{+} U(s) \,\mathrm{d}s ,
\end{align}
\noindent where we have used (\ref{05132049.1}) and
\begin{align}\label{05132049.2}
\max_{s\in [0,t]} e^{\frac{\lambda^2 (t-s) e^{2\beta s}}{2}} =e^{\frac{\lambda^2}{4\beta}e^{2\beta t-1}}.
\end{align}
For the term $II$, we estimate as follows
{\allowdisplaybreaks \begin{align}\label{05132036}
II \leq & c_1\sup_{t\leq T, x\in\mathbb{R}}\bigg\{\int_0^t \sup_{y\in\mathbb{R}}\left(|u(s,y)|e^{-\lambda |y|e^{\beta s}}\right) \nonumber\\
&~~~~~~~~~\times \left(e^{\frac{\lambda^2 (t-s) e^{2\beta s}}{2}}e^{\lambda |x| e^{\beta s}} \lambda |x| e^{\beta s} + C_{\lambda,\beta, t}e^{\lambda |x|e^{\beta s}} \right) \mathrm{d}s \times e^{-\lambda |x|e^{\beta t}}\bigg\} \nonumber\\
\leq & c_1\sup_{t\leq T, x\in\mathbb{R}}\bigg\{\sup_{s\leq t, y\in\mathbb{R}}\left(|u(s,y)|e^{-\lambda |y|e^{\beta s}}\right) \times \frac{1}{\beta}\sup_{s\leq t}\left(e^{\frac{\lambda^2 (t-s) e^{2\beta s}}{2}}\right)\nonumber\\
&~~~~~~~~~~\times \int_0^t \frac{\mathrm{d}}{\mathrm{d}s} e^{\lambda |x| e^{\beta s}} \,\mathrm{d}s\times e^{-\lambda |x|e^{\beta t}}\bigg\} \nonumber\\
& + c_1\sup_{t\leq T}\bigg\{C_{\lambda,\beta, t} \int_0^t \sup_{r\leq s, y\in\mathbb{R}}\left(|u(r,y)|e^{-\lambda |y|e^{\beta r}}\right) \mathrm{d}s \bigg\} \nonumber\\
\leq & \frac{c_1}{\beta}e^{\frac{\lambda^2}{4\beta}e^{2\beta T-1}} U(T) + C_{\lambda, c_1, T}\int_0^T U(s)\,\mathrm{d}s ,
\end{align}
}
\noindent where we have used (\ref{05161718}),
(\ref{05132049.2}), and that the constant $C_{\lambda,\beta,t}$ is increasing with respect to $t>0$. Note that
\begin{align}\label{05132040.1}
\frac{c_1}{\beta}e^{\frac{\lambda^2}{4\beta}e^{2\beta T-1}}\leq \frac{1}{2} \iff T\leq T^*(\lambda, c_1)= \frac{1}{2\beta}\left[1+\log\left(\frac{4\beta}{\lambda^2}\log\frac{\beta}{2c_1}\right)\right] .
\end{align}
Hence for $T\leq T^*(\lambda, c_1)$,
\begin{align}\label{05132040.2}
II \leq & \frac{1}{2} U(T) + c_1 C_{\lambda, \beta, T}\int_0^T U(s)\,\mathrm{d}s .
\end{align}
\noindent Combining (\ref{200404.1559}), (\ref{05161111}), (\ref{200404.1600}), (\ref{05161114}), (\ref{05161112}) and (\ref{05132040.2}) together, we obtain that for $T\leq T^*(\lambda, c_1)$,
\begin{align}
U(T)\leq & 2e^{\frac{\lambda^2 T}{2}} \sup_{y\in\mathbb{R}}\left(|u_0(y)|e^{-\lambda |y|}\right) + \sup_{t\leq T, x\in\mathbb{R}}\left(|V(t,x)| e^{-\lambda |x|}\right) \nonumber\\
& + c_2 T +2c_1 e^{\frac{\lambda^2}{4\beta}e^{2\beta T-1}} \int_0^T U(s)\log_{+} U(s) \,\mathrm{d}s \nonumber\\
& + \frac{1}{2} U(T) + C_{\lambda, c_1, T}\int_0^T U(s)\,\mathrm{d}s .
\end{align}
Subtracting $\frac{1}{2}U(T)$ on both sides of the above inequality, and then applying the log Gronwall inequality (see Lemma \ref{A.1}), (\ref{2.3}) is deduced.
$\blacksquare$
\section{Existence of weak solutions}
\setcounter{equation}{0}
In this section, we assume that (H1) is satisfied and that $\sigma$ is bounded, continuous.
We will approximate the coefficients $b$ and $\sigma$ by Lipschitz continuous functions and establish the existence of weak solutions of the stochastic reaction-diffusion equation.
Let $\varphi$ be a nonnegative smooth function on $\mathbb{R}$ such that the support of $\varphi$ is contained in $(-1,1)$ and $\int_{\mathbb{R}}\varphi(x)\,\mathrm{d}x=1$. Let $\{\eta_n\}_{n\geq 1}$ be a sequence of symmetric smooth functions such that for any $n\geq 1$, $0\leq \eta_n \leq 1$, $\eta_n(x)=1$ if $|x|\leq n$, and $\eta_n(x)=0$ if $|x|\geq n+2$.
Define
\begin{align}
b_n(x):=n\int_{\mathbb{R}}b(y)\varphi(n(x-y))\,\mathrm{d}y \times \eta_n(x) , \\
\sigma_n(x):=n\int_{\mathbb{R}}\sigma(y)\varphi(n(x-y))\,\mathrm{d}y \times \eta_n(x) .
\end{align}
Assume that $\sigma$ is bounded by a constant $K_{\sigma}$, that is
\begin{align}
|\sigma(z)|\leq K_{\sigma}, \quad \forall\, z\in\mathbb{R}.
\end{align}
Then it is easy to check that there exist constants $L_n, L_b$ and $K_n$ such that for any $x,y\in\mathbb{R}$,
\begin{align}
\label{05160929.1} |b_n(x)-b_n(y)|\leq & L_n|x-y| , \\
\label{05160929.2} |b_n(x)|\leq & c_1 |x|\log_{+}|x| + L_b(|x|+1 ), \\
\label{210122.1854} |\sigma_n(x)-\sigma_n(y)|\leq & K_n|x-y| , \\
\label{210122.1855} |\sigma_n(x)|\leq & K_{\sigma} ,
\end{align}
where the constant $c_1$ is same as that in condition (H1), and the constant $L_b$ is independent of $n$. Moreover, if $x_n\rightarrow x$ in $\mathbb{R}$, then
\begin{align}
\label{05160930.1}
b_n(x_n)\rightarrow b(x) , \\
\label{210126.2140}
\sigma_n(x_n)\rightarrow \sigma(x) .
\end{align}
For $n\geq 1$, consider the following stochastic equation on the real-line $\mathbb{R}$,
\begin{align}\label{05150908}
u_n(t,x)= & P_t u_0(x)+\int_0^t\int_{\mathbb{R}} p_{t-s}(x,y) b_n(u_n(s,y)) \,\mathrm{d}s\mathrm{d}y \nonumber\\
& + \int_0^t\int_{\mathbb{R}} p_{t-s}\sigma_n(u_n(s,y)) \,W(\mathrm{d}s,\mathrm{d}y).
\end{align}
It is known (see \cite{S,MPS,MP}) that for each $n\geq 1$, there exists a unique solution $u_n$ to the above equation. Moreover, the sample paths of $u_n$ are a.s. in $C(\mathbb{R}_{+},C_{tem})$.
The following result is a uniform bound for the solutions $u_n$.
\begin{lemma}\label{lemma 05151406}
Assume $u_0 \in C_{tem}$ and (H1). Suppose that $\sigma$ is bounded and continuous.
Then for any $p\geq 1$ and $\lambda, T >0$, we have
\begin{align}\label{05142144}
\sup_{n\geq 1} \mathbb{E}\left[\sup_{t\leq T, x\in\mathbb{R}} \left(|u_n(t,x)|e^{-\lambda |x|}\right)^p\right] <\infty .
\end{align}
\end{lemma}
\noindent {\bf Proof}. It suffices to prove this lemma for sufficiently large $p$ and sufficiently small $\lambda$. Fix $T>0$. As $\lim_{\lambda\rightarrow 0}T^*(\lambda, c_1)=\infty$, there exists a positive constant $\lambda_T$ such that $T\leq T^*(\lambda, c_1)$ for all $\lambda\leq \lambda_T$. For any fixed $\lambda\leq\lambda_T$, we choose $\lambda_0 >0$ so that $2\lambda_0 e^{\beta(\lambda_0, c_1) T} = \lambda$.
In the following we write $\beta$ for $\beta(\lambda_0, c_1)$ to simplify the notation.
Define
\begin{align}
U_n(r):=\sup_{t\leq r, x\in\mathbb{R}}\left(|u_n(t,x)|e^{-\lambda_0 |x|e^{\beta t}}\right),\quad r\in[0, T].
\end{align}
It remains to prove
\begin{align}\label{210225.1905}
\sup_{n\geq 1}\mathbb{E} [U_n(T)]^p <\infty .
\end{align}
From (\ref{05150908}) we have
\begin{align}
U_n(r) \leq & \sup_{t\leq r, x\in\mathbb{R}} \left(|P_t u_0(x)|e^{-\lambda_0 |x| e^{\beta t}}\right) \nonumber\\
& + \sup_{t\leq r, x\in\mathbb{R}}\left\{\left|\int_0^t\int_{\mathbb{R}}p_{t-s}(x,y)b_n(u_n(s,y))\,\mathrm{d}s\mathrm{d}y\right|e^{-\lambda_0 |x|e^{\beta t}}\right\} \nonumber\\
& + \sup_{t\leq r, x\in\mathbb{R}}\left\{\left|\int_0^t\int_{\mathbb{R}} p_{t-s}(x,y) \sigma_n(u_n(s,y))\,W(\mathrm{d}s,\mathrm{d}y)\right|e^{-\lambda_0 |x|e^{\beta t}}\right\} .
\end{align}
In the above inequality,
the first term can be estimated the same as (\ref{05161111}).
Let
\begin{align}
V_n(t,x) = \int_0^t\int_{\mathbb{R}} p_{t-s}(x,y) \sigma_n(u_n(s,y))\,W(\mathrm{d}s,\mathrm{d}y) .
\end{align}
Then by (\ref{210122.1855}) and Lemma \ref{210121.1036}, we have for any $p>0$,
\begin{align}\label{210124.2114}
& \sup_{n\geq 1}\mathbb{E} \sup_{t\leq T, x\in\mathbb{R}}\left\{\left|V_n(t,x)\right|e^{-\lambda_0 |x|e^{\beta t}}\right\}^p \nonumber\\
\leq & \sup_{n\geq 1}\mathbb{E}\sup_{t\leq T, x\in\mathbb{R}}\left\{\left|V_n(t,x)\right|e^{-\lambda_0 |x| }\right\}^p \nonumber\\
\leq & C_{\lambda_0, K_{\sigma}, T, p} <\infty.
\end{align}
On the other hand,
\begin{align}\label{4.0}
& \sup_{t\leq r, x\in\mathbb{R}}\left\{\left|\int_0^t\int_{\mathbb{R}}p_{t-s}(x,y)b_n(u_n(s,y))\,\mathrm{d}s\mathrm{d}y\right|e^{-\lambda_0 |x|e^{\beta t}}\right\} \nonumber\\
\leq & \sup_{t\leq r, x\in\mathbb{R}}\bigg\{\int_0^t\int_{\mathbb{R}}p_{t-s}(x,y)[ c_1|u_n(s,y)|\log_+ |u_n(s,y)| \nonumber\\
& ~~~~~~~~~~~~ +L_b(|u_n(s,y)|+1) ]\,\mathrm{d}s\mathrm{d}y\times e^{-\lambda_0 |x|e^{\beta t}}\bigg\} \nonumber\\
\leq & c_1\sup_{t\leq r, x\in\mathbb{R}}\left\{\int_0^t\int_{\mathbb{R}}p_{t-s}(x,y)|u_n(s,y)|\log_+ |u_n(s,y)|\,\mathrm{d}s\mathrm{d}y\times e^{-\lambda_0 |x|e^{\beta t}}\right\} \nonumber\\
& + L_b \sup_{t\leq r, x\in\mathbb{R}}\left\{\int_0^t\int_{\mathbb{R}}p_{t-s}(x,y) |u_n(s,y)|\,\mathrm{d}s\mathrm{d}y\times e^{-\lambda_0 |x|e^{\beta t}}\right\} + L_b r .
\end{align}
\noindent Now using (\ref{4.0}) and following a similar proof as that of Lemma \ref{lemma 3.1} we obtain
\begin{align}
U_n(T)\leq & C_{\lambda_0, c_1, L_b, T}\times \bigg\{1+2L_b T+4e^{\frac{\lambda_0^2 T}{2}}\sup_{x\in\mathbb{R}}\left(|u_0(x)|e^{-\lambda_0 |x|}\right) \nonumber\\
& + 2\sup_{t\leq T, x\in\mathbb{R}}\left(|V_n(t,x)|e^{-\lambda_0 |x|}\right)\bigg\}^{e^{4c_1 T e^{\frac{\lambda_0^2}{4\beta} e^{2\beta T -1}}}} .
\end{align}
Hence it follows from (\ref{210124.2114}) that
\begin{align}\label{210225.1854}
\mathbb{E} [U_n(T)^p] \leq & C_{\lambda_0, c_1, L_b, T, p}\left[1+\sup_{x\in\mathbb{R}}\left(|u_0(x)|e^{-\lambda_0 |x|}\right)^{p e^{4c_1 T e^{\frac{\lambda_0^2}{4\beta} e^{2\beta T -1}}}}\right] \nonumber\\\
& + C_{\lambda_0, c_1, L_b, T, p} \,\mathbb{E}\left[\sup_{t\leq T, x\in\mathbb{R}}\left(\left|V_n(t,x)\right|e^{-\lambda_0 |x| }\right)^{p e^{4c_1 T e^{\frac{\lambda_0^2}{4\beta} e^{2\beta T -1}}}}\right] \nonumber\\
\leq & C_{\lambda_0, c_1, L_b, K_{\sigma}, T, p, \|u_0\|_{\lambda_0,\infty}} ,
\end{align}
where $\|u_0\|_{\lambda_0,\infty}:= \sup_{x\in\mathbb{R}}(|u_0(x)|e^{-\lambda_0 |x|})$. Note that the last constant in (\ref{210225.1854}) is independent of $n$.
Hence (\ref{210225.1905}) is proved, completing the proof of the lemma.
\vskip 0.6cm
We will apply a Kolmogorov type tightness criterion (see Lemma 6.3 of \cite{S}) to establish the tightness of the law of $\{u_n\}$ in $C(\mathbb{R}_+, C_{tem})$. This is given in the following lemma.
Define
\begin{align}
X_n(t,x):=\int_0^t\int_{\mathbb{R}} p_{t-r}(x,z) b_n(u_n(r,z)) \,\mathrm{d}r\mathrm{d}z ,\quad n\geq 1 .
\end{align}
\begin{lemma}\label{05152151}
Let $u_0 \in C_{tem}$. Assume that (H1) holds and that $\sigma$ is continuous with $K_{\sigma}:=\sup_{z\in\mathbb{R}}|\sigma(z)| < \infty$.
Then for any $\lambda, T>0$, $p\geq 1$ and $\theta\in(0,1)$, there exist constants $C_{\lambda, c_1, L_b, K_{\sigma}, T, p, \theta, u_0}$ and $C_{K_{\sigma}, T, p}$ independent of $n$ such that
\begin{align}
\label{05152147}
\mathbb{E}\left(|X_n(t,x)-X_n(s,y)|^p e^{-\lambda |x|}\right)\leq & C_{\lambda, c_1, L_b, K_{\sigma}, T, p, \theta, u_0}\left(|t-s|^{\theta p}+|x-y|^p\right), \\
\label{210125.2123}
\mathbb{E}\left(|V_n(t,x)-V_n(s,y)|^p e^{-\lambda |x|}\right)\leq & C_{K_{\sigma}, T, p}\left(|t-s|^{\frac{p}{4}}+|x-y|^{\frac{p}{2}}\right) ,
\end{align}
for any $s,t\in [0,T]$ and $x,y\in\mathbb{R}$ with $|x-y|\leq 1$. In particular, the family $\{u_n\}$ is tight in $C(\mathbb{R}_+, C_{tem})$.
\end{lemma}
\noindent {\bf Proof}. It suffices to prove this lemma for sufficiently large $p$ and sufficiently small $\lambda$.
Fix $T>0$. As $\lim_{\lambda\rightarrow 0}T^*(\lambda, c_1)=\infty$, there exists a positive constant $\lambda_T$ such that $T\leq T^*(\lambda, c_1)$ for all $\lambda\leq \lambda_T$. For any fixed $\lambda\leq\lambda_T$, we choose $\lambda_0 >0$ so that
\begin{align}\label{210225.1936}
2\lambda_0 e^{\beta(\lambda_0, c_1) T} = \lambda .
\end{align}
In the following we write $\beta$ for $\beta(\lambda_0, c_1)$ to simplify the notation.
Without loss of generality, we assume $t\geq s$ .
{\allowdisplaybreaks\begin{align}\label{05152143}
& |X_n(t,x)-X_n(s,y)|e^{-\lambda |x| } \nonumber\\
\leq & |X_n(t,x)-X_n(s,x)|e^{-\lambda |x| } + |X_n(s,x)-X_n(s,y)|e^{-\lambda |x| }\nonumber\\
\leq & \left|\int_0^s\int_{\mathbb{R}} [p_{t-r}(x,z)-p_{s-r}(x,z)]b_n(u_n(r,z))\,\mathrm{d}r\mathrm{d}z\right|e^{-\lambda |x| } \nonumber\\
& + \left|\int_s^t\int_{\mathbb{R}} p_{t-r}(x,z) b_n(u_n(r,z))\,\mathrm{d}r\mathrm{d}z\right|e^{-\lambda |x| } \nonumber\\
& + \left|\int_0^s\int_{\mathbb{R}} [p_{s-r}(x,z)-p_{s-r}(y,z)] b_n(u_n(r,z))\,\mathrm{d}r\mathrm{d}z\right|e^{-\lambda |x| } \nonumber\\
=: & \, J_1 +J_2 +J_3 .
\end{align}}
\noindent By (i) of Lemma \ref{A.4}, (\ref{05161718}) and (\ref{05132049.1}), we have
\begin{align}
J_1
\leq & \int_0^s\int_{\mathbb{R}} |p_{t-r}(x,z)-p_{s-r}(x,z)|\times \nonumber\\
& \left[ c_1 |u_n(r,z)|\log_+ |u_n(r,z)|+ L_b(|u_n(r,z)|+1) \right]\,\mathrm{d}r\mathrm{d}z \times e^{-\lambda |x|} \nonumber\\
\leq & \int_0^s\int_{\mathbb{R}} \frac{(2\sqrt{2})^\theta |t-s|^{\theta}}{(s-r)^{\theta}}\big(p_{s-r}(x,z)+p_{t-r}(x,z)+p_{2(t-r)}(x,z)\big) \nonumber\\
& \times \bigg\{ c_ 1 e^{\lambda_0 |z|e^{\beta r}}\lambda_0 |z|e^{\beta r}\times \sup_{z\in\mathbb{R}}\left(|u_n(r,z)|e^{-\lambda_0 |z|e^{\beta r}}\right) \nonumber\\
& + c_1 e^{\lambda_0 |z|e^{\beta r}}\times\sup_{z\in\mathbb{R}}\left[|u_n(r,z)|e^{-\lambda_0 |z|e^{\beta r}}\log_+\left(|u_n(r,z)|e^{-\lambda_0 |z|e^{\beta r}}\right)\right] \nonumber\\
& + L_b e^{\lambda_0 |z|e^{\beta r}}\times \sup_{z\in\mathbb{R}}\left(|u_n(r,z)|e^{-\lambda_0 |z|e^{\beta r}}\right) +L_b \bigg\} \,\mathrm{d}z\mathrm{d}r \times e^{-\lambda |x| } \nonumber\\
\leq & (2\sqrt{2})^{\theta}|t-s|^{\theta}\times \int_0^s \frac{dr}{(s-r)^{\theta}}\times \bigg\{\sup_{r\leq T, z\in\mathbb{R}}\left(|u_n(r,z)|e^{-\lambda_0 |z|e^{\beta r}}\right) \nonumber\\
& \times C_{\lambda_0,c_1,L_b,T}\left(e^{\lambda_0 |x|e^{\beta r}}\lambda_0 |x|e^{\beta r}+ e^{\lambda_0 |x|e^{\beta r}}\right) \times e^{-\lambda|x| } \nonumber\\
& + \sup_{r\leq T, z\in\mathbb{R}}\left[|u_n(r,z)|e^{-\lambda_0 |z|e^{\beta r}}\log_+\left(|u_n(r,z)|e^{-\lambda_0 |z|e^{\beta r}}\right)\right]\nonumber\\
& \times C_{\lambda_0,c_1,L_b,T}\times e^{\lambda_0 |x|e^{\beta r}}\times e^{-\lambda|x|} + 3L_b \bigg\} \,\mathrm{d}r.
\end{align}
By the choice of $\lambda_0$,
\begin{align}\label{200405.2053}
e^{\lambda_0 |x|e^{\beta r}}\lambda_0 |x|e^{\beta r}\times e^{-\lambda|x|} \leq 1, \quad \forall\, r\in[0,T].
\end{align}
Therefore
\begin{align}\label{4.1}
J_1\leq & (2\sqrt{2})^{\theta}|t-s|^{\theta}\times \int_0^s \frac{dr}{(s-r)^{\theta}}\,\mathrm{d}r \times C_{\lambda_0,c_1,L_b,T}\nonumber\\
&\times
\Bigg\{\sup_{r\leq T, z\in\mathbb{R}}\left(|u_n(r,z)|e^{-\lambda_0 |z|e^{\beta r}}\right) + \sup_{r\leq T, z\in\mathbb{R}}\left(|u_n(r,z)|e^{-\lambda_0 |z|e^{\beta r}}\right) \nonumber\\
& \times \log_+\left[\sup_{r\leq T, z\in\mathbb{R}}\left(|u_n(r,z)|e^{-\lambda_0 |z|e^{\beta r}}\right)\right] + 1\Bigg\} .
\end{align}
By Lemma \ref{lemma 05151406}, (\ref{210225.1854}) and the fact that $\theta\in(0,1)$ , we deduce from (\ref{4.1}) that
\begin{align}\label{05152144.1}
E J_1^p\leq C_{\lambda_0,c_1,L_b,K_{\sigma},T,p,\theta,\|u_0\|_{\lambda_0,\infty}} |t-s|^{\theta p}.
\end{align}
Similarly, we can show that
\begin{align}\label{05152144.2}
E J_2^p\leq C_{\lambda_0,c_1,L_b,K_{\sigma},T,p,\|u_0\|_{\lambda_0,\infty}}|t-s|^{ p}.
\end{align}
Now we estimate $J_3$. By (ii)-(iv) of Lemma \ref{A.4}, we have
\begin{align}
J_3
\leq & \int_0^s\int_{\mathbb{R}} |p_{s-r}(x,z)-p_{s-r}(y,z)|\times \nonumber\\
& \left[ c_1 |u_n(r,z)|\log_+ |u_n(r,z)|+ L_b(|u_n(r,z)|+1) \right] \,\mathrm{d}r\mathrm{d}z \times e^{-\lambda |x|} \nonumber\\
\leq & \int_0^s\int_{\mathbb{R}} |p_{s-r}(x,z)-p_{s-r}(y,z)| \nonumber\\
& \times \bigg\{c_1 e^{\lambda_0 |z|e^{\beta r}}\lambda_0 |z|e^{\beta r}\times \sup_{z\in\mathbb{R}}\left(|u_n(r,z)|e^{-\lambda_0 |z|e^{\beta r}}\right) \nonumber\\
& + c_1 e^{\lambda_0 |z|e^{\beta r}}\times\sup_{z\in\mathbb{R}}\left[|u_n(r,z)|e^{-\lambda_0 |z|e^{\beta r}}\log_{+}\left(|u_n(r,z)|e^{-\lambda_0 |z|e^{\beta r}}\right)\right] \nonumber\\
& + L_b e^{\lambda_0 |z|e^{\beta r}}\times \sup_{z\in\mathbb{R}}\left(|u_n(r,z)|e^{-\lambda_0 |z|e^{\beta r}}\right) +L_b \bigg\} \,\mathrm{d}z\mathrm{d}r \times e^{-\lambda |x|} \nonumber\\
\leq & |x-y|\times\int_0^s \frac{1}{\sqrt{s-r}} \times \bigg\{\sup_{r\leq T, z\in\mathbb{R}}\left(|u_n(r,z)|e^{-\lambda_0 |z|e^{\beta r}}\right) \nonumber\\
& \times C_{\lambda_0, c_1, L_b,T} \Big[e^{\lambda_0 (|x|+|x-y|)e^{\beta r}}\lambda_0 (|x|+|x-y|)e^{\beta r} + e^{\lambda_0 (|x|+|x-y|)e^{\beta r}}\Big] \nonumber\\
& + \sup_{r\leq T,z\in\mathbb{R}}\left(|u_n(r,z)|e^{-\lambda_0 |z|e^{\beta r}}\right)\times \log_{+}\left[\sup_{r\leq T,z\in\mathbb{R}}\left(|u_n(r,z)|e^{-\lambda_0 |z|e^{\beta r}}\right)\right]\nonumber\\
& \times C_{\lambda_0,c_1,L_b,T}\times e^{\lambda_0(|x|+|x-y|)e^{\beta r}} + \sqrt{2}L_b \bigg\} \,\mathrm{d}r \times e^{-\lambda |x|} .
\end{align}
Due to the fact that $|x-y|\leq 1$ and the choice of $\lambda_0$, we have
\begin{align}\label{200405.20531}
& e^{\lambda_0 (|x|+|x-y|) e^{\beta r}}\lambda_0 (|x|+|x-y|) e^{\beta r}\times e^{-\lambda|x|} \nonumber\\
\leq & e^{\lambda_0 (|x|+1) e^{\beta r}}\lambda_0 (|x|+1) e^{\beta r}\times e^{-\lambda(|x|+1)} e^{\lambda} \nonumber\\
\leq & e^{\lambda}, \quad \forall\, r\in[0,T].
\end{align}
Hence in view of (\ref{210225.1854}), we see that
\begin{align}\label{05152144.3}
E J_3^p \leq C_{\lambda_0,c_1,L_b,K_{\sigma},T,p,\|u_0\|_{\lambda_0,\infty}} |x-y|^p .
\end{align}
Combining (\ref{05152143}), (\ref{05152144.1}), (\ref{05152144.2}) and (\ref{05152144.3}) together yields
\begin{align}
\mathbb{E} \left( |X_n(t,x)-X_n(s,y)|e^{-\lambda |x|} \right)^p \leq C_{\lambda_0,c_1,L_b,K_{\sigma},T,p,\theta,\|u_0\|_{\lambda_0,\infty}}\left(|t-s|^{\theta p}+|x-y|^p\right) ,
\end{align}
where the constant $\lambda_0$ is determined by $\lambda, T, c_1$ according to (\ref{210225.1936}). Thus,
(\ref{05152147}) is proved.
Now we prove (\ref{210125.2123}). Observe that
\begin{align}
& |V_n(t,x)-V_n(s,y)| \nonumber\\
\leq & \bigg|\int_0^s\int_{\mathbb{R}}[p_{t-r}(x,z)-p_{s-r}(x,z)]\sigma_n(u_n(r,z))\,W(\mathrm{d}r,\mathrm{d}z)\bigg| \nonumber\\
& + \bigg|\int_s^t\int_{\mathbb{R}}p_{t-r}(x,z)\sigma_n(u_n(r,z))\,W(\mathrm{d}r,\mathrm{d}z)\bigg| \nonumber\\
& + \bigg|\int_0^t\int_{\mathbb{R}}[p_{s-r}(x,z)-p_{s-r}(y,z)]\sigma_n(u_n(r,z))\,W(\mathrm{d}r,\mathrm{d}z)\bigg| \nonumber\\
=: & \, I_1 +I_2 + I_3 .
\end{align}
So
\begin{align}\label{210126.2129}
|V_n(t,x)-V_n(s,y)|^p \leq 3^{p-1} (I_1^p + I_2^p + I_3^p) .
\end{align}
By the BDG inequality, (\ref{210122.1855}), and (v) of Lemma \ref{A.4}, we get
\begin{align}\label{210126.2130}
\mathbb{E} I_1^p \leq & \mathbb{E}\bigg[\int_0^s\int_{\mathbb{R}}\big|p_{t-r}(x,z)-p_{s-r}(x,z)\big|^2\sigma_n(u_n(r,z))^2 \,\mathrm{d}z\mathrm{d}r\bigg]^{\frac{p}{2}} \nonumber\\
\leq & K_{\sigma}^p \times\bigg[\int_0^s\int_{\mathbb{R}}\big|p_{t-r}(x,z)-p_{s-r}(x,z)\big|^2 \,\mathrm{d}z\mathrm{d}r\bigg]^{\frac{p}{2}} \nonumber\\
\leq & C_{K_{\sigma},T,p} |t-s|^{\frac{p}{4}}.
\end{align}
Similarly, we have
\begin{align}\label{210126.2131}
\mathbb{E}I_3^p \leq C_{K_{\sigma},T,p} |x-y|^{\frac{p}{2}} .
\end{align}
For the term $I_2$, the uniform boundedness of $\sigma_n$ gives
\begin{align}\label{210126.2132}
\mathbb{E} I_2^p \leq & \mathbb{E}\bigg[\int_s^t\int_{\mathbb{R}} p_{t-r}(x,z)^2\sigma_n(u_n(r,z))^2 \,\mathrm{d}z\mathrm{d}r\bigg]^{\frac{p}{2}} \nonumber\\
\leq & K_{\sigma}^p \times\bigg[\int_s^t\int_{\mathbb{R}}\big|p_{t-r}(x,z)\big|^2 \,\mathrm{d}z\mathrm{d}r\bigg]^{\frac{p}{2}} \nonumber\\
\leq & K_{\sigma}^p \times\bigg[\int_s^t \frac{1}{2\sqrt{\pi(t-r)}} \,\mathrm{d}r\bigg]^{\frac{p}{2}} \nonumber\\
\leq & C_{K_{\sigma},T,p} |t-s|^{\frac{p}{4}} .
\end{align}
Combining (\ref{210126.2129}), (\ref{210126.2130}), (\ref{210126.2131}) and (\ref{210126.2132}) together, we obtain (\ref{210125.2123}) .
This completes the proof of Lemma \ref{05152151}.
$\blacksquare$
\vskip 0.6cm
\noindent {\bf Proof of Theorem \ref{thm1}}.
We have established in Lemma \ref{05152151} that $\{u_n\}$ is tight in $C(\mathbb{R}_+,C_{tem})$.
By Prokhorov's theorem and the modified version of Skorokhod's representation theorem whose proof can be found in Appendix C of [BHR], we may assume that $d(u_n, u)\rightarrow 0$ (not relabelled) a.s. in $C(\mathbb{R}_+,C_{tem})$ for some process $u$ on some probability space $(\widetilde{\Omega}, \widetilde{\mathcal{F}}, \widetilde{\mathbb{P}})$, in other words, for any $\lambda>0$, $T\geq 0$,
\begin{align}\label{05160930.2}
\sup_{t\leq T, x\in\mathbb{R}}\left(|u_n(t,x)-u(t,x)|e^{-\lambda|x|}\right)\rightarrow 0, \quad \widetilde{\mathbb{P}}-a.s..
\end{align}
By the dominated convergence theorem, and using (\ref{05160930.2}), (\ref{05160930.1}), (\ref{210126.2140}), (\ref{05160929.2}) and (\ref{210122.1855}), one can deduce that for any $(t,x)\in\mathbb{R_+}\times\mathbb{R}$,
\begin{align}
\int_0^t\int_{\mathbb{R}} p_{t-r}(x,z)b_n(u_n(r,z))\,\mathrm{d}r\mathrm{d}z\rightarrow \int_0^t\int_{\mathbb{R}} p_{t-r}(x,z)b(u(r,z))\,\mathrm{d}r\mathrm{d}z ,
\end{align}
$\widetilde{\mathbb{P}}$-a.s. as $n\rightarrow\infty$, and
\begin{align}
\int_0^t\int_\mathbb{R} p_{t-r}(x,z)\sigma_n(u_n(r,z))\,W(\mathrm{d}r,\mathrm{d}z)\rightarrow \int_0^t\int_{\mathbb{R}} p_{t-r}(x,z)\sigma(u(r,z))\,W(\mathrm{d}r,\mathrm{d}z),
\end{align}
in the sense of $L^p(\widetilde{\Omega})$ for any $p\geq 1$ as $n\rightarrow\infty$.
Therefore, we see that $u$ is a weak solution of (\ref{1.a}).
$\blacksquare$
\section{Pathwise uniqueness}
In this section, we prove the pathwise uniqueness of solutions to (\ref{1.a}) and hence obtain the strong solution.
\vskip 0.3cm
\noindent {\bf Proof of Theorem \ref{thm3}}. Since condition (H2) implies condition (H1), there exists a weak solution to (\ref{1.a}) according to Theorem \ref{thm1}. We only show the pathwise uniqueness for solutions of (\ref{1.a}). The existence of strong solutions then follows from the Yamada-Watanabe theorem.
Suppose that $u,v$ are two solutions of equation (\ref{1.a}) that belong to the space $C(\mathbb{R}_{+}, C_{tem})$. We are going to show that $u=v$.
Fix $T>0$, and take $\lambda>0$ sufficiently small so that $T\leq T^*(\lambda, c_4)$. In this section, we write $\beta$ for $\beta(\lambda, c_4)$ for simplicity.
Let $M>0 $ and $0<\delta\leq e^{-1}$.
Define stopping times
\begin{align*}
\tau_M := & \inf\left\{t>0: \sup_{x\in\mathbb{R}}\left(|u(t,x)|e^{-\lambda|x|e^{\beta t}}\right)\geq M\right\} \\
&\wedge \inf\left\{t>0: \sup_{x\in\mathbb{R}}\left(|v(t,x)|e^{-\lambda|x|e^{\beta t}}\right)\geq M\right\} , \\
\tau^{\delta}:=& \inf\left\{t>0: \sup_{x\in\mathbb{R}}\left(|u(t,x)-v(t,x)|e^{-\lambda|x|e^{\beta t}}\right)\geq \delta\right\} , \\
\tau_M^{\delta}:=& \tau_M\wedge\tau^{\delta}\wedge T ,
\end{align*}
with the convention that $\inf \emptyset = +\infty$.
Define also
\begin{align}
Z(r):= \mathbb{E}\sup_{t\leq r\wedge\tau_M^{\delta}, x\in\mathbb{R}}\left(|u(t,x)-v(t,x)|e^{-\lambda |x|e^{\beta t}}\right) .
\end{align}
\noindent Obviously
\begin{align}\label{210224.1634}
& Z(r) \nonumber\\
\leq & \,\mathbb{E}\sup_{t\leq r\wedge\tau_M^{\delta}, x\in\mathbb{R}} \left\{ \int_0^t\int_{\mathbb{R}} p_{t-s}(x,y) |b(u(s,y))-b(v(s,y)) | \,\mathrm{d}s\mathrm{d}y \times e^{-\lambda |x|e^{\beta t}} \right\}\nonumber\\
& + \mathbb{E}\sup_{t\leq r\wedge\tau_M^{\delta}, x\in\mathbb{R}} \left\{ \left| \int_0^t\int_{\mathbb{R}} p_{t-s}(x,y) [\sigma(u(s,y))-\sigma(v(s,y))] \,W(\mathrm{d}s,\mathrm{d}y) \right| e^{-\lambda |x|e^{\beta t}} \right\} \nonumber\\
=: & \, I + J .
\end{align}
Now we estimate the term $I, J$ separately. By condition (H2), we have
{\allowdisplaybreaks\begin{align}\label{210224.1635}
I \leq & \,\mathbb{E}\sup_{t\leq r\wedge\tau_M^{\delta}, x\in\mathbb{R}}\bigg\{\int_0^t \int_{\mathbb{R}} p_{t-s}(x,y) c_3|u(s,y)-v(s,y)| \nonumber\\
& ~~~~~~~\times \log_{+}\frac{1}{|u(s,y)-v(s,y)|} \,\mathrm{d}s\mathrm{d}y \times e^{-\lambda |x| e^{\beta t}} \bigg\} \nonumber\\
& + \mathbb{E}\sup_{t\leq r\wedge\tau_M^{\delta}, x\in\mathbb{R}}\bigg\{\int_0^t \int_{\mathbb{R}} p_{t-s}(x,y) c_4 \log_{+}\big(|u(s,y)|\vee |v(s,y)|\big) \nonumber\\
& ~~~~~~~ \times |u(s,y)-v(s,y)| \,\mathrm{d}s\mathrm{d}y \times e^{-\lambda |x| e^{\beta t}} \bigg\} \nonumber\\
& + \mathbb{E}\sup_{t\leq r\wedge\tau_M^{\delta}, x\in\mathbb{R}}\bigg\{\int_0^t \int_{\mathbb{R}} p_{t-s}(x,y) c_5 |u(s,y)-v(s,y)| \,\mathrm{d}s\mathrm{d}y \times e^{-\lambda |x| e^{\beta t}} \bigg\} \nonumber\\
=: & \, I_1 + I_2 + I_3 .
\end{align}}
\noindent First, we estimate the term $I_1$. By the fact that the function $x\mapsto x\log\frac{1}{x}$ is increasing and concave on $(0,e^{-1})$, (\ref{05132049.1}) and (\ref{05132049.2}), we get
{\allowdisplaybreaks\begin{align}
I_1 \leq & c_3 \,\mathbb{E}\sup_{t\leq r\wedge\tau_M^{\delta}, x\in\mathbb{R}}\bigg\{\int_0^t \int_{\mathbb{R}} p_{t-s}(x,y) e^{\lambda |y| e^{\beta s}} |u(s,y)-v(s,y)| e^{-\lambda |y| e^{\beta s}} \nonumber\\
& ~~~~~~~\times \log_{+}\frac{1}{|u(s,y)-v(s,y)| e^{-\lambda |y| e^{\beta s}}} \,\mathrm{d}y\mathrm{d}s \times e^{-\lambda |x| e^{\beta t}} \bigg\} \nonumber\\
\leq & c_3 \,\mathbb{E}\sup_{t\leq r\wedge\tau_M^{\delta}, x\in\mathbb{R}}\bigg\{\int_0^t \sup_{y\in\mathbb{R}} \bigg[|u(s,y)-v(s,y)| e^{-\lambda |y| e^{\beta s}} \nonumber\\
& ~~~~~~~ \times \log_{+}\frac{1}{|u(s,y)-v(s,y)| e^{-\lambda |y| e^{\beta s}}}\bigg] \nonumber\\
& ~~~~~~~ \times \int_{\mathbb{R}} p_{t-s}(x,y) e^{\lambda |y| e^{\beta s}} \,\mathrm{d}y\mathrm{d}s \times e^{-\lambda |x| e^{\beta t}} \bigg\} \nonumber\\
\leq & 2c_3 e^{\frac{\lambda^2}{4\beta} e^{2\beta r-1}} \int_0^t \mathbb{E} \bigg\{ \sup_{\rho\leq s\wedge\tau_M^{\delta}, y\in\mathbb{R}} \Big( |u(\rho,y)-v(\rho,y)| e^{-\lambda |y| e^{\beta \rho}} \Big) \nonumber\\
& \times \log_{+} \frac{1}{\sup_{\rho\leq s\wedge\tau_M^{\delta}, y\in\mathbb{R}}\big(|u(\rho,y)-v(\rho,y)| e^{-\lambda |y| e^{\beta \rho}}\big)} \bigg\} \,\mathrm{d}s \nonumber\\
\leq & 2c_3 e^{\frac{\lambda^2}{4\beta} e^{2\beta r-1}} \int_0^r Z(s)\log_{+}\frac{1}{Z(s)} \,\mathrm{d}s,
\end{align}}
\noindent where (\ref{05132049.1}) was used.
Note that
\begin{align*}
\log_{+}(ab) \leq \log_{+}a + \log_{+} b .
\end{align*}
By the definition of $\tau_M^{\delta}$, we have
{\allowdisplaybreaks\begin{align}
I_2
\leq & c_4 \,\mathbb{E}\sup_{t\leq r\wedge\tau_M^{\delta}, x\in\mathbb{R}}\bigg\{\int_0^t \int_{\mathbb{R}} p_{t-s}(x,y) \bigg[ \log_{+}\Big(e^{\lambda|y| e^{\beta s} }\Big) \nonumber\\
& + \log_{+}\Big(\big(|u(s,y)|e^{-\lambda|y| e^{\beta s}}\big)\vee \big(|v(s,y)| e^{-\lambda|y| e^{\beta s}}\big) \Big) \bigg]\nonumber\\
& \times \Big(|u(s,y)-v(s,y)|e^{-\lambda|y| e^{\beta s}}\Big) e^{\lambda|y| e^{\beta s}} \,\mathrm{d}y\mathrm{d}s \times e^{-\lambda |x| e^{\beta t}} \bigg\} \nonumber\\
\leq & c_4 \,\mathbb{E}\sup_{t\leq r\wedge\tau_M^{\delta}, x\in\mathbb{R}}\bigg\{\int_0^t \sup_{y\in\mathbb{R}} \Big(|u(s,y)-v(s,y)|e^{-\lambda|y| e^{\beta s}}\Big)\nonumber\\
& \times \int_{\mathbb{R}} p_{t-s}(x,y) e^{\lambda|y| e^{\beta s}} \lambda|y| e^{\beta s} \,\mathrm{d}y\mathrm{d}s \times e^{-\lambda |x| e^{\beta t}} \bigg\} \nonumber\\
& + c_4 \log_{+}(M) \mathbb{E}\sup_{t\leq r\wedge\tau_M^{\delta}, x\in\mathbb{R}}\bigg\{\int_0^t \sup_{y\in\mathbb{R}} \Big(|u(s,y)-v(s,y)|e^{-\lambda|y| e^{\beta s}}\Big)\nonumber\\
& \times \int_{\mathbb{R}} p_{t-s}(x,y) e^{\lambda|y| e^{\beta s}} \,\mathrm{d}y\mathrm{d}s \times e^{-\lambda |x| e^{\beta t}} \bigg\} \nonumber\\
\leq & \frac{1}{2} Z(r) + c_4 C_{\lambda, \beta, M, r} \int_0^r Z(s)\,\mathrm{d}s ,
\end{align}}
\noindent where the last inequality holds for the same reason as the derivation of (\ref{05132036})-(\ref{05132040.2}) with constant $c_1$ replaced by constant $c_4$.
Similarly,
\begin{align}
I_3 \leq 2 c_5 e^{\frac{\lambda^2}{4\beta} e^{2\beta r-1}} \int_0^r Z(s) \,\mathrm{d}s .
\end{align}
\noindent For the term $J$, we use the estimate established in Proposition \ref{estimates 003} to obtain
\begin{align*}
J \leq & \epsilon \,\mathbb{E} \sup_{s\leq r\wedge\tau_M^{\delta}, y\in\mathbb{R}}\left(|\sigma(u(s,y))-\sigma(v(s,y))| e^{-\lambda|y|e^{\beta s}}\right) \nonumber\\
& + C_{\epsilon,\lambda, \beta, r} \,\mathbb{E} \int_0^{r\wedge\tau_{M}^{\delta}}\int_{\mathbb{R}}|\sigma(u(s,y))-\sigma(v(s,y))| e^{-\lambda |y|e^{\beta s}} \,\mathrm{d}y\mathrm{d}s ,
\end{align*}
where the constant $C_{\epsilon,\lambda, \beta, r}$ is the constant $C_{\epsilon,p,h(T),T}$ appeared in (\ref{101.2}) by taking $p=1$, $T=r$ and $h(T)=\lambda e^{\beta r}$.
Since $\sigma$ is bounded and Lipschitz, there exists two nonnegative constants $K_{\sigma}$ and $L_{\sigma}$ such that
\begin{align*}
|\sigma(x)| \leq & K_{\sigma}, \quad \forall\, x\in\mathbb{R}, \\
|\sigma(x)-\sigma(y)| \leq & L_{\sigma}|x-y|, \quad \forall\, x,y\in\mathbb{R}.
\end{align*}
Hence for any $0< \theta <1$, we have
\begin{align}\label{210217.2114}
J \leq & \epsilon L_{\sigma} Z(r) + C_{\epsilon,\lambda, \beta, r}\,\mathbb{E} \int_0^{r\wedge\tau_{M}^{\delta}} \sup_{y\in\mathbb{R}} \bigg\{ \left(|\sigma(u(s,y))-\sigma(v(s,y))| e^{-\lambda |y|e^{\beta s}} \right)^{\theta} \nonumber\\
& \times \int_{\mathbb{R}} |\sigma(u(s,y))-\sigma(v(s,y))|^{1-\theta} e^{-(1-\theta)\lambda |y|e^{\beta s}} \,\mathrm{d}y \bigg\} \,\mathrm{d}s \nonumber\\
\leq & \epsilon L_{\sigma} Z(r) + \frac{(2K_{\sigma})^{1-\theta} L_{\sigma}^{\theta} C_{\epsilon,\lambda, \beta, r}}{(1-\theta)\lambda} \mathbb{E} \int_0^{r\wedge\tau_{M}^{\delta}} \sup_{y\in\mathbb{R}} \left(| u(s,y)- v(s,y)| e^{-\lambda |y|e^{\beta s}} \right)^{\theta} \mathrm{d}s \nonumber\\
\leq & \epsilon L_{\sigma} Z(r) + \frac{(2K_{\sigma})^{1-\theta} L_{\sigma}^{\theta} C_{\epsilon,\lambda, \beta, r}}{(1-\theta)\lambda} \int_0^r Z(s)^{\theta} \,\mathrm{d}s .
\end{align}
Combining (\ref{210224.1634})-(\ref{210217.2114}) together, we obtain that
\begin{align}
Z(r)\leq & \left(\frac{1}{2} + \epsilon L_{\sigma} \right) Z(r) + C_{\lambda,M, c_4, c_5, r} \int_0^r Z(s)\,\mathrm{d}s \nonumber\\
& + 2c_3 e^{\frac{\lambda^2}{4\beta}e^{2\beta r-1}} \int_0^r Z(s)\log_{+}\frac{1}{Z(s)}\,\mathrm{d}s +\frac{(2K_{\sigma})^{1-\theta} L_{\sigma}^{\theta} C_{\epsilon,\lambda, \beta, r}}{(1-\theta)\lambda} \int_0^r Z(s)^{\theta} \,\mathrm{d}s .
\end{align}
Taking for example $\epsilon= \frac{1}{4L_{\sigma}}$, subtracting $\left(\frac{1}{2} + \epsilon L_{\sigma} \right) Z(r)$ from both sides of the above inequality, and then applying the special Gronwall-type inequality established in Lemma \ref{A.3}, we obtain
\begin{align}
Z(r)\equiv 0, \quad \forall\, r\geq 0 .
\end{align}
Since the solutions of (\ref{1.a}) don't blowup, let $M\rightarrow\infty$ to obtain $\mathbb{P}$-a.s.,
\begin{align}
u(t,x)=v(t,x), \quad \forall\, x\in \mathbb{R}, \ \forall\, t\in [0, \tau^{\delta}\wedge T].
\end{align}
This implies that $\tau^{\delta} \geq T $, $\mathbb{P}$-a.s., otherwise it contradicts the definition of $\tau^{\delta}$. By the arbitrariness of $T$, we obtain that for $\mathbb{P}$-a.s.,
\begin{align}
u(t,x)=v(t,x), \quad \forall\, (t,x)\in \mathbb{R}_{+}\times\mathbb{R}.
\end{align}
This completes the proof the pathwise uniqueness.
$\blacksquare$
\vskip 0.6cm
\noindent{\bf Acknowledgement}. We are grateful to Robert Dalang and Davar Khoshnevisan for their useful suggestions and comments. This work is partially supported by by NSFC (No. 11971456, 11721101, No. 12001516).
|
\section{\textcolor{PREPRINTCOL}{#1}}}
\def\ssec#1{\subsection{\textcolor{PREPRINTCOL}{#1}}}
\def\sssec#1{\subsubsection{\textcolor{PREPRINTCOL}{#1}}}
\def\usec#1{\section*{\textcolor{PREPRINTCOL}{#1}}}
\def\shd#1{{\bf \color{PREPRINTCOL}#1}}
\begin{document}
\preprint{Redd and Cancio, preprint 2020}
\title{ Asymptotic Analysis of the Pauli Potential for Atoms }
\author{Jeremy J. Redd}
\email{<EMAIL>}
\affiliation{Department of Physics, Utah Valley University, Orem, UT 84058}
\author{Antonio C. Cancio}
\affiliation{Department of Physics and Astronomy, Ball State University, Muncie, Indiana 47306}
\date{\today}
\begin{abstract}
ABSTRACT:
Modeling the Pauli energy, the contribution to the kinetic energy caused by Pauli statistics, without using orbitals is the open problem of orbital-free density functional theory.
An important aspect of this problem is correctly reproducing the Pauli
potential, the response of the Pauli kinetic energy to a change in density.
We analyze the behavior of the Pauli potential of non-relativistic neutral atoms under
Lieb-Simon scaling -- the process of taking nuclear charge and particle number
to infinity, in which the kinetic energy tends to the Thomas-Fermi limit.
We do this by mathematical analysis of the near-nuclear region and by calculating the exact
orbital-dependent Pauli potential using the approach of Ouyang and Levy for closed-shell atoms
out to element Z=976.
In rough analogy to Lieb and Simon's own findings for the charge density, we find that the
potential does not converge smoothly to the Thomas-Fermi limit on a point-by-point basis
but separates into several distinct regions of behavior.
Near the nucleus, the potential approaches a constant given by the difference
in energy between the lowest and highest occupied eigenvalues.
We discover a transition region in the
outer core where the potential deviates unexpectedly and predictably from both
the Thomas-Fermi potential and the gradient expansion correction to it.
These results may provide insight into semi-classical description of Pauli statistics, and new
constraints to aid the improvement of orbital-free DFT functionals.
\end{abstract}
\keywords{Density functional theory, orbital-free Density functional theory, Pauli potential,
Electronic structure}
\maketitle
\section{Introduction}\label{Introduction}
The most generally accurate and widely used method for
predicting electronic structure is the Kohn-Sham (KS) approach to
density functional theory (DFT).~\cite{KS}
By introducing auxiliary orbitals into the definition of particle density,
the KS functional allows for an accurate representation of the
energy of the exact many-body Hamiltonian by the energy of a simpler
noninteracting system.~\cite{Martin, IRG}
This greatly simpliifies the mathematics and speeds up computations
as compared to many-body or Hartee-Fock calculations.~\cite{karasiev}
However, the use of orbitals still comes with increasing computational cost as
the number of particles in the system is up-scaled.
This means that for systems that require the calculation of many orbitals
such as mesoscale systems where quantum properties may be
important\cite{akimov2015large}
and warm dense matter,~\cite{graziani2014frontiers, rosnerbasic, karasievWDM}
in which many states become thermally activated,
the computational cost of the KS method becomes prohibitive.
The Hohenburg-Kohn theorem, however, states that the ground state of
any Hamiltonian system can be uniquely characterized by the particle density
alone.~\cite{HK} This means that exact Hamiltonian solutions can be expressed as functionals of exclusively the density, eliminating the need
for orbitals.\cite{karasiev,TranWes}
Crucially, this theorem applies to any piece of the
energy, so not only is the true interacting Kinetic energy a
functional of the density but also the KS kinetic energy:
as conventionally defined, this is a functional of the KS orbitals,
nevertheless a more general orbital-free expression should exist.
Recent years have seen a growing number of approaches to constructing orbital-free DFT (OFDFT)
approximations to the KS method that allow for improved computational scaling for systems with
large numbers of orbitals.\cite{TranWes, witt2018orbital} The orbital-free philosophy has also
been introduced successfully into
conventional KS DFT, in the form of ``de-orbitalizing'' exchange-correlation
functionals that depend explicitly upon the KS kinetic energy density (KED),
replacing it with an equivalent expression in terms of the density and
its derivatives.~\cite{trickey_deorbitalization}
However the challenge of developing a robust OFDFT model with reasonable
predictive accuracy for a variety of systems is severe.
In conventional Kohn-Sham DFT one needs to approximate
the exchange-correlation energy describing the difference
in energy between interacting and noninteracting systems for the
same external potential, normally a small correction.
OFDFT must approximate the kinetic energy, which is of the order
of the energy itself
and must therefore be modeled to high accuracy.
The most basic OF theory, Thomas-Fermi (TF)
theory,~\cite{thomas1927calculation, F27}
uses the KE of the homogeneous electron gas applied to the local density,
in analogy to the LDA of the KS method.
But unlike the LDA, which produces at least qualitatively good
structural predictions, TF theory does not permit chemical
binding at all.~\cite{teller,cartermol,finzel2018chemical}
The simplest functional beyond Thomas-Fermi,
the gradient expansion (GE), does very well for atoms, but still not
so well for molecular binding.
Many attempts have been made to build on this foundation to develop
semilocal or ``single-point'' functionals using the local density
and its gradient as ingredients,~\cite{TranWesolowski, LacksGordon94,
Thakkar92, APBE, VT84F, BorgooMol, LKT}
sometimes adding the Laplacian of the
density,~\cite{PC,LCPB,cancioredd,CFS2018,CFS20194thorder} and
the electronic Hartree potential.~\cite{constantin2017modified}
These more complex models generally share the problems of their predecessors,
but can be competitive~\cite{CFS2018} with more expensive empirical
nonlocal functionals for some solids.
Two-point nonlocal functionals have been somewhat more
successful.~\cite{WangCarterNew,huangcarter,pavanello}
These are based on the
Lindhard formula for linear response of the homogeneous
electron gas.
However, the Lindhard function is not an appropriate reference
point for finite systems
and systems with surfaces.~\cite{ConstLindhard}
At least to date,
such functionals require system-dependent empirical parameters to succeed.
The challenge of OFDFT is modeling the kinetic energy due to the Pauli
exclusion principle -- the orbital dependence in the KS functional is a
consequence of Pauli statistics.
One considers the total kinetic energy as a sum of this Pauli KE contribution
and the von Weizs\"{a}cker KE -- the KE
of a fictitious Bose system with the same density as the real system.
Minimizing the OFDFT energy then generates
an Euler equation for the density for this fictitious Bose system
where the contribution of Pauli statistics appears as an effective
Pauli potential, that forces this fictitious system to have the
same density as the true fermionic one.
This Pauli potential thus is analogous
to the Kohn-Sham potential for the conventional Kohn-Sham method.
The Pauli potential thus plays an important role
in guaranteeing the stability and accuracy of structural calculations;
nonetheless, like the Kohn-Sham potential, it gets much attention in
developing functionals than the Pauli energy.
A notable exception to this tendency is the use of the non-negativity of the potential
as a constraint -- a significant feature
of at least one family of functionals.~\cite{VT84F, LKT}
Nonetheless, a quite pleasing property of the
exact Pauli potential is that it can be easily constructed in terms of KS orbitals in a simple fashion.\cite{levy1988exact} Essentially one can use the
orbital definition of density to solve the KS problem and equivalent Euler problem simultaneously. This allows one to compare the results
of model Pauli potentials to the exact potential for any system of interest.
Exact Pauli potentials have been constructed in this way for example
atoms,~\cite{levy1988exact, gritsenko1994, van1995step, baerends1997quantum, KraislerAxel20,
FinzelMolecules20}
Approximate Pauli potentials play a key role in a recently developed OF
method~\cite{finzel2018chemical, Finzel19, FinzelMolecules20}.
These rely on the orbitals of isolated atoms and an
orbital-free description of the bond, thus constituting a hybrid approach to deorbitalizing the
KS problem.
One way to generate useful constraints for functional development -- whether
on the total energy or the potential -- is to consider the behavior of the
functional under scaling of the system.
A particularly fruitful example is
Lieb-Simon scaling, the best
known example of which is the scaling of the KE of neutral atoms as nuclear charge
tends to infinity.~\cite{liebsimon, BCGP16,LCPB}
This should not be confused with Levy-Perdew scaling which more closely
resembles the scaling of nuclear charge to infinity with constant
particle number.\cite{levyperdew}
The lower bound of this scaling behavior is the von Weizs\"acker (VW) solution for
hydrogen and helium, trivially convertible to orbital-free form because
it involves only one occupied orbital.
The upper bound is less simple but more powerful. The leading order in $Z$ of the
total and kinetic energy as $Z\to\infty$ is given by TF theory~\cite{liebsimon}.
The gradient expansion approximaton (GEA) contributes corrections of smaller
order in $Z$ to the large $Z$ limit of the energy.~\cite{BCGP16,LCPB}
The limiting behavior of total energies is reflected in the kinetic energy
density, which is locally approximated by a variant of the gradient
expansion in the core region of the atom.~\cite{cancioredd}
The physical property that has not been explored carefully in the Lieb-Simon limit
is the Pauli potential.
Even though the total Pauli kinetic energy should be well described by
TF theory in this limit, the same does not
necessarily hold point for point for the potential. And it is unknown to what
extent functionals that are successful in describing total energies
work for the potential.
In this paper, we analyze the behavior of the exact KS Pauli potential
for nonrelativistic neutral atoms as a function of $Z$ up to $Z=976$, large
enough to extract limiting behavior
and exact constraints that may be of aid to the development of KE functionals.
We find an exact constraint in the near-nucleus limit
and an unexpected deviation from the Thomas Fermi limit for a
the outer shells of the large-$Z$ atom.
The paper is organized as follows:
Section~\ref{Theory} discusses the theoretical background of Pauli potentials, both in the context of KS and of OFDFT approximations.
Section~\ref{Methods} describes the methods and algorithms used for calculations and validation of the results.
Section~\ref{Results} details the visual results of extending the exact Pauli response functionals and Pauli potentials to large-$Z$, as compared
to approximations.
Section~\ref{Conclusions} discusses the
ramifications of our findings and possible future work.
\section{Theory} \label{Theory}
In Kohn-Sham theory, the total energy of an electronic system as
a functional of the density $\rho$ is given by
\begin{equation}
\label{eq:Eks}
E[\rho]=T_{KS}[\rho]+U[\rho]+\int v_{ext}(\boldsymbol{r})\rho(\boldsymbol{r}) d^3r+E_{xc}[\rho],
\end{equation}
where $T_{KS}[\rho]$ is the noninteracting contribution to the KE, $U[\rho]$ is the static electron-electron interaction, $v_{ext}$ is an external potential, and $E_{xc}$ is the energy of exchange and correlation effects. The last term contains the difference in energy
between the true interacting system and the fictitious noninteracting one.
The KS density is given by
\begin{equation}
\label{eq:rhoks}
\rho(\boldsymbol{r})=\sum_{i}f_{i}|\phi_{i}(\boldsymbol{r})|^2,
\end{equation}
where $\phi_i$ are the auxiliary single-particle orbital that describe the noninteracting system and
$f_i$ is the occupation number.
The KS kinetic energy is then given by
\begin{equation}
\label{eq:tauks}
T_{KS}=\int\tau_{KS}d^3r
=\int\frac{1}{2}\sum_{i}f_{i}
|\nabla \phi_{i}(\boldsymbol{r})|^2 d^3r,
\end{equation}
where $\tau_{KS}$ is a positive-definite kinetic energy density.
The density is determined
by the functional minimization of the energy with respect
to each orbital, with the constraint
of preserving orbital normalization. This generates the effective Kohn-Sham equation for each orbital:
\begin{equation}
\label{eq:eks}
\left[\frac{1}{2}\nabla^2+v_{KS}(\boldsymbol{r})\right]\phi_{i}(\boldsymbol{r}) = \epsilon_{i}\phi_{i}(\boldsymbol{r})
\end{equation}
where $\epsilon_i$ is an auxiliary eigen value.
The KS potential is determined from the functional derivative of the energy:
\begin{equation}
\label{eq:vsham}
v_{KS}=\frac{\delta}{\delta\rho}\left(U[\rho]+E_{xc}[\rho]\right)+v_{ext}.
\end{equation}
In order to generate an orbital-free version of the Kohn-Sham functional,
we define the Pauli KE
as the difference between KS and vW kinetic energies.
\begin{equation}
\label{eq:Tp}
T_{p}=T_{KS}-T_{vW},
\end{equation}
and define a Pauli KE density, the integral over which yields the Pauli KE, similarly:
\begin{equation}
\label{eq:taup}
\tau_{p}=\tau_{KS}-\tau_{vW}.
\end{equation}
As discussed in the introduction, the vW kinetic energy, in the spirit of
the KS idea, is the kinetic energy of a fictitious
Bose system that has the same energy and density as the true, fermionic system.
In this case, all particles occupy the ground state, $\rho = N|\psi_0|^2$,
so that the associated KE density is
\begin{equation}
\label{eq:tauvw}
\tau_{vW}=-\frac{1}{2}\left| \nabla\sqrt{\rho(r)}\right|^2=
\frac{1}{8}\frac{|\nabla \rho(\boldsymbol{r})|^2}{\rho(\boldsymbol{r})}.
\end{equation}
This is strictly correct for the true system only if $N\leq 2$.
The Pauli KE then measures the additional kinetic
energy due to Fermi statistics. This has to be approximated somehow by a functional
of the density, in a way similar to how the XC energy incorporating
electron interactions is approximated in KS theory.
One can now generate an orbital-free Euler expression of the KS problem.
It calculates the non-interacting Bose KE explicitly and considers effects of the Pauli contribution to the KE
to come from a positive definite Pauli potential $v_p$.
Minimizing $E - \mu \int \rho(\mathbf{r}) d^3r$, one finds
\begin{equation}
\label{eq:tsqrtrho}
\left[-\frac{1}{2}\nabla^2+v_\mathrm{eff}(\mathbf{r})\right]\sqrt{\rho(\mathbf{r})}=\mu\sqrt{\rho(\mathbf{r})}.
\end{equation}
Like the KS procedure, this generates an effective potential $v_\mathrm{eff}$, and solves for the
density and a single eigenvalue $\mu$, the chemical potential.
The effective potential is given by $v_\mathrm{eff}=v_{KS}+v_p$, with
the addition to the Kohn-Sham potential -- the Pauli potential $v_p$ --
given by\cite{levy1988exact}
\begin{equation}
\label{eq:vp}
v_{p}(\mathbf{r})=\frac{\delta T_{p}[\rho]}{\delta\rho(\boldsymbol{r})}.
\end{equation}
It can be interpreted as the potential needed to make the density of the
fictitious Bose system calculated with Eq.~[\ref{eq:tsqrtrho}]
equal the density of the fermionic Kohn-Sham system.
The Pauli potential may be determined exactly in terms of the KS orbitals as~\cite{levy1988exact}
\begin{equation}
\label{eq:vpauli}
v_p(\mathbf{r})=\frac{\tau_p(\mathbf{r})}{\rho(\boldsymbol{r})}+v_{r}(\mathbf{r})
\end{equation}
where $v_{r}$ is the response of the effective potential, and consequently the KE, to an arbitrary change in density. The exact response potential is given by \cite{levy1988exact}
\begin{equation}
\label{eq:vrespexact}
v_r(\mathbf{r})=\frac{2}{\rho(\boldsymbol{r})} \sum_{j=1}^{M}(\epsilon_{M}-\epsilon_{j})\phi_{j}^{*}(\boldsymbol{r})\phi_{j}(\boldsymbol{r}).
\end{equation}
Eq.~\ref{eq:vpauli} and~\ref{eq:vrespexact} can be derived by simultaneously
solving Eq.~\ref{eq:eks} and Eq.~\ref{eq:tsqrtrho} using the same density Eq.~\ref{eq:rhoks}
where the occupation $f_i =2$.
The primary tool for our study of atomic Pauli potentials is Lieb and Simon's
$\zeta$ scaling of the kinetic energy of
neutral atoms~\cite{liebsimon, lieb1977thomas, BCGP16, LCPB, cancioredd}
~The Lieb-Simon theorem scales the potential and particle number of a system simultaneously:
\begin{eqnarray}
\label{eq:densscale}
N_\zeta&=&\zeta N_1, \\
\label{eq:potscale}
v_\zeta(r)&=&-\frac{\zeta}{r}.
\end{eqnarray}
This yields the neutral atoms for integer values $Z$ of the continuous variable
$\zeta$ with the choice $N_1=1$.
Particle distance is then scaled in units of the Thomas-Fermi atomic radius $\sim Z^{-1/3}a_0$
so that formally, the potential scales as
\begin{equation}
\label{eq:volscale}
v_\zeta=\zeta^{4/3} v_1(\zeta^{1/3}r).
\end{equation}
The key result for this paper is that in the limit $\zeta \to \infty$,
(for atoms, $Z\to \infty$) the total energy and
thus also kinetic energy in the Thomas-Fermi approximation becomes
relatively exact:
\begin{equation}
\lim_{\zeta\to\infty} \frac{T_{KS} - T_{TF}}{T_{KS}} \to 0.
\label{eq:LiebSimonlimit}
\end{equation}
Secondly, in the case of atoms,
the TF energy and the leading corrections in the $Z\to \infty$ limit
are exactly known and form an expansion in powers of $Z^{1/3}$:
\begin{equation}
\label{eq:TofZexpansion}
T_{KS}=c_0 Z^{7/3}+c_1 Z^2+c_2 Z^{5/3}+...
\end{equation}
Here the leading order $c_0=0.768745$ is predicted by
TF theory,~\cite{thomas1927calculation}
$c_1=-1/2$,~\cite{scott1952, schwinger1980thomas} and $c_2=0.269900$.~\cite{schwinger1981thomas}
Any candidate for an orbital free KE functional ought to satisfy this
scaling behavior,
but this not a trivial task.~\cite{LCPB} The second correction $c_2$
is generated by the standard gradient expansion. However,
the Scott correction, which scales as $Z^2$, is a larger effect and although it
may be modeled with a gradient expansion, it explicitly deals with the
KE near the Coulomb singularity, where the GE is
not legitimate. Not surprisingly, very few GGA's or metaGGA's get this
limit correctly.~\cite{PC,LCPB}
Given the importance of the gradient expansion model for the large $Z$
expansion, we will compare our results to functionals of this form;
keeping in mind its limitations, we explore a number of variations on the
theme.
The leading order term of the large-$Z$ expansion is, as per Eq.~(\ref{eq:LiebSimonlimit}), given by the
the Thomas-Fermi KED~\cite{spruch1991pedagogic} -- the KED in
the limit of a homogeneous electron gas, applied to the local density $\rho(\boldsymbol{r})$:
\begin{equation}
\label{eq:tautf}
\tau_{TF}(\boldsymbol{r})=\frac{3}{10}k_F(\boldsymbol{r})^2\rho(\boldsymbol{r})=\frac{3}{10}(3\pi^2)^{2/3}\rho(\boldsymbol{r})^{5/3}.
\end{equation}
(It should be noted that this and subsequent model equations are defined
for the Kohn-Sham and not the Pauli KED).
The subsequent orders depend on the gradient expansion of the kinetic energy for
the slowly varying electron gas.
The gradient expansion may be formally derived as an expansion in orders
of $\hbar$, good for large values of the local fermi energy -- in effect,
large numbers of occupied states.
To second order it is given by~\cite{kirzhnitsND}
\begin{equation}
\label{eq:taugea2}
\tau_{GEA}=\left[1+\frac{5}{27}p+\frac{20}{9}q\right]\tau_{TF},
\end{equation}
where
\begin{equation}
\label{eq:p}
p=\frac{|\nabla \rho|^2}{4{k_F}^2{\rho}^2},
\end{equation}
and
\begin{equation}
\label{eq:q}
q=\frac{\nabla^2 \rho}{4{k_F}^2\rho}.
\end{equation}
The fourth order~\cite{hodges73} correction improves on this
model for atoms~\cite{LCPB,JandG} but will not be considered in this paper.
Although this ``canonical" GEA yields a reasonable description of the
large $Z$ expansion, it is not perfect.
It neither appears to be the best candidate for describing the total
KE of atoms~\cite{LCPB} nor the
local KED.~\cite{lindmaa14,cancioredd}
In fact, a modification of the GEA ($\mathrm{Loc}$) can be
determined by fitting the local KED of the core shells of
large-$Z$ atoms~\cite{cancioredd}, which yields
\begin{equation}
\label{eq:gealoc}
\tau_{Loc}=\left(1-0.275p+2.895q\right)\tau_{TF}.
\end{equation}
Although this does not yield good total energies, it is of obvious interest to model
the \textit{potential}
which is also a local quantity and is related to the KED by Eq.~\ref{eq:vpauli}.
It is noteworthy that the gradient term of the $\mathrm{Loc}$ GE has
a sign opposite to that of the canonical GE, and thus a net correction
to the TF energy which is negative, rather than positive.
Ref.~\onlinecite{cancioredd} also introduces a
$Z$-dependent near-nuclear correction to this
that does yield good total energies, labelled $\mathrm{NNloc}$ in the results.
A similar local model is that of Lindmaa, Armiento and Mattsson,~\cite{lindmaa14} given
by:
\begin{equation}
\label{eq:tauairy}
\tau_{Airy}=\left[1-\frac{5}{27}p+\frac{30}{9}q\right]\tau_{TF},
\end{equation}
This gradient expansion is derived from the analysis of
the ``edge electron gas" or Airy gas~\cite{KohnMattsson} that is
constructed by taking a linear potential with hard wall boundary, in the
limit that the hard wall is moved to infinity. Thus it is meant to be valid
for surfaces, and not necessarily as global functional.
It has been shown to be a good approximation for the KED of model systems
including jellium droplets and the Bohr atom.
As an atom
is necessarily a system with a surface region, it is of interest to see
how it fares here.
A final variant of the GE is introduced by Tsirelson et al. (Ref.~\onlinecite{Ts}) which uses an estimate of the chemical
potential to modify the large-$r$ limit. The most relevant portion of their model
is the response function which is fit to the following form:
\begin{equation}
\label{eq:TSprefit}
v_{r,Ts} = \frac{(3\pi^2)^{2/3}}{5}\rho^{2/3}+a\frac{|\nabla\rho|^2}{\rho^2}+b\frac{\nabla^2\rho}{\rho},
\end{equation}
where $a=0.05$ and $b=0.14$.
We can now consider the functional derivative of the gradient expansion KE,
to generate gradient expansion formulae for $v_{p}$ and $v_r$.
The kinetic energy using second-order differentials of the density may be written as
\begin{equation}
\label{eq:tksapprox}
T^{approx}_{KS}=\int\tau^{approx}_{KS}\left[\rho(\boldsymbol{r}),\nabla \rho(\boldsymbol{r}),
\nabla^2\rho(\boldsymbol{r})\right]d^3r.
\end{equation}
The functional derivative of this form is
\begin{equation}
\label{eq:vksapprox}
\frac{\delta T^{approx}_{KS}}{\delta \rho} =
\frac{\partial\tau_{KS}^{approx}}{\partial\rho}-\nabla\cdot\frac{\partial\tau_{KS}^{approx}}{\partial\nabla\rho}+\nabla^2\frac{\partial\tau_{KS}^{approx}}{\partial\nabla^2\rho}.
\end{equation}
Now, consider an arbitrary second order GE of the form
\begin{equation}
\label{eq:genge}
T^{approx}_{KS} = \int \left(1+\eta_Q q+\eta_P p\right )\tau_{TF} \,d^3r
\end{equation}
To get the Pauli KE, we
subtract the von Weizsacker kinetic energy $\int (5p/3) \tau_{TF} \,d^3r$
from Eq.~(\ref{eq:genge}).
Then applying the functional derivative [Eq.~(\ref{eq:vksapprox})]
yields the GE approximation of the Pauli potential:
\begin{equation}
\label{eq:vpauliGE}
v_{p}^{GE}=\left[\frac{5}{3} + (\eta_P-5/3) p - 2(\eta_P-5/3) q\right]\frac{\tau_{TF}}{\rho}.
\end{equation}
which is independent of $\eta_Q$ because the Laplacian
term $\sim q$ does not contribute to the KE or its functional derivative.
The response function follows trivially from Eq.~(\ref{eq:vpauli}):
\begin{equation}
\label{eq:vrespGE}
v_{r}^{GE}=\left[\frac{2}{3} - (\eta_Q+ 2(\eta_P-5/3))
q\right]\frac{\tau_{TF}}{\rho}.
\end{equation}
We finish by considering the regions of general behavior in atomic electron densities for large-$Z$ atoms proposed by the analysis of
Lieb and Simon~\cite{liebsimon} and augmented by Heilman and Lieb.~\cite{heilmannlieb}
Moving outward from the center, there is first a region near the nucleus where TF behavior breaks down, consisting
of electrons whose behavior can be described described by the Bohr atom of non-interacting electrons.~\cite{heilmannlieb}
There is an inner core region, the density of which should behave as a slowly varying electron gas obeying TF theory [Eq.~(\ref{eq:tautf})].
The characteristic length of this core scales as $Z^{-1/3}$ and
the density scales as $Z^2$.
There is a ``mantle of the core" also with length scale $Z^{-1/3}$, and in which the density decays as $1/r^6$.
In the infinite-$Z$ atom the ratio of electrons outside core and mantle
to those inside drops to zero.
Then, there is a ``complicated transition region''~\cite{liebsimon} and a valence
region of outer shells with a length scale presumably of order 1.
Finally there is an evanescent region where the electron density decays exponentially.
One check of how well we have approached the $Z\to\infty$ limit may be how many of these regions we can actually detect in our data.
\section{Methods} \label{Methods}
In order to calculate Kohn-Sham orbitals and eigenvalues needed for the
calculation of exact KEDs and response potentials we use
the atomic code FHI98PP\cite{FHI98PP}
in its all-electron, non-relativistic mode.
FHI98PP computes wave functions on a logarithmic grid, with spacing
between successive points increasing by a geometric factor $\gamma$.
We use the default $\gamma=0.0247$ which yields inappreciably different
results from 0.0123.
For simplicity, the exchange-correlation functional used was the PW91
LSDA.\cite{numGGA}
The disagreement between LDA and exact Kohn-Sham calculations is known
to disappear in the large-$Z$ limit;~\cite{BCGP16} in practice,
our kinetic energies agree with exact OEP calculations
within 0.67\% for Ne and 0.055\% for Rn. See Supplemental Material
for further details.
For differentiation of functions we employ Lagrange interpolating polynomials, similar to Gauss quadrature, with polynomials up to twelfth-order, while
for integration we use the composite Simpson's method. Details are given
in Ref.~\onlinecite{thesis}.
The construction of extremely large atoms should also be discussed.
For the nonrelativistic case, one naively extends the Aufbau principle out to infinity.\cite{pyykko}
For atoms with highly degenerate valence energy shells, the lanthanum and actinium series for instance, this is likely a poor assumption, because
completing such a shell might take preference over filling a lower energy
shell with low degeneracy.
However for atoms in the eight principal columns of the periodic table, all
highly degenerate shells are already completely filled and thus do not
influence the filling order.
We have extended FHI98PP using the Aufbau principle out to element number 976,
with a 16p valence shell.
The validity of this extension has been tested by comparing the total energy of Aufbau-constructed shells versus several other shell configurations for
elements 976 (filled 16p), 970 (filled 15d), and 816 (filled 16s).
For all cases tested, the Aufbau construction proves to be the nonrelativistic
ground state for these atoms.
To check the quality of our numerical solutions,
completely independent calculations were done with a second
atomic DFT code, OPMKS,~\cite{opmks} for atoms with $Z<400$.
The results are indistinguishable with those of FHI98PP within machine error.
A table of highest occupied atomic orbital (HOAO) eigenvalues and
kinetic energies of large $Z$ atoms
from both methods is given in the Supplemental Material.
\section{Results} \label{Results}
\subsection{Verifying densities}
As a partial confirmation of our method,
we compare densities generated by FHI98PP and the Aufbau principle for
the mathematical element with $Z=976$ to the Thomas-Fermi density using the
numerical parameterization of Ref.~\onlinecite{LCPB}.
Fig~\ref{fig:denstest} shows the
KS density (blue line), the TF density (red dotted line), and the TF limit of
the density (black dotted line) versus scaled radius for element number 976.
The scaling of $Z^{1/3}$ reflects the radius of the atom in the TF approximation, with peak radial density occurring for $Z^{1/3}r\sim 1$.
Note the high agreement between the KS density and the TF density over a large
range in scaled radius.
However, though suppressed by the log plot, shell structure is evident in the
KS density as oscillations around the TF density, wih the density deviating
from the TF limit especially for the last oscillation or two ($Z^{1/3}r>10$).
As expected the density diverges from the TF density for very large
values of scaled radius, in the region of exponential decay beyond the last
occupied shell. The Thomas-Fermi model assumes an infinite number of
particles and continues indefinitely with the density decaying as $1/r^6$.
The KS density never quite reaches the $1/r^6$ large-$r$ limit of the TF
density
(the ``mantle'' of the core of Ref.~\onlinecite{liebsimon}) and
in this sense has not completely reached the TF limit.
\begin{figure}[!htbp] \includegraphics[width=1.0\linewidth,height=0.41\textheight,keepaspectratio]{denstest-eps-converted-to.pdf}
\caption{\label{fig:denstest}Comparison of KS density (KS), the TF density (TF), and the asymptotic TF limit (TF limit) for the Z=976 neutral atom.}
\end{figure}
\subsection{Pauli Potential: near the nucleus \label{sec:nearnucleus}}
As suggested by Lieb and Simon's schema for describing the
large-$Z$ atom, it is helpful to investigate Pauli potentials for separate
regions of space. We thus examine first, the potential of the one
or two electron shells nearest the nucleus, then that of
the core and valence shells,
and finally the evanescent behavior far from the nucleus.
Fig~\ref{fig:vpBe} shows the Pauli potential and the two components that are used to construct it via Eq.~(\ref{eq:vpauli}) -- the Pauli KED divided by the density and the response potential. The constant $\epsilon_{M}-\epsilon_{0}$ is shown as a solid red line.
\begin{figure}[!htbp]
\includegraphics[width=1.0\linewidth,height=0.41\textheight,keepaspectratio]{Vp_vs_r_Be-eps-converted-to.pdf}
\caption{\label{fig:vpBe} Various contributions to the Pauli
potential for Be. Blue is the Pauli KED divided by the density,
close black dotted is the response potential, and red dotted is the Paul
potential.
The eigenvalue difference $\epsilon_M-\epsilon_0$ between
highest and lowest occupied orbitals is shown as a solid black line.
}
\end{figure}
Beryllium is a usefully didactic system since it has only two shells -- it is in effect a two-state system and the simplest atomic structure that has a non-zero Pauli
contribution. The Pauli KED is nonzero only in the transition region between the 1s and 2s shells. Further out, it is zero because only the 2s shell effectively contributes to the KED -- it becomes effectively a single state system indistinguishable from the bosonic case. Inside, the issue is more complicated. The 2s shell has a small nonzero piece and so naively one would expect the Pauli KED to be nonzero, but as we discuss in detail below, the Pauli KED is exactly zero at the nucleus as a consequence of the nuclear cusp condition on the density.
The response potential $v_r$ for Be is essentially a step function with a
single step from the 1s shell, having an eigenvalue close to the hydrogenic 1s
value, to the 2s shell.
The second shell is the highest occupied energy shell and thus, given the
definition of $v_r$ [Eq.~(\ref{eq:vrespexact})], makes zero contribution to
the numerator of the response potential. The response potential of the
lowest energy shell agrees reasonably well near the nucleus by the two-state
energy difference
$\epsilon_M-\epsilon_0 = \epsilon_{2s} - \epsilon_{1s}.$\cite{baerends1997quantum}
The net effect on $v_p$ of the two contributions to it in Eq.~(\ref{eq:vpauli}) is instructive. Recall the conceptual definition of $v_p$:
given a system of fermions in an external potential that one wishes
to replace with a fictitous system of bosons with the same ground state
density, then $v_p$ is the potential that one needs to add to the
external potential in the bosonic system to achieve this.
Here the goal is to make the density from a single bosonic state
$\psi \sim \sqrt{\rho}$ duplicate the two shells of the fermionic system.
This is done by creating a potential step (due to $v_r$) that pushes density out of
the 1s shell region into the 2s shell region, while an additional barrier (due
to $\tau_p/\rho$) separates the charge into two distinct shells. Finally, one
may note that the values of $v_p$ and $v_r$ at contact with the nucleus
are equal to each other and slightly less than $\epsilon_M-\epsilon_0$.
As there are known asymptotic behaviors for both total energy and near nuclear energy densities related with large Z scaling for the Kohn-Sham KE,
it is of interest to analyze the large Z scaling of $v_p$.
Fig~\ref{fig:vpRn} plots the same quantities as Fig~\ref{fig:vpBe}, but for Rn. In addition we plot three gradient expansion models for the response potential discussed Sec~\ref{Theory} -- the canonical GEA (purple dashed), the fit to the KED of high-Z atoms ($v_{Loc}$) (green dash), and the model of Ref.~\onlinecite{Ts} (yellow dashed). Note that every potential visually has a three step structure with transitions at $Z^{1/3}r=0.1$ and $Z^{1/3}r=1$, related to the three innermost of six occupied shells (the remaining three shells are too small to see in this plot). In comparison to Be, $v_r$ seems to retain the step structure and $\tau_p$ has weak local maxima in between shells, but the shell structure overall is less pronounced.
Note that for Rn, $v_p$ is almost exactly $\epsilon_m-\epsilon_0$ in the
near-nuclear region.
This trend continues to improve as $Z\to\infty$, however visually Rn
essentially shows complete agreement between $\epsilon_m-\epsilon_0$ and $v_p(r=0)$, so no larger Z atoms are plotted in this fashion.
The actual contact value for $v_p(0)$ is much larger than that for
Be -- the energy scale is roughly $Z^2$, that of the noninteracting
hydrogen-like system.
At the same time $\tau_p(0)$ is definitely non-zero and so $v_r(0)$ and
$v_p(0)$ are no longer the same value.
The GEA models all trend toward $-\infty$ as $\boldsymbol{r}\to0$. This is due to the charge singularity at the origin, resulting in a Laplacian of the density that
diverges in this limit.
This is a flaw in any GEA model
and is caused by the divergence term in Eq.~(\ref{eq:vksapprox}).
It is notable that the TS model does come close to predicting the turning point for $v_r$.
\begin{figure}[!htbp]
\includegraphics[width=1.0\linewidth,height=0.41\textheight,keepaspectratio]{Vp_vs_r_Rn-eps-converted-to.pdf}
\caption{\label{fig:vpRn}
Components of the Pauli potential,
$\frac{\tau_p}{\rho(\boldsymbol{r})}$, $v_p$, and $v_r$, for Rn, compared to
response potential from several DFT models.
GEA is the gradient expansion approximation [Eq.~(\ref{eq:taugea2})],
Airy, the Airy gas gradient expansion [Eq.~(\ref{eq:gealoc})],
loc, the fit to the local KED for high-$Z$ atoms [Eq.~(\ref{eq:tauairy})],
Ts, the Tsirelson model [Eq.~(\ref{eq:TSprefit})].
$\epsilon_M-\epsilon_0$ is difference between highest and lowest occupied eigenvalues.
}
\end{figure}
\subsubsection{Analytic analysis of nuclear region}
It seems from Fig.~\ref{fig:vpRn} that as $Z\to\infty$, the value of
$v_p$ near the nucleus approaches a constant equal to
$\epsilon_M-\epsilon_0$. At the same time, $\tau_p/\rho$ does not seem to
converge to zero here, but rather to a value about 10$\%$ of $v_p(0)$;
thus $v_{r}(0)$ falls short of $v_p(0)$ by the same amount.
These asymptotic behaviors can be proven mathematically.
Naively, if one considers Eq.~\ref{eq:vrespexact}
as $\boldsymbol{r}\to0$, and assume that the 1s orbital is the primary contribution
to this equation, one gets
\begin{equation}
\label{eq:vrespzeroapprox}
\lim_{r\to0}v_{r}\approx(\epsilon_M-\epsilon_0).
\end{equation}
This can not be exactly true because every s orbital has a contribution at the nucleus.
To improve the description, we consider what happens near the nucleus in the
limit that nuclear charge $Z$ and electron number both go to infinity.
Even for finite $Z$ the effect of electron-electron interactions
becomes very small compared to the nuclear potential and thus
they can be ignored. Low-lying energy eigenvalues approach in energy and
degeneracy those of the corresponding noninteracting system --
a Coulombic nuclear potential with charge $Z$.
One thus can use hydrogenic wavefunctions to construct both $\tau_p$ and
$v_{r}$ near the nucleus. This should be accurate out to a radius of
order $r/Z$ where Lieb and Simon~\cite{liebsimon} show that the electron density starts to resemble that of the Thomas-Fermi atom.
To build a model for $v_r$ based on this picture we first note that
for a hydrogenic central potential, the value at the nucleus of the radial
component $R_{nl}$ of an eigenfunction is
\begin{equation}
\label{eq:Rhydend}
R_{nl}(0) = \sqrt{4\left(\frac{Z}{n}\right)^3} \delta_{l0}.
\end{equation}
Here the identity $L_{n-1}^{1}(0)=n$ for Laguerre polynomials has been used.
Next, we apply this result to the exact expression for $v_{r}$
[Eq.~(\ref{eq:vrespexact})] by defining the net density of an angular
momentum subshell,
\begin{equation}
\rho_{nl}(r) = \frac{(2l + 1)}{4\pi} \left|R_{nl}(r)\right|^2
\label{eq:rhonl}
\end{equation}
and, following Ref.~\onlinecite{heilmannlieb}, the density of a complete
energy shell:
\begin{equation}
\rho_{n}(r) = \sum_{l=0}^{n-1} \rho_{nl}(r).
\end{equation}
For a hydrogenic system, given the degeneracy in energy over angular
momentum quantum number $l$, one has
\begin{equation}
v_{r}(r) = \sum_{n=1}^M
\left( \epsilon_M - \epsilon_n \right) \rho_n(r)/\rho(r).
\label{eq:vresphydrogen}
\end{equation}
Since we are considering the limit $Z\to\infty$, it is appropriate also to
take the limit that $M\to\infty$, that is, the perfect ``Bohr atom'' where
all orbitals are filled and all are given by those of the noninteracting
hydrogen atom.
At the origin, only $l=0$ contributes, so $\rho_{n}(0) = \rho_{n0}(0)$.
Thus Eq.~(\ref{eq:vresphydrogen}) reduces to
\begin{equation}
\label{eq:vrespzero}
v_{r}(0) = \braket{\epsilon_{n0}}
= {\displaystyle -\frac{\sum_{n=1}^{\infty}
\epsilon_{n0} \rho_{n0}(0)}
{\sum_{n=1}^{\infty} \rho_{n0}(0)}},
\end{equation}
where we assume $\epsilon_{\infty}=0$.
Substituting in Eqs.~\ref{eq:rhonl} and~\ref{eq:Rhydend} and summing over
$n$ gives
\begin{equation}
\label{eq:rhohydrogen}
\rho(0)=\sum_{n=1}^{\infty}\frac{2}{\pi}\left(\frac{Z}{n}\right)^{3}.
\end{equation}
Similarly, we use $\epsilon_n = -Z^2/2n^2$ for a hydrogenic atom and
repeat this process to get the numerator of Eq.~\ref{eq:vrespzero}.
With a bit of manipulation
one can write the ratio as
\begin{equation}
\label{eq:vrespzerolargeZ}
{\displaystyle
v_{r}(0) = {\frac{Z^2}{2} \frac{ \zeta(5)}{ \zeta(3)},
}
}
\end{equation}
where
\begin{equation}
\zeta(s)=\sum _{n=1}^{\infty }{\frac {1}{n^{s}}}
\end{equation}
is the Riemann-Zeta function.
Given that $\zeta(5)=1.03692$ and $\zeta(3)=1.20205$, one has as $Z\to\infty$
\begin{equation}
v_{r}(0)=0.862626\frac{Z^2}{2}=-0.862626\epsilon_0.
\end{equation}
The Pauli KED near the nucleus can be analyzed in a similar fashion, with
more difficulty, since it necessarily involves derivatives of orbitals.
It is fairly straightforward to show that the contribution of $s$-orbitals
to the KS KED exactly equals the von Weizsacker KED in this region.
In effect, this
describes the connection between the cusp conditions near the nucleus
obeyed by $s$-orbitals, and that of the total density. Somewhat
counterintuitively, $p$-orbitals also have a nonzero contribution to the KS KED
near the nucleus, both radially and from their nonzero angular momentum\cite{cancioredd, thesis, acharya, constantin2016kinetic}.
It is the contribution from these orbitals that cause $\tau_p$ to be nonzero
near the nucleus.
The Bohr atom model used here for $v_{r}$ has recently been used by
Constantin et al. to analyze the large-$Z$ limit of $\tau_p$ at the nucleus.
In this case, they show (Eq.~(20) of Ref.~\onlinecite{constantin2016kinetic})
\begin{eqnarray}
\label{eq:taupconst}
\lim_{Z\to\infty} \tau_p(0)&=&
\sum_{n=1}^{\infty} 3\tau_{vW}[\rho_{n1}](0) \\
&=& \sum_{n=1}^{\infty} \frac{(n^2 - 1)Z^5}{\pi n^5},
\end{eqnarray}
where $\tau_{vW}[\rho_{nl}]$ is the vW KED evaluated using the density of
the $(nl)$ angular momentum subshell.
The end result is closely related to that for the $Z\to\infty$ limit of
$v_{r}(0)$:
\begin{equation}
\label{eq:taupoverrho}
\frac{\tau_p(0)}{\rho(0)} =
\frac{Z^2}{2} \frac{\zeta(3)-\zeta(5)}{\zeta(3)}
\end{equation}
and therefore, using Eq.~(\ref{eq:vpauli})
\begin{equation}
\label{eq:vpepsilonnot}
\lim_{Z\to\infty} v_p(0)= \frac{Z^2}{2}
\end{equation}
This may be recast in a form that is more robust as well as conceptually
revealing. First we note that the limit $Z^2/2$ is shared with the
lowest orbital eigenvalue:
\begin{equation}
\lim_{Z\to\infty} \epsilon_0 = -\frac{Z^2}{2}.
\end{equation}
Then, observing that the form of the response potential
involves a difference between the highest occupied eigenvalue
$\epsilon_M$ and the other occupied eigenvalues,
and noting that $\epsilon_M$ is a small energy independent of $Z$,
we posit the general limit for $v_p(0)$:
\begin{equation}
\label{eq:vpepsilon}
v_p(0)\sim \epsilon_M-\epsilon_0.
\end{equation}
We verify these assumptions first by plotting, in
Fig.~\ref{fig:vpZenot}, ($\epsilon_M-\epsilon_0)/Z^2$ (red dashed)
and $-\epsilon_0/Z^2$ (blue) for alkali metals and noble gases from
He to $Z=976$.
These are plotted against the small parameter $Z^{-1/3}$
that characterizes the large-$Z$ expansion of atomic energies.
We fit this trend with a polynomial form $(ax^2+bx+0.5)$ with $x = Z^{-1/3}$.
A least squares regression results in $a\!=\!-0.879 \pm 0.017$ and $b\!=\!-0.091 \pm 0.004$.
The fit is highly accurate for large Z atoms, starts to deviate from the
observed around $Z<64$, but is still within 10\% of the true value for Ne.
One may note that dropping $\epsilon_M$ from the approximation for $v_p(0)$
affects the result primarily for He where in fact $\epsilon_M=\epsilon_0$.
At the same time, the value of $-\epsilon_0$ is significantly
off the hydrogenic value of 0.5 for any realistic value of $Z$.
\begin{figure}[!htbp]\centering
\includegraphics[width=1.0\linewidth,height=0.41\textheight,keepaspectratio]{Vp_vs_Z_enot-eps-converted-to.pdf}
\caption{\label{fig:vpZenot}
Analysis of lowest energy eigenvalue of large-$Z$ atoms
versus the small parameter
$Z^{-1/3}$ for the large-$Z$ expansion of atomic energies.
We show $-\epsilon_0+\epsilon_M$, $-\epsilon_0$, and a curve fit of the
form $ax^2+bx+0.5$ versus
$x=Z^{-1/3}$ for noble gases from He to $Z=976$. }
\end{figure}
To test our assumptions of the finite-$Z$ value
of $v_p(0)$ [Eqs.~(\ref{eq:vpepsilonnot}) and (\ref{eq:vpepsilon})]
work,
we next plot in Fig.~\ref{fig:vpZ} the value of
$v_p - v_r$ (blue) at the origin for all atoms in columns 2, 13, and 18 of
the periodic table, extended to $n=16$ ($Z=976$.)
This is again scaled by $Z^2$ and plotted versus $Z^{-1/3}$. Subtracting
off $v_r$ removes the large majority of the Pauli potential at the origin,
leaving a relatively small piece (equal to $\tau_p/\rho$) which makes
the error in our limiting ansatz readily visible.
Then we compare to the difference between the large $Z$
limit $\epsilon_M-\epsilon_0$ and $v_r$ (black dashed line) and repeat for the
the less accurate limit $-\epsilon_0$ (red dashed).
Data is taken from three columns of the periodic table, and differentiated
by plotting points of different types.
A complete table of data used to generate fig.~\ref{fig:vpZenot} and fig.~\ref{fig:vpZ} /is included in supplemental materials.
Note that all three functions of Z approach the same limit as $Z\to \infty$,
converging to less than 1\% error in $v_p(0)$ by roughly $Z=36$.
Furthermore this dependence is column independent-- curves from each column plotted fall onto the same trend after just one shell. This makes sense since we are measuring the Pauli
potential at the nucleus, where presumably the effects of a variably filled
valence shell should be minimal.
The effect of including $\epsilon_M$ in our model is felt most for
single-shell systems like He,
where it retrieves the exact value of zero for $V_p(0)$.
We make a parabolic fit of the data to the trend
$cx^2+dx+0.07$ where $x=Z^{1/3}$ (black dash-dotted line).
A least-squares regression of our data at large $Z$
results in $c=-0.223 \pm 0.009$ and $d=-0.0491 \pm 0.0024$.
The value of 0.07 for the $y$ intercept is determined using
Eq.~\ref{eq:taupoverrho}.
The fit has a very weak linear term, indicating
that the contact value of $v_p$ roughly varies with nuclear charge
as $0.5 Z^2 + cZ^{4/3}$.
\begin{figure}[!htbp]\centering
\includegraphics[width=1.0\linewidth,height=0.41\textheight,keepaspectratio]{Vp_vs_Z-eps-converted-to.pdf}
\caption{\label{fig:vpZ}Contact values for $\tau_p/\rho$,~$-\epsilon_0-v_{r}$, and $\epsilon_M-\epsilon_0-v_{r}$ as a function of $Z^{-1/3}$. Values taken atoms from columns 2 and 13 and 18 respectively,
extended to $Z=976$, and compared to a parabolic fit $cx^2+dx+0.07$.
}
\end{figure}
\subsection{Core and Valence}
Next we consider the behavior of the Pauli potential
and its constituents $\tau_p/\rho$ and $v_{r}$, away from the nucleus.
For the ease of visualization across many shells,
we employ unitless, scale-invariant quantities. For the kinetic energy
density, it is common to do so by defining an enhancement factor, $F$, relative
to the Thomas-Fermi KED:
\begin{equation}
F = \tau_{KS} / \tau_{TF}
\end{equation}
and equivalently, a Pauli enhancement factor defined by
\begin{equation}
\tau_{p} = F_p\tau_{TF}
\end{equation}
so that $F_p$ is given by
\begin{equation}
\label{eq:enhancement}
F_p = \left(\tau-\tau_{vW}\right)/\tau_{TF}.
\end{equation}
For any model for the Kohn-Sham kinetic energy density,
model Pauli enhancement factors may be similarly defined.
Beyond the obvious advantages of scale invariance, the Pauli enhancement
factor for the KS KED is closely related to the Electron Localization Factor (ELF)~\cite{BeckeEdgecombe} and is equal to the $\alpha$ term used in
meta-GGA functionals.~\cite{Becke98,SXR12}
In order to produce a unitless representation of the
Pauli potential and its components, we scale each quantity by
$\tau_{TF}/\rho$, the ratio of KE and particle densities in the
TF model. This is $3/5$ of the local fermi energy $\epsilon_F$ in the
TF picture.
Fig.~\ref{fig:Fresp_Rn}(a) plots
the Pauli enhancement factor $F_p$ for the Kohn-Sham KE density of radon.
This is compared to various GE approximations:
the standard gradient expansion, the Airy gas model~\cite{lindmaa14}, the local fit to the gradient expansion of Ref.~\onlinecite{cancioredd} and the model of
Ref.~\onlinecite{Ts}. These are plotted against the scaled distance
$x = Z^{1/3}r$, chosen so that the peak radial probability density in the
TF model for any atom occurs at roughly $x\!=\!1$.
The constant line at one shows the TF limit for $F_p$.
A notable feature of these plots is the nearly periodic oscillation
of the exact KE density and the GE models about the TF limit.
This behavior reflects the shell structure of the atom:
a value of $F_p < 1$ indicates a region dominated by a single shell,
producing a value for $\tau_{KS}$ lower than that predicted by TF theory, while
the opposite is true for $F_p > 1$.
Thus each minimum indicates a different principal quantum
shell.
The five maxima show the regions of transition between the six shells of Rn,
while the last exponentially divergent tail at large $r$ is the classically
forbidden evanescent region outside the atom.
It is interesting the oscillations have a roughly equal period in a
semi-log plot, suggesting exponential growth in the period of quantum
oscillations.
\begin{figure}[!htbp]\centering
\subfigure[]{
\includegraphics[width=1.0\linewidth,height=0.41\textheight,keepaspectratio]{F_resp_models_Rn-eps-converted-to.pdf}
}
\subfigure[]{
\includegraphics[width=1.0\linewidth,height=0.41\textheight,keepaspectratio]{F_resp_Rn-eps-converted-to.pdf}
}
\caption{\label{fig:Fresp_Rn}
Unitless representation of KE potentials for radon.
(a) Pauli enhancement factor $F_p$ for the Kohn Sham KED, compared to
several variations of the gradient expansion approximation, versus scaled
radius $Z^{1/3}r$. Models shown are the standard gradient expaonsion (GEA),
Airy gas model (Airy), fit to the local KED of atoms (Loc) and the Tsirelson
model (Ts);
the Thomas-Fermi limit is shown as the solid horizontal line.
(b) Pauli potential $v_p$ (solid blue line), response potential $v_r$ (dashed blue line), and gradient expansion approximations of the same,
scaled by $\tau_P/\rho$.
Each model is shown with the same color and dashing for
$v_p$ and $v_r$ but center on different TF limiting cases -- 5/3 for the
former, 2/3 for the latter.
}
\end{figure}
All versions of the gradient expansion recover the main qualitative features
of the Pauli enhancement factor away from the nucleus and for the most part
are quite accurate quantitatively.
The quantum oscillations of the Tsirelson model start to deviate from the
TF limit in the outer three shells; also the Airy gas model
overestimates the true enhancement factor by a scaling factor of roughly 10/9.
Fig.~\ref{fig:Fresp_Rn}(b) plots unitless representations of the Pauli potential $\rho v_p/\tau_{TF}$ and response potential $\rho v_r/\tau_{TF}$ versus
scaled radial distance $Z^{1/3}r$ for radon. These
are compared to various gradient expansion models, as before.
The Pauli potential is not plotted for the Tsirelson model
because of its dependence on the functional derivative of other quantities
like exchange and correlation.
Referring to Eqs.~(\ref{eq:vpauliGE}) and~(\ref{eq:vrespGE}), we see that
the Thomas-Fermi limit of the scaled Pauli potential is the constant 5/3, and
that of the response potential, 2/3, shown as black horizontal lines.
Note that the scaled potentials show the same shell structure as the Pauli
enhancement factor, oscillating about their TF limis with peaks and minima
in nearly the same locations. The center of the oscillations starts
to deviate from the TF line slightly for the outer three shells.
This trend is better matched by the ``non-canonical'' GE's like the Loc and Airy models than by the standard GEA.
Comparing gradient expansion models for the Pauli potential
[Fig.~\ref{fig:Fresp_Rn}(b)] we note that the Loc and Airy gas models
outperform the conventional GE over most of the atom. Notably, the gradient
expansion of the Pauli potential [Eq.~(\ref{eq:vpauliGE})] only depends upon
the coefficient $\eta_P$ for the contribution of the gradient expansion
from the gradient variable $p[\rho(r)]$. Thus this data supports the
use of a negative $\eta_P$ coefficient, as in the Airy gas and Loc models,
in contrast to the positive coefficient of the standard gradient expansion.
In addition the \textit{a priori} Airy gas potential
is almost as good as the empirically fit Loc model.
At the same time, only the Loc model provides a close fit for the separate
pieces of the Pauli potential, $\tau_p/\rho$ and $v_r$ (the Airy gas
model is particularly poor for $v_r$, which underestimates the size of quantum
oscillations by a factor of 3). Thus Loc has the
best description of the coefficient $\eta_Q$ of the Laplacian term of the
expansion.
We now consider the
trends in atomic data as $Z$ is taken to be as large as practical in order
to try to piece out the high-$Z$ limit.
Fig~\ref{fig:Fresp_976}, similarly to Fig~\ref{fig:Fresp_Rn} (b), plots the unitless representation of $v_p$ and $v_r$, as compared to various GE
approximations, for element 976. This is the ``noble gas'' for row 16 of the periodic table mathematically extended using
the Aufbau principle -- thus there are 16 oscillations and 16 shells.
Gratifyingly, these oscillations have considerably less amplitude than for radon, indicating passage towards the high-$Z$ limit. Note that although the inner
eight shells of $v_p$ and $v_r$ oscillate about the TF limit,
there is now an unmistakable trend away from the TF limit in the outer shells.
This deviation is markedly absent in all the gradient
expansion models. The deviation seems to be linear, and does not start
until the middle shell of the atom is reached, around $Z^{1/3}r=1$.
The last oscillation in the potential, demarcating the valence shell,
occurs at $Z^{1/3}r\approx40$.
Similar plots are made for the Column 2, 10 and 13 atoms from row 16 in
the Supplemental Material.
\begin{figure}[!htbp]\centering
\includegraphics[width=1.0\linewidth,height=0.41\textheight,keepaspectratio]{F_resp_N76-eps-converted-to.pdf}
\caption{\label{fig:Fresp_976}
Unitless representation of the Pauli potential $v_p$, response potential $v_r$, gradient expansion approximation of the response potential $GEA$, Airy gas response potential $Airy$, the local variation of the GEA response potential $loc$, and the Tsirelson response potential $Ts$ for element 976.}
\end{figure}
To investigate further this unexpected behavior, we show
in Fig.~\ref{fig:Fresp_976_error} the error in
the Pauli enhancement factor $F_p^{model} - F_p^{exact}$ for the various
GEA models, for the $Z=976$ atom. This highlights the
``non asymptotic'' piece in $F_p$ -- the part that has not yet converged
to the Thomas-Fermi asymptote.
The scaled Pauli energy density $F_p$ shows no unexpected behavior
whatsoever. The non-asymptotic remnant
can be fit extremely well by a zero line, and its amplitude of oscillation
is quite small. This is consistent with the Lieb-Simon theorem
that the kinetic energy, obtained by integrating over the KE density must
tend to the Thomas-Fermi limit for large $Z$.
\begin{figure}[!htbp]\centering
\includegraphics[width=1.0\linewidth,height=0.41\textheight,keepaspectratio]{F_resp_models_N76_error-eps-converted-to.pdf}
\caption{\label{fig:Fresp_976_error} Difference between the exact
Pauli enhancement factor $F_p$ and various GEA models for element 976, as
function of scaled radius.}
\end{figure}
In contrast,
Fig.~\ref{fig:Fresp_pauli_row} plots the difference between the exact response potential and the standard GEA approximation [Eq.~(\ref{eq:taugea2})] for noble
gas elements with even principle quantum number $n$ up to $Z=976$.
The odd rows left out show a similar trend but with oscillations
out of phase with the those of even rows.
The deviation from the TF limit seen in
Fig.~\ref{fig:Fresp_976} here forms part of a trend,
apparent at radon and growing consistently with $Z$,
along a linear trendline versus $\log(Z^{1/3}r)$,
from about $Z^{1/3}r \!=\! 2$ out to the edge of each atom.
As $Z$ increases, the potential does not converge to the Thomas-Fermi
limit -- rather it deviates further away from it along this limiting
trend.
The valence shell (the last dip and peak before each curve drops to negative
infinity) deviates from the trend of the inner shells. However it forms its
own predictable linear trend away from the GE prediction, starting perhaps
with Kr, with same slope as the inner shells.
The difference between the exact response potential and the standard GEA model
for element 976 was fitted to a form $y = a\log(x/x_0)$ for $x\!=\!Z^{1/3}r$
and $y\! = \!\rho v_r / \tau_{TF}$. The results of a linear regression are
$a = 0.194\pm0.007$ and $x_0 = 1.39\pm 0.06$.
In comparison, $x=a_{TF}=0.88$ is the position of peak radial density for
the TF atom.
An overall model of deviation from the TF limit can thus be extracted:
\begin{equation}
v_r(r) = v_r^{GEA}(r) + 0.194 \frac{\tau_{TF}}{\rho} \log{(Z^{1/3}r/1.39)}
\end{equation}
A similar fit performed
for $v_p$ yields results that agree within the fit standard deviation.
(Details of the fit of the anomalies in $v_p$ and $v_r$ can be found
in the Supplemental Material).
\begin{figure}[!htbp]\centering
\includegraphics[width=1.0\linewidth,height=0.41\textheight,keepaspectratio]{F_pvsq_even_nobles-eps-converted-to.pdf}
\caption{\label{fig:Fresp_pauli_row} Scaled difference between $v_r$ and GEA response potential for noble gases with even principle quantum numbers up to $n=16$. The Thomas Fermi limit (TF) is narrow purple line; Ne, brown dashed; Kr, black dashed; Rn, green dashed; element 168, gold dot-dashed; 290, purple dot-dashed; 460, black dotted; 686, red dotted; and 976, blue solid.
}
\end{figure}
It is worthwhile to analyze this deviation in terms of the
scaled gradient and Laplacian of density in this region.
One would
expect these to become small nearly everywhere as $Z\to\infty$ and
the total energy tends to the TF limit,
but our
results with the potential call this expectation into question.
To this end, Fig.~\ref{fig:p_vs_q_even} shows parametric plots of the
scale-invariant quantities
$p(\boldsymbol{r})$ vs $q(\boldsymbol{r})$ for noble gases with even principle quantum
numbers. Parametric plots for other large-$Z$ atoms can
be found in the Supplemental Material.
There are three clear regions of behavior in this
plot.\cite{thesis, cancioredd}
The asymptotic approach of $q(\boldsymbol{r})\to -\infty$ indicates the
nuclear cusp.
The tail where $p(\boldsymbol{r})$ and $q(\boldsymbol{r})$ both tend to infinity
indicates the evanescent region far from the atom.
The loops or ``orbits'' come from the atomic core,
reflecting the oscillations in $F_p$ and $v_p$ seen in Figs.~
\ref{fig:Fresp_Rn} and~\ref{fig:Fresp_976}.
Each orbit represents a new shell, with $p$ and $q$
moving outwards to their largest values in regions between shells,
and approaching the TF limit $p\!=q\!=0$ in the center of each shell.
Thus one loop is seen for Ne with two shells.
For increasing $Z$ one may
in general see most loops shrinking towards $p\!=q\!=0$,
indicating that the TF limit is being approached locally.
But surprisingly, the process
stops for the outermost loop, starting with radon.
Subsequently this outer loop, formed by the valence shell and transition
to the next shell is largely invariant with atomic number. A second loop
seems largely stabilized by $Z=168$, and so on.
For element 976, the inner shells are all close to the TF limit.
However the outer orbitals gradually deviate from the TF limit,
and show no sign of ever converging to this limit as $Z \to \infty$.
So the deviation from the TF limit of $v_r$ for $Z^{1/3}>1$ is
indicated by a similar deviation in the GE variables, suggesting
that the GE will never become accurate for these shells.
Similar behavior can be seen for other columns of the periodic table, with
considerable differences in the last shell or two -- data for columns
2, 12 and 13 of the periodic table are shown in the Supplemental Material.
\begin{figure}[!htbp]\centering
\includegraphics[width=1.0\linewidth,height=0.41\textheight,keepaspectratio]{pvsq_evens-eps-converted-to.pdf}
\caption{\label{fig:p_vs_q_even}
Parametric plot of $q(r)$ versus $p(r)$ for noble gas atoms with
even principle quantum numbers.}
\end{figure}
It is interesting to note that Eq~\ref{eq:tsqrtrho} implies that
\begin{equation}
\label{eq:vpts}
v_p=\mu-v_{KS}-v_{vW},
\end{equation}
where $v_{vW} = \frac{\delta T_{vW}}{\delta\rho}$ and $\mu\sim 0$ may
be taken for a large-$Z$ atom. This implies that the
unexpected behavior in the Pauli potential ought to mirrored in the KS
potential as well. We find that this is indeed the case, calculating
the XC potential using
the Leeuwen-Baerends exchange potential~\cite{van1994exchange},
and the Perdew-Zunger LDA correlation potential~\cite{PerdewZunger}.
The combined Hartree plus XC potential veers off the
TF limit of $-5\tau_{TF}/3\rho$, to nearly cancel the Pauli potential, with each piece contributing about half of the net effect. The vW contribution
does not show any unexpected behavior.
\subsection{Evanescent Region}
Fig.~\ref{fig:tail_row_16} allows one to examine the evanescent region
more closely.
It plots the response potentials for elements
976 (red), 971 (black), 970 (blue), and 816 (gold), that is,
the atom for column 2, 12, 13 and 18 for the 16th row of the
extended nonrelativistic periodic table.
Elements 976 and 971 both have p shells as their highest occupied atomic
orbital (HOAO) and tend asymptotically to infinity.
Elements 970 and 816 both have s shells as their HOAO and tend to zero.
(Thus elements with s shells as their HOAO lack the last local minimum
in $v_r$ present in the other cases.)
The same trends are shown by the Pauli potential and the Pauli
enhancement factor $F_p$.
\begin{figure}[!htbp]\centering
\includegraphics[width=1.0\linewidth,height=0.41\textheight,keepaspectratio]{F_resp_tail_row-eps-converted-to.pdf}
\caption{\label{fig:tail_row_16}Evanescent behavior of $v_r$ for the alkali atom ($\!=\!816$), closed d shell atom (970), column 13 atom (971), and noble gas atom (976) with highest principle quantum number 16, as well as
the GEA for each (dotted lines).}
\end{figure}
The standard GEA to the response potential for each element is
shown with the same color and a dashed line.
Other GE-like models (Loc, Airy gas) make very similar predictions.
Note that the GEA shows roughly the
same limiting behavior regardless of the column of the periodic table.
They make reasonably accurate predictions for elements that have
p shells as their HOAO but completely fail for the others.
One can partly explain the asymptotic trend of $v_r$ as $r\to \infty$
with a simple analysis of Eq.~\ref{eq:vrespexact}. The highest energy
shell that contributes to an atom with $M$ shells is the $M-1$ shell.
Defining $\rho_M$ as the density of the $M$-th energy shell,
the assumption that $\rho(r)\approx\rho_M(r)$ as $r\to\infty$
gives the following approximation:
\begin{equation}
\label{eq:vrevanend}
\lim_{r\to\infty} \frac{\rho(r) v_r(r)}{\tau_{TF}(r)}
\approx\frac{2(\epsilon_M-\epsilon_{M-1})\rho_{M-1}(r)}{[\rho_M(r)]^{5/3}}.
\end{equation}
The slope of the asymptotic behavior
of Fig.~\ref{fig:tail_row_16} should thus be predicted by the logarithm
of this result.
The long-range exponential decay constant for the HOAO orbital is proportional
to the square root of its eigenvalue, $\sqrt{2|\epsilon_M|}$, and this
dominates the behavior of the density in the evanescent region.~\cite{mparrlevy}
A reasonable expectation for the decay constant for the second HOAO, at least for
the local or semilocal exchange models employed in most DFT's is,
$\sqrt{2|\epsilon_{M-1}|}$.~\cite{evanescent} In this case, the roughly
linear behavior in Fig.~\ref{fig:tail_row_16} can be explained by a decay rate
$\kappa \sim \sqrt{|\epsilon_{M-1}|} - (5/3)\sqrt{|\epsilon_{M}|}$,
as shown in Table~\ref{table:evanescenteigen}. The
predicted rates agree closely with the observed behavior in
Fig.~\ref{fig:tail_row_16}. In
general, the main predictor of the evanescent rate is whether the
HOAO and second HOAO have the same or different principal quantum numbers.
The former case leads to a $v_r$ that dies off slowly and tends to
infinity relative to the local fermi energy $\sim \tau_{TF}/\rho$,
while the latter case shows the opposite effect.
Finally we note that the same qualitative trend in asymptotic behavior
occurs for the full Pauli potential and the additional
term $\tau_P/\rho$ that contributes to it, despite centrifugal terms~\cite{constantin2015atom} that contribute to these quantities.
\begin{table}[htb]
\caption{\label{table:evanescenteigen} The two highest
eigenvalues ($\epsilon_M$ and $\epsilon_{M-1}$) belonging to the
HOAO and second HOAO states respectively,
and the predicted exponential decay rate $\kappa$,
for atoms in the
16th row of the extended periodic table. Negative $\kappa$ indicates
exponential growth.}
\begin{ruledtabular}
\begin{tabular}{l|ccc}
Atom & $\epsilon_M$ & $\epsilon_{M-1}$ & $\kappa$ \\ \hline\hline
816 &-0.0859& -0.4004 & 0.1443\\ \hline
970 &-0.1389 & -0.4110 & 0.0199\\ \hline
971 &-0.0928 & -0.1827 & -0.0802\\ \hline
976 &-0.2113 & -0.3572 & -0.1685\\
\end{tabular}
\end{ruledtabular}
\end{table}
\section{Discussion and Conclusions}\label{Conclusions}
In this paper, we have mapped the behavior of Pauli potentials of
closed-shell atoms up to $Z=976$ or 16 complete energy shells.
This represents a partial traversal of Lieb-Simon scaling
to infinite $Z$, a process that transforms the Hamiltonian and expectations
of real atoms to a limit where Thomas-Fermi theory is relatively exact for
energies.
Unlike energy expectations, expectations that are functions of position --
the electron density and, in this paper, the Pauli potential -- do not have
to go to the Thomas-Fermi limit uniformly, and the electron density is
richly structured even in the Thomas-Fermi limit.
We find that this is true of the Pauli potential as well.
In comparison to the six regions of the large-$Z$
limit of electron density defined by Refs.~\onlinecite{liebsimon} and
~\onlinecite{heilmannlieb}, we can identify in our results perhaps five:
\begin{enumerate}
\item There is a near-nuclear region of constant Pauli potential.
\item An inner core region consisting of half of the occupied energy
shells where the potential oscillates about the TF limit.
\item An outer core region, where the potential experiences an unexpected
departure from the TF limit.
\item A small valence region where the last oscillation occurs,
where the potential deviates slightly from the anomalous trend, varying
somewhat between columns of the periodic table.
\item An evanescent region that
also varies for each column of the periodic table.
The slope of this evanescent region is related to the eigenvalues of the
last two shells.
\end{enumerate}
Notably,
the $1/r^6$ limiting behavior of the TF atom density is just barely hinted at
for the largest atom we study.
The Lieb-Simon
limit is understandably harder to reach for local features like the Pauli
potential than for globally integrated quantities like the kinetic energy.
In the near-nuclear region (1) we find a
constraint on the Pauli potential, analogous to
that on the Pauli KED found in Ref.~\onlinecite{constantin2015atom}.
The Pauli potential in the limit $r\to 0$ tends to the difference between the
highest and lowest occupied eigenvalues.
This result is consistent with the interpretation of the OFDFT Euler
equation [Eq.~(\ref{eq:tsqrtrho})] as solving for the density of a system of
fictitious bosons
constrained to have the same density as the true fermionic system.
Then the action of the Pauli potential in this region is to shift
the energy of ones fictitious bosonic system from
the lowest energy level of the actual fermionic system to that of the
chemical potential $\mu$.
It is interesting that the response potential in this region is nearly
constant. Using the theorems of Ref.~\onlinecite{heilmannlieb},
it should be possible to prove that the slope of the response potential
and thus the Pauli potential is zero at the nucleus.
The inner core shells are the only region where the Pauli potential
clearly tends to the Thomas-Fermi limit as $Z\to\infty$. It is
well fit by the gradient expansion in whatever variant, with the
best candidate being the Loc GEA, which was fit to the KED in this region.
This helps justify this model
because KED by itself is ambiguously defined -- it is the total energy
and the potential that are the physical measurables of the system.
The key point for an optimal fit seems to be a negative value to the
coefficient for the $s^2$ term in the expansion --
the Airy gas model with a similar negative coefficient
performs about as well.
Nevertheless, the fact that \textit{all} variants of the gradient
expansion work nearly equally well tells us that the dominant contribution
of the Pauli potential comes from the removal of the von Weizs\"{a}cker
KE from the Kohn-Sham kinetic energy. In fact, having no gradient
correction at all, i.e., a Thomas-Fermi energy, would produce quite a
good Pauli potential for atoms, if the real density could be used.
That is, the Thomas-Fermi energy is not so much the problem here
as the Thomas-Fermi density.
For outer half of the core, there is a surprising deviation from the TF
limit as the system is scaled to large $Z$.
This effect grows with $Z$ in a consistent way
even as the size of quantum oscillations decreases,
indicating that the potential in the Lieb-Simon limit diverges from the
TF potential.
This process seems not to be an artifact of numerical methods, and
since it involves half of the energy shells, cannot depend much
on minor errors in the Aufbau principle for determining the
order of occupying the outermost shells.
The fact that the trend occurs over nearly half of the shells
suggests a clue as to the origin of this effect.
Within the Aufbau principle, the inner core is composed of energy
shells that are completely filled; that is, all possible angular
momentum subshells of a given quantum number are filled.
The outer core shells are incrementally less complete, and the shells
become dominated by orbitals with large numbers of radial nodes.
The system structurally slowly trends from a fully three-dimensional,
homogeneous gas, and towards something like the radial variant of a
one-dimensional gas.
In fact, if we associate each oscillation in the Pauli potential
with a separate energy shell, the observed discrepancy is linear in the
number of unoccupied angular momentum subshells for that energy. This is
zero for the first roughly half of the energy shells of a noble gas atom,
and increases linearly with each shell beyond that -- exactly the behavior
observed here.
This deviation also raises an interesting issue -- does the Lieb-Simon limit
for the total energy have a local equivalent for the potential,
and if so where? For the large-$Z$ atom, it is common knowledge that
the density diverges from the TF limit at
the nucleus and asymptotically. Our work indicates that there is
a finite but not ignorable deviation of the Pauli potential
for a finite fraction of electrons.
The final, evanescent region is poorly described by all GE approximations.
The response function for normal DFT models has a dependence upon the
difference between the two highest energy eigenvalues that leads to an
extreme range of asymptotic behaviors for the Pauli potential.
The GE, is, at best, roughly comparable to the behavior
of atoms with a $p$-shell valence and fails to capture the
asymptotic trend of any other system.
There are a number of potential avenues along which to take this work
further.
One obvious track is to model the various deviations of the Pauli
potential from the GE prediction in large-$Z$ atoms.
A good question would be how to
implement the near-nucleus constraint defined in Sec.~\ref{sec:nearnucleus}.
This does not seem possible in a standard semilocal or one-point model
for the Pauli kinetic energy, at least using only the local density,
gradient and Laplacian.
It may be possible however to construct an accurate
correction that explicitly depends on the nuclear charge $Z$ in the spirit
of Ref.~\onlinecite{acharya}.
Such a correction could not be
easily made self-consistent, but might be worth the loss of self-consistency
to produce a physically reasonable potential in this region.
Secondly, deriving a GGA model for the
deviation of the Pauli potential from the TF limit in the outer core
would be of interest, since this region is more directly involved in bonding.
Preliminary calculations indicate that including fourth-order gradient
expansion terms or simple generalized gradient approximations fail to
reproduce the anomalous trend in the potential found for the outer half
of the core shells.
Such models can improve the potential in the outermost shell of the atom,
which may be useful for improving binding, but do not capture
the physics of the atom as a whole.
The most intractable region to model seems to be the evanescent region, since
the asymptotic behavior of the response potential depends on the eigenvalue
spectrum.
But this sensitivity is partly an artifact of the
character of DFT orbitals -- the asymptotic behavior of
HF and higher-rung DFT orbitals depends upon the HOMO eigenvalue only, which
may simplify the task for OFDFT considerably.
Finally, it would be of considerable interest to extend this study to
relativistic systems.~\cite{Oganesson}
We have studied nonrelativistic atoms because of the
well-known large-$Z$ limiting behavior for this case.
For relativistic systems, spin-orbit coupling spreads out eigenvalues
with the same principal quantum number and grows more important
as $Z$ increases.
For the largest-$Z$ atoms
fabricated in the lab, such as Oganesson, it is believed that this effect
kills shell structure altogether in the outer core.~\cite{Jerabek2018}
As this leads to a more homogeneous density, one might expect
less deviation of the Pauli potential from the the Thomas-Fermi limit
in this region than what we report here.
In contrast, we
expect that the basic physics underlying the value
of the Pauli potential near the nucleus would be unaltered by
relativistic corrections. The Pauli potential should still be given
by the difference in energy between highest and lowest occupied eigenvalues,
although of course these would be very different in value from the
nonrelativistic case.
\section*{Supplementary Material}
See supplementary material for tables and plots of kinetic energies using
various models discussed in the paper;
fits for the outer core of Pauli and response potentials and additional $p$ versus $q$
parametric plots. Additional data is available on request.
\begin{acknowledgments}
\end{acknowledgments}
The author would like to thank Sam Trickey and Kieron Burke for helpful
discussions, Eberhard Engel for use of his atomic DFT code, OPMKS, and
Thomas Baker for help with the intricacies of gnuplot.
|
\section{Introduction}
Haze is a common atmospheric phenomenon that not only has a serious adverse effect on human visual perception, but also has a serious impact on the performance of modern computer vision systems for various visual tasks, such as image classification, object detection and video surveillance. Since hazy conditions like fog, aerosols, sands and mists can scatter and adsorb light, images taken in the hazy environment suffer from obscured visibility, reduced contrast, color cast and many other degradations. Therefore, it is crucial to develop effective solutions to reduce the impact of image distortion caused by environmental conditions through dehazing.
\begin{figure}[tbp]
\centering
\renewcommand{\tabcolsep}{1pt}
\renewcommand{\arraystretch}{1}
\begin{center}
\begin{tabular}{ccc}
\includegraphics[width=0.32\linewidth]{Figure/Intro/BJ_Baidu_606_hazy.jpeg} &
\includegraphics[width=0.32\linewidth]{Figure/Intro/BJ_Baidu_606_tnet.jpg} &
\includegraphics[width=0.32\linewidth]{Figure/Intro/BJ_Baidu_606_stack.jpg} \\
\includegraphics[width=0.32\linewidth]{Figure/Intro/MSK_Bing_129_hazy.jpeg} &
\includegraphics[width=0.32\linewidth]{Figure/Intro/MSK_Bing_129_tnet.jpg} &
\includegraphics[width=0.32\linewidth]{Figure/Intro/MSK_Bing_129_stack.jpg} \\
\includegraphics[width=0.32\linewidth]{Figure/Intro/MLS_Google_197_hazy.jpeg} &
\includegraphics[width=0.32\linewidth]{Figure/Intro/MLS_Google_197_tnet.jpg} &
\includegraphics[width=0.32\linewidth]{Figure/Intro/MLS_Google_197_stack.jpg} \\
(a) Hazy image &
(b) T-Net &
(c) Stack T-Net\\
\end{tabular}
\end{center}
\caption{Dehazed results on real-world hazy images. We show the results of T-Net and Stack T-Net, which both perform well on these images. Compared with T-Net, the results of Stack T-Net are cleaner and brighter and exhibit finer details.}
\label{fig: intro}
\end{figure}
In the past ten years, single image dehazing has attracted widespread attention in the computer vision community. Its goal is to restore clean scenes from hazy images. According to the atmosphere scattering model~\cite{1977Optics,2002Vision,2016Haze}, the hazy process could be approximated as,
\begin{equation}
I(x)=J(x)t(x)+A(x)(1-t(x)),
\end{equation}
where $I(x)$ and $J(x)$ separately denote a hazy image and its clean scene, $A(x)$ represents the global atmospheric light, $t(x)$ describes the transmission map and $x$ is the pixel location. And the transmission map $t(x)$ is related to the depth $d(x)$ of the image, represented as $t(x)=e^{-\beta{d(x)}}$, where $\beta$ is the atmospheric scattering coefficient and controls the scale of the haze. Many methods based on this physical model restore clean scene $J(x)$ from a hazy image $I(x)$ by estimating the global atmospheric light and the transmission map of the hazy image.
Early prior-based methods attempt to estimate the transmission map by using the difference in statistical properties of hazy and clear images, such as contrast prior~\cite{4587643}, dark channel prior (DCP)~\cite{5567108}, color-lines~\cite{2651362}, color attenuation prior~\cite{7128396}, haze-lines~\cite{Berman_2016_CVPR}, and color ellipsoidal prior~\cite{8101508}. However, these hypothetical priors are not suitable for all of the real-world images, and it is easy to obtain inaccurate approximations of the transmission map, which will in turn cause the quality degradation of the restored image. To solve this problem, driven by the unprecedented success in recent years of deep learning in low level vision tasks~\cite{2016Image,2016Deeply,Ledig_2017_CVPR,2017Image,2016Beyond}, Convolution Neural Networks (CNNs) have been adopted for the task of image dehazing.
Earlier deep learning based methods, including the DehazeNet~\cite{2016DehazeNet}, the multi-scale CNN (MSCNN)~\cite{2016Single}, and the residue learning technique~\cite{8451663}, use deep networks to estimate the transmission map and exploit the conventional DCP's~\cite{5567108} bright pixels method to estimate the atmospheric light. However, since the estimation of the atmospheric light is coarse, the results of these methods are not sufficiently satisfactory. Thereby some methods try to use deep networks to estimate the transmission map and atmospheric light, simultaneously or respectively. For example, Li et al.~\cite{8237773} propose a network named AOD-Net to estimate a new variable which combines the transmission map with the atmospheric light. And Zhang et al.~\cite{8578435} use a two-stream network to estimate these two variables respectively.
In addition to the methods based on the atmospheric scattering model, there are some works that use end-to-end deep neural networks to directly estimate clean images instead of performing explicit estimation of the transmission map and atmospheric light. Considering that the estimation of the transmission map and atmospheric light sometimes deviates from the real hazy image, directly estimating the clean image is able to avoid sub-optimal restoration. Inspired by the successes of the generative adversarial networks (GANs) in synthesizing realistic images, many methods use GANs as the framework to transform hazy images to clean images. For instance, Qu et al.~\cite{8953692} propose a GAN based on pix2pix to realize image-to-image translation for dehazing.
Image restoration methods based on deep learning are driven by data, requiring large-scale synthetic or real-world datasets for training. These datasets are created by using the atmosphere scattering model to simulate the process of image degradation. Although proper guidance by the transmission map and atmospheric light is beneficial for effective dehazing under a wide range of haze density and for dealing with color distortions caused by haze, the algorithms based on the estimation of the transmission map and atmospheric light also rely on the same physical model. These algorithms may suffer from inherent performance loss on real-world images due to the over-fitting of the algorithms on synthetic datasets. In comparison, the methods of direct mapping exhibit higher robustness on real-world hazy images.
For this reason, we design an end-to-end network named T-Net for single image dehazing, which directly predicts clean images without using the atmospheric scattering model. Our T-Net is a symmetrical T-shaped architecture which is composed of two components, a backbone module and a dual attention module. The backbone module is a network based on the U-Net~\cite{2015UNet} architecture, in which residual dense block, upsampling block and downsampling block are used as basic blocks. In order to obtain distinctive features conducive to dehazing, a dual attention module~\cite{Fu_2019_CVPR} is embedded in the backbone module, which can also enhance the robustness of the network. Meanwhile, in view of the effectiveness of multi-scale features for image restoration~\cite{2016Single,7327182,Yang_2017_CVPR,Zhang_2018_CVPR,8575288,Liu_2019_ICCV,9237759}, we use skip connections with a new fusion strategy~\cite{NIPS2017_3f5ee243} in T-Net to realize adaptive multi-scale feature fusion. Moreover, inspired by the recurrent structure of~\cite{Yang_2017_CVPR} and~\cite{Ren_2019_CVPR}, we propose Stack T-Net by repeatedly unfolding our T-Net to further improve the performance of dehazing. Each stage of the Stack T-Net uses the output of the previous stage and the original hazy image as input and outputs clean image. Fig.~\ref{fig: intro} shows the dehazed results of our proposed T-Net and Stack T-Net on exemplar real-world images.
To summarize, our work has three-fold contributions as follows.
1) T-Net, a novel network with a dual attention module based on the U-Net architecture, is proposed to realize efficient information exchange across the multi-scale features from different levels to directly predict clean images.
2) Stack T-Net, which is created by repeatedly unfolding T-Net, can further improve the performance of dehazing by taking both the stage-wise result and the original rainy image as input to each stage.
3) Extensive experiments show that our plain T-Net and Stack T-Net perform favorably against the state-of-the-art methods on both synthetic and real-world hazy images. Meanwhile, an ablation study is conducted to demonstrate the effects of different modules in the proposed network.
The rest of the paper is organized as follows. Section~\ref{sec: rel} discusses the related work of single image dehazing. Section~\ref{sec: tnet} describes the structure of the plain network T-Net, including a backbone module and a dual attention module. Section~\ref{sec: stack} explains the details of our overall dehazing network Stack T-Net and defines the loss function for training. Section~\ref{sec: exp} introduces the datasets used in the experiments, and presents the results of the ablation study and the quantitative and qualitative comparisons between our approach and the state-of-the-art methods on synthetic and real-world images. Section~\ref{sec: concl} concludes the paper and discusses the direction of the future work.
\section{Related Work}
\label{sec: rel}
In this section, we briefly discuss the single image dehazing methods, which could be roughly divided into two categories, prior-based methods and learning-based methods as mentioned above.
\subsection{Prior based methods}
Single image dehazing is a highly ill-posed problem in computer vision, and it is difficult to dehaze images without additional information. To address this challenge, different priors or assumptions obtained through observations and statistics on a large amount of real data have been used in many dehazing methods. Except the earliest methods which are based on image enhancement algorithms~\cite{557356,663733,736004,2791513}, most prior-based dehazing methods are proposed based on the atmosphere scattering model, obtaining clean images by estimating the transmission map and the atmospheric light to invert Eq.(1). Representative works following this route include~\cite{1360671,4587643,5567108,2651362,7128396,Berman_2016_CVPR,8101508}, etc.
Fattal et al.~\cite{1360671} propose a dehazing technique by estimating the albedo under the assumption that the transmission map and surface shading are locally uncorrelated. Tan et al.~\cite{4587643} propose a local contrast-maximization method based on Markov Random Field (MRF), due to the observation that the contrast of clear images is higher than the contrast of the corresponding hazy images. Observing that the intensity of at least one color channel in local regions of natural haze-free images is close to zero and that the pixel intensity increases as haze increases, He et al.~\cite{5567108} propose the dark channel prior (DCP) as the approximation of haze distribution to estimate the transmission map. Meanwhile, this work also proposes a popular way to estimate the atmosphere light by averaging the top 0.1$\%$ of brightness pixels in hazy images. Many subsequent works make improvement on the DCP to refine the estimation of the transmission map by using different edge-preserving smoothing filters~\cite{8030116,6957555,7279117,6319316}. Discovering that haze causes the color-lines~\cite{1315267} to deviate from the origin, Fattal et al.~\cite{2651362} propose a new method to recover the transmission map, where color-lines are proposed by observing that pixels of small image patches typically exhibit an one-dimensional distribution. The color attenuation prior adopted in the linear model of~\cite{7128396} is based on the assumption that, as haze increases, the brightness of images increases but saturation decreases. With the assumption that several hundreds of distinct colors can well represent colors of a haze-free image, Berman et al.~\cite{Berman_2016_CVPR} observe that these colors in hazy images form clusters along lines in the RGB space and these lines pass through the coordinate value corresponding to the atmospheric light, where these lines are named haze-lines. Based on this observation, they proposed a method to estimate the transmission map by calculating the haze-lines with the predicted atmospheric light which is computed with the brightness pixels method of DCP. Biu and Kim~\cite{8101508} construct color ellipsoids by statistically fitting haze pixel clusters in the RGB space and then calculate a prior vector through color ellipsoid geometry to obtain the transmission map. In spite of different degrees of success which have been obtained, the prior based methods are limited by the hypothetical priors themselves. There is a certain gap between these priors and the reality, which can cause that the performance of these dehazing methods is not sufficiently satisfactory.
\subsection{Deep learning based methods}
Recently, deep learning achieves significant success in low-level vision tasks such as image super-resolution \cite{ledig2017photo,johnson2016perceptual,niu2021blind}, deblurring \cite{zhang2020deblurring,zhang2018adversarial}, deraining \cite{zhang2020beyond,zhang2021beyond}, desnowing \cite{liu2018desnownet,zhang2021deep}, which also include dehazing \cite{2016DehazeNet,2016Single}.
At present, there are two kinds of main ideas about the dehazing methods based on deep learning. One is to estimate the transmission map and atmospheric light according to the atmospheric scattering model and the other uses deep learning network to directly predict clean images.
The deep learning dehazing methods according to the physical model use the same strategy as the prior based methods to restore clean images, but generally estimate the transmission map and atmospheric light by using specially designed CNNs instead of priors so as to avoid the limitation on performance caused by the gap between hand-crafted priors and the reality. Cai et al.~\cite{2016DehazeNet} propose a dehazing model, DehazeNet, to estimate the transmission map. Ren et al.~\cite{2016Single} design a Multi-Scale CNN (MSCNN) to estimate the transmission map with a coarse-to-fine strategy. Zhang et al.~\cite{8578435} employ a densely connected pyramid dehazing network (DPCN) which is a two-stream network to predict the transmission map and the atmospheric light, respectively. Li et al.~\cite{8237773} create a reformulation of the atmosphere scattering model by using a new variable to integrate the transmission map and the atmospheric light and design an end-to-end neural network named AOD-Net to estimate this variable. Zhang et al.~\cite{8753731} introduce Famed-net to estimate the same variable of~\cite{8237773}. Dudhane et al.~\cite{8802288} propose PYF-Net, which consists of a YNet for the estimation of the transmission map in the RGB and YCbCr space and a FNet to fuse two transmission maps. Recently, there are some works that combine the traditional method DCP with deep learning. For example, Golts et al.~\cite{8897130} design a new loss based on the DCP, which is used for the training of an unsupervised deep network to estimate the transmission map. Chen et al.~\cite{8954465} exploit a Patch Map Selection Net to adaptively set the patch size of dark channel corresponding to each pixel.
The deep learning dehazing methods of direct mapping regard dehazing as an image-to-image translation problem. In recent years, generative adversarial networks (GANs) have shown great success in image generation and translation~\cite{Isola_2017_CVPR}~\cite{Zhu_2017_ICCV}~\cite{Ledig_2017_CVPR}, and are subsequently used in the field of dehazing. Qu et al.~\cite{8953692} propose a pix2pix GAN with two enhancing blocks to predict the clean images and enhance the detail and color. Raj et al.~\cite{9198400} design a condition GAN based on the U-Net architecture for dehazing. Engin et al.~\cite{8575279} design an end-to-end dehazing network based on CycleGAN with no need for paired hazy and corresponding clean images for training. Du et al.~\cite{Du2018PerceptuallyOG} design a GAN with an adaptive loss to facilitate end-to-end perceptual optimization and propose a new post processing method for halo artifacts removal using guide filters. Shao et al.~\cite{Shao_2020_CVPR} apply an end-to-end network which is made up of two sub-networks, a bidirectional translation network based on CycleGAN to bridge the gap between the synthetic and real domains and a dehazing network to restore clean images from the hazy images before and after translation. In addition to GAN, some works are proposed based on other architectures. Ren et al.~\cite{Ren_2018_CVPR} introduce an end-to-end network based on the encoder-decoder architecture, which adopts a novel fusion-based strategy that derives three inputs by using three pre-processing methods. Liu et al.~\cite{Liu_2019_ICCV} propose a grid network GridDehazeNet for single image dehazing, which consists of three modules, pre-processing, backbone, and post-processing. Li et al.~\cite{Li_2019_ICCV} introduce a level-aware progressive network (LAP-Net), of which each stage learns different levels of haze with different supervision and the final output is yielded with an adaptive integration strategy. Dong et al.~\cite{Dong_2020_CVPR} propose a Multi-Scale Boosted Dehazing Network (MSBDN) based on the U-Net architecture with two principles, boosting and error feedback to realize dense feature fusion.
\section{T-Net}
\label{sec: tnet}
In this section, we describe the proposed T-Net which is a symmetrical T-shaped network consisting of two sub-modules, a backbone module based on the U-Net architecture and a dual attention module. Fig.~\ref{fig: T-Net} illustrates the architecture of T-Net, and the details are given in the following.
\begin{figure}[t]
\centering
\renewcommand{\tabcolsep}{1pt}
\renewcommand{\arraystretch}{1}
\includegraphics[width=8.5CM]{Figure/Net/T-Net.png}
\caption{The architecture of T-Net. It is an end-to-end network to directly obtain the dehazed result, consisting of a backbone module and a dual attention module (dark yellow rectangular block). The red path is the trunk road of the network, and the gray paths are skip connections with weights.}
\label{fig: T-Net}
\end{figure}
\subsection{Backbone module}
The backbone module of T-Net is based on the U-Net architecture, and we adopt a new fusion strategy to make use of the skip connections. As shown in Fig.~\ref{fig: T-Net}, the network mainly includes three kinds of basic blocks, residual dense block (RDB)~\cite{Zhang1_2018_CVPR}, upsampling block and downsampling block. And the detailed structure of these three kinds of blocks is shown in Fig.~\ref{fig: Blocks}.
\begin{figure}[ht]
\centering
\renewcommand{\tabcolsep}{1pt}
\renewcommand{\arraystretch}{1}
\includegraphics[width=8.5cm]{Figure/Net/Blocks.png}
\caption{The detailed structure of three kinds of blocks in the backbone module of the T-Net. The yellow, blue and green doted boxes separately represent the RDB block, the downsampling block and the upsampling block, where we illustrate the variation of the feature size and the number of channels in each layer of each kind of block.}
\label{fig: Blocks}
\end{figure}
Extensive research has demonstrated that the use of multi-scale features is beneficial to various tasks of image understanding. The high-level features are of the downsampled spatial resolution but compress more semantic contextual information that is necessary for scene understanding of images. In contrast, the low-level features are of higher resolution to help localize objects but contain less semantic contextual information. Thereby the feature fusion of different levels can simultaneously preserve spatial information from low-level features and exploit the semantic contextual information from high-level features for image understanding.
Actually, image restoration includes the process of image understanding to determine what should be kept on the images and what should be removed. In order to make full use of the information from different levels of features, multi-scale feature fusion has been applied to many image restoration tasks such as image derainning~\cite{Yang_2017_CVPR,Zhang_2018_CVPR}, image debluring~\cite{Kupyn_2019_ICCV,Nah_2017_CVPR} and image dehazing~\cite{8575288,Liu_2019_ICCV,Dong_2020_CVPR}, which significantly improves the performance of the algorithms. The U-Net~\cite{2015UNet} architecture is originally proposed for semantic segmentation, consisting of a contracting path to capture contextual cues and a symmetric expanding path for precise localization as well as multiple lateral connections between these contracting paths and their symmetric expanding paths for multi-scale feature fusion. We design our T-Net based on this architecture and make several improvements for the need of dehazing.
As shown in Fig.~\ref{fig: T-Net}, the backbone is a symmetric network, which mainly includes three pairs of RDB blocks and four pairs of upsampling and downsampling blocks in the trunk road (the red path in Fig.~\ref{fig: T-Net}), and three RDB blocks in the lateral connections. In addition, there is a convolutional layer without activation function at the beginning of the network, which generates 16 linear feature maps as the learned input from hazy image. And there is another convolutional layer at the end, which is symmetrical to the beginning and used to generate high-quality dehazed images.
The RDB~\cite{Zhang1_2018_CVPR} block which keeps the number of feature maps unchanged can extract abundant local features via densely connected convolutional layers, so we choose it as the basic block for feature generation. As shown in Fig.~\ref{fig: Blocks}, we use five convolutional layers in each RDB and set the growth rate to 16. Every convolutional layer takes the concatenation of the output features of all convolutional layers before it as input, and the final output of each RDB block is the combination of the output of the last layer and the input of this RDB block through channel-wise addition, where the first four convolutional layers are used to extract features and the last layer (kernel size = 1, 1x1 convolution) is used to keep the number of channels of the output feature the same as that of the input feature. In order to reduce information loss, we use convolutional layers to realize upsampling and downsampling, and the detailed structure is shown in Fig.~\ref{fig: Blocks}. The upsampling block and the downsampling block both consist of two convolutional layers of which the first layer adjusts the size of feature maps by setting different kernel sizes of convolutional layers and the other layer uses 1x1 convolution to change the number of channels. In each upsampling block, the number of feature maps decreases by half as the size of feature maps increases by two times, which is the reverse in each downsampling block.
We use a new fusion strategy to realize skip connections, for the consideration that features from different scales may not be equally important. Motivated by~\cite{NIPS2017_3f5ee243}, we set two trainable fusion parameters for each skip connection, where every fusion parameter is a $n$-dimensional vector ($n$ is the number of channels of the features before fusion). Moreover, RDB blocks instead of 1x1 convolution are used in the lateral connection to obtain more feature combinations, which improves the possibility of obtaining more effective information for dehazing. The backbone includes four pairs of upsampling and downsampling blocks, so the feature from feature fusion could be expressed as
\begin{equation}
\begin{split}
\breve{F}_i^j=\alpha &f_r(F_i^j)+\beta f_u(f_x(F_{i+1}^j)),\\
&F_{i+1}^j=f_d(F_i^j),\\
f_x(\cdot)&=
\begin{cases}
f_r(\cdot),\quad i < m-1,\\
f_a(\cdot),\quad i = m-1.
\end{cases}\\
i=0,1,&\cdots,m-1;\quad j=0,1,\cdots,n,
\end{split}
\end{equation}
where $\alpha$ and $\beta$ are the fusion parameters for the two features to be fused, $f_r(\cdot)$, $f_u(\cdot)$, $f_d(\cdot)$ and $f_a(\cdot)$ separately stand for the functions of RDB, upsampling, downsampling and dual attention, $F^j_i$ represents the $j$-th feature channel after the $i$-th downsampling, $\breve{F}_i^j$ represents the fused feature symmetric with $F^j_i$, $m$ is the number of the upsampling and downsampling blocks pairs, and $n$ is the number of channels of the features.
\subsection{Dual attention module}
As shown in Fig.~\ref{fig: T-Net}, in addition to the backbone module, we use a dual attention module in the T-Net, which is embedded in the middle of the backbone module and includes two blocks, the position attention block and the channel attention block. Fig.~\ref{fig: Dual} is an overview of the dual attention module.
\begin{figure}[ht]
\centering
\renewcommand{\tabcolsep}{1pt}
\renewcommand{\arraystretch}{1}
\includegraphics[width=8.5cm]{Figure/Net/Dual.png}
\caption{An overview of the dual attention module. The final output of the dual attention module is the sum of the outputs of the position attention block and the channel attention block, and has the same feature size and number of channels as the input.}
\label{fig: Dual}
\end{figure}
Discriminant feature representations are the key for scene understanding, which could be obtained by capturing long-range contextual information. The dual attention module is originally proposed in~\cite{Fu_2019_CVPR} for scene segmentation, which could adaptively integrate local features with their global dependencies. The position attention block encodes a wider range of contextual information into local features by a weighted sum of the features at all positions, thus enhancing their representation capability. The channel attention block emphasizes interdependent feature channels by integrating associated features among all channels, to further improve the feature representation of specific semantics. These two attention blocks could be simply regarded as two different dimension feature enhancers, and expand the range of semantic contextual information from two different dimensions.
As shown in Fig.~\ref{fig: Dual}, the position attention block generates feature maps of spatial long-range contextual through three steps. Firstly, generating a spatial attention matrix by a softmaxed product of two new features of which one is obtained by reshaping the input feature to $\mathbb{R}^{C\times (H\times W)}$, and the other is the transposition of first new feature, where $(H\times W)$ is the number of pixels. Secondly, getting an attention increment by performing a matrix multiplication between the attention matrix and the second new feature. Thirdly, obtaining the final representations reflecting long-range contexts through an element-wise sum of the attention increment and the original input feature. The channel attention block captures the channel relationship through similar steps, except for the first step, in which channel attention matrix is calculated in channel dimension instead of position dimension.
As explained above, it is obvious that these two attention blocks focus on the content similarity of features, that is, the interdependence between features. They can adaptively emphasize the effective information in the features under the guidance of the loss function and obtain discriminant feature representations according to different input images, without being limited by the network structure itself. Thereby, we believe combining the U-Net architecture with the dual attention module can further improve the robustness of the network, and the discriminant features obtained by the dual attention module can guide the parameter learning of the network. As shown in Fig.~\ref{fig: T-Net}, we embed the dual attention module in the middle of the last pair of upsampling and downsampling blocks to extract discriminant feature representations from high-level semantic features.
\section{Stack T-Net}
\label{sec: stack}
\begin{figure*}
\centering
\renewcommand{\tabcolsep}{1pt}
\renewcommand{\arraystretch}{1}
\includegraphics[width=18cm]{Figure/Net/Stack.png}
\caption{The architecture of Stack T-Net. (a) The detailed structure. (b)The information flow and the feature resolution change among different stages. We use T-Net as the sub-network to build Stack T-Net by repeatedly unfolding it several times. Each dotted line of (b) represents different resolution, and the resolution of each dotted line is 1/4 of it of the above one. Since we use T-Net with four pairs of upsampling and downsampling blocks as sub-network, there are five dotted lines in (b) (containing the original resolution).}
\label{fig: Stack}
\end{figure*}
In this section, we introduce the overall dehazing network Stack T-Net which is constructed by repeatedly unfolding the plain T-Net, and the loss function used in our method. The first subsection below explains the details of the architecture of Stack T-Net, and Fig.~\ref{fig: Stack} is the illustration a $K$-stage Stack T-Net, where sub-figure (a) exhibits the detailed structure and sub-figure (b) shows the information flow and the feature resolution change among different stages. The second subsection introduces the loss function we used for training the full network.
\subsection{The details of the architecture}
Inspired by~\cite{Yang_2017_CVPR} and~\cite{Ren_2019_CVPR}, we introduce the recurrent structure to our method. In~\cite{Yang_2017_CVPR}, Yang et al. use recurrent network for deraining but just use the deraining result of the previous stage as the input of the next stage, while the performance of deraining is limited by the deraining result of each stage. So we use the strategy of~\cite{Ren_2019_CVPR}, using the concatenation of the output of the previous stage and the original hazy image as input, where the original hazy image can supplement the information loss caused by each dehazing stage. As shown in Fig.~\ref{fig: Stack}, we use T-Net with the same structure in each stage. The inference of Stack T-Net at the $k$-th stage can be formulated as
\begin{equation}
\begin{split}
x^k&=f_{in}(x^0,y^{k-1}),\\
&y^k=f^k_T(x^k),
\end{split}
\end{equation}
where $x^k$ and $y^k$ separately represent the input and the output of the $k$-th stage, $x^0$ is the original hazy image and $y^0=x^0$, $f_{in}$ stands for the operation of concatenation, $f_T^k$ denotes the mapping of T-Net at the $k$-th stage, and meanwhile $k\in \left\{1,\cdots,K\right\}$. T-Net of each stage plays the role of a dehazing sub-network, which directly learns the mapping from hazy image to dehazy results. Thereby, the input of each stage is a six-channel image which is the concatenation of two three-channel RGB images, and the output is a three-channel RGB image. Meanwhile, in order to be consistent with other stages, the first stage uses the concatenation of two identical original hazy images as input. And we choose the output of the last stage as the final dehazing result.
However, the network of recurrent structure has considerably large amount of parameters to learn, which take a lot of memory and decrease the efficiency of dehazing. And the more complex the structure of network is, the more likely it leads to over-fitting. Witnessing the success of recursive computation in image restoration~\cite{Kim_2016_CVPR,Li_2018_ECCV,Tai_2017_CVPR}, we apply it to our method. In our practice, we utilize the inter-stage recursive computation instead of repeatedly unfolding the plain network. Since inter-stage recursive computation requires each stage to share the same network parameters, each stage can be expressed as
\begin{equation}
f_T^k(\cdot)=f_T(\cdot).
\end{equation}
Experimental results verify that this recursive strategy can improve the dehazing effectiveness without designing a deeper and more complex network.
\subsection{Loss function}
The loss function of our method measures the error of the dehazing results at each-stage, and is used to train the proposed network. The MSE loss has smooth function curve, which is convenient for the use of the gradient descent algorithm. But it is more sensitive to outliers and easier to cause gradient explosions than the $L_1$ loss. Therefore, we employ the smooth $L_1$ loss which combines the MSE loss and the $L_1$ loss to enhance the robustness of the network.
Let $I_c(x)$ denote the intensity of the $c$-th color channel at pixel $x$ in the ground truth, and $J_c(x)^k$ represents the intensity of the same position of the dehazed image in the $k$-th stage. $K$ and $N$ separately denote the number of the stages and the number of pixels. The smooth $L_1$ loss of the $k$-th stage ($L_{SL_1}^{k}$) can be defined as
\begin{equation}
L_{SL_1}^{k}=\frac{1}{N}\sum_{x=1}^{N}\sum_{c=1}^{3}f_{SL_1}(\left|J_c^k(x)-I_c(x)\right|),
\end{equation}
where
\begin{equation}
f_{SL_1}(e)=
\begin{cases}
0.5e^2,\quad 0\leq e<1,\\
e-0.5,\quad e\ge1.
\end{cases}
\end{equation}
When the error between the dehazed result and the ground truth exceeds the limited value, the $L_1$ norm is used instead of the $L_2$ norm to reduce the influence of outliers on the network parameters.
The perceptual loss~\cite{2016Perceptual} measures image visual similarities between the dehazed image and the ground truth more effectively than the pixel-wise loss, for which we take advantage of the perceptual loss to strengthen the finer details of the dehazed images. By leveraging multi-scale features extracted from VGG16~\cite{simonyan2015deep} pre-trained on ImageNet~\cite{2015ImageNet}, the perceptual loss of the $k$-th dehazed stage can be defined as
\begin{equation}
L_p^k=\sum_{j=1}^3\frac{1}{C_jH_jW_j}\left\|F_{j}(J^k)-F_j(I)\right\|_2^2,
\end{equation}
where $J^k$ represents the dehazed result of the $k$-th stage, $I$ represents the ground truth, each $F_{j}(\cdot)$ ($j =\left\{1,2,3\right\}$) separately denotes the output feature of relu1-2, relu2-2, relu3-3 layer of VGG-16 and $C_j$, $H_j$ and $W_j$ specify the dimension of these features.
By combining the smooth L1 loss and the perceptual loss, the total loss of all stages is defined as
\begin{equation}
L=\sum_{k=1}^KL_{SL_1}^k+\lambda \sum_{k=1}^{K}L_p^k,
\end{equation}
that is,
\begin{equation}
L=L_{SL_1}+\lambda L_P,
\end{equation}
where $\lambda$ is set to $0.04$ for controlling the relative weights on the two loss components.
\section{Experiments}
\label{sec: exp}
In this section, quantitative and qualitative experimental results are shown to demonstrate the effectiveness of the proposed method. We conduct a series of experiments to compare the performance of our method with the state-of-the-art approaches on both synthetic and real-world datasets. Moreover, we carry out an ablation study to demonstrate the effectiveness of each module of our network.
\subsection{Datasets}
In the training phase, we use the datasets of~\cite{Shao_2020_CVPR} as training set. This dataset contains $6000$ synthetic hazy images from the RESIDE dataset~\cite{8451944}, where the RESIDE dataset~\cite{8451944} contains both synthesized and real-world hazy/clean image pairs of indoor and outdoor scenes and is split into five subsets, namely, ITS (Indoor Training Set), OTS (Outdoor Training Set), SOTS (Synthetic Object Testing Set), URHI (Unannotated Real Hazy Images), and RTTS (Real Task-driven Testing Set). Among the $6000$ images, $3000$ are chosen from the ITS and the rest is from the OTS. All the images are randomly cropped to $256\times256$ and randomly flipped for data augmentation, and then the pixel values are normalized to $[-1, 1]$. Since our training set is a subset of ITS and OTS, we use SOTS as the total testing set in the test phase to compare the performance of our proposed method with other methods. Furthermore, real-world hazy images are chosen from URHI to evaluate the generalization of our method in the real-world scenery.
\subsection{Network implementation}
We implement our framework with Pytorch~\cite{Paszke2017AutomaticDI} and exploit the Adam optimizer~\cite{kingma2017adam} with a batch size $14$ to train the network, where the momentum $\beta_1$ and $\beta_2$ adopt the default values of $0.9$ and $0.999$, respectively. We train every model for $2000$ epochs in total. The learning rate is initially set to $0.001$, reduced by half every $20$ epochs, and kept fixed as $0.0001$ from the $80$-th epoch. Our proposed method is evaluated against the following state-of-the-art approaches: DCP~\cite{5567108}, MSCNN~\cite{2016Single}, DehazeNet~\cite{2016DehazeNet}, NLD~\cite{Berman_2016_CVPR}, AOD-Net~\cite{8237773}, GFN~\cite{Ren_2018_CVPR}, DCPDN~\cite{8578435}, EPDN~\cite{8953692}, and DA$\_$dehazing~\cite{Shao_2020_CVPR}. We adopt peak-signal-to-noise ratio (PSNR) and structural similarity measure (SSIM) as quantitative evaluation indexes, where PSNR can measure the pixel to pixel difference and SSIM can quantify the structural difference between a dehazed output image and the corresponding ground truth. The results of all experiments are shown in the following.
\subsection{Ablation study}
To evaluate the effectiveness of several key modules in our network, we perform ablation studies with the following three strategies.
\textbf{The first ablation study} is conducted to determine the configurations of the backbone module of the proposed T-Net, that is, the number of upsampling and downsampling blocks pairs and the number of RDB blocks pairs in the trunk road. RDB blocks are used as feature generator in our network. Simply put, the more RDB block pairs we use, the deeper the network and the deeper the features we can extract. The position attention block contains a matrix multiplication of $(N, C)\times(C, N)$, where $N$ is the number of the pixels in each channel of the feature. So it is easy to cause out of memory if the input feature size of this block is too large. We set the initial pairs of upsampling and downsampling blocks as 2 to avoid this problem. Table~\ref{tab: tabel1} shows the performance of T-Net with different configurations on the SOTS dataset, where $m$ and $n$ respectively represent the number of upsampling and downsampling block pairs and the number of RDB block pairs in the trunk road.
\newcolumntype{I}{!{\vrule width 0.8pt}}
\newcommand{\PreserveBackslash}[1]{\let\temp=\\#1\let\\=\temp}
\newcolumntype{C}[1]{>{\PreserveBackslash\centering}p{#1}}
\newcolumntype{R}[1]{>{\PreserveBackslash\raggedleft}p{#1}}
\newcolumntype{L}[1]{>{\PreserveBackslash\raggedright}p{#1}}
\renewcommand\arraystretch{1.25}
\begin{table}[tbp]
\centering
\caption{Comparison on the SOTS dataset for T-Net with different configurations.}
\begin{tabular}{Ic|c|c|cI}
\Xhline{0.8pt}
\multicolumn{2}{Ic|}{Configuration} & \multicolumn{2}{cI}{SOTS}\\
\Xhline{0.8pt}
\multicolumn{1}{IC{1.2cm}|}{$m$} &\multicolumn{1}{C{1.2cm}|}{$n$} & \multicolumn{1}{C{1.5cm}|}{PSNR} & \multicolumn{1}{C{1.5cm}I}{SSIM} \\
\Xhline{0.8pt}
\multirow{2}[3]{*}{2} & 1 &22.90 &0.8954 \\
\cline{2-4} & 2 &26.86 &0.9451 \\
\cline{2-4} & 3 &27.43 &0.9473 \\
\hline
\multirow{2}[3]{*}{3} & 1 &22.97 &0.8949 \\
\cline{2-4} & 2 &27.49 &0.9520 \\
\cline{2-4} & 3 &28.13 &0.9536 \\
\hline
\multirow{2}[3]{*}{4} & 1 &22.86 &0.8932\\
\cline{2-4} & 2 &28.30 &0.9535\\
\cline{2-4} & 3 &\textbf{28.55} &\textbf{0.9543}\\
\Xhline{0.8pt}
\end{tabular}%
\label{tab: tabel1}
\end{table}%
There is a tendency shown in Table~\ref{tab: tabel1} that the average PSNR and SSIM values increase as $m$ and $n$ increase. The exception happens when $n=1$, i.e., the performance of T-Net does not get improved with the increase of $m$ in this case. The reason is that, the representation ability of features is influenced by the number of RDB block pairs in the trunk road. When $n$ is set to 1, the features extracted by our network contains insufficient deep discriminant information, which makes the fitting ability of neural network badly limited. Moreover, when $n>1$, the performance of adding a pair of upsampling and downsampling blocks is better than adding a pair of RDB blocks (see e.g., (3,2) versus (2,3), (4,2) versus (3,3) in the form of (m,n)), which verifies the effectiveness of multi-scale features. Compared with single-scale features, multi-scale features can provide more discriminant information which is helpful for image understanding. As shown in Table~\ref{tab: tabel1}, the average PSNR and SSIM values are the highest when $m=4, n=3$.Therefore, we use a T-Net with four pairs of upsampling and downsampling blocks and three pairs of RDB blocks in the trunk road as our network in the following experiments.
\begin{figure*}[tbp]
\scriptsize
\centering
\renewcommand{\tabcolsep}{1pt}
\renewcommand{\arraystretch}{1}
\begin{center}
\begin{tabular}{cccccccccc}
\includegraphics[width=0.095\linewidth]{Figure/Syn/Hazy/0108_1_0.2_hazy.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/NLD/0108_1_0.2_nld.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DehazeNet/0108_1_0.2_dehaze.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/AOD-Net/0108_1_0.2_aod.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DCPDN/0108_1_0.2_dcpdn.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/EPDN/0108_1_0.2_epdn.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DA/0108_1_0.2_da.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/T-Net/0108_1_0.2_tnet.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/Stack T-Net/0108_1_0.2_stack.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/GT/0108.jpg} \\
\includegraphics[width=0.095\linewidth]{Figure/Syn/Hazy/0087_0.9_0.16_hazy.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/NLD/0087_0.9_0.16_nld.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DehazeNet/0087_0.9_0.16_dehaze.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/AOD-Net/0087_0.9_0.16_aod.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DCPDN/0087_0.9_0.16_dcpdn.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/EPDN/0087_0.9_0.16_epdn.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DA/0087_0.9_0.16_da.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/T-Net/0087_0.9_0.16_tnet.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/Stack T-Net/0087_0.9_0.16_stack.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/GT/0087.jpg} \\
\includegraphics[width=0.095\linewidth]{Figure/Syn/Hazy/0152_0.8_0.2_hazy.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/NLD/0152_0.8_0.2_nld.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DehazeNet/0152_0.8_0.2_dehaze.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/AOD-Net/0152_0.8_0.2_aod.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DCPDN/0152_0.8_0.2_dcpdn.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/EPDN/0152_0.8_0.2_epdn.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DA/0152_0.8_0.2_da.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/T-Net/0152_0.8_0.2_tnet.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/Stack T-Net/0152_0.8_0.2_stack.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/GT/0152.jpg} \\
\includegraphics[width=0.095\linewidth]{Figure/Syn/Hazy/0333_0.95_0.2_hazy.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/NLD/0333_0.95_0.2_nld.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DehazeNet/0333_0.95_0.2_dehaze.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/AOD-Net/0333_0.95_0.2_aod.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DCPDN/0333_0.95_0.2_dcpdn.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/EPDN/0333_0.95_0.2_epdn.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DA/0333_0.95_0.2_da.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/T-Net/0333_0.95_0.2_tnet.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/Stack T-Net/0333_0.95_0.2_stack.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/GT/0333.jpg} \\
\includegraphics[width=0.095\linewidth]{Figure/Syn/Hazy/1400_10_hazy.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/NLD/1400_10_nld.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DehazeNet/1400_10_dehaze.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/AOD-Net/1400_10_aod.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DCPDN/1400_10_dcpdn.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/EPDN/1400_10_epdn.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DA/1400_10_da.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/T-Net/1400_10_tnet.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/Stack T-Net/1400_10_stack.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/GT/1400.jpg} \\
\includegraphics[width=0.095\linewidth]{Figure/Syn/Hazy/1401_10_hazy.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/NLD/1401_10_nld.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DehazeNet/1401_10_dehaze.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/AOD-Net/1401_10_aod.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DCPDN/1401_10_dcpdn.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/EPDN/1401_10_epdn.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DA/1401_10_da.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/T-Net/1401_10_tnet.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/Stack T-Net/1401_10_stack.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/GT/1401.jpg} \\
\includegraphics[width=0.095\linewidth]{Figure/Syn/Hazy/1409_10_hazy.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/NLD/1409_10_nld.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DehazeNet/1409_10_dehaze.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/AOD-Net/1409_10_aod.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DCPDN/1409_10_dcpdn.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/EPDN/1409_10_epdn.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DA/1409_10_da.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/T-Net/1409_10_tnet.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/Stack T-Net/1409_10_stack.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/GT/1409.jpg} \\
\includegraphics[width=0.095\linewidth]{Figure/Syn/Hazy/1421_10_hazy.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/NLD/1421_10_nld.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DehazeNet/1421_10_dehaze.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/AOD-Net/1421_10_aod.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DCPDN/1421_10_dcpdn.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/EPDN/1421_10_epdn.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/DA/1421_10_da.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/T-Net/1421_10_tnet.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/Stack T-Net/1421_10_stack.jpg} &
\includegraphics[width=0.095\linewidth]{Figure/Syn/GT/1421.jpg} \\
(a) Hazy image &
(b) NLD &
(c) DehazeNet &
(d) AOD-Net &
(e) DCPDN &
(f) EPDN &
(g) DA\_dehaze &
(h) T-Net &
(i) Stack T-Net &
(j) GT \\
\end{tabular}
\end{center}
\caption{Visual comparisons on the SOTS dataset. We show the visual results of eight methods on the synthetic data including a kind of prior-based methods and seven kinds of deep-learning-based methods. The images of the first and the last columns represent the hazy images and their corresponding ground truths, respectively. The results of our proposed T-Net and Stack T-Net are separately shown in the eighth and ninth columns.}
\label{fig: SOTS}
\end{figure*}
\textbf{The second ablation study} is conducted to verify the effectiveness of each module by comparing the performance of several variants of the T-Net. We set the trunk road without the dual attention module (where we use a RDB block to replace the dual attention module) as the basic model. Starting from this basic model, we create other variants by gradually injecting our modifications which include usual skip connections realized by 1x1 convolutions, usual skip connections realized by RDB blocks, our skip connections, as well as our skip connections and dual attention module (T-Net). Table~\ref{tab: tabel2} shows the performance of each variant of T-Net on the SOTS dataset.
\begin{table}[tbp]
\centering
\caption{Comparison on the SOTS dataset of several variants of T-Net.}
\begin{tabular}{Ip{5.1cm}|c|cI}
\Xhline{0.8pt}
\multicolumn{1}{Ic|}{\multirow{1}[4]{*}{Variant}} & \multicolumn{2}{cI}{SOTS } \\
\cline{2-3} & \multicolumn{1}{c|}{PSNR} & \multicolumn{1}{cI}{SSIM} \\
\Xhline{0.8pt}
the basic model &24.19 &0.8228\\
\hline
with usual skip connections (1x1 conv) &27.42 &0.9504\\
\hline
with usual skip connections (RDB block) &27.61 &0.9520\\
\hline
with our skip connections &27.82 &0.9530\\
\hline
T-Net &\textbf{28.55} &\textbf{0.9543}\\
\Xhline{0.8pt}
\end{tabular}%
\label{tab: tabel2}
\end{table}%
It is clear that the performance is further improved every time a new component is added to the basic model, which justifies the overall design. As shown in Table~\ref{tab: tabel2}, the comparison among these four variants show the effectiveness of our skip connections with a new fusion strategy. It is demonstrated that RDB blocks used in the lateral connection instead of 1x1 convolution can extract more complex semantic information for dehazing, and that different scale features have different importances in feature fusion. Furthermore, the addition of the dual attention module has the greatest effect on the dehazing performance improvement, which effectively enhances the robustness and generalization ability of the network. This ablation study demonstrates the contribution of each component in our T-Net.
\textbf{The third ablation study} is conducted to evaluate our proposed Stack T-Net with different recursive stage number $K$ and demonstrates the effectiveness of our recursive strategy. Limited by physical memory, we just set $K = 1, 2, 3$. Table~\ref{tab: tabel3} shows the performance on the SOTS dataset of the Stack T-Net with different recursive stage numbers.
\begin{table}[tbp]
\centering
\caption{Comparison on the SOTS dataset of the Stack T-Net with different $K$ stages.}
\begin{tabular}{Ic|c|cI}
\Xhline{0.8pt}
\multirow{1}[4]{*}{Recursive stage number} & \multicolumn{2}{cI}{SOTS}\\
\cline{2-3} & \multicolumn{1}{C{1.5cm}|}{PSNR} & \multicolumn{1}{C{1.5cm}I}{SSIM}\\
\Xhline{0.8pt}
$K=1$ &28.55 &0.9543 \\
\hline
$K=2$ &28.71 &0.9556 \\
\hline
$K=3$ &\textbf{28.83} &\textbf{0.9551} \\
\Xhline{0.8pt}
\end{tabular}%
\label{tab: tabel3}
\end{table}%
\begin{figure*}[tbp]
\scriptsize
\centering
\renewcommand{\tabcolsep}{1pt}
\renewcommand{\arraystretch}{1}
\begin{center}
\begin{tabular}{ccccccccc}
\includegraphics[width=0.107\linewidth]{Figure/Real/Hazy/HEB_Bing_016_hazy.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/NLD/HEB_Bing_016_nld.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DehazeNet/HEB_Bing_016_dehaze.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/AOD-Net/HEB_Bing_016_aod.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DCPDN/HEB_Bing_016_dcpdn.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/EPDN/HEB_Bing_016_epdn.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DA/HEB_Bing_016_da.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/T-Net/HEB_Bing_016_tnet.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/Stack T-Net/HEB_Bing_016_stack.jpg} \\
\includegraphics[width=0.107\linewidth]{Figure/Real/Hazy/HF_Bing_639_hazy.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/NLD/HF_Bing_639_nld.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DehazeNet/HF_Bing_639_dehaze.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/AOD-Net/HF_Bing_639_aod.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DCPDN/HF_Bing_639_dcpdn.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/EPDN/HF_Bing_639_epdn.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DA/HF_Bing_639_da.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/T-Net/HF_Bing_639_tnet.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/Stack T-Net/HF_Bing_639_stack.jpg} \\
\includegraphics[width=0.107\linewidth]{Figure/Real/Hazy/XG_Bing_030_hazy.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/NLD/XG_Bing_030_nld.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DehazeNet/XG_Bing_030_dehaze.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/AOD-Net/XG_Bing_030_aod.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DCPDN/XG_Bing_030_dcpdn.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/EPDN/XG_Bing_030_epdn.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DA/XG_Bing_030_da.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/T-Net/XG_Bing_030_tnet.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/Stack T-Net/XG_Bing_030_stack.jpg} \\
\includegraphics[width=0.107\linewidth]{Figure/Real/Hazy/SFC_Google_493_hazy.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/NLD/SFC_Google_493_nld.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DehazeNet/SFC_Google_493_dehaze.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/AOD-Net/SFC_Google_493_aod.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DCPDN/SFC_Google_493_dcpdn.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/EPDN/SFC_Google_493_epdn.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DA/SFC_Google_493_da.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/T-Net/SFC_Google_493_tnet.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/Stack T-Net/SFC_Google_493_stack.jpg} \\
\includegraphics[width=0.107\linewidth]{Figure/Real/Hazy/TJ_Baidu_518_hazy.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/NLD/TJ_Baidu_518_nld.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DehazeNet/TJ_Baidu_518_dehaze.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/AOD-Net/TJ_Baidu_518_aod.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DCPDN/TJ_Baidu_518_dcpdn.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/EPDN/TJ_Baidu_518_epdn.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DA/TJ_Baidu_518_da.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/T-Net/TJ_Baidu_518_tnet.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/Stack T-Net/TJ_Baidu_518_stack.jpg} \\
\includegraphics[width=0.107\linewidth]{Figure/Real/Hazy/IRQ_Google_040_hazy.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/NLD/IRQ_Google_040_nld.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DehazeNet/IRQ_Google_040_dehaze.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/AOD-Net/IRQ_Google_040_aod.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DCPDN/IRQ_Google_040_dcpdn.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/EPDN/IRQ_Google_040_epdn.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DA/IRQ_Google_040_da.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/T-Net/IRQ_Google_040_tnet.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/Stack T-Net/IRQ_Google_040_stack.jpg} \\
\includegraphics[width=0.107\linewidth]{Figure/Real/Hazy/HEB_Bing_613_hazy.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/NLD/HEB_Bing_613_nld.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DehazeNet/HEB_Bing_613_dehaze.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/AOD-Net/HEB_Bing_613_aod.jpeg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DCPDN/HEB_Bing_613_dcpdn.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/EPDN/HEB_Bing_613_epdn.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/DA/HEB_Bing_613_da.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/T-Net/HEB_Bing_613_tnet.jpg} &
\includegraphics[width=0.107\linewidth]{Figure/Real/Stack T-Net/HEB_Bing_613_stack.jpg} \\
(a) Hazy image &
(b) NLD &
(c) DehazeNet &
(d) AOD-Net &
(e) DCPDN &
(f) EPDN &
(g) DA\_dehaze &
(h) T-Net &
(i) Stack T-Net \\
\end{tabular}
\end{center}
\caption{Visual comparisons on the URHI dataset. We show the visual results of different methods on the real-world data. The first column shows the hazy images, and other columns represent the dehazed results of different methods. The results of our proposed T-Net and Stack T-Net are separately shown in the last two columns.}
\label{fig: URHI}
\end{figure*}
As shown in Table~\ref{tab: tabel3}, Stack T-Net usually achieves higher average PSNR and SSIM values as the recursive stage number increases. However, the SSIM value of Stack T-Net with three stages is a bit lower than the value of Stack T-Net with two stages, which is because the best model is saved according to the highest PSNR value, instead of the SSIM value. Actually, the highest SSIM value of Stack T-Net with three stages is higher than that of Stack T-Net with two stages in the experiment. This ablation study demonstrates that our recursive strategy is effective.
\subsection{Performance comparison on synthetic dataset}
The proposed method is tested on the same synthetic dataset SOTS, to qualitatively and quantitatively compare with the state-of-the-art methods which include DCP~\cite{5567108}, NLD~\cite{Berman_2016_CVPR}, MSCNN~\cite{2016Single}, DehazeNet~\cite{2016DehazeNet}, AOD-Net~\cite{8237773}, GFN~\cite{Ren_2018_CVPR}, DCPDN~\cite{8578435}, EPDN~\cite{8953692}, and DA\_dehaze~\cite{Shao_2020_CVPR}. Apart from the DCP and NLD which are prior based methods, the others are deep learning based methods. We test two models of our algorithm, namely T-Net and Stack T-Net with three stages in the comparison experiment.
Table~\ref{tab: tabel4} presents the quantitative comparison results of different dehazed methods on the SOTS dataset in terms of the average PSNR and SSIM values. Moreover, the data in this table except the last two rows is all quoted from the work of Shao~\cite{Shao_2020_CVPR}. It is clear that our work outperforms the state-of-the-art methods by a wide margin. Not only Stack T-Net but also our plain T-Net outperform the state-of-the-art DA\_haze~\cite{Shao_2020_CVPR} in terms of both the average PSNR and SSIM values on the SOTS dataset.
\begin{table}[tbp]
\centering
\caption{Quantitative comparison on the SOTS dataset between the state-of-the-art dehazing methods.}
\begin{tabular}{IC{3.0cm}|C{1.5cm}|C{1.5cm}I}
\Xhline{0.8pt}
\multirow{1}[4]{*}{Methods} & \multicolumn{2}{cI}{SOTS} \\
\cline{2-3} & PSNR & SSIM\\
\Xhline{0.8pt}
DCP~\cite{5567108} & 15.49 & 0.64 \\
\hline
NLD~\cite{Berman_2016_CVPR} & 17.27 & 0.75 \\
\hline
MSCNN~\cite{2016Single} & 17.57 & 0.81 \\
\hline
DehazeNet~\cite{2016DehazeNet} & 21.14 & 0.85 \\
\hline
AOD-Net~\cite{8237773} & 19.06 & 0.85 \\
\hline
DCPDN~\cite{8578435} & 19.39 & 0.65 \\
\hline
GFN~\cite{Ren_2018_CVPR} & 22.30 & 0.88 \\
\hline
EPDN~\cite{8953692} & 23.82 & 0.89 \\
\hline
DA\_dehaze~\cite{Shao_2020_CVPR} & 27.76 & 0.93 \\
\hline
T-Net & 28.55 & 0.95 \\
\hline
Stack T-Net &\textbf{28.83} &\textbf{0.96} \\
\Xhline{0.8pt}
\end{tabular}%
\label{tab: tabel4}
\end{table}%
Fig.~\ref{fig: SOTS} shows the qualitative comparison results among different dehazing methods, where the synthetic images are chosen from the SOTS datasets. The haze density and scene of the shown images in Fig.~\ref{fig: SOTS} are different, where we choose four images from the indoor and outdoor subset of SOTS, respectively. We note that the first three methods~\cite{Berman_2016_CVPR,2016DehazeNet,8237773} perform poorly on most images and are easy to cause under-dehazing, color oversaturation and other distortion problems. The dehazing results of DCPDN~\cite{8578435} and EPDN~\cite{8953692} seem better, but DCPDN~\cite{8578435} tends to make the brighter areas in images overexposed and EPDN~\cite{8953692} tends to reduce the brightness of the darker areas in images, which both cause the loss of many details in the images (see e.g., the first and third rows in Fig.~\ref{fig: SOTS} (e) (f)). DA\_dehaze~\cite{Shao_2020_CVPR} performs well on most images, but it tends to cause color distortions (see e.g., the first, second, fifth and sixth rows in Fig.~\ref{fig: SOTS} (g) of which the color is a little different from the ground truth images) and cause halo artifacts sometimes (see e.g., the sky area of the third row in Fig.~\ref{fig: SOTS} (g)), where other methods except our methods have the same problem.
Compared with the state-of-the-art methods, our T-Net and Stack T-Net have the best performance in terms of haze removal and are effective in suppressing halo artifacts and color distortions. The dehazed images are visually most similar to their ground truth ones (see e.g., Fig.~\ref{fig: SOTS} (h) (i)). In addition, we note that Stack T-Net can further eliminate hazy areas on the base of T-Net (see e.g. the first row in Fig.~\ref{fig: SOTS} (h) in which the haze in the middle is eliminated in the same row of (i)), demonstrating that the recursive strategy is helpful for image dehazing.
\subsection{Performance comparison on real-world hazy images}
We further compare our method with the state-of-the-art approaches on real-world images to evaluate the generalization ability of our method in the real-world scenery, where the real-world images used in the experiment are chosen from the URHI dataset. There is not quantitative comparison since the clean ground truth images of the real-world hazy images are not available, and the qualitative comparison results are shown in Fig.~\ref{fig: URHI}.
We note that the results on real-world data are primarily the same as those on the synthetic data. The dehazed images of NLD suffer from severe color distortions and halo artifacts (see e.g., Fig.~\ref{fig: URHI} (b)), as well as other problems such as overexposure (see e.g., the second, fifth rows in Fig.~\ref{fig: URHI} (b)). DehazeNet and AOD-Net tend to under-dehaze images (see e.g., the second, third, fifth and last rows in Fig.~\ref{fig: URHI} (c) (d)). DCPDN, EPDN and DA\_dehaze have better dehazing performance but tend to cause halo artifact in the sky area and the object edges with large color differences (see e.g., the sky area of the first, second, third, fourth and sixth rows and the edge of the airplane of the third row in Fig.~\ref{fig: URHI} (e) (f) (g)). Compared with the state-of-the-art ones, our methods have the best visual effect. As shown in Fig.~\ref{fig: URHI}, in addition to effective haze removal, T-Net and Stack T-Net can also suppress color distortions and halo artifacts. There are almost no artifacts or distortions in our dehazed images of which the sky area and the object edges are clean and smooth (see e.g., the first four rows in Fig.~\ref{fig: URHI} (h) (i)). Moreover, our method can restore the color of the image better than other methods (see e.g., the first row in Fig.~\ref{fig: URHI} of which the color is most similar to the hazy image). Although T-Net and Stack T-Net both have good performance on real world images, Stack T-Net can further improve the quality of the images (see e.g., the fifth row in Fig.~\ref{fig: URHI} (h) (i), where the image of (i) has higher brightness and looks cleaner than the same image of (g)). Specially, even for images with severe haze and deep depth (see e.g., the second row in Fig.~\ref{fig: URHI}), our method can still remove a certain amount of haze while maintaining the authenticity of the images, without causing image distortion due to halo artifacts like other methods. Overall, our methods achieve a good balance between dehazing and maintaining image authenticity.
\section{Conclusion}
\label{sec: concl}
In this work, we have proposed a new end-to-end dehazing network named T-Net, which is built based on the U-Net architecture and contains a backbone module and a dual attention module. Inspired by the recursive strategy, we have further proposed Stack T-Net by repeatedly unfolding the plain T-Net. Through the ablation studies, we verify that the overall design of T-Net is effective and the recursive strategy is helpful for dehazing. Experimental results on both synthetic and real-world images demonstrate that our T-Net and Stack T-Net perform favorably against the state-of-the-art dehazing algorithms, and that our Stack T-Net could further improve the dehazing performance.
The real-world images under severe haze lose amount of texture details and color information, which causes the performance degradation of the state-of-the-art dehazing methods or makes the color and details of the dehazing results deviate from the original images. Our future work will focus on this problem, and further study the inherent latent features of invariance.
\section*{Acknowledgment}
This work was partially supported by National Natural Science Foundation of China (Nos. 61771319, 62076165, 61871154), Natural Science Foundation of Guangdong Province (No. 2019A1515011307), Shenzhen Science and Technology Project (No. JCYJ20180507182259896) and the other project (Nos. 2020KCXTD004, WDZC20195500201).
\bibliographystyle{IEEEtran}
|
\section{Introduction}
In the early universe, primordial curvature perturbations can generate overdense regions. The gravitational collapse of the overdense areas may form primordial black holes (PBHs) after the horizon re-entry \cite{Hawking:1971,Carr:1974}. In contrast to the stellar black holes (BHs), PBHs have a broad mass range, and so they can explain the binary BH detected by the LIGO and Virgo collaborations \cite{Abbott:2016-a,Abbott:2016-b,Abbott:2017-a,Abbott:2017-b,Abbott:2017-c}. The PBHs formed in the early universe can be considered as a candidate for dark matter (DM) \cite{Ivanov:1994,Khlopov1:2005,Frampton:2010,Belotsky:2014,Clesse:2015,Carr:2016,Inomata:2017}. Sasaki et al. \cite{Sasaki:2016} showed that PBHs with a merger rate of around $12-213~ {\rm Gpc^{-3} yr^{-1}}$ and a mass scale of about $ 10 M_{\odot} $ can make up roughly ${\mathcal{O}(1)} \% $ of the total DM. Also, this type of PBHs is well consistent with the LIGO and Virgo observations \cite{Abbott:2017-a}.
Recently, the ultra-short timescale microlensing events were detected in the OGLE data and they prepare an allowed region for PBHs formation. The favored area of OGLE data shows that the abundance of PBHs with a mass scale of ${\mathcal{O}(10^{-5})}M_\odot $ can reach ${\mathcal{O}(10^{-2})}$ \cite{OGLE}. The observations tightly constrain the wide mass range of PBHs as DM, but there is no constraint on the mass scale from ${\mathcal{O}(10^{-13})}M_\odot$ to ${\mathcal{O}(10^{-11})}M_\odot$ \cite{Ali:2017,WD,HSC,EGG,femto,kepler,EROS,CMB-a,CMB-b,Katz:2018,Montero:2019}. Therefore, PBHs in this mass range can demonstrate all the DM in the universe.
The production of PBHs by the collapse of the overdense regions demands an amplitude of the primordial curvature perturbation $A_s$ in order of $\sim{\mathcal{O}(10^{-2})}$ at small scales \cite{sato:2019}. The cosmic microwave background (CMB) anisotropy measurements show that $A_s =2.1 \times 10^{-9}$ at the pivot scale $k_{*}=0.05{\rm Mpc}^{-1}$ \cite{akrami:2018}. It means that $A_s$ must increase nearly seven orders of magnitude at small scales for PBH formation.
A large abundance of PBHs cannot be generated in slow-roll inflationary models \cite{Motohashi:2017,Passaglia:2019}. Therefore, the violation of the slow-roll conditions is needed for PBHs formation. Recent studies have suggested various scenarios for PBHs production \cite{Cai:2018,Ballesteros:2019,Ballesteros:2020a,Ballesteros:2020b,Kamenshchik:2019,Inomata:2018,Ezquiaga:2018,Germani:2017,Di:2018,Ballesteros:2018,Dalianis:2019,
chen:2019,Ozsoy:2018,Atal:2019,mishra:2020,fu:2019,lin:2020,Khlopov:2010,Belotsky1:2014,Belotsky:2019,Braglia:2020,Braglia2:2020,shiPi:2018,Fumagalli:2020a,Sypsas:2020,Dalianis:2020}. For instance, a parametric resonance, which is due to the oscillating sound speed square, can enhance the primordial curvature perturbations \cite{Cai:2018,chen:2019}.
Also, in the single-field model with a non-canonical kinetic term, the curvature perturbations may increase if the sound speed becomes zero \cite{Ballesteros:2019,Kamenshchik:2019}.
Additionally, in the multi-field inflationary models, turning trajectories may lead to the power spectrum enhance at small scales and consequently the PBHs can be produced \cite{Fumagalli:2020a,Braglia:2020,Sypsas:2020}.
One of the most common approach for PBHs formation is using the inflationary models which have an inflection point. A single field model with an inflection point can lead to the violation of the slow-roll regime \cite{Germani:2017,Di:2018,Ezquiaga:2018,Dalianis:2019}.
Around the inflection point, the inflaton velocity decreases significantly, and the amplitude of the primordial curvature perturbation enhance firmly \cite{fu:2019,lin:2020}. For increasing the primordial curvature perturbation, we require fine-tuning of the model parameters. It should be mentioned that the total number of $e$-fold should remain between 50 and 60 \cite{Passaglia:2019,Sasaki:2018}. Moreover, the models must be consistent with the Planck observations at the CMB scale \cite{akrami:2018}.
Furthermore, the PBHs formation is followed by the generation of the induced gravitational waves (GWs), when the primordial curvature perturbations enhance significantly \cite{Matarrese:1998,Mollerach:2004,Saito:2009,Garcia:2017,Cai:2019-a,Cai:2019-b,Cai:2019-c,Bartolo:2019-a,Bartolo:2019-b,Wang:2019,Fumagalli:2020b,Domenech:2020a,Domenech:2020b,Hajkarim:2019,Kohri:2018,Xu:2020}. In other words, after the horizon re-entry, the collapse of the overdense regions can generate large metric perturbations besides PBHs. In the second-order, the scalar and tensor perturbations may be coupled to each other. The scalar metric perturbations, through the second-order effect, can generate the stochastic GW background \cite{Cai:2019-a,Cai:2019-b,Cai:2019-c,Bartolo:2019-a,Bartolo:2019-b,Wang:2019,Fumagalli:2020b}. Thus, induced GW signal detection indicates a novel approach to search for PBHs.
The Galileon inflation model is one of the most popular models in the inflation context, which is placed in the subset of Horndeski's theory \cite{Horndeski:1974,kobayashi:2010,Burrage:2010,Tumurtushaa:2019,teimoori:2018}.
In this scenario, the Galileon field acts like a scalar field which is responsible for inflation. In the Minkowski spacetime, the action of this specific scalar field is invariant under the Galilean symmetry $\partial_{\mu}\phi\rightarrow \partial_{\mu}\phi+ b_{\mu}$ \cite{Nicolis:2009,Deffayet:2009a,Deffayet:2009b}. Also, in the Galileon inflation models, the CMB anomalies and the decaying of the CMB power spectrum at the largest scales can be explained when inflaton undergoes the ultra slow-roll phase \cite{Hirano:2016}.
In addition, in \cite{lin:2020} it was shown that in the inflationary model driven by a suitable Galileon term $G(\phi)$, the scalar power spectrum can be enhanced and PBHs are produced. In \cite{lin:2020}, the authors studied the possibility of PBH formation with the Galileon term $G(\phi)=-\frac{1}{2}\frac{d}{\sqrt{\big(\frac{\phi-\phi_c}{c}\big)^2+1}} $ for the power-law potential $V(\phi)=\lambda \phi^{p}$, where $p=2/5$.
In this paper, our main goal is to investigate the possibility of PBH formation in the framework of inflation with field-dependent kinetic term for quartic and natural potentials. The structure of the paper is as follows. We review the inflation model with field-dependent kinetic term. in Sec. \ref{sec2}. The mechanism of PBH formation is explained in Sec. \ref{sec3}. In Sec. \ref{sec4}, we study the consequences of reheating to verify whether the primordial curvatures re-enter the horizon during reheating or after. In Sec. \ref{sec5}, we estimate the abundance of PBHs. The induced GWs are investigated in Sec. \ref{sec6}. Finally, Sec. \ref{sec7} is dedicated to our conclusions.
\section{Inflation with field-dependent kinetic term}\label{sec2}
The action of our model is given by \cite{lin:2020}
\begin{equation}
\label{action}
S=\int {\rm d}^4x\sqrt{-g}\left[\frac{M_{\rm pl}^{2}}{2}R+\big(1-2G(\phi) \big)X-V(\phi)\right],
\end{equation}
where $X\equiv -\frac{1}{2}g^{\mu\nu}\phi_{,\mu}\phi_{,\nu}$. In addition,
$g$ and $R$ are the determinant of the metric $g_{\mu \nu}$ and Ricci scalar, respectively.
Also, $G$ are considered as general functions of the scalar field $\phi$.
From the action (\ref{action}), for a spatially flat Friedmann-Robertson-Walker (FRW) universe the Friedmann equations are obtained as follows \cite{lin:2020,kobayashi:2010,Ohashi:2012}
\begin{gather}
\label{eom1}
3H^2=\frac{1}{2}\dot{\phi}^2+V(\phi)-\dot{\phi}^2G(\phi),\\
\label{eom2}
2\dot{H}+3H^2+\frac{1}{2}\dot{\phi}^2-V(\phi)-\dot{\phi}^2G(\phi)=0,
\end{gather}
where we take $M_{\rm{pl}}=1/\sqrt{8\pi G}=1$.
Using the action (\ref{action}), the equation of motion governing the scalar field $\phi$ reads
\begin{gather}
\label{eom3}
\ddot{\phi}+3H\dot{\phi}+\frac{V_{,\phi}-\dot{\phi}^2G_{,\phi}}{1-2G(\phi)}=0,
\end{gather}
where $,_{\phi}\equiv {\rm d}/{\rm d}\phi$, and the dot describe the derivative with respect to the cosmic time.
The quadratic action for curvature perturbation ${\cal R}$ at the first order approximation is given by \cite{kobayashi:2010}
\begin{eqnarray}\label{2ndaction}
S^{(2)}=\frac{1}{2} \int{\rm d}\tau{\rm d}^3x
\tilde{z}^2(1-2G_{,\phi})\left[({\cal R}_\phi')^2-(\Vec{\nabla}{\cal R}_\phi)^2\right],
\end{eqnarray}
where
$\tilde{z}= \frac{a\dot\phi}{H},$
and the prime indicates the derivative with respect to the conformal time $\tau$ \cite{kobayashi:2010}.
In the Fourier space, the Mukhanov-Sasaki (MS) equation can be calculated by varying the action (\ref{2ndaction}) with the respect to the curvature perturbation ${\cal R}$ as follows
\begin{eqnarray}\label{MS_Eq}
u''_k+\left(k^2-\frac{z''}{z}\right)u_k=0,
\end{eqnarray}
where
$ z=(1-2G)^{1/2}\tilde{z}$, and
$u_k = z{\cal R}_{\phi,k}$ \cite{lin:2020,kobayashi:2010}.
Consequently, the scalar power spectrum can be obtained as
\begin{equation}
{\cal P}_{\cal R}(k)=(2\pi^{2})^{-1}k^{3}\vert u_{k}/z\vert^{2}.
\end{equation}
The slow-roll parameters here are defined as
\begin{equation}
\varepsilon_1\equiv-\frac{\dot{H}}{H^2},\ \varepsilon_2\equiv-\frac{\ddot{\phi}}{H\dot{\phi}},\
\varepsilon_3\equiv\frac{G_{,\phi}\dot{\phi}^2}{V_{\phi}},
\end{equation}
where $|\varepsilon_i|\ll1$ for $i=1,2,3$.
In the slow-roll conditions, the background Eqs. (\ref{eom1}) and (\ref{eom3}) can be turned to
\begin{gather}
3H^2\simeq V, \\
3H\dot{\phi}(1-2G)+V_{\phi}\simeq0.
\end{gather}
Under the slow-roll approximation, the power spectrum of curvature perturbations is given by \cite{lin:2020}
\begin{equation}
\label{eq:ps}
P_{\cal R}=\frac{H^2}{8\pi^2\varepsilon_{1}}\simeq\frac{V^3}{12\pi^2V_{\phi}^2}(1-2G),\\
\end{equation}
and, the scalar spectral index can be estimated as
\begin{equation}
\label{nseq1}
n_s-1=\frac{1}{1-2G}\left(2\eta_V-6\varepsilon_V+
\frac{2G_{\phi}}{1-2G}\sqrt{2\varepsilon_V}\right),
\end{equation}
where $\varepsilon_V\equiv\frac{1}{2}(V'/V)^2$ and $\eta_V\equiv V''/V$. The recent value of the scalar spectral index measured by the Planck satellite is $n_s = 0.9627 \pm 0.0060$ (68$\%$ CL, Planck 2018 TT+lowE) \cite{akrami:2018}.
The tensor power spectrum is given by \cite{lin:2020}
\begin{equation}
\label{ptspec}
P_T=\frac{H^2}{2\pi^2},
\end{equation}
and the tensor-to-scalar ratio can be written as
\begin{equation}
\label{req1}
r\equiv\frac{P_T}{P_{\cal R}}=\frac{16 X(1-2G)}{H^2}.
\end{equation}
There is an upper bound on the tensor-to-scalar ratio provided by the Planck observation as $r< 0.0654$ (68$\%$ CL, Planck 2018 TT+lowE) \cite{akrami:2018}. In the following section, we solve numerically the background equations (\ref{eom2})-(\ref{eom3}) to obtain evolution of the both Hubble parameter and the scalar field. Then, with the help of numerical solution of the MS equation (\ref{MS_Eq}), we estimate the exact value of the scalar power spectrum. Note that in our numerical calculations, we use the slow roll solutions as initial conditions.
\section{PBH formation mechanism}\label{sec3}
A proper kinetic term can amplify the curvature perturbations at small scales. Also, the proposed function must cause the model to be consistent with the Planck measurements at the pivot scale $k_{*}=0.05{\rm Mpc}^{-1}$. To this aim, we suggest the kinetic function $G(\phi)$ to be parameterized as follows
\begin{equation}\label{g}
G(\phi)=g_I(\phi)\big(1+g_{II}(\phi)\big),
\end{equation}
where
\begin{equation}\label{gI}
g_I(\phi)=-\left(\frac{\phi}{M}\right)^{\alpha},
\end{equation}
\begin{equation}\label{gII}
g_{II}(\phi)=\frac{1}{2}\frac{d}{\sqrt{\big(\frac{\phi-\phi_c}{c}\big)^2+1}}\,.
\end{equation}
In the above, $g_I(\phi)$ is the base kinetic term, and we utilize it to sure that the model can satisfy the observational constraints on $n_s$ and $r$. Also $M$ is a constant with the mass dimension and for $M \rightarrow \infty$, our model recovers the standard inflation.
Note that the term $g_{II}(\phi)$ is responsible for generating the peak in the scalar power spectrum at $\phi=\phi_c $. Also, the value of $g_{II}(\phi)$ is vanishing at distances far from $\phi=\phi_c$. Here, the parameters $\alpha$ and $d$ are dimensionless constants. Also $\phi_c$ and $c$ have the dimension of mass. Additionally, $d$ and $c$ control the height and width of the scalar power spectrum at peak position, respectively.
Note that Lin et al. \cite{lin:2020} has already been studied a similar model like (\ref{g}) in which the selected kinetic term has only one part as $G(\phi)=-\frac{1}{2}\frac{d}{\sqrt{\big(\frac{\phi-\phi_c}{c}\big)^2+1}}$. Also, they considered the power-law potential $V(\phi)=\lambda \phi^{p}$, where $p=2/5$.
\subsection{Quartic inflation with field-dependent kinetic term}
Here, we are interested in investigating the possibility of PBH formation in our setup with quartic potential. The inflationary quartic potential has the following form
\begin{equation}\label{h-potential}
V(\phi)=\frac{\lambda}{4}\phi^{4},
\end{equation}
where $\lambda$ is a constant parameter. At the horizon exit, we set $N_{*}=0$ and take $\phi_{*}\simeq 0.765 $ at the pivot scale, which is estimated by slow-roll approximation. Note that in the framework of standard inflation, i.e. the Einstein gravity, the quartic potential (\ref{h-potential}) is completely rolled out by the Planck 2018 data \cite{akrami:2018}. This motivates us to examine the quartic potential in this scenario to check viability of the model in light of the Planck observations.
In the case of quartic inflation, the model is described by six free parameters ($\alpha$, $\lambda$, $d$, $c$, $\phi_{c}$, $M$). By setting suitable values for $M$ and $\alpha$ one can get the observational parameters $n_s$ and $r$ compatible with the Planck 2018 data. Also, the parameter $\lambda$ is fixed by the amplitude of scalar power spectrum at the CMB scale $k_{*}=0.05~{\rm Mpc}^{-1}$ \cite{akrami:2018}. Thus, $d$, $c$, and $\phi_{c}$ are the remaining free parameters of our model, and they can affect the PBH production.
\begin{table}[H]
\centering
\caption{The parameter sets for PBHs production in quartic inflation model. The parameters $\phi_c$ and $c$ are in units of $M_{\rm{pl}}=1$. Here, $M=0.46$ and $\alpha=15$.}
\begin{tabular}{cccc}
\hline
Sets \quad &\quad $\phi_{c}$ \quad & \quad $d$\quad &$c$\quad \\ [0.5ex]
\hline
\hline
$\rm{Case~}A_{\rm H}$ \quad &\quad $0.707$ \quad &\quad $2.659\times 10^{8}$ \quad &\quad $2.3\times10^{-11}$ \quad\\[0.5ex]
\hline
$\rm{Case~}B_{\rm H}$ \quad &\quad $0.739$ \quad &\quad $1.415\times 10^{8}$ \quad &\quad $2.1\times10^{-11}$ \quad\\[0.5ex]
\hline
$\rm{Case~}C_{\rm H}$ \quad &\quad $0.753$ \quad &\quad $1.265\times 10^{8}$ \quad &\quad $1.8\times10^{-11}$ \quad\\[0.5ex]
\hline
\end{tabular}
\label{tab1}
\end{table}
\begin{table}[H]
\centering
\caption{The values of $n_s$, $r$, $k_{\text{peak}}$, ${\cal P}_{\cal R}^\text{peak}$, $M_{\text{PBH}}^{\text{peak}}$ and $f_{\text{PBH}}^{\text{peak}}$ for the cases listed in Table \ref{tab1}.}
\begin{tabular}{ccccccc}
\hline
Sets \quad & \quad $n_{s}$\quad &$r$\quad & \quad$k_{\text{peak}}/\text{\rm Mpc}^{-1}$ \quad &\quad ${\cal P}_{\cal R}^\text{peak}$ \quad & \quad $M_{\text{PBH}}^{\text{peak}}/M_{\odot}$ \quad & \quad $f_{\text{PBH}}^{\text{peak}}$\\ [0.5ex]
\hline
\hline
$\rm{Case~}A_{\rm H}$ \qquad &\quad $0.963$ \quad &\quad $0.057$ \quad &\quad $2.29\times10^{12}$ \quad &\quad $0.035$ \quad &\quad $4.77\times 10^{-13}$ \quad &$0.91$ \\[0.5ex]
\hline
$\rm{Case~}B_{\rm H}$ \quad &\quad $0.965$ \quad &\quad $0.057$ \quad &\quad $4.52\times10^{8}$ \quad &\quad $0.042$ \quad &\quad $1.82\times10^{-5}$ \quad & $0.031$ \\[0.5ex]
\hline
$\rm{Case~}C_{\rm H}$ \quad &\quad $0.972$ \quad &\quad $0.056$ \quad &\quad $2.61\times10^{5}$ \quad &\quad $0.052$ \quad &\quad $36.9$ \quad &$0.0017$\\[0.5ex]
\hline
\end{tabular}
\label{tab2}
\end{table}
The slow-roll regime fails on the small scales where the peak of the power spectrum appears. It means that, for calculating the exact value of the scalar power spectrum ${\cal P}_{\cal R}(k)=(2\pi^{2})^{-1}k^{3}\vert u_{k}/z\vert^{2}$, we need to solve the MS equation (\ref{MS_Eq}), numerically. We find three sets of parameters, which are listed in Table \ref{tab1}. Here, we set $M=0.46$ and $\alpha=15$ to keep our model predictions consistent with the Planck measurements for $n_s$ and $r$ \cite{akrami:2018}. As shown in Table \ref{tab2}, in the cases $A_H$ and $B_H$ the predictions of the model for $n_s$ and $r$ take place inside the 68$\%$ CL region of the Planck 2018 TT+lowE data \cite{akrami:2018}. In the case $C_H$, the results of $n_s$ and $r$, respectively, are compatible with the 95$\%$ and 68$\%$ CL of the Planck 2018 data \cite{akrami:2018}. Note that these results are in contrast with the result of quartic potential in the standard model of inflation, in which the prediction of the model is completely rolled out in light of the Planck observations \cite{akrami:2018}. The parameter $d$ should be at least in order of ${\cal O}(10^8)$ to the power spectrum peak enhances seven
orders of magnitude at peak position. For fine tuning the parameter $c$ which controls the width of the scalar power spectrum at peak position, we have two restrictions. In one hand, the total number of $e$-folds should remain between 50 and 60. On the other hand, the quantity $g_{II}(\phi)$ should be negligible away from the peak to the usual slow-roll is guaranteed. Using these limitations, we set the parameter $c$ as in order of ${\cal O}(10^{-11})$.
The values of the power spectrum and the corresponding PBHs abundance for these sets are shown in Table \ref{tab2}.
The evolution of the scalar field $\phi$ versus the $e$-fold number $N$ is depicted in Fig. \ref{fig-phi-h1} for the parameter set $A_{\rm H}$.
The flat region in Fig. \ref{fig-phi-h1} is because of the decreasing inflaton velocity in this area where the inflaton undergoes the ultra slow-roll (USR) phase. The value of $\varepsilon_1$ is reduced critically in the USR stage, as shown in the Fig. \ref{fig-eps-h1}. The intense reduction can give a substantial accretion in the scalar power spectrum. In Fig. \ref{fig-eta-h1}, we see that during the USR period, the slow-roll condition $|\varepsilon_2|\ll 1$ is violated.
\begin{figure}[H]
\begin{minipage}[b]{1\textwidth}
\subfigure[\label{fig-phi-h1} ]{ \includegraphics[width=0.45\textwidth]%
{phiN_H1.eps}}\hspace{.1cm}
\subfigure[\label{fig-eps-h1}]{ \includegraphics[width=.45\textwidth]%
{eps_H1.eps}}
\centering
\subfigure[\label{fig-eta-h1}]{
\includegraphics[width=.45\textwidth]%
{eta_H1.eps}}
\end{minipage}
\caption{Evolutions of (a) the scalar filed $\phi$, (b) the first slow-roll parameter $\varepsilon_1$ and (c) the second slow-roll parameter $\varepsilon_2$ versus the $e$-fold number $N$ for the quartic inflation model. The dashed line shows the violation of the slow-roll condition. The auxiliary parameters are given by the parameter set of case $A_{\rm H}$ in Table \ref{tab1} and we take $N_{*}=0$ at the horizon exit.
}\label{linear}
\end{figure}
Figure \ref{fig-pr-higgs} presents the scalar power spectrum for the three parameter sets tabulated in Table \ref{tab1}. As displayed in Fig. \ref{fig-pr-higgs}, the scalar power spectrum grows seven orders from $\sim{\cal O}(10^{-9})$ at the CMB scale to $\sim{\cal O}(10^{-2})$ at the small scales, which is ideally suitable for PBHs formation. Also, the results of our model for the PBHs production are in good agreement with the observational constraints like the observations of CMB $\mu$-distortion, big bang nucleosynthesis (BBN), and pulsar timing array (PTA) \cite{Inomata:2019-a,Inomata:2016,Fixsen:1996}.
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{pr_Higgs.eps}
\vspace{-0.5em}
\caption{The scalar power spectrum ${\cal P}_{\cal R}$ in terms of wavenumber $k$ for the quartic inflation model. The blue, red and green lines are corresponding to the cases $A_{\rm H}$, $B_{\rm H}$ and $C_{\rm H}$, respectively. The CMB observations exclude the light-green shaded area \cite{akrami:2018}. The orange zone shows the $\mu$-distortion of CMB \cite{Fixsen:1996}. The cyan area represents the effect on the ratio between neutron and proton during the big bang nucleosynthesis (BBN) \cite{Inomata:2016}. The yellow region demonstrates the constraint from the PTA observations \cite{Inomata:2019-a}. }
\label{fig-pr-higgs}
\end{figure}
\subsection{Natural inflation with field-dependent kinetic term}
In this section, we study the PBHs formation in the framework of inflation with field-dependent kinetic term for the natural potential given by
\begin{equation}\label{potential}
V(\phi)=\lambda^4\left[1+\cos\left(\frac{\phi}{f}\right)\right],
\end{equation}
where $\lambda$ and $f$ are constant. Here we fix the parameter $\lambda$ by using $P_{\cal R}(k_{*})=2.1\times 10^{-9}$ at the CMB scale $k_{*}=0.05~{\rm Mpc}^{-1}$ \cite{akrami:2018}. With the help of slow-roll approximation, we obtain the scalar field at the pivot scale as $\phi_{*}\simeq 0.23 $. Also, we set $N_{*}=0$ at the horizon exit. Note that in the standard inflation setting, the prediction of the natural potential (\ref{potential}) is not in very good agreement with the latest observations, because its result takes place inside the 95$\%$ CL constraints of Planck 2018 TT+lowE data \cite{akrami:2018}. This motivates us to investigate whether the results of the natural potential in this scenario can be improved in light of the Planck 2018 observations.
For the case of natural potential (\ref{potential}), we have seven free parameters ($\alpha$, $\lambda$, $f$, $d$, $c$, $\phi_{c}$, $M$) for our model.
By choosing the proper values for $M$, $\alpha$, and $f$, one can adjust the values of $n_s$ and $r$ to be compatible with the Planck observations.
Also, the parameter $\lambda$ is fixed by the amplitude of scalar power spectrum at the pivot scale $k_{*}=0.05~{\rm Mpc}^{-1}$ \cite{akrami:2018}. As a result, the free parameters $d$, $c$, and $\phi_{c}$ are only remaining parameters which may affect the generation of PBHs.
For the natural inflation model we obtain the three parameter sets listed in Table \ref{tab3} which can lead to the PBHs generation. By setting $M=1.7$, $\alpha=-1$, and $f=1$, our model is compatible with the Planck observations for $n_s$ and $r$. In the cases $A_{\rm N}$ and $B_{\rm N}$, the results of $n_{s}$ and $r$ are compatible with the Planck observations at the $68\% $ CL \cite{akrami:2018}. Also in the case $C_{\rm N}$, we get $n_{s}=0.961$ which satisfies the 95$\%$ CL constraint, but the value of $r=0.005$ falls in the $68\%$ CL region of Planck 2018 data \cite{akrami:2018}.
Table \ref{tab4} shows the values of power spectrum and the PBHs abundances corresponding to the three cases of Table \ref{tab3}. In Fig. \ref{fig-phi-n1}, evolution of the scalar field as a function of the $e$-fold number is plotted for the case $A_{\rm N}$ in Table \ref{tab3}. In this figure, the plateau-like region at $\phi=\phi_{c}$ leads to a severe enhancement of the primordial curvature perturbations (see Fig. \ref{fig-pr-nat}). In this region, the slow-roll condition is violated because $|\varepsilon_2| > 1$, as shown in Fig. \ref{fig-eta-n1}.
Figure \ref{fig-pr-nat} shows the scalar power spectrum for the all cases of Table \ref{tab3}. For the PBHs formation, one needs the scalar power spectrum to increase seven orders at the small scales, and Fig. \ref{fig-pr-nat} exhibits such growth. Furthermore, Fig. \ref{fig-pr-nat} shows that the scalar power spectrum of this model is in good agreement with the observational bounds. These constraints include the observations of CMB $\mu$-distortion, big bang nucleosynthesis (BBN), and pulsar timing array (PTA) \cite{Inomata:2019-a,Inomata:2016,Fixsen:1996}.
\begin{table}[H]
\centering
\caption{The parameter sets for PBHs production in natural inflation model. The parameters $\phi_c$ and $c$ are in units of $M_{\rm{pl}}=1$. Here, $M=1.7$, $\alpha=-1$, and $f=1$.}
\begin{tabular}{cccc}
\hline
Sets \quad &\quad $\phi_{c}$ \quad & \quad $d$\quad &$c$\quad \\ [0.5ex]
\hline
\hline
$\rm{Case~}A_{\rm N}$ \quad &\quad $0.74$ \quad &\quad $5.34\times 10^{9}$ \quad &\quad $1.9\times10^{-11}$ \quad\\[0.5ex]
\hline
$\rm{Case~}B_{\rm N}$ \quad &\quad $0.41$ \quad &\quad $2.272\times 10^{9}$ \quad &\quad $1.3\times10^{-11}$ \quad\\[0.5ex]
\hline
$\rm{Case~}C_{\rm N}$ \quad &\quad $0.28$ \quad &\quad $1.311\times 10^{9}$ \quad &\quad $1.07\times10^{-11}$ \quad\\[0.5ex]
\hline
\end{tabular}
\label{tab3}
\end{table}
\begin{table}[H]
\centering
\caption{The values of $n_s$, $r$, $k_{\text{peak}}$, ${\cal P}_{\cal R}^\text{peak}$, $M_{\text{PBH}}^{\text{peak}}$ and $f_{\text{PBH}}^{\text{peak}}$ for the cases listed in Table \ref{tab3}.}
\begin{tabular}{ccccccc}
\hline
Sets \quad & \quad $n_{s}$\quad &$r$\quad & \quad$k_{\text{peak}}/\text{\rm Mpc}^{-1}$ \quad & \quad ${\cal P}_{\cal R}^\text{peak}$\quad &\quad $M_{\text{PBH}}^{\text{peak}}/M_{\odot}$ \quad & \quad $f_{\text{PBH}}^{\text{peak}}$\\ [0.5ex]
\hline
\hline
$\rm{Case~}A_{\rm N}$ \qquad &\quad $0.964$ \quad &\quad $0.007$ \quad &\quad $2.03\times10^{12}$ \quad &\quad $0.036$ \quad &\quad $6.05\times 10^{-13}$ \quad &$0.92$ \\[0.5ex]
\hline
$\rm{Case~}B_{\rm N}$ \quad &\quad $0.965$ \quad &\quad $0.006$ \quad &\quad $4.8\times10^{8}$ \quad &\quad $0.044$ \quad &\quad $1.60\times10^{-5}$ \quad & $0.035$ \\[0.5ex]
\hline
$\rm{Case~}C_{\rm N}$ \quad &\quad $0.961$ \quad &\quad $0.005$ \quad &\quad $2.68\times10^{5}$ \quad &\quad $0.053$ \quad &\quad $34.8$ \quad &$0.0015$\\[0.5ex]
\hline
\end{tabular}
\label{tab4}
\end{table}
\begin{figure}[H]
\begin{minipage}[b]{1\textwidth}
\subfigure[\label{fig-phi-n1} ]{ \includegraphics[width=0.45\textwidth]%
{phiN_N1.eps}}\hspace{.1cm}
\subfigure[\label{fig-eps-n1}]{ \includegraphics[width=.45\textwidth]%
{eps_N1.eps}}
\centering
\subfigure[\label{fig-eta-n1}]{
\includegraphics[width=.45\textwidth]%
{eta_N1.eps}}
\end{minipage}
\caption{Same as Fig. \ref{linear}, but for the natural inflation model. The auxiliary parameters are given by the parameter set of case $A_{\rm N}$ in Table \ref{tab3} and we take $N_{*}=0$ at the horizon exit.
}\label{linear}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{pr_Natt.eps}
\vspace{-0.5em}
\caption{Same as Fig. \ref{fig-pr-higgs}, but for the natural inflation model. The blue, red and green lines are corresponding to the cases $A_{\rm N}$, $B_{\rm N}$ and $C_{\rm N}$, respectively. }
\label{fig-pr-nat}
\end{figure}
\section{Reheating stage}\label{sec4}
After inflation, the scalar field starts to oscillate around the minimum value of the potential. Consequently, inflaton is decayed to the particles of the Standard Model. This process is known as the reheating, which joins the supercooled universe at the end of inflation to the thermalized universe at the radiation dominated (RD) epoch.
The decay rate of the inflaton can influence the reheating era duration. On the other hand, the reheating temperature $T_{\rm reh}$ can affect the decay rate of the inflaton. Hence, the large and small reheating temperatures lead to the short and prolonged reheating eras, respectively. The horizon re-entry can occur before the RD era if reheating phase has a prolonged period. Depending on the re-entry horizon era, the mathematical formalisms applied to determine the mass fraction and the energy density of GWs will be different. Here, we discuss whether PBHs produce in the RD era or reheating epoch \cite{mahbub:2020}.
In this paper, we use the method introduced in \cite{Dalianis:2019} to estimate the time of horizon re-entry. For a scale $k^{-1}$ that exits the horizon $\Delta N_{k}$ $e$-fold before the end of inflation, we can write
\begin{equation}\label{deltank}
\left( \frac{a_{k,\text{re}}}{a_{\text{end}}} \right)^{\frac{1}{2}(1+3w)}=e^{\Delta N_{k}},
\end{equation}
where $a_{k,\text{re}}$ denotes the scale factor at the horizon re-entry moment, and $w$ is the equation of state parameter. The number of $e$-folds between the end of inflation and horizon re-entry can be defined as
\begin{equation}\label{nk}
\tilde{N}_{k}\equiv \ln\left( \frac{a_{k,\text{re}}}{a_{\text{end}}} \right).
\end{equation}
Consequently, Eqs. (\ref{deltank}) and (\ref{nk}) can be related to each other as follows
\begin{equation}\label{deltaNtoN}
\tilde{N}_{k}=\left(\frac{2}{1+3w}\right)\Delta N_{k},
\end{equation}
where $w>-1/3$. The number of $e$-folds throughout the reheating period is
$\tilde{N}_{\text{reh}}\equiv \ln\left( a_{\text{reh}}/a_{\text{end}} \right)=\left[3(1+w_{\text{reh}}) \right]^{-1} \ln\left(\rho_{\text{end}}/\rho_{\text{reh}} \right) $,
in which $a_{reh}$ shows the value of the scale factor at the end of reheating \cite{mahbub:2020,Dalianis:2019}. Also, $\rho_{\text{end}}=3H_{\text{end}}^{2}M_{p}^2$ is the energy density at the end of inflation epoch.
The values of $\tilde{N}_{k}$ and $\tilde{N}_{\text{reh}}$ can specify the time of horizon re-entry. If $\tilde{N}_{k}>\tilde{N}_{\text{reh}}$, the scale $k^{-1}$ re-enters during the RD era. On the other hand, the re-entering of the scale $k^{-1}$ with $\tilde{N}_{k}<\tilde{N}_{\text{reh}}$ occurs during the reheating stage. It is possible to assume $w_{\rm reh}\simeq 0$ because the reheating phase can be recognized as an early matter-dominated era. The amount of observable inflation $\Delta N$ is given by \cite{mahbub:2020}
\begin{equation}\label{deltaNfinal}
\Delta N\simeq 57.3+\frac{1}{4}\ln\left( \frac{\varepsilon_{*}V_{*}}{\rho_{\text{end}}} \right)-\frac{1}{4}\tilde{N}_{\text{reh}}~,
\end{equation}
where $\varepsilon_{*}$ and $V_{*}$ show the values of the first slow-roll parameter and the potential at the pivot scale, respectively.
For a scale such as $k^{-1}$, which re-enters the horizon just at the end of the reheating era, we have $\tilde{N}_{k}=\tilde{N}_{\text{reh}}$. Hence, $\Delta N_{\text{k}}$ and $\tilde{N}_{\text{reh}}$ can be associated with $\Delta N_{\text{k}}=\tilde{N}_{\text{reh}}/2$ in which we used Eq. (\ref{deltaNtoN}) and assumed $w_{\text{reh}}\simeq 0$. Therefore, the scale $k^{-1}$ re-enters horizon after reheating phase if $\Delta N_{\text{k}} > \tilde{N}_{\text{reh}}/2$.
Now for the both models studied in this paper, we can calculate the values of $\Delta N_{\text{k}}^{\rm peak}$ and $\tilde{N}_{\text{reh}}$, where $\Delta N_{\text{k}}^{\rm peak}$ describes the number of $e$-folds after the end of inflation until the scale $k^{-1}_{\text{peak}}$ re-enters the horizon. The co-moving wavenumber of the scale $k^{-1}$ that re-enters the horizon at the end of reheating phase, $k^{-1}_{\rm reh}$, is given by $k_{\rm reh}=e^{-\frac{\tilde{N}_{\rm reh}}{2}} k_{\rm end}$, where $k^{-1}_{\rm end}$ is the scale that exits the horizon at the end of inflation \cite{Dalianis:2019}.
Our results in Table \ref{tab5} indicate that for the all cases in the both models we have $\Delta N_{\text{k}}^{\rm peak} > \tilde{N}_{\text{reh}}/2$. This means that in our models, the PBHs formation occurs in the RD era.
For instance, in Figs. \ref{fig-reh-higgs} and \ref{fig-reh-nat} we plot the scalar power spectrum associated with the cases $A_{\rm H}$ and $A_{\rm N}$, respectively.
The shaded regions describe the scales that re-enters the horizon during the reheating stage.
As shown in this figure, the duration of the reheating phase is short, and the peak scales re-enter the horizon after reheating epoch.
Hence, In the following section, we use the mathematical formalisms, which are verified in the RD era to calculate PBHs abundance and energy density of the induced GWs.
\begin{table}[H]
\centering
\caption{The values of $N_{\text{peak}}$, $\Delta N_{\text{k}}^{\rm peak}$ and $\tilde{N}_{\text{reh}}$ for the all cases in Tables \ref{tab1} and \ref{tab3}.}
\begin{tabular}{cccccc}
\hline
Sets \quad & \quad $\Delta N$\quad & $k_{\text{peak}}$ \quad & \quad $N_{\text{peak}}$ \quad &\quad $\Delta N_{\text{k}}^{\rm peak}$ \quad & \quad$\tilde{N}_{\text{reh}}$ \quad \\ [0.5ex]
\hline
$\rm{Case~}A_{\rm H}$\quad & \quad 54.5 \quad & \quad $2.29\times 10^{12}$ \quad & \quad 31.6 \quad & \quad 16.42 \quad & \quad 6.98 \quad\\ \hline
$\rm{Case~}B_{\rm H}$ \quad & \quad 54.4 \quad & \quad $4.52\times 10^{8}$ \quad & \quad 23 \quad & \quad 22.16 \quad & \quad 7.44 \quad\\ \hline
$\rm{Case~}C_{\rm H}$ \quad & \quad 54.3 \quad & \quad $2.61\times 10^{5}$ \quad & \quad 15.5 \quad & \quad 26.56 \quad & \quad 7.85 \quad\\ \hline\hline
$\rm{Case~}A_{\rm N}$ \quad & \quad 54 \quad & \quad $2.03\times 10^{12}$ \quad & \quad 31.4 \quad & \quad 16.5 \quad & \quad 8.4 \quad\\ \hline
$\rm{Case~}B_{\rm N}$ \quad & \quad 53.5 \quad & \quad $4.8\times 10^{8}$ \quad & \quad 23 \quad & \quad 22.04 \quad & \quad 9.2 \quad\\ \hline
$\rm{Case~}C_{\rm N}$ \quad & \quad 53.5 \quad & \quad $2.68\times 10^{5}$ \quad & \quad 15.5 \quad & \quad 26.83 \quad & \quad 9.2 \quad\\ \hline
\end{tabular}
\label{tab5}
\end{table}
\begin{figure}[H]
\begin{minipage}[b]{1\textwidth}
\subfigure[\label{fig-reh-higgs} ]{ \includegraphics[width=.46\textwidth]%
{phi4reh.eps}}\hspace{.1cm}
\subfigure[\label{fig-reh-nat}]{ \includegraphics[width=.46\textwidth]%
{natreh.eps}}
\end{minipage}
\caption{The scalar power spectrum for the cases (a) $A_{\rm H}$ and (b) $A_{\rm N}$, respectively, in quartic and natural inflation models. The shaded areas denote
the scales that re-enter the horizon during the reheating phase }\label{reh-fig}
\end{figure}
\section{Abundance of primordial black holes}\label{sec5}
During the RD era, when primordial curvature perturbations re-enter the horizon, the gravitational collapse may generate PBHs. The PBH mass at the production time is given by $M=\gamma M_{\rm H}$, where $M_{\rm H}$ is the horizon mass and $\gamma=0.2$ is the collapse efficiency parameter \cite{Inomata:2017,Sasaki:2018}. The current ratio of the PBH mass to the total DM reads \cite{Sasaki:2018}
\begin{equation}\label{fpbheq}
f_{\rm PBH}(M_{\rm PBH})=1.68\times 10^{8} \left(\frac{\gamma}{0.2} \right)^{\frac{1}{2}} \left(\frac{g_{*}}{106.75} \right)^{-\frac{1}{4}} \left(\frac{M_{\rm PBH}}{M_{\odot}} \right)^{-\frac{1}{2} }\beta(M_{\rm PBH}),
\end{equation}
where $g_{*}\simeq106.75$ is the effective degrees of freedom at the formation time of the PBH \cite{Motohashi:2017}. The mass fraction of PBH $\beta$ in Eq. (\ref{fpbheq}) can be estimated by \cite{Sasaki:2018,young:2014,harada:2013,Musco:2013,Germani:2019,Shibata:1999,Polnarev:2007,Musco:2009}
\begin{equation}\label{betaeq}
\beta(M_{\rm PBH})=\gamma \frac{\sigma_{M_{\rm PBH}}}{\sqrt{2\pi}\delta_{\rm th}}\exp{\left(-\frac{\delta_{\rm th}^{2}}{2 \sigma_{M_{\rm PBH}}^{2}} \right)},
\end{equation}
where $\delta_{\rm th}=0.4$ shows the threshold density contrast for PBH formation. Also, $\sigma_{M_{\rm PBH}}$ is the variance of density contrast at the comoving horizon scale and it is obtained as \cite{young:2014}
\begin{equation}\label{sigmaeq}
\sigma_{k}^{2}=\left(\frac{4}{9} \right)^{2} \int \frac{{\rm d}q}{q} W^{2}(q/k)(q/k)^{4} P_{\cal R}(q),
\end{equation}
where
$W(x)=\exp{\left(-x^{2}/2 \right)} $
is the Gaussian window function.
It is possible to associate the mass of the PBHs and the corresponding wavenumber as \cite{Motohashi:2017,mishra:2020,Sasaki:2018}
\begin{equation}\label{masseq}
M_{\rm PBH}=1.13\times 10^{15}\left(\frac{\gamma}{0.2} \right) \left(\frac{g_{*}}{106.75} \right)^{-\frac{1}{6}}\left(\frac{k_{\rm PBH}}{k_{*}} \right)^{-2} M_{\odot}.
\end{equation}
With the help of Eqs. (\ref{fpbheq}) and (\ref{masseq}), one can estimate the abundance and the mass of PBHs. For the both quartic and natural inflationary models, the results of $M_{\text{PBH}}^{\text{peak}}$ and $f_{\text{PBH}}^{\text{peak}}$ calculated at the peak scale of the PBH formation are listed in Tables \ref{tab2} and \ref{tab4}. Also in Fig. \ref{fpbh-figs}, we plot variations of $f_{\text{PBH}}$ versus $M_{\text{PBH}}$ for the both models. In this figure, the shaded areas reveal the observational bounds on the abundance of PBH. Figure \ref{fpbh-figs} shows that (i) the PBHs with masses $\sim10^{-13}M_\odot$, $\sim10^{-5}M_\odot$, and $\sim10M_\odot$ which are compatible with the observations can be produced in the aforementioned models. (ii) Surprisingly enough is that for the cases $A_{\rm H}$ and $A_{\rm N}$, we obtain $f_{\rm PBH}^{\rm peak}=0.91$ and $f_{\rm PBH}^{\rm peak}=0.92$, respectively, which can explain most of DM in the universe. (iii) For the cases ${\rm B_{H}}$ and ${\rm B_{N}}$, the abundance peak of the PBHs can grow up to $f_{\text{PBH}}^{\text{peak}}= 0.031$ and $f_{\text{PBH}}^{\text{peak}}= 0.035$, respectively, for the mass scales $1.82\times10^{-5}M_\odot$ and $1.60\times10^{-5}M_\odot$. The results obtained for the cases ${\rm B_{H}}$ and ${\rm B_{N}}$ are located at the allowed region of microlensing events in the OGLE data.
(iv) For the cases ${\rm C_{H}}$ and ${\rm C_{N}}$, the peaks of the PBH abundance are located at $36.9M_{\odot}$ and $34.8M_{\odot}$ and enhance to $f_{\text{PBH}}^{\text{peak}}= 0.0017$ and $f_{\text{PBH}}^{\text{peak}}= 0.0015$, respectively. This result is compatible with the upper limit bounds on the LIGO merger rate and consequently can describe BHs observed by LIGO and Virgo collaboration.
\begin{figure}[H]
\begin{minipage}[b]{1\textwidth}
\subfigure[\label{fig-fpbh-higgs} ]{ \includegraphics[width=.46\textwidth]%
{fpbh_Higgs.eps}}\hspace{.1cm}
\subfigure[\label{fig-fpbh-nat}]{ \includegraphics[width=.46\textwidth]%
{fpbh_natt.eps}}
\end{minipage}
\caption{(a) The PBH abundance $f_{\rm PBH}$ versus the mass $M$ in the quartic inflation model for the cases $A_{\rm H}$ (blue line), $B_{\rm H}$ (red line), and $C_{\rm H}$ (green line). (b) Same as Fig. \ref{fig-fpbh-higgs} but in the natural inflation model for the cases $A_{\rm N}$ (blue line), $B_{\rm N}$ (red line), and $C_{\rm N}$ (green line). The observational constraints on PBH abundance are shown by the shaded areas. The brown shaded area illustrates the allowed region of PBH abundance in the OGLE data \cite{OGLE}. The red dots show the upper bound on the PBH abundance due to the upper limit on the LIGO event merger rate \cite{Ali:2017}. Other shaded regions present the recent observational constraints containing accretion constraints from CMB \cite{CMB-a,CMB-b}, with EROS/MACHO (EROS/MACHO) \cite{EROS}, with the Kepler satellite (Kepler) \cite{kepler}, microlensing events with Subaru HSC (Subaru HSC) \cite{HSC}, white dwarf explosion (WD) \cite{WD}, femtolensing of gamma-ray burst (Femto) \cite{femto} and extragalactic gamma rays from PBH evaporation (EG$\gamma$) \cite{EGG}.}\label{fpbh-figs}
\end{figure}
\section{Induced Gravitational Waves}\label{sec6}
The induced GWs can produce together with PBHs at the horizon re-entry of the primordial density perturbations \cite{Matarrese:1998,Mollerach:2004,Saito:2009,Garcia:2017,Cai:2019-a,Cai:2019-b,Cai:2019-c,Bartolo:2019-a,Bartolo:2019-b,Wang:2019,Fumagalli:2020b,Domenech:2020a,Domenech:2020b,Hajkarim:2019,Kohri:2018,Xu:2020}.
The induced GWs can be tested by the GWs observatory like LISA \cite{lisa}. To investigate the induced GWs, we start with the perturbed FRW metric in the conformal Newtonian gauge which is written as \cite{Ananda:2007}
\begin{eqnarray}
ds^2=a(\eta)^2\left[ -(1+2\Psi)d\eta^2 +\left((1-2\Psi)\delta_{ij}+\frac{h_{ij}}{2} \right)dx^idx^j \right],
\end{eqnarray}
where $a$ is the scale factor and $\eta$ represents the conformal time. Also $\Psi$ denotes the first-order scalar perturbations, and
$h_{ij}$ is the perturbation of the second-order transverse-traceless tensor. After inflation, the universe should be thermalized in the reheating era. Hence, the inflaton will be decayed into the light particles, and the universe will be dominated by radiation, consequently. In the RD era, the effect of the inflation field has almost ignorable on the cosmic evolution. Accordingly, to study the scalar induced GWs during the RD era, we can use the standard Einstein equation. So, the equation of motion for second-order tensor perturbations $h_{ij}$ satisfies \cite{Ananda:2007,Baumann:2007}
\begin{eqnarray}\label{EOM_GW}
h_{ij}^{\prime\prime}+2\mathcal{H}h_{ij}^\prime - \nabla^2 h_{ij}=-4\mathcal{T}^{lm}_{ij}S_{lm}\;,
\end{eqnarray}
where $\mathcal{H}\equiv a^{\prime}/a$ is the conformal Hubble parameter. The quantity $\mathcal{T}^{lm}_{ij}$ denotes the transverse-traceless projection operator, and $S_{ij}$, which is the GW source term, is obtained as
\begin{eqnarray}
S_{ij}=4\Psi\partial_i\partial_j\Psi+2\partial_i\Psi\partial_j\Psi-\frac{1}{\mathcal{H}^2}\partial_i(\mathcal{H}\Psi+\Psi^\prime)\partial_j(\mathcal{H}\Psi+\Psi^\prime)\; .
\end{eqnarray}
During the RD era, the scalar metric perturbation $\Psi$ takes the form \cite{Baumann:2007}
\begin{eqnarray}
\Psi_k(\eta)=\psi_k\frac{9}{(k\eta)^2}\left(\frac{\sin(k\eta/\sqrt{3})}{k\eta/\sqrt{3}}-\cos(k\eta/\sqrt{3}) \right)\;,
\end{eqnarray}
where $k$ indicates the comoving wavenumber, and the primordial perturbation $\psi_k$ is given by
\begin{eqnarray}
\langle \psi_{\bf k}\psi_{ \tilde{\bf k}} \rangle = \frac{2\pi^2}{k^3}\left(\frac{4}{9}\mathcal{P}_\mathcal{R}(k)\right)\delta(\bf{k}+ \tilde{\bf k})\;.
\end{eqnarray}
The energy density of the induced GWs in the RD era can be estimated by \cite{Kohri:2018}
\begin{eqnarray}\label{OGW}
&\Omega_{\rm{GW}}(\eta_c,k) = \frac{1}{12} {\displaystyle \int^\infty_0 dv \int^{|1+v|}_{|1-v|}du } \left( \frac{4v^2-(1+v^2-u^2)^2}{4uv}\right)^2\mathcal{P}_\mathcal{R}(ku)\mathcal{P}_\mathcal{R}(kv)\left( \frac{3}{4u^3v^3}\right)^2 (u^2+v^2-3)^2\nonumber\\
&\times \left\{\left[-4uv+(u^2+v^2-3) \ln\left| \frac{3-(u+v)^2}{3-(u-v)^2}\right| \right]^2 + \pi^2(u^2+v^2-3)^2\Theta(v+u-\sqrt{3})\right\}\;,
\end{eqnarray}
where $\Theta$ is the Heaviside theta function, and $\eta_{c}$ is the time in which the growing of $\Omega_{\rm{GW}}$ is stopped. Also the scalar power spectrum $\mathcal{P}_\mathcal{R}$ is estimated by the MS equation (\ref{MS_Eq}). The current energy density of the induced GWs can be obtained as follows \cite{Inomata:2019-a}
\begin{eqnarray}\label{OGW0}
\Omega_{\rm{GW},0}h^2 = 0.83\left( \frac{g_c}{10.75} \right)^{-1/3}\Omega_{\rm{r},0}h^2\Omega_{\rm{GW}}(\eta_c,k)\;,
\end{eqnarray}
where $\Omega_{\rm{r},0}h^2\simeq 4.2\times 10^{-5}$ is the radiation density parameter at the present time. Also $g_c\simeq106.75$ indicates the effective degrees of freedom in the energy density at $\eta_c$. The frequency and comoving wavenumber can be related as follows
\begin{eqnarray}\label{k_to_f}
f=1.546 \times 10^{-15} \left(\frac{k}{{\rm Mpc}^{-1}}\right){\rm Hz}.
\end{eqnarray}
Now, with numerical solving of Eq. (\ref{OGW}) and using (\ref{OGW0})-(\ref{k_to_f}) one can estimate the present energy density of induced GWs. The results of $\Omega_{\rm{GW},0}$ for the quartic and natural models are shown in Fig. \ref{omega-figs}. The blue, red, and green curves correspond to the PBHs with the mass in the order of $ 10^{-13}M_\odot$, $ 10^{-5}M_\odot$, and $ 10M_\odot$, respectively. The shape and the amplitude of the current energy spectra of the induced GWs are almost the same for all cases, but the peak frequencies are different as depicted in Figs. \ref{fig-omega-higgs} and \ref{fig-omega-nat}.
Figure \ref{omega-figs} shows that (i) for the cases $A_{\rm H}$ and $A_{\rm N}$ with $M \sim {\cal O}(10^{-13})M_\odot$, the peak frequency of $\Omega_{GW,0}$ is around the $\rm{mHz}$ band, which means both of them can place in the sensitive region of space-based observatories like LISA, $\rm{TianQin}$, and Taiji.
(ii) In the cases $B_{\rm H}$ and $B_{\rm N}$ with PBH mass around $10^{-5}M_\odot$, the $\Omega_{GW,0}$ has a peak in $f \sim 10^{-6} \rm{Hz}$ band. (iii) The energy density of induced GWs produced from the cases $C_{\rm H}$ and $C_{\rm N}$ has a peak in frequencies in the order of $10^{-10} \rm{Hz}$. Hence, the cases $B_{\rm H}$, $C_{\rm H}$, $B_{\rm N}$, and $C_{\rm N}$ can be tested by SKA observation.
\begin{figure}[H]
\begin{minipage}[b]{1\textwidth}
\subfigure[\label{fig-omega-higgs} ]{ \includegraphics[width=.49\textwidth]%
{omega_fit_higgs.eps}}\hspace{0.01cm}
\subfigure[\label{fig-omega-nat}]{ \includegraphics[width=.49\textwidth]%
{omega_nat_fit.eps}}
\end{minipage}
\caption{(a) The induced GWs energy density parameter $\Omega_{\rm GW{,0}}$ versus the frequency in the quartic inflation model for the cases $A_{\rm H}$ (blue line), $B_{\rm H}$ (red line), and $C_{\rm H}$ (green line). (b) Same as Fig. \ref{fig-omega-higgs} but in the natural inflation model for the cases $A_{\rm N}$ (blue line), $B_{\rm N}$ (red line), and $C_{\rm N}$ (green line). The black dashed line denotes the broken power-law behaviour of $\Omega_{GW}$. The dashed curves describe the sensitivity of GWs observatories, such as the European PTA (EPTA) \cite{EPTA-a,EPTA-b,EPTA-c,EPTA-d}, the Square Kilometer Array (SKA) \cite{ska}, the Advanced Laser Interferometer Gravitational Wave Observatory (aLIGO) \cite{ligo-a,ligo-b}, the Laser Interferometer Space Antenna (LISA) \cite{lisa,lisa-a}, Taiji \cite{taiji} and TianQin \cite{tianqin}.}\label{omega-figs}
\end{figure}
Recent studies confirm that the energy density can be parameterized as $\Omega_{\rm GW} (f) \sim f^{n} $ \cite{Xu:2020,Fu:2020,Kuro:2018}.
In this regard, for the case $A_{\rm H}$ we estimate $\Omega_{\rm GW} \sim f^{1.78}$ for $f<f_{c}=3.54\times 10^{-3}{\rm Hz}$, and $\Omega_{\rm GW} \sim f^{-2.69}$ for $f>f_{c}$. Also, for the case $A_{\rm N}$, our calculations indicate that $\Omega_{\rm GW} \sim f^{1.68}$ for $f<f_{c}=3.73\times 10^{-3}{\rm Hz}$, and $\Omega_{\rm GW} \sim f^{-2.69}$ for $f>f_{c}$. In addition, the results in both cases confirm $\Omega_{\rm GW} \sim f^{3-2/\ln(f_c/f)}$ in the infrared limit $f\ll f_{c}$, which is entirely compatible with the results of \cite{Yuan:2020,shipi:2020}.
\section{Conclusions}\label{sec7}
In this paper, we investigated the possibility of PBHs formation in the context of inflation with field-dependent kinetic term for two types of the inflationary potentials (i.e. quartic and natural potentials). Utilizing a proper kinetic term may increase the primordial perturbations to ${\cal O}(10^{-2})$ at small scales, which is required for PBHs formation. On the other hand, the value of the $P_{\cal R}(k)$ should be consistent with the Planck observation at the pivot scale. In our model, we use a kinetic function that contains two parts as $G(\phi)=g_I(\phi)\big(1+g_{II}(\phi)\big)$. The first part $g_{I}(\phi)$ guarantees that the model is compatible with the Planck measurements at the CMB scale. Moreover, the fine-tuning of the parameters of the second section $g_{II}(\phi)$ can severely increase the $P_{\cal R}(k)$ at small scales.
We find three sets of parameters for each potential which are listed in Tables \ref{tab1} and \ref{tab3}. Our calculations demonstrate that the scales re-enter the horizon after the reheating stage. Therefore, we can apply the mathematical formalism that is valid in the RD epoch.
In the following, we estimated the PBHs abundance for all cases in Tables \ref{tab2} and \ref{tab4}. In our model, the noteworthy sets are $A_{\rm H}$ and $A_{\rm N}$, which their $f_{\rm PBH}$ are in order $ {\cal O}(1)$ and can explain the total DM in the universe. For the cases $B_{\rm H}$ and $B_{\rm N}$, the peak values of $f_{\rm PBH}$ are as $ {\cal O}(10^{-2})$, which locate in the allowed region of ultrashort timescale in OGLE data. Also, the cases of $C_{\rm H}$ and $C_{\rm N}$ are compatible with the upper limit of LIGO, and $f_{\rm PBH}$ grows to $ {\cal O}(10^{-3})$.
In the last section, we studied the propagation of secondary GWs in our model. For the cases $A_{\rm H}$ and $A_{\rm N}$, the peaks of $\Omega_{\rm GW}$ are formed in the mHz frequency range and can be examined by the observations of LISA, Taiji, and TianQin. Moreover, as shown in the Fig. \ref{omega-figs}, in the cases $B_{\rm H}$, $C_{\rm H}$, $B_{\rm N}$, and $C_{\rm N}$ the GWs energy density parameter can be examined by the SKA observation. In addition, our numerical calculations expose that the GWs energy density parameter can be parameterized as a power-law function $\Omega_{\rm GW} (f) \sim f^{n} $. The results show that $\Omega_{\rm GW} \sim f^{1.78}$ for $f<f_{c}=3.54\times 10^{-3}{\rm Hz}$, and $\Omega_{\rm GW} \sim f^{-2.69}$ for $f>f_{c}$ in the case of $A_{\rm H}$, and $\Omega_{\rm GW} \sim f^{1.68}$ for $f<f_{c}=3.73\times 10^{-3}{\rm Hz}$, and $\Omega_{\rm GW} \sim f^{-2.69}$ for $f>f_{c}$ in the case of $A_{\rm N}$. Also, in the infrared limit $f\ll f_{c}$, for the both cases, the power index of the GWs energy density satisfies the relation $n=3-2/\ln(f_c/f) $, which is well consistent with the analytical result obtained in \cite{Yuan:2020,shipi:2020}.
\subsection*{Acknowledgements}
The authors thank the referee for his/her valuable comments.
|
\section{Introduction}
\label{intro}
\begin{figure}[!htb]
\centering
\adjincludegraphics[Clip=0 0 0 {0.404\height}, width=0.8\linewidth]{images/detector.jpg}
\caption{The central detector of JUNO. An acrylic sphere with a diameter of 35.4 meters filled with 20 kt of liquid scintillator. The detector contains \num{17612} large (20 inches) PMTs and \num{25600} small PMTs (3 inches).}
\label{fig:detector}
\end{figure}
JUNO is a neutrino observatory under construction in southern China. Its physical program covers a wide range of problems~
\cite{JUNO}. The main goals are to determine the neutrino mass ordering and to accurately measure the parameters of neutrino oscillations $\sin^2{\theta_{12}}, \Delta m_{21}^2, \Delta m^{2}_{31}$. JUNO will detect reactor neutrinos from the Yangjiang and Taishan nuclear power plants. Simultaneously JUNO will be able to observe neutrinos from supernovae, atmospheric neutrinos, solar neutrinos and geoneutrinos.
Figure~\ref{fig:detector} shows the detector design. The detector is a transparent acrylic sphere with a diameter of 35.4 meters that is located underground in a cylindrical water pool. The sphere is filled with 20 kt of liquid scintillator. The detector is equipped with a huge number of photo-multiplier tubes (PMTs) of two types: \num{17612} large PMTs (20 inches) and \num{25600} small PMTs (3 inches). Neutrinos, which are produced in nuclear reactors, interact with the protons of the scintillator in the detector via the inverse beta-decay (IBD) channel: $\overline \nu_{e} + p \rightarrow e^{+} + n$. The scintillator then produces visible light upon the interaction of the ejected positron with the media. The amount of emitted photons is tightly related to the neutrino energy. The neutron, after some time, is captured by a hydrogen atom of liquid scintillator, producing 2.2 MeV de-excitation gammas. Thus, the time coincidence of signals from the positron and the neutron makes it possible to separate the event from backgrounds. The information collected by PMTs is used for estimation of the neutrino energy.
To resolve the neutrino mass ordering the energy resolution must be $\sigma \leqslant 3\%$ at 1 MeV, which is very close to the statistical limit corresponding to the light yield in JUNO, about 1300 detected photons (hits) at 1~MeV. The energy nonlinearity uncertainty should be < 1\%~\cite{JUNO}.
Machine Learning (ML) methods are very popular in science today, including high energy physics, in particular, neutrino experiments~\cite{Psihas} and collider experiments~\cite{Guest}. We use ML approach for energy reconstruction in the JUNO experiment. Our problem is a regression supervised learning problem. The data (time and charge information) collected by PMTs is used as input for supervised training of ML model. Earlier we demonstrated that the ML approach can have the quality required for the JUNO experiment on our data and also has the advantage of speed and ease of application~\cite{CNNs}.
In this work we use Boosted Decision Trees (BDT)~\cite{Friedman} for energy reconstruction in the energy range of 0–10 MeV covering the region of interest for IBD events from reactor electron antineutrinos. Compared to~\cite{CNNs} we designed and studied new features and achieved much better resolution with BDT, which is now comparable to the resolution of more complex models.
\section{Data description}
\label{DataDesc}
\begin{figure}[!htb]
\centering
\subfloat[]{\includegraphics[width=0.49\textwidth]{images/event_vis.pdf}\label{fig:data_vis_ch}}
\subfloat[]{\includegraphics[width=0.49\textwidth]{images/event_vis_ht.pdf}\label{fig:data_vis_ht}}
\caption{Example of an event seen by 17612 large ($20''$) PMTs for a positron event of 5.5 MeV. Only fired PMTs are shown. In fig (a) color represents the accumulated charge in PMTs: yellow points show the channels with more hits, red points — the channels with fewer hits. In fig (b) color indicates PMT activation time --- darker blue color shows an earlier activation. The primary vertex is shown by the gray sphere.}
\label{fig:data_vis}
\end{figure}
The dataset is generated by the full detector Monte Carlo method using the official JUNO software \cite{Huang}. The detector simulation is based on the Geant4 framework~\cite{Allison} with the geometry implemented in detail. The train and test datasets are described as follows:
\begin{enumerate}
\item \textbf{Training dataset} consists of 5 million events, uniformly distributed in kinetic energy from 0 to 10 MeV and in the volume of the central detector (in liquid scintillator).
\item \textbf{Testing dataset} consists of subsets with discrete kinetic energies of 0 MeV, 0.1 MeV, 0.2 MeV, ..., 1 MeV, 2 MeV, ..., 10 MeV. Each subset contains about 10 thousand events. This dataset is used to estimate performance after the end of training.
\end{enumerate}
Our data have four configurations: 1) without electronics effects;
2) taking into account the transit time spread (TTS) of PMTs;
3) taking into account the dark noise (DN) of PMTs; and 4) taking into account both effects.
TTS occurs due to the stochasticity of the path of photo-electron from the photo-cathode to the anode and effectively smears the time information. DN effect gives spontaneous hits on PMTs. In further TTS and DN are always enabled if not specified otherwise.
Figure~\ref{fig:data_vis} illustrates an example of accumulated charge in the PMT channels (left) and the evolution in time of the same signal in terms of the first hit time distribution (right).
\section{Boosted Decision Trees}
\label{BDT}
BDT is the an ensemble model, where a simple and quickly learning Decision Tree (DT) model is used as the base algorithm. DTs in BDT are trained sequentially. Each subsequent DT is trained to correct errors of previous DTs in the ensemble.
In this work we use the XGBRegressor implementation of BDT from the XGBoost library~\cite{XGB}.
DT is built recursively starting from the root node, splitting the source set into two subsets (left and right) based on the values of the input features. To build a tree, we need a principle based on which we will split the original set of objects into subsets. XGBRegressor uses Gain maximization to splitting input data into subsets.
In XGBoost the objective function contains a two parts: the training loss and the regularization term:
\begin{equation}\label{eq:obj}
\mathcal{L}(\phi) = \sum_{i}{l \left(\hat y_{i}, y_{i} \right)} + \sum_{k} {\Omega (f_k)}
\end{equation}
A tree is penalized if the sum of the norm of values in its leaves is very large. Therefore, the regularization term is introduced here as follows:
\begin{equation}
\Omega(f) = \gamma T + \frac{1}{2}\lambda \sum^{T}_{j=0}{\omega^2_j},
\end{equation}
where $T$ is the number of leaves, $\omega_j$ are values in the leaves, $\gamma$ and $\lambda$ are numerical parameters of the regularization. In~\cite{XGB} the authors showed that the optimization of an objective function \eqref{eq:obj} reduces to maximizing of the Gain. And Gain is defined as:
\begin{equation}
{\rm Gain} = \frac{1}{2}\left[ \frac{G_l^2}{H_l^2 + \lambda} + \frac{G_r^2}{H_r^2 + \lambda} - \frac{(G_l + G_r)^2}{H^2_l + H^2_l + \lambda}\right] - \gamma,
\end{equation}
where $G$, $H$ are the corresponding sums of the first and second derivatives of the objective function for a given partition and the indices $l$ and $r$ mean the left and right partition.
\section{Feature Engineering}
\label{FeatEng}
The basic features for the energy reconstruction are the following aggregated features:
\begin{enumerate}
\item Total number of detected photo-electrons (hits): \texttt{nHits}.
In the first approximation, the total number of hits is proportional to the event energy.
\item Coordinate components of the center of charge:
\begin{equation}
(x_{cc},\ y_{cc},\ z_{cc}) = \mathbf{r}_{\rm cc} = \frac{\sum_i^{N_{\rm PMTs}} \mathbf{r}_{{\rm PMT}_i} n_{{\rm p.e.}, i}}{\sum_i^{N_{\rm PMTs}}n_{{\rm p.e.}, i}},
\end{equation}
and its radial component:
\begin{equation}
R_{cc} = \sqrt{x_{cc}^2 + y_{cc}^2 + z_{cc}^2}.
\end{equation}
Coordinate components of the center of charge are rough approximations of the location of the energy deposition. These features are important for energy reconstruction since the number of hits depends on the location of the energy deposition.
\item Coordinate components of the center of first hit time:
\begin{equation}
(x_{cht},\ y_{cht},\ z_{cht}) = \mathbf{r}_{\rm cht} = \frac{1}{\sum_i^{N_{\rm PMTs}} \frac{1} {t_{{\rm ht},i} + c}} \sum_i^{N_{\rm PMTs}} \frac{\mathbf{r}_{{\rm PMT}_i}} {t_{{\rm ht},i} + c},
\end{equation}
and its radial component:
\begin{equation}
R_{cht} = \sqrt{x_{cht}^2 + y_{cht}^2 + z_{cht}^2}.
\end{equation}
Here the constant $c$ is required to avoid division by zero.
These features bring extra information on the location of the energy deposition.
\item Mean and standard deviation of the first hit time distributions: \texttt{ht\_mean}, \texttt{ht\_std}.
\end{enumerate}
For ML models including Boosted Decision Trees, it is often useful to engineer new features from the existing features~\cite{Heaton}. We use the following extra synthetic features:
\begin{gather}
\gamma_{z}^{cc} = \frac{z_{cc}}{\sqrt{x_{cc}^2 + y_{cc}^2}},\
\gamma_{y}^{cc} = \frac{y_{cc}}{\sqrt{x_{cc}^2 + z_{cc}^2}},\
\gamma_{x}^{cc} = \frac{x_{cc}}{\sqrt{z_{cc}^2 + y_{cc}^2}}; \\
\theta_{cc} = \arctan{\frac{\sqrt{x_{cc}^2 + y_{cc}^2}}{z_{cc}}},\
\phi_{cc} = \arctan{\frac{y_{cc}}{x_{cc}}}; \\
J_{cc} = R_{cc}^2 \cdot \sin{\theta_{cc}},\
\rho_{cc} = \sqrt{x_{cc}^2 + y_{cc}^2}.
\end{gather}
And some trigonometric functions of angles $\theta_{cc}$, $\phi_{cc}$: $ \sin{\theta_{cc}},\ \cos{\theta_{cc}},\ \sin{\phi_{cc}},\ \cos{\phi_{cc}}$. We also use 11 similar features for the center of first hit time.
In addition, we prepare five more features related to the location of the PMT received the maximum number of photo-electrons: \texttt{x\_max}, \texttt{y\_max}, \texttt{z\_max}, \texttt{theta\_max}, \texttt{phi\_max}, the maximum number of photons on PMT \texttt{npe\_max}, and the average number of photons on PMTs \texttt{npe\_mean}.
Also we added the following features: \texttt{entries1}, \texttt{entries2}. Here, \texttt{entries1} is the percentage of PMTs with only 1~hit, \texttt{entries2} --- with 2~hits. And the one more feature is \texttt{nPMTs} --- the total number of fired PMTs.
Now let's take a closer look at the first hit time distribution. Consider what fraction of fired PMTs received at least one photon depending on time, which is, in essence, the cumulative distribution function (CDF) of the first hit time distribution. Figure~\ref{fig:cdf_and_pdf_ht} illustrates an example of a 7~MeV event. The entire event typically lasts for about \num{1000}~ns, but the majority of photo-electrons are recorded by the PMTs in the first hundred nanoseconds and then mainly dark hits are recorded.
In Figure~\ref{fig:cdf_and_pdf_ht} one also can see that at the beginning of events there is a short period of time $\Delta t$ during which the photons have not reached the PMTs and only dark noise is recorded.
Figure~\ref{fig:cdf_and_pdf_ht_diff_R} shows how the CDFs for the events with the energy of 7~MeV change depending on their location, closer to the edge (large R) or closer to the center of the detector (small R).
In the case where R is quite small, it takes more time for the photons to be detected by PMTs and also it takes more time to ``saturate''.
\begin{figure}[!htb]
\centering
\subfloat[Example for a specific R.]{\includegraphics[width=0.49\textwidth]{images/cdf_and_pdf_ht.pdf}\label{fig:cdf_and_pdf_ht}}
\subfloat[Example for different R.]{\includegraphics[width=0.49\textwidth]{images/cdf_and_pdf_ht_diff_R.pdf}\label{fig:cdf_and_pdf_ht_diff_R}}
\caption{Examples of CDFs and PDFs (bottom right) of the first hit time distribution for events with energy equal to 7~MeV.}
\label{fig:ht_profiles}
\end{figure}
Finally, the idea is to simply decompose the entire curve into a set of percentiles, and then select those that suit best for energy reconstruction. X\%-percentile indicates how long it takes to register X\% of the first PMT hits. We use the following set of percentiles: \{1\%, 2\%, ..., 10\%, 15\%, ..., 90\%, 91\%, ..., 99\%\}.
\section{Selection of event time window}
\label{EventTimeWinSel}
One can also see in Figure~\ref{fig:ht_profiles} that the signal hits arrive within the first few hundred nanoseconds, while the dark hits form quasi-constant pedestal. Therefore, a time window can be selected, based on the data, in a way to contain mainly signal events. For this purpose we trained Boosted Decision Trees model with different bounds of window: \{75ns, 125ns, 175ns, 250ns, 500ns, 750ns, 1500ns\} and always started from $t=0$. It was trained on a 200k dataset and using all new features.
Table~\ref{tab:bounds_results} shows the Mean Absolute Error (MAE), the Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE) on the test dataset for different window bounds. The best one was found to be 500~ns, however all the windows longer than 175~ns showed similar performance.
\begin{table}[!htb]
\centering
\caption{MAE, RMSE, MAPE metrics for BDT models for different window bounds.}
\label{tab:bounds_results}
\begin{tabular}{llll}
\hline
Bound, ns& MAE, MeV & RMSE, MeV & MAPE, \% \\
\hline
75 & 0.1019 & 0.1664 & 3.705 \\
125 & 0.0562 & 0.0781 & 1.858 \\
175 & 0.0505 & 0.0698 & 1.662 \\
250 & 0.0487 & 0.0676 & 1.594 \\
\textbf{500} & \textbf{0.0477} & \textbf{0.0660} & \textbf{1.569} \\
750 & 0.0480 & 0.0664 & 1.585 \\
1500 & 0.0479 & 0.0662 & 1.589 \\
\hline
\end{tabular}
\end{table}
\section{Feature selection}
\label{FeatSel}
Finally, we have a large set of features, but many of them are highly correlated, so we expect that a small set of them contain all the information and provide a performance close to the best possible one. Thus, the next task is to get a subset of the most informative features from the available set of all 78 features. For this purpose we use a dataset with 1M events.
To select the most informative features, we use the following algorithm. First, we train a model on all features and computed RMSE on the validation dataset with 150k events. Then we take an empty list and start populating it with features. On each step we pick a feature which provide the best improvement of the model in terms of RMSE calculated on the validation dataset and put it to the end of the list. We continue while this RMSE value differs from the RMSE value for the model trained on all features by more than $\varepsilon$, chosen to be 0.0002. This procedure results to the following set of features (sorted by importance): \[\texttt{nHits},\ \texttt{ht\_20p},\ \texttt{jacob\_cc},\ \texttt{ht\_2p},\ \texttt{ht\_35p},\ \texttt{R\_cc},\ \texttt{ht\_75p}\]
Not surprisingly, for the energy reconstruction, the most informative feature is the total number of hits \texttt{nHits}, because its strongly correlated with energy, but at the same time it is hard to interpret the order of the rest features. The subset of the selected features contains \texttt{jacob\_cc} and \texttt{R\_cc}, which bring the spatial information allowing to recover the non-uniformity of detector response.
We checked simpler combinations of features for the center of charge position. It turned out that the combination of \texttt{rho\_cc} and \texttt{R\_cc} gives the same result as \texttt{jacob\_cc} and \texttt{R\_cc}, so we have chosen them as they are more intuitive. Our final set of features is: \[\texttt{nHits},\ \texttt{ht\_20p},\ \texttt{rho\_cc},\ \texttt{ht\_2p},\ \texttt{ht\_35p},\ \texttt{R\_cc},\ \texttt{ht\_75p}\]
Figure~\ref{fig:percentiles_vis} illustrates the selected percentiles of the CDFs for an event with energy of 7~MeV and for different radial positions. As one can see \texttt{ht\_2p} contains information about the beginning of the event, that is about the moment when the number of photons PMT hits begins to grow sharply. The remaining percentiles contain information about the shape of the CDF curve and help us to separate one curve from another. The 75\% percentile is close to the moment of ``saturation'' of the CDF curve.
\begin{figure}[!htb]
\centering
\includegraphics[width=\textwidth]{images/percentiles_vis.pdf}
\caption{The selected percentiles of the CDFs of the first hit time distributions for a 7~MeV event for different radial positions: 2\% (top left), 20\% (top right), 35\% (bottom left) and 75\% (bottom right).}
\label{fig:percentiles_vis}
\end{figure}
\section{Results}
\label{Res}
To evaluate the quality of the model, we use two metrics: resolution and bias. These metrics are obtained as a result of the Gaussian fit of $E_{\rm pred} - E_{\rm true}$ distribution. The resolution is defined as $\sigma / E_{\rm true}$ and the bias --- as $\mu / E_{\rm true}$, where $\sigma$ and $\mu$ are the standard deviation and the mean of the Gaussian distribution respectively. The performance is shown dependent on the so called visible energy, i.e.\ the maximal energy that can be converted into light: $E_{\rm vis} = E_{\rm kin} + 1.022 \text{ MeV}$. This procedure is described in more details in~\cite{CNNs}.
Figure~\ref{fig:res_and_bias_diff_sizes_of_datasets} illustrates the results for BDT models trained on datasets that contain different amount of events: 100k, 1M, 5M. We obtained that 1M events can provide the best possible accuracy of the model, providing only a little improvement compared to 100k events. This illustrates that fast learning is one of the advantages of the BDT model: one can get an acceptable quality already on a relatively small number of events in the dataset.
Figure~\ref{fig:res_and_bias_diff_options} shows a comparison for the BDT model trained on the 5M dataset for different options: without TTS \& DN, with TTS only, with DN only, with TTS \& DN. One can see that DN worsens the resolution, TTS --- almost does not.
A comparison of the BDT model with other more complex deep learning models (ResNet, VGG, and GNN)~\cite{CNNs} is shown in Figure~\ref{fig:res_and_bias_all_models}. All the models are trained on the dataset with 5M events. We can see that the performance of the energy reconstruction with BDT model is practically similar to the complex deep learning models. At the same time the computations required for training and prediction are much faster due to the minimalistic nature of BDT.
\begin{figure}[!htb]
\centering
\subfloat[Comparision of different sizes of datasets.]{\includegraphics[width=0.48\textwidth]{images/res_and_bias_diff_sizes_of_datasets}\label{fig:res_and_bias_diff_sizes_of_datasets}}
\hspace{1em
\subfloat[Comparison of different TTS \& DN options for a dataset with 5M events.]{\includegraphics[width=0.48\textwidth]{images/res_and_bias_diff_options}\label{fig:res_and_bias_diff_options}}
\caption{Results of the energy reconstruction for the BDT model: resolution (upper panel) and bias (lower panel). Note that the first point corresponds to 1.122 MeV.}
\label{fig:results_bdt}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.7\textwidth]{images/res_and_bias_all_models}
\caption{Energy reconstruction performance: resolution (upper panel) and bias (lower panel) obtained with BDT, ResNet-J, VGG-J and GNN-J models. The plots are offset along X-axis within $\pm$0.06 MeV for better readability. Note that the first point corresponds to 1.122 MeV.}
\label{fig:res_and_bias_all_models}
\end{figure}
\section{Summary}
\label{Sum}
In this work we have presented the use of Boosted Decision Trees for energy reconstruction in the JUNO experiment in the relevant energy range. We have designed and investigated a large set of features and have selected a small subset providing the performance nearly equal the one obtained with the full set of features. Using such a minimalistic and fast model as BDT we achieved a performance similar to the one of more complex models like ResNet, VGG, GNN.
\section*{Acknowledgements}
\begin{acknowledgement}
We are immensely grateful to Yury Malyshkin for his invaluable contribution to this work. We would like to thank Weidong Li, Jiaheng Zou, Tao Lin, Ziyan Deng, Guofu Cao and Miao Yu for their tremendous contribution to the development of JUNO offline software and to Xiaomei Zhang and Jo\~ao Pedro Athayde Marcondes de Andr\'e for production of the MC samples. We are grateful to N.~Kutovskiy, N.~Balashov for providing an extensive IT support and computing resources of JINR cloud services~\cite{Baranov}.
Fedor Ratnikov is supported by the Russian Science Foundation under grant agreement \textnumero{17-72-20127}.
\end{acknowledgement}
|
\section{Introduction}
\label{sec-intro}
Recently, transformer models have been introduced into the field of computer vision and achieved high performance in many tasks such as object recognition~\cite{dosovitskiy2020image}, image process~\cite{chen2020pre}, and video analysis~\cite{kim2018spatio}. Compared with the convolutional neural networks~(CNNs), the transformer architecture introduces less inductive biases and hence has larger potential to absorb more training data and generalize well on more diverse tasks~\cite{dosovitskiy2020image,touvron2020training,liu2021post,yuan2021tokens,tang2021augmented,han2022pyramidtnt}. However, similar to CNNs, vision transformers also suffer high computational cost, which blocks their deployment on resource-limited devices such as mobile phones and various IoT devices. To apply a deep neural network in such real scenarios, massive model compression algorithms have been proposed to reduce the required computational cost~\cite{ye2018rethinking,li2016pruning}. For example, quantization algorithms approximate weights and intermediate features maps in neural networks with low-bit data~\cite{courbariaux2016binarized,rastegari2016xnor}. Knowledge distillation improves the performance of a compact network by transferring knowledge from giant models~\cite{hinton2015distilling,lan2018knowledge}.
In addition, network pruning is widely explored and used to reduce the neural architecture by directly removing useless components in the pre-defined network~\cite{han2015learning, han2015deep, liu2017learning,liu2018frequency}. Structured pruning discards whole contiguous components of a pre-trained model, which has attracted much attention in recent years, as it can realize acceleration without specific hardware design. In CNNs, removing a whole filter for improving the network efficiency is a representative paradigm, named channel pruning (or filter pruning)~\cite{he2017channel,liu2017learning}. For example, Liu~\textit{et al.}~\cite{liu2017learning} introduce scaling factors to control the information flow in the neural network and filters with small factors will be removed. Although the aforementioned network compression methods have made tremendous efforts for deploying compact convolutional neural networks, there are only few works discussing how to accelerate vision transformers.
Different from the paradigm in conventional CNNs, the vision transformer splits the input image into multiple patches and calculates the features of all these patches in parallel. The attention mechanism will further aggregate all patch embeddings into visual features as the output. Elements in the attention map reflect the relationship or similarity between any two patches, and the largest attention value for constructing the feature of an arbitrary patch is usually calculated from itself. Thus, we have to preserve this information flow in the pruned vision transformers for retaining the model performance, which cannot be guaranteed in the conventional CNN channel pruning methods. Moreover, not all the manually divided patches are informative enough and deserve to be preserved in all layers, \eg, some patches are redundant with others. Hence we consider developing a patch slimming approach that can effectively identify and remove redundant patches.
In this paper, we present a novel patch slimming algorithm for accelerating the vision transformers. In contrast to existing works focusing on the redundancy in the network channel dimension, we aim to explore the computational redundancy in the patches of a vision transformer~(as shown in Figure~\ref{fig-patch}. The proposed method removes redundant patches from the given transformer architecture in a top-down framework, in order to ensure the retained high-level features of discriminative patches can be well calculated. Specifically, the patch pruning will execute from the last layer to the first layer, wherein the useless patches are identified by calculating their importance scores to the final classification feature (\ie, class token). To guarantee the information flow, a patch will be preserved if the patches in the same spatial location are retained by deeper layers. For other patches, the importance scores determine whether they are preserved, and patches with lower scores will be discarded. The whole pruning scheme for vision transformers is conducted under a careful control of the network error, so that the pruned transformer network can maintain the original performance with significantly lower computational cost. Extensive experiments validate the effectiveness of the proposed method for deploying efficient vision transformers. For example, our method can reduce more than 45\% FLOPs of the ViT-Ti model with only 0.2\% top-1 accuracy loss on the ImageNet dataset.
\begin{figure}
\centering
\small
\includegraphics[width=0.7\columnwidth]{patch_pr}
\caption{The diagram of patch slimming for vision transformers.}
\label{fig-patch}
\end{figure}
\section{Related work}
\textbf{Structure pruning for CNNs. }
Channel pruning discards the entire convolution kernels~\cite{lebedev2016fast} to accelerate the inference process and reduce the required memory cost~\cite{lebedev2016fast,DBLP:conf/iclr/SuYH00Z021,liu2017learning,he2019filter,su2021bcnet}. To identify the redundant filters, massive methods have been proposed. Wen~\textit{et al.}~\cite{wen2016learning} add a group-sparse regularization on the filters and remove filters with small norm. Beyond imposing sparsity regularization on the filters directly, Liu~\textit{et al.}~\cite{liu2017learning} introduce extra scaling factors to each channel and these scaling factors are trained to be sparse. Filters with small scaling factors has less impact on the network output and will be removed for accelerating inference. He~\textit{et al.}~\cite{he2019filter} rethink the criterion that filters with small norm values are less important and propose to discard the filters having larger similarity to others. To maximally excavate redundancy, Tang~\cite{tang2020scop} set up a scientific control to alleviate the distribution of irrelevant factors and remove filters with little relation to the given task. In the conventional channel pruning for CNNs, channels in different layers have no one-to-one relationship, and then the choice of effective channels in a layer has little impact on that in other channels.
\textbf{Structure pruning for transformers. }
In the transformer model for NLP tasks, a series of works focus on reducing the heads in the multi-head attention (MSA) module. For example, Michel~\textit{et al.}~\cite{NEURIPS2019_2c601ad9} observes that removing a large percentages of heads in the pre-trained BERT~\cite{devlin2018bert} models has limited impact on its performance. Voita~\textit{et al.}~\cite{voita2019analyzing} analyze the role of each head in the transformer and evaluate their contribution to the model performance. Those heads with less contributions will be reduced. Besides the MSA module, the neurons in the multilayer perceptron (MLP) module are also pruned in \cite{brown2020language}. Designed for vision transformers, VTP~\cite{zhu2021visual} reduces the number of embedding dimensions by introducing control coefficients and removes neurons with small coefficients. Different from them, the proposed patch slimming explores the redundancy from a new perspective by considering the information integration of different patches in a vision transformer. Actually, reducing patches can be also combined with pruning in other dimensions to realize higher acceleration.
\section{Patch Slimming for Vision Transformer}
In this section, we introduce the scheme of pruning patches in vision transformers. We first review the vision transformer briefly and then introduce the formulation of patch slimming.
In vision transformer, the input image is split into $N$ patches and then fed into transformer model for representation learning. For an $L$-layer vision transformer model, the multihead self-attention~(MSA) modules and multi-layer perceptron (MLP) modules are its main components occupying most of the computational cost. Denoting $Z_{l-1}, Z'_l \in \mathbb{R}^{N\times d}$ as the input and the intermediate features of the $l$-th layer, the MSA and MLP modules can be formulated as:
\begin{equation}
\small
\label{eq-blk}
\begin{aligned}
&{\rm MSA}(Z_{l-1})\\
&\quad = {\rm Concat }\left[ {\rm softmax }\left( \frac{Q_l^h {K_l^h}^{\top}}{\sqrt{d}} \right)V_l^h \right]_{h=1}^H W_l^o, \\
&{\rm MLP} (Z'_l)= \phi (Z'_lW_l^a)W_l^b,
\end{aligned}
\end{equation}
where $d$ is embedding dimension, $H$ is the number of heads, $Q_l^h=Z_{l-1}W^{hq}_l$, $K_l^h=Z_{l-1}W^{hk}_l$, and $V_l^h=Z_{l-1}W^{hv}_l$ are the query, key and value of the $h$-th head in the $l$-th layer, respectively. $W_l^a$, $W_l^b$ are the weights for linear transformation and $\phi(\cdot)$ is the non-linear activation function~(\eg, GeLU). Most of recent vision transformer models are constructed by stacking MSA and MLP modules alternately and a block $\mathcal{B}_l (\cdot)$ is defined as $\mathcal{B}_l (Z_{l-1})={\rm MLP}({\rm MSA}(Z_{l-1})+Z_{l-1})+Z'_l$.
As discussed above, there is considerable redundant information existing in the patch level of vision transformers. To further verify this phenomenon, we calculate the average cosine similarity between patches within a layer, and show how similarity vary \wrt layers in Figure~\ref{fig-sim}. The similarity between patches increase rapidly as layers increase, and the average similarity even exceed 0.8 in deeper layers. The high similarity implies that patches are redundant especially in the deeper layers and removing them will not obviously affect the feature calculation.
Patch slimming aims to recognize and discard redundant patches for accelerating the inference process~(as shown in Figure~\ref{fig-patch}). Here we use a binary vector ${\bm{m}}_l \in \{0,1\}^N$ to indicate whether a patch is preserved or not, the pruned MSA and MLP modules can be formulated as follows:
\begin{equation}
\small
\label{eq-mblk}
\begin{aligned}
&\widehat {\rm MSA}_l(\widehat{Z}_{l-1})
\\&\ = {\rm Concat }\left[ {\mbox{diag}}({\bm{m}}_l) {\rm softmax }\left( \frac{{Q}_l^h {K_l^h}^{\top}}{\sqrt{d}} \right) {V}_l^h \right]_{h=1}^H W_l^o,\\
&\widehat {\rm MLP}_l (\widehat{Z}'_l)= {\mbox{diag}}({\bm{m}}_l) \phi (\widehat{Z}'_lW_l^a)W_l^b,
\end{aligned}
\end{equation}
where ${\mbox{diag}}({\bm{m}}_l)$ is a diagonal matrix whose diagonal line is composed of elements in ${\bm{m}}_l$. Specifically, ${\bm{m}}_{l,i}=0$ indicates that the $i$-th patch in the $l$-th layer is pruned. $\widehat{Z}_{l-1}$, $\widehat{Z}'_l$ are the input and the intermediate features of the $l$-th layer in a pruned vision transformer.
Then the pruned block is defined as $\widehat{\B}_l(\widehat{Z}_{l-1})=\widehat {\rm MLP}_l(\widehat {\rm MSA}_l(\widehat{Z}_{l-1})+ \widehat{Z}_{l-1})+\widehat{Z}'_l$.
In practical implementation, only the effective patches of input feature $\widehat{Z}_{l-1}$ are selected to calculate queries, and then all the subsequent operations are only implemented on these effective patches.
Thus, the computation of the pruned patches can be avoided \footnote{According to ${\bm{m}}_l$, only effective patches from the shortcut branch are added to the output of pruned MSA, while the output of pruned MLP is padded with zeros before added to the shortcut.}.
\textbf{Computation Efficiency.} Compared with the original block $\mathcal{B}_l(\cdot)$, the pruned $\widehat{\B}_l(\cdot)$ can save a large amount of computational cost. Given a block $\mathcal{B}(\cdot)$ with $N$ patches and $d$-dimension embedding, the computational costs of MLP~(2-layers with hidden dimension $d'$) and MSA are $(2Ndd')$ and $(2N^2d+4Nd^2)$, respectively. After pruning $\eta\%$ patches, all the computational components in MLP are pruned, and then $\eta\%$~FLOPs in the MLP module are reduced. For the MSA module, the cost of calculating query, attention map and output projection can be reduced, and then $\eta\%(2N^2d+ 2Nd^2)$ FLOPs is reduced.
\section{Excavating Redundancy via Inverse Pruning}
In this section, we present the top-down framework to prune patches in the vision transformer, and provide an effective importance score estimation of each patch.
\begin{figure}[htp]
\centering
\small
\includegraphics[width=1.0\columnwidth]{cnn_vit}
\caption{The comparison between channel pruning in CNNs and patch pruning in vision transformers.}
\label{fig-cvt}
\end{figure}
\begin{figure}[htp]
\centering
\small
\includegraphics[width=0.6\columnwidth]{sim}
\caption{The average similarity of different patches varies \wrt network depth in the ViT-Base model.}
\label{fig-sim}
\vspace{-2mm}
\end{figure}
\subsection{Top-Down Pruning}
\label{sec-proc}
For patch slimming in vision transformer, we adopt a top-down manner to prune patches layer-by-layer. It is a natural choice with two reasons as described in the following.
For a CNN model, pruning channels in different layers independently can achieves high performance ~\cite{liu2017learning,tang2020scop}. However, this paradigm cannot work well in vision transformers. The main reason is that patches in different layers of a vision transformer are one-to-one corresponding.
Figure~\ref{fig-cvt} compares pruning channels in CNNs and pruning patches in vision transformers.
As own in Figure~\ref{fig-cvt}(a), channels in adjacent layers of a CNN model are fully connected by learnable weights, and each channel contains information from the entire image. However, in the vision transformer~(Figure~\ref{fig-cvt}(b)), different patches communicate with others by an attention map, which reflects the similarity between different patches. If patch $i$ and patch $j$ are more similar, the corresponding value $A_{lh}^{ij}$ tends to have a larger value. The diagonal elements $A_{lh}^{ii}$ usually plays a dominant role, that is, a patch pays highest attention to the input at the position of itself. Besides, the shortcut connection directly copies the feature in the $l$-layer to the corresponding patches in the next layer. This one-to-one correspondence inspires us to preserve some important patches in the same spatial locations of different layers, which can guarantee the information propagation across layers.
Another characteristic of vision transformer is that deeper layers tend to have more redundant patches. The attention mechanism in the MSA module aggregates different patches layer-by-layer, and a large number of similar patches are produced in the process (as shown in Figure~\ref{fig-sim}). It implies that more redundant patches can be safely removed in deeper layers, and fewer in shallower layers.
Based on the above analysis, we start the pruning procedure from the output layer, and then prune previous layers by transmitting the selected effective patches from top to down. Specially, all the patches preserved in the $(l+1)$-th layer will be also preserved in the $l$-th layer.
Thus, this top-down pruning procedure can guarantee that shallow layers maintain more patches than the deep layers, which is consistent with the redundancy characteristic of vision transformer.
\subsection{Impact Estimation}
\label{sec-impact}
With the patch pruning scheme described in the above section, all that's left is to recognize redundant patches in a vision transformer, \ie, find the optimal mask ${\bm{m}}_l$ in each layer. Our goal is to prune patches as many as possible to realize maximal acceleration, while maintaining the representation ability of the output feature. Actually, only a part of patch embeddings in the last layer are used to predict the labels of input images for a specific task. For example, in the image classification task, only a patch related to classification (\ie, class token) is sent to the classifier for predicting labels. Other patches in the output layer can be removed safely without affecting network output. Supposing the first patch is the class token, we can get the mask in the last layer, \ie, ${\bm{m}}_{L,1}=1$, and ${\bm{m}}_{L,i}=0, \forall~i=2,3,\cdots,N$. Then for the other layers, the optimization object is formulated as follows:
\begin{equation}
\begin{aligned}
\label{eq-obj}
&\min_{{\bm{m}}_1, {\bm{m}}_2, \cdots, {\bm{m}}_{L-1}} \sum_{l=1}^{L-1}\|{\bm{m}}_l\|_0,\\
&\quad \quad{\rm s.t.} \ {\bm{m}}_l\in\{0,1\}^N , \\
&\quad \quad\mathcal{E}_L=\norm{{\mbox{diag}}({\bm{m}}_L)\left(\widehat{Z}_L-Z_L\right)}_F^2\le \epsilon,
\end{aligned}
\end{equation}
where $\|\cdot\|_0$ is the $\ell_0$-norm of a vector, \ie, the number of non-zero elements. $\|\cdot \|_F$ is the Frobenius norm of a matrix. $\epsilon$ is the tolerable error. $\widehat{Z}_L$ and $Z_L$ are output features of the pruned and unpruned transformers. Eq.~\ref{eq-obj} is hard to optimize directly, as it involve $\ell_0$ optimization under constraint, which is non-convex, NP hard and requires combinatorial search~\cite{liu2017learning}. To solve Eq.~\ref{eq-obj}, we firstly define a signification score by approximating the impact of a patch on the reconstruction error $\mathcal{E}_L$ and then develop a pruning procedure.
The attention mechanism aggregates information from different patches to one patch, which is the main cause to produce redundant patch features. To focus on the attention layer for excavating redundant patches, we reformulate the definition of a block $\mathcal{B}(\cdot)$ in a simple formulation. Denoting $P_l^h= {\rm softmax }\left( {Q^{h}_l{K^{h}_l}^{\top}}/{\sqrt{d}}\right)$, the
MSA module in Eq.~\ref{eq-blk} can be formulated as:
\begin{equation}
\small
\begin{aligned}
{\rm MSA}(Z_l)& = {\rm Concat }\left [P_l^h V_l^h\right]_{h=1}^H W^{o}_l \\
&= \sum_{h=1}^H P_l^h V_l^h W^{ho}_l
= \sum_{h=1}^H P_l^h Z_{l-1} W_l^{hv} W^{ho}_l,
\end{aligned}
\end{equation}
where $W^{o}_l=[W^{1o}_l;W^{2o}_l; \cdots; W^{Ho}_l]$, and $W^{ho}_l\in\mathbb{R}^{\frac{d}{H}\times d}$.
Then the original block $\mathcal{B}_l(Z_{l-1})$ and pruned block $\widehat{\B}_l(Z_{l-1})$ can be represented as:
\begin{equation}
\small
\label{eq-sblk}
\begin{aligned}
&\mathcal{B}_l(Z_{l-1})=\O(\sum_{h=1}^HP_l^hZ_{l-1},\{W_l\}), \\ &\widehat{\B}_l(Z_{l-1},{\bm{m}}_l)=\O(\sum_{h=1}^H{\mbox{diag}}({\bm{m}}_l)P_l^hZ_{l-1},\{W_l\})
\end{aligned}
\end{equation}
where $\O(\cdot,W_l)$ is composed of multiple linear projection matrices $\{W_l\}$ in the MSA and MLP module, as well as non-linear activation functions (\eg, GeLU).
Based on the simplified formulation of a block~(Eq.~\ref{eq-sblk}), we here explore how a patch in the $t$-th layer affects the error $\mathcal{E}_L$~(Eq~\ref{eq-obj}) of effective patches in the last layer. We reverse the transformer and prune it from the last to the first layer sequentially. Thus when it comes to the $t$-th layer, all the deeper layers have been pruned. To approximate the significance of each token, we have the following theorem.
\begin{theorem}
The impact of the $t$-th layer's patch on the final error $\mathcal{E}_L$ can be reflected by a significance metric ${\bm{s}}_t\in\mathbb{R}^N$. For the $i$-th patch in the $t$-th layer, we have
\begin{equation}
\label{eq-imp}
{\bm{s}}_{t,i}=\sum_{h\in [H]^{L\sim t+1}} \norm{A_t^h[:,i] U_t^h[i,:]}_F^2,
\end{equation}
where $A_t^h=\prod_{l=t+1}^L{\mbox{diag}}({\bm{m}}_l)P_l^h $ and $U_t^h=P_t^h \abs{Z_{t-1}}$. $A_t^h[:,i]$ denotes the $i$-th column of $A_t^h$ and $U^h_t[i,:]$ is $i$-th row of $U_t^h$. $[H]^{L\sim t+1}$ denotes all the attention heads in the $(t+1)$-th to $L$-th layer.
\end{theorem}
\begin{proof}
We use $\widehat{\F}_{l\sim t}(Z_{t-1},\{{\bm{m}}_t\}_t^L)~(l>t)$ to denote feature of the $l$-th layer in a vision transformer, whose layers behind $t$-th layer have been pruned, while the previous layers has not pruned yet, \ie, $\widehat{\F}_{L\sim t}(Z_{t-1},\{{\bm{m}}_t\}_t^L) = \hat \mathcal{B}_L \circ \hat \mathcal{B}_{L-1} \circ \cdots \circ \hat \mathcal{B}_{t}(Z_{t-1})$.
When pruning the patch in the $t$-th layer, we compare effective patches of the last layer from two transformers to decide whether the $t$-th layer has been pruned. Then the error $\mathcal{E}_L$ is calculated as:
\begin{equation}
\small
\begin{aligned}
\mathcal{E}_L &=||{\mbox{diag}}({\bm{m}}_L)[\widehat{\F}_{L\sim (t+1)}(\widehat{\B}_t(Z_{t-1})) \\
&- \widehat{\F}_{L\sim (t+1)}(\mathcal{B}_t(Z_{t-1}))]||_F^2.
\end{aligned}
\end{equation}
The error $\mathcal{E}_L$ in the last layer can be represented by the patches in the $(L-1)$-th layer, \ie,
\begin{equation}
\small
\label{eq-el1}
\begin{aligned}
\mathcal{E}_L& = ||{\mbox{diag}}({\bm{m}}_L) P_{L}^h [\O(\sum_{h=1}^H \widehat{\F}_{(L-1)\sim (t+1)}(\widehat{\B}_t(Z_{t-1})))\\
&-\O( \sum_{h=1}^H \widehat{\F}_{(L-1)\sim (t+1)}(\mathcal{B}_t(Z_{t-1})) ) ] ||_F^2 \\
&\le C_L || \sum_{h=1}^H {\mbox{diag}}({\bm{m}}_L) P_{L}^h |\widehat{\F}_{(L-1)\sim (t+1)}(\widehat{\B}_t(Z_{t-1}))\\
&- \widehat{\F}_{(L-1)\sim (t+1)}(\mathcal{B}_t(Z_{t-1}))|||_F^2,
\end{aligned}
\end{equation}
where $|\cdot|$ is the element-wisely absolute value. The inequality above comes the Lipschitz continuity~\cite{funahashi1993approximation, dupuis1991lipschitz} of function $\O(\cdot)$ and $C_L$ is the Lipschitz constant. Recalling that $\O(\cdot)$ is compose of multiple linear projections and non-linear activation function, the condition of Lipschitz continuity is satisfied~\cite{funahashi1993approximation}. $E_L$ can be further transmitted to previous layers, and for the $t$-th layer we have
\begin{align}
\small
\mathcal{E}_L & \le \prod_{l=t+1}^L C_l || \sum_{h\in [H]^{L\sim t}} \prod_{l=t+1}^L{\mbox{diag}}({\bm{m}}_l)P_l^h \\
& \quad |\widehat{\B}_t(Z_{t-1})-\mathcal{B}_t(Z_{t-1})| ||_F^2 \\
& \le \prod_{l=t}^L C_l || \sum_{h\in [H]^{L\sim t+1}} \prod_{l=t+1}^L{\mbox{diag}}({\bm{m}}_l) \\
& \quad P_l^h \left (I_N-{\mbox{diag}}({\bm{m}}_l)\right)P_t^h \abs{Z_{t-1}} ||_F^2 \\
&=C'_t ||\sum_{h\in [H]^{L\sim t+1}}A_t^h (I_N -{\mbox{diag}}({\bm{m}}_t)) U_t^h ||_F^2, \label{eq-ela}
\end{align}
where $A_t^h=\prod_{l=t+1}^L{\mbox{diag}}({\bm{m}}_l)P_l^h \in \mathbb{R}^{N\times N}$, $U_t^h=P_t^h \abs{Z_{t-1}} \in \mathbb{R}^{N\times d}$, and $C'_t=\prod_{l=t}^L C_l$. $[H]^{l\sim t}$ denotes all the attention heads in the $t$-th to $l$-th layer. To investigate how each patch in the $t$-th layer affect the final error $E_L$, we expand Eq.~\ref{eq-ela} \wrt each element in the indicator ${\bm{m}}_l$. Denoting ${\bm{m}}_{l,i}$ as the $i$-th element in ${\bm{m}}_l$, $A_t^h[:,i]$ is the $i$-th column of $A_t^h$ and $U^t_t[i,:]$ is $i$-th row of $U_t^h$, Eq.~\ref{eq-ela} can be written as:
\begin{align}
\small
\mathcal{E}_L & \le C'_t ||\sum_{h\in [H]^{L\sim t+1}} \sum_{i=1}^N A_t^h[:,i] (1-{\bm{m}}_{t,i}) U_t^h[i,:] ||_F^2 \\
& \le C'_t \sum_{i=1}^N(1-{\bm{m}}_{t,i})\sum_{h\in [H]^{L\sim t+1}} \norm{A_t^h[:,i] U_t^h[i,:]}_F^2 \label{eq-ele}.
\end{align}
Then we get the importance of each patch, \ie, $ {\bm{s}}_{t,i}=\sum_{h\in [H]^{L\sim t+1}} \norm{A_t^h[:,i] U_t^h[i,:]}_F^2$.
\end{proof}
\begin{table*}[t]
\centering
\small
\caption{Comparison of the pruned vision transformers with different methods on ImageNet. `FLOPs~$\downarrow$' denotes the reduction ratio of FLOPs.}
\vspace{-2mm}
\begin{tabular}{c|c|c|c|c|c|c|c}
\toprule[1.5pt]
\multirow{2}{*}{Model} & \multirow{2}{*}{Method} & {Top-1 } &Top-5 & FLOPs & FLOPs &Throughput& Throughput \\
&&Acc. (\%)&Acc. (\%)&(G)&$\downarrow$ (\%)&(image / s)& $\uparrow$ (\%)\\
\hline
\multirow{6}{*}{ViT~(DeiT)-Ti}
& Baseline& 72.2&91.1& 1.3 & 0& 2536&0\\
&SCOP~\cite{tang2020scop} &68.9 (-3.3)&89.0 (-2.1)&0.8&38.4&3372&33.0 \\
&PoWER~\cite{goyal2020power}&69.4 (-2.8)&89.2 (-1.9)&0.8&38.4&3304&30.3\\
&HVT~\cite{pan2021scalable} & 69.7 (-2.5) & 89.4 (-1.7) & 0.7 & 46.2&3524&38.9\\
& PS-ViT (Ours) &\textbf{72.0} (-0.2)&\textbf{91.0} (-0.1)&0.7&46.2&3576&41.0\\
& DPS-ViT (Ours) &\textbf{72.1} (-0.1)&\textbf{91.1} (-0.0) &0.6&53.8&3639&43.5 \\
\hline
\multirow{6}{*}{ViT~(DeiT)-S}
& Baseline& 79.8&95.0& 4.6 & 0&940&0\\
&SCOP~\cite{tang2020scop} &77.5 (-2.3)&93.5 (-1.5)&2.6&43.6&1310&39.4 \\
&PoWER~\cite{goyal2020power} &78.3 (-1.5)&94.0 (-1.0)& 2.7 & 41.3&1295&37.8\\
&HVT~\cite{pan2021scalable} & 78.0 (-1.8) & 93.8 (-1.2) & 2.4 & 47.8&1335&42.1 \\
&PS-ViT (Ours) & \textbf{79.4} (-0.4) & \textbf{94.7} (-0.3) & 2.6 & 43.6&1321&40.5 \\
&DPS-ViT (Ours) & \textbf{79.5} (-0.3) & \textbf{94.8} (-0.2) & 2.4 & 47.8&1342&42.8 \\
\hline
\multirow{6}{*}{ViT~(DeiT)-B}
& Baseline&81.8 &95.6 & 17.6&0&292&0\\
&SCOP~\cite{tang2020scop} &79.7 (-2.1)&94.5 (-1.1)&10.2& 42.0&403&38.1\\
&PoWER~\cite{goyal2020power}&80.1 (-1.7)&94.6 (-1.0)&10.4&39.2&397&35.8\\
&VTP~\cite{zhu2021visual}&80.7 (-1.1)&95.0 (-0.6)&10.0& 43.2&412&41.0\\
&PS-ViT (Ours) & \textbf{81.5} (-0.3) & \textbf{95.4} (-0.2)& 9.8 &44.3&414&41.8\\
&DPS-ViT (Ours) & \textbf{81.6} (-0.2) & \textbf{95.4} (-0.2)& 9.4 &46.6&413&41.3\\
\hline
\multirow{4}{*}{T2T-ViT-14}
& Baseline&81.5 &95.4& 5.2 &0&764&0\\
&PoWER~\cite{goyal2020power} &79.9 (-1.6)&94.4 (-1.0)&3.5&32.7&991&29.7\\
&PS-T2T (Ours) &\textbf{81.1} (-0.4)&\textbf{95.2} (-0.2)&3.1&40.4&1055&38.1\\
&DPS-T2T (Ours) &\textbf{81.3} (-0.2)&\textbf{95.3} (-0.1)&3.1&45.4&1078&41.1\\
\bottomrule[1.5pt]
\end{tabular}
\vspace{-4mm}
\label{tab-img}
\end{table*}
For the $i$-th patch in the $t$-th layer, ${\bm{s}}_{t,i}$ reflects its impact on the effective output of the final layer.
A larger ${\bm{s}}_{t,i}$ implies the corresponding patch has larger impact to the final error, which can reflect the importance of a patch to the model performance. The calculation of ${\bm{s}}_{t,i}$ involves all the attention maps in behind layers and the input feature of the current layer. Before pruning the current layer, we randomly sample a subset of training dataset to calculate the significance scores ${\bm{s}}_t$ and the average ${\bm{s}}_t$ over these data is adopted. The obtained ${\bm{s}}_t$ can be viewed as the real-number score for binary ${\bm{m}}_t$.
\begin{algorithm}[h]
\caption{Patch Slimming for Vision Transformers.}
\label{alg}
\begin{algorithmic}[1]
\REQUIRE{Training dataset $\mathcal{D}$, vision transformer $\mathcal{T}$ with L layers, patch masks $\{{\bm{m}}_l\}_{l=1}^L$, tolerant value $\epsilon$, preserved patch's number $r$ and search granularity $r'$. }
\STATE Initialize ${\bm{m}}_{L,0}$ as 1 and other elements as 0.
\FOR{ $l=L-1, \cdots, 1$ }
\STATE Randomly sample a subset of training data to get the significance score ${\bm{s}}_l$ in the $l$-th layer;
\STATE Set ${\bm{m}}_l={\bm{m}}_{l+1}$, $\mathcal{E}_l=+\infty$, $r=0$;
\WHILE{$\mathcal{E}_{l+1} > \epsilon$ }
\STATE Set $r$ elements in ${\bm{m}}_{l,i}$ to 1 according to positions of the largest $r$ scores ${\bm{s}}_{l,i}$.
\STATE Fine-tune $l$-th layer $\mathcal{B}_l(Z_{l-1})$ for a few epochs.
\STATE Calculate error $\mathcal{E}_{l+1}$ in the $(l+1)$-th layer.
\STATE $ r=r+ r'$.
\ENDWHILE
\ENDFOR
\ENSURE{The pruned vision transformer.}
\end{algorithmic}
\end{algorithm}
\vspace{-3mm}
\subsection{Pruning Procedure}
Here we conclude the overall pipeline of the proposed patch slimming method.
We start from the output layer and prune the previous layers layer-by-layer from top to down. Specially, all the patches preserved in the $(l+1)$-th layer will be also preserved in the $l$-th layer. The other patches are greedily selected according to their impact scores ${\bm{s}}_{l,i}$, wherein patches with larger scores are preserved preferentially. Considering the reconstruction error $\mathcal{E}_{l+1}$ in the $(l+1)$-th layer is directly affected by the patch selection in the $l$-th layer, we use it to determine whether the $l$-th layer has already enough patches. In practice, we iteratively select $r'$ important patches in each step and continue the selection process in the current layer until $\mathcal{E}_{l+1}$ is less than the given tolerate value $\epsilon$. To make $\mathcal{E}_{l+1}$ well maintain the representation ability of current preserved patches, we fine-tune the current block $\widehat{\B}_l$ for a few epochs after each step of patch selection. Taking the original feature $Z_{l-1}$ in the $(l-1)$-th layer as input, and the reconstruction error $\mathcal{E}_{l+1}$ as the objective, the parameters in the current block $\widehat{\B}_l$ are optimized. Note that the block $\widehat{\B}_l$ is a very small model with only one MSA and one MLP modules, the fine-tune process is very fast. After pruning, the mask ${\bm{m}}_l$ is fixed, and weight parameters in the vision transformer is further fine-tuned to be compatible with the efficient architecture. The procedure of patching slimming for vision transformer is summarized in Algorithm~\ref{alg}.
\subsection{A Dynamic Variant
In the above procedure, whether a patch will be preserved is determined by the statistics over the training dataset. It exploits the commonalities of redundant filters from different input adequately. Besides, dynamic pruning is the improved version of static pruning methods, which selects different patches for each input image. The dynamic strategy has been widely explored for reducing channels of CNN models~\cite{gao2018dynamic,hua2018channel, tang2021manifold}. Similarly, the proposed patch slimming paradigm can be easily extended to the dynamic variant~( dubbed as DPS-ViT), and here we present a simple implementation. Following \cite{gao2018dynamic}, we insert a small module $\mathcal{G}$ in each block to predict which patch is effective. The module $\mathcal{G}$ composes of a downsampling layer, linear layer and activation functions, which takes the input feature $Z_{l-1}\in\mathbb{R}^{N\times d}$ as input and outputs the approximate significance score $\hat {\bm{s}}_l \in \mathbb{R}^{N}$. Recalling that the score of a patch ${\bm{s}}_l$~(Eq.~\ref{eq-imp}) actually depends on the input images, the module $\mathcal{G}$ is trained to fit ${\bm{s}}_l$ calculated for each input instance in the training phase. At inference, only patches with large score $\hat {\bm{s}}_{l,i}$ are required to calculate. The dynamic strategy finds redundant patches of vision transformers depending on input data, which can excavate patch redundancy more adequately.
\vspace{-1mm}
\section {Experiments}
\label{sec-exp}
\vspace{-1mm}
In this section, we empirically investigate the effectiveness of the proposed patch slimming methods for efficient vision transformers~(PS-ViT). We evaluate our method on the benchmark ImageNet~(ILSVRC2012)~\cite{imagenet} dataset, which contains 1000-class natural images, including 1.2M training images and 5k validation images. The proposed method is compared with SOTA pruning methods and we also conduct extensive ablation studies to better understand our method.
\subsection{Experiments on ImageNet}
We conduct experiments on the standard ViT models~\cite{dosovitskiy2020image} (DeiT~\cite{touvron2020training}), an improved variant network T2T-ViT~\cite{yuan2021tokens} and the state-of-the art LV-ViT~\cite{jiang2021all}.
\textbf{Implementation details.} For a fair comparison, we follow the training and testing settings in the original papers~\cite{touvron2020training, yuan2021tokens, yuan2021tokens}, and the patch slimming is implemented based on the official pre-trained models. The global tolerant error is select from \{0.01, 0.02\} to get models with different acceleration rates, and the search granularity $r$ is set to 10. We fine-tune the current block for 3 epochs after each iteration of patch selection. After determining the proper patches in each layer, the pruned transformers are fine-tuned following the training strategy in \cite{touvron2020training}. All the experiments are conducted with PyTorch~\cite{paszke2017automatic} and MindSpore~\cite{mindspore} on NVIDIA V100 GPUs.
\textbf{Competing methods.} We compare our patch slimming with several representative model pruning methods including CNN channel pruning methods~\cite{tang2020scop} and BERT pruning methods~\cite{goyal2020power}. SCOP~\cite{tang2020scop} is a SOTA network pruning method for reducing the channels of CNNs, and we re-implement it to reduce the patches in vision transformers.
PoWER~\cite{goyal2020power} accelerates BERT inference by progressively eliminating word-vector. HVT~\cite{pan2021scalable} directly designs efficient vision transformer architectures by progressively reducing the spatial dimensions through pooling operations.
\begin{table}
\centering
\small
\caption{Comparisons with SOTA transformer models on ImageNet.}
\vspace{-2mm}
\label{tab-sota}
\begin{tabular}{l | c | c}
\toprule[1.5pt]
Model & FLOPs (G) & Top-1 Acc. (\%) \\ \hline
DeiT-S~\cite{touvron2020training} & 4.6 & 79.8 \\
DeiT-B~\cite{touvron2020training} & 17.5 & 81.8 \\ \hline
PVT-Small~\cite{wang2021pyramid} & 3.8 & 79.8 \\
PVT-Medium~\cite{wang2021pyramid} & 6.7 & 81.2 \\
PVT-Large~\cite{wang2021pyramid} & 9.8 & 81.7 \\ \hline
T2T-ViT-14~\cite{yuan2021tokens} &5.2&81.5\\
T2T-ViT-19~\cite{yuan2021tokens} &8.9&81.9\\
T2T-ViT-24~\cite{yuan2021tokens} &14.1&82.3\\ \hline
TNT-S~\cite{han2021transformer} & 5.2 & 81.5 \\
TNT-B~\cite{han2021transformer} & 14.1 & 82.9 \\ \hline
Swin-T~\cite{liu2021Swin} & 4.5 & 81.3 \\
Swin-S~\cite{liu2021Swin} & 8.7 & 83.0 \\
Swin-B~\cite{liu2021Swin} & 15.4 &83.5 \\ \hline
LV-ViT-S~\cite{jiang2021all} & 6.6 & 83.3 \\
LV-ViT-M~\cite{jiang2021all} & 16.0 &84.1 \\ \hline
PS-LV-ViT-S~(Ours) & 4.7 & 82.4\\
DPS-LV-ViT-S~(Ours) & 4.5 & 82.9 \\
PS-LV-ViT-M~(Ours) & 8.6 & 83.5\\
DPS-LV-ViT-M~(Ours) & 8.3 & 83.7\\
\bottomrule[1.5pt]
\end{tabular}
\vspace{-4mm}
\end{table}
\textbf{Experimental results.} The experimental results are shown in Table~\ref{tab-img}, where `PS-' and `DPS-' denote the proposed patch pruning method and its dynamic variant, respectively. We evaluate on three versions of DeiT~\cite{touvron2020training} with different model sizes, \ie, DeiT-Ti, DeiT-S, and DeiT-B. Our method achieve obviously higher performance compared to the existing methods. The SCOP method~\cite{tang2020scop} designed for CNNs achieve poor performance when applied for reducing patches in a vision transformer, implying simply migrating the channel pruning methods cannot work well. PoWER~\cite{goyal2020power} has a larger accuracy drop than our method, indicating the model compression method for NLP models is not optimal for CV models. Compared to the vision transformer structure pruning method VTP~\cite{zhu2021visual}, our method investigates a new prospective by pruning patches and achieve higher accuracy with similar FLOPs.
As for T2T-ViT model, our method can reduce the FLOPs by 40.4\% and only have a small accuracy decrease (0.4\%), which is much better than the compared PoWER method. This indicates that the patch-level redundancy exists in various vision transformer models and our method can well excavate the redundancy.
We further conduct experiments on a SOTA transformer model, LV-ViT~\cite{jiang2021all}, and show the results in Table~\ref{tab-sota}. The results show that our patch pruning method also work well on LV-ViT, \eg, the dynamic patch slimming reduces the FLOPs of LV-ViT-M from 16.0G to 8.3G, still achieving 83.7\% top-1 accuracy. Its performance is also superior to other SOTA models such as Swin transformer~\cite{liu2021Swin}.
\subsection{Ablation Study}
We conduct extensive ablation studies on ImageNet to verify the effectiveness of each component in our method. The DeiT-S model on the ImageNet dataset is used as the base model.
\begin{table}[t]
\centering
\small
\caption{Learned patch pruning \textit{vs.}~uniform pruning.}
\vspace{-2mm}
\begin{tabular}{c|c|c|c}
\toprule[1.5pt]
\multirow{2}{*}{Method} & {Top-1} &Top-5 & FLOPs \\
& {Acc. (\%)} &Acc. (\%) & (G) \\
\hline
Baseline& 79.8&95.0& 4.6 \\
Uniform pruning & 77.2 & 93.8 & 2.6 \\
Ours & 79.4 & 94.7 & 2.6 \\
\bottomrule[1.5pt]
\end{tabular}
\label{tab-even}
\vspace{-4mm}
\end{table}
\begin{figure*}[t]
\centering
\small
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=0.9\linewidth]{flops}
\caption{The ImageNet accuracy and FLOPs of the pruned DeiT-S \wrt tolerant error $\epsilon$.}
\label{fig-eps}
\end{minipage}
\hspace{0.5em}
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=0.9\linewidth]{width}
\caption{The architecture of pruned DeiT-S on ImageNet.}
\label{fig-even}
\end{minipage}
\hspace{0.5em}
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=0.9\columnwidth]{pr_ratio} %
\caption{Pruning different layers of the DeiT-S model on ImageNet.}
\label{fig-layers}
\vspace{-5mm}
\end{minipage}
\vspace{-2mm}
\end{figure*}
\begin{figure*}[h]
\centering
\small
\begin{subfigure}{0.3\linewidth}
\centering
\includegraphics[width=0.9\linewidth]{layer1} %
\caption{The 3-th layer.}
\end{subfigure}
\hspace{0.5em}
\begin{subfigure}{0.3\linewidth}
\centering
\includegraphics[width=0.9\linewidth]{layer5}
\caption{The 7-th layer.}
\end{subfigure}
\hspace{0.5em}
\begin{subfigure}{0.3\linewidth}
\centering
\includegraphics[width=0.9\linewidth]{layer9}
\caption{The 11-th layer.}
\end{subfigure}
\vspace{-2mm}
\caption{Accuracy \wrt the pruning rate of patches in a single layer.}
\label{fig-acc}
\vspace{-4mm}
\end{figure*}
\textbf{The effect of global tolerant error $\epsilon$.}
The tolerant error $\epsilon$ affects the balance between computational cost and accuracy of the pruned model, which is empirically investigated in Figure~\ref{fig-eps}. Increasing $\epsilon$ implies larger reconstructed error between features of the pruned DeiT and original DeiT, while more patches can be pruned to achieve higher acceleration rate. When the reduction of FLOPs is less than 45\%, there is almost no accuracy loss~(less than 0.4\%), which is because that a large number of patches are redundant.
\textbf{Learned patch pruning \textit{vs.}~uniform pruning.}
In our method, the number of patches required in a specific layers is determined automatically via the global tolerant value $\epsilon$. The architecture of the pruned DeiT model is shown in Figure~\ref{fig-even}. We can see that a pyramid-like architecture is obtained, where most of the patches in deep layers are pruned while more patches are preserved in shallow layers. To validate the superiority of the learned pyramid architecture, we also implement a baseline that uniformly prunes all the layers with the similar pruning rate. We compare the results of the proposed patch slimming method and uniform pruning in Table~\ref{tab-even}. The accuracy of uniform pruning is only 77.2\%, which incurs a large accuracy drop (-2.6\%).
To better understand the behavior of patch pruning in the vision transformer, we prune patches in a single layer to see how the test accuracy change.
The experiments are conducted with DeiT-S model on ImageNet.
\textbf{Redundancy \wrt depth.} We test the patch redundancy of different layers to verify the motivation of top-down patch slimming procedure. We prune a single layer and keep the same pruning ratio for different layers. Figure~\ref{fig-layers} shows the accuracy of the pruned model after pruning patches of a certain layer, and each line denotes pruning patches with a given pruning rate. In deeper layers, more patches can be safely removed without large impact on the final performance. However, removing a patch in lower layers usually incurs obvious accuracy drop. The patch redundancy is extremely different across layers and deeper layers have more redundancy, which can be attributed to that the attention mechanism aggregates features from different patches and the deeper patches have been fully communicated with each other. This phenomenon is different from the channel pruning in CNNs, where lower layers are observed to have more channel-level redundancy~(Figure~4 in \cite{he2017channel}).
\textbf{Effectiveness of impact estimation.} We define the scores ${\bm{s}}_l$ in Eq.~\ref{eq-imp} to approximate significance of a patch by propagating the reconstruction error of effective patches in output layer. To validate its effectiveness, we compare it with two baseline scores: `Random' denotes removing patches in the layer randomly, and `Attn' approximates the importance of a patch only with the norm of its attention map in the current layer. We compare the three scores by utilizing them to prune patches in different layer. The results are presented in Figure~\ref{fig-acc}, where $y$-axis is the test accuracy of the pruned models~(without fine-tuning). From the results, our impact estimation manner suffers less accuracy loss than the others with the same pruning rate~(\eg, 50\%). It implies that our method can effectively identify patches that really make contributions to the final prediction.
\vspace{-1mm}
\section{Conclusion}
\vspace{-1mm}
We propose to accelerate vision transformers by reducing the number of patches required to calculate. Considering that the attention mechanism aggregates different patches layer-by-layer, a top-down framework is developed to excavate the redundant patches. The importance of each patch is also approximated according to its impact on the effective output features.
After pruning, a compact vision transformer with a pyramid-like architecture is obtained. Extensive experiments on benchmark datasets validate that the proposed method can effectively reduce the computational cost. In the future, we plan to combine the patch slimming methods with more compression technologies~(\eg, weight pruning, model quantization) to explore extremely efficient vision transformers.
\noindent\textbf{Acknowledgment.} This work is supported by National Natural Science Foundation of China under Grant No.61876007, Australian Research Council under Project DP210101859 and the University of Sydney SOAR Prize.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}\label{1}
Topological insulators are band insulators with topologically protected boundary states and insulating bulk states~\cite{Moore2010,Hasan2010,Qi2011,Shen2012,Wang2017,Qin2020,Chen2021prl,Chen2021prr}. A well-known topological insulator paradigm is the quantum Hall insulator (QHI) phase~\cite{ZhangExp2009}.
Distinct from the QHI phase, there also exists a quantized version of the Hall effects in a magnetically doped topological insulator without applying any an external magnetic field.
In an intrinsic magnetic topological insulator~\cite{XueExp2013,CheckelskyExp2014,QAHIexp2014,QAHIexp2015,XueExp2015,YoshimiExp2015,ChangExp2016,GrauerExp2017,XueExp2018,ZhangExp2020,QAHIexp2020,ZhaoExp2020,MogiExp2021}, a band gap on the surface states of the topological materials is opened by the time-reversal symmetry breaking, which is essential for the realization of the quantum anomalous Hall insulator (QAHI) phase~\cite{XueExp2013,Qi2008,Yu2010,Chu2011}.
Physically, how to realize the QAHI phase has attracted much attention in the past few decades~\cite{Qiao2010,Qi2006,Onoda2006,Nomura2011,Checkelsky2012}. One of the promising physical schemes is on the basis of the topological insulators doped with magnetic impurities~\cite{Hor2010,Chen2010,Wray2011,Wang2013,Wang2014,Law2017,Kim2018,Kawamura2018}, where the interplay of the magnetic exchange interaction and the spin-orbit coupling gives rise to the band inversion between the conduction and valence bands.
More importantly, the experimental realization of the QAHI phase has been reported in the thin films of Cr-doped (Bi,Sb)$_2$Te$_3$, which is an intrinsic magnetic topological insulator~\cite{XueExp2013}.
The quantum spin Hall insulator (QSHI) phase is another family member of the Hall effects~\cite{Kane2005,Bernevig2006,Haldane2006,Prodan2009,Xing2010,Xing2011,Chen2017}. To explain QSHI, we take the Kane and Mele model~\cite{Kane2005} for example. In this model, the spin-up and -down
electrons exhibit opposite Chern numbers due to the spin-orbit coupling, so that the total Chern number vanishes but the spin Chern number is nonzero. Particularly, the QSHI phase has been observed experimentally in HgTe~\cite{QSHIexp2006,QSHIexp2007}. Usually, the QSHI phase was considered to be protected by the time-reversal symmetry. However, it has been found that the QSHI phase with nonzero spin Chern numbers persists when the time-reversal symmetry is broken and this is called the time-reversal-symmetry-broken QSHI phase~\cite{Xing2011,Chen2017}.
Based on the Floquet theory that allows one to map a time-dependent problem into a stationary one, it has been shown that a periodic perturbation can induce topological phase transitions in a topological trivial insulator~\cite{Oka2009prb,Calve2015prb,Kitagawa2011prb,Lindner2011np,Rechtsman2013nature,Wang2018prl,Xu2021as,Zhu2014,lightexp2013,lightexp2016,Leon2013,Katan2013,Inoue2010,Kitagawa2010,Yan2016,Narayan2015,Saha2016,Chen20181,Chen20182,Pervishko2018,Kyriienko2019,Huang20prl,Hu20prl,Du21arxiv,Ning21arxiv}.
In particular, the quantum Hall effect induced by circularly polarized field in Dirac materials was predicted decade ago by Oka and Aoki~\cite{Oka2009prb}. Further, the theory given by Oka and Aoki was generalized for various topological insulators in the previous Refs.~\cite{Calve2015prb,Kitagawa2011prb,Lindner2011np,Rechtsman2013nature,Wang2018prl,Xu2021as,Zhu2014,lightexp2013,lightexp2016,Leon2013,Katan2013,Inoue2010,Kitagawa2010,Yan2016,Narayan2015,Saha2016,Chen20181,Chen20182,Pervishko2018,Kyriienko2019,Huang20prl,Hu20prl,Du21arxiv,Ning21arxiv}. Later the corresponding Floquet topological insulators have been observed by experiments~\cite{lightexp2013,lightexp2016}. Also, it has been shown that an intense high-frequency linearly polarized light can be used to manipulate the value of a gap in of a non-doped topological insulator thin film~\cite{Pervishko2018}. However, the linearly polarized light cannot include the contribution of $1/\omega$, so that they only estimated the terms proportional to $1/\omega^2$~\cite{Pervishko2018}, where $\omega$ is the frequency of the polarized light. Noteworthy, the circularly polarized light can include the contribution of $1/\omega$.
Therefore, the impact of high-frequency pumping with the circularly polarized light on the thin films of topological insulators needs to be studied comprehensively.
In this work, we investigate the topological phase transitions in the atomically thin flakes of an intrinsic magnetic topological insulator with high-frequency pumping. Based on the results, it is found that the intensity of the circularly polarized light can be used as a knob to drive a topological transition. Different from the situation in the absent of optical field, there exist four different phases: the normal insulator (NI) phase, the time-reversal-symmetry-broken QSHI phase, and two different QAHI phases with opposite Chern numbers $C=\pm1$, respectively. In particular, for a special given layer thickness, one can apparently detect the tendency towards light-induced band inversion upon increasing optical field intensity and passing through the QSHI phase region. This means that the energy gap of the surface states can be tuned by adjusting the intensity of the driving optical field in an experimentally accessible range.
The paper is organized as the following: In Sec.~\ref{2}, we
give the model Hamiltonian. In Sec.~\ref{3}, we introduce the Floquet theory for a time-periodic Hamiltonian. In Sec.~\ref{4}, we give the polarized light and Floquet Hamiltonian which is used in the following calculations.
Furthermore, we study the basis states at the $\Gamma$ point in Sec.~\ref{5}.
Moreover, we calculate the high-frequency pumping-induced topological properties of the effective Floquet Hamiltonian of the thin film in Sec.~\ref{6}. In addition, we give the light-induced topological phase diagram in Sec.~\ref{7}. Finally, we summarize in Sec.~\ref{8}.
\section{Model}\label{2}
We take the periodic boundary conditions in the $x-y$ plane such that $k_x$ and $k_y$ are good quantum numbers,
and denote the thickness of the thin film along $z$ direction as $L$.
In the basis $(|p1_{z}^{+},\uparrow\rangle , |p2_{z}^{-},\uparrow\rangle , |p1_{z}^{+},\downarrow\rangle , |p2_{z}^{-},\downarrow\rangle )$ which
are the hybridized states of Se $p_{z}$ orbital and Bi $p_{z}$ orbital, with even ($+$) and odd ($-$) parities, up ($\uparrow$) and down ($\downarrow$) spins, the low-energy three-dimensional Hamiltonian for Cr-doped Bi$_{2}$Se$_{3}$ is given by~\cite{Shan2010,Lu2010,Xing2010,Lu2013,Liu2019,Sun2020,Liu2010prb,Dabiri2021prb1,Dabiri2021prb2}
\begin{align}\label{eq:H}
H({\bf k}) = H_{0}({\bf k}) + H_{X}(z),
\end{align} where
\begin{align}\label{eq:H0}
H_{0}({\bf k}) &\!=\! \epsilon_{0}({\bf k})I_{4} \nonumber\\
&~\!+\!
\!\left(\!
\begin{array}{cccc}
\!M({\bf k})\sigma_{z} \!-\! iA_{1}\partial_{z}\sigma_{x} & A_{2}k_{-}\sigma_{x} \\
A_{2}k_{+}\sigma_{x} & M({\bf k})\sigma_{z} \!+\! iA_{1}\partial_{z}\sigma_{x}\!
\end{array}
\!\right)\! \nonumber\\
&\!=\! \epsilon_{0}({\bf k})\tau_{0}\otimes\sigma_{0} \!+\! M({\bf k})\tau_{0}\otimes\sigma_{z} \!+\! \!A_{1}\!k_{z}\tau_{z}\otimes\sigma_{x} \nonumber\\
&~\!+\! A_{2}k_{x}\tau_{x}\otimes\sigma_{x} \!+\! A_{2}k_{y}\tau_{y}\otimes\sigma_{x}.
\end{align} Here $\sigma_{x,y,z}$ are the Pauli matrices for the orbital degree of freedom, $k_z=-i\partial_{z}$, $I_{4}$ is the $4\times 4$ identity matrix, $k_{\pm}=k_{x}\pm ik_{y}$, $\epsilon_{0}({\bf k})=C_{0}-D_{1}\partial_{z}^{2}+D_{2}(k_{x}^{2}+k_{y}^{2})$, $M({\bf k})=M_{0}+B_{1}\partial_{z}^{2}-B_{2}(k_{x}^{2}+k_{y}^{2})$, $C_0$, $D_i$, $M_0$, $B_i$, and $A_i$ are model parameters with $i=1,2$.
The parameters for Bi$_{2}$Se$_{3}$ are adopted as~\cite{ZhangExp2009,XueExp2013}: $C_{0}=-0.0068$ eV, $D_{1}=1.3$ eV\AA$^2$, $D_{2}=19.6$ eV\AA$^2$, $A_{1}=2.2$ eV\AA, $A_{2}=4.1$ eV\AA, $M_{0}=0.28$ eV, $B_{1}=10$ eV\AA$^2$, and $B_{2}=56.6$ eV\AA$^2$.
The exchange field reads~\cite{Lu2013,Liu2019}
\begin{align}\label{eq:HX}
H_{X}(z) = m_{0}\tau_{z}\otimes\sigma_{0},
\end{align}
where $m_{0}$ is the magnitude of the bulk magnetic moment~\cite{XueExp2013}, $\tau_{z}$ is the $z$ Pauli
matrix for the spin degree of freedom, $\sigma_{0}$ is a $2\times2$ unit matrix, and the magnetization energy along the $z$ direction is given by $m_{0}$, i.e.,
$m_{0}$ is the exchange field from the magnetic dopants.
\section{Floquet formula}\label{3}
The Floquet theory can be applied to a time-periodic Hamiltonian $H(t) = H(t + T )$ with the period
$T = 2\pi/\omega$ and the frequency $\omega$ of the light. By employing the Floquet theory, the wave function of the time-periodic
Schr\"odinger equation $i\partial_{t}\Psi(t) = H(t)\Psi(t)$, has the form $\Psi(t)=\sum_{m}\psi_{m}e^{-i(\epsilon/\hbar + m\omega)t}$,
where $\epsilon$ is the quasienergy and $m$ is an integer. With a Fourier series expansion, we find that $\sum_{m}H_{n,m}\psi_{m} = \epsilon \psi_{n}$, where
\begin{align}\label{eq:HFnm}
H_{n,m} = n\hbar\omega\delta_{n,m} + \frac{1}{T} \int_{0}^{T} H(t) e^{i(n-m)\omega t}dt,
\end{align}
which is a block Hamiltonian of the Floquet state, $n$ and $m$ are integers. If $\Psi(t)$ is an eigenvector with the quasienergy $\epsilon$, $e^{in\omega t}\Psi(t)$ is also an eigenvector of the system with the quasienergy $\epsilon + n\hbar\omega$.
From Eq.~(\ref{eq:HFnm}), one can have
\begin{align}\label{eq:Hnm}
H_{n,m}=
\left(
\begin{array}{ccccc}
\cdots & \cdots & \cdots & \cdots & \cdots \\
\cdots & H_{-1,-1} & H_{-1,0} & H_{-1,1} & \cdots \\
\cdots & H_{0,-1} & H_{0,0} & H_{0,1} & \cdots \\
\cdots & H_{1,-1} & H_{1,0} & H_{1,1} & \cdots\\
\cdots & \cdots & \cdots & \cdots & \cdots
\end{array}
\right),
\end{align} where $H_{n,m} =H_{-m,-n}$ with $n\neq m$.
\section{Polarized light and Floquet Hamiltonian}\label{4}
The time-dependent Hamiltonian can be experimentally introduced by normally illuminating with elliptically polarized light described by a time-varying gauge field (or the vector potential) ${\bf A}(t) = A( \sin(\omega t), \sin(\omega t + \varphi) )$~\cite{Saha2016,Zhu2014}, which gives the optical field as ${\bf E}(t) = \partial{\bf A}(t)/\partial t = E_{0}( \cos(\omega t), \cos(\omega t + \varphi) )$, where $E_{0}=A/\omega$ is the amplitude of the optical field, $\omega$ is the frequency of the optical field, and the phase $\varphi$ controls the polarization: when $\varphi = 0$ or $\pi$, the optical field is linearly polarized; when $\varphi=\pm\pi/2$, the optical field is circularly polarized; for example, ${\bf A}(t) = A( \sin(\omega t), \cos(\omega t) )$ with $\varphi=\pi/2$; when $\varphi$ takes other values, the optical field is elliptically polarized.
From the point of view of the source~\cite{wikipedia_Circular}, with $\varphi=\pi/2$, the corresponding electric field vector is ${\bf E}(t) = E_{0}(\cos(\omega t), \cos(\omega t + \pi/2)) = E_{0}(\cos(\omega t), -\sin(\omega t))$, which is a left-handed circularly polarized wave and corresponds to ${\bf A}(t) = A( \sin(\omega t), \cos(\omega t) )$. For $\varphi=-\pi/2$, the corresponding electric field vector is ${\bf E}(t) = E_{0}(\cos(\omega t), \cos(\omega t - \pi/2)) = E_{0}(\cos(\omega t), \sin(\omega t))$, which is a right-handed circularly polarized wave and corresponds to ${\bf A}(t) = A( \sin(\omega t), -\cos(\omega t) )$. The vector potentials for right- and left-handed circularly polarized lights are different from each other and different vector potential will give different time-dependent Hamiltonian (\ref{eq:Ht}). In the following calculations, we focus on the left-handed circularly polarized light with $\varphi=\pi/2$.
As a high-frequency laser light, we choose the photon energy, for example, to be $\hbar\omega\approx0.15$ eV ($\omega \sim 2.2789\times10^{2}$ THz), which is close to the typical values in recent optical pump-probe experiments. For example, the typical amplitude of light $A_{0}=eA/\hbar=eE_{0}/(\hbar\omega)$ is about 0.03 \AA$^{-1}$,
and the corresponding electric field strength $E_{0}=\hbar\omega A_{0}/e$ is $4.5\times10^{7}$ V/m, which is within the experimental accessibility~\cite{lightexp2013,lightexp2016}.
By use of the Peierls substitution, the time-dependent Hamiltonian is obtained as
\begin{align}\label{eq:Ht}
H(t)= H\left({\bf k}_{\perp} - \frac{e}{\hbar}{\bf A}(t), -i\partial_{z}\right) ,
\end{align} where ${\bf k}_{\perp} = (k_x, k_y)$.
Making use of the Floquet theory~\cite{Zhu2014,Pervishko2018,Chen20181,Chen20182} in the high-frequency limit, the periodically driven system can be described by a static effective Hamiltonian as~\cite{high-frequency1982,high-frequency1988,high-frequency20031,high-frequency20032,high-frequency2014,high-frequency20151,high-frequency20152}
\begin{align}\label{eq:HF0}
H^{(F)} = H_{0,0} + \frac{[H_{0,-1}, H_{0,1}]}{\hbar\omega} ,
\end{align} where
\begin{align}\label{eq:H00}
H_{0,0}
= H({\bf k}) + D_{2}A_{0}^{2} - B_{2}A_{0}^{2}I_{2}\otimes\sigma_{z},
\end{align}
\begin{align}\label{eq:H0-1}
H_{0,-1}
&\!=\! -iD_{2}A_{0}(k_{x} \!+\! e^{-i\varphi}k_{y}) \nonumber\\
&\!+\!
\!\left(\!
\begin{array}{cccc}
iB_{2}A_{0}(k_{x} \!+\! e^{-i\varphi}k_{y})\sigma_{z} & \!-\!\frac{i}{2}A_{2}A_{0}(1 \!-\! ie^{-i\varphi})\sigma_{x}\! \\
\!-\!\frac{i}{2}A_{2}A_{0}(1 \!+\! ie^{-i\varphi})\sigma_{x} & iB_{2}A_{0}(k_{x} \!+\! e^{-i\varphi}k_{y})\sigma_{z}\!
\end{array}
\!\right)\!,
\end{align}
\begin{align}\label{eq:H01}
H_{0,1}
&\!=\! iD_{2}A_{0}(k_{x} + e^{i\varphi}k_{y}) \nonumber\\
&\!+\!
\!\left(\!
\begin{array}{cccc}
\!-\!iB_{2}A_{0}(k_{x} \!+\! e^{i\varphi}k_{y})\sigma_{z} & \frac{i}{2}A_{2}A_{0}(1 \!-\! ie^{i\varphi})\sigma_{x} \\
\frac{i}{2}A_{2}A_{0}(1 \!+\! ie^{i\varphi})\sigma_{x} & \!-\!iB_{2}A_{0}(k_{x} \!+\! e^{i\varphi}k_{y})\sigma_{z}\!
\end{array}
\!\right)\!.
\end{align}
From Eq.~(\ref{eq:HF0}), the Floquet Hamiltonian is
\begin{align}\label{eq:HF}
H^{(F)} &\!=\!
\!\left(\!
\begin{array}{cccc}
\tilde{M}({\bf k})\sigma_{z} \!-\! iA_{1}\partial_{z}\sigma_{x} & A_{2}k_{-} \sigma_{x} \\
A_{2}k_{+} \sigma_{x} & \tilde{M}({\bf k})\sigma_{z} \!+\! iA_{1}\partial_{z}\sigma_{x}\!
\end{array}
\!\right)\! \nonumber\\
&\!+\! \frac{A_{0}^{2}A_{2}\sin\varphi}{\hbar\omega} \!\left(\!
\begin{array}{cccc}
\!-\!A_{2}I_{2} & \!-\! B_{2}k_{-}(\sigma_{+} \!-\! \sigma_{-})\! \\
\!B_{2}k_{+}(\sigma_{+} \!-\! \sigma_{-}) & A_{2}I_{2}
\end{array}
\!\right)\! \nonumber\\
&\!+\! \tilde{\epsilon}_{0}({\bf k})I_{4} ,
\end{align} where $\tilde{\epsilon}_{0}({\bf k})=\epsilon_{0}({\bf k}) + A_{0}^{2}D_{2}$, $\tilde{M}({\bf k})=M({\bf k}) - A_{0}^{2}B_{2}$, $\sigma_{+}=\sigma_{x} + i\sigma_{y}$, and $\sigma_{-}=\sigma_{x} - i\sigma_{y}$.
To give a simple quantitative estimation of the validity of the theoretical formalism
developed here, we evaluate the maximum instantaneous energy of the time-dependent Hamiltonian (\ref{eq:Ht}) averaged over
a period of the field $\frac{1}{T}\int_{0}^{T}dt~\text{max}\left\{\big|\big|H(t)\big|\big|\right\}<\hbar\omega$.
Therefore, in the vicinity of the $\Gamma$ point, the field parameters have to meet the condition $A_{0}A_{2}/(\hbar\omega)<1$. Particularly, in the high-frequency regime for an external pumping $\hbar\omega\approx0.4$ eV ($\omega \sim 6.07707\times10^{2}$ THz), one can estimate $A_{0}\lesssim0.09$ \AA$^{-1}$.
Experimentally, from the spectroscopic ellipsometry measurements of the bulk Bi$_2$Se$_3$~\cite{McIver2012}, the penetration depth of the laser light is about $25$ nm~\cite{McIver2012}.
At the same time, from the transient reflectivity/transmission measurements of the bulk Bi$_2$Se$_3$~\cite{Glinka2013}, the penetration depth of the laser light is about $21$ nm~\cite{Glinka2013}.
Both $21$ nm and $25$ nm are reasonably larger than the film thickness ($2\sim8$ nm) of the bulk Bi$_2$Se$_3$.
It is important that the penetration depth of the laser light should be reasonably larger than the film thickness, so that the vector potential ${\bf A}(t)$ of the laser light can properly change the momentum ${\bf k}_{\perp}$ as in the thin-film Hamiltonian~(\ref{eq:Ht}). Otherwise, the effect of the vector potential will not go into the bulk state of the thin film due to the skin effect.
\section{Basis states at the $\Gamma$ point}\label{5}
To establish an effective model for the surface states, we
first find the four solutions to the surface states of the model
in Eq.~(\ref{eq:H}) at the $\Gamma$ point ($k_x=k_y= 0$) as~\cite{Shan2010,Lu2010}
\begin{align}\label{eq:H02}
H_{0} =
\left(
\begin{array}{cccc}
h(A_{1}) & 0 \\
0 & h(-A_{1})
\end{array}
\right),
\end{align} where
\begin{align}
h(A_{1})
&\!=\! (C_{0}-D_{1}\partial_{z}^{2})I_{2}+(M_{0}+B_{1}\partial_{z}^{2})\sigma_{z} - iA_{1}\partial_{z}\sigma_{x} \nonumber\\
&\!=\!
\left(
\begin{array}{cccc}
C_{0}+M_{0}-D_{-}\partial_{z}^{2} & -iA_{1}\partial_{z} \\
-i A_{1}\partial_{z} & C_{0}-M_{0}-D_{+}\partial_{z}^{2}
\end{array}
\right),
\end{align} and $D_{\pm}=D_{1}\pm B_{1}$.
$H_{0}$ in Eq.~(\ref{eq:H02}) is block-diagonal and its solution can be found by solving each block separately, i.e.,
\begin{align}
&h(A_{1}) \Psi_{\uparrow}(z) = E \Psi_{\uparrow}(z),\label{eq:up} \\
&h(-A_{1}) \Psi_{\downarrow}(z) = E \Psi_{\downarrow}(z).
\end{align} Because the lower block is the ``time'' reversal of the upper block, the solutions satisfy $\Psi^{\downarrow}(z) = \Theta\Psi^{\uparrow}(z)$, where
$\Theta=-i\sigma_{y}\mathcal{K}$ is the time-reversal operator and $\mathcal{K}$ is the complex conjugation operation. Equivalently, we can replace $A_1$ by $-A_1$ in all the results for the upper block, to obtain those for the lower block. Therefore, we only need to solve $h(A_1)$.
The solution of the block-diagonal $H_{0}$ can be found by putting a two-component trial solution into the eigenequation (\ref{eq:up}) of the upper block with $\Psi_{\uparrow}(z) = \Psi^{\uparrow}_{\lambda}(z) e^{\lambda z}$, where $\lambda$ is the trial coefficients defining the behavior of the wave functions and $E$ is the trial eigenenergy.
Therefore, one can have
\begin{align}
E_{\pm} &= C_{0} - D_{1}\lambda^{2} \nonumber\\
&~~\pm \sqrt{(M_{0} + B_{1}\lambda^{2} - A_{1}\lambda ) (M_{0} + B_{1}\lambda^{2} + A_{1}\lambda )},
\end{align} and
\begin{align}\label{eq:eigenvector1}
\Psi^{\uparrow}_{\lambda} = \left(
\begin{array}{cc}
D_{+}\lambda^{2} - L_{-} + E_{\pm} \\
-i A_{1}\lambda
\end{array}
\right).
\end{align}
Note that the trial coefficients may have multiple solutions, the final solution should be a linear superposition of these solutions with the superposition coefficients determined by boundary conditions. Then the problem becomes a straightforward calculation of the Schr\"odinger equation or the secular equation $\text{det}|h(A_{1}) - E|=0$ which gives four solutions of $\lambda_{\alpha}(E)$, denoted as $\beta\lambda_{\alpha}(E)$, with $\alpha\in \{1, 2\}$, $\beta\in \{+, -\}$ and
$\lambda_{\alpha}$ define the behavior of the wave functions along $z$ axis and are functions of the energy $E$ as
\begin{align}
\!\lambda_{\alpha}(E) \!=\! \sqrt{\frac{\!-\! F \!+\! (-1)^{\alpha-1}\sqrt{R}}{2D_{+}D_{-}}},
\end{align}
where we have defined
\begin{align}
F&=A_{1}^{2} + 2D_{1}(E - C_{0}) - 2B_{1}M_{0}\nonumber\\
&=A_{1}^{2} + D_{+}(E - L_{+}) + D_{-}(E - L_{-}), \\
R&=F^{2} - 4(D_{1}^{2} -B_{1}^{2})[(E-C_{0})^{2} - M_{0}^{2}]\nonumber\\
&=F^{2} - 4D_{+} D_{-}(E - L_{+}) (E - L_{-}),
\end{align} $D_{\pm}=D_{1}\pm B_{1}$, and $L_{\pm}=C_{0}\pm M_{0}$.
With Eq.~(\ref{eq:eigenvector1}), the general solution is a linear combination of the four linearly independent two-component vectors
\begin{align}\label{eq:wavefunctionup1}
\Psi_{\uparrow}(z) &= \sum_{\alpha=1,2} \sum_{\beta=+,-} C_{\alpha\beta} \Psi^{\uparrow}_{\alpha\beta} e^{\beta\lambda_{\alpha}z} \nonumber\\
&= \sum_{\alpha=1,2} \sum_{\beta=+,-} C_{\alpha\beta} \left(
\begin{array}{cc}
D_{+}\lambda_{\alpha}^{2} - L_{-} + E_{\pm} \\
-i A_{1}(\beta\lambda_{\alpha} )
\end{array}
\right) e^{\beta\lambda_{\alpha}z},
\end{align} where the superposition coefficients $C_{\alpha\beta}$ are determined by boundary conditions.
\section{Effective model of the thin film}\label{6}
In this section, we derive the effective low-energy continuous model for the thin film of the three-dimensional topological insulators.
\subsection{Finite-thickness boundary conditions}\label{sec:Finite-thickness}\label{6.1}
Now, we turn to discuss the gap in a thin film with both top and bottom open surfaces.
When the thickness $L$ of the film is comparable with the characteristic length $1/\lambda$ of the
surface states, there is a coupling between the states on opposite surfaces. One has to consider
the boundary conditions at both surfaces simultaneously. Without loss of generality, we will
consider that the top surface is located at $z = L/2$ and the bottom surface at $-L/2$. The
boundary conditions are given as
\begin{align}\label{eq:finite-thickness}
\Psi_{\uparrow}\left(z=\pm\frac{L}{2} \right) = 0,
\end{align} where $L=N_{L}d$ is the thickness of the film with the number $N_{L}$ of the layers and the thickness $d$ of a layer, and $-L/2\leqslant z \leqslant L/2$.
With the boundary conditions in Eq.~(\ref{eq:finite-thickness}), the trial wave function can be given by
\begin{align}\label{eq:finite-wavefunction}
\Psi_{\uparrow}(z) = \left(
\begin{array}{cc}
\psi_{1}(z) \\
\psi_{2}(z)
\end{array}
\right),
\end{align} where
\begin{align}
\psi_{1}(z) &= c_{+}f_{+}(z) + c_{-}f_{-}(z), \\
\psi_{2}(z) &= d_{+}f_{+}(z) + d_{-}f_{-}(z),
\end{align}
\begin{align}
f_{+}(z) &= \frac{\cosh(\lambda_{1}z)}{\cosh(\lambda_{1}L/2)} - \frac{\cosh(\lambda_{2}z)}{\cosh(\lambda_{2}L/2)}, \\
f_{-}(z) &= \frac{\sinh(\lambda_{1}z)}{\sinh(\lambda_{1}L/2)} - \frac{\sinh(\lambda_{2}z)}{\sinh(\lambda_{2}L/2)}.
\end{align}
Substituting Eq.~(\ref{eq:finite-wavefunction}) into (\ref{eq:up}): $h(A_{1})\Psi_{\uparrow}(z) \!=\! E \Psi_{\uparrow}(z)$, we have
\begin{align}\label{eq:finite-eigenequation}
\!\left(\!
\begin{array}{cccc}
\!C_{0}\!+\!M_{0}\!-\!E\!-\!D_{-}\partial_{z}^{2}\! & -iA_{1}\partial_{z} \\
-i A_{1}\partial_{z} & \!C_{0}\!-\!M_{0}\!-\!E\!-\!D_{+}\partial_{z}^{2}\!
\end{array}
\!\right)\! \!\left(\!
\begin{array}{cc}
\!\psi_{1}(z)\! \\
\!\psi_{2}(z)\!
\end{array}
\!\right)\!=\!0.
\end{align}
With Eq.~(\ref{eq:finite-eigenequation}), we have
\begin{align}
\frac{c_{+}}{d_{-}} &= \frac{C_{0}-M_{0}-E-D_{+}\lambda_{1}^{2}}{ i A_{1}\lambda_{1}} \frac{\cosh(\lambda_{1}L/2)}{\sinh(\lambda_{1}L/2)}, \\
\frac{c_{+}}{d_{-}} &= \frac{C_{0}-M_{0}-E-D_{+}\lambda_{2}^{2}}{ i A_{1}\lambda_{2}} \frac{\cosh(\lambda_{2}L/2)}{\sinh(\lambda_{2}L/2)}, \\
\frac{c_{-}}{d_{+}} &= \frac{C_{0}-M_{0}-E-D_{+}\lambda_{1}^{2}}{ i A_{1}\lambda_{1}} \frac{\sinh(\lambda_{1}L/2)}{\cosh(\lambda_{1}L/2)}, \\
\frac{c_{-}}{d_{+}} &= \frac{C_{0}-M_{0}-E-D_{+}\lambda_{2}^{2}}{ i A_{1}\lambda_{2}} \frac{\sinh(\lambda_{2}L/2)}{\cosh(\lambda_{2}L/2)}.
\end{align}
Furthermore, the secular equation of the nontrivial solution to the superposition coefficients $C_{\alpha\beta}$ leads to the transcendental equations
\begin{align}
&\frac{[C_{0} \!-\! M_{0} \!-\! E_{+}^{0} \!-\! D_{+}\lambda_{1}^{2}(E_{+}^{0}) ]\lambda_{2}(E_{+}^{0})}{[C_{0} \!-\! M_{0} \!-\! E_{+}^{0} \!-\! D_{+}\lambda_{2}^{2}(E_{+}^{0}) ]\lambda_{1}(E_{+}^{0})} \!=\! \frac{\tanh(\lambda_{2}(E_{+}^{0})L/2)}{\tanh(\lambda_{1}(E_{+}^{0})L/2)}, \label{eq: transcendentalequation21} \\
&\frac{[C_{0} \!-\! M_{0} \!-\! E_{-}^{0} \!-\! D_{+}\lambda_{2}^{2}(E_{-}^{0}) ]\lambda_{1}(E_{-}^{0})}{[C_{0} \!-\! M_{0} \!-\! E_{-}^{0} \!-\! D_{+}\lambda_{1}^{2}(E_{-}^{0}) ]\lambda_{2}(E_{-}^{0})} \!=\! \frac{\tanh(\lambda_{2}(E_{-}^{0})L/2)}{\tanh(\lambda_{1}(E_{-}^{0})L/2)}. \label{eq: transcendentalequation22}
\end{align}
The numerical solutions of the transcendental equations (\ref{eq: transcendentalequation21}) and (\ref{eq: transcendentalequation22}) are shown in Fig.~\ref{Fig:E0zf}.
\begin{figure}[htpb]
\centering
\includegraphics[width=0.47\textwidth]{E0zf}
\caption{The eigen energies $E_{+}^{0}$ and $E_{-}^{0}$ as a function of the finite thickness $L$ for Bi$_{2}$Se$_{3}$. The red solid star points are for $E_{+}^{0}$ , and the blue solid circle points are for $E_{-}^{0}$. The numerical solutions are obtained by solving the transcendental equations (\ref{eq: transcendentalequation21}) and (\ref{eq: transcendentalequation22}).} \label{Fig:E0zf}
\end{figure}
In particular, for $L\rightarrow+\infty$ (infinitely thick case), we have the analytic solutions for Eqs.(\ref{eq: transcendentalequation21}) and (\ref{eq: transcendentalequation22}) as (the detail derivations can be found in the Appendix A)
\begin{align}\label{eq:E0zf}
E_{\pm}^{0}\xrightarrow{L\rightarrow+\infty} C_{0} + \frac{D_{1}M_{0} }{B_{1} } = 0.0296~({\rm eV}),
\end{align} which is consistent with the numerical results shown in Fig.~\ref{Fig:E0zf}.
For the infinitely thin case ($L\rightarrow0$), there will be no surface state which corresponds to no boundary condition. Besides, there is no wave function in the bulk and surface of the thin film physically. Therefore, there is no analytic solutions in the infinitely thin case.
Therefore, the eigen wavefunctions for $E_{+}^{0}$ and $E_{-}^{0}$ are, respectively,
\begin{align}\label{eq:wavefunctionup-finite}
\varphi(A_{1}) \equiv \Psi_{\uparrow}^{+} &= \bar{C}_{+} \left(
\begin{array}{cc}
- D_{+}\eta_{1}^{+}f_{-}^{+} \\
i A_{1}f_{+}^{+}
\end{array}
\right), \\
\chi(A_{1}) \equiv \Psi_{\uparrow}^{-} &= \bar{C}_{-} \left(
\begin{array}{cc}
- D_{+}\eta_{2}^{-}f_{+}^{-} \\
i A_{1}f_{-}^{-}
\end{array}
\right),
\end{align} where $\bar{C}_{\pm}$ is the normalization factor. The superscripts of $f_{\pm}^{\pm}$ and $\eta_{1,2}^{\pm}$ stand for $E_{\pm}^{0}$
and the subscripts of $f_{\pm}^{\pm}$ for parity, respectively. The expressions for $f_{\pm}^{\pm}$ and $\eta_{1,2}^{\pm}$ are given by
\begin{align}
f_{+}^{\pm}(z) &= \left[ \frac{\cosh(\lambda_{1}z)}{\cosh(\lambda_{1}L/2)} - \frac{\cosh(\lambda_{2}z)}{\cosh(\lambda_{2}L/2)} \right]_{E=E_{\pm}^{0}}, \\
f_{-}^{\pm}(z) &= \left[ \frac{\sinh(\lambda_{1}z)}{\sinh(\lambda_{1}L/2)} - \frac{\sinh(\lambda_{2}z)}{\sinh(\lambda_{2}L/2)} \right]_{E=E_{\pm}^{0}} , \\
\eta_{1}^{\pm} &= \frac{\lambda_{1}^{2} - \lambda_{2}^{2}}{\lambda_{1}\coth(\lambda_{1}L/2) - \lambda_{2}\coth(\lambda_{2}L/2) } \bigg|_{E=E_{\pm}^{0}}, \\
\eta_{2}^{\pm} &= \frac{\lambda_{1}^{2} - \lambda_{2}^{2}}{\lambda_{1}\tanh(\lambda_{1}L/2) - \lambda_{2}\tanh(\lambda_{2}L/2) } \bigg|_{E=E_{\pm}^{0}} .
\end{align}
\subsection{Effective Floquet Hamiltonian of the thin film}\label{6.2}
\begin{figure}[htpb]
\centering
\includegraphics[width=0.47\textwidth]{Fig1newnew.pdf}
\caption{A thin film of Bi$_{2}$Se$_{3}$ pumped by a circularly polarized optical field propagating to its surface. Here we suppose that $L\ll\lambda_{L}$ with the thickness $L$ of the thin film and the wavelength $\lambda_{L}$ of the driving optical field.} \label{Fig:Fig1}
\end{figure}
The energy spectra and wavefunctions of the lower block $h(A_{1})$ of $H_{0}$ can be obtained directly
by replacing $A_{1}$ by $- A_{1}$. Based on the above discussions, the four eigenstates of $H_{0}$ can be given by
\begin{align}
\Phi_{1}&= \left(
\begin{array}{cc}
\varphi(A_{1}) \\
0
\end{array}
\right),~~~~~~~\Phi_{2} = \left(
\begin{array}{cc}
\chi(A_{1}) \\
0
\end{array}
\right), \\
\Phi_{3} &= \left(
\begin{array}{cc}
0 \\
\varphi(-A_{1})
\end{array}
\right),~~~~~\Phi_{4} = \left(
\begin{array}{cc}
0 \\
\chi(-A_{1})
\end{array}
\right),
\end{align} with $\Phi_{1}\rightarrow\Phi_{3}$ and $\Phi_{2}\rightarrow\Phi_{4}$ under the time-reversal operation.
We should emphasize that these four solutions are for the surface states, and we use the four states as the basis states.
With the help of the four states, at the $\Gamma$ point, we can expand the Hamiltonian
Eq.~(\ref{eq:HF}) to obtain a new Floquet Hamiltonian of the thin film
\begin{align}\label{eq:HFfilm}
\!H_{\text{film}}^{(F)} &\!=\! \int_{\!-\!L/2}^{L/2}\!dz\!\{\Phi_{1}, \Phi_{4}, \Phi_{2}, \Phi_{3}\}^{\dagger} [\!H^{(F)}\!(\!{\bf k}\!)\!] \{\Phi_{1}, \Phi_{4}, \Phi_{2}, \Phi_{3}\} \nonumber\\
&\!=\! \left(
\begin{array}{cccc}
h_{f1}(k_{\perp}) & 0 \\
0 & h_{f2}(k_{\perp})
\end{array}
\right),
\end{align}
where
\begin{align}
\!h_{f1}(k_{\perp}) &\!=\! \tilde{E}_{1}^{0} \!-\! Dk_{\perp}^{2} \!+\! \!\left(\!
\begin{array}{cccc}
\frac{\tilde{\Delta}}{2} \!-\! Bk_{\perp}^{2} \!+\! \tilde{m} & i\gamma_{1} k_{-} \\
- i\gamma_{1} k_{+} & - \frac{\tilde{\Delta}}{2} \!+\! Bk_{\perp}^{2} \!-\! \tilde{m}
\end{array}
\!\right)\! \nonumber\\
&\!=\! (\tilde{E}_{1}^{0} \!-\! Dk_{\perp}^{2})\sigma_0 \!+\! \left(\frac{\tilde{\Delta}}{2} \!-\! Bk_{\perp}^{2} \!+\! \tilde{m} \right)\sigma_z \nonumber\\
&~\!+\! i\gamma_{1} k_{-}\sigma_{+} \!-\! i\gamma_{1} k_{+}\sigma_{-}, \\
\!h_{f2}(k_{\perp}) &\!=\! \tilde{E}_{2}^{0} \!-\! Dk_{\perp}^{2} \!+\! \!\left(\!
\begin{array}{cccc}
-\frac{\tilde{\Delta}}{2} \!+\! Bk_{\perp}^{2} \!+\! \tilde{m} & i\gamma_{2} k_{-} \\
- i\gamma_{2} k_{+} & \frac{\tilde{\Delta}}{2} \!-\! Bk_{\perp}^{2} \!-\! \tilde{m}
\end{array}
\!\right)\! \nonumber\\
&\!=\! (\tilde{E}_{2}^{0} \!-\! Dk_{\perp}^{2})\sigma_0 \!+\! \left(-\frac{\tilde{\Delta}}{2} \!+\! Bk_{\perp}^{2} \!+\! \tilde{m} \right)\sigma_z \nonumber\\
&~\!+\! i\gamma_{2} k_{-}\sigma_{+} \!-\! i\gamma_{2} k_{+}\sigma_{-},
\end{align} $\sigma_{\pm}=(\sigma_{x}\pm i\sigma_{y})/2$, $\tilde{E}_{1}^{0}\!=\!E^{0} \!-\! A_{0}^{2}D \!+\! \frac{m_{1} \!-\! m_{3}}{2}$, $\tilde{E}_{2}^{0}\!=\!E^{0} \!-\! A_{0}^{2}D \!-\! \frac{m_{1} \!-\! m_{3}}{2}$, $\frac{\tilde{\Delta}}{2}\!=\!\frac{\Delta}{2} \!-\! A_{0}^{2}B$, $\tilde{m}\!=\!\frac{m_{1} \!+\! m_{3}}{2} \!-\! A_{0}^{2}A_{2}^{2}\sin\varphi/(\hbar\omega)$, $E^{0}\!=\!(E_{+}^{0} \!+\! E_{-}^{0})/2$, $\Delta\!=\!E_{+}^{0} \!-\! E_{-}^{0}$, $B\!=\!(\tilde{B}_{1} \!-\! \tilde{B}_{2})/2$, $D\!=\!(\tilde{B}_{1} \!+\! \tilde{B}_{2})/2 \!-\! D_{2}$, $$\tilde{B}_{1} \!=\! B_{2}|\bar{C}_{+}|^{2} \int_{-L/2}^{L/2}dz \left( D_{+}^{2}|\eta_{1}^{+}f_{-}^{+}|^{2} \!-\! A_{1}^{2}|f_{+}^{+}|^{2} \right),$$ $$\tilde{B}_{2} \!=\! B_{2}|\bar{C}_{-}|^{2} \int_{-L/2}^{L/2}dz \left( D_{+}^{2}|\eta_{2}^{-}f_{+}^{-}|^{2} \!-\! A_{1}^{2}|f_{-}^{-}|^{2}\right),$$ $\gamma_{1}\!=\!\gamma \!-\! \gamma_{f}A_{0}^{2}\sin\varphi/(\hbar\omega)$, $\gamma_{2}\!=\!\gamma \!+\! \gamma_{f}A_{0}^{2}\sin\varphi/(\hbar\omega)$, $\gamma \!=\! A_{1}A_{2}\bar{C}_{+}^{*}\bar{C}_{-}D_{+} \int_{-L/2}^{L/2}dz \left( \eta_{1}^{+}f_{-}^{+}f_{-}^{-} \!+\! \eta_{2}^{-}f_{+}^{+}f_{+}^{-} \right)$, $\gamma_{f} = 2A_{1}A_{2}B_{2}\bar{C}_{+}^{*}\bar{C}_{-}D_{+} \int_{-L/2}^{L/2}dz \left( \eta_{1}^{+}f_{-}^{+}f_{-}^{-} - \eta_{2}^{-}f_{+}^{-}f_{+}^{+} \right)$, $m_{1} \!=\! m_{0}|\bar{C}_{+}|^{2} \int_{-L/2}^{L/2}dz~\left( D_{+}^{2}|\eta_{1}^{+}f_{-}^{+}|^{2} + A_{1}^{2}|f_{+}^{+}|^{2} \right) \!=\! m_{0}$, and $m_{3} \!=\! m_{0}|\bar{C}_{-}|^{2} \int_{-L/2}^{L/2}dz~\left( D_{+}^{2}|\eta_{2}^{-}f_{+}^{-}|^{2} \!+\! A_{1}^{2}|f_{-}^{-}|^{2} \right) \!=\! m_{0}$.
From the time-dependent Hamiltonian (\ref{eq:Ht}), it is found that the circularly polarized light is acting as an effective ``magnetic field''. Furthermore, from the Floquet thin-film Hamiltonian (\ref{eq:HFfilm}), we find that the ``effective field'' is causing kind of a Zeeman effect with the term $\tilde{m}\tau_{0}\otimes\sigma_{z}$ in (\ref{eq:HFfilm}) and an effective Zeeman field $\tilde{m}\!=\!m_{0}\!-\! A_{0}^{2}A_{2}^{2}\sin\varphi/(\hbar\omega)$.
The dispersions are
\begin{align}
E_{f1\pm}&\!=\!\tilde{E}_{1}^{0} \!-\! Dk_{\perp}^{2} \!\pm\! \sqrt{ \left(\frac{\tilde{\Delta}}{2} \!-\! Bk_{\perp}^{2} \!+\! \tilde{m}\right)^{2} \!+\! \gamma_{1}^{2}k_{\perp}^{2} }, \\
E_{f2\pm}&\!=\!\tilde{E}_{2}^{0} \!-\! Dk_{\perp}^{2} \!\pm\! \sqrt{ \left(\frac{\tilde{\Delta}}{2} \!-\! Bk_{\perp}^{2} \!-\! \tilde{m}\right)^{2} \!+\! \gamma_{2}^{2}k_{\perp}^{2} }.
\end{align}
At the $\Gamma$ point with $k_{\perp}=0$, the surface gap $\Delta_{\text{sur}}$ is defined as the minimum energy gap between the conduction and valence bands, i.e.,
\begin{align}\label{eq:sur}
\Delta_{\text{sur}}\!=\!\text{Min}\!\left[\!2\bigg|\frac{\tilde{\Delta}}{2} \!+\! \tilde{m}\bigg|,2\bigg|\frac{\tilde{\Delta}}{2} \!-\! \tilde{m}\bigg|,\left(\!\bigg|\frac{\tilde{\Delta}}{2} \!+\! \tilde{m}\bigg|\!+\!\bigg|\frac{\tilde{\Delta}}{2} \!-\! \tilde{m}\bigg|\!\right) \!\right]\!,
\end{align} where we use $m_{1}=m_{3}=m_{0}$.
Further, the wavefunctions of the two valence bands $E_{f1-}$ and $E_{f2-}$ are found to be
\begin{align}
\psi_{f1-} \!=\! G_{f1}\!\left(\!
\begin{array}{cccc}
b_{f1}(k_{\perp}) \\
-i\gamma_{1} k_{+} \\
0 \\
0
\end{array}
\!\right),~
\psi_{f2-} \!=\! G_{f2}\!\left(\!
\begin{array}{cccc}
0 \\
0 \\
b_{f2}(k_{\perp}) \\
i\gamma_{2} k_{+}
\end{array}
\!\right)\! ,
\end{align} where
\begin{align}
\!b_{f1}(\!k_{\perp}\!)\! &\!=\! \frac{\tilde{\Delta}}{2} \!-\! Bk_{\perp}^{2} \!+\! \tilde{m} \!-\! \sqrt{ \left(\!\frac{\tilde{\Delta}}{2} \!-\! Bk_{\perp}^{2} \!+\! \tilde{m}\!\right)^{2} \!+\! \gamma_{1}^{2}k_{\perp}^{2} }, \\
\!b_{f2}(\!k_{\perp}\!)\! &\!=\! \frac{\tilde{\Delta}}{2} \!-\! Bk_{\perp}^{2} \!-\! \tilde{m} \!+\! \sqrt{ \left(\!\frac{\tilde{\Delta}}{2} \!-\! Bk_{\perp}^{2} \!-\! \tilde{m}\!\right)^{2} \!+\! \gamma_{2}^{2}k_{\perp}^{2} }, \\
\!G_{f1} &\!=\! \frac{1}{\sqrt{ b_{f1}^{2}(k_{\perp}) \!+\! \gamma_{1}^{2}k_{\perp}^{2} }},~
\!G_{f2} \!=\! \frac{1}{\sqrt{ b_{f2}^{2}(k_{\perp}) \!+\! \gamma_{2}^{2}k_{\perp}^{2} }}.
\end{align}
\subsection{Chern number of a thin film with high-frequency pumping}\label{7.3}
In principle, we can find the Hall conductance for each $h_{f1/f2}(k_{\perp})$.
Note that $h_{f1/f2}(k_{\perp})$ in Eq.~(\ref{eq:HFfilm}) can be explicitly written as
\begin{align}
h_{f1}(k_{\perp})&\!=\!h_{f+}(k_{\perp})
\!=\! \tilde{E}_{1}^{0} \!-\! Dk_{\perp}^{2} \!+\! \sum_{i=x,y,z}d_{i}\sigma_{i},\\
h_{f2}(k_{\perp})&\!=\!h_{f-}(k_{\perp})
\!=\! \tilde{E}_{2}^{0} \!-\! Dk_{\perp}^{2} \!+\! \sum_{i=x,y,z}d_{i}\sigma_{i},
\end{align} where the subscripts are $f1=f+$, $f2=f-$, $\sigma_{i}$ are the Pauli matrices and the $\textbf{d}(k_{\perp})$ vectors are
\begin{align}
d_{x} &\!=\! \!-\! \left(\!-\! \gamma \!\pm\! \frac{A_{0}^{2}}{\hbar\omega}\gamma_{f}\sin\varphi \right) k_{y},\label{eq:dfx}\\
d_{y} &\!=\! \left(\!-\! \gamma \pm \frac{A_{0}^{2}}{\hbar\omega}\gamma_{f}\sin\varphi \right) k_{x},\label{eq:dfy}\\
d_{z} &\!=\!\pm \left( \frac{\tilde{\Delta}}{2} \!-\! Bk_{\perp}^{2}\right) \!+\! \tilde{m}.\label{eq:dfz}
\end{align}
For the $2\times2$ Hamiltonian in terms of the $\textbf{d}(k_{\perp})$ vectors and Pauli matrices, the Kubo formula for the Hall conductance can be generally expressed as~\cite{Lu2010,Qi2006}
\begin{align}\label{eq:HallconductanceF1}
\sigma_{xy} = \frac{e^{2}}{2\hbar} \int \frac{d^{2}\textbf{k}}{(2\pi)^{2}} \frac{(f_{k,c} - f_{k,\nu})}{d^{3}} \epsilon_{\alpha\beta\gamma} \frac{\partial d_{\alpha}}{\partial k_{x}} \frac{\partial d_{\beta}}{\partial k_{y}} d_{\gamma},
\end{align} where we use $\textbf{k}=(k_x,k_y)$, $k=k_{\perp}$, $\epsilon_{\alpha\beta\gamma}$ is the Levi-Civita anti-symmetric tensor, $d=\sqrt{d_{x}^{2} + d_{y}^{2} + d_{z}^{2} }$ is the norm of $(d_{x}, d_{y}, d_{z})$, $\hbar=h/(2\pi)$ is the reduced Planck's constant, $-e$ is the electron charge, $f_{k,c/\nu} = 1/\{\exp[(\varepsilon_{c/\nu}(k) - \mu)/(k_{B}T)] + 1\}$ is the Fermi distribution function of the conduction ($c$) and valence ($\nu$) bands with the chemical potential $\mu$, the Boltzmann constant $k_{B}$, and the temperature $T$ .
At zero temperature and when the chemical potential $\mu$ lies between $(-|\Delta_{\text{sur}}|/2, |\Delta_{\text{sur}}|/2)$, the Fermi functions reduce to $f_{k,c}= 0$ and $f_{k,\nu}= 1$. By substituting Eqs.~(\ref{eq:dfx})-(\ref{eq:dfz}) into (\ref{eq:HallconductanceF1}), we arrive at
\begin{align}\label{eq:HallconductanceF2}
\!\sigma_{xy}^{(\pm)} &\!=\! \!-\! \left(\! \pm\frac{e^{2}}{2h} \!\right) \nonumber\\
&\!\times\!\int_{0}^{\infty}\! \frac{\left(\!-\! \gamma \!\pm\! \frac{A_{0}^{2}}{\hbar\omega}\gamma_{f}\sin\varphi \!\right)^{2}\!\left(\! \frac{\tilde{\Delta}}{2} \!+\! Bk^{2} \pm \tilde{m} \!\right)\! kdk}{\left[\! \left(\! \!-\! \gamma \pm \frac{A_{0}^{2}}{\hbar\omega}\gamma_{f}\sin\varphi \!\right)^{2}\!k^{2} \!+\! \left(\! \frac{\tilde{\Delta}}{2} \!-\! Bk^{2} \!\pm\! \tilde{m} \!\right)^{2} \!\right]^{3/2}} ,
\end{align} where $+$ ($-$) corresponds to pseudo-spin-$\uparrow$ (pseudo-spin-$\downarrow$) or $h_{f+}(k_{\perp})$ ($h_{f-}(k_{\perp})$).
By defining
\begin{align}
\!\cos\theta \!=\! \frac{\left( \frac{\tilde{\Delta}}{2} \!-\! Bk^{2} \!\pm\! \tilde{m} \right)}{\!\left[\! \left( \!-\! \gamma \!\pm\! \frac{A_{0}^{2}}{\hbar\omega}\gamma_{f}\sin\varphi \right)^{2}k^{2} \!+\! \left( \frac{\tilde{\Delta}}{2} \!-\! Bk^{2} \!\pm\! \tilde{m} \!\right)^{2}\!\right]^{1/2}\!},
\end{align}
Eq.~(\ref{eq:HallconductanceF2}) can be transformed into
\begin{align}\label{eq:HallconductanceF3}
\sigma_{xy}^{(\pm)} = \pm\frac{e^{2}}{2h} \int_{0}^{\infty} d k^{2} \frac{\partial \cos\theta}{\partial k^{2}} .
\end{align}
The value of $\cos\theta$ at $k= 0$ and $k\rightarrow \infty$ only depends on the signs of $\frac{\tilde{\Delta}}{2}\pm\tilde{m}$ and $B$, respectively. As a result, in the insulating regime $-|\Delta_{\text{sur}}|/2 \leqslant \mu \leqslant |\Delta_{\text{sur}}|/2$, we find that the anomalous Hall conductance for each hyperbola has the form
\begin{align}\label{eq:anomalousHallConductance}
\sigma_{xy}^{(\pm)} =- \left( \pm\frac{e^{2}}{2h} \right) [\text{sgn}(\Delta_{\pm}) + \text{sgn}(B)],
\end{align} where $\Delta_{\pm}=\Delta_{c} \pm \tilde{m}$,
$\Delta_{c} = \frac{\tilde{\Delta}}{2} = \frac{\Delta}{2} - A_{0}^{2}B$, and
$\tilde{m} = \frac{m_{1}+m_{3}}{2} - A_{0}^{2}A_{2}^{2}\sin\varphi/(\hbar\omega)$.
In particular, the information of the interplay between the circularly polarized light and the bulk magnetic moment can be extracted from the analytical expression for the anomalous Hall conductance Eq.~(\ref{eq:anomalousHallConductance}). The competition between the polarized light and the bulk magnetic moment can change the first sign function [$\text{sgn}(\Delta_{\pm})$] in Eq.~(\ref{eq:anomalousHallConductance}). This interplay has significant impact on the anomalous Hall conductance or the topological phase diagram of Cr-doped Bi$_{2}$Se$_{3}$ thin film as shown in Fig.~\ref{Fig:PhaseDiagram3D}(b).
Therefore, the spin Chern numbers of the valence bands are given by
\begin{align}
C_{\pm} = - \left( \pm\frac{1}{2} \right) [\text{sgn}(\Delta_{\pm}) + \text{sgn}(B)].
\end{align}
Further, the total Chern number is given by
\begin{align}
C=C_{+}+C_{-}.
\end{align}
Since a net magnetic moment should also break time-reversal symmetry, it is important that the excitation decreases the magnetic moment as an effective magnetic moment $\tilde{m}\!=\!m_{0}\!-\! A_{0}^{2}A_{2}^{2}\sin\varphi/(\hbar\omega)$, where we choose $\varphi=\pi/2$. Therefore, the effective magnetic moment decreases with the increasing of the light intensity $A_{0}$.
\section{Phase diagram}\label{7}
\begin{figure}[htpb]
\centering
\includegraphics[width=0.48\textwidth]{PhaseDiagram3Dm0_new.pdf} \\
\includegraphics[width=0.48\textwidth]{PhaseDiagram3Dm30_new.pdf}
\caption{The total Chern number as functions of the thin-film thickness and the light intensity.
(a) Topological phase diagram for the non-doped Bi$_{2}$Se$_{3}$ with $m_{0}=0$.
(b) Topological phase diagram for the Cr-doped Bi$_{2}$Se$_{3}$ with $m_{0}=30$ meV.
The other parameters are given as $\varphi=\pi/2$ and $\hbar\omega\approx150$ meV ($\omega \sim 2.2789\times10^{2}$ THz).} \label{Fig:PhaseDiagram3D}
\end{figure}
As shown in Fig.~\ref{Fig:PhaseDiagram3D}, we calculate the total Chern number as functions of the thin-film thickness and the light intensity for Bi$_{2}$Se$_{3}$.
For the non-doped Bi$_{2}$Se$_{3}$ with $m_{0}=0$ as shown in Fig.~\ref{Fig:PhaseDiagram3D}(a), by modulating the strength of the polarized optical field, there are three different regions: the NI phase with the spin Chern numbers $C_{\pm}=0$, the QAHI phase with the spin Chern numbers $C_{+}=1$ and $C_{-}=0$, and the QSHI phase with the spin Chern numbers $C_{\pm}=\pm1$. In the absent of optical field with $A_{0}=0$, there are only two phases: NI phase and QSHI phase. The result reveals that the circularly polarized light breaks the time-reversal symmetry and induces the QAHI phase. Moreover, the light-induced time-reversal-symmetry-broken QSHI phase is different from the time-reversal-symmetry QSHI phase in the absent of optical field.
For the Cr-doped Bi$_{2}$Se$_{3}$ with $m_{0}=30$ meV, it is indicated from Fig.~\ref{Fig:PhaseDiagram3D}(b) that there are four different regions: the NI phase, the time-reversal-symmetry-broken QSHI phase, and two different QAHI phases with opposite nonzero Chern numbers. Here the interplay between the light and bulk magnetic moment can separate two different QAHI phases with opposite Chern numbers.
For a given layer thickness $L=3.2$ nm, one can apparently detect the tendency towards light-induced band inversion upon increasing optical field intensity and passing through the QSHI phase region marked by gray.
In particular, the gray regions between the fixed points are the QSHI phases with the Chern number $C=0$. With increasing the thickness $L$, the gray regions will vanish.
\begin{figure}[htpb]
\centering
\includegraphics[width=0.48\textwidth]{E_thinfilm_together}
\caption{Dispersions of the thin film for the Cr-doped Bi$_{2}$Se$_{3}$ under open boundary condition along $y$ direction and periodic boundary conditions along $x$ direction. (a) Bands for the Floquet Hamiltonian. (b) Bands for the pseudo-spin-$\uparrow$ (``+'') Floquet Hamiltonian. (c) Bands for the pseudo-spin-$\downarrow$ (``-'') Floquet Hamiltonian.
The other parameters are given as $A_{0}=0.036$ \AA$^{-1}$, $L=3$ nm, $a_x=a_y=1$ \AA, $N_y=200$, $m_{0}=30$ meV, $\varphi=\pi/2$, and $\hbar\omega\approx0.15$ eV ($\omega \sim 2.2789\times10^{2}$ THz) for chromium-doped Bi$_{2}$Se$_{3}$. Here, $a_x$ and $a_y$ are the lattice constants along $x$ and $y$ directions respectively.} \label{Fig:E_thinfilm}
\end{figure}
Furthermore, we calculate the energy dispersions of the thin film for the Cr-doped Bi$_{2}$Se$_{3}$ with $A_{0}=0.036$ \AA$^{-1}$ and $L=3$ nm under open boundary condition along $y$ direction and periodic boundary conditions along $x$ direction (the detail derivations can be found in the Appendix B).
As shown in Fig.~\ref{Fig:E_thinfilm}, it is found that there exist the edge states in the thin film as shown in Fig.~\ref{Fig:E_thinfilm}(a), and the edge states come from the pseudo-spin-$\uparrow$ (``$+$'') Floquet Hamiltonian in the chosen parameters here as shown in Fig.~\ref{Fig:E_thinfilm}(b). However, there is no edge states in the energy spectra of the pseudo-spin-$\downarrow$ (``$-$'') Floquet Hamiltonian as shown in Fig.~\ref{Fig:E_thinfilm}(c).
\section{Probation of the topological transitions}\label{8}
\begin{figure}[htpb]
\centering
\includegraphics[width=0.48\textwidth]{nonlocal_new.pdf}\\
\includegraphics[width=0.48\textwidth]{nonlocal_2_new.pdf}
\caption{
(a) QAHI with $C_{-}=0$ and $C_{+}=1$ which corresponds to pseudo-spin-$\uparrow$ (``$+$'') current on the side edges. Here, the arrow denotes the direction of the current.
(b) QAHI with $C_{+}=0$ and $C_{-}=-1$ which corresponds to pseudo-spin-$\downarrow$ (``$-$'') current on the side edges.
(c) QSHI has opposite currents with different pseudo-spins on the side edges.
(d) NI with no pseudo-spin currents.
(e) The proposed device for nonlocal edge measurements with the electrodes (green cuboids).
(f) Schematics of the experimental setup with the electrodes (green rectangles). Here, $n_L$ and $n_R$ are the coefficients of the transmission probability matrix, and the number of the clockwise and anti-clockwise pseudo-spin currents are determined by $n_L$ and $n_R$.
(g)-(i) Electric-circuit diagrams with a bias voltage leads on electrodes 1 and 4, and the currents which leads on electrodes 2, 3, 5, and 6 are set to zero. In particular, the electric-circuit diagram (i) is used to measure the Hall resistance.} \label{Fig:nonlocal}
\end{figure}
Physically, as shown in Fig.~\ref{Fig:nonlocal}(a)-(d), the quantum anomalous Hall insulator with the Chern number $C=1$ has a pseudo-spin-$\uparrow$ (``$+$'') current on the side edges~\cite{XueExp2013,Qi2008,Yu2010,Chu2011,Qiao2010,Qi2006,Onoda2006,Nomura2011,Checkelsky2012,Hor2010,Chen2010,Wray2011,Wang2013,Wang2014,Law2017,Kim2018,Kawamura2018},
the quantum anomalous Hall insulator with the Chern number $C=-1$ has a pseudo-spin-$\downarrow$ (``$-$'') current on the side edges, and the quantum spin Hall insulator has antiparallel currents with different pseudo-spins on the side edges~\cite{Roth2009,Chen2021prb,Piskunow2021prl}. Here, each current is responsible for the $e^2/h$ Hall conductance as shown in Eq.~(\ref{eq:anomalousHallConductance}). However, the normal insulator has no pseudo-spin currents which correspond to a zero Hall conductance.
Detecting the pseudo-spin current can help to identify the topological transitions.
An experimental set up to measure the pseudo-spin current is given in Fig.~\ref{Fig:nonlocal}(d): The side surface of the device has six electrodes as 1-6 (green cuboids). Further, one needs to measure the nonlocal edge transport~\cite{Roth2009,Chen2021prb,Piskunow2021prl,Chen2012prb}, i.e., the nonlocal resistances. As shown in Fig.~\ref{Fig:nonlocal}(e) and (f), a current $I_{ij}$ is applied between electrodes $i$ and $j$ on the side surface, and a nonlocal voltage $V_{ij}$ between other two electrodes $i'$ and $j'$ is measured to define a nonlocal resistance $R_{ij,i'j'} = V_{i'j'}/I_{ij}$~\cite{Roth2009,Chen2021prb,Piskunow2021prl,Chen2012prb,Datta1997book}.
Within the Landauer-B{\"u}ttiker-Fisher-Lee formalism~\cite{Landauer1970pm,Buttiker1988prb,Buttiker1986prl,Fisher1981prb}, the current $I_{i}$ flowing out of the $i$th electrode into the sample region is expressed as~\cite{Roth2009,Chen2021prb,Piskunow2021prl,Chen2012prb,Datta1997book}
\begin{align}\label{eq:LBFL}
I_{i}=\frac{e^2}{h}\sum_{j,j\neq i}T_{ij}(E_F)(V_{i} - V_{j}),
\end{align}
where $V_i$ is the voltage on the $i$th electrode and $T_{ij}(E_F)$ is the transmission probability matrix from the $i$th to the $j$th electrode at Fermi energy $E_F$. Notice that, in a time-reversal-invariant system, the transmission
coefficients satisfy the condition $T_{ij}=T_{ji}$~\cite{Roth2009}.
For the present case here, there are six electrodes on the side surface, so $T_{ij}$ is a $6\times 6$ matrix.
Concretely, Eq.~(\ref{eq:LBFL}) can be written in a matrix form
\begin{widetext}\begin{align}
\begin{pmatrix}
I_1 \\
I_2 \\
I_3 \\
I_4 \\
I_5 \\
I_6
\end{pmatrix}=\frac{e^2}{h}\begin{pmatrix}
\sum_{j=1}^{6}T_{1j} & -T_{12} & -T_{13} & -T_{14} & -T_{15} & -T_{16} \\
-T_{21} & \sum_{j=1}^{6}T_{2j} & -T_{23} & -T_{24} & -T_{25} & -T_{26} \\
-T_{31} & -T_{32} & \sum_{j=1}^{6}T_{3j} & -T_{34} & -T_{35} & -T_{36} \\
-T_{41} & -T_{42} & -T_{43} & \sum_{j=1}^{6}T_{4j} & -T_{45} & -T_{46} \\
-T_{51} & -T_{52} & -T_{53} & -T_{54} & \sum_{j=2}^{6}T_{5j} & -T_{56} \\
-T_{61} & -T_{62} & -T_{63} & -T_{64} & -T_{65} & \sum_{j=1}^{6}T_{6j}
\end{pmatrix}
\begin{pmatrix}
V_1 \\
V_2 \\
V_3 \\
V_4 \\
V_5 \\
V_6
\end{pmatrix},
\end{align}\end{widetext} where $T_{11}=T_{22}=T_{33}=T_{44}=T_{55}=T_{66}=0$.
If the six electrodes are identical, we have $T_{j+1,j}=T_{1,6}=n_R$ and $T_{j,j+1}=T_{6,1}=n_L$ with $j=1,2,\cdots,5$; $T_{j+2,j}=T_{2,6}=T_{1,5}$ and $T_{j,j+2}=T_{6,2}=T_{5,1}$ with $j=1,2,3,4$; $T_{j+3,j}=T_{3,6}$ and $T_{j,j+3}=T_{6,3}$, i.e., $T_{j+3,j}=T_{j,j+3}$ with $j=1,2,3$. For the non-neighboring terminals, we set that $T_{ij}=0$. Therefore, the corresponding transmission matrix has the form
\begin{align}
&T=\nonumber\\
&\!\begin{pmatrix}
n_L\!+\!n_R & \!-\!n_L & 0 & 0 & 0 & \!-\!n_R \\
\!-\!n_R & n_L\!+\!n_R & \!-\!n_L & 0 & 0 & 0 \\
0 & \!-\!n_R & n_L\!+\!n_R & \!-\!n_L & 0 & 0 \\
0 & 0 & \!-\!n_R & n_L\!+\!n_R & \!-\!n_L & 0 \\
0 & 0 & 0 & \!-\!n_R & n_L\!+\!n_R & \!-\!n_L \\
\!-\!n_L & 0 & 0 & 0 & \!-\!n_R & n_L\!+\!n_R
\end{pmatrix}.\!
\end{align}
As shown in Fig.~\ref{Fig:nonlocal}(g)-(i), we can apply a bias voltage leads on electrodes 1 and 4, and the currents which leads on electrodes 2, 3, 5, and 6 are set to zero. Owing to the current conservation, one finds that $I_1=-I_4\equiv I_{14}$. We set $V_5 = 0$, which allows us to truncate the fifth row and the fifth column of the matrix and write
\begin{align}\label{eq:Vs}
\begin{pmatrix}
V_1 \\
V_2 \\
V_3 \\
V_4 \\
V_6
\end{pmatrix}=\frac{h}{e^2}T_{d}^{-1}\begin{pmatrix}
I_{14} \\
0 \\
0 \\
-I_{14} \\
0
\end{pmatrix},
\end{align} where
\begin{align}
&T_{d}=\nonumber\\
&\!\begin{pmatrix}
n_L\!+\!n_R & \!-\!n_L & 0 & 0 & \!-\!n_R \\
\!-\!n_R & n_L\!+\!n_R & \!-\!n_L & 0 & 0 \\
0 & \!-\!n_R & n_L\!+\!n_R & \!-\!n_L & 0 \\
0 & 0 & \!-\!n_R & n_L\!+\!n_R & 0 \\
\!-\!n_L & 0 & 0 & 0 & n_L\!+\!n_R
\end{pmatrix}\!.
\end{align}
Then the voltages of the six terminals are given by Eq.~(\ref{eq:Vs}) as
\begin{align}
V_1&=\frac{hI_{14}}{e^2}\frac{n_R}{n_L^2+n_R^2-n_Ln_R},\\
V_2&=\frac{hI_{14}}{e^2}\frac{n_R^2+n_Ln_R-n_L^2}{n_L^3+n_R^3},\\
V_3&=\frac{hI_{14}}{e^2}\frac{n_R-n_L}{n_L^2+n_R^2-n_Ln_R},\\
V_4&=-\frac{hI_{14}}{e^2}\frac{n_L^2}{n_L^3+n_R^3},\\
V_5&=0,\\
V_6&=\frac{hI_{14}}{e^2}\frac{n_Ln_R}{n_L^3+n_R^3},
\end{align}
where $n_L$ and $n_R$ are the coefficients of the transmission probability matrix. They determine the number of the
clockwise and anti-clockwise pseudo-spin currents as shown in the Fig.~\ref{Fig:nonlocal}(f).
For the QAHI with $C=1$, we have $n_L = 1$ and $n_R = 0$. For the QAHI with $C=-1$, we have $n_L = 0$ and $n_R = 1$. For the QSHI, we have $n_L = n_R = 1$. For the NI, we have $n_L = n_R = 0$. As shown in Fig.~\ref{Fig:nonlocal}(g)-(i), the nonlocal resistances can be analytically found as
\begin{align}
R_{14,14}&=\frac{V_1 - V_4}{I_{14}}=\frac{h}{e^2}\frac{n_L^2 + n_R^2 + n_Rn_L}{n_L^3 + n_R^3},\\
R_{14,23}&=\frac{V_2 - V_3}{I_{14}}=\frac{h}{e^2}\frac{n_{L}n_{R}}{n_L^3 + n_R^3},\\
R_{14,26}&=\frac{V_2 - V_6}{I_{14}}=\frac{h}{e^2}\frac{n_R-n_L}{n_L^2 + n_R^2 - n_{L}n_{R}}.
\end{align}
For the QAHI with $C=1$, we have $R_{14,14} = \frac{h}{e^2}$, $R_{14,23} = 0$, and $R_{14,26} = -\frac{h}{e^2}$. For the QAHI with $C=-1$, we have $R_{14,14} = \frac{h}{e^2}$, $R_{14,23} = 0$, and $R_{14,26} = \frac{h}{e^2}$. For the QSHI, we obtain $R_{14,14} = \frac{3h}{2e^2}$, $R_{14,23} = \frac{h}{2e^2}$, and $R_{14,26} = 0$. For the NI, we have $R_{14,14} = R_{14,23} = R_{14,26} = 0$.
\section{Summary}\label{9}
We investigate the impact of high-frequency pumping on a ferromagnetic topological insulator with the circularly polarized optical field. It is found that the intensity of the circularly polarized light can be used as a knob to drive a topological transition.
With modulating the strength of the polarized optical field in an experimentally accessible range, there are four different regions: the NI phase, the time-reversal-symmetry-broken QSHI phase, and two different QAHI phases. This is different from the situation in the absent of optical field.
We propose an experimental scheme to manipulate the topological phases in Cr-doped Bi$_{2}$Se$_{3}$ with high-frequency pumping light. Our proposal can be realized in an experimentally accessible range. Particularly, to realize the light driven topological phases, the frequency and intensity of the light are both within the experimental accessibility~\cite{lightexp2013,lightexp2016}. In most of the recent experiments~\cite{XueExp2013,CheckelskyExp2014,QAHIexp2014,QAHIexp2015,XueExp2015,YoshimiExp2015,ChangExp2016,GrauerExp2017,XueExp2018,ZhangExp2020,QAHIexp2020,ZhaoExp2020,MogiExp2021}, people focus on using the magnetic fields to manipulate the topological phases of the system.
However, the topological phases obtained in this way may be confused with the quantum Hall effect. Luckily, our proposal avoids this. Therefore, the theoretically investigations we put forward will be helpful to the future experiments.
Note Added. When finishing this paper, we became aware of another preprint, Ref.~\cite{Liu2021arxive}, which also did the Floquet engineering of magnetism in topological insulator thin films.
\begin{acknowledgments}
Fang Qin and Rui Chen contribute equally to this work. We acknowledge helpful discussions with Dong-Hui Xu, Hai-Peng Sun, Xiao-Bin Qiang, De-Huan Cai, and Hui Xiong.
This work was supported by the National Natural Science Foundation of China (Grants No.~11534001, No.~11925402, and No.~11404106), the Strategic Priority Research Program of Chinese Academy of Sciences (Grant No.~XDB28000000), Guangdong province (Grants No.~2016ZT06D348 and No.~2020KCXTD001), the National Key R \& D Program (Grant No.~2016YFA0301700), Shenzhen High-level Special Fund (Grants No.~G02206304 and No.~G02206404), and the Science, Technology and Innovation Commission of Shenzhen Municipality (No.~ZDSYS20190902092905285, No.~ZDSYS20170303165926217, No.~JCYJ20170412152620376, and No.~KYTDPT20181011104202253).
F.Q. acknowledges support from the project funded by the China Postdoctoral Science Foundation (Grant No.~2019M662150 and No.~2020T130635) and the SUSTech Presidential Postdoctoral Fellowship.
R.C. acknowledges support from the project funded by the China Postdoctoral Science Foundation (Grant No.~2019M661678) and the SUSTech Presidential Postdoctoral Fellowship.
\end{acknowledgments}
|
\section{Introduction}
Efficient numerical methods for
electromagnetic wave propagation are central to a wide range of
applications in science and technology \cite{assous_ciarlet_labrunie_2018a,griffiths_1999a}.
For wave phenomena with harmonic time dependence, governed by a single angular frequency $\omega > 0$, the electromagnetic wave field satisfies time-harmonic Maxwell's equations in a
domain $\Omega \subset \mathbb R^3$: Given a current density $\boldsymbol j: \Omega \to \mathbb C^3$,
we seek two vector fields $\boldsymbol e,\boldsymbol h: \Omega \to \mathbb C^3$ such that
\begin{subequations}
\label{eq_maxwell_freq_strong}
\begin{equation}
\label{eq_maxwell_freq_strong_volume}
\left \{
\begin{array}{rcl}
i\omega \boldsymbol \varepsilon \boldsymbol e + \boldsymbol \sigma \boldsymbol e + \grad \times \boldsymbol h &=& \boldsymbol j,
\\
i\omega \boldsymbol \mu \boldsymbol h - \grad \times \boldsymbol e &=& \boldsymbol 0,
\end{array}
\right .
\end{equation}
inside the computational domain $\Omega$, where the first-order tensors $\boldsymbol \varepsilon$, $\boldsymbol \sigma$ and $\boldsymbol \mu$
are the permittivity, conductivity and permeability
of the medium in $\Omega$. At the boundary $\partial \Omega$
of $\Omega$, divided into two disjoint sets ${\Gamma_{\rm P}}$ and ${\Gamma_{\rm I}}$, we impose
the boundary conditions
\begin{equation}
\label{eq_maxwell_freq_strong_boundary}
\left \{
\begin{array}{rcll}
\boldsymbol e \times \boldsymbol n &=& \boldsymbol 0 & \text{ on } {\Gamma_{\rm P}},
\\
\boldsymbol e \times \boldsymbol n + \boldsymbol Z \boldsymbol h_\tau &=& \boldsymbol g & \text{ on } {\Gamma_{\rm I}},
\end{array}
\right .
\end{equation}
\end{subequations}
where $\boldsymbol n$ stands for the outward unit normal to $\partial \Omega$
and $\boldsymbol h_\tau := \boldsymbol n \times (\boldsymbol h \times \boldsymbol n)$. Here, the
first-order tensor $\boldsymbol Z$, defined on ${\Gamma_{\rm I}}$, describes a surface impedance
while $\boldsymbol g: {\Gamma_{\rm I}} \to \mathbb C^3$ typically represents
incident electromagnetic field. The PEC condition on ${\Gamma_{\rm P}}$ corresponds to the surface of
a perfectly conducting material whereas the impedance
boundary condition on ${\Gamma_{\rm I}}$ either models the boundary of an imperfect
conductor or corresponds to an approximation of the Silver-M\"uller radiation
condition \cite{colton_kress_2012a}. Note that ${\Gamma_{\rm P}}$ or ${\Gamma_{\rm I}}$ may be empty.
In heterogeneous media with intricate
geometries, Galerkin discretizations based on variational formulations of
\eqref{eq_maxwell_freq_strong}, such as curl-conforming finite elements or
discontinous Galerkin (DG) methods \cite{li_lanteri_perrussel_2014a,monk_2003a},
probably are the most flexible and competitive approaches
currently available. If $\omega$ is ``large'' and
the computational domain spans many wavelengths,
resolving the wavelength and limiting dispersion errors requires the
use of highly refined meshes coupled with high-order elements
\cite{chaumontfrelet_nicaise_2019a,melenk_sauter_2020a}. Hence, the high-frequency regime
typically leads to large, sparse, indefinite and ill-conditioned linear systems which need to be solved numerically
by direct or iterative methods.
Although considerable progress has been achieved over the past decades
\cite{amestoy_ashcraft_boiteau_buttari_lexcellent_weisbecker_2015a,amestoy_duff_lexcellent_2000a},
the parallel implementation of scalable direct solvers remains a challenge when the number of
unknowns is large. On the other hand, the design of robust and efficient preconditioners for
iterative solvers is a delicate task \cite{ernst_gander_2012a}. Recent developments
include domain decomposition \cite{bonazzoli_dolean_graham_spence_tournier_2019a,li_lanteri_perrussel_2014a},
shifted-laplacian \cite{gander_graham_spence_2015a}, and sweeping \cite{tsuji_engquist_ying_2012a}
preconditioners. Still, the efficient solution of 3D time-harmonic Maxwell's equations with hetereogeneous
coefficients remains to this day a formidable challenge, especially in the high-frequency regime.
To avoid these difficulties, we instead transform \eqref{eq_maxwell_freq_strong} back to the time-domain and
consider its time-dependent counterpart
\begin{equation}
\label{eq_maxwell_time_strong}
\left \{
\begin{array}{rcll}
\boldsymbol \varepsilon \dot \boldsymbol E + \boldsymbol \sigma \boldsymbol E + \grad \times \boldsymbol H &=& \boldsymbol J & \text{ in } \mathbb R_+ \times \Omega,
\\
\boldsymbol \mu \dot \boldsymbol H - \grad \times \boldsymbol E &=& \boldsymbol 0 & \text{ in } \mathbb R_+ \times \Omega,
\\
\boldsymbol E \times \boldsymbol n &=& \boldsymbol 0 & \text{ on } \mathbb R_+ \times {\Gamma_{\rm P}},
\\
\boldsymbol E \times \boldsymbol n + \boldsymbol Z \boldsymbol H_\tau &=& \boldsymbol G &\text{ on } \mathbb R_+ \times {\Gamma_{\rm I}},
\end{array}
\right .
\end{equation}
with time-harmonic forcing
$\boldsymbol J(t,\boldsymbol x) := \Re \left \{\boldsymbol j(\boldsymbol x) e^{i\omega t}\right \}$,
$\boldsymbol G(t,\boldsymbol x) := \Re \left \{\boldsymbol g(\boldsymbol x) e^{i\omega t}\right \}$,
and initial conditions $\boldsymbol E|_{t=0} = \boldsymbol E_0$ and $\boldsymbol H|_{t=0} = \boldsymbol H_0$ yet to
be specified. The key advantage of this strategy is that it only requires the
solution of a time evolution problem for which efficient numerical schemes, such as
finite differences \cite{taflove_hagness_2005a,yee_1966a} or DG
\cite{fezoui_lanteri_lohrengel_piperno_2005a,grote_schneebeli_schotzau_2007a,hesthaven_warburton_2002a}
discretizations coupled with explicit time integration, can be utilized.
As these algorithms are inherently parallel with a low memory footprint,
they are extremely attractive on modern computer architectures.
In this context, a simple and common approach follows from the limiting amplitude principle
\cite{morawetz_1962a}, which states under suitable assumptions that the solution of \eqref{eq_maxwell_time_strong} ``converges'' to the time-harmonic solution
in the sense that $\boldsymbol E(t,\boldsymbol x) \to \Re \left \{\boldsymbol e(\boldsymbol x)e^{i\omega t}\right \}$ and
$\boldsymbol H(t,\boldsymbol x) \to \Re \left \{\boldsymbol h(\boldsymbol x)e^{i\omega t}\right \}$
as $t \to +\infty$. Thus, to solve \eqref{eq_maxwell_freq_strong} one can simply
simulate time-dependent Maxwell's equations
for a ``sufficiently long'' time and eventually extract the time-harmonic solution.
However, as the final simulation time
required to obtain
an accurate approximation may be very large, especially near resonances or in the presence of trapping geometries, the usefulness of this approach is somewhat limited \cite{bardos_rauch_1994a}.
Both controllability methods and fixed-point iterations
have been proposed to accelerate convergence and determine initial conditions ($\boldsymbol E_0,\boldsymbol H_0$)
which render the time-dependent solution $T$-periodic with period $T := 2\pi/\omega$.
Inspired by the seminal work in \cite{lions_1988a}, controllability
methods (CM) \cite{bristeau_glowinski_periaux_1994a,bristeau_glowinski_periaux_1998a} reformulate
the controllability problem as a minimization problem
for a quadratic cost functional $J(\boldsymbol E_0,\boldsymbol H_0)$, which measures the
misfit between $(\boldsymbol E_0,\boldsymbol H_0)$ and the time-dependent solution $(\boldsymbol E(T),\boldsymbol H(T))$
after one period. Then, the functional $J$ is minimized
by a conjugate gradient (CG) iteration, which leads to the combined
controllability method-CG algorithm, or CMCG for short. Alternatively, fixed-point iterations
determine the $T$-periodic solution by applying a judicious filtering operator at
each iteration to achieve convergence \cite{AppeloGarciaRunborg, peng2021emwaveholtz}. As the convergence of fixed-point iterations can be slow
near resonances or in the presence of trapping geometries, an outer CG or GMRES Krylov subspace method
must be applied, depending on boundary condititions.
When using the controllability approach, one faces two central questions:
efficient computation of the gradient $J'$ and uniqueness of the time-periodic solution.
As early work on CMCG methods was restricted to scattering problems
from acoustics \cite{bristeau_glowinski_periaux_1994a,bristeau_glowinski_periaux_1998a}
or electromagnetics \cite{bristeau:inria-00073072} in second-order formulation, the computation
of $J'$ always required the solution of a strongly elliptic (coercive) problem.
To avoid solving that additional
elliptic problem at each CG iteration, the controllability method was later applied to the Helmholtz
equation in first-order formulation \cite{kahkonen_glowinski_rossi_makinen_2011a} using
Raviart-Thomas FE for the spatial discretization; due to the lack of available mass-lumping,
however, the mass-matrix then needed to be inverted at each time-step during the
time integration. By combining a first-order formulation with a DG discretization, a scalable
parallel formulation was recently derived \cite{grote_nataf_tang_tournier_2020a},
which completely avoids the need for solving any elliptic problem or inverting the mass-matrix.
In general, the $T$-periodic solution of \eqref{eq_maxwell_time_strong} is not unique
and hence does not necessarily yield the desired (unique) time-harmonic solution of
\eqref{eq_maxwell_freq_strong}. For sound-soft acoustic scattering, where Dirichlet
and impedance conditions are imposed on distinct parts of the boundary, the $T$-periodic
solution in fact is unique and the one-to-one correspondence is therefore immediate.
For other boundary-value problems, however,
such as sound-hard scattering or problems in bounded physical domains,
the periodic solution is generally no longer unique, as it may contain additional
($T$-periodic) spurious modes. Two ideas have been
proposed as a remedy to extend the CMCG approach to arbitrary boundary conditions.
First, uniqueness can be restored by modifying $J$, though at a small price in
the computation of its gradient \cite{bardos_rauch_1994a,grote_tang_2019a}. Alternatively,
a cheap filtering operator can be applied as a post-processing step to any minimizer of $J$,
which removes any spurious modes \cite{grote_nataf_tang_tournier_2020a,tang_2020a}
and thus restores uniqueness using the original cost functional $J$.
Here we propose a CMCG method for time-harmonic
Maxwell's equations \eqref{eq_maxwell_freq_strong} in their first order formulation,
which completely avoids the solution of any elliptic problem, and combine it with a
post-processing filtering step to guarantee uniqueness, regardless of the boundary conditions.
Moreover, thanks to a DG discretization in space, the mass-matrix is automatically block-diagonal.
Hence, the resulting CMCG algorithm is inherently parallel and scalable but also
guaranteed to converge to the time-harmonic solution
starting from any initial guess, as long as time-harmonic Maxwell's equations
\eqref{eq_maxwell_freq_strong} are well-posed for the frequency $\omega$ under
consideration.
The remainder of this work is organized as follows. We provide a formal
description of the algorithm and a discussion of our key theoretical
results in Section \ref{section_description}. As the mathematical framework required to
rigorously define and analyze Maxwell's equations is rather involved,
the precise description and preliminary results are postponed to Section \ref{section_model}.
Section \ref{section_perio} contains the bulk of the theory, where we carefully
analyze the relation between the time-harmonic and time-periodic solutions.
Here, our contributions are twofold. On the one hand, we identify configurations
of boundary conditions and right-hand sides for which the unique time-periodic solution
coincides with the time-harmonic solution. On the other hand, we show that
the filtering procedure introduced in \cite{grote_nataf_tang_tournier_2020a,tang_2020a}
always recovers the time-harmonic solution from any minimizer, as long as
\eqref{eq_maxwell_freq_strong} is well-posed.
In Section \ref{section_CMCG}, we describe in detail our CMCG method and establish
its convergence toward the time-harmonic solution.
In Section \ref{section_numer},
we present various numerical experiments highlighting the performance of the proposed CMCG algorithm.
Here, we benchmark the proposed CMCG algorithm against the limiting amplitude principle,
where pure time-marching (without controllability) is utilized, as both methods are
non-invasive and easily integrated with any existing time-marching code; in contrast, efficient preconditioners typically require an important and dedicated implementation effort. Finally, we provide
in Section \ref{section_conclusion} some concluding remarks.
\section{Main results}
\label{section_description}
Throughout this work, we adopt the notation $U = (\boldsymbol e,\boldsymbol h)$ for a time-harmonic
electromagnetic field, while the calligraphic font $\mathscr U = (\boldsymbol E,\boldsymbol H)$ is reserved
for time-dependent fields. It is easily seen that if $U$ is a time-harmonic
field solution to \eqref{eq_maxwell_freq_strong} with right-hand side $\boldsymbol j$ and $\boldsymbol g$, then
$\mathscr U(t,\boldsymbol x) := \Re \{U(\boldsymbol x) e^{i\omega t}\}$ is the solution of time-dependent
Maxwell's equations \eqref{eq_maxwell_time_strong} with right-hand side
$\boldsymbol J(t,\boldsymbol x) := \Re \{\boldsymbol j(\boldsymbol x) e^{i\omega t}\}$, $\boldsymbol G(t,\boldsymbol x) := \Re \{\boldsymbol g(\boldsymbol x) e^{i\omega t}\}$,
and initial condition $\mathscr U_0 := \Re U$.
The CMCG algorithm hinges on an idea that is essentially the converse of the above
statement. Namely, we seek an initial condition $\mathscr U_0$ such that the resulting time-dependent
field $\mathscr U$ (with right-hand sides $\boldsymbol J$ and $\boldsymbol G$ as above) is time-periodic, with period
$T := 2\pi/\omega$.
Let $P_{\boldsymbol j,\boldsymbol g,\omega}: \mathscr U_0 \to \mathscr U(T)$ denote the (affine) operator mapping the initial
condition $\mathscr U_0$ to the solution $\mathscr U$ of \eqref{eq_maxwell_time_strong} with
time-harmonic right-hand sides $\boldsymbol J$ and $\boldsymbol G$ evaluated at time $T$. Then, the
``controllability method'' corresponds to solving (linear) equation
$P_{\boldsymbol j,\boldsymbol g,\omega} \mathscr U_0 = \mathscr U_0$.
At this point, three main questions arise. First, if the time-dependent solution with
initial condition $\mathscr U_0$ is periodic, can we ensure that $\mathscr U_0 = \Re U$, where $U$
is the corresponding frequency-domain solution? Second, can we design an efficient
algorithm to solve for $P_{\boldsymbol j,\boldsymbol g,\omega} \mathscr U_0 = \mathscr U_0$? Finally, can we prove the
convergence of this algorithm?
\subsection{The structure of periodic solutions}
Our first set of results characterizes those initial conditions $\mathscr U_0$ such that
$\mathscr U_0 = P_{\boldsymbol j,\boldsymbol g,\omega} \mathscr U_0$. In essence, we establish that
\begin{equation*}
\mathscr U_0 = \Re \left ([\boldsymbol p,\boldsymbol q] + U + \sum_{|\ell| \geq 2} U_\ell\right ),
\end{equation*}
where $U$ is the unique time-harmonic solution, $\boldsymbol p$ and $\boldsymbol q$ are two curl-free fields with
$\boldsymbol p \times \boldsymbol n = \boldsymbol q \times \boldsymbol n =0$ on ${\Gamma_{\rm I}}$, and for all $|\ell| \geq 2$, $U_\ell$ is any
time-harmonic solution with frequency $\ell\omega$ and vanishing right-hand sides.
Thus, if time-harmonic problem \eqref{eq_maxwell_freq_strong} is well-posed for
all multiples $\ell\omega$ of $\omega$, then we simply have
$\mathscr U_0 = \Re \left ([\boldsymbol p,\boldsymbol q] + U\right )$,
which holds whenever the problem features dissipation ($\operatorname{supp} \boldsymbol \sigma \neq \emptyset$
and/or $|{\Gamma_{\rm I}}| > 0$). Moreover, we show that if both $\mathscr U_0$ and $\boldsymbol j$ are
orthogonal to curl-free fields, then $\boldsymbol p = \boldsymbol q = \boldsymbol o$, so that $\mathscr U_0 = \Re U$.
In fact, if $\Omega$ is simply connected, we have $\boldsymbol p = \boldsymbol \nabla p$
and $\boldsymbol q = \boldsymbol \nabla q$ for two scalar functions $p$ and $q$, while the condition on
$\mathscr U_0$ and $\boldsymbol j$ simply means that they are divergence-free.
Our second set of results concerns the post-processing of periodic solutions
by the filtering operator
\begin{equation}
\label{eq_filter_rhs}
F_{\boldsymbol j,\boldsymbol g,\omega} \mathscr U_0 := \frac{2}{T} \int_0^T \mathscr U(t) e^{-i \omega t} dt,
\end{equation}
where $\mathscr U$ is the solution to time-dependent Maxwell's equations \eqref{eq_maxwell_time_strong}
with initial condition $\mathscr U_0$ and right-hand sides $\boldsymbol J$ and $\boldsymbol G$. Note that $F_{\boldsymbol j,\boldsymbol g,\omega}$
may be easily computed ``on the fly'' during time-marching while computing $P_{\boldsymbol j,\boldsymbol g,\omega}$
without storing the time-history of $\mathscr U(t)$. Then, our key result states that
$U = F_{\boldsymbol j,\boldsymbol g,\omega} \mathscr U_0$ for any initial condition $\mathscr U_0$ satisfying
$\mathscr U_0 = P_{\boldsymbol j,\boldsymbol g,\omega} \mathscr U_0$, as long as time-harmonic problem \eqref{eq_maxwell_freq_strong}
is well-posed for the frequency $\omega$,
In fact, we prove the slightly stronger result that for any
initial condition $\mathscr U_0$, $F_{\boldsymbol j,\boldsymbol g,\omega} \mathscr U_0$ solves time-harmonic Maxwell's
equations with a modified right-hand side, where the misfit $(I-P_{\boldsymbol j,\boldsymbol g,\omega})\mathscr U_0$ is added
to the physical source terms. This result enables us to control the error
$U - F_{\boldsymbol j,\boldsymbol g,\omega} \mathscr U_0$ by the misfit $\mathscr U_0 - P_{\boldsymbol j,\boldsymbol g,\omega} \mathscr U_0$.
It is also central for subsequently analyzing the convexity of the cost functional.
\subsection{The CMCG algorithm}
\label{sec_CMCG_algo}
To determine an initial condition $\mathscr U_0$ that leads to a time-periodic solution,
i.e. $\mathscr U_0=P_{\boldsymbol j,\boldsymbol g,\omega}\mathscr U_0$, we minimize the ``energy functional''
\begin{equation*}
J(\mathscr U_0)
:=
\frac{1}{2} \|\mathscr U(T)-\mathscr U_0\|_{\boldsymbol \varepsilon,\boldsymbol \mu}^2
=
\frac{1}{2} \|(I-P_{\boldsymbol j,\boldsymbol g,\omega})\mathscr U_0\|^2_{\boldsymbol \varepsilon,\boldsymbol \mu}
\end{equation*}
which measures the ($\boldsymbol \varepsilon,\boldsymbol \mu$-weighted) $L^2(\Omega)$-misfit between the initial condition
and the solution after one period. Since $P_{\boldsymbol j,\boldsymbol g,\omega}$ is an affine operator,
it can be decomposed as $P_{\boldsymbol j,\boldsymbol g,\omega} \mathscr U_0 = P_\omega \mathscr U_0 + \mathscr G$, where
$\mathscr G := P_{\boldsymbol j,\boldsymbol g,\omega} 0$ and the operator
$P_\omega := P_{\boldsymbol{0},\boldsymbol{0},\omega}$, which corresponds to the propagation
of the initial condition $\mathscr U_0$ a time $T$ with zero right-hand side, is now linear.
Hence
\begin{equation*}
J(\mathscr U_0)
=
\frac{1}{2} \|(I-P_\omega) \mathscr U_0 - \mathscr G\|_{\boldsymbol \varepsilon,\boldsymbol \mu}^2,
\end{equation*}
is a standard quadratic functional.
The gradient is given by
\begin{equation*}
J'(\mathscr U_0) = (I-P_\omega^\star)(I-P_\omega) \mathscr U_0 - \mathscr G^\star,
\qquad
\mathscr G^\star := (I-P_\omega^\star) \mathscr G,
\end{equation*}
where $P_\omega^\star$ denotes the adjoint of $P_\omega$, which
actually maps the final condition $\mathscr W_T$ to $\mathscr W(0)$ by back-propagation.
In practice the action of $P_\omega$ and $P_\omega^\star$ on any $\mathscr U_0$ is
simply obtained by solving \eqref{eq_maxwell_time_strong} numerically in the
time-domain for one period. Hence, after the initialization
step described in Algorithm \ref{al_init}, we simply compute the gradient of $J$ by one
forward and one backward solve as listed in Algorithm \ref{al_grad}.
Once we have an efficient algorithm to compute $J'$, we may choose any quadratic minimization
algorithm \cite{ciarlet_1989a}. Here, we employ the conjugate gradient method,
resulting in Algorithm \ref{al_cmcg}. Note that in practice the evaluation
of the scalar product $(\mathscr U_0,\mathscr V_0)_{\boldsymbol \varepsilon,\boldsymbol \mu}$ simply amounts to computing
$\mathbb V^{\rm T} \mathbb M \mathbb U$, where $\mathbb M$ is the mass matrix
arising from space discretization, and $\mathbb U$ (resp. $\mathbb V$) is
the discrete vector of degrees of freedom representing $\mathscr U_0$ (resp. $\mathscr V_0$).
\begin{algorithm}[t]
\begin{algorithmic}[1]
\REQUIRE right-hand sides $\boldsymbol j$ and $\boldsymbol g$
\STATE compute $\mathscr G = P_{\boldsymbol j,\boldsymbol g,\omega} 0$ by time-marching for one period
\STATE compute $\mathscr G_T = P_\omega^\star \mathscr G$ by back-propagating over one peroid
\STATE set $\mathscr G^\star = \mathscr G-\mathscr G_T$
\RETURN $\mathscr G^\star$
\end{algorithmic}
\caption{Initialization}
\label{al_init}
\end{algorithm}
\begin{algorithm}[t]
\begin{algorithmic}[1]
\REQUIRE real-valued eletromagnetic field $\mathscr U_0$, precomputed $\mathscr G^\star$
\STATE compute $\mathscr U_T = P_\omega \mathscr U$ by time-marching for one period
\STATE set $\mathscr W_T = \mathscr U_T-\mathscr U_0$.
\STATE compute $\mathscr W_0 = P_\omega^\star \mathscr W_T$ by back-propagation over one period
\STATE set $J'(\mathscr U_0) = \mathscr W_T-\mathscr U_0-\mathscr G^\star$.
\RETURN $J'(\mathscr U_0)$
\end{algorithmic}
\caption{Gradient evaluation}
\label{al_grad}
\end{algorithm}
\begin{algorithm}[t]
\caption{CMCG Algorithm}
\label{al_cmcg}
\begin{algorithmic}[1]
\REQUIRE
right-hand sides $\boldsymbol j$ and $\boldsymbol g$,
initial guess $\mathscr U_0^{(0)}$,
tolerance $\delta$,
maximum iteration $\cgit_{\rm max}$
\STATE compute $\mathscr G^\star$ from $\boldsymbol j$ and $\boldsymbol g$ with Algorithm \ref{al_init}
\STATE compute $\mathscr J' = J'(\mathscr U_0^{(0)})$ with Algorithm \ref{al_grad}
\STATE set $\mathscr R^{(0)} = \mathscr J'$, $\mathscr D^{(0)} = \mathscr J'$
\FOR{$\cgit=0,\dots,\cgit_{\rm max}-1$}
\IF{$\|\mathscr R^{(\cgit)}\|_{\boldsymbol \varepsilon,\boldsymbol \mu} \leq \delta \|\mathscr R^{(0)}\|_{\boldsymbol \varepsilon,\boldsymbol \mu}$}
\RETURN $\mathscr U_0^{(\cgit)}$
\ENDIF
\STATE compute $\mathscr A = J'(\mathscr D^{(\cgit)})+\mathscr G^\star$ with Algorithm \ref{al_grad}
\STATE set $\alpha = \|\mathscr R^{(\cgit)}\|_{\boldsymbol \varepsilon,\boldsymbol \mu}^2 / (\mathscr D^{(\cgit)},\mathscr A)_{\boldsymbol \varepsilon,\boldsymbol \mu}$
\STATE set $\mathscr U_0^{(\cgit+1)} = \mathscr U_0^{(\cgit)} + \alpha \mathscr D^{(\cgit)}$
\STATE set $\mathscr R^{(\cgit+1)} = \mathscr R^{(\cgit)} - \alpha \mathscr A$
\STATE set $\beta = \|\mathscr R^{(\cgit+1)}\|_{\boldsymbol \varepsilon,\boldsymbol \mu}^2/\|\mathscr R^{(\cgit)}\|_{\boldsymbol \varepsilon,\boldsymbol \mu}^2$
\STATE set $\mathscr D^{(\cgit+1)} = \mathscr R^{(\cgit)} + \beta \mathscr D^{(\cgit)}$
\ENDFOR
\RETURN $\mathscr U_0^{(\cgit_{\rm max})}$
\end{algorithmic}
\end{algorithm}
\subsection{Convexity of the functional and convergence}
Finally, we address the convexity of the energy functional, which immediately
relates to the convergence of the CMCG algorithm. It has been previously established
that $J$ is strongly convex for the case of sound-soft scattering by a convex obstacle,
but that it is \emph{not} necessarily so for general geometries \cite{bardos_rauch_1994a}.
Here, we show that $J$ is strongly convex in an appropriate sense
as long as time-harmonic problem \eqref{eq_maxwell_freq_strong} is well-posed, thereby
ensuring the convergence of the proposed algorithm. To do so, we introduce
a second filtering operator $F_\omega \mathscr U_0 := F_{\boldsymbol 0,\boldsymbol 0,\omega} \mathscr U_0$
that is defined as \eqref{eq_filter_rhs}, but with right-hand sides $\boldsymbol j = \boldsymbol g = \boldsymbol 0$.
Our key result is that $J$ is continuous, uniformly-Lipschitz
and strictly convex on the space of initial conditions modulo the kernel of $F_\omega$.
This quotient space is only used as a technical tool in the proofs, and, in practice,
if $\mathscr U_0^{(\cgit)}$ is the initial condition at iteration $\cgit$ in the CG algorithm, then
$F_{\boldsymbol j,\boldsymbol g,\omega} \mathscr U_0^{(\cgit)} \to U$ for any initial guess $\mathscr U_0^{(0)}$.
\section{Settings and preliminary results}
\label{section_model}
This section provides the mathematical framework needed to
rigorously anayze the CMCG algorithm.
\subsection{Domain and coefficients}
We consider time-harmonic Maxwell's equations set in a Lipschitz domain
$\Omega \subset \mathbb{R}^3$. The boundary $\Gamma := \partial \Omega$ of $\Omega$
is partitioned into two relatively open disjoint subsets ${\Gamma_{\rm P}}$ and ${\Gamma_{\rm I}}$.
We assume that $\overline{{\Gamma_{\rm P}}} \cap \overline{{\Gamma_{\rm I}}} = \emptyset$,
which is not mandatory, but simplifies the analysis. Figure \ref{figure_boundary} presents
a possible configuration.
To avoid the proliferation of necessary notation to handle both two
and three-dimensional problems at the same time, we restrict our theoretical
investigations to three-dimensional domains However, our analysis also applies
to two-dimensional problems in any polarization with natural modifications.
For the sake of simplicity, we also avoid dealing with boundary sources
in our theoretical analysis, and focus on volumic sources.
Still, our numerical experiments show, that our CMCG method applies equally well
with both types of sources.
\begin{figure}
\begin{center}
\input{figures/boundary_condition_settings}
\end{center}
\caption{Example of boundary condition settings}
\label{figure_boundary}
\end{figure}
We consider three measurable symmetric tensor-valued functions
$\boldsymbol \varepsilon,\boldsymbol \mu,\boldsymbol \sigma: \Omega \to \mathbb S(\mathbb{R}^3)$ which respectively represent
the electric permittivity, the magnetic permeability, and the conductivity of
the material contained in $\Omega$. These tensors are assumed to be uniformly
bounded. We require that $\boldsymbol \varepsilon$ and $\boldsymbol \mu$ are uniformly elliptic in $\Omega$.
For the conductivity, we assume that $\boldsymbol \sigma = \boldsymbol 0$ outside some
set $\Omega_{\boldsymbol \sigma} \subset \Omega$ with Lipschitz boundary
$\Gamma_{\boldsymbol \sigma} := \partial \Omega_{\boldsymbol \sigma}$ with $\boldsymbol \sigma$ uniformly elliptic
in $\Omega_{\boldsymbol \sigma}$.
On ${\Gamma_{\rm I}}$, we consider a symmetric tensor-valued ``impedance'' function
$\boldsymbol Z: {\Gamma_{\rm I}} \to \mathbb S(\mathbb{R}^3)$ which is assumed to be measurable with
respect to the surface measure, uniformly bounded and elliptic. We
also assume that $\boldsymbol Z$ is tangential, i.e., for all $\boldsymbol \xi \in \mathbb{R}^3$
and a.e. $\boldsymbol x \in {\Gamma_{\rm I}}$, $\boldsymbol \xi \cdot \boldsymbol n(\boldsymbol x) = 0$ implies that $\boldsymbol Z(\boldsymbol x) \cdot \boldsymbol \xi = 0$.
Finally, $\boldsymbol Y := \boldsymbol Z^{-1}$ denotes the inverse of $\boldsymbol Z$.
\subsection{Functional spaces}
\label{section_model_functional}
If $\mathbb K = \mathbb{R}$ or $\mathbb{C}$, $L^2(\Omega,\mathbb K)$
denotes the space of measurable square integrable functions
mapping $\Omega$ to $\mathbb K$ \cite{adams_fournier_2003a}.
Similarly, $L^2({\Gamma_{\rm I}},\mathbb K)$ is the space of functions from
${\Gamma_{\rm I}}$ to $\mathbb K$ that are square integrable with respect to
the surface measure of ${\Gamma_{\rm I}}$. For vector-valued function, we write
$\boldsymbol L^2(\Omega,\mathbb K) := \left (L^2(\Omega,\mathbb K)\right )^3$
and
$\boldsymbol L^2({\Gamma_{\rm I}},\mathbb K) := \left (L^2({\Gamma_{\rm I}},\mathbb K)\right )^3$.
We denote by $(\cdot,\cdot)_\Omega$ and $(\cdot,\cdot)_{\Gamma_{\rm I}}$
the inner-products of these spaces. If $\boldsymbol \phi$ is a measurable essentially
bounded tensor, we employ the notations
$\|\cdot\|_{\boldsymbol \phi,\Omega}^2 = (\boldsymbol \phi\cdot,\cdot)_\Omega$
and $\|\cdot\|_{\boldsymbol \phi,{\Gamma_{\rm I}}}^2 = (\boldsymbol \phi \cdot,\cdot)_{\Gamma_{\rm I}}$.
As usual, $H^1(\Omega)$ stands for the first-order Sobolev space
\cite{adams_fournier_2003a}. If $\gamma \subset \partial \Omega$ is a relatively
open subset, $H^1_\gamma(\Omega,\mathbb K)$ is the subset of functions of
$H^1(\Omega,\mathbb K)$ with vanishing trace on $\gamma$.
For the analysis, we also need Sobolev spaces of vector-valued functions with ``well-defined''
curl, denoted by
$\boldsymbol{\CH}(\boldsymbol{\operatorname{curl}},\Omega,\mathbb K)
:=
\left \{
\boldsymbol v \in \boldsymbol L^2(\Omega,\mathbb K)
\; | \;
\grad \times \boldsymbol v \in \boldsymbol L^2(\Omega,\mathbb K)
\right \}$,
see \cite{girault_raviart_1986a}. Following \cite{fernandes_gilardi_1997a}, we can define
the tangential trace of a function $\boldsymbol v \in \boldsymbol{\CH}(\boldsymbol{\operatorname{curl}},\Omega,\mathbb K)$
on ${\Gamma_{\rm P}}$ and ${\Gamma_{\rm I}}$, and introduce
$\boldsymbol{\CX}(\Omega,\mathbb K)
:=
\left \{
\boldsymbol v \in \boldsymbol{\CH}(\boldsymbol{\operatorname{curl}},\Omega,\mathbb K)
\; | \;
\boldsymbol v_\tau|_{\Gamma_{\rm I}} \in \boldsymbol L^2({\Gamma_{\rm I}},\mathbb K)
\right \}$
and
$\boldsymbol{\CX}_{\Gamma_{\rm P}}(\Omega,\mathbb K)
:=
\left \{
\boldsymbol v \in \boldsymbol{\CX}(\Omega,\mathbb K)
\; | \;
\boldsymbol v_\tau|_{\Gamma_{\rm P}} = \boldsymbol 0
\right \}$.
To simplify the discussion below, we finally introduce the product spaces
$L(\Omega) := \boldsymbol L^2(\Omega,\mathbb{C}) \times \boldsymbol L^2(\Omega,\mathbb{C})$,
$\mathscr L(\Omega) := \boldsymbol L^2(\Omega,\mathbb{R}) \times \boldsymbol L^2(\Omega,\mathbb{R})$,
$V(\Omega) := \boldsymbol{\CX}_{\Gamma_{\rm P}}(\Omega,\mathbb{C}) \times \boldsymbol{\CX}(\Omega,\mathbb{C})$
and $\mathscr V(\Omega) := \boldsymbol{\CX}_{\Gamma_{\rm P}}(\Omega,\mathbb{R}) \times \boldsymbol{\CX}(\Omega,\mathbb{R})$.
In the remaining of this work, we follow the convention introduced
above: if $Y(\Omega)$ is a space of complex-valued
electromagnetic fields, $\mathscr Y(\Omega)$ always denotes its real-valued counterpart.
The spaces $L$ and $\mathscr L$ are equipped with the inner product
\begin{equation}
\label{eq_innerproduct}
([\boldsymbol v,\boldsymbol w],[\boldsymbol v',\boldsymbol w'])_{\boldsymbol \varepsilon,\boldsymbol \mu}
:=
(\boldsymbol \varepsilon\boldsymbol v,\boldsymbol v')_\Omega + (\boldsymbol \mu\boldsymbol w,\boldsymbol w')_\Omega
\end{equation}
for all $[\boldsymbol v,\boldsymbol w],[\boldsymbol v',\boldsymbol w'] \in L(\Omega)$
and the associated norm $\|\cdot\|_{\boldsymbol \varepsilon,\boldsymbol \mu}^2 = (\cdot,\cdot)_{\boldsymbol \varepsilon,\boldsymbol \mu}$,
while we introduce the energy norm
\begin{align}
\label{eq_energy_norm}
\enorm{[\boldsymbol v,\boldsymbol w]}^2
&:=
\omega^2\|\boldsymbol v\|_{\boldsymbol \varepsilon,\Omega}^2
+
\|\boldsymbol v_\tau\|_{\boldsymbol Y,{\Gamma_{\rm I}}}^2
+
\|\grad \times \boldsymbol v\|_{\boldsymbol \mu^{-1},\Omega}^2
+
\|\boldsymbol \sigma\boldsymbol v\|_{\boldsymbol \varepsilon^{-1},\Omega}^2
\\
\nonumber
&\;+
\omega^2 \|\boldsymbol w\|_{\boldsymbol \mu,\Omega}^2
+
\|\boldsymbol w_\tau\|_{\boldsymbol Z,{\Gamma_{\rm I}}}^2
+
\|\grad \times \boldsymbol h\|_{\boldsymbol \varepsilon^{-1},\Omega}^2
\end{align}
for all $[\boldsymbol v,\boldsymbol w] \in V(\Omega)$. We also introduce the subspace
\begin{equation*}
\LV_{\rm I}(\Omega)
:=
\left \{
[\boldsymbol e,\boldsymbol h] \in \mathscr V(\Omega)
\; | \;
\boldsymbol e \times \boldsymbol n + \boldsymbol Z \boldsymbol h_\tau = \boldsymbol 0
\;
\text{ on } {\Gamma_{\rm I}}
\right \},
\end{equation*}
of fields satisfying impedance condition \eqref{eq_maxwell_freq_strong_boundary} on ${\Gamma_{\rm I}}$.
Finally, if $\mathscr Y(\Omega)$ is any of the aforementioned real-valued
spaces, then $C^0(0,T;\mathscr Y(\Omega))$ and $C^1(0,T;\mathscr Y(\Omega))$ contain
functions from $[0,T]$ to $\mathscr Y(\Omega)$.
\subsection{Variational formulation}
We introduce the sesquilinear form
\begin{equation}
\label{eq_weak_maxwell}
a([\boldsymbol e,\boldsymbol h],[\boldsymbol v,\boldsymbol w])
:=
(\boldsymbol \sigma \boldsymbol e,\boldsymbol v)
+
(\boldsymbol Y \boldsymbol e_\tau,\boldsymbol v_\tau)_{\Gamma_{\rm I}}
+
(\boldsymbol Z \boldsymbol h_\tau,\boldsymbol w_\tau)_{\Gamma_{\rm I}}
+
(\boldsymbol h,\grad \times \boldsymbol v)
-
(\boldsymbol e,\grad \times \boldsymbol w)
\end{equation}
for all $[\boldsymbol e,\bh],[\boldsymbol v,\boldsymbol w] \in V(\Omega)$. Then,
the weak formulation of \eqref{eq_maxwell_freq_strong} is:
Find $[\boldsymbol e,\bh] \in V(\Omega)$ such that
\begin{equation*}
i\omega ([\boldsymbol e,\boldsymbol h],[\boldsymbol v,\boldsymbol w]) + a([\boldsymbol e,\boldsymbol h],[\boldsymbol v,\boldsymbol w])
=
(\boldsymbol j,\boldsymbol v) + (\boldsymbol Y \boldsymbol g \times \boldsymbol n,\boldsymbol v_\tau)_{\Gamma_{\rm I}} + (\boldsymbol Z \boldsymbol g,\boldsymbol w_\tau)_{\Gamma_{\rm I}}
\end{equation*}
for all $[\boldsymbol v,\boldsymbol w] \in V(\Omega)$. By using integration by parts,
we easily verify that
\begin{equation}
\label{eq_adjoint}
a([\boldsymbol v,\boldsymbol w],[\boldsymbol e,\boldsymbol h]) = \overline{a([\boldsymbol e,-\boldsymbol h],[\boldsymbol v,-\boldsymbol w])}
\end{equation}
for all $[\boldsymbol v,\boldsymbol w],[\boldsymbol e,\boldsymbol h] \in V_{\rm I}(\Omega)$.
\subsection{Well-posedness}
Throughout this work, we assume that the time-harmonic
problem under consideration is well-posed for the chosen angular
frequency $\omega$.
\begin{assumption}[Well-posedness]
\label{eq_assumption_well_posedness}
For all $\phi \in L(\Omega)$, there exists a unique
$S_\omega \phi \in V(\Omega)$ such that
\begin{equation}
\label{eq_maxwell_weak}
i\omega (S_\omega \phi,w)_{\boldsymbol \varepsilon,\boldsymbol \mu} + a(S_\omega \phi,w)
=
(\phi,w)_{\boldsymbol \varepsilon,\boldsymbol \mu}
\quad \forall w \in V(\Omega).
\end{equation}
In addition, the stability estimate
\begin{equation}
\label{eq_frequency_stability}
\enorm{S_\omega \phi} \leq C_{\rm stab} \|\phi\|_{\boldsymbol \varepsilon,\boldsymbol \mu}
\end{equation}
holds true.
\end{assumption}
In \eqref{eq_frequency_stability}, $C_{\rm stab}$ is a dimensionless
constant that depends on the frequency $\omega$, the shape
of the boundaries ${\Gamma_{\rm P}}$ and ${\Gamma_{\rm I}}$, and the physical coefficients $\boldsymbol \varepsilon$,
$\boldsymbol \mu$ and $\boldsymbol \sigma$. Unless the entire domain contains a conductive materials
(i.e. $\Omega_{\boldsymbol \sigma} = \Omega$), the stability constant will increase with
the frequency. In the most favorable case of a non-trapping configuration
\cite{hiptmair_moiola_perugia_2010a,moiola_spence_2019a}, we have
\begin{equation*}
C_{\rm stab} \simeq \frac{\omega d_\Omega}{c},
\end{equation*}
where $c := 1/\sqrt{\varepsilon_{\max}\mu_{\max}}$ is the (minimal) wavespeed
and $d_\Omega$ is the diameter of the computational domain. If
$\lambda := c/\omega$ denotes the wavelength, $C_{\rm stab}$ is actually proportional
to the number of wavelengths $N_\lambda := d_\Omega/\lambda$
across $\Omega$. The stability constant can however
exhibit ``arbitrarily bad'' behaviour in more complicated geometries
(close to a resonance frequency when $\Omega_{\boldsymbol \sigma} := \emptyset$ and ${\Gamma_{\rm I}} := \empty \emptyset$
for instance). We also mention that when considering two-dimensional geometries,
the two possible polarizations are equivalent to scalar Helmholtz problems, for
which a vast body of literature is now available (see, e.g., \cite{graham_pembery_spence_2019a}
and the references therein).
For future references, we note that the ``converse'' estimate to
\eqref{eq_frequency_stability}, namely
\begin{equation}
\label{eq_frequency_continuity}
\|\phi\|_{\boldsymbol \varepsilon,\boldsymbol \mu} \leq \enorm{S_\omega \phi},
\end{equation}
holds true, as can be seen from the strong form of
time-harmonic Maxwell's equations \eqref{eq_maxwell_freq_strong} and
definition \eqref{eq_energy_norm} of the energy norm.
We finally observe that in view of \eqref{eq_adjoint}, the operator $S_\omega^\star$
defined for all $\phi \in L(\Omega)$ by the variational equation
\begin{equation*}
i\omega(w,S_\omega^\star \phi)_{\boldsymbol \varepsilon,\boldsymbol \mu} + a(w,S_\omega^\star \phi)
=
(w,\phi)_{\boldsymbol \varepsilon,\boldsymbol \mu}
\quad \forall w \in L(\Omega),
\end{equation*}
has a very similar structure to $S_\Omega$. In particular, \eqref{eq_frequency_stability}
and \eqref{eq_frequency_continuity} hold true for $S_\omega^\star$ too.
\subsection{Time-harmonic solution}
Henceforth, we consider a fixed right-hand side
$\psi \in L(\Omega)$, and denote by $U \in V(\Omega)$ the
associated solution satisfying
\begin{equation}
\label{eq_frequency_solution}
i\omega(U,w) + a(U,w) = (\psi,w)_{\boldsymbol \varepsilon,\boldsymbol \mu} \quad \forall w \in V(\Omega),
\end{equation}
whose existence and uniqueness follows from Assumption \ref{eq_assumption_well_posedness}.
\subsection{Time-dependent solutions}
\label{section_model_time}
Although existence and uniqueness results for the time-dependent Maxwell's equations
\eqref{eq_maxwell_time_strong} are fairly standard, we provide some detail here,
since the final controllability method seeks an initial condition lying only in
the space $\mathscr L(\Omega)$, so that solutions to \eqref{eq_maxwell_time_strong} can only
be defined in a very weak sense.
Following Sections 4.3.1 and 5.2.4 of \cite{assous_ciarlet_labrunie_2018a},
we introduce the unbounded operator
\begin{equation*}
A:
\LV_{\rm I}(\Omega) \ni [\boldsymbol e,\boldsymbol h]
\to
[\boldsymbol \varepsilon^{-1}\boldsymbol \sigma\boldsymbol e+\boldsymbol \varepsilon^{-1} \grad \times \boldsymbol h,-\boldsymbol \mu^{-1} \grad \times \boldsymbol e] \in \mathscr L(\Omega).
\end{equation*}
Then Hille-Yosida's theorem \cite[Theorem 4.3.2]{assous_ciarlet_labrunie_2018a} shows
that for all $\mathscr U_0 \in \LV_{\rm I}(\Omega)$ and $\mathscr F \in C^1(0,T,\mathscr L(\Omega))$,
there exists a unique $\mathscr U \in C^1(0,T,\mathscr L(\Omega)) \cap C^0(0,T,\LV_{\rm I}(\Omega))$
such that
\begin{equation}
\label{eq_ode_maxwell_time}
\left \{
\begin{array}{rcll}
\dot \mathscr U(t) + A\mathscr U(t) &=& \mathscr F(t) & t \in [0,T],
\\
\mathscr U(0) &=& \mathscr U_0, &
\end{array}
\right .
\end{equation}
and the estimate
\begin{equation}
\label{eq_est_maxwell_time}
\|\mathscr U(T)\|_{\boldsymbol \varepsilon,\boldsymbol \mu}
\leq
\|\mathscr U_0\|_{\boldsymbol \varepsilon,\boldsymbol \mu}
+
\int_0^T \|\mathscr F(t)\|_{\boldsymbol \varepsilon,\boldsymbol \mu} dt
\end{equation}
holds true. Owing to the regularity of $\mathscr U$, simple manipulations then show that
we can rewrite the first line of \eqref{eq_ode_maxwell_time} as
\begin{equation}
\label{eq_var_maxwell_time}
(\dot \mathscr U(t),v)_{\boldsymbol \varepsilon,\boldsymbol \mu} + a(\mathscr U(t),v) = (\mathscr F(t),v)_{\boldsymbol \varepsilon,\boldsymbol \mu} \quad \forall t \in [0,T]
\end{equation}
for all $v \in \mathscr V(\Omega)$.
So far, we have defined solutions to \eqref{eq_maxwell_time_strong} in a
variational sense for sufficiently smooth initial data $\mathscr U_0 \in \LV_{\rm I}$,
where the link between \eqref{eq_maxwell_weak} and \eqref{eq_var_maxwell_time}
is clear. This is not entirely sufficient since as previously explained, the
functional framework for the controllability method is set in $\mathscr L(\Omega)$.
By density of $\LV_{\rm I}(\Omega)$ in $\mathscr L(\Omega)$ however, estimate \eqref{eq_est_maxwell_time}
enables us to define, for any fixed $\mathscr F$, the operator $\mathscr U_0 \to \mathscr U(T)$ for
all $\mathscr U_0 \in \mathscr L(\Omega)$ by continuity, thereby defining a continuous affine operator
mapping $\mathscr L(\Omega)$ into itself. This observation is linked to the fact that
when $\mathscr F := 0$, the operator $A$ is the infinitesimal generator of a $C_0$ semigroup
on $\mathscr L(\Omega)$, see \cite{pazy_1983a}.
Although $\mathscr U(T)$ can be defined for rough initial data
$\mathscr U_0 \in \mathscr L(\Omega)$, the corresponding solution $\mathscr U$ only solves \eqref{eq_ode_maxwell_time}
in a very weak sense as we only have $\mathscr U \in C^0(0,T;\mathscr L(\Omega))$. In particular,
\eqref{eq_var_maxwell_time} does not hold. In the proofs below, we circumvent this difficulty
by establishing our results first for initial data in $\LV_{\rm I}(\Omega)$, and then extend them
to the general case by continuity owing to the dense inclusion $\LV_{\rm I}(\Omega) \subset \mathscr L(\Omega)$.
Finally, we note that in view of \eqref{eq_adjoint}, for all
$\mathscr U_0 \in \LV_{\rm I}(\Omega)$, there exists a unique
$\mathscr U^\star \in C^1(0,T;\mathscr L(\Omega)) \cap C^0(0,T,\LV_{\rm I}(\Omega))$ such that
\begin{equation}
\label{eq_adj_maxwell_time}
(v,\dot \mathscr U^\star(t))_{\boldsymbol \varepsilon,\boldsymbol \mu} + a(v,\mathscr U^\star(t)) = 0 \quad \forall t \in [0,t]
\end{equation}
and $\mathscr U^\star(0) = \mathscr U_0$. Here, we can also extend
the notion of (weak) solutions to \eqref{eq_adj_maxwell_time} to any $\mathscr U_0 \in \mathscr L(\Omega)$,
as for \eqref{eq_var_maxwell_time}.
\section{Properties of time-periodic solutions}
\label{section_perio}
Here, we introduce the key operators at involved in the controllability method.
We also discuss in detail the link between periodic solutions to time-dependent
Maxwell's equations \eqref{eq_maxwell_time_strong} and the time-harmonic solution
to \eqref{eq_maxwell_freq_strong}.
\subsection{Key operators}
First, we introduce the filtering and propagator operators,
which are the building blocks of the energy functional and
the associated CMCG method.
\subsubsection{Filtering}
Let $T := \omega/(2\pi)$ denote the period associated with the frequency $\omega$.
The filtering operator $F_\omega$ is defined by
\begin{equation}
\label{eq_filter}
F_\omega \mathscr U := \frac{2}{T} \int_0^T \mathscr U(t) e^{-i\omega t}dt
\end{equation}
for all $\mathscr U \in C^0(0,T;\mathscr L(\Omega))$. Clearly,
$F_\omega$ continuously maps $C^0(0,T;\mathscr L(\Omega))$ into $L(\Omega)$.
and $C^0(0,T;\mathscr V(\Omega))$ into $V(\Omega)$. In addition, when $\mathscr U \in C^1(0,T;\mathscr L(\Omega))$,
integration by parts easily shows that
\begin{equation}
\label{eq_filter_derivative}
F_\omega \dot \mathscr U = i\omega F_\omega \mathscr U + \frac{\omega}{\pi} \jmp{\mathscr U}_T,
\end{equation}
where, for $\mathscr W \in C^0(0,T,\mathscr L(\Omega))$, we have introduced the notation
$\jmp{\mathscr W}_T := \mathscr W(T)-\mathscr W(0)$.
\subsubsection{Propagators}
\label{section_perio_propagators}
Following the discussion in Section \ref{section_model_time},
if $\mathscr U_0 \in \LV_{\rm I}(\Omega)$ and $\phi \in L(\Omega)$,
there exists a unique element $\mathscr U \in C^1(0,T;\mathscr L(\Omega)) \cap C^0(0,T;\LV_{\rm I}(\Omega))$
such that
\begin{equation}
\label{eq_def_propagator}
\left \{
\begin{array}{rcll}
(\dot \mathscr U(t),v)_{\boldsymbol \varepsilon,\boldsymbol \mu} + a(\mathscr U(t),v)
&=&
(\Re (\phi \; e^{i\omega t}), v)_{\boldsymbol \varepsilon,\boldsymbol \mu}
&
\forall v \in \mathscr V, \quad t \in (0,T)
\\
\mathscr U(0) &=& \mathscr U_0, &
\end{array}
\right .
\end{equation}
and we define forward propagator $P_{\phi,\omega} \mathscr U_0 := \mathscr U(T)$.
When $\phi := 0$, we simply write $P_\omega := P_{0,\omega}$.
Similarly, we define a backward propagator. For $\mathscr W_T \in \LV_{\rm I}(\Omega)$,
there exists a unique element $\mathscr W \in C^1(0,T,\mathscr L(\Omega)) \cap C^0(0,T,\LV_{\rm I}(\Omega))$
such that
\begin{equation}
\label{eq_def_backward_propagator}
\left \{
\begin{array}{rcll}
-(v,\dot \mathscr W(t))_{\boldsymbol \varepsilon,\boldsymbol \mu} + a(v,\mathscr W(t))
&=&
0
&
\forall v \in \mathscr V, \quad t \in (0,T)
\\
\mathscr W(T) &=& \mathscr W_T, &
\end{array}
\right .
\end{equation}
and we set $P_\omega^\star \mathscr W_T := \mathscr W(0)$.
Notice that $\mathscr W$ is indeed well-defined, since the change of variable $\widetilde t := T-t$
transforms \eqref{eq_def_backward_propagator} into \eqref{eq_adj_maxwell_time}.
Together with \eqref{eq_adjoint}, this remark shows that the same time-stepping algorithm may be
used to compute $P_{\phi,\omega}$ and $P_\omega^\star$ simply by changing the sign of
the magnetic field.
Again, while the above definitions of $P_{\phi,\omega}$
and $P_\omega^\star$ require $\LV_{\rm I}(\Omega)$-regularity of the initial data,
semigroup theory allows us to extend the definitions of
$P_{\phi,\omega}$ and $P_\omega^\star$ as operators continuously mapping $\mathscr L(\Omega)$
into itself \cite{pazy_1983a}.
Next, we remark that $P_\omega$ is linear, whereas $P_{\phi,\omega}$ is affine,
since
\begin{equation}
\label{eq_decomposition_propagator}
P_{\phi,\omega} \mathscr U_0 = P_\omega \mathscr U_0 + P_{\phi,\omega} 0
\qquad
\forall \mathscr U_0 \in \mathscr L(\Omega).
\end{equation}
\begin{lemma}
The operator $P_\omega^\star$ is the adjoint of $P_\omega$ for the
$\mathscr L(\Omega)$ inner-product, i.e.
\begin{equation}
\label{eq_dual_propagators}
(P_\omega \mathscr U_0,\mathscr W_T)_{\boldsymbol \varepsilon,\boldsymbol \mu}
=
(\mathscr U_0,P_\omega^\star\mathscr W_T)_{\boldsymbol \varepsilon,\boldsymbol \mu}
\end{equation}
for all $\mathscr U_0,\mathscr W_T \in \mathscr L(\Omega)$.
\end{lemma}
\begin{proof}
We only need to show \eqref{eq_dual_propagators} in $\LV_{\rm I}(\Omega)$;
the general case follows by density. Hence, we consider $\mathscr U_0,\mathscr W_T \in \LV_{\rm I}(\Omega)$
and denote by $\mathscr U,\mathscr W \in C^1(0,T,\mathscr L(\Omega)) \cap C^0(0,T,\LV_{\rm I}(\Omega))$ the associated
solutions to \eqref{eq_def_propagator} and \eqref{eq_def_backward_propagator}. Owing to the
time-regularity of $\mathscr U$ and $\mathscr W$, integration by parts shows that
\begin{equation*}
\int_0^T (\dot \mathscr U(t),\mathscr W(t))_{\boldsymbol \varepsilon,\boldsymbol \mu} dt
=
\left [
(\mathscr U(t),\mathscr W(t))_{\boldsymbol \varepsilon,\boldsymbol \mu}
\right ]_0^T
-
\int_0^T (\mathscr U(t),\dot \mathscr W(t))_{\boldsymbol \varepsilon,\boldsymbol \mu} dt,
\end{equation*}
which we rewrite as
\begin{equation}
\label{tmp_dual_propagator_0}
\int_0^T (\dot \mathscr U(t),\mathscr W(t))_{\boldsymbol \varepsilon,\boldsymbol \mu} dt
+
\int_0^T (\mathscr U(t),\dot \mathscr W(t))_{\boldsymbol \varepsilon,\boldsymbol \mu} dt
=
(P_\omega \mathscr U_0,\mathscr W_T)_{\boldsymbol \varepsilon,\boldsymbol \mu}
-
(\mathscr U_0,P_\omega^\star \mathscr W_T)_{\boldsymbol \varepsilon,\boldsymbol \mu}.
\end{equation}
The left-hand side of \eqref{tmp_dual_propagator_0} vanishes, since
\begin{multline*}
\int_0^T (\dot \mathscr U(t),\mathscr W(t))_{\boldsymbol \varepsilon,\boldsymbol \mu} dt
+
\int_0^T (\mathscr U(t),\dot \mathscr W(t))_{\boldsymbol \varepsilon,\boldsymbol \mu} dt
\\
=
\int_0^T (\dot \mathscr U(t),\mathscr W(t))_{\boldsymbol \varepsilon,\boldsymbol \mu} + a(\mathscr U(t),\mathscr W(t)) dt
+
\int_0^T (\mathscr U(t),\dot \mathscr W(t))_{\boldsymbol \varepsilon,\boldsymbol \mu} - a(\mathscr U(t),\mathscr W(t)) dt
\end{multline*}
which is zero due to \eqref{eq_def_propagator} and \eqref{eq_def_backward_propagator}.
\end{proof}
\subsubsection{Filtering of initial conditions}
If $\mathscr U_0 \in \mathscr L(\Omega)$ and $\phi \in L(\Omega)$,
we introduce the notation $F_{\phi,\omega} \mathscr U_0 := F_\omega \mathscr U$,
where $\mathscr U \in C^0(0,T,\mathscr L(\Omega))$ solves \eqref{eq_def_propagator}
in a weak sense, see \ref{section_model_time}. For
$\phi := 0$, we simply write $F_\omega \mathscr U_0 := F_{0,\omega} \mathscr U_0$.
\subsubsection{Energy functional}
Let $J: \mathscr L(\Omega) \to \mathbb{R}$ denote the ``energy functional''
\begin{equation}
\label{eq_energy_functional}
J(\mathscr U_0) := \frac{1}{2} \|P_{\psi,\omega} \mathscr U_0 - \mathscr U_0\|_{\boldsymbol \varepsilon,\boldsymbol \mu}^2
\quad
\forall \mathscr U_0 \in \mathscr L(\Omega).
\end{equation}
Using \eqref{eq_decomposition_propagator}, we can rewrite \eqref{eq_energy_functional} as
\begin{equation}
\label{eq_energy_functional_linear}
J(\mathscr U_0) = \frac{1}{2} \|(I-P_\omega) \mathscr U_0 - \mathscr G\|_{\boldsymbol \varepsilon,\boldsymbol \mu}^2
\quad
\forall \mathscr U_0 \in \mathscr L(\Omega),
\end{equation}
where $\mathscr G := P_{\psi,\omega} 0$. Note that $J$ is continuous
over $\mathscr L(\Omega)$ thanks to the discussions in Sections \ref{section_model_time}
and \ref{section_perio_propagators}.
\subsection{Structure of the minimizers}
For $U$, the (unique) time-harmonic solution to \eqref{eq_frequency_solution},
$\mathscr U_0 := \Re U$ is a minimizer of $J$ since $J(\mathscr U_0) = 0$. However, depending
on the boundary conditions, and properties of the right-hand sides, $\mathscr U_0$
may not be the only minimizer of $J$. In this section, we
analyze the properties satisfied by the minimizers of $J$
and exhibit the structure of the minimization set. We also
identify situations in which the minimizer of $J$ is unique.
The starting point of our analysis is the following model decomposition result.
\begin{lemma}[Modal decomposition]
\label{lemma_modal_decomposition}
Let $\mathscr U_0 \in \LV_{\rm I}(\Omega)$ satisfy $J(\mathscr U_0) = 0$. Then, we have
\begin{equation}
\label{eq_fourier_U0}
\mathscr U_0 = \Re \left (U_0 + U + \sum_{\ell \geq 2} U_\ell \right ),
\end{equation}
where $U_0 \in \ker a$, $U$ is the unique solution to \eqref{eq_frequency_solution},
and for $\ell \geq 2$, $U_\ell$ is an element of $V(\Omega)$ satisfying
\begin{equation}
\label{eq_higher_harmonics}
i\ell\omega (U_\ell,v) + a(U_\ell,v) = 0 \quad \forall v \in V(\Omega).
\end{equation}
\end{lemma}
\begin{proof}
Since the proof closely follows along the lines of \cite[Theorem 6]{tang_2020a},
we omit details for the sake of brevity. Consider $\mathscr U_0 \in \LV_{\rm I}(\Omega)$ such
that $J(\mathscr U_0) = 0$, and let $\mathscr U \in C^1(0,T,\mathscr L(\Omega)) \cap C^0(0,T,\LV_{\rm I}(\Omega))$
be the solution to \eqref{eq_def_propagator} with initial condition $\mathscr U_0$ and
right-hand side $\psi$. By assumption, $J(\mathscr U_0) = 0$ since $\mathscr U$ is $T$-periodic.
Hence, we can expand $\mathscr U$ in Fourier series as
\begin{equation}
\label{tmp_fourier_U0}
\mathscr U(t) = \Re \left (\sum_{\ell \geq 0} U_\ell e^{i\ell\omega t} \right )
\quad
\forall t \in (0,T)
\end{equation}
where
\begin{equation}
\label{tmp_fourier_coef}
U_0 := \frac{1}{T} \int_0^T \mathscr U(t) dt \; \in V(\Omega),
\qquad
U_\ell := \frac{2}{T} \int_0^T \mathscr U(t) e^{-i\ell\omega t} dt, \quad \ell \geq 1,
\end{equation}
Then, we obtain \eqref{eq_fourier_U0} by setting $t=0$ in \eqref{tmp_fourier_U0}.
After multiplying \eqref{eq_def_propagator} by $e^{-i\ell\omega t}$ and integrating
over $(0,T)$, we see that $U_0 \in \ker a$, $U_1 = U$, and that $U_\ell$ satisfies
\eqref{eq_higher_harmonics} for $\ell \geq 2$.
\end{proof}
Equipped with Lemma \ref{lemma_modal_decomposition}, we need a further understanding
of the kernel
\begin{equation*}
\ker a
:=
\left \{
u \in V(\Omega) \; | \; a(u,v) = 0 \quad \forall v \in V(\Omega)
\right \}
\end{equation*}
and the space
\begin{equation*}
K(\Omega)
:=
\left \{
[\boldsymbol e,\boldsymbol h] \in V(\Omega)
\left |
\begin{array}{l}
\boldsymbol e \times \boldsymbol n = \boldsymbol h \times \boldsymbol n = \boldsymbol 0 \text{ on } {\Gamma_{\rm I}}
\\
\grad \times \boldsymbol e = \grad \times \boldsymbol h = \boldsymbol 0 \text{ in } \Omega
\end{array}
\right .
\right \}
\end{equation*}
will play an important role. To characterize its structure, we introduce the set of gradients
$G(\Omega) := \boldsymbol \nabla H^1_\Gamma(\Omega,\mathbb C) \times \boldsymbol \nabla H^1_{\Gamma_{\rm I}}(\Omega,\mathbb C)$
and its orthogonal complement (with respect to the $(\cdot,\cdot)_{\boldsymbol \varepsilon,\boldsymbol \mu}$ inner-product)
$Z(\Omega) := G^\perp(\Omega)$, which consists of divergence-free functions. Then, we have
$K(\Omega) = G(\Omega) \oplus H(\Omega)$, where $H(\Omega) := K(\Omega) \cap Z(\Omega)$
is a ``cohomology'' space associated with $\Omega$. The structure of
$H(\Omega)$ is well-characterized \cite{fernandes_gilardi_1997a}. In particular,
it is finite-dimensional, and even trivial when $\Omega$ is simply-connected.
Similar properties hold for the real-valued counterparts of these spaces.
\begin{lemma}[Characterization of $\ker a$]
\label{lemma_kernel}
We have
\begin{equation*}
\ker a
=
\left \{
[\boldsymbol e,\boldsymbol h] \in K(\Omega)
\; | \;
\boldsymbol e = \boldsymbol 0 \text{ on } \Omega_{\boldsymbol \sigma}
\right \}.
\end{equation*}
\end{lemma}
\begin{proof}
Let $W := [\boldsymbol e,\boldsymbol h] \in V(\Omega)$. For all smooth, compactly supported, vector
valued-function $\boldsymbol \phi \in \boldsymbol{\mathcal D}(\Omega)$, we have
\begin{equation*}
a([\boldsymbol e,\boldsymbol h],[\boldsymbol \phi,0]) = (\boldsymbol \sigma \boldsymbol e,\boldsymbol \phi) + (\boldsymbol h,\grad \times \boldsymbol \phi) = 0,
\qquad
a([\boldsymbol e,\boldsymbol h],[0,\boldsymbol \phi]) = -(\boldsymbol e,\grad \times \boldsymbol \phi) = 0,
\end{equation*}
which implies that $\grad \times \boldsymbol h = -\boldsymbol \sigma \boldsymbol e$ and $\grad \times \boldsymbol e = \boldsymbol 0$. As a consequence,
we have
\begin{align*}
0
&=
\Re a([\boldsymbol e,\boldsymbol h],[\boldsymbol e,\boldsymbol h])
\\
&=
(\boldsymbol \sigma \boldsymbol e,\boldsymbol e) + (\boldsymbol Y \boldsymbol e_\tau,\boldsymbol e_\tau)_{{\Gamma_{\rm I}}} + (\boldsymbol Z \boldsymbol h_\tau,\boldsymbol h\tau)_{{\Gamma_{\rm I}}}
+
(\boldsymbol h,\grad \times \boldsymbol e)-(\boldsymbol e,\grad \times \boldsymbol h)
\\
&=
2(\boldsymbol \sigma \boldsymbol e,\boldsymbol e) + (\boldsymbol Y \boldsymbol e_\tau,\boldsymbol e_\tau)_{{\Gamma_{\rm I}}} + (\boldsymbol Z \boldsymbol h_\tau,\boldsymbol h\tau)_{{\Gamma_{\rm I}}},
\end{align*}
from which we conclude that $\boldsymbol e \times \boldsymbol n = \boldsymbol h \times \boldsymbol n = \boldsymbol 0$
on ${\Gamma_{\rm I}}$ and $\boldsymbol e = \boldsymbol 0$ in $\Omega_{\boldsymbol \sigma}$. This last equality
also implies that $\grad \times \boldsymbol h = \boldsymbol 0$.
\end{proof}
The first key result of this section applies to the case where the time-harmonic
problem is well-posed for all multiplies $\ell\omega$ of the original frequency
$\omega$. It is an immediate consequence of Lemmas \ref{lemma_modal_decomposition}
and \ref{lemma_kernel} and of the decomposition of $K(\Omega)$
discussed above, so that its proof is omitted.
\begin{theorem}[Decomposition for well-posed problems]
\label{thm_decomposition_well_posed}
Assume that time-harmonic equations \eqref{eq_maxwell_weak} are well-posed for all
frequencies $\ell\omega$, $\ell \in \mathbb N^\star$. Then, we have
\begin{equation*}
\mathscr U_0 = \Re \left ([\boldsymbol \nabla p,\boldsymbol \nabla q] + \theta + U\right )
\end{equation*}
where $p \in H^1_\Gamma(\Omega,\mathbb{C})$ and $q \in H^1_{{\Gamma_{\rm I}}}(\Omega,\mathbb{C})$ and
$\theta \in H(\Omega)$.
\end{theorem}
Next, we show that if the right-hand side of the problem satisfies suitable
conditions, the ``stationary part'' $U_0$ of the minimizer must vanish.
\begin{theorem}[Decomposition of divergence-free minimizers]
\label{thm_decomposition_divergence_free}
Assume that $\psi \in K^\perp(\Omega)$ and that $\mathscr U_0 \in \mathscr V(\Omega) \cap \mathscr K^\perp(\Omega)$.
Then, we have
\begin{equation*}
\mathscr U_0 = \Re \left (U + \sum_{\ell \geq 2} U_\ell\right ).
\end{equation*}
\end{theorem}
\begin{proof}
Let $\mathscr U$ be the time domain solution with initial condition $\mathscr U_0$,
and introduce $[\boldsymbol E_0,\boldsymbol H_0] := \mathscr U_0$ and $[\boldsymbol E,\boldsymbol H] := \mathscr U$.
For any test functions $[\boldsymbol v,\boldsymbol 0],[\boldsymbol 0,\boldsymbol w] \in \mathscr K(\Omega)$, we have
\begin{equation*}
(\boldsymbol \varepsilon\dot\boldsymbol E,\boldsymbol v)_{\widetilde \Omega_{\boldsymbol \sigma}} = (\boldsymbol \mu\dot \boldsymbol H,\boldsymbol w)_\Omega = 0,
\end{equation*}
which implies that $[\boldsymbol E(t),\boldsymbol H(t)] \in \mathscr K^\perp(\Omega)$. Therefore, $U_0 \in K^\perp(\Omega)$.
It follows that $U_0 \in K(\Omega) \cap K^\perp(\Omega)$ and hence, vanishes.
\end{proof}
We finally observe that if the assumptions of Theorems
\ref{thm_decomposition_well_posed} and \ref{thm_decomposition_divergence_free}
are both satisfied, we indeed have $\mathscr U_0 = \Re U$. Since
$\mathscr K^\perp(\Omega) = \mathscr Z(\Omega) \cap \mathscr H^\perp(\Omega)$, we see
that the assumptions on $\mathscr U_0$ and $\psi$ in the statement of
\eqref{thm_decomposition_divergence_free} mean that these fields are
divergence-free and orthogonal to the (finite-dimensional) space $\mathscr H(\Omega)$.
Note that this last requirement is null for simply connected domains, since
$\mathscr H(\Omega) = \{0\}$ in this case. Similarly to \cite[Theorem 1]{grote_nataf_tang_tournier_2020a}
in the acoustic case, it is always possible to explicitly compute the time independent
components $[\boldsymbol \nabla p,\boldsymbol \nabla q]$ and $\theta$ by solving Poisson problems.
\subsection{Filtering of periodic solutions}
\label{section_filtering}
In the previous section, we exhibited the structure of the
minimizing set of $J$ using Fourier theory. As the filtering
operator essentially selects one specific Fourier mode, modal
decomposition \eqref{eq_fourier_U0} can be used to show how filtering acts on minimizers
of $J$. In fact, this technique was used in \cite{grote_nataf_tang_tournier_2020a} to show
that for any minimizer $\mathscr U_0$ of $J$, we recover the time-harmonic solution $U$ after filtering.
Here, we develop an alternate proof technique, that actually
does not rely on the development of the previous section. This idea
appears to be new, and enables to quantify how well
initial conditions $\mathscr U_0$ leading to ``approximately periodic''
time-dependent solution approximate the time-harmonic
solution $U$ after filtering. The proof improves similar concepts
used in \cite[Theorem 10]{tang_2020a} for the acoustic Helmholtz equation
formulated using a second-order in time framework.
\begin{theorem}[Alternate characterization of filtered solutions]
\label{theorem_filtering}
Let $\phi \in L(\Omega)$. Then, for all $\mathscr U_0 \in \mathscr L(\Omega)$,
we can characterize $F_\omega \mathscr U_0$ as the unique element of $V(\Omega)$ such that
\begin{equation}
\label{eq_characterization_filtering}
i\omega(F_{\phi,\omega} \mathscr U_0,v)_{\boldsymbol \varepsilon,\boldsymbol \mu} + a(F_{\phi,\omega} \mathscr U_0,v)
=
(\phi,v)_{\boldsymbol \varepsilon,\boldsymbol \mu} + \frac{\omega}{\pi} (\mathscr U_0-P_\omega\mathscr U_0,v)_{\boldsymbol \varepsilon,\boldsymbol \mu}
\end{equation}
for all $v \in V(\Omega)$. As a direct consequence, we have
\begin{equation}
\label{eq_mismatch_stability}
\enorm{U-F_{\psi,\omega} \mathscr U_0}
\leq
\frac{\omega}{\pi} C_{\rm stab} \|(I-P_{\psi,\omega})\mathscr U_0\|_{\boldsymbol \varepsilon,\boldsymbol \mu}.
\end{equation}
for all $\mathscr U_0 \in \mathscr V(\Omega)$.
\end{theorem}
\begin{proof}
We first discuss the case where $\mathscr U_0 \in \LV_{\rm I}(\Omega)$. Thus,
let $\mathscr U$ be as in \eqref{eq_def_propagator} with initial condition
$\mathscr U_0$ and right-hand side $\phi \in L(\Omega)$.
For all $v \in \mathscr V(\Omega)$, we have
\begin{equation}
\label{tmp_proof_mismatch}
\frac{2}{T}
\int_0^T \left \{(\dot \mathscr U,v)_{\boldsymbol \varepsilon,\boldsymbol \mu} + a(\mathscr U,v)
\right \} e^{-i\omega t} dt
=
\frac{2}{T} \int_0^T (\Re (\phi e^{i\omega t}), v)_{\boldsymbol \varepsilon,\boldsymbol \mu} e^{-i\omega t} dt.
\end{equation}
Since $\boldsymbol \varepsilon,\boldsymbol \sigma,\boldsymbol \mu$ and $v$ are time-independent, we can write
\begin{equation*}
\frac{2}{T}
\int_0^T \left \{(\dot \mathscr U,v)_{\boldsymbol \varepsilon,\boldsymbol \mu} + a(\mathscr U,v)
\right \} e^{-i\omega t} dt
=
(F_\omega \dot \mathscr U,v)_{\boldsymbol \varepsilon,\boldsymbol \mu} + a (F_\omega \mathscr U,v),
\end{equation*}
and \eqref{eq_filter_derivative} shows that
\begin{equation*}
\frac{2}{T}
\int_0^T \left \{(\dot \mathscr U,v)_{\boldsymbol \varepsilon,\boldsymbol \mu} + a(\mathscr U,v)
\right \} e^{-i\omega t} dt
=
i\omega (F_\omega \mathscr U,v)_{\boldsymbol \varepsilon,\boldsymbol \mu} + a(F_\omega \mathscr U,v)
+
\frac{\omega}{\pi} (\jmp{\mathscr U}_T,v)_{\boldsymbol \varepsilon,\boldsymbol \mu}.
\end{equation*}
Similarly, since $\phi$ is time-independent, we have
\begin{equation*}
\frac{2}{T} \int_0^T (\Re (\phi \; e^{i\omega t}), v)_{\boldsymbol \varepsilon,\boldsymbol \mu} e^{-i\omega t} dt
=
(\phi,v)_{\boldsymbol \varepsilon,\boldsymbol \mu},
\end{equation*}
and as a result
\begin{equation*}
i\omega (F_\omega \mathscr U,v)_{\boldsymbol \varepsilon,\boldsymbol \mu} + a(F_\omega \mathscr U,v)
=
(\phi,v)_{\boldsymbol \varepsilon,\boldsymbol \mu} - \frac{\omega}{\pi} (\jmp{\mathscr U}_T,v)_{\boldsymbol \varepsilon,\boldsymbol \mu},
\end{equation*}
so that \eqref{eq_characterization_filtering} follows whenever $\mathscr U_0 \in \LV_{\rm I}(\Omega)$,
recalling that $F_{\phi,\omega} \mathscr U_0 := F_\omega \mathscr U$
and $\jmp{\mathscr U}_T := P_{\phi,\omega} \mathscr U_0 - \mathscr U_0$.
For the general case where $\mathscr U_0 \in \mathscr L(\Omega)$, we first observe that we may
equivalently rewrite \eqref{eq_characterization_filtering} as
\begin{equation}
\label{tmp_characterization_L2}
F_{\phi,\omega} \mathscr U_0
=
S_\omega\left (\phi + \frac{\omega}{\pi}(I-P_\omega) \mathscr U_0 \right ).
\end{equation}
At that point, identity \eqref{tmp_characterization_L2} is already established in $\LV_{\rm I}(\Omega)$.
But then, since \eqref{tmp_characterization_L2} involves continuous operators
from $L(\Omega)$ into itself, the density of $\LV_{\rm I}(\Omega)$ into $L(\Omega)$ implies
the general case.
To conclude the proof, letting $\phi = \psi$ and recalling the definition
\eqref{eq_frequency_solution} of $U$, we obtain
\begin{equation*}
i\omega(U-F_{\psi,\omega} \mathscr U_0,v)_{\boldsymbol \varepsilon,\boldsymbol \mu}
+
a(U-F_{\psi,\omega} \mathscr U_0,v)
=\frac{\omega}{\pi}((P_{\psi,\omega}-I)\mathscr U_0,v)_{\boldsymbol \varepsilon,\boldsymbol \mu},
\end{equation*}
so that \eqref{eq_mismatch_stability} follows from \eqref{eq_frequency_stability}.
\end{proof}
Using \eqref{eq_maxwell_weak}, we may rewrite \eqref{eq_characterization_filtering}
in compact form as
\begin{equation}
\label{eq_characterization_filtering_compact}
F_\omega \mathscr U_0 = \frac{\omega}{\pi} S_\omega \circ (I-P_\omega) \mathscr U_0
\qquad \forall \mathscr U_0 \in \mathscr L(\Omega).
\end{equation}
Taking again advantage of the similarity between the
original and adjoint problems, we can also show that
\begin{equation}
\label{eq_characterization_filtering_compact_adjoint}
F_\omega \mathscr W_T = \frac{\omega}{\pi} S_\omega^\star \circ (I-P_\omega^\star) \mathscr W_T
\qquad
\forall \mathscr W_T \in \mathscr L(\Omega).
\end{equation}
Stability estimate \eqref{eq_mismatch_stability} is of particular interest,
since it shows that filtering ``nearly periodic'' solutions yields
good approximations of the time-harmonic solution. It also suggests that the
misfit $\mathscr U_0-P_{\psi,\omega} \mathscr U_0$ may be used as a stopping criterion for iterative
methods, but the dependency on the frequency must be taken into account.
\section{Controllability Method}
\label{section_CMCG}
In this section, we build upon the results of the previous section
to introduce our controllability method, that we couple with a conjugate
gradient minimization algorithm.
We seek an initial condition $\mathscr U_0 \in \mathscr L(\Omega)$ satisfying $P_{\psi,\omega} \mathscr U_0 = \mathscr U_0$,
or maybe more explicitly, such that
\begin{equation}
\label{eq_controlability}
(I-P_\omega)\mathscr U_0 = \mathscr G,
\end{equation}
where $P_{\psi,\omega}$, $P_{\psi}$ and $\mathscr G$ are respectively introduced at
\eqref{eq_def_propagator}, \eqref{eq_decomposition_propagator}
and \eqref{eq_energy_functional_linear}.
Clearly, $\mathscr U_0 := \Re U$ is one solution to \eqref{eq_controlability}
but it may not be unique. Nevertheless, we always have $U = F_{\psi,\omega} \mathscr U_0$.
In addition, estimate \eqref{eq_mismatch_stability}
implies that for any approximate solution $\mathscr U_0$ to \eqref{eq_controlability},
$F_\omega \mathscr U_0$ is an approximate solution to \eqref{eq_frequency_solution}.
\subsection{The conjugate gradient method}
After space discretization, \eqref{eq_controlability} corresponds to a finite-dimensional
linear system. In principle, the matrix corresponding to $P_\omega$ could therefore be
(approximately) assembled by running a time-domain solver for one period for every possible initial
conditions. However, this approach is prohibitively expensive in practice. Instead, we opt
for the matrix-free conjugate gradient iteration, which only requires
evaluating $P_\omega \mathscr U_0$ for a limited number of initial conditions.
We thus reformulate controllability equation \eqref{eq_controlability} as the optimization problem
\begin{equation}
\label{eq_controllability_formulation}
\min_{\mathscr U_0 \in \mathscr L(\Omega)} J(\mathscr U_0),
\end{equation}
where $J$ is the energy functional introduced in \eqref{eq_energy_functional}.
From \eqref{eq_energy_functional_linear}, we recall that
$J$ corresponds to a ``standard'' quadratric form and, as result,
its gradient and Hessian are easily derived. The proof of the result below is omitted,
as it follows from standard algebraic manipulations.
\begin{theorem}[Structure of the energy functional]
For all $\mathscr U_0,\mathscr V_0 \in \mathscr L(\Omega)$, we have
\begin{align*}
J(\mathscr U_0+\mathscr V_0)
=
J(\mathscr U_0)
&+
\Re ((I-P_\omega^\star)(I-P_\omega)\mathscr U_0-(I-P_\omega^\star)\mathscr G,\mathscr V_0)_{\boldsymbol \varepsilon,\boldsymbol \mu}
\\
&+
\frac{1}{2}
((I-P_\omega)\mathscr V_0,(I-P_\omega)\mathscr V_0)_{\boldsymbol \varepsilon,\boldsymbol \mu}.
\end{align*}
It follows that
\begin{equation}
\label{eq_functional_gradient}
J'(\mathscr U_0) = (I-P_\omega^\star)(I-P_\omega) \mathscr U_0 - (I-P_\omega^\star) \mathscr G
\end{equation}
and
\begin{equation}
\label{eq_functional_hessian}
\left (J''(\mathscr U_0)\right )(\mathscr V_0,\mathscr V_0)
=
\|(I-P_\omega) \mathscr V_0\|_{\boldsymbol \varepsilon,\boldsymbol \mu}^2.
\end{equation}
\end{theorem}
Next, we show that $J$ is continuous, uniformly Lipschitz, and strongly convex
over the quotient space $\mathscr L(\Omega) / \ker F_\omega$. These properties ensure
the uniqueness of the minimizer of $J$ up to an element of $\ker F_\omega$ and
also implies the convergence of gradient-based algorithms \cite{ciarlet_1989a}.
\begin{theorem}[Convexity of energy functional]
For $\mathscr U_0 \in \mathscr L(\Omega)$, we have
\begin{equation}
\label{eq_functional_continuous}
J(\mathscr U_0)
=
\frac{1}{2}
\left \|\frac{\pi}{\omega} S_\omega^{-1}F_\omega \mathscr U_0 - \mathscr G \right \|_{\boldsymbol \varepsilon,\boldsymbol \mu}^2.
\end{equation}
In addition, for all $\mathscr U_0,\mathscr V_0 \in \mathscr L(\Omega)$, the estimates
\begin{equation}
\label{eq_functional_lipschitz}
\|J'(\mathscr U_0)-J'(\mathscr V_0)\|_{\boldsymbol \varepsilon,\boldsymbol \mu}
\leq
\frac{\omega^2}{\pi^2} \enorm{F_\omega(\mathscr U_0-\mathscr V_0)}
\end{equation}
and
\begin{equation}
\label{eq_functional_convex}
\left (J''(\mathscr U_0)\right )(\mathscr V_0,\mathscr V_0)
\geq
\frac{\pi^2}{\omega^2}\frac{1}{C_{\rm stab}^2} \enorm{F_\omega \mathscr V_0}^2
\end{equation}
hold true.
\end{theorem}
\begin{proof}
Identity \eqref{eq_functional_continuous} is a direct
consequence of \eqref{eq_characterization_filtering}.
Then, estimate \eqref{eq_functional_lipschitz} follows
from \eqref{eq_functional_gradient},
characterizations \eqref{eq_characterization_filtering_compact}
and \eqref{eq_characterization_filtering_compact_adjoint} of
$(I-P_\omega)$ and $(I-P_\omega^\star)$, and the continuity estimate
\eqref{eq_frequency_continuity}. Finally, we obtain convexity estimate
\eqref{eq_functional_convex} from \eqref{eq_functional_hessian},
\eqref{eq_characterization_filtering} and \eqref{eq_frequency_stability}.
\end{proof}
This result is to be compared with \cite[Theorem 3]{bardos_rauch_1994a},
where a convexity result is established under specific assumptions on the
spectrum. The use of the filtering allows to bypass this limitation.
In practice, it is not necessary to introduce the quotient space
$\mathscr L(\Omega) / \ker F_\omega$. Indeed, a careful examination of standard convergence
proofs (see, e.g., \cite[Theorem 8.4.4]{ciarlet_1989a}) shows that properties
\eqref{eq_functional_lipschitz} and \eqref{eq_functional_convex} are sufficient to ensure
the convergence of $F_{\psi,\omega} \mathscr U_0^{(\cgit)}$ to $U$ starting from any initial guess
$\mathscr U_0^{(0)} \in \mathscr L(\Omega)$, where $\mathscr U_0^{(\cgit)}$ denotes a minimizing sequence.
In addition, a reduction factor of the form
\begin{equation*}
\enorm{U-F_{\psi,\omega} \mathscr U_0^{(\cgit+1)}}
\leq
\left (1-C_{\rm stab}^{-4}\right ) \enorm{U-F_{\psi,\omega} \mathscr U_0^{(\cgit)}}
\end{equation*}
can be obtained.
Among the possible gradient descent techniques, we select the usual CG iteration
(see \cite[Section 8.5]{ciarlet_1989a}) to solve \eqref{eq_controllability_formulation}.
\subsection{Discretization}
In our computations, we use an upwind-flux discontinuous Galerkin
method to discretize Maxwell's equations \eqref{eq_def_propagator} and
\eqref{eq_def_backward_propagator} in space, while explicit Runge-Kutta schemes are
employed for time integration. We restrict our numerical experiments to
two-dimensional examples, and the required notation is briefly presented below.
\subsubsection{Two-dimensional setting}
Here, we consider two-dimensional Maxwell's equations in a bounded domain
$\Omega\subset\mathbb{R}^2$. Specifically, we consider three-dimensional Maxwell's equations
\eqref{eq_maxwell_freq_strong} in the domain $\Omega \times I$ for some interval $I$,
under the assumption that the electromagnetic field $(\boldsymbol e,\boldsymbol h)$ does not depend on the
third space variable. There are two uncoupled polarizations, and we focus on the
``transverse magnetic'' case where $\boldsymbol h = (\boldsymbol h_1,\boldsymbol h_2,0)$ and $\boldsymbol e = (0,0,\boldsymbol e_3)$.
The other polarization can be dealt with similarly by swapping the roles of $\boldsymbol h$ and $\boldsymbol e$.
Employing the notation $\boldsymbol h$ for the 2D vector gathering the magnetic field component and $e$
for the only non-zero component of the electric field. This, time-harmonic Maxwell's equations
reduce to
\begin{equation}
\label{eq_Maxwell_TM}
\left\{
\begin{array}{rclll}
i\omega\varepsilon e + \sigma e + \operatorname{curl} \boldsymbol h &=& j & \text{ in } \Omega,
\\
i\omega \mu \boldsymbol h - \boldsymbol{\scurl} \; e &=& \boldsymbol 0 & \text{ in } \Omega,
\\
e &=& 0 & \text{ on } {\Gamma_{\rm P}},
\\
e + Z \boldsymbol h_\tau &=& g &\text{ on } {\Gamma_{\rm I}},
\end{array}
\right.
\end{equation}
where $\varepsilon,\sigma,\mu$ and $Z$ are now scalar-valued functions,
and the two-dimensional curl operators are given by
\begin{equation*}
\operatorname{curl} \boldsymbol v = \partial_1 \boldsymbol v_2 - \partial_2 \boldsymbol v_1
\qquad
\boldsymbol{\scurl} \; v = (\partial_2 v,-\partial_1 v)
\end{equation*}
for any vector-valued and scalar-valued function $\boldsymbol v$ and $v$.
The corresponding time-domain Maxwell's equations are given by
\begin{subequations}
\label{eq_time_maxwell_2D}
\begin{equation}
\label{eq_time_maxwell_2D_volume}
\left \{
\begin{array}{rcl}
\varepsilon \dot E + \sigma E + \operatorname{curl} \boldsymbol H &=& J,
\\
\mu \dot \boldsymbol H - \boldsymbol{\scurl} \; E &=& \boldsymbol 0,
\end{array}
\right.
\end{equation}
in $\Omega$ and
\begin{equation}
\label{eq_time__maxwell_2D_surface}
\left \{
\begin{array}{rcll}
E &=& 0 & \text{ on } {\Gamma_{\rm P}},
\\
E + Z \boldsymbol H \times \boldsymbol n &=& G & \text{ on } {\Gamma_{\rm I}},
\end{array}
\right.
\end{equation}
\end{subequations}
for all $t \in [0,T]$.
\subsubsection{Discontinuous Galerkin discretization}
Following \cite{fezoui_lanteri_lohrengel_piperno_2005a,hesthaven_warburton_2002a},
we discretize \eqref{eq_time_maxwell_2D} with a first-order discontinuous Galerkin
(DG) method. The computational domain $\Omega$ is thus partitioned into a mesh
$\mathcal T_h$ consisting of triangular elements $K$. For any element $K \in \mathcal T_h$,
$\rho_K$ denote the diameter of the largest circle contained in $K$.
For the sake of simplicity, we assume that $\mathcal T_h$ is conforming in the sense that
the intersection $\overline{K_+} \cap \overline{K_-}$ of two distinct elements
$K_\pm \in \mathcal T_h$ is either empty, a single vertex, or a full face of both elements.
Note that the considered DG method is very flexible, and can, in principle, accommodate
non-conforming meshes with hanging nodes and/or different types of elements.
Next, we denote by $\mathcal F_h$ the set of faces associated to $\mathcal T_h$,
and we assume that each boundary face $F \in \mathcal F_h$ with $F \subset \partial \Omega$
either entirely belongs to ${\Gamma_{\rm I}}$ or ${\Gamma_{\rm P}}$. The sets ${\CF_{{\rm I},h}},{\CF_{{\rm P},h}} \subset \mathcal F_h$
gather those faces respectively lying in ${\Gamma_{\rm I}}$ and $\mathcal P$, whereas ${\CF_{{\rm int},h}}$ gathers
the remaining ``interior'' faces. We associate with each face $F \in \mathcal F_h$ a fixed
normal unit normal vector $\boldsymbol n_F$ chosen such that $\boldsymbol n_F = \boldsymbol n$ when $F \subset \partial \Omega$.
For internal faces, the orientation is arbitrary. We also employ the
notation $\boldsymbol t_F$ for the unit tangential to $F$ obtained from $\boldsymbol n_F$
by a $+\pi/2$ rotation.
For a given integer $q \in \mathbb N$, $\mathcal P_q(\mathcal T_h)$
stands for scalar-valued functions $v: \Omega \to \mathbb R$
such that $v|_K$ is a polynomial of degree less than or equal to $q$
for all $K \in \mathcal T_h$. Note that the elements of $\mathcal P_q(\mathcal T_h)$ are,
in general, discontinuous across the faces $F \in \mathcal F_h$ of the mesh.
Similarly $\boldsymbol{\CP}_q(\mathcal T_h)$ is the space of vector-valued functions
$\boldsymbol v := (\boldsymbol v_1,\boldsymbol v_2): \Omega \to \mathbb R^2$ such that $\boldsymbol v_1,\boldsymbol v_2 \in \mathcal P_q(\mathcal T_h)$.
If $v \in \mathcal P_q(\mathcal T_h)$ and $F \in {\CF_{{\rm int},h}}$, the notations
\begin{equation*}
\avg{v}_F := v_+|_F + v_-|_F
\qquad
\jmp{v}_F := v_+|_F (\boldsymbol n_+ \cdot \boldsymbol n_F) + v_-|_F (\boldsymbol n_- \cdot \boldsymbol n_F)
\end{equation*}
stand for the usual average and jump of $v$ across $F$, where we used
$v_\pm := v|_{K_\pm}$ and $\boldsymbol n_\pm = \boldsymbol n_{K_\pm}$,
for any to elements $K_-$ and $K_+$ of $\mathcal T_h$ such that $F = \partial K_- \cap \partial K_+$.
For external faces, we simply set $\avg{v}_F := \jmp{v}_F := v|_F$.
In addition, if $\boldsymbol w \in \boldsymbol{\CP}_q(\mathcal T_h)$ the same notations have to be understood
component-wise.
Given $E_{h,0} \in \mathcal P_q(\mathcal T_h)$ and $\boldsymbol H_{h,0} \in \boldsymbol{\CP}_q(\mathcal T_h)$,
the semi-discrete DG scheme consists in finding $E_h(t) \in \mathcal P_q(\mathcal T_h)$ and
$\boldsymbol H_h(t) \in \boldsymbol{\CP}_q(\mathcal T_h)$ by solving the system of ODE
for $t \in (0,T)$,
\begin{equation}
\label{eq_DG_semidiscrete}
\left \{
\begin{array}{rcll}
(\varepsilon \dot E_h(t),v_h)_\Omega
+
(\sigma E_h(t),v_h)_\Omega
+
(\boldsymbol H_h(t),\boldsymbol{\scurl} \; v_h)_\Omega
+
(\widehat \boldsymbol H_h(t) \times \boldsymbol n_F,\jmp{v_h})_{\mathcal F_h}
&=&
(J(t),v_h)
\\
(\boldsymbol \mu \dot \boldsymbol H_h(t),\boldsymbol w_h)_\Omega
+
(E_h(t),\operatorname{curl} \boldsymbol w_h)_\Omega
+
(\widehat E_h(t),\jmp{\boldsymbol w_h} \times \boldsymbol n_F)_{\mathcal F_h}
&=&
0
\end{array}
\right .
\end{equation}
for all $v_h \in \mathcal P_q(\mathcal T_h)$ and $\boldsymbol w_h \in \mathcal P_q(\mathcal T_h)$,
with initial conditions $E_h(0) = E_{h,0}$ and $\boldsymbol H_h(t) = \boldsymbol H_{h,0}$.
In \eqref{eq_DG_semidiscrete}, $(\cdot,\cdot)_{\mathcal F_h} := \sum_{F \in \mathcal F_h} (\cdot,\cdot)_F$,
while $\widehat E_h(t)$ and $\widehat \boldsymbol H_h(t)$
are the upwind ``numerical fluxes''
\begin{equation*}
\widehat E_h|_F
:=
\frac{1}{\avg{Y_{\rm flux}}} \left (
\avg{Y_{\rm flux} E_h}_F + \frac{1}{2} \jmp{\boldsymbol H_h}_F \times \boldsymbol n_F
\right )
\quad
\widehat
\boldsymbol H_h|_F
:=
\frac{1}{\avg{Z_{\rm flux}}} \left (
\avg{Z_{\rm flux}\boldsymbol H_h}_F - \frac{1}{2} \jmp{E_h}_F \boldsymbol t_F
\right ),
\end{equation*}
where $Z_{\rm flux} := \sqrt{\mu/\varepsilon}$, $Y_{\rm flux} = 1/Z_{\rm flux}$,
whenever $F \in {\CF_{{\rm int},h}}$. For the remaining faces, we set
\begin{equation*}
\widehat E_h|_F
:=
0
\qquad
\widehat
\boldsymbol H_h|_F
:=
-Y E_h \boldsymbol t_F + \boldsymbol H_h
\end{equation*}
when $F \in {\CF_{{\rm P},h}}$ and
\begin{equation*}
\widehat E_h|_F
:=
\frac{1}{2} \left (E_h + Z\boldsymbol H_h \times \boldsymbol n + G \right )
\qquad
\widehat
\boldsymbol H_h|_F
:=
\frac{Y}{2} \left ( Z\boldsymbol H_h - E_h \boldsymbol t_F - G\boldsymbol t_F \right )
\end{equation*}
if $F \in {\CF_{{\rm I},h}}$. This choice introduces some numerical dissipation,
leading to stable discretizations when coupled with Runge-Kutta time-integration.
To simplify further discussions, we introduce the compact notation
$\mathscr U_h(t) := (E_h(t),\boldsymbol H_h(t))$, and we
denote by ${\mathbb U}_h(t)$ the coefficients of $\mathscr U_h(t)$ expanded in the
nodal basis of $\mathcal P_q(\mathcal T_h)$, to rewrite \eqref{eq_DG_semidiscrete} as
\begin{equation*}
\mathbb M \dot {\mathbb U}_h(t) + \mathbb K {\mathbb U}_h(t)
=
\Re \left (\mathbb M \mathbb J e^{i\omega t}\right ),
\end{equation*}
where $\mathbb M$ and $\mathbb K$ are the usual mass and stiffness matrices.
A key asset of DG discretizations is that $\mathbb M$ is block-diagonal, so that
the inverting $\mathbb M^{-1}$ is cheap. Hence, we may reformulate
the above ODE system as
\begin{equation}
\label{eq_ode_system}
\dot {\mathbb U}_h(t)
=
\Phi(t,{\mathbb U}_h(t)),
\qquad
\Phi(t,{\mathbb U}_h(t))
:=
\Re \left (\mathbb J e^{i\omega t}\right )
+
\mathbb{B} {\mathbb U}_h(t),
\qquad
\mathbb B := \mathbb M^{-1} \mathbb K.
\end{equation}
\subsection{Time integration scheme}
We integrate \eqref{eq_ode_system} using a standard second-order explicit Runge-Kutta (RK2)
method with $\mathcal P_1$ elements, or a fourth-order explicit Runge-Kutta (RK4) method with $\mathcal P_3$
elements. Both are stable under a ``CFL condition'' on the time-step $\delta t$:
\begin{equation}
\label{eq_cfl}
\delta t \leq c_q \min_{K \in \mathcal T_h} \left (\sqrt{\mu_K\varepsilon_K} \rho_K\right ),
\end{equation}
where the constant $c_q$ only depends on the polynomial degree $q$ and the shape-regularity
of the mesh. In our computations, we use $c_1 := 0.24$ and $c_3 := 0.12$, which we empirically
found to be near the stability limit.
We thus select a time-step $\delta t := T/M$, where $M$ is the smallest positive integer
such that \eqref{eq_cfl} holds, and iteratively compute approximation $\mathscr U_{h,m}$ to $\mathscr U_h(t_m)$
for $1 \leq m \leq M$, where $t_m := m \delta t$. Since there are no ``physical''
initial conditions, we are free to choose the initial condition as
piecewise polynomial function and therefore, there are no requirements to interpolate
or project the initial condition to define $\mathscr U_{h,0}$ and the associated dof
vector $\mathbb U_{h,0}$. We either use the RK2 or the RK4 scheme to
compute $\mathbb U_{h,m+1}$ from $\mathbb U_{h,m}$. Both time integration schemes are
standard but for the sake of completeness, there are briefly listed in Algorithms \ref{al_rk2}
and \ref{al_rk4}.
\begin{algorithm}[t]
\caption{Explicit second-order Runge-Kutta (RK2) method}
\label{al_rk2}
\begin{algorithmic}[1]
\REQUIRE{$\mathbb U_{h,m}$ an approximation of ${\mathbb U}_h(t_m)$, $m\ge0$}
\STATE ${\mathbb K}_{h,1} \ :=\ \Phi(t_m,\mathbb U_{h,m})$
\STATE ${\mathbb K}_{h,2} \ :=\ \Phi(t_m+(\delta t/2),\mathbb U_{h,m} + (\delta t/2) {\mathbb K}_{h,1})$
\RETURN $\mathbb U_{h,m+1} \ :=\ \mathbb U_{h,m} + \delta t {\mathbb K}_{h,2}$
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[t]
\caption{Explicit fourth-order Runge-Kutta (RK4) method}
\label{al_rk4}
\begin{algorithmic}[1]
\REQUIRE{$\mathbb U_{h,m}$ an approximation of ${\mathbb U}_h(t_m)$, $m\ge0$}
\STATE ${\mathbb K}_{h,1} \ := \ \Phi(t_m,\mathbb U_{h,m})$
\STATE ${\mathbb K}_{h,2} \ := \ \Phi(t_m+(\delta t/2),\mathbb U_{h,m} + (\delta t/2) {\mathbb K}_{h,1})$
\STATE ${\mathbb K}_{h,3} \ := \ \Phi(t_m+(\delta t/2),\mathbb U_{h,m} + (\delta t/2) {\mathbb K}_{h,2})$
\STATE ${\mathbb K}_{h,4} \ := \ \Phi(t_m+\delta t,\mathbb U_{h,m} + \delta t {\mathbb K}_{h,3})$
\RETURN $\mathbb U_{h,m+1} \ := \ \mathbb U_{h,m}+(\delta t/6)\big({\mathbb K}_{h,1}+2{\mathbb K}_{h,2}+2{\mathbb K}_{h,3}+{\mathbb K}_{h,4}\big)$
\end{algorithmic}
\end{algorithm}
\subsection{Implementation of the filtering}
In this section, we briefly discuss the implementation of the filtering operator $F_\omega$
defined in \eqref{eq_filter}. For the RK2 scheme, we may simply employ the trapezoidal rule
\begin{equation}
\label{eq_filtering_RK2}
F_\omega \mathbb U_h
\simeq
\frac{\delta t}{2} \sum_{m=1}^M \left (
\mathbb U_{h,m-1} e^{-i\omega t_{m-1}} + \mathbb U_{h,m} e^{-i\omega t_m}
\right ),
\end{equation}
since it is second-order accurate. The situation is slightly more delicate for the RK4
scheme, as employing \eqref{eq_filtering_RK2} would deteriorate the convergence rate
of the method. Instead, we employ a method based on Hermite interpolation. This method
is especially efficient, because the RK algorithm computes the vectors $\Phi(t,\mathbb U_{h,m})$
anyways which are natural approximations to $\dot {\mathbb U}_{h,m}$. We thus let
\begin{equation*}
\mathbb I_{h,m}(t)
:=
\mathbb U_{h,{m-1}} p_{00}(t) + \mathbb U_{h,m} p_{01}(t)
+
\Phi(t_{m-1},\mathbb U_{h,{m-1}}) p_{10}(t) + \Phi(t_m,\mathbb U_{h,m}) p_{11}(t),
\end{equation*}
where the Hermite polynomials $p_{ij}$ are the only elements of $\mathcal P_3(t_{m-1},t_m)$
satisfying $p_{ij}^{(\ell)}(t_{m+k}) = \delta_{ik} \delta_{j\ell}$ for $0 \leq k,\ell \leq 1$.
Since Hermite polynomials are explicitly available, we can evaluate
\begin{equation*}
\xi_{ij} := \int_{t_{m-1}}^{t_m} p_{ij}(t) e^{-i\omega t} dt
\end{equation*}
analytically, which yields
\begin{align}
\label{eq_filtering_RK4}
F_\omega \mathbb U_h
&\simeq
\sum_{m=1}^M \int_{t_{m-1}}^{t_m} \mathbb I_{h,m}(t) e^{-i\omega t}
\\
\nonumber
&=
\mathbb U_{h,{m-1}} \xi_{00} + \mathbb U_{h,m} \xi_{01}
+
\Phi(t_{m-1},\mathbb U_{h,{m-1}}) \xi_{10} + \Phi(t_m,\mathbb U_{h,m}) \xi_{11}.
\end{align}
We emphasize that \eqref{eq_filtering_RK2} and \eqref{eq_filtering_RK4}
only require the solutions $\mathbb U_{h,{m-1}}$ and $\mathbb U_{h,m}$.
In fact, we can easily reformulate the above formula to only require
$\mathbb U_{h,m}$ at a single time, and this readily compute $F_\omega \mathbb U_h$
on the fly.
\section{Numerical examples}
\label{section_numer}
This section gathers numerical examples where we compare our CMCG algorithm
against a limiting amplitude principle, where ``naive'' time-stepping is employed
until convergence. The latter algorithm is denoted by FW (for full wave).
We utilize the DG method described in Section \ref{section_CMCG} in both cases,
so that a fair measure of the cost is the number of periods that need to be simulated
to reach a given accuracy level. We chose to start both algorithm with $\mathscr U_0^{(0)} = 0$
in all the considered experiments. It is known that this strategy is not optimal, since
transient right-hand sides generally improves the perfomance of FW, and the convergence of CMCG
can be accelerated, if it is applied after a ``run-up'' phase of a few FW iterations
(see, e.g. \cite{bristeau_glowinski_periaux_1994a,tang_2020a}). Nevertheless,
we restrict ourselves to zero initialization for a fair comparison.
Another question we address is the comparison of the solution obtained after convergence
of the CMCG or FW method against the solution given by the same frequency-domain
DG scheme. In this case we solve the linear system
$(i\omega \mathbb M + \mathbb K) \mathbb U_h = \mathbb M \mathbb J_h$,
with the direct solver implemented in the software package {\tt MUMPS}
\cite{amestoy_ashcraft_boiteau_buttari_lexcellent_weisbecker_2015a,amestoy_duff_lexcellent_2000a}.
We use the notation FS (frequency solver) to refer to this solution.
This is a subtle point, because the CMCG and FW algorithm will converge to
a (slightly) different approximation, due to the error from time discretization.
Whenever the exact solution is available, we chose the mesh $\mathcal T_h$ and
polynomial degree $q$ so that the FS relative error, measured as
\begin{equation*}
\text{error} := \|U-U_h\|_{\boldsymbol \varepsilon,\boldsymbol \mu}/\|U\|_{\boldsymbol \varepsilon,\boldsymbol \mu},
\end{equation*}
where $U$ is the exact solution and $U_h$ the FS solution,
is of the order of a few percents, which seems realistic for typical applications.
For the CMCG and FW method, the main figure of merit is then the relative error
\begin{equation*}
\text{error} := \|U - F_{\psi,\omega} \mathscr U_{0,h}^{(\cgit)}\|_{\boldsymbol \varepsilon,\boldsymbol \mu}/\|U\|_{\boldsymbol \varepsilon,\boldsymbol \mu},
\end{equation*}
where $\mathscr U_{h,0}^{(\cgit)}$ is the current iterate in the CMCG or FW algorithm. Specifically
$\mathscr U_{h,0}^{(\cgit)}$ denotes the solution obtained after $\cgit$ iterations of the CMCG algorithm,
or the solution in the FW algorithm after simulating $\cgit$ periods. Note that CMCG requires
twice many time-periods to compute $\mathscr U_{h,0}^{(\cgit)}$ as FW, which is accounted for in
the graphs below. In the last experiment, where the analytical solution is not available,
we monitor
\begin{equation*}
\text{error} := \|U_h - F_{\psi,\omega} \mathscr U_{0,h}^{(\cgit)}\|_{\boldsymbol \varepsilon,\boldsymbol \mu}/\|U_h\|_{\boldsymbol \varepsilon,\boldsymbol \mu},
\end{equation*}
when comparing CMCG against FW.
In all examples we set $\sigma := 0$, $\mu := 1$, and $Z := 1$.
For $\theta \in [0,2\pi)$, we denote by $\boldsymbol d_\theta := (\cos \theta,\sin \theta)$
the direction associated with $\theta$ and $\xi_\theta(\boldsymbol x) := e^{i\omega \boldsymbol d \cdot \boldsymbol x}$
($\boldsymbol x \in \mathbb{R}^2$) is the plane wave travelling along the direction $\boldsymbol d$.
Sometimes, we employ structured meshes based on Cartesian grids.
In this case, an ``$N \times M$ Cartesian mesh'' is obtained by starting
from a grid of $N \times M$ rectangles and then dividing each rectangle
into four triangles by joining each of its vertices with its barycentre.
\subsection{Plane wave in free space}
In this experiment, we set $\theta = 45^o$ and consider the propagation of a plane wave,
traveling along the direction $\boldsymbol d_\theta$ in the square $\Omega := (0,1)^2$.
A Silver-M\"uller absorbing boundary condition is imposed on the whole boundary,
so that ${\Gamma_{\rm I}} := \partial \Omega$ and ${\Gamma_{\rm P}} := \emptyset$. We set $\varepsilon := 1$,
$j := 0$ and $g = \boldsymbol \nabla \xi_\theta \cdot \boldsymbol n + i\omega \xi_\theta$.
The solution then reads $(e,\bh) := (\xi_\theta, \xi_\theta \boldsymbol d^\perp)$, with
$\boldsymbol d^\perp := (-\sin \theta,\cos \theta)$.
We consider the two frequencies $\omega = 10\pi$ and $40\pi$.
We employ a $32 \times 32$ Cartesian meshes in both cases with
$\mathcal P_1$ elements for $\omega=10\pi$, and $\mathcal P_3$ elements for $\omega=40\pi$.
Figure \ref{figure_planewave_error} shows the evolution of the error.
In this particular experiment, FW outperforms CMCG. When using $\mathcal P_1$ elements, the error
achieved by both FW and CMCG is indistinguishable from the FS error. On the other hand,
the error slightly increases in both FW and CMCG when using $\mathcal P_3$ elements.
\input{figures/planewave_error}
\subsection{Half open waveguide}
We now consider a rectangular domain $\Omega := (0,4) \times (0,1)$,
where the bottom, top and left sides are perfectly conducting,
while an impedance boundary condition is imposed on right side. Hence,
we have ${\Gamma_{\rm P}} := (0,4) \times \{0,1\} \cup \{0\} \times (0,1)$ and
${\Gamma_{\rm I}} := \{4\} \times (0,1)$. Then, we solve \eqref{eq_Maxwell_TM} with
$\varepsilon := 1$, $j := 0$, $g := \xi_\theta$ and $\theta = 30^o$.
We obtain a semi-analytical solution by first performing the expansion
\begin{equation}
\label{eq_series_waveguide}
e = \sum_{n \geq 0} e_n(\boldsymbol x_1) \sin(n\pi \boldsymbol x_2),
\end{equation}
that is justified by the fact that the top and bottom boundary conditions
are ``Dirichlet-like''. Then, $e_n$ can be analytically found as the
solution of linear ordinary differential equation with constant coefficients.
In practice, we truncate the expansion \eqref{eq_series_waveguide} at $n=50$,
which is sufficient since the convergence is exponential. $\boldsymbol h$ is easily recovered
by (analytically) differentiating \eqref{eq_series_waveguide}.
First, we consider $\omega = 2\pi$ with a $64 \times 16$ Cartesian mesh and $\mathcal P_1$ elements.
Then, for $\omega = 6\pi$ we use $\mathcal P_3$ elements on a $32 \times 8$ Cartesian mesh.
Figures \ref{figure_waveguide_error} shows the convergence history of the FW and CMCG solver.
CMCG converges significantly faster than FW. In particular, for $\omega = 6\pi$, the
FW solver does not reach convergence within 1000 simulated periods. As in the previous experiment,
CMCG achieves the same accuracy than FS for $\mathcal P_1$ elements, while the error is slightly
increased for $\mathcal P_3$ elements.
\input{figures/waveguide_error}
\subsection{Cavity problem}
\label{section_cavity_experiment}
We next consider an interior problem in a closed cavity $\Omega := (0,1)^2$
surrounded by a conducting material. We thus set ${\Gamma_{\rm P}} := \partial \Omega$
and ${\Gamma_{\rm I}} := \emptyset$. We apply a source $j := 1$ and set $g := 0$.
This problem features resonances at frequencies $\omega_{{\rm r},n,m}^2 := (n^2 + m^2) \pi^2$,
for all $n,m \geq 0$, with associated eigenmodes $u_{n,m} := \sin(n\pi \boldsymbol x_1)\sin(m\pi \boldsymbol x_2)$.
Again, we obtain a semi-analytical solution by truncating the Fourierexpansion.
We examine the behaviour of FW and CMCG when the frequency
$\omega$ is relatively far or close to a resonant frequency $\omega_{\rm r}$.
Hence, for a fixed resonant frequency $\omega_{\rm r}$, we consider a frequency
of the form $\omega_\delta := \omega_{\rm r} + \sqrt{2}\pi \delta$
with $\delta = 1/8$ or $1/64$.
We first take $\omega_{\rm r} := 3\sqrt{2}\pi$ with $\mathcal P_1$ elements
and a $32 \times 32$ Cartesian mesh. Then, we use $\mathcal P_3$ elements on
an $8 \times 8$ Cartesian mesh for $\omega_{\rm r} := 5\sqrt{2}\pi$.
Figures \ref{figure_cavity_error_O00} and \ref{figure_cavity_error_O02}
depict the convergence history of FW and CMCG. The FW algorithm fails to converge
even in the favorable case where $\delta = 1/8$. The CMCG algorithm converges in all cases,
and the convergence rate is only slightly affected for the smaller value of $\delta$.
\input{figures/cavity_error}
\subsection{Dipole source in a trapping medium}
The goal of this experiment is to modelize the electromagnetic field
generate by a dipole source emitting inside a body $G \subset \Omega := (-1,1)^2$.
We set ${\Gamma_{\rm P}} := \emptyset$ and ${\Gamma_{\rm I}} := \partial \Omega$.
The permittivity is not constant, and instead, we assume that
\begin{equation*}
\varepsilon(\boldsymbol x) := \left |
\begin{array}{ll}
4 & \text{ if } \boldsymbol x \in G,
\\
1 & \text{ otherwise},
\end{array}
\right .
\end{equation*}
this choice is made so that $G$ traps rays: Snell's law ensures that rays
crossing the interface with incident angle less that $60^o$ are totally reflected inside the
$G$. We modelize the dipole with $j(\boldsymbol x) := \exp \left (-|\boldsymbol x-\boldsymbol c|^2/s^2\right )$
where $s := 0.05$ and $\boldsymbol c \in G$ is the dipole localization. We consider two configurations.
In the first case, the trapping body $G := [-0.5,0.5]^2$ is squared, $\boldsymbol c := (0.25,0)$ and
$\omega := 10\pi$. In the second case $G := \{ \boldsymbol x \in \mathbb R^2 \; | \; |\boldsymbol x| < 0.5 \}$
is a disk, $\boldsymbol c := (\sqrt{2}/4,1/2-\sqrt{2}/4)$ and $\omega := 20\pi$.
We employ unstructured meshes generated with {\tt GMSH} \cite{geuzaine_remacle_2009a}.
For the square case, we impose a maximum element size $h=0.05$ leading
to a 3636 elements mesh. For the circular trap, the condition $h=0.02$
leads to a 22294 triangles mesh. In both cases, $\mathcal P_3$ elements are used respectively
resulting in 109k and 668k degrees of freedom. Figure \ref{figure_trap_solution} represents
the solutions while Figure \ref{figure_trap_error} shows the behaviour of the error. Again,
CMCG clearly outperforms FW.
\input{figures/trap_solution}
\input{figures/trap_error}
\section{Conclusion}
\label{section_conclusion}
We propose a controllability method (CM) to solve Maxwell's equations in the frequency-domain
in their first-order formulation. By minimizing a quadratic
cost functional $J$ using a conjugate gradient iteration (CG), the CMCG method determines
a time-periodic solution in the time-domain. At each CG iteration, the gradient $J'$ is computed
simply by running a time-domain solver forward and backward over one period, without
the need for solving any additional linear system. Hence, our CMCG algorithm automatically inherits
the parallelism, scalability, and low memory footprint of the underlying DG
time-domain solver. The full CMCG Algorithm 2.3 is listed in Section \ref{sec_CMCG_algo}.
In general, there exist several time-periodic solutions to Maxwell's equations,
distinct from the desired time-harmonic solution,
so that the minimizer of $J$ may not be unique. To remove those spurious modes and thus extract
the time-harmonic solution from any minimizer, we apply a cheap filtering operator computed
``on the fly'' as a final post-processing step. In Theorem \ref{theorem_filtering},
we establish that $J$ combined with the filtering operator is
strongly convex in an appropriate sense, which ensures the convergence of the CMCG method
to the desired time-harmonic solution from any initial guess. In Section \ref{section_filtering},
we also show that nearly periodic solutions already provide good approximations to
the time-harmonic solution after filtering. Hence, by monitoring the misfit, the CG iteration
may be stopped as soon as the desired accuracy has been reached.
Comparison with a direct frequency-domain solver shows that the additional
error due to time discretization is hardly visible for the low-order
$\mathcal P_1$-RK2 discretization and very small for the higher order $\mathcal P_3$-RK4 discretization.
In these numerical experiments, we also compare the CMCG method against the limiting amplitude
principle, where one simply lets the time-domain solver run until the time-harmonic regime
is reached. For simple plane wave propagation, the
limiting amplitude principle in fact slightly outperforms CMCG. For all other examples however,
CMCG significantly outperforms the limiting amplitude approach.
For the cavity experiment in Section
\ref{section_cavity_experiment}, in particular, the convergence of
CMCG is hardly affected by the trapping geometry,
whereas the limiting amplitude principle utterly fails.
Our CMCG method is non-intrusive and easily integrated
into any existing time-domain code. It is
not limited to DG discretizations; thus, we expect similar performance
using solvers based on finite differences \cite{taflove_hagness_2005a,yee_1966a}.
Although we have only used simple first-order Silver-Müller absorbing boundary conditions
in our computations, the CMCG approach immediately extends to other more accurate absorbing
conditions or perfectly matched layers \cite{tang_2020a}.
In the presence of complex geometry and local mesh refinement, local time-stepping methods
permit to overcome the stringent local CFL stability condition without sacrificing explicitness
\cite{grote_mehlin_mitkova_2015a,grote_tang_2019a}. The CMCG approach can also compute solutions
for multiple frequencies in ``one shot'', that is at the cost of a single solve, as proposed in
\cite{tang_2020a}.
|
\section{Introduction}
\label{introduction}
\begin{figure}
\begin{center}
\includegraphics[width=1\linewidth]{graph/results-style-illustration1.pdf}\\
\end{center}
\caption{Illustration of our motivation. If we want to put a police car into these images with different visual \emph{style} , we must ensure that the car is compatible with the background images (small-sized images with red boundaries in the \emph{top row}). Simple cut-and-paste operations introduce unrealistic results (\emph{top row}). Our method aims to adaptively learn high-level visual \emph{style} from different backgrounds and produce harmonious composite images (\emph{bottom row}).}
\label{fig:toyexample}
\end{figure}
Image composition is one of the most common operations in image editing~\cite{xue2012understanding,cun2020improving} and data augmentation~\cite{dwibedi2017cut,zhang2020learning}, \etc. However, generating a realistic composite image by taking an object from one image and combining it with a new background image usually requires professional compositors to adjust the appearance of the foreground objects by photo editing software like Adobe Photoshop, and ensure the realism of the generated image. To alleviate this burden, image harmonization is introduced for adjusting the foreground and making it seamlessly integrated into the new image with less human involvement, especially for non-expert users.
However, what makes a composite image appear more realistic? In this paper, we present a new perspective for image harmonization. Let us take Fig.~\ref{fig:toyexample} for example. Fig.~\ref{fig:toyexample} shows three different real photos (small-sized images with red border) that hold different visual properties. When an unbefitting foreground object with special visual properties is pasted into a new image with incompatible visual features, we can easily distinguish it from real photos. This is an unsolved problem and has emerged for years, which we call visual \emph{style} discrepancy. Specifically, in this paper, we define the visual \emph{style} of an image as visual properties including illumination, color temperature, saturation, hue, texture \etc, which varies from image to image. To make a composite image look more realistic, we must ensure a more consistent visual \emph{style} between the foreground and the background.
Abundant image harmonization approaches have been proposed for improving the realism of composite images. Traditional methods address the harmonization problem by transferring statistics of hand-crafted features between foreground and background regions, such as color~\cite{pitie2007linear,reinhard2001color,xue2012understanding,sunkavalli2010multi}. However, these methods only work in simple cases where the foreground image is already consistent with the background image. Recently, more deep learning-based methods~\cite{cong2020dovenet,cun2020improving,tsai2017deep,zhu2015learning} have been proposed for generating harmonious images in an end-to-end manner. Zhu \etal~\cite{zhu2015learning} propose to adopt a discriminative model to predict the realism of a compsite image and assist optimization of color adjustment. Tsai \etal~\cite{tsai2017deep} propose an end-to-end learning approach for image harmonization while only constraining semantic information learning in the encoder. Cun \etal~\cite{cun2020improving} adopt a spatial-separated attention module to enforce the network to learn the foreground and background features separately, failing to ensure the \emph{style} consistency between these two parts. To sum up, none of these methods really consider the realism from the perspective of visual \emph{style} consistency. Cong \etal~\cite{cong2020dovenet} propose to use a domain verification discriminator and adversarial loss~\cite{goodfellow2014generative} to improve domain-consistency between foreground and background regions but neglect to explicitly transform the foreground features in the generator. However, performance improvement brought by such an auxiliary discriminator is limited (\ie, 0.27dB for PSNR, which is revealed in~\cite{cong2020dovenet}).
To address these issues, in this work, we reframe image harmonization as a background-to-foreground \emph{style} transfer problem, where we render the foreground image to hold similar visual \emph{style} of the background image. Taking \emph{style} guidance from background information is of great importance because the foreground image should be converted to own different appearances when pasted into different background images (as illustrated in Fig.~\ref{fig:toyexample}). To generate style-consistent and realistic-looking composite images, we expect a unified transferring operation to adaptively adjust the \emph{style} of the foreground objects to be in perfect harmony with new background images even collected in different environments. Therefore, in this work, we propose a learnable layer, named Region-aware Adaptive Instance Normalization (RAIN) layer, to learn the style from background images and apply it to the foreground objects. By taking convolutional features and the foreground mask as input, the RAIN layer aligns the channel-wise mean and variance of the foreground activation to match those learned from the background. The details of the proposed RAIN module are presented in Fig.~\ref{fig:rain}. It is worth mentioning that our RAIN layer can be easily applied to existing image harmonization networks and encourage performance improvements.
The contributions of this work are as follows.
1) To the best of our knowledge, we are the first to introduce the \emph{style} concept of background images and regard the image harmonization task as a \emph{style} transferring problem.
2) We propose a novel Region-aware Adaptive Instance Normalization (RAIN) method, which captures the \emph{style} information only from the background features and applies it to the foreground for image harmonization tasks. Our RAIN module is simple yet effective and can be used as a \emph{plug-and-play} module for existing image harmonization networks to enhance their performance.
3) Extensive experiments demonstrate that our method surpasses the state-of-the-art methods by a large margin.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.9\linewidth]{graph/model-Generator.pdf}
\end{center}
\caption{Overview of the proposed generator. We provide a detailed structure of our RainNet to ensure better understanding and reproducibility. The bottom legend: Conv.= Convolution, Trans. = Transposed. }
\label{fig:model_attentioned}
\end{figure*}
\section{Related work}
\noindent
\textbf{Image harmonization} aims to adjust a foreground image to seamlessly match a background image. Traditional methods mainly focus on matching the appearance of the foreground with background regions based on handful of hand-crafted heuristics, such as color statistics~\cite{reinhard2001color,pitie2007linear,xue2012understanding}, gradient information~\cite{jia2006drag,perez2003poisson,tao2013error}, multi-scale statistical features~\cite{sunkavalli2010multi}, semantic information~\cite{tsai2017deep,tsai2016sky}. These methods directly match appearance to harmonize a composite image while paying less attention to visual realism. Johnson \etal~\cite{johnson2010cg2real} introduce a data-driven approach to improve the realism of computer-generated images by retrieving a small number of real images from an image dataset and transfer the features of color, tone, texture, \etc. Lalonde \etal~\cite{lalonde2007using} predict the realism of images by learning global and local statistics from natural images. With the advances of deep learning, more deep learning-based methods~\cite{cun2020improving,cong2020dovenet,tsai2017deep,zhu2015learning} draw much attention due to their impressive results.
Different from these works, we start from the perspective of background-to-foreground \emph{style} transfer, and push the limit of image harmonization performance by introducing a novel RAIN module, which separates our approach from previous methods.
\medskip
\noindent
\textbf{Neural style transfer} is designed to render a photo with special visual style captured from artistic creations while retaining the content information from the original image. Earlier style transfer methods concentrate on texture synthesis or transfer~\cite{efros2001image,elad2017style,li2016precomputed,ulyanov2016texture}. Gatys \etal~\cite{gatys2016image} first introduce a method to match feature statistics in pre-trained convolutional networks and demonstrate impressive artistic style transfer. To achieve the goal of real-time style transfer, Johnson \etal~\cite{johnson2016perceptual} propose a novel feed-forward perceptual loss with a pre-trained VGG network~\cite{simonyan2014very}. Later, Huang \etal~\cite{huang2017arbitrary} propose Adaptive Instance Normalization (AdaIN) to achieve arbitrary style transfer from the perspective of feature normalization. Besides AdaIN, other normalization methods~\cite{dumoulin2016learned,ulyanov2016instance} were also proposed for fast stylization and later adopted in various vision tasks~\cite{huang2018multimodal,li2020advancing,liu2019few,zakharov2019few,xue2020realistic}.
\medskip
\noindent
\textbf{Normalization layers} include unconditional normalization (Batch Normalization (BN)~\cite{ioffe2015batch}, Instance Normalization (IN)~\cite{ulyanov2016instance}, Layer Normalization (LN)~\cite{ba2016layer}, Group Normalization (GN)~\cite{wu2018group}, \etc) and conditional normalization (Conditional Batch Normalization (CBN)~\cite{de2017modulating}, Conditional Instance Normalization (CIN)~\cite{dumoulin2016learned}, SPADE~\cite{park2019semantic}, Region Normalization (RN)~\cite{yu2020region}, and AdaIN~\cite{huang2017arbitrary}, \etc). Note that unconditional normalization aligns the mean and variance of feaures without guidance from external data. On the contrary, conditional normalization~\cite{de2017modulating,dumoulin2016learned,huang2017arbitrary,park2019semantic} requires external data to provide affine parameters, which embed new information from the external data. SPADE~\cite{park2019semantic} applies spatially-varying transformations from semantic masks for image synthesis, which cannot be used in our image harmonization task due to the irregular shapes of foreground objects. RN~\cite{yu2020region} is designed for image inpainting which aims to alleviate the mean and variance shift problem but it does not consider the semantic connection between the background and the foreground. AdaIN~\cite{huang2017arbitrary} is proposed for real-time image stylization which uses a pre-trained VGG network to extract style code. However, it is not practical for our task because the \emph{style} defined in this work is considered to be consistent with image realism instead of texture. Besides, the background image with one region removed cannot be extracted by a pre-trained network, which will introduce new problems of mean and variance shift. In this paper, we seek ways to establish a connection between the background and the foreground. Therefore, we regard image harmonization as a new \emph{style} transfer task in which we transfer \emph{style} from the background to the foreground instance.
\section{Our approach}
\label{sec:approach}
Our goal is to learn a mapping network for the foreground image and ensure that the foreground image is compatible with the background. To achieve this goal, we introduce our Region-aware Adaptive Instance Normalization (RAIN) for improving the performance of basic networks.
\subsection{Problem formulation}
\label{subsec:problem_formulation}
We consider a foreground image and a background image as $I_{f}$ and $I_{b}$ respectively. The foreground mask is denoted by $M$, which indicates the region to be harmonized in the composite image $I_{c}$. Accordingly, the background mask is $\bar{M}=1-M$. The object composition process is formulated as $I_{c} = M\circ I_{f} + (1-M) \circ I_{b}$,
where $\circ$ is the Hadamard product. In this paper, we define the harmonization model as generator $G$, and the harmonized image as $\hat{I} = G(I_{c}, M)$, where $G$ is a learnable model that we expect to optimize for making $\hat{I}$ close to the ground truth image $I$ by $\|G(I_{c}, M) - I \|_{1}$.
\subsection{Region-aware Adaptive Instance Normalization (RAIN)}
\label{subsec:RAIN}
The input of our normalization module consists of two parts, \ie, the foreground mask, and the convolutional features (see in Fig.~\ref{fig:rain}). Without loss of generality, we take the RAIN module in the $i$-th layer of $G$ for example. Let $F^{i}\in \mathbb{R}^{H^{i}\times W^{i}\times C^{i}}$ be the activations and $M^{i}\in \mathbb{R}^{H^{i}\times W^{i}}$ be the resized foreground mask in the $i$-th layer, where $H^{i}, W^{i}, C^{i}$ denote the height, width, and number of channels of feature $F^{i}$, respectively. We propose a simple yet effective normalizing method called Region-aware Adaptive Instance Normalization (RAIN).
As depicted in Fig.~\ref{fig:rain}, we first multiply the input features $F^{i}$ by the foreground mask and its corresponding background mask. Then we normalize the foreground features by IN~\cite{ulyanov2016instance}, and then affine the normalized features with learned scale and bias from the background features. The new activation value $\bar{F^{i}}$ at site ($h, w, c$) in the foreground region is computed by:
\begin{equation}
\label{equ:rain}
\bar{F^{i}}_{h, w, c}=\gamma_{c}^{i}\frac{F^{i}_{h,w,c}-\mu_{c}^{i}}{\sigma_{c}^{i}} + \beta_{c}^{i},
\end{equation}
where $\mu_{c}^{i}$ and $\sigma_{c}^{i}$ are the channel-wise mean and variance of the foreground feature in $i$-th layer:
\begin{equation}
\label{equ:mean_foreground}
\mu_{c}^{i} = \frac{1}{\#\{M^{i}=1\}}\sum_{h,w}F^{i}_{h,w,c}\circ M^{i}_{h,w},
\end{equation}
\begin{equation}
\label{equ:var_foreground}
\sigma_{c}^{i} = \sqrt{\frac{1}{\#\{M^{i}=1\}}\sum_{h,w}(F^{i}_{h,w,c}\circ M^{i}_{h,w}-\mu_{c}^{i})^2 + \epsilon} .
\end{equation}
The expression $\#\{x=k\}$ means the number of pixels which equal to value $k$ in $x$. The $\gamma_{c}^{i}$ and $\beta_{c}^{i}$ are the mean and standard deviation of the activations of the background in channel $c$ of layer $i$:
\begin{equation}
\label{equ:mean_background}
\gamma_{c}^{i} = \frac{1}{\#\{\bar{M}^{i}=1\}}\sum_{h,w}F^{i}_{h,w,c}\circ \bar{M}^{i}_{h,w}
\end{equation}
\begin{equation}
\label{equ:var_background}
\beta_{c}^{i} = \sqrt{\frac{1}{\#\{\bar{M}^{i}=1\}}\sum_{h,w}(F^{i}_{h,w,c}\circ \bar{M}^{i}_{h,w}-\gamma_{c}^{i})^2 + \epsilon}
\end{equation}
where $\bar{M}^{i}$ is the background mask in $i$-th layer.
\begin{figure}
\begin{center}
\includegraphics[width=1\linewidth]{graph/model-RAIN.pdf}
\end{center}
\caption{Our RAIN module takes the input feature $F^i$ and resized mask $M^i$ as input. Then we obtain the statistical style parameters $\gamma^i$ and $\beta^i$ from only background features. The produced $\gamma^i$ and $\beta^i$ are multiplied and added to the normalized foreground features in a channel-wise manner.
}
\label{fig:rain}
\end{figure}
Our method is different from AdaIN in two aspects. First, our method focuses on transferring the visual \emph{style} from background to foreground only within the same image while AdaIN considers the style of features from another whole external image. Second, AdaIN uses a pre-trained VGG network to extract and calculate the statistics of the features, which cannot be directly employed in our task. Contrarily, our RAIN is designed and trained for image harmonization, such that the style parameters are better fitted for the foreground adjustment operations. Moreover, comprehensive experimental results demonstrate the efficacy of the proposed method.
\noindent
\textbf{RainNet.} We take a simple U-Net~\cite{ronneberger2015u,isola2017image} alike network without any feature normalization layers as our basic network architecture. Following~\cite{cong2020dovenet,cun2020improving}, in this work, we add three attention blocks in the decoder part for our \textbf{Baseline} network. Theoretically, our RAIN module can be applied in any layers of the basic network. In this work, we train our baseline with different normalization methods and exploit the design strategy of implementing our RAIN module to obtain the best model, denoted as RainNet. The structure of our RainNet is depicted in Fig.~\ref{fig:model_attentioned}.
\noindent
\textbf{Why is RAIN effective?} Briefly, RAIN helps the model to capture the visual \emph{style} information from the background image and inject it into the foreground, so that the generated foreground objects are more compatible with the new background.
Consider a simple case with Region Normalization (RN)~\cite{yu2020region} that performs feature normalization for the foreground features and the background features separately. In each normalization layer, the background features will not provide any guidance for the model to transform the foreground features. Consequently, the model can only transform the foreground image to hold the average back-ground visual statistics in the training data, leading to unsatisfactory harmonizing results. However, when performing normalization with BN or IN, the foreground features will be normalized with the same mean and variance as the background features, where the mean and variance are statistically measured from the whole global feature map. Unfortunately, the styles of background features will be shifted by those statistics from the foreground and limit the style consistency learning in subsequent layers.
In contrast with other normalization methods, our RAIN module only transfers the statistics from the background features to the normalized foreground features, without the influences from inconsistent foreground objects. As plotted in Fig.~\ref{fig:with_normalization}, IN and BN outperform RN, while our RAIN outperforms IN and BN by a large margin, demonstrating the reasonableness of our aforementioned analysis.
\begin{table*}[!htp]
\small
\begin{center}
\begin{tabular}{rcccccc}
\toprule
{Method} & {Venue} & {HCOCO} & {HAdobe5k} & {HFlickr} & {Hday2night} & {Average} \\
\midrule
Input composite & - & 33.94 & 28.16 & 28.32 & 34.01 & 31.63 \\
Lalonde and Efros~\cite{lalonde2007using} & ICCV'07 & 31.14 & 29.66 & 26.43 & 29.80 & 30.16 \\
Xue \etal~\cite{xue2012understanding} & TOG'12 & 33.32 & 28.79 & 28.32 & 31.24 & 31.40 \\
Zhu \etal~\cite{zhu2015learning} & ICCV'15 & 33.04 & 27.26 & 27.52 & 32.32 & 30.72 \\
DIH~\cite{tsai2017deep} & CVPR'17 & 34.69 & 32.28 & 29.55 & 34.62 & 33.41 \\
S$^2$AM~\cite{cun2020improving} & TIP'20 & 35.47 & 33.77 & 30.03 & 34.50 & 34.35 \\
DoveNet~\cite{cong2020dovenet} & CVPR'20 & \textcolor{blue}{\underline{35.83}} & \textcolor{blue}{\underline{34.34}} & \textcolor{blue}{\underline{30.21}} & \textcolor{red}{\textbf{35.18}} & \textcolor{blue}{\underline{34.75}} \\
\midrule
Baseline & This work & 35.03 & 33.35 & 29.50 & \textcolor{blue}{\underline{35.02}} & 33.92 \\
RainNet & Ours & \textcolor{red}{\textbf{37.08}} & \textcolor{red}{\textbf{36.22}} & \textcolor{red}{\textbf{31.64}} & 34.83 & \textcolor{red}{\textbf{36.12}} \\
\bottomrule
\end{tabular}
\end{center}
\caption{Quantitative performance comparisons of PSNR metric on the four sub-datasets of iHarmoni4~\cite{cong2020dovenet}. The numbers in \textcolor{red}{\textbf{red}} and \textcolor{blue}{\underline{blue}} represent the best and second best performance. As can be found from the results, our approach performs favorably against other methods. }
\label{tab:performance_on_4datasets}
\end{table*}
\begin{table*}[!htp]
\small
\begin{center}
\begin{tabular}{rccccccccc}
\toprule
\multirow{2}{*}{Method} & \multirow{2}{*}{Venue} & \multicolumn{2}{c}{0$\%\sim$5$\%$} & \multicolumn{2}{c}{5$\%\sim$15$\%$} & \multicolumn{2}{c}{15$\%\sim$100$\%$} & \multicolumn{2}{c}{Average} \\
\cmidrule(r){3-4}
\cmidrule(r){5-6}
\cmidrule(r){7-8}
\cmidrule(r){9-10}
& & MSE & fMSE & MSE & fMSE & MSE & fMSE & MSE & fMSE \\
\midrule
Lalonde and Efros~\cite{lalonde2007using} & ICCV'07 & 41.52 & 1481.59 & 120.62 & 1309.79 & 444.65 & 1467.98 & 150.53 & 1433.21 \\
Xue \etal~\cite{xue2012understanding} & TOG'12 & 31.24 & 1325.96 & 132.12 & 1459.28 & 479.53 & 1555.69 & 155.87 & 1141.40 \\
Zhu \etal~\cite{zhu2015learning} & ICCV'15 & 33.30 & 1297.65 & 145.14 & 1577.70 & 682.69 & 2251.76 & 204.77 & 1580.17\\
DIH~\cite{tsai2017deep} & CVPR'17 & 18.92 & 799.17 & 64.23 & 725.86 & 228.86 & 768.89 & 76.77 & 773.18 \\
S$^2$AM~\cite{cun2020improving} & TIP'20 & 15.09 & 623.11 & 48.33 & 540.54 & 177.62 & 592.83 & 59.67 & 594.67 \\
DoveNet~\cite{cong2020dovenet} & CVPR'20 & \textcolor{blue}{\underline{14.03}} & \textcolor{blue}{\underline{591.88}} & \textcolor{blue}{\underline{44.90}} & \textcolor{blue}{\underline{504.42}} & \textcolor{blue}{\underline{152.07}} & \textcolor{blue}{\underline{505.82}} & \textcolor{blue}{\underline{52.36}} & \textcolor{blue}{\underline{549.96}} \\
\midrule
Baseline & This work & 19.21 & 841.61 & 64.54 & 749.36 & 241.15 & 803.05 & 79.97 & 808.68 \\
RainNet & Ours & \textcolor{red}{\textbf{11.66}} & \textcolor{red}{\textbf{550.38}} & \textcolor{red}{\textbf{32.05}} & \textcolor{red}{\textbf{378.69}} & \textcolor{red}{\textbf{117.41}} & \textcolor{red}{\textbf{389.80}} & \textcolor{red}{\textbf{40.29}} & \textcolor{red}{\textbf{469.60}} \\
\bottomrule
\end{tabular}
\end{center}
\caption{We measure the error of different methods in foreground ratio range based on the whole test set. fMSE indicates the mean square error of the foreground region. The numbers in \textcolor{red}{\textbf{red}} and \textcolor{blue}{\underline{blue}} indicate the best and second-best results. }
\label{tab:performance_on_foreground_ratio}
\end{table*}
\section{Implementation}
\label{sec:implementation}
\noindent
\textbf{Datasets.} To demonstrate the efficacy of our approach, we analyze the performance of our model against previous methods on the benchmark dataset iHarmony4~\cite{cong2020dovenet}. According to~\cite{cong2020dovenet}, iHarmony4 consists of 4 sub-datasets (\ie, HCOCO, HAdobe5K, HFlicker and Hday2night), and 73147 pairs of synthesized composite images and corresponding ground truth images are provided. In our experiments, we follow the train-test split as~\cite{cong2020dovenet} suggested.
\noindent
\textbf{Training.} We trained the model by Adam~\cite{kingma2014adam} optimizer with a learning rate of 0.0002, and optimized our model with the same objective that DoveNet~\cite{cong2020dovenet} uses. Our model was optimized for 100 epochs on an Nvidia GTX 2080Ti GPU, with input images resized to 256$\times$256 and batch size set to 12. Detailed training objectives of our model are presented in the supplementary materials.
\begin{figure*}[!htbp]
\begin{center}
\includegraphics[width=1\linewidth]{graph/Comparison1.pdf}
\end{center}
\caption{\textbf{Qualitative comparison}. We present example results of our RainNet against three state-of-the-art methods. The samples are taken from the testing dataset of iHarmony4~\cite{cong2020dovenet}. }
\label{fig:comparison1}
\end{figure*}
\section{Experimental Results}
\label{sec:results}
In this section, we conduct extensive experiments to demonstrate the efficacy of our method. We first compare our best model (RainNet) to current state-of-the-art methods both qualitatively and quantitatively in Sec.~\ref{subsec:comparison}. Then, we investigate the design choice of RAIN for our generator in Sec.~\ref{subsec:ablation}. Subjective evaluations and further discussions are presented in Sec.~\ref{subsec:user_study} and Sec.~\ref{subsec:limitation_discussion}, respectively.
\begin{figure*}[!htbp]
\begin{center}
\includegraphics[width=1\linewidth]{graph/Comparison2.pdf}
\end{center}
\caption{\textbf{Example results on real composite images.}. We present real composite images, foreground mask, the results of three state-of-the-art methods, and the proposed model. The samples are taken from the testing dataset of~\cite{tsai2017deep}. Our method achieves better harmonized visual results than competing methods. }
\label{fig:comparison2}
\end{figure*}
\begin{table*}[!htp]
\footnotesize
\begin{center}
\begin{tabular}{rp{0.5cm}<{\centering}p{0.6cm}<{\centering}p{0.6cm}<{\centering}p{0.5cm}<{\centering}p{0.6cm}<{\centering}p{0.6cm}<{\centering}p{0.6cm}<{\centering}p{0.6cm}<{\centering}p{0.6cm}<{\centering}p{0.5cm}<{\centering}p{0.6cm}<{\centering}p{0.6cm}<{\centering}p{0.6cm}<{\centering}p{0.6cm}<{\centering}p{0.6cm}<{\centering}}
\toprule
\multirow{2}{*}{Method} & \multicolumn{3}{c}{0$\%\sim$5$\%$} & \multicolumn{3}{c}{5$\%\sim$15$\%$} & \multicolumn{3}{c}{15$\%\sim$30$\%$} & \multicolumn{3}{c}{30$\%\sim$100$\%$}& \multicolumn{3}{c}{Average} \\
\cmidrule(r){2-4}
\cmidrule(r){5-7}
\cmidrule(r){8-10}
\cmidrule(r){11-13}
\cmidrule(r){14-16}
& fL1 & PSNR & SSIM & fL1 & PSNR & SSIM & fL1 & PSNR & SSIM & fL1 & PSNR & SSIM & fL1 & PSNR & SSIM \\
\midrule
Baseline & 21.76 & 37.99 & 0.9951 & 20.55 & 32.05 & 0.9838 & 20.97 & 27.85 & 0.9631 & 21.49 & 24.39 & 0.9285 & 21.31 & 33.92 & 0.9824 \\
+ IN~\cite{ulyanov2016instance} & 18.61 & 39.08 & 0.9959 & 16.53 & 33.75 & 0.9870 & 16.34 & 29.77 & 0.9711 & 17.97 & 25.97 & 0.9384 & 17.69 & 35.32 & 0.9855 \\
+ BN~\cite{ioffe2015batch} & 17.81 & 39.48 & 0.9962 & 16.79 & 33.60 & 0.9876 & 17.76 & 29.15 & 0.9704 & 19.32. & 25.10 & 0.9395 & 17.65 & 35.34 & 0.9859 \\
+ RN~\cite{yu2020region} & 18.85 & 38.74 & 0.9959 & 17.54 & 32.85 & 0.9864 & 18.77 & 28.42 & 0.9673 & 20.55 & 24.37 & 0.9326 & 18.62 & 34.57 & 0.9842 \\
\midrule
+ RAIN-1 & \textcolor{red}{\textbf{17.10}} & \textcolor{red}{\textbf{39.67}} & \textcolor{red}{\textbf{0.9963}} & \textcolor{blue}{\underline{14.70}} & \textcolor{blue}{\underline{34.69}} & \textcolor{blue}{\underline{0.9882}} & 14.20 & 31.02 & {0.9742} & 14.92 & 27.36 & 0.9478 & \textcolor{red}{\textbf{15.88}} & \textcolor{blue}{\underline{36.06}} & \textcolor{blue}{\underline{0.9873}} \\
+ RAIN-2 & {17.71} & {39.39} & {0.9961} & {14.88} & 34.52 & \textcolor{blue}{\underline{0.9882}} & 13.89 & \textcolor{blue}{\underline{31.19}} & 0.9737 & \textcolor{blue}{\underline{14.39}} & \textcolor{blue}{\underline{27.72}} & {0.9491} & {16.16} & {36.01} & {0.9871} \\
+ RAIN-3 & 17.97 & 39.28 & 0.9960 & 15.00 & {34.54} & {0.9881} & \textcolor{blue}{\underline{13.82}} & \textcolor{blue}{\underline{31.19}} & \textcolor{blue}{\underline{0.9743}} & \textcolor{red}{\textbf{14.21}} & \textcolor{red}{\textbf{27.75}} & \textcolor{blue}{\underline{0.9493}} & 16.30 & 35.95 & {0.9872}\\
+ RAIN-4 & 17.95 & 39.27 & 0.9959 & 14.95 & 34.51 & 0.9878 & \textcolor{red}{\textbf{13.75}} & \textcolor{red}{\textbf{31.23}} & 0.9735 & 14.75 & 27.51 & 0.9469 & 16.31 & 35.96 & 0.9868 \\
+ RAIN-Encoder & 19.29 & 38.81 & 0.9957 & 16.64 & 33.79 & 0.9869 & 15.96 & 30.15 & 0.9719 & 16.40 & 26.72 & 0.9449 & 17.89 & 35.31 & 0.9861 \\
+ RAIN-Decoder & \textcolor{blue}{\underline{17.41}} & \textcolor{blue}{\underline{39.50}} & \textcolor{blue}{\underline{0.9962}} & \textcolor{red}{\textbf{14.32}} & \textcolor{red}{\textbf{34.89}} & \textcolor{red}{\textbf{0.9889}} & 14.18 & 31.01 & \textcolor{red}{\textbf{0.9746}} & {14.75} & 27.60 & \textcolor{red}{\textbf{0.9507}} & \textcolor{blue}{\underline{15.92}} & \textcolor{red}{\textbf{36.12}} & \textcolor{red}{\textbf{0.9877}} \\
\bottomrule
\end{tabular}
\end{center}
\caption{Ablation studies. The numbers in \textcolor{red}{\textbf{red}} and \textcolor{blue}{\underline{blue}} represent the best and second-best performance. }
\label{tab:performance_on_ablation}
\end{table*}
\subsection{Comparison with existing methods}
\label{subsec:comparison}
\noindent
\textbf{Performance on different sub-datasets.} To quantitatively validate our approach, we adopt the evaluation protocols from previous work~\cite{cong2020dovenet,tsai2017deep,cun2020improving}. We first train our model on the whole training set. Then we evaluate the trained model on given testing images by measuring mean square error (MSE) and PSNR score for the synthesized images. The results of all previous methods as well as our RainNet are given in Table~\ref{tab:performance_on_4datasets}. It can be observed that the baseline model attains comparable performance of DIH~\cite{tsai2017deep}.
Benefiting from the proposed RAIN module, our RainNet improves the baseline by a reduction of 39.68 in MSE metric, and a performance gain of 2.2 in PSNR for all datasets. Although DoveNet~\cite{cong2020dovenet} is slightly favorable to our approach in Hday2night dataset, our model achieves the best results on HCOCO, HAdobe5k, and HFlickr and outperforms~\cite{cong2020dovenet} by a large margin in average performance.
\noindent
\textbf{Influence of foreground ratios.} We next examine the influence of different foreground ratios on the harmonization models. Following~\cite{cong2020dovenet}, we split the images into three groups according to different foreground ratio ranges, \ie, 0$\%\sim$5$\%$, 5$\%\sim$15$\%$, and 15$\%\sim$100$\%$. We compare the performance by metrics of MSE and fMSE. For fMSE, we only calculate the MSE of the foreground regions. The comparison results are presented in Table~\ref{tab:performance_on_foreground_ratio}. As can be found, on one hand, the model performance in terms of MSE downgrades as the foreground ratios increases while fMSE is less likely to be influenced by foreground ratios. On the other hand, our model outperforms~\cite{cong2020dovenet} by 80.36 in the fMSE metric and improves the performance of the baseline model by 39.68, 339.08 in MSE, fMSE, respectively.
\noindent
\textbf{Qualitative comparisons.}
We proceed to take a closer look at model performance and provide qualitative comparisons with the previous competing methods. From the sample results in Fig.~\ref{fig:comparison1}, it can be easily observed that our method better integrates the foreground objects into the background image, achieving much better visual consistency compared to other methods. For instance, in the second row of Fig.~\ref{fig:comparison1}, the background image is underexposed, while the foreground objects (balloons) are much brighter, leading to unrealistic visual results. Both DIH and DoveNet cannot adjust the foreground to be compatible with the dim backgrounds, while S$^2$AM generates the least realistic result. Our RainNet achieves more photorealistic results with context consistency by adaptively learning the \emph{style} features from the background and applying to the foreground objects. Fig.~\ref{fig:comparison2} gives another three typical samples picked from 99 real composited images evaluated in~\cite{tsai2017deep}. Although there is no ground truth image as a reference, we can still observe significant improvements of visual \emph{style} consistency achieved by our approach.
\begin{figure}
\begin{center}
\includegraphics[width=0.9\linewidth]{graph/with_normalization.pdf}
\end{center}
\caption{Comparisons of different normalization methods on PSNR metric. Without normalization (labeled by \emph{None}), the model performance heavily deteriorates. }
\label{fig:with_normalization}
\end{figure}
\subsection{Ablation study}
\label{subsec:ablation}
In this section, we conduct comprehensive ablation studies to demonstrate the effectiveness of our RAIN module. Different from Sec.~\ref{subsec:comparison}, we resort to three alternative measures (\ie, foreground L1 norm (fL1), PSNR, and SSIM~\cite{wang2004imagessim}) for quantitative evaluation.
\noindent
\textbf{Efficacy of RAIN.} We first investigate the performance gain brought by our RAIN module compared to other normalization methods, \ie, RN, IN, and BN. To begin with, we apply RN to the baseline model and observe stable model training curves and better performance than that without noralization layers (See in Table~\ref{tab:performance_on_ablation} and Fig.~\ref{fig:with_normalization}). Note that RN only performs batch normalization for the background (foreground) features within all background (foreground) regions, respectively. This operation splits the background and foreground features and prevents the network from propagating information from the background to the foreground, thus cannot generalize well in image harmonization tasks.
We proceed to add IN and BN to the baseline. As can be found in Table~\ref{tab:performance_on_ablation} and Fig.~\ref{fig:with_normalization} (the purple and green curves), the baseline+IN/BN outperforms the baseline method and baseline+RN by a large margin. Potential explanations can be analyzed from two aspects. On one hand, feature normalizing operations can help to stabilize and benefit the training process of deep neural networks, yielding better convergence. On the other hand, performing feature normalization with IN or BN enables the foreground features to be modified by the mean and variance statistically measured from both the foreground features and the background features. Therefore, the model can learn to adjust the visual properties of the foreground objects somehow.
Furthermore, we replace the normalization layer in the decoder network with RAIN while setting the normalization layer to IN in the encoder, then train the network under the same settings. The results are plotted in Fig.~\ref{fig:with_normalization} (red curve). Obviously, thanks to our novel RAIN module, the model with RAIN-Decoder outperforms other normalization methods and achieves the best performance on average.
\noindent
\textbf{Which layer to add RAIN?} In order to exploit the best implementation strategy for RAIN, we conduct experiments by gradually adding and removing the RAIN layers in the RainNet network. Here we compare several variants that are boosted by RAIN module in different convolutional stages (more variants and comparisons are presented in the supplementary materials). Note that in the middle layers of the generator, the spatial size of convolutional features decreases significantly. For instance, when we resize the foreground mask to 4$\times$4, the valid pixels of the foreground mask are rather rare. Under these circumstances, our RAIN downgrades to Instance Normalization. So we gradually remove RAIN layers from the 4 outermost layers in the encoder and decoder. \textbf{(a) Baseline+RAIN-Decoder}: we add RAIN layer to the decoder and IN to the encoder. \textbf{(b) Baseline+RAIN-Encoder}: in contrast to \textbf{(a)}, we use RAIN a layer only for the encoder and use IN for the decoder. \textbf{(c) Baseline+RAIN-k}: we add k (k=1,2,3,...) RAIN layers to the outermost four layers of the encoder and decoder, and IN to the remaining layers.
The quantitative comparison results are provided in Table~\ref{tab:performance_on_ablation} and Fig.~\ref{fig:how_normalize}. Our observations can be summarized as follows:
\begin{figure}
\begin{center}
\includegraphics[width=0.9\linewidth]{graph/how_normalize.pdf}
\end{center}
\caption{Comparisons of different implementation strategies of RAIN on PSNR metric. }
\label{fig:how_normalize}
\end{figure}
\noindent
\textbf{1}) Baseline+RAIN-Encoder achieves comparable performance of that with IN, while Baseline+RAIN-Decoder outperforms RAIN-Encoder by a large margin. The differences indicate the better choices of RAIN for the decoder and IN for the encoder.
\noindent
\textbf{2}) Starting from Baseline+RAIN-Decoder, we decrease the number of RAIN layers in the decoder, while adding as many RAIN layers to the outermost parts of the encoder, \ie, Baseline+RAIN-4. The model attains dropped performance but still better than Baseline+IN.
\noindent
\textbf{3}) Baseline+RAIN-1 slightly outperforms Baseline+RAIN-2, Baseline+RAIN-3, and Baseline+RAIN-4 by minor improvements. However, when compared to IN, BN, and RN, the improvements brought by our RAIN are significant.
From the experimental results, we conclude that adopting RAIN in the decoder and IN in the encoder or using the similar structure as Baseline+RAIN-$k$ are better choices. One probable reason is that some visual-consistency related features (\eg, color tone, illumination \etc) are likely to be related to the low-level features extracted in the shallow layers of convolutional neural networks, so the layers that are closest to the network’s input and output impose greater impacts on estimation error. Another reason is that the deployment of the RAIN in the symmetrical layers of the encoder and decoder helps the concatenated features have the same mean-variance in the background and foreground regions, which is helpful for the filters to stabilize the training and converge to better performance.
\noindent
\textbf{Adding RAIN to previous work.} To apply RAIN in existing methods, we conduct experiments with DIH~\cite{tsai2017deep}. We first implement DIH (with segmentation branch) in Pytorch~\cite{paszke2017automatic} and then train the basic network. In order to add RAIN to DIH, we replace BN with IN in the encoder, and RAIN with BN in the harmonization decoder. The performance of DIH model reaches to 33.36dB of PSNR while the new model with RAIN achieves 33.84dB (+0.48dB). Detailed illustrations can be found in the supplementary materials.
\begin{table}
\footnotesize
\begin{center}
\begin{tabular}{p{1.3cm}<{\centering}p{0.6cm}<{\centering}p{0.8cm}<{\centering}p{1cm}<{\centering}p{1.1cm}<{\centering}p{1.0cm}<{\centering}}
\toprule
Method & Input & DIH~\cite{tsai2017deep} & S$^2$AM~\cite{cun2020improving} & DoveNet~\cite{cong2020dovenet} & RainNet\\
\midrule
Total votes & 113 & 203 & 193 & 226 & \textbf{354} \\
Preference & 10.4$\%$ & 18.6$\%$ & 17.7$\%$ & 20.8$\%$ & \textbf{32.5}$\%$ \\
\bottomrule
\end{tabular}
\end{center}
\caption{Comparisons between our method and other competing methods under user study.}
\label{tab:user_study}
\end{table}
\subsection{User study}
\label{subsec:user_study}
Table~\ref{tab:user_study} shows the user evaluation results on real-world composited images collected by DIH~\cite{tsai2017deep}. Specifically, we invited 11 volunteers to rate and choose the most realistic harmonized images from 5 given images. Those 5 images include the original composite image and its corresponding 4 harmonized versions created by DIH, S$^2$AM, DoveNet, and Ours. We randomly shuffle the displaying order of 5 images to ensure that the users do not know which model each image belongs to. Each user is asked to evaluate for the whole set (99 images). As shown in the Table~\ref{tab:user_study}, RainNet attains more votes than the rest, which demonstrates the effectiveness of the proposed approach.
\subsection{Discussions and limitations}
\label{subsec:limitation_discussion}
\noindent\textbf{Discussions.} Obviously, benefiting from RAIN module, RainNet achieves a higher PSNR score and lower estimation error than previous DoveNet~\cite{cong2020dovenet} by 1.37dB and 12.07, respectively. Although we found that parts of these improvements are attributed to our generator settings, in which we only learn to modify the foreground image and copy the background pixels from the input, thus reducing the error of the background, we attain lower foregroud estimation errors (fMSE). fMSE is fair for all methods. Furthermore, comparing to IN, RainNet remarkably improves the performance of a baseline model and achieves the best scores on average, which demonstrates the superiority of the proposed RAIN module.
\noindent\textbf{Limitations.} Despite the improvements, our proposed approach still faces with two major confusions. First, it is not very clear why applying RAIN only in the encoder brings little improvement. Second, our model will soften the sharp foreground object and reduce the visual style discrepancy in the samples with dark background and sharp foreground objects. Future investigation in these issues should be required.
\section{Conclusion}
\label{sec:conclusion}
In this paper, we propose to solve the visual \emph{style} inconsistency problem in image harmonization and present a simple yet effective Region-aware Adaptive Instance Normalization (RAIN) module, which outperforms previous normalization methods by a large margin. We have also exploited the best implementation choice of RAIN for the baseline network. Moreover, we demonstrate the efficacy of RAIN by applying RAIN into existing networks, \eg, DIH, and observe performance gains over these models.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
\label{sec-intro}
Magnetic reconnection is an important plasma physical process because it can rapidly convert magnetic energy to particle kinetic energy; it does so by rearranging the magnetic field configuration, ``breaking'' and subsequently ``reconnecting'' magnetic field lines
\citep[for a review, see, e.g.,][]{Zweibel_Yamada-2009,Yamada_etal-2010}.
Magnetic reconnection is thus believed to play a fundamental role in a wide variety of rapid---and sometimes violent and spectacular---releases of magnetic energy resulting in particle energization and radiation, from solar flares to X-ray emission in coronae of accreting black holes to flaring TeV emission in blazar jets of active galactic nuclei (AGN).
In some high-energy astrophysical sources, such as pulsar wind nebulae (PWN) and blazar jets, reconnection may operate in a somewhat extreme regime of relativistically-hot plasma containing positrons instead of, or in addition to, ions.
In such environments, reconnection in the relativistic regime has become a promising explanation for high-energy emission,
especially since particle-in-cell (PIC) simulations have convincingly demonstrated that,
in addition to heating
plasma, reconnection can accelerate a significant fraction of particles to very high energies, yielding a nonthermal power-law energy distribution of particles
\citep[e.g.,][]{Zenitani_Hoshino-2001,Zenitani_Hoshino-2005a,Zenitani_Hoshino-2005b,Zenitani_Hoshino-2007,Zenitani_Hoshino-2008,Jaroschek_etal-2004a,Lyubarsky_Liverts-2008,Liu_etal-2011,Sironi_Spitkovsky-2011,Sironi_Spitkovsky-2014,Bessho_Bhattacharjee-2012,Cerutti_etal-2012b,Cerutti_etal-2013,Cerutti_etal-2014a,Cerutti_etal-2014b,Kagan_etal-2013,Guo_etal-2014,Guo_etal-2015,Guo_etal-2016,Guo_etal-2016b,Guo_etal-2019,Nalewajko_etal-2015,Sironi_etal-2015,Sironi_etal-2016,Dahlin_etal-2015,Dahlin_etal-2017,Werner_etal-2016,Werner_etal-2018,Werner_etal-2019,Werner_Uzdensky-2017,Petropoulou_Sironi-2018,Ball_etal-2018,Ball_etal-2019,Li_etal-2019,Schoeffler_etal-2019,Mehlhaff_etal-2020,Sironi_Beloborodov-2020,Kilian_etal-2020,Guo_etal-2020arxiv,Hakobyan_etal-2019,Hakobyan_etal-2021,Zhang_etal-2021arxiv}.
High-energy nonthermal particle acceleration (NTPA) followed by synchrotron or inverse Compton emission could explain observed nonthermal radiation spectra.
Large 2D PIC simulations of reconnection in collisionless, relativistic pair plasma have observed fast reconnection, with reconnection rates around $E_{\rm rec}\sim 0.1 B_0 v_A/c$ (where $B_0$ is the upstream, ambient, reconnecting magnetic field, and $v_A$ is the corresponding Alfv\'{e}n speed), resulting in rapid conversion of magnetic energy to plasma energy.
In these simulations, NTPA yields power-law electron energy distributions $f(\gamma)\sim \gamma^{-p}$ (where electron energy is $\gamma m_e c^2$) with a range of slopes $p$; it appears that, depending on the environment and system size, $p$ can take on values greater than~1, and $p$ may even approach~1 in highly magnetically-dominated reconnection \citep[e.g.,][]{Guo_etal-2014,Guo_etal-2015,Werner_etal-2016}.
Since $p$ can in principle be inferred from observed radiation, these simulation results may potentially elucidate plasma parameters (such as magnetization) in astrophysical sources.
Besides the power-law slope $p$, an important result of these simulations is the determination of the high-energy cutoff $\gamma_c$ of the power law, and in particular its scaling with system size $L$, since real astrophysical systems are usually much larger (with respect to kinetic scales) than we can possibly simulate.
For systems with small $L$, simulations observe ``extreme acceleration'' consistent with particles being accelerated in reconnection electric field $E_{\rm rec}\simeq 0.1(v_A/c)B_0$ (where $B_0$ is the reconnecting magnetic field) over system size $L$ \citep{Hillas-1984,Aharonian_etal-2002}, so that particles reach energies $\gamma_c m_e c^2 \simeq eE_{\rm rec} L$ \citep{Werner_etal-2016}.
However, for larger systems, this rapid acceleration seems to stall around
$\gamma_c \sim 4\sigma$, where $\sigma\equiv B_0^2/(4\upi n_{b0} m_e c^2)$ is
the ``cold'' magnetization parameter involving the reconnecting field $B_0$ and
the ambient plasma rest-mass energy density $n_{b0} m_e c^2$ \citep{Werner_etal-2016}.
More recent PIC simulations found that acceleration can continue well beyond this limit in extremely large systems, albeit at a significantly slower rate so that cutoff energies grow with the square root of time, resulting in $\gamma_c \sim \sqrt{L}$ \citep{Petropoulou_Sironi-2018,Hakobyan_etal-2021}.
The precise mechanism of reconnection-driven particle acceleration has received much attention \citep[e.g.,][]{Zenitani_Hoshino-2001,Drake_etal-2006,Zenitani_Hoshino-2007,Jaroschek_etal-2004a,Lyubarsky_Liverts-2008,Liu_etal-2011,Bessho_Bhattacharjee-2012,Cerutti_etal-2012a,Cerutti_etal-2013,Cerutti_etal-2014b,Kagan_etal-2013,Guo_etal-2014,Guo_etal-2015,Guo_etal-2019,Bessho_etal-2015,Nalewajko_etal-2015,Dahlin_etal-2016,Sironi_etal-2016,Werner_etal-2016,Petropoulou_Sironi-2018,Ball_etal-2018,Ball_etal-2019,Li_etal-2019,Kilian_etal-2020,Guo_etal-2020arxiv,Zhang_etal-2021arxiv};
a number of mechanisms have been considered, and it remains a matter of ongoing research to determine precisely which mechanism operates most effectively in which regime.
Most studies of reconnection in relativistic plasmas have relied on 2D simulations; an important outstanding question remains whether these are applicable to 3D events in nature.
A much smaller number of 3D simulations of reconnection in relativistic pair plasmas have been done \citep{Zenitani_Hoshino-2005b,Zenitani_Hoshino-2008,Yin_etal-2008,Liu_etal-2011,Kagan_etal-2013,Cerutti_etal-2014b,Sironi_Spitkovsky-2014,Guo_etal-2014,Guo_etal-2015,Werner_Uzdensky-2017,Guo_etal-2020arxiv,Zhang_etal-2021arxiv} \citep[there have also been 3D PIC simulations of non-relativistic electron-ion reconnection, e.g.,][]{Hesse_etal-2001,Pritchett_Coroniti-2001,Pritchett_Coroniti-2004,Lapenta_etal-2006,Daughton_etal-2011,Daughton_etal-2014,Markidis_etal-2012,Lapenta_etal-2014,Lapenta_etal-2020,Nakamura_etal-2013,Wendel_etal-2013,Dahlin_etal-2015,Dahlin_etal-2017,Lapenta_etal-2017,Le_etal-2018,Le_etal-2019,Pucci_etal-2018,Li_etal-2019,Stanier_etal-2019,Lapenta_etal-2020}.
Because of computational expense, 3D simulations have been smaller in physical system size, and have covered a much narrower range of regimes.
The first 3D PIC studies of relativistic reconnection focused on the competition between the tearing instability that leads to reconnection and the relativistic drift-kink instability (RDKI), which is forbidden in 2D reconnection simulations but in 3D can grow as fast as the tearing instability \citep{Zenitani_Hoshino-2005a,Zenitani_Hoshino-2007,Zenitani_Hoshino-2008,Yin_etal-2008}.
Later, simulations with sufficiently large system sizes observed NTPA, and moreover (despite clear manifestations of RDKI) found substantial similarities between 2D and 3D relativistic reconnection in both NTPA and general reconnection dynamics, including magnetic energy release and reconnection rates \citep{Liu_etal-2011,Kagan_etal-2013,Sironi_Spitkovsky-2014,Guo_etal-2014,Guo_etal-2015,Werner_Uzdensky-2017,Guo_etal-2020arxiv,Zhang_etal-2021arxiv}.
Our previous work \citep{Werner_Uzdensky-2017} systematically compared 2D and 3D simulations in the magnetically-dominated regime (where the upstream magnetic energy dominates over the upstream plasma thermal and rest-mass energy), and found magnetic energy conversion and NTPA in 2D and 3D to be nearly indistinguishable.
This was found to be the case over a range of guide magnetic fields,
$0 \leq B_{gz}/B_0 \leq 1$ (although both energy conversion and NTPA were suppressed by stronger guide field).
In the nonrelativistic regime, where NTPA occurs but may not yield power laws, 2D and 3D reconnection (with guide field) have also been found to be similar, but with slightly enhanced NTPA in 3D \citep{Dahlin_etal-2015}.
We note in passing that all of these 3D simulations began with a thin initial Harris or force-free current sheet.
The close resemblance between 2D and 3D reconnection justifies the use of 2D simulations to model natural 3D systems, allowing us to simulate larger system sizes with lower computational expense.
However, some have supposed that 3D reconnection might behave somewhat or even drastically differently, because of 3D instabilities and turbulence \citep[e.g.,][]{Lazarian_Vishniac-1999,Zenitani_Hoshino-2008,Takamoto_etal-2015,Lazarian_etal-BookChapter2016,Beresnyak-2017,Beresnyak-2018,Munoz_Buechner-2018,Takamoto-2018,ZhouX_etal-2018,Boozer-2019,Lapenta_etal-2020,Lazarian_etal-2020}
\vspace{\baselineskip}
{\bf In this paper,} we use PIC simulation to study both 2D and 3D magnetic reconnection in an ultrarelativistically-hot pair plasma in the moderately-magnetized regime where the upstream magnetic energy is comparable to the upstream plasma thermal energy (i.e., $\sigma_h=1$, where the ``hot'' magnetization $\sigma_h$ is defined in~\S\ref{sec:setup}).
In both 2D and 3D, we systematically explore the effects of different initial current sheet configurations, as well as the effect of guide magnetic field; in 2D, where very large simulations are feasible, we also vary the overall system size, and in 3D, we vary the system size in the third dimension (i.e., $L_z$).
The moderately-magnetized, ultrarelativistic pair plasma regime is of interest for several reasons.
First, it is likely relevant to astrophysical sources such as AGN jets and PWN (e.g., the Crab Nebula) in which energy might be expected to be roughly equally partitioned between plasma and magnetic field.
Second, this regime lies between magnetically-dominated relativistic reconnection and non-relativistic reconnection, both of which have been much more studied (especially in 2D); this study provides an important bridge between these regimes and also connects to reconnection in semi- or trans-relativistic \emph{electron-ion} plasma, where particles may be transrelativistic or ions may be subrelativistic while electrons are relativistic \citep[e.g.,][]{Rowan_etal-2017,Rowan_etal-2019,Ball_etal-2018,Ball_etal-2019,Werner_etal-2018}.
Third, this is perhaps the most computationally tractable regime, because the kinetic (microphysical) plasma length scales all collapse to roughly the same values, allowing the greatest dynamic range between system size and the kinetic scales, with less computational expenditure.
Using pair plasma means that electron and ``ion'' (i.e., positron) scales are the same; in the ultrarelativistic limit, the collisionless skin depth $d_e$ and Debye length $\lambda_D$ are nearly equal, $d_e=\sqrt{3}\lambda_D$; and the moderately-magnetized regime implies that the gyroradius $\rho_{b}$ in the ambient plasma is nearly the same as $d_e$.
Thus the grid spacing $\Delta x$ can be chosen just a little smaller than all these kinetic scales (which are nearly the same), maximizing the separation between the largest (hence all the) kinetic scales and the system size $L$ (for a given computational cost).
This facilitates the study of macroscopic, large-system behaviour, such as plasmoid formation and NTPA.
Reconnection in mildly relativistic, moderately magnetized pair plasma has been previously studied with 3D PIC simulations \citep{Yin_etal-2008,Liu_etal-2011,Kagan_etal-2013,Guo_etal-2020arxiv}.
However, this work features the first extensive systematic exploration of several important parameters (such as the initial current sheet configuration and guide magnetic field strength) both in 2D and 3D, running for long times with sufficiently large system sizes that allow us to investigate NTPA.
A detailed discussion of this current work in the context of those previous studies can be found in~\S\ref{sec:discussionPrevWork}.
We will show that
the evolution of current sheets can be substantially different in 2D and 3D, in particular in the way that they evolve and convert magnetic energy to plasma energy---but despite these differences, NTPA is not much changed (if anything, it is enhanced in 3D).
First, however, we will systematically characterize 2D reconnection in the
ultrarelativistic moderately-magnetized ($\sigma_h=1$) pair-plasma regime, as a number of parameters are varied---including initial current sheet configuration, system size, and guide magnetic field.
Then, we will characterize current sheet evolution in 3D across a similar range of parameters, comparing with 2D reconnection.
We will identify patches in 3D simulations that exhibit signatures of classical 2D reconnection, including outflows from thin current sheets, non-ideal electric fields (parallel to and/or larger than the magnetic field), and energy transfer from fields to plasma.
However, we will see that 3D reconnection is vastly more complicated to study than 2D reconnection, displaying a greater variety of behaviours along with greater sensitivity to the initial configuration (e.g., the initial current sheet thickness or field perturbation).
The rates of magnetic energy depletion and upstream magnetic flux reduction are generally slower in~3D, but reconnection continues for much longer times.
Ultimately, more magnetic energy is released in~3D than in~2D, because
in 2D, magnetic energy in outflows is trapped forever in plasmoids, whereas in 3D the plasmoid-like structures can decay.
We also find that, in 3D, there is a new channel for magnetic energy conversion that does not necessarily involve classical reconnection at all: the kinking of the current sheet in 3D can grow to such large amplitudes that the current sheet becomes violently and chaotically deformed, resulting in rapid magnetic energy release and turbulent thickening of the current layer.
This process is not inevitable; two initially-similar simulations can end up behaving very differently over long times because one triggers such deformation and energy release, while the other does not, even though the upstream conditions may be identical in both cases.
This sensitivity to initial conditions complicates the study of 3D reconnection by increasing the parameter space that needs to be explored.
\vspace{\baselineskip}
{\bf This paper is organized as follows.}
We will define the simulation parameters and set-up in~\S\ref{sec:setup}.
Then in~\S\ref{sec:diagnostics} we describe diagnostics used to characterize current sheet evolution, starting with precise definitions of important terms (\S\ref{sec:terminology}) and followed by detailed descriptions of diagnostics, such as how we measure the amount of unreconnected flux in a simulation and how we characterize NTPA.
In~\S\ref{sec:2d} we will describe 2D reconnection in the moderately-magnetized relativistically-hot pair-plasma regime.
We will see that in 2D, this regime exhibits behaviour that is qualitatively very similar to, e.g., magnetically-dominated relativistic reconnection, and we will use this opportunity to review the basic picture of plasmoid-dominated (multiple X-line) reconnection, to be contrasted later with 3D simulations.
In addition, we quantify the effects of varying a number of parameters (such as initial current sheet thickness, system size, and guide magnetic field) on magnetic energy conversion and NTPA.
Besides expanding our knowledge of 2D reconnection in the moderately-magnetized relativistically-hot regime, this provides a baseline for our study of 3D reconnection in~\S\ref{sec:3d}.
There we find that current sheet evolution can differ distinctly from 2D reconnection, and we investigate the effects of varying the initial current sheet configuration, varying system length $L_z$ in the third dimension, and varying guide magnetic field.
We discuss differences between the moderately-magnetized regime studied in this paper and the magnetically-dominated regime (studied in 3D in previous work) in~\S\ref{sec:highSigmah}.
We list the most important findings in~\S\ref{sec:summary} before discussing some possible impacts on plasma physics and astrophysical modelling in~\S\ref{sec:discussion}, and finally conclude with~\S\ref{sec:conclusion}.
For reference, in Appendix~\ref{sec:res2d} we present a resolution convergence study justifying the cell size~$\Delta x$ used in this work.
\section{Simulation setup}
\label{sec:setup}
The simulations in this work use a standard double-periodic
simulation box initialized with a twice-reversing magnetic field
balanced by two thin, oppositely-directed Harris current sheets
\citep[e.g.,][]{Werner_etal-2016,Werner_Uzdensky-2017}.
The Harris sheets (which contain drifting plasma) are superimposed upon a uniform, stationary, relativistically-hot background plasma.
Once the initial state is set, the simulation code {\tt Zeltron} evolves the plasma with no external input (e.g., no driving) according to a standard explicit, relativistic electromagnetic PIC algorithm \citep{Cerutti_etal-2013}, with periodic boundary conditions; in 2D simulations, all quantities are assumed to be uniform in the third dimension ($z$) at all times---e.g., $\partial B_x/\partial z=0$, since all derivatives with respect to $z$ are zero.
In this section, we first describe the background plasma, and then the
drifting plasma that forms the Harris current sheets. At the end,
we summarize the parameters describing the set-up, including whether they are fixed or varied throughout this study.
Ideally, we hope to associate reconnection behaviour with the uniform
background (or upstream) pair plasma,
described by
\begin{itemize}[leftmargin=3mm, itemindent=0mm, labelsep=1mm]
\item $n_{b0}$, the initial (electron plus positron) particle number density,
\item $\theta_b \equiv T_b/m_e c^2$, the normalized temperature,
\item $B_0$, the (reconnecting) magnetic field in the~$x$ direction, and
\item $B_{gz}\hat{\boldsymbol{z}}$, the initially-uniform guide magnetic field.
\end{itemize} \noindent
($\boldsymbol{B}$ is initially uniform except for $B_x$ reversing around the current sheets.)
It is useful to express $n_{b0}$, $\theta_b$, $B_0$, and $B_{gz}$ in
terms of a nominal gyroradius~$\rho_0$ and three dimensionless parameters, $\sigma$, $\sigma_h$, and $B_{gz}/B_0$.
We define the nominal relativistic gyroradius
$\rho_0\equiv m_ec^2/eB_0$ in place of $B_0$; thus a particle with
ultrarelativistic Lorentz factor $\gamma \gg 1$ has a gyroradius of order~$\gamma \rho_0$ in
field $B_0$, hence typical background particles have gyroradii around
$3\theta_b \rho_0$ (for $\theta_b \gg 1$).
The density~$n_{b0}$ can then be expressed in terms of $\rho_0$ and
the ``cold'' magnetization
parameter $\sigma \equiv B_0^2/(4\upi n_{b0} m_e c^2)$, which is twice
the ratio of the reconnecting magnetic field's energy density, $B_0^2/8\upi$, to plasma rest-energy density, $n_{b0}m_e c^2$.
If all magnetic energy were given to particles, their Lorentz factors would increase by $\Delta \gamma = \sigma/2$---i.e., ~$\sigma$ determines the available magnetic energy per particle. Thus
we generally normalize lengths to a characteristic scale $\rho_c \equiv \sigma \rho_0$,
which is on the order of the gyroradii of typical energized particles.
Finally, we re-express the temperature $\theta_b$ in terms of the
hot magnetization
$\sigma_h \equiv B_0^2/(16\upi \theta_b n_{b0} m_e c^2) = \sigma/(4\theta_b)$,
the ratio of magnetic enthalpy density $B_0^2/4\upi$ to plasma
enthalpy density (including rest-mass energy), which is
$h=4\theta_b n_{b0} m_e c^2$ in the relativistic limit $\theta_b\gg 1$ \citep{Melzani_etal-2014b}.
We note that (for $B_{gz}=0$) the Alfv\'{e}n 4-velocity is $c\sqrt{\sigma_h}$;
i.e., the Alfv\'{e}n velocity is $v_A=c\sqrt{\sigma_h/(1+\sigma_h)}$.
In this paper we focus on the specific case of reconnection with moderate magnetization, or $\sigma_h=1$ (the majority of studies of reconnection in pair plasma have explored $\sigma_h \gg 1$, and most electron-ion reconnection studies have been nonrelativistic with $\sigma_h \ll 1$).
The upstream plasma is then completely defined by specifying, in addition to $\sigma_h=1$, the values of $\rho_0$, $B_{gz}/B_0$, and $\sigma$.
The particular value of $\rho_0$ is irrelevant; the entire simulation scales trivially with $\rho_0$.
For astrophysical relevance and theoretical simplicity,
we choose to study the ultrarelativistically-hot regime in which $\sigma \gg 1$ and $\theta_b\gg 1$---as with $\rho_0$, the particular value of $\sigma$ does not matter (the simulation scales trivially with $\rho_c\equiv \sigma \rho_0$ as long as we stay in the ultrarelativistically-hot regime).
When $\sigma_h=1$, the upstream magnetic and plasma energy are roughly comparable; even complete conversion of magnetic energy would not drastically increase the plasma energy.
For $B_{gz}=0$, $\sigma_h=1$ implies that the plasma beta is $\beta_{\rm plasma}=1/2$ and the
expected bulk reconnection outflow velocity is $v_A=c/\sqrt{2}$.
Important plasma length scales, in terms of $\sigma\rho_0$ and $\sigma_h$, include the average background gyroradius
$\rho_b\equiv 3\theta_b \rho_0 = (3/4)\sigma_h^{-1} \sigma \rho_0$,
the background Debye length $\lambda_D =(1/2) \sigma_h^{-1/2} \sigma\rho_0 $,
and the collisionless skin depth
$d_e = \sqrt{3\theta_b m_e c^2/4\upi n e^2} =\sqrt{3} \lambda_D$.
Over time, the background plasma should (at least in 2D) dominate reconnection
dynamics.
However, to start in a state susceptible to reconnection, we
need a reversing magnetic field and its associated current sheet \citep{Kirk_Skjaeraasen-2003}.
Actually we use \emph{two} oppositely-directed
Harris current sheets to allow periodic boundary conditions in all
directions.
The simulation box, containing both current sheets, has size $L_x\times L_y \times L_z$
where $x$ is the direction of reconnecting
magnetic field, $y$ is perpendicular to the initial current sheet,
and $z$ is the ``third'' dimension, parallel to the initial current.
We initialize this box with a (doubly) reversing magnetic field:
\begin{eqnarray}
\boldsymbol{B}(t=0) &=& \hat{\boldsymbol{z}} B_{gz} +
\hat{\boldsymbol{x}} B_0 \times
\left\{ \begin{array} {r@{\quad}l}
-{\rm tanh} [(y\:-\,\phantom{3}L_y/4)/\delta] & \textrm{for } y < L_y/2 \\
\phantom{-}{\rm tanh} [(y\:-\,3L_y/4)/\delta] & \textrm{for } y > L_y/2
\end{array} \right.
\end{eqnarray}
where $\delta$ is the half-thickness of the initial current sheet.
All simulations in this work use $L_y=2L_x$;
although the lower and upper halves of the simulation can interact,
the separation $L_y/2=L_x$ between current sheets has been found
to limit the interaction so that, e.g., the reconnection rate
is the same as in simulations with much larger $L_y$ (also, in 3D, the current sheets kink and stray more in the $y$ direction; it is important that $L_y$ be large enough that they never touch).
When presenting results, we often focus on just one current sheet (the lower one) for simplicity,
but display results involving total energy and particle spectra for the
entire simulation.
The system length $L_z$ in the third dimension will be varied to investigate 3D effects, but the fiducial value for 3D simulations will be $L_z=L_x$.
To balance the magnetic field reversal (according to Ampere's law),
as well as to provide pressure balance against the upstream magnetic
field, we add a ``drifting'' plasma in addition to the background
plasma \citep{Kirk_Skjaeraasen-2003}.
This initially-drifting plasma has a total particle density strongly
peaked around the magnetic field reversal,
$n_d(y) = n_{d0} \textrm{cosh}^{-2}[(y-y_c)/\delta]$ (measured in
the simulation frame), where $y_c$ is the centre of the current sheet;
$y_c=L_y/4$ for the lower sheet, or $y_c=3L_y/4$ for the upper.
Within each sheet, the electrons and positrons drift in opposite $\pm \hat{\boldsymbol{z}}$
directions with
speed $\beta_d c$ [hence Lorentz factor $\gamma_d=(1-\beta_d^2)^{-1/2}$], and the initial temperature in the co-moving
frame is $\theta_d m_e c^2$.
The initial current sheet configuration can be specified by the drift $\beta_d$ and the overdensity $\eta\equiv n_{d0}/n_{b0}$, with
the remaining parameters, $\theta_d$ and $\delta$, determined by
pressure balance and Ampere's law:
$n_{d0} \theta_d/\gamma_d = B_0^2/8\upi m_e c^2 = \sigma n_{b0}/2$
and $B_0/\delta = 4\upi e n_{d0} \beta_d$.
In terms of the characteristic length $\sigma \rho_0$, the sheet thickness is
$\delta = (n_{b0}/n_{d0})\sigma \rho_0/\beta_d$; and in terms of the gyroradius $\rho_d \equiv 3\theta_d \rho_0$ of drifting particles,
$\delta = 2\rho_d/(3\beta_d \gamma_d)$.
In previous (especially 2D) reconnection simulations, the initial
current sheets have often been slightly perturbed to trigger reconnection
faster (reducing computation time) and to create an initial X-point in
a predetermined location. Although such a perturbation can have an effect in
2D \citep{Ball_etal-2019}, we find that small perturbations leave 2D
simulations basically unchanged in many important ways
(cf.~\S\ref{sec:pertAndEta2d}) as the simulation becomes dominated by
the background plasma; however, the effect in 3D can be more significant (cf.~\S\ref{sec:pertAndEta3d}).
While most of our simulations begin with zero perturbation, we have
explored in a few cases the effect of a small magnetic field perturbation---a tearing-type perturbation with a single X-point and a single magnetic island with a small height (in $y$) but a long width (in~$x$).
The perturbation, expressed in the vector potential, is
$\boldsymbol{A}=\hat{\boldsymbol{z}} A_z$:
\begin{eqnarray} \label{eq:pert}
A_z(x,y) &=& \left[1 +
0.877 a \frac{\delta}{L_y} \cos \left( \frac{2\upi x}{L_x} \right)
\cos^2 \left( \frac{2\upi (y-y_c)}{L_y} \right) \right]
\nonumber \\
& & \qquad \times
B_0 \delta
\left[ \ln \textrm{cosh} \left( \frac{L_y}{4\delta} \right)
- \ln \textrm{cosh} \left( \frac{y-y_c}{\delta} \right)
\right]
\end{eqnarray}
where $a$ is the perturbation strength.
Although $a$ is useful for initializing $A_z(x,y)$, it is also useful to characterize the perturbation in terms of $s$,
the half-height of the initial magnetic island.
For $\delta \ll L_y$ and $s \ll L_y$, $s$ and $a$ are related approximately by
$\textrm{cosh}(s/\delta) \approx \exp (0.44 a)$.
In the above, $a$ is thus normalized so that when
$a=1$, we have $s/\delta=1$; i.e., the separatrix extends roughly as far as the initial current sheet.
For $a \lesssim 1$, the initial separatrix lies within the initial
current sheet ($s \lesssim \delta$).
We note that the ``wavelength'' of the perturbation in $x$ is $L_x$---the
longest that can fit in the simulation.
For 3D simulations, this perturbation is uniform in $z$.
Unless specifically noted, $a=0$.
At the beginning of each simulation, particle velocities for the background (and drifting) populations are initialized
randomly, drawn from the appropriate (drifting) relativistic Maxwell-J\"{u}ttner
distribution. Because of this randomness, two simulations that are
identical in all macroscopic initial parameters (i.e., differing only in random initial positions and velocities) will not
yield completely identical results, even for global quantities such as the total
magnetic energy depletion or total particle energy spectrum.
Although computational expense prohibits running large statistical ensembles of simulations for every macroscopic parameter choice, one goal of this study is the estimation of ``stochastic variation'' due to randomized particle initialization (especially in 3D; cf.~\S\ref{sec:variability3d}).
In a limited number of cases, we will therefore show the results of multiple macroscopically-identical simulations as a means of gauging stochastic variability, and distinguishing it from systematic effects correlated with macroscopic parameters.
In summary, the reconnection simulation setup is described by the following physical parameters:
\begin{itemize}[leftmargin=3mm, itemindent=0mm, labelsep=1mm]
\item The nominal gyroradius $\rho_0=m_ec^2/eB_0$ is fixed for all simulations; its value is irrelevant---we can trivially scale simulation results to any other value of $\rho_0$.
\item The upstream (cold) magnetization (excluding guide field) $\sigma = B_0^2/(4\upi n_{b0} m_e c^2) = 10^4 \gg 1$ is fixed in all simulations; its particular value is irrelevant, as long as $\sigma \gg 1$.
\item The upstream hot magnetization (excluding guide field) $\sigma_h = \sigma/(4\theta_b) = 1$ is fixed
in all simulations; its value puts this study in the moderately-magnetized regime. This implies (for $B_{gz}=0$) plasma $\beta=1/2$ and $v_A=c/\sqrt{2}$.
(Previous studies of reconnection in pair plasma have mostly concentrated on the $\sigma_h\gg 1$ regime, while nonrelativistic reconnection studies have $\sigma_h \ll 1$.)
\item The initial guide magnetic field $B_{gz}$ is zero except in~\S\ref{sec:Bz2d} and in~\S\ref{sec:Bz3d}, when it is varied up to $B_{gz}=4B_0$ in 2D and up to $B_{gz}=B_0$ in 3D, respectively.
\item The initial current sheet overdensity, $\eta=n_{d0}/n_{b0}$, is set to $\eta=5$, except when varied systematically in~\S\ref{sec:pertAndEta2d}, \S\ref{sec:pertAndEta3d}, and~\S\ref{sec:RDKIamp}.
\item The average drift speed of particles forming the initial Harris sheet is $\beta_d c = 0.3 c$, except when varied in \S\ref{sec:pertAndEta2d}, \S\ref{sec:pertAndEta3d}, and~\S\ref{sec:RDKIamp}.
\item The initial current sheet half-thickness $\delta=(\eta\beta_d)^{-1} \sigma\rho_0$ is $(2/3)\sigma\rho_0$ except when $\eta$ and $\beta_d$ are varied as noted above.
\item The initial magnetic field perturbation strength $a$ (see~Eq.~\ref{eq:pert}) is zero except when varied in \S\ref{sec:pertAndEta2d} and \S\ref{sec:pertAndEta3d}.
\item The system size $L_x$ determines the overall system size, and is desired to be as large as possible, but limited by computational resources; its value will be noted in each subsection, where we generally compare simulations with the same $L_x$. However, the effect of varying $L_x$ is specifically investigated in~\S\ref{sec:Lx2d}.
\item The system aspect ratio $L_y/L_x = 2$ in all simulations.
\item The system aspect ratio $L_z/L_x$ is considered to be zero in
all 2D simulations; in 3D simulations the value of $L_z/L_x$ is noted in each case, although we most commonly use $L_z=L_x$;
we systematically explore the effect of varying $L_z/L_x$ in~\S\ref{sec:Lz3d}.
\item For 2D simulations, a simulation time $T\approx 10L_x/c$ is usually sufficient; 3D simulations were run for longer times, 20--50$L_x/c$ (the longest runs are shown in~\S\ref{sec:Lz3d}).
\end{itemize}
For the various choices of above parameters, important plasma length scales will be summarized in table~\ref{tab:etaEffect} (in~\S\ref{sec:pertAndEta2d}).
Using periodic boundary conditions offers theoretical and numerical simplicity; and it is usually the simplest way to simulate a mesoscopic box---i.e., a small part of a system whose global inhomogeneity scale along the layer is much larger than the computational box.
For application to real systems, the effect of these or any boundary conditions must ultimately be studied, e.g., by examining simulation size dependence or by comparing simulations with different boundary conditions; alternatively, running simulations for less than one light-crossing time ($t\lesssim 1\,L_x/c$) can lessen effects of boundaries, but at the cost of exacerbating the effect of the initial conditions.
In this work, we run simulations well beyond~$1\,L_x/c$ for two primary reasons:
(1) the very early evolution, $t \lesssim 1L_x/c$, may be heavily influenced by initial conditions (which are uniform in the third dimension) and may fail to capture the most interesting and important 3D phenomena, and (2) the evolution in 3D can be relatively slow.
Over long times, boundary conditions might well be expected to affect system evolution significantly, and, this being the case, there is a trade-off between possibly more realistic but complicated boundaries \citep[like ``open'' boundaries, e.g.,][]{Daughton_etal-2006} and the simpler, better-understood periodic boundaries.
However, the effect of periodic boundaries is arguably realistic for long simulations, e.g., $t\sim 30\,L_x/c$, if the inhomogeneity scale of the real system is larger than $\sim 30L_x$.
Determining the effect of various boundary conditions on current sheet evolution in~3D is important, but beyond the scope of this work.
Finally, we summarize the numerical parameters.
The grid resolution $\Delta x=\Delta y=\Delta z=\sigma\rho_0/3$ marginally resolves the background Debye length and skin depth
$d_e/\sqrt{3}=\lambda_D=\sigma \rho_0/2$ as well as the background
gyroradius $\rho_b = (3/4)\sigma \rho_0$ and the initial current sheet
[usually $\delta=(2/3)\sigma \rho_0$].
This marginal resolution allows us to simulate the largest system sizes
with the computational resources available; Appendix~\ref{sec:res2d} shows
that this resolution is sufficient.
The timestep $\Delta t=\Delta x/(c\sqrt{d})$, where the dimensionality
$d$ is 2 or 3,
is determined by the Courant-Friedrichs-Lewy condition (all 2D simulations compared directly against 3D simulations were run with $d=3$).
Simulations were initialized with $20\times 4=80$ macroparticles per cell: 20 background electrons and positrons, plus 20 ``drifting'' electrons and positrons.
The initially-drifting particles were weighted (depending on their $y$-position) to represent the non-uniform current sheet density, and particles with negligibly-low weights were deleted from the simulation.
Thus our largest simulation ($L_x=L_z=512\sigma\rho_0$, with $1536\times 3072\times 1536$ cells, cf.~\S\ref{sec:overview3d}) contained about 300 billion macroparticles.
Energy is not precisely conserved by the PIC algorithm, but with these numerical parameters it is approximately conserved in all simulations to better than 1~per~cent---and in almost all but the larger 2D simulations, energy is conserved to better than 0.1~per~cent.
\section{Diagnostics}
\label{sec:diagnostics}
In this section we will define terms used to characterize reconnection (in~\S\ref{sec:terminology}) and then provide detailed descriptions of diagnostics that we will use:
the central surface of the ``layer'' or current sheet, $y_c(x,z,t)$ (\S\ref{sec:sheetCenter});
global volume-integrated characteristics as functions of time---unreconnected flux (\S\ref{sec:unreconnectedFlux}) and various energy components (\S\ref{sec:diagEnergy});
the reconnection onset time $t_{\rm onset}$ (\S\ref{sec:onsetTime});
the local bulk velocity of the plasma (\S\ref{sec:bulkVel});
and the power-law index and high-energy cutoff of a particle energy distribution (\S\ref{sec:NTPAfit}).
\subsection{Terminology}
\label{sec:terminology}
Before describing diagnostics for reconnection, we offer explicit definitions or clarifications of some often-used terms, so that we can avoid lengthy qualifications in the text.
{\bf Transverse:} the $x$ and $y$ directions, transverse to $z$, the direction of the guide magnetic field.
{\bf Transverse magnetic field energy $U_{Bt}$:} the energy in the $B_x$ and $B_y$ components of the magnetic field---the volume integral of $(B_x^2+B_y^2)/8\upi$ over the entire system.
Even in simulations with substantial guide magnetic field, it is mainly the transverse magnetic energy (and not energy in $B_z$) that is depleted during reconnection.
{\bf Guide magnetic field energy $U_{Bz}$:} the energy in the $B_z$ components of the magnetic field.
Because~$B_z(x,y,z)$ is initially uniform and the flux $\int B_z dx dy$ through any transverse plane is exactly conserved in the simulation, $U_{Bz}$ can only increase from its initial value (in practice it does not increase much and the increase can often be neglected).
{\bf Plasma energy:} the total kinetic energy of individual plasma particles.
{\bf Magnetic energy conversion:} the conversion of (transverse) magnetic field energy $U_{Bt}$ to plasma energy. Because the guide magnetic field energy~$U_{Bz}$ cannot decrease, it is essentially equivalent to refer to magnetic energy conversion or transverse magnetic energy conversion.
{\bf Plasma energization:} the conversion of (transverse) magnetic energy to plasma energy.
{\bf Released magnetic energy:} (transverse) magnetic energy that has been
converted to plasma energy.
{\bf Magnetic energy depletion:} (transverse) magnetic energy conversion, with emphasis on the reduction in magnetic energy (as it is converted to plasma energy).
This is nearly a synonym for ``dissipation,'' except that ``dissipation'' refers to irreversible heating, whereas ``depletion'' or ''conversion'' also include conversion to bulk flow kinetic energy (in this paper, however, we will not measure the difference between depletion and dissipation).
{\bf Unreconnected (magnetic) field (line):} a magnetic field line that crosses the entire simulation domain in the $x$ direction (possibly diagonally), without reversing (in $x$). As we discuss in~\S\ref{sec:unreconnectedFlux}, this is an imperfect, approximate, but practical definition.
{\bf ``Reconnected'' (magnetic) field (line):} any magnetic field line that is not unreconnected (according to the definition above). In 2D reconnection, a reconnected field line is a closed loop around one or more magnetic islands or plasmoids, and our definition of a ``reconnected'' field line identifies most closed loops accurately. In 3D, however, there may be many field lines that are ``not unreconnected'' but do not resemble closed loops or even spirals around flux ropes; we nevertheless call them ``reconnected'' and will often use quotation marks as a reminder that they do not necessarily resemble reconnected field lines in 2D reconnection.
{\bf Upstream:} the upstream region includes all points on ``unreconnected'' field lines. By extension, upstream plasma is the plasma in this region, upstream magnetic energy is the magnetic energy in this region, etc.
{\bf Upstream value:} When we refer to specific upstream values, such as the upstream $\sigma_h$, $B_0$, $n_{be}$, or $\theta_b$, we mean the asymptotic or far upstream values.
{\bf Current sheet (or sheets):} the region containing currents that support the reversal in magnetic field. We often refer to the initial current sheet, which is well defined, and otherwise use the term to refer to sheet-like regions where $J_z$ is strong.
{\bf The layer (reconnection layer, current layer):} the complement of the upstream region---i.e., the region containing ``reconnected'' field lines. This region contains the current sheet (or current sheets) as well as reconnection outflows and plasmoids. We think of ``the layer'' as the evolved current sheet. Each (double-periodic) simulation contains two layers---the upper and lower layers.
{\bf Separatrix:} the surface between the ``upstream'' region and ``the layer,'' i.e., that separates ``reconnected'' and ``unreconnected'' magnetic field lines.
In 2D this is a smooth, well-defined surface. In 3D that may not be the case, but none of our analysis will be sensitive to the precise location of the separatrix.
{\bf Unreconnected flux:} the integral of $B_x$ over the $x=0$ plane intersected with the ``upstream'' region; i.e., the integral is over all (and only) unreconnected field. See~\S\ref{sec:unreconnectedFlux} for more detail.
{\bf Reconnected flux:} the flux of magnetic field around the major (largest) plasmoid. This concept is well defined and easily measured in 2D, where the unreconnected plus reconnected flux is conserved (cf.~\S\ref{sec:unreconnectedFlux}). It is nontrivial, however, to define and measure this in 3D, and we will not do so.
{\bf Upstream magnetic energy $U_{Bt,\rm up}$:} the transverse magnetic energy in the upstream region. We exclude guide field energy from this quantity, because we are primarily interested in conversion of magnetic energy to plasma energy.
{\bf Magnetic energy in the layer $U_{Bt,\rm layer}$:}, the transverse magnetic energy in the layer; $U_{Bt,\rm up} + U_{Bt,\rm layer}=U_{Bt}$.
We exclude guide field energy from this quantity.
{\bf Reconnection:} technically, ``reconnection'' should involve the conversion/rearrangement of ``unreconnected field lines'' to ``reconnected field lines'' in a way that conserves the total unreconnected plus reconnected flux. In 2D reconnection, this conservation is easily verified.
However, in 3D it is nontrivial to decide precisely how to characterize reconnection, much less to measure true reconnection rates. Rather than argue for any particular precise definition of reconnection, we will focus on more tangible properties of the current layer, such as magnetic energy and plasma energy. We will use the words ``3D reconnection'' loosely to refer to any magnetic-energy-depleting evolution of thin current sheets that, in principle, could or would undergo 2D-like reconnection.
However, any discussion of reconnecting \emph{flux} or \emph{field lines} will refer specifically to the 2D-like flux-conserving process.
{\bf Flux annihilation:} the utter disappearance of upstream (unreconnected) magnetic flux by unspecified means. In contrast, 2D reconnection does not annihilate flux (it conserves flux as described above). For example, magnetic diffusion in a plasma with finite resistivity can directly annihilate flux; although this process is far too slow to explain observed magnetic energy releases (assuming neat laminar current sheets), it could be considerably enhanced, e.g., by effective turbulent magnetic diffusivity.
Incidentally, the flux $\int B_x dy dz$ through any $x$-plane in the simulation is exactly conserved---however, by virtue of the reversing magnetic field, this flux is zero, and it tells us nothing about how much flux is reconnected or annihilated.
{\bf (Upstream) flux depletion:} the depletion of the upstream magnetic flux, without regard to the process (e.g., reconnection or annihilation).
\subsection{Current sheet (layer) central surface}
\label{sec:sheetCenter}
It is useful to analyse some field quantities at the current sheet ``centre'' (in $y$), even as the layer kinks and deforms.
In the initial state, the current sheet (or layer) central surface $y_c(x,z,t=0)$ is the location
where $B_x(x,y_c,z,t=0)=0$, i.e., where $B_x$ crosses through zero in the $y$ direction, reversing sign.
At later times, we continue to define $y_c(x,z,t)$ in the same way, with the following complication.
The layer may become extremely distorted by the nonlinear development of the kink instability so that it folds over on itself, resulting in multiple field reversals.
Because the upstream magnetic field never changes sign, $B_x(x,y,z,t)$ will
always reverse sign at least once along $y$; however, it may reverse sign an odd number of times at each layer.
If there are multiple reversals in $y$ at some $(x,z)$ and time $t$, we take $y_c(x,z,t)$ to be the middle reversal [e.g., if there are three reversals, $y_c(x,z,t)$ will be at the second reversal].
We define the displacement of the central surface, $\Delta y_c(x,z,t) \equiv y_c(x,z,t) - y_c(x,z,0)$.
In~\S\ref{sec:Lz3d} we look at~$\tilde{y}_c(k_z,t)$, the Fourier spectrum of $\Delta y_c(x,z,t)$ in~$z$, averaged over~$x$, at a given time~$t$.
To find the ``Fourier spectrum in $z$ averaged over $x$,''
we take the power spectrum in $z$ of $\Delta y_c(x,z,t)$
at every $x$ (for a given time $t$),
then average the power spectra over $x$, and take the
square root. We normalize the result $\tilde{y}_c(k_z,t)$ so that
it indicates the amplitude of the sinusoidal component with $k_z$.
E.g., if $\Delta y_c(x,z,t)=A\sin(k_z z)$, then $\tilde{y}_c(k_z,t)=A$.
\subsection{Unreconnected flux and reconnection rate}
\label{sec:unreconnectedFlux}
Reconnection breaks upstream (unreconnected) field lines
and reconnects them in a different configuration in the downstream
outflows.
Defining and measuring unreconnected flux is straightforward in 2D using the vector potential $A_z(x,y)$, which is constant along field lines.
In 3D, however, it is much more difficult to define unreconnected flux precisely, and so we adopt the following practical definition.
An unreconnected field line is a field line that runs from $x=0$ to $x=L_x$ without changing direction in $x$ ($B_x$ never changes sign).
In 2D, this definition is unambiguous and almost always agrees with the method using $A_z$ (disagreement could occur in the rare case of an unreconnected field line with an $\Omega$-shaped kink; we believe such cases are negligibly rare in 2D, but we have not studied whether they are also rare in 3D).
In 2D, a field line that runs from $x=0$ to $x=L_x$ wraps around via periodic boundary conditions exactly onto itself; therefore, this definition unambiguously determines whether the entire field line has the same sign for $B_x$.
In 3D, this is not necessarily true; nevertheless this method captures the most obviously-unreconnected upstream field lines with $|B_y|\ll |B_x|$, that
run across the simulation box with little deflection in the $y$ direction.
Specifically, we trace a field line from each grid node $(0,y,z)$ in the $x=0$ plane, following it in the $+x$ direction;
if the field line ever reverses direction in $x$,
then the field line is considered to be ``reconnected.''
If the sign of $B_x$ does not reverse, then the field line must eventually reach $x=L_x$, at which point we stop and consider the field line to be ``unreconnected.''
The unreconnected flux is then estimated as $\sum_{y,z} B_x(0,y,z) \Delta y \Delta z$ where the sum is over nodes $(y,z)$ on unreconnected field lines,
and $\Delta y \Delta z$ is the cell face area.
We add the flux for all unreconnected field lines between the central surfaces of the two layers (cf.~\S\ref{sec:sheetCenter}) to obtain the total unreconnected flux.
During this process of tracing field lines, we note the index of every cell that is penetrated by an unreconnected field line and consider any such cell to be part of the ``upstream'' region; any cell not penetrated by any unreconnected field line is considered part of the ``layer.''
In this way we can calculate, for example, the magnetic energy in upstream and layer regions.
The boundary between regions of unreconnected and reconnected field lines is (an approximation of) the separatrix, which can be clearly seen in Fig.~\ref{fig:ndePerturb2dLx320}; (in 2D, at least) the separatrix bounds magnetic islands (O-points or plasmoids) and goes through X-points.
This analysis is rigorously based on 2D reconnection; in 3D it still paints a useful picture, even though the notions of ``unreconnected'' and ``reconnected'' are no longer precisely well-defined, and the analysis does not distinguish between reconnection and annihilation of upstream flux (i.e., between reconnected flux and annihilated flux).
In 2D, any field line that is not unreconnected is reconnected, forming a closed loop in a magnetic island (when $B_z$ is ignored). Moreover, we find that in 2D, flux is conserved---specifically, the flux between the major O-points of the two layers (in the double-periodic simulation) is conserved. This flux can be divided into two parts: the upstream unreconnected flux between the two layers, plus the reconnected flux between the major O-point and separatrix for each layer, and the sum of these parts remains constant, at least within our measurement precision, which is better than 1~per~cent of the initial flux.
Therefore, in 2D, flux is not annihilated or destroyed (in a more resistive plasma, flux would be annihilated; and even in a ``collisionless'' PIC simulation, numerical resistivity will eventually annihilate flux, but only over extremely long times).
In 3D, field lines that are ``not unreconnected'' do not necessarily form a closed loop around an O-point or flux rope,
but again, for brevity we will still use the term ``reconnected'' for such field lines.
Periodically we include quotation marks around ``reconnected'' as a reminder that it really means ``not unreconnected.''
In addition (spoiler alert!), we will find that in 3D, flux is not conserved in the way it is in 2D: some flux is outright annihilated.
When we refer to a ``reconnection rate,'' we mean the rate at which the upstream magnetic flux decays---regardless of whether this flux depletion occurs due to reconnection (as always in 2D collisionless reconnection) or to annihilation (as might be happening in 3D simulations).
The reconnection rate represents the rate at which upstream flux is changing, and therefore represents an electric field along an X-line by Faraday's law, if a suitable X-line (or reasonable approximation) exists.
We use the symbol $\beta_{\rm rec}$ for the dimensionless reconnection rate normalized to $cB_0$:
$\beta_{\rm rec} \equiv -(cB_0L_z)^{-1} d\psi/dt$, where $\psi(t)$ is the
flux upstream of one layer (or half the flux between the two layers).
Usually we obtain values of $\beta_{\rm rec}$ averaged over some
time---for example the time over which $\psi(t)$ falls from~$0.9\psi_0$ to~$0.8\psi_0$, or alternatively, 0.8--0.7$\psi_0$ (the choice between intervals often depends on whether all simulations being compared actually reached $0.7\psi_0$).
Sometimes we also quote~$(c/v_A)\beta_{\rm rec}$, the dimensionless reconnection rate normalized to $B_0 v_A$, where $v_A$ is the Alfv\'{e}n velocity.
\subsection{Energy}
\label{sec:diagEnergy}
Energy is a powerful diagnostic because of its physical importance, because
it is conserved (to a good approximation in these simulations), and
because it is generally straightforward and unambiguous to measure.
We will calculate various energy components integrated over the entire simulation volume (at any given time $t$), including total particle
energy $U_{\rm plasma}(t)$ and electric field energy $U_{E}(t)$.
The component of most frequent interest is the magnetic field energy $U_B(t)$, and more specifically the transverse magnetic field energy, $U_{Bt}(t)\equiv \int dV\, (B_x^2+B_y^2)/8\upi$;
it is mainly $U_{Bt}$
that gets converted to particle energy $U_{\rm plasma}$, while
$U_{Bz}\equiv U_{B}-U_{Bt}$ remains relatively constant (or at least negligibly small) over the course of a simulation.
For zero guide field, $U_{Bt}\approx U_B$ and it does not much matter which we use;
when comparing simulations with different guide fields, however, it is more illuminating to compare $U_{Bt}$.
We also calculate the upstream magnetic energy $U_{Bt,\rm up}$ in transverse magnetic
field components of unreconnected field lines, and then define
the magnetic energy in the layer
(i.e., in ``reconnected'' field lines) to be
$U_{Bt,\rm layer}=U_{Bt} - U_{Bt,\rm up}$
(see~\S\ref{sec:unreconnectedFlux}).
\subsection{Onset time $t_{\rm onset}$}
\label{sec:onsetTime}
Sometimes the initial current sheet configuration strongly affects the time it takes reconnection to start; when comparing time evolution of simulations with different initial configurations, it is sometimes helpful to compare different cases relative to the onset time rather than $t=0$.
We define the onset time as the time $t_{\rm onset}$ when the transverse magnetic energy $U_{Bt}(t)$ has declined by 1 per~cent from its initial value $U_{Bt0}$.
Although this value is somewhat arbitrary, no results will depend
strongly on this choice, as long as it is used consistently.
Occasionally the time axis of graphs will be shifted by $t_{\rm onset}$ to allow more revealing comparison between different simulations, which may have similar time evolution apart from different onset times.
\subsection{Bulk Velocity}
\label{sec:bulkVel}
We compute electron bulk velocities $\boldsymbol{v}_e=\boldsymbol{J}_e/\rho_e$ (in each grid cell),
where $\boldsymbol{J}_e$ and $\rho_e$ are the current and charge density due to electrons; similarly, the positron velocity is $\boldsymbol{v}_i=
\boldsymbol{J}_i/\rho_i$.
The plasma velocity is then computed as the average of electron and ion
velocities, $\boldsymbol{v}\equiv (\boldsymbol{v}_e + \boldsymbol{v}_i)/2$.
More accurately, this average should be weighted by the mass (or $\gamma m$) of particles in each cell; however, in this paper, quasineutrality and the symmetry between electrons and positrons make the simple average a reasonable approximation.
\subsection{Particle energy spectra and power-law fitting}
\label{sec:NTPAfit}
Particle energy distributions $f(\gamma)$ are shown integrated over the entire simulation box at single time snapshots (where the Lorentz factor
$\gamma$ is a proxy for particle energy $\gamma m_e c^2$).
The energy spectra often display power laws $f(\gamma)\sim \gamma^{-p}$ extending to high energies (well above the average energy).
We determine the power-law index~$p$ by finding the longest, straightest segment on a log-log plot \citep{Werner_etal-2018}.
We compute the local slope, $p(\gamma)=-d\ln f/d\ln \gamma$ and search exhaustively for the interval $[\gamma_1,\gamma_2]$ with the largest $\gamma_2/\gamma_1$ such that $p(\gamma)$ remains within a range of $\Delta p$ over the entire interval.
We then choose ``the'' power-law index~$p$ to be the median~$p(\gamma)$ over~$[\gamma_1,\gamma_2]$.
We do this separately for $\Delta p=$0.1, 0.2, and~0.4, as well as using spectra at different nearby times, considering variation in~$p$---whether due to time variation or choice of~$\Delta p$---as uncertainty in the measurement.
In this paper, we display ``error bars'' on~$p$ comprising the middle 68~per~cent of all the values of~$p$ measured.
To find the cutoff of the high-energy power law, we fit $f(\gamma)$ over $[\gamma_1,\gamma_2]$ to the form $A\gamma^{-p}$, where~$p$ is determined as above (thus only the normalization~$A$ must be found).
We then consider the cutoff~$\gamma_c$ to be the energy at which~$f(\gamma_c)=e^{-1} A\gamma_c^{-p}$.
\section{2D reconnection with moderate magnetization: basic evolution and NTPA}
\label{sec:2d}
Before exploring 3D reconnection, we investigate reconnection---in
ultrarelativistically-hot pair plasma with $\sigma_h=1$---in 2D,
systematically varying a number of parameters.
We will begin with an overview of 2D reconnection (\S\ref{sec:overview2d});
then we will investigate the effects of different initial current sheet
configurations in~\S\ref{sec:pertAndEta2d}. In~\S\ref{sec:Lx2d} we
study system-size effects, and finally in~\S\ref{sec:Bz2d} we report
on the effects of guide magnetic field.
This 2D reconnection study serves multiple purposes.
First, it is of interest in its own right to characterize 2D reconnection
in the ultrarelativistically-hot $\sigma_h=1$ regime (most previous reconnection studies have focused on $\sigma_h \ll 1$, usually in nonrelativistic electron-ion plasma, or $\sigma_h \gg 1$).
Second, much larger simulations are possible in 2D; only in 2D, therefore, can we really explore system-size dependence.
And last, to determine whether 3D reconnection is different, we need to compare with 2D simulation; this 2D study provides a baseline for the subsequent 3D study (\S\ref{sec:3d}).
\subsection{Overview of 2D reconnection with $\sigma_h=1$}
\label{sec:overview2d}
We begin by reviewing the familiar behaviour of plasmoid-dominated reconnection in 2D, for a single representative simulation with
$L_x=1280\sigma \rho_0$; other initial parameters are: $B_{gz}=0$, $\eta=5$, $\beta_d=0.3$, $\delta=(2/3)\sigma\rho_0$, $a=0$.
This simulation exhibits the familiar behaviour of 2D plasmoid-dominated reconnection---it is qualitatively similar to plasmoid-dominated reconnection at small and large $\sigma_h$.
The simulation lingers in its initial state until the
tearing instability triggers reconnection, resulting in a chain of plasmoids (magnetic islands or O-points) in each layer.
(In the following, we describe just one of the layers in the double-periodic system; both layers behave qualitatively similarly.)
Reconnection occurs at X-points in thin, elementary current sheets between plasmoids: at X-points, upstream magnetic field lines are broken, and reconnected in a different topology, with reconnected field lines wrapping around plasmoids in the reconnection outflows.
In this process, some of the upstream magnetic energy is converted to particle/plasma energy, while some of it ends up as magnetic energy in the layer (i.e., in ``reconnected'' field in plasmoids).
Secondary tearing breaks up the elementary current sheets when they become too long and too thin, detaching the small plasmoids from the X-point that fed them magnetic energy and reconnected flux. Thereafter, those plasmoids no longer grow via reconnection, but instead grow by merging with other plasmoids via the coalescence instability, which conserves the flux around O-points.
A hierarchy of different-sized plasmoids thus develops, and
eventually a single monster plasmoid dominates the simulation, continuing to grow as it consumes (i.e., merges with) smaller plasmoids.
If $L_y/L_x$ is large enough, this major plasmoid eventually grows to a size of order $L_x$, and reconnection can no longer continue; the angle of the magnetic separatrix at the major X-point opens to 90 degrees, and reconnection effectively stops (actually, systems often oscillate with low amplitudes about the final state, with magnetic energy and flux sloshing between the major plasmoid and the upstream).
If $L_y \gg L_x$ (even if $L_y \gtrsim 2L_x$), there will still be significant upstream ``unreconnected'' magnetic field that could in principle be reconnected were it not prevented from doing so by reaching a stable magnetic configuration (at least on reconnection timescales; magnetic diffusion might deplete additional magnetic energy on much, much longer timescales).
During this evolution, upstream magnetic flux is reconnected and some upstream magnetic energy is converted into plasma energy, while some ends up permanently stored in the magnetic field of the major plasmoid.
To describe this process more quantitatively, we will rely heavily on several important diagnostics.
These will be essential throughout the rest of the paper, as we explore reconnection over a very large parameter space, to help us compare and contrast reconnection with different parameters including $a$, $\eta$, $\beta_d$, $L_x$, $B_{gz}$, and (in 3D) $L_z$.
We will measure, for example, the amount of magnetic flux that is reconnected over time (cf.~\S\ref{sec:unreconnectedFlux}).
However, perhaps of even more importance is the behaviour of various energy components---for example, the plasma or magnetic energy versus time (cf.~\S\ref{sec:diagEnergy}).
In addition, we further decompose the plasma energy into the energy distribution of particles to distinguish thermal heating from NTPA (cf.~\S\ref{sec:NTPAfit}).
In the following, we describe these diagnostics for the single representative simulation.
\begin{figure}
\centering
\includegraphics*[width=0.325\textwidth]{fig01a.pdf}%
\hfill
\includegraphics*[width=0.325\textwidth]{fig01b.pdf}%
\hfill
\includegraphics*[width=0.325\textwidth]{fig01c.pdf}%
\caption{ \label{fig:energyFluxVsTime2D}
The left panel shows the unreconnected flux (normalized to the initial unreconnected flux~$\psi_0$) diminishing over time as reconnection occurs. Horizontal dashed lines mark the levels $\psi=0.9\psi_0$, $0.8\psi_0$, and $0.7\psi_0$;
the dimensionless reconnection rates averaged over each of these $0.1\psi_0$ drops are $\beta_{\rm rec}=$0.03 and~0.02
(cf.~\S\ref{sec:unreconnectedFlux}),
and dashed lines with those slopes are drawn.
The middle panel shows
the energy in various components versus time (normalized to
$\int dV \, B_0^2/8\upi$), and the right panel shows the changes in
those components from their initial values.
Energy components shown are
total energy ($U_{\rm total}$, dash-dotted black line),
magnetic ($U_B$, solid blue),
plasma/particle ($U_{\rm plasma}$, dashed green), and
electric ($U_E$, dotted red---in the middle panel, $U_E$ is magnified by a factor of 20).
}
\end{figure}
The behaviour of magnetic flux and energy is qualitatively similar to that observed in other 2D regimes,
such as $\sigma_h\gg 1$ \citep[e.g.,][]{Guo_etal-2014,Werner_etal-2016,Werner_Uzdensky-2017,Werner_etal-2018}.
Figure~\ref{fig:energyFluxVsTime2D} shows unreconnected flux versus time
as well as the energy (and energy changes) versus time in various components, including the energy in magnetic fields and particles.
The unreconnected (upstream) magnetic flux $\psi(t)$
decreases over the course of reconnection, finally down to about 55~per~cent of its initial value (i.e., reconnecting 45~per~cent of upstream flux).
(We note that these values are specific to this representative simulation; reconnection with other parameters may yield different values while being qualitatively similar.)
During this time, the magnetic energy falls to about 70~per~cent of its
initial value, with the 30~per~cent loss converted almost entirely to plasma energy.
This occurs within a time of roughly 8--12$\:L_x/c$ over which the rate of reconnection continually slows (in this closed system).
The dimensionless reconnection rates (cf.~\S\ref{sec:unreconnectedFlux}), averaged
over the time it takes for unreconnected flux $\psi(t)$ to fall from~$0.9\psi_0$ to~$0.8\psi_0$, and from~$0.8\psi_0$ to~$0.7\psi_0$
\citep[as measured in ][]{Werner_etal-2018}, are
$\beta_{\rm rec}\equiv -(cB_0 L_z)^{-1} \dot{\psi}=0.032$ and~$0.22$, respectively (here the dimensionless rate is normalized to $B_0 c$).
Normalized to $B_0 v_A$ instead of $B_0 c$, where
$B_0 v_A=B_0 c\sqrt{\sigma_h/(1+\sigma_h)}=B_0 c/\sqrt{2}$, the
reconnection rates over these same intervals are
$(c/v_A)\beta_{\rm rec}=0.045$ and~$0.031$.
The normalized-to-$B_0 v_A$ reconnection rate around 0.03--0.05 is on the order of, but less than the oft-cited nominal value of 0.1 for fast reconnection \citep{Cassak_etal-2017}---a value that is realized for ultrarelativistic reconnection with $\sigma_h \gg 1$.
By way of comparison, we note that, with the same aspect ratio $L_y/L_x=2$, a similar setup with $\sigma_h\gg 1$
depletes roughly 40~per~cent of magnetic field energy
\citep{Werner_Uzdensky-2017} and reconnects 55~per~cent of the initial flux
\citep{Werner_etal-2018}.
Energy conversion in high-$\sigma_h$ reconnection occurs within
about 3$\:L_x/c$, with normalized reconnection rates around 0.10--0.12.
An important aspect of 2D reconnection is that the upstream field lines are indeed broken and reconnected in the following precise sense:
the decrease in upstream flux equals the increase in flux around the O-points in magnetic islands (plasmoids).
In 2D simulations we can measure these fluxes accurately using the $z$-component of the magnetic vector potential (see~\S\ref{sec:unreconnectedFlux}), and we find that the sum of these fluxes---which is precisely equivalent to the total flux between the major O-points in each layer---is conserved to better than~1~per~cent.
This is not surprising; the only way the total flux can change is (by Faraday's law) if $E_z\neq 0$ at the major O-points
(e.g., $E_z$ could be non-zero due to resistivity, which would allow annihilation of magnetic flux via magnetic diffusion, but typically the rate of annihilation is very slow, especially in a collisionless plasma).
The reconnection rate measurements quoted in this paper are defined in terms of an upstream field of $B_0$, although in these closed systems, the actual
(far) upstream field diminishes over time (although not very much for large systems).
For simulations presented in this paper, the upstream magnetic field decays only to $\approx 0.9B_0$,
which affects the calculation less than the measurement
uncertainty.
There are also questions about whether ``the upstream field'' should be far upstream or at some upstream point closer to the current sheet \citep{Liu_etal-2017,Cassak_etal-2017}, and the difference between far upstream and near upstream might potentially depend on $\sigma_h$.
However, we must leave a better understanding of the precise $\sigma_h$-dependence of the 2D reconnection rate to future work.
When $\sigma_h$ is not large, particles cannot gain much energy \emph{on average} during reconnection.
E.g., for $\sigma_h=1$ and 30~per~cent of magnetic energy depleted, the average particle energy (averaged over the entire simulation) increases by a relative fraction of only $\Delta \gamma/\gamma_b = 0.30 (2/3) \sigma_h = 0.20$
(where $\gamma_b = 3\theta_b = 0.75\sigma$ is the initial average Lorentz factor).
Nevertheless, \emph{some} particles reach energies well above the average energy, $\bar{\gamma} \approx 0.7$--$0.9\sigma$,
as shown by the particle energy spectra in Fig.~\ref{fig:spectra2D}.
Reconnection in the $\sigma_h=1$ regime clearly generates NTPA;
a high-energy nonthermal power-law $f(\gamma)\sim \gamma^{-p}$
becomes apparent within a time of 1--2$L_x/c$, with
approximately the same power-law index $p\approx 4$ as at later times;
the power law extends up to a cutoff energy exceeding $4\sigma$ \citep[cf.][]{Werner_etal-2016}.
The power-law index of $p\approx 4$ is quite steep; to distinguish small changes
(when comparing other simulations), we will usually graph
$\gamma^4 f(\gamma)$, so that if $p$ were exactly 4, the graph would be
a horizontal line.
The high-energy cutoff of the power law has been observed, for simulations with high $\sigma_h\gg 1$ and very large $L_x$, to grow rapidly to an energy $\sim 4\sigma$ \citep{Werner_etal-2016}, and then to continue to grow slowly (sublinearly) in time \citep{Petropoulou_Sironi-2018,Hakobyan_etal-2021}; our results are consistent with this.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig02a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig02b.pdf}%
}
\caption{ \label{fig:spectra2D}
The particle energy spectra $f(\gamma)$ at logarithmically-spaced times, compensated by $\gamma^2$ to
show the relative amount of energy stored in different energy ranges (left)
and by $\gamma^4$ to enhance differences (right). Here particle energy is $\gamma m_e c^2$.
The average particle energy $\bar{\gamma}$ is indicated at $t=0$ and $tc/L_x=9.3$. A power law $\gamma^{-4.4}$ is also shown (in the right panel, $f\sim \gamma^{-4}$ would be horizontal).
}
\end{figure}
Importantly, the high-energy cutoff measured in our simulations is far below the extreme particle acceleration limit for this simulation size and duration---i.e., below the energy a particle could gain in the nominal reconnection electric field of $E_{\rm rec}$ over the simulation duration $T$ (which is usually proportional to simulation size $L_x$).
Such an ``extremely-accelerated'' particle would gain maximum energy
$m_ec^2 \Delta \gamma_{\rm ext} = eE_{\rm rec}cT$.
We can estimate $\Delta \gamma_{\rm ext}$ using $\beta_{\rm rec}$ to relate $E_{\rm rec}$ and $B_0$, i.e.,
$E_{\rm rec}=\beta_{\rm rec}B_0$.
Since $m_e c^2/eB_0\equiv \rho_0$,
$\Delta \gamma_{\rm ext} \sim \beta_{\rm rec} cT/\rho_0
\sim 0.02 (cT/L_x) L_x/\rho_0$.
In our simulations, $\beta_{\rm rec}$ decreases over time (as reconnection slows to a stop), but above we measured $\beta_{\rm rec}\approx 0.02$ in the middle of active reconnection; and in this simulation, $L_x=1280\sigma\rho_0$,
so $\Delta \gamma_{\rm ext} \sim 26 (cT/L_x)\sigma $.
Thus, after $T\sim 10L_x/c$, the maximum energy gain would be around
$\Delta \gamma_{\rm ext}\sim 250 \sigma$.
However, rather than estimate $\beta_{\rm rec}$ and the active reconnection time $T$, it is convenient to consider that by Faraday's law,
$\int_0^T E_{\rm rec} dt = -\Delta \psi/(cL_z)$ where $\Delta \psi$ is the change in unreconnected upstream flux; in a 2D simulation, the simulation size $L_z$ in the unsimulated third dimension is arbitrary, but the upstream flux is also proportional to $L_z$---the flux upstream of the initial current sheet is $\psi_0 \approx B_0 L_z L_y/4$.
Therefore, $\Delta \gamma_{\rm ext} \approx (L_y/4 L_x) (\Delta \psi/\psi_0) (eB_0 L_x/m_e c^2) = 0.5(0.45)(L_x/\rho_0)$, or
$\Delta \gamma_{\rm ext}\approx 0.2(1280\sigma)\approx 250\sigma$.
In this simulation, we observe $\Delta \gamma \lesssim 40\sigma$, far below the extreme acceleration limit, $\Delta \gamma_{\rm ext}\approx 250 \sigma$.
However, some particles undergo extreme acceleration up to at least $\gamma \sim 4\sigma$ (but not beyond $\gamma \approx 20\sigma$) in a time consistent with constant acceleration in $E_{\rm rec}$.
We can infer from Fig.~\ref{fig:spectra2D} that a tiny fraction of particles reach nearly $\gamma\approx 20\sigma$ by time $0.44L_x/c$.
At this time, $\Delta \psi/\psi_0 = 0.039$, so we estimate
$\gamma_{\rm ext} = 0.5(0.039)(1280\sigma)=25\sigma$, close to the observed $\gamma \approx 20\sigma$ (which we emphasize is not the power-law cutoff, but the maximum energy gained by any particle in the simulation). This suggests that particles are in fact ``extremely-accelerated'' up to some energy less than but on the order of $\gamma \approx 20\sigma$. However, no particle ever exceeds $\gamma \approx 40$ significantly, even after much longer times.
We note that: (1) we measured $\Delta \psi$ with a time cadence of $0.44 L_x/c$, so we cannot examine extreme acceleration on smaller timescales, and (2) ideally we would prefer to know about extreme acceleration in the middle of the simulation rather than at the beginning, but we present this result for the very beginning because it is the only time that we can know that a particle with $\gamma\approx 20$ accelerated to that energy within $0.44L_x/c$.
Thus our results are consistent with particles experiencing extreme acceleration up to an energy around $\gamma \sim 4\sigma$ \citep{Werner_etal-2016}, and experiencing slower acceleration thereafter \citep[as described by][]{Petropoulou_Sironi-2018,Hakobyan_etal-2021}.
Incidentally, we also note that all particle gyroradii are much smaller than the system size. A particle with energy $\gamma$ has gyroradius $\sim \gamma\rho_0$ (in field $B_0$), and so a particle would need $\gamma=640 \sigma$ for its gyroradius to equal $L_x/2$; therefore, the system size is much larger than the gyro-orbits of even the most energetic particles.
In the rest of this paper, we will be comparing the flux reconnected versus time, as well as the conversion of magnetic to plasma energy and the resulting particle energy spectra, for reconnection simulations with different initial parameters (including 3D simulations with different lengths~$L_z$).
In the following subsections, we consider 2D simulations all with the same upstream plasma conditions as in this subsection, but with different initial current sheet configurations (\S\ref{sec:pertAndEta2d}), and different system sizes (\S\ref{sec:Lx2d}); then, we consider the addition of different guide field strengths (\S\ref{sec:Bz2d}).
We will find that for large systems, the initial current sheet configuration and system size $L_x$ have relatively weak effects on reconnection, including on energy conversion and NTPA; in contrast, (sufficiently strong) guide magnetic field will significantly slow energy conversion and inhibit NTPA.
\subsection{2D reconnection: dependence on initial current sheet configuration}
\label{sec:pertAndEta2d}
In this subsection we show that 2D reconnection
behaviour is
largely (but not completely) determined by the background (upstream)
plasma and not the configuration of the initial current sheet.
We will compare results from 2D simulations with varying initial magnetic perturbation strength $a$, and varying initial current sheet parameters (i.e., density $n_{d0}$ or, equivalently, $\eta=n_{d0}/n_{b0}$, temperature $\theta_d$, drift speed $\beta_d c$, and half-thickness $\delta$).
All simulations in this subsection have identical background plasma (i.e., $\sigma=10^4$, $\sigma_h=1$, hence upstream $\beta_{\rm plasma}=0.5$,
and $B_{gz}=0$, as well as $L_y/L_x=2$).
The initial magnetic field perturbation $a$ [cf. Eq.~(\ref{eq:pert})], can be varied independently of all other parameters, as in the following substudy (2D-a); however, increasing $a$ shifts the system away from the Harris equilibrium.
In contrast, we always maintain the initial equilibrium when altering $\eta$, $\theta_d$, $\beta_d$, and $\delta$.
The equilibrium satisfies pressure balance, or $\eta \theta_d / \gamma_d = \sigma/2$, and Ampere's law, which requires $\delta = \sigma \rho_0 / (\eta \beta_d)$.
Thus, the initial current sheet, though fully specified through four parameters, has only two independent degrees of freedom in this study; usually we specify $\eta$ and $\beta_d$.
For reference we have listed the relevant parameters
corresponding to different values of $\eta$ and $\beta_d$ in table~\ref{tab:etaEffect}.
\begin{table}
\centering
\begin{tabular}{l|rrrrrc|@{\hspace{0.05in}}crrrrr}
$\eta=n_{d0}/n_{b0}$ & 0.5 & {\bf 1} & 3.1& {\bf 5} & 10
&& 0.33 & 0.5 & 1 & 2 & 4 \\
$\beta_d$ & \multicolumn{5}{c}{--- 0.3 ---}
&& 0.9 & 0.6 & 0.3 & 0.15 & 0.075 \\
\hline
$\theta_d/\theta_b$ & 4.2 & 2.1 & 0.68 & 0.42 & 0.21
&& 14 & 5.0 & 2.1 & 1.01 & 0.50 \\
$\rho_d/\sigma \rho_0$ & 3.1 & 1.6 & 0.5 &0.31 &0.16
&& 10 & 3.8 & 1.6 & 0.76 & 0.38 \\
$d_{e,d}/\sigma \rho_0$ & 2.6 & 1.2 & 0.40 &0.24 &0.12
&& 5.5 & 2.8 & 1.2 & 0.62 & 0.31 \\
$\delta/\sigma \rho_0$ & 6.7 & 3.3 & 1.1 &0.67 & 0.33
&& \multicolumn{5}{c}{--- 3.3 ---} \\
$\delta/\rho_d$ & \multicolumn{5}{c}{--- 2.1 ---}
&& 0.32 & 0.89 & 2.1 & 4.4 & 8.9 \\
$\delta/d_{e,d}$ & \multicolumn{5}{c}{--- 2.7 ---}
&& 0.58 & 1.2 & 2.7 & 5.4 & 11 \\
\end{tabular}
\caption{\label{tab:etaEffect}
Initial temperature and length scales of the drifting (current sheet) plasma resulting from various combinations of initial peak current sheet density~$n_{d0}=\eta n_{b0}$ and drift speed~$\beta_d c$ used in this study.
Listed are (normalized) values of
the initial temperature $\theta_d m_e c^2$,
the average gyroradius $\rho_d=3\theta_d\rho_0$, and the skin depth of the drifting plasma forming the current sheet, $d_{e,d}=\sqrt{3}\lambda_{Dd}=[3\theta_d m_e c^2/(4\upi n e^2)]^{1/2}$,
as well as the initial sheet half-thickness~$\delta$ (relative to the fiducial upstream plasma length scale $\sigma \rho_0$ and drifting plasma length scales~$\rho_d$ and~$d_{e,d}$).
We note that
$\theta_d/\theta_b=2\gamma_d \sigma_h/\eta$
and $\eta \delta = \sigma\rho_0/\beta_d$.
For comparison, the background plasma gyroradius is $\rho_b=0.75 \sigma\rho_0$ and the skin depth $d_e=\sqrt{3}\lambda_D=0.87 \sigma\rho_0$; the cell size is $\Delta x =0.33 \sigma\rho_0$.
Boldface entries indicate the two configurations used to study the effect of varying perturbation strength~$a$.
}
\end{table}
It is useful to compare the sheet half-thickness $\delta$ with the characteristic gyroradii of the upstream plasma and the drifting (initial current sheet) plasma.
The characteristic gyroradius of the upstream plasma is $\rho_b=3\theta_b\rho_0=(3/4)\sigma\rho_0\sim\sigma\rho_0$; the average gyroradius of the drifting plasma is $\rho_d=3\theta_d\rho_0=3\gamma_d/(2\eta)$.
The ratio of sheet thickness to upstream plasma scales is $\delta/\sigma\rho_0 = 1/(\eta\beta_d)$, and the ratio to drifting plasma scales is $\delta/\rho_d=2/(3\gamma_d\beta_d)$.
In substudy (2D-b), below, we
vary $\eta$ while keeping $\beta_d$ constant, which changes $\delta$ with respect to upstream scales but not with respect to drifting plasma scales. On the other hand, in substudy (2D-c), we vary $\beta_d\propto \eta^{-1}$, changing $\delta/\rho_d$ but leaving $\delta/\sigma\rho_0$ constant.
\emph{(2D-a) Varying the initial perturbation.}
For $a=0$, the magnetic field lines are initially entirely parallel to $x$.
Increasing $a$ perturbs the field so that there is a magnetic island and an X-point in the layer.
As described in~\S\ref{sec:setup}, when $a=1$, the initial island half-height $s$ (the maximum height of the separatrix above the initial midplane) equals the current sheet half-thickness, $s=\delta$; for $a\lesssim 1$, the separatrix that bounds the island is contained within the initial layer.
We investigate the effect of varying $a$ over a wide range, for
two different initial current sheets: the first [$\beta_d=0.3$, $\eta=5$, $\delta=(2/3)\sigma\rho_0$, $\theta_d=0.10\sigma$] was chosen to be like most simulations in this paper, marginally resolving the initial sheet with $\delta/\Delta x=2$;
the second [$\beta_d=0.3$, $\eta=1$, $\delta=(10/3)\sigma\rho_0$, $\theta_d=0.52\sigma$] was chosen to resolve the current sheet with $\delta/\Delta x=10$.
In both cases we used modest system sizes with $L_x=320\sigma\rho_0$.
To help distinguish between the effect of $a$ and stochastic variation,
we ran two simulations for each $a$, identical except for random
initial particle velocities (the source of randomness is discussed later in detail in~\S\ref{sec:variability3d}).
For the denser, thinner sheet ($\beta_d=0.3$, $\eta=5$, used in most simulations here),
we ran simulations with
$a \in \{$0, 0.22, 0.55, 1.1, 2.2, 5.5, 11, 22, 55, 110$\}$, corresponding to $s/\delta \in \{$0, 0.44, 0.72, 1.1, 1.6, 3.1, 5.4, 10, 23, 43$\}$, respectively.
For $a < 0.3$, we note that $s < \Delta x$, so the simulation can hardly tell the difference between
$a=0$ and $a=0.22$;
and for $a=110$, the separatrix extends out to $s=43\delta=29 \sigma\rho_0 \approx 0.1L_x$.
In Fig.~\ref{fig:perturb2dLx320}(a) we observe that the evolution of
magnetic energy
versus time is nearly the same for the two simulations with $a < 0.3$; it is also
the same, up to stochastic variation, within the group of simulations with
$0.55 \lesssim a \lesssim 22$.
The simulations with $a<0.3$ exhibit slightly slower magnetic energy conversion for $1 \lesssim tc/L_x \lesssim 4L_x/c$ than the simulations with $0.55 \lesssim a \lesssim 22$, but by $t\gtrsim 5L_x/c$ they have all converted the same amount of energy, so while noticeable, this is a very minor difference.
For $a\gtrsim 55$ (very large perturbations),
we observe an increasingly quick onset of reconnection and more rapid
magnetic depletion.
The particle energy spectra in Fig.~\ref{fig:perturb2dLx320}(b) are
all very
similar [considering that differences are enhanced by
showing $\gamma^4 f(\gamma)$ instead of $f(\gamma)$],
although very large perturbations result in an almost
imperceptibly harder spectrum with a slightly smaller fraction of
particles reaching the very highest energies.
We conclude that for all but very large perturbations, the initial perturbation
does not have a very substantial effect on energy conversion or NTPA.
For the less dense, initially-thick current sheet, [$\beta_d=0.3$, $\eta=1$, $\delta=(10/3)\sigma\rho_0$, $\theta_d=0.52\sigma$],
we explored $a \in \{$0, 0.04, 0.11, 0.22, 0.44, 1.1, 2.2, 4.4, 11, 22$\}$,
corresponding to separatrix heights $s/\delta\in \{$0, 0.2, 0.31, 0.44, 0.63, 1, 1.6, 2.5, 5.1, 9$\}$.
The results are shown in Fig.~\ref{fig:perturb2dLx320}(c,d).
Here, $s>\Delta x$ as long as $a > 0.011$ (i.e., for all the nonzero $a$ used); for $a=22$, the separatrix extends to $s=9\delta = 30\sigma\rho_0\approx 0.1L_x$.
The less dense, thicker sheet
leads to a slower onset of reconnection; increasing
the initial perturbation strength tends to hasten reconnection onset.
However, once reconnection is triggered, the energy evolves very similarly for zero and large perturbations, as seen in
Fig.~\ref{fig:perturb2dLx320}(c), which shows magnetic energy versus time,
shifted relative to $t_{\rm onset}$,
the time at which the magnetic energy has fallen by 1~per~cent (cf.~\S\ref{sec:onsetTime}).
With this time shift, the magnetic energy curves are almost identical
in all cases except for the largest perturbation, $a=22$.
Not surprisingly, the particle spectra in Fig.~\ref{fig:perturb2dLx320}(d)
are similarly identical (within stochastic variation).
We again conclude that the strength of perturbation
(at least for $a\lesssim 20$) has negligible effect on energy conversion
and NTPA---aside from having a strong effect on the time to reconnection
onset.
\begin{figure}
\centering
\fullplot{
\raisebox{0.315\textwidth}{\large (a)}%
\includegraphics*[width=0.455\textwidth]{fig03a.pdf}%
\hfill
\raisebox{0.315\textwidth}{\large (b)}%
\includegraphics*[width=0.455\textwidth]{fig03b.pdf}\\
\raisebox{0.315\textwidth}{\large (c)}%
\includegraphics*[width=0.455\textwidth]{fig03c.pdf}%
\hfill
\raisebox{0.315\textwidth}{\large (d)}%
\includegraphics*[width=0.455\textwidth]{fig03d.pdf}\\
}
\caption{ \label{fig:perturb2dLx320}
(a) Transverse magnetic energy versus time, and (b) electron energy spectra
$f(\gamma)$, compensated by $\gamma^4$, at times~$t$ when $U_{Bt}(t)=0.75U_{B0}$---marked by the horizontal grey line in (a)---for
different initial magnetic field perturbations in
2D simulations with an initially dense, thin current sheet
($\eta=5$, $\delta/\sigma \rho_0=2/3$).
Panels~(c) and (d) show the same for a less dense, thicker layer
($\eta=1$, $\delta/\sigma \rho_0=10/3$), with spectra at $U_{Bt}(t)=0.77U_{B0}$.
}
\end{figure}
For both overdensities $\eta=1$ and $\eta=5$,
the initial perturbation has significant consequences for the evolution of the plasmoid hierarchy, even though this does not seem to alter the overall energy conversion rate and NTPA.
Without perturbation (Fig.~\ref{fig:ndePerturb2dLx320}, left),
a number of roughly equal-sized plasmoids form at the onset of
reconnection, with sizes determined by the fastest-growing
wavelength of the tearing instability.
These plasmoids undergo the coalescence instability and merge in pairs, resulting in a smaller
number of larger plasmoids,
each of which contains a similar
fraction of the initially-drifting particles that made up the initial
current sheet.
In contrast, even a very small initial perturbation will favour one plasmoid (the
``major'' plasmoid), which is always larger than the others
(Fig.~\ref{fig:ndePerturb2dLx320}, right);
it may contain many or most of the initially-drifting particles.
In this case, subsequent plasmoid mergers tend to be between different-sized plasmoids;
plasmoids form at the major X-point, and move toward the major
plasmoid (with mergers between smaller plasmoids
sometimes occurring before they reach the major plasmoid).
We note that the final state has just a single monster plasmoid, regardless of initial perturbation.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=\textwidth]{fig04a.pdf}%
}
\caption{ \label{fig:ndePerturb2dLx320}
Magnetic field lines superimposed over
the density $n_{de}$ of the initially-drifting electrons that form the initial
current sheet (i.e., excluding background electrons), for cases (left) without an initial magnetic perturbation ($a=0$), and (right) with a small
initial perturbation $a=5.5$ or $s/\delta=3$ [cf. Fig.~\ref{fig:perturb2dLx320}(a)],
at times $t=0.44L_x/c$ (top) and $t=1.2L_x/c$ (bottom).
In both cases, the drifting particles are well contained in plasmoid cores.
With zero perturbation (left), the first plasmoids formed are roughly equal in size, and they generally merge in pairs so that at any time, there is some number of equal-sized largest plasmoids, each containing a similar fraction of the drift particles.
In contrast, a small
perturbation (right) favours one plasmoid, which remains always larger than all other plasmoids, and rapidly gathers most and then all of the drifting particles.
Magnetic field lines in the layer (i.e., within the separatrix) are solid green, with the separatrix being solid cyan; upstream of the separatrix, field lines are dashed blue.
These simulations have size $L_x=320\sigma \rho_0$, with
an initially thin current sheet, $\delta=(2/3) \sigma \rho_0$.
}
\end{figure}
Summarizing (2D-a), we see that (for $L_x\gtrsim 320\sigma\rho_0$) the
initial perturbation (if not terribly large, i.e., $a\lesssim 20$, or
$s/\delta \lesssim 9$)
does not significantly alter global energy evolution (after reconnection
onset) or NTPA.
However, the initial perturbation can reduce the time it takes reconnection to start, and
it can also have a significant effect on the nature of plasmoid
formation and subsequent evolution.
Similar differences in plasmoid evolution caused by artificial triggering with a locally reduced pressure have been previously studied in 2D semirelativistic electron-ion reconnection \citep{Ball_etal-2019}; there NTPA was found to be similarly insensitive to triggering, at least for weak guide field.
\emph{(2D-b) Varying current sheet density $\eta$ with fixed $\beta_d=0.3$; i.e., varying $\delta/\sigma\rho_0$ for fixed $\delta/\rho_d=2.1$}.
We now look at the effects of different initial current sheets, keeping $\beta_d=0.3$ constant and varying $\eta \in \{$0.5, 1, 3.1, 5, 10$\}$
(considering $\eta < 1$ is unusual in reconnection research, but we include $\eta=0.5$ to show that it is not that different from $\eta=1$, although
it does start to indicate a trend for the limit $\eta \rightarrow 0$).
Table~\ref{tab:etaEffect} shows how other quantities such as $\delta$ vary as $\eta$ is varied in this study.
Increasing $\eta$ directly increases the current sheet density $n_{d0}$ (relative to the upstream plasma density $n_{b0}$, which remains the same).
As $\eta$ increases, the temperature $\theta_d$ decreases to maintain pressure balance across the current sheet against the upstream magnetic field, $\theta_d = \gamma_d \sigma / (2\eta) \approx 0.52 \sigma/\eta$; this reduces the fundamental length scales (and hence also time scales) of the current sheet.
(We note, for context, that if background plasma with $n_{b0}$ and $\theta_b$ were
adiabatically compressed with adiabatic index 4/3, appropriate for
relativistically-hot plasma, to a density $n_{\rm ad}$ and temperature
$\theta_{\rm ad}$
sufficient to balance the upstream magnetic pressure $B_0^2/8\upi$ as well as the upstream plasma pressure $n_{b0}\theta_b m_ec^2$, this would yield $\eta_{\rm ad}\equiv n_{\rm ad}/n_{b0}=2.28$ and
$\theta_{\rm ad}/\theta_b=1.3$.)
As $\eta$ varies (with fixed $\beta_d$), we have $\theta_d \propto \eta^{-1}$ and so $\delta$ as well as the fundamental length scales of the current sheet plasma (e.g., both the gyroradius $\rho_d$ and collisionless skin depth $d_{ed}$) all scale $\propto \eta^{-1}$ (as long as $\gamma_d \approx 1$).
Therefore, $\delta/\rho_d=2.1$ and $\delta/d_{ed}=2.7$ remain constant (see table~\ref{tab:etaEffect}), although the current sheet thickness changes with respect to upstream plasma scales (e.g., $\rho_b=0.75\sigma\rho_0$) as well as $L_x$ and $\Delta x$.
For this investigation we use a larger system size, $L_x=1280\sigma\rho_0$,
so that $L_x,L_y \gg \delta$ for all $\delta$ explored; other fixed parameters are: $\beta_d=0.3$ and $B_{gz}=0$.
We used the exact same initial magnetic field $\boldsymbol{B}(x,y)$ in all cases, with the same small
perturbation: $a \delta /L_x=0.0052$ --- the dependence $a\sim \delta^{-1}$ ensures that the initial field is independent of $\delta$---i.e., $s/L_x$ is fixed [cf. Eq.~(\ref{eq:pert})].
Based on (2D-a) above, we expect our conclusions for this subsection would be the same for other small perturbations, including zero perturbation.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig05a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig05b.pdf}
}
\caption{ \label{fig:eta2dLx1280}
For simulations with varying $\eta$ (hence varying initial
current sheet half-thicknesses $\delta$;
cf.~table~\ref{tab:etaEffect}):
(left) transverse magnetic energy versus time (shifted by $t_{\rm onset}$; cf.~\S\ref{sec:onsetTime}),
and (right) electron energy
spectra times $t$ when $U_{Bt}(t)=0.75 U_{B0}$ (indicated by the horizontal grey line in the left panel). Dotted grey-on-yellow lines indicate power-law slopes of 4.1 and 4.8.
All simulations have size
$L_x=1280\sigma \rho_0$.
}
\end{figure}
Figure~\ref{fig:eta2dLx1280} shows (left) the magnetic energy evolution and
(right) NTPA spectra for simulations
with a range
of $\eta=0.5$--$10$, corresponding to $\delta/\sigma\rho_0=6.7$--$0.33$.
We first note, as above, that reconnection onset is delayed by less dense, thicker
current sheets, but that once magnetic energy depletion begins,
energy evolution and NTPA are very similar and almost independent of~$\eta$.
However, closer examination shows a tentative trend of decreasing
NTPA efficiency with decreasing $\eta$; this trend is very small for $\eta >1$, but becomes noticeable for
the $\eta =0.5$ case (with $\theta_d/\theta_b=4.2$ and
$\delta\approx 7\sigma\rho_0$), which reconnects a little slower and generates slightly less efficient NTPA.
Specifically, $\eta =0.5$ yields a power law $p=4.6\pm 0.1$, whereas denser layers with $\eta \geq 1$ yield $p=4.4\pm 0.1$.
Also, denser initial sheets seem to lead to slightly more overall
magnetic energy depletion.
This weak $\eta$-dependence
hints at a more severe behaviour in the limit of underdense current sheets: in a simulation (not shown here) with $\eta=0.1$ and $\delta=33\sigma\rho_0$, reconnection did not even start, despite a simulation duration longer than $10L_x/c$---hence no magnetic energy was depleted.
We thus conclude that the initial current sheet density has a negligible effect on total energy conversion and NTPA efficiency as long as $\eta \gg 0.1$; however, for small $\eta$, NTPA efficiency and total energy conversion decrease as $\eta$ decreases, and the thick, low-density current sheet may eventually become effectively stable against tearing and reconnection.
\emph{(2D-c) Varying $\beta_d\propto \eta^{-1}$; i.e., varying $\delta/\rho_d$ for fixed $\delta/\sigma\rho_0=(10/3)\sigma\rho_0$.}
The motivation for this particular parameter space exploration
is a little complicated.
Essentially, we would like to study the influence of $\beta_d$ on reconnection.
After substudy (2D-b), which varied $\eta$ but not $\beta_d$, it would be natural to hold $\eta$ constant while varying $\beta_d$.
However, we will instead vary $\eta \propto \beta_d^{-1}$ to keep $\eta\beta_d$ constant.
Importantly, this fixes $\delta=\sigma\rho_0/(\eta\beta_d)$ with respect to upstream plasma scales.
In particular, $\delta/\Delta x=10$ and $L_x/\delta=384$ will be the same in every case, so that, as $\beta_d$ and $\eta$ vary, the initial current sheet will remain well-resolved but very thin compared with $L_y=2L_x$.
Satisfying these two criteria will be important in 3D (\S\ref{sec:pertAndEta3d}--\ref{sec:RDKIamp}) when we will focus on the evolution of the current layer at early times as it undergoes large deformations due to RDKI, requiring $\delta/\Delta x\gg 1$ and $\delta/L_y \ll 1$; in addition, it will be more straightforward to compare layer deformation and its interaction with the upstream plasma for different simulations when the current sheets begin with the same thickness.
Although neither RDKI nor extreme layer deformation will occur in 2D, this substudy provides a baseline against which to compare the 3D simulations.
Therefore, we simultaneously vary $\eta$ along with $\beta_d$ to keep the product $\eta\beta_d$ constant.
Maintaining constant $\eta\beta_d$ fixes $\delta/\sigma \rho_0$, while essentially varying the drifting plasma length scales [such as gyroradius $\rho_d=3\theta_d\rho_0$] with respect to $\delta$---i.e., varying $\delta/\rho_d = 2/(\gamma_d \beta_d)$.
We compare five configurations with the same background plasma (with zero guide field), system size $L_x=1280\sigma\rho_0=384\delta$, and zero initial magnetic perturbation ($a=0$).
We vary $\beta_d\in \{$0.075, 0.15, 0.3, 0.6, 0.9$\}$; to keep $\eta \beta_d$ constant, the overdensities $\eta$ are $\{$4, 2, 1, 0.5, 0.33$\}$, yielding the same $\delta=(10/3)\sigma \rho_0=10\Delta x$ for all five cases.
Other parameters are (see also table~\ref{tab:etaEffect}): $\theta_d/\theta_b \in \{0.5$, 1.0, 2.1, 5.0, 14$\}$ and $\delta/\rho_d \in \{8.9$, 4.4, 2.1, 0.89, 0.32$\}$, respectively.
At $\beta_d=0.075$, $\eta=4$, the gyroradius and collisionless skin depth of the drifting plasma are at their smallest values of $\rho_d=0.38\sigma\rho_0$ and $d_{ed}=0.31\sigma\rho_0$, both near the grid resolution $\Delta x$, and the half-thickness $\delta$ is 10 times larger than those scales.
At the other end of the parameter scan, $\beta_d=0.9$, $\eta=0.33$, the microphysical scales of the initial current sheet become somewhat larger than $\delta$: $\delta \approx 0.3 \rho_d \approx 0.6 d_{ed}$.
To characterize the effects of varying $\beta_d$ (with $\eta$), we focus on the time history of magnetic and plasma energy, and on the resulting NTPA; and again, we find little variation among this set of simulations
(see Fig.~\ref{fig:betad2dLx1280}).
All cases show very similar magnetic depletion rates and NTPA (within stochastic variation), except for the case $\beta_d=0.9$ ($\eta=0.33$), which is a little different; this case ends up with a bit more energy in plasmoids, although the rate at which \emph{upstream} magnetic energy and unreconnected flux decrease is similar to the others.
The $\beta_d=0.9$ case has significantly hotter (though lower density) drift particles (cf.~table~\ref{tab:etaEffect}), which results in correspondingly hotter and less dense plasmoids cores, since the initially-drifting particles, being close to the initial midplane, are among the first to be swept into plasmoids and remain near the plasmoid cores (the original drift kinetic energy may also be converted to thermal energy, contributing to even hotter cores).
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig06a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig06b.pdf}
}
\caption{ \label{fig:betad2dLx1280}
(Left) Transverse magnetic energy versus (shifted) time,
and (right) electron energy
spectra when 23~per~cent of the initial magnetic energy has been converted (see the horizontal grey line in the left panel), for simulations with varying initial current drift $\beta_d$, but constant initial sheet thickness $\delta=(10/3)\sigma\rho_0$. (Table~\ref{tab:etaEffect} shows how initial temperatures and gyroradii vary with $\beta_d$.)
Dotted grey-on-yellow lines indicate power-law slopes of~4.2 and~4.6.
All simulations have size
$L_x=1280\sigma \rho_0$.
}
\end{figure}
\emph{Summary of 2D-a,b,c.}
We have seen that in 2D, the details of the initial current sheet can affect the time delay before the onset of reconnection, as well as the evolution of the plasmoid hierarchy. Nevertheless, the initial current sheet configuration has relatively little effect on the long-term magnetic energy conversion and NTPA in 2D reconnection; after a short initial stage, reconnection energetics are dominated by the upstream plasma, regardless of the very early current sheet evolution.
This conclusion may seem to be at odds with that from \citet{Ball_etal-2019}, which found subtle but significant differences in NTPA due to initial perturbation and varying initial sheet thickness;
however, \citet{Ball_etal-2019} studied transrelativistic electron-proton reconnection, with low upstream plasma beta and non-zero guide magnetic field ($B_{gz}/B_0=$0.1 and 0.3).
They used a small pressure imbalance as an initial perturbation, and found that its influence diminished with diminishing guide field, potentially consistent with our zero-guide-field results here.
The impact of varying sheet thickness (keeping $\eta$ constant) was also seen at the larger guide field, $B_{gz}=0.3B_0$, so there are several potentially important differences between these two studies: transrelativistic electron-proton versus ultrarelativistic pairs, low upstream plasma beta versus beta of 0.5 ($\sigma_h=1$), weak-to-moderate versus zero guide field, and pressure versus magnetic field perturbation.
\subsection{2D reconnection: system-size dependence}
\label{sec:Lx2d}
In this subsection we investigate how reconnection depends on system size~$L_x$, to estimate the applicability of our kinetic simulations to astrophysical systems, which may be much larger than even the largest supercomputers can possibly simulate while including full kinetics.
We present results for system sizes ranging over
a factor of 30, from $L_x/\sigma \rho_0=80$ to~2560.
These simulations achieve a substantial scale separation between the system size~$L_x$ and all the microphysical scales, which are of order $\sigma\rho_0$ (cf.~\S\ref{sec:setup}).
Such large scale separations are feasible because, for $\sigma_h=1$, all the kinetic scales are comparable, hence they can all be resolved by a cell size~$\Delta x$ only marginally smaller than $\sigma\rho_0$.
All simulations presented in this subsection are 2D with zero guide field,
starting with an initial current sheet half-thickness
$\delta=(2/3)\sigma\rho_0$, $\eta=5$, $\beta_d=0.3$, and zero perturbation.
Before discussing how~$L_x$ affects typical reconnection behaviour, we note that stochastic evolution of the plasmoid hierarchy affects energy conversion and NTPA, especially over short timescales.
For example, in the $L_x=1280\sigma\rho_0$ simulation, the magnetic energy
depletion slows (compared with earlier times) between~$t\approx 3L_x/c$ and~$t \approx 5L_x/c$, so that $U_{Bt}(t)$ is distinctly higher at $t=5L_x/c$ than in all the other simulations (Fig.~\ref{fig:energyFluxVsLx2d} left, inset).
At~$t\approx 3L_x/c$, the lower layer in the simulation domain for $L_x=1280\sigma\rho_0$ has two large,
roughly equal-sized plasmoids, separated by~$\approx L_x/2$
(while the upper layer already has only one final plasmoid at this time).
In principle, one of the plasmoids could either move left to merge with the
other, or (since the simulation is periodic in~$x$) it could move right.
Thus they reach a ``Buridan's Ass'' metastable state that seems to slow reconnection evolution for a short time.
However, by $4.4L_x/c$, the two plasmoids have
decided which way to move, and by $4.8L_x/c$ they have
started to merge, causing a burst of magnetic energy depletion; by the time the merger is nearly complete at $6.1L_x/c$, $U_{Bt}(t)$ has ``caught up'' to all the other simulations.
Over short times, the reconnection dynamics thus depends on the detailed behaviour of large plasmoids, which has an element of randomness;
however, the random plasmoid behaviour does not alter the long-term behaviour of reconnection.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig07a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig07b.pdf}
}
\caption{ \label{fig:energyFluxVsLx2d}
For different system sizes $L_x$:
(left) transverse magnetic energy versus time, with an inset showing~$L_x=1280\sigma\rho_0$ alone, and (right) unreconnected flux versus time.
The horizontal dashed lines indicate the periods over which the average reconnection rate is given in Fig.~\ref{fig:reconRateAndTotalDissipationVsLx2d},
and slopes corresponding to $\beta_{\rm rec}=0.04$ and~0.02 (cf.~\S\ref{sec:unreconnectedFlux}) are shown as grey dotted lines.
}
\end{figure}
Figure~\ref{fig:energyFluxVsLx2d} shows the evolution of transverse magnetic energy~$U_{Bt}(t)$ and unreconnected flux~$\psi(t)$ for a range of~$L_x$.
Although~$U_{Bt}(t)$ often decreases fitfully, $\psi(t)$ tends to decrease more smoothly.
Initially, $\psi(t)$ falls very rapidly as reconnection is dominated by the initial current sheet, and the rate of fall slows a little as the upstream plasma begins to dominate reconnection (the larger~$L_x/\delta$ is, the sooner this happens relative to the total duration of reconnection).
Eventually, the system reaches a stable magnetic configuration and $\psi(t)$ approaches a constant value;
therefore, the reconnection rate must decrease over time, eventually toward zero (in a closed system).
The reconnection rates over two different time intervals are shown in Fig.~\ref{fig:reconRateAndTotalDissipationVsLx2d}(left), namely for the two intervals $0.9 > \psi(t)/\psi_0 > 0.8$ and $0.8 > \psi(t)/\psi_0 > 0.7$, i.e., between the horizontal dashed lines in Fig.~\ref{fig:energyFluxVsLx2d}(right).
As the system size~$L_x/\sigma\rho_0$ increases from~160 to~2560, $\beta_{\rm rec}$ (cf.~\S\ref{sec:unreconnectedFlux}) falls from~$\approx 0.04$ to~$\approx 0.03$ in the earlier interval, and from~$\beta_{\rm rec}\approx 0.03$ to~$\beta_{\rm rec}\approx 0.02$ in the later interval.
However, $\beta_{\rm rec}$ stabilizes within the uncertainty of measurement by $L_x\gtrsim 640\sigma\rho_0$.
Reconnection continues until approximately~$0.27U_{B0}$ and~$0.42\psi_0$ have been depleted; the fractions of lost energy and flux are practically independent of~$L_x$ (Fig.~\ref{fig:reconRateAndTotalDissipationVsLx2d}, right), although we note that they are specific to the fixed aspect ratio~$L_y/L_x=2$.
The largest simulation, $L_x=2560\sigma\rho_0$ shows slightly less loss in~$U_{Bt}$ and~$\psi$, likely because it had not quite finished reconnecting when it halted around $10L_x/c$ (cf.~Fig.~\ref{fig:energyFluxVsLx2d}).
Importantly, the \emph{upstream} magnetic energy and flux are both depleted by the same amount---42~per~cent.
However, of that~$0.42U_{B0}$ initially associated with magnetic field lines that eventually undergo reconnection, only~$0.27U_{B0}$ is converted to plasma energy; the remaining~$0.15 U_{B0}$ ends up stored in the ``reconnected'' magnetic field around and within plasmoids.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig08a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig08b.pdf}%
}
\caption{ \label{fig:reconRateAndTotalDissipationVsLx2d}
(Left) The
average reconnection rate (normalized to $B_0 c$ on the left scale, and to $B_0 v_A$
on the right scale) versus system size $L_x$, over the period when the unreconnected flux falls from 0.9--0.8$\psi_0$ (blue squares), and also from 0.8-0.7$\psi_0$ (red circles)---i.e., between the horizontal dashed lines in Fig.~\ref{fig:energyFluxVsLx2d}(right).
(Right) The fraction of initial (transverse) magnetic field energy converted to plasma energy (blue diamonds) and of upstream flux depleted (red plusses) during reconnection versus
system size $L_x$ (for aspect ratio $L_y/L_x=2$, with two layers).
The largest system size, $L_x=1280\sigma\rho_0$ had not quite finished reconnection when the simulation halted around $10L_x/c$; it would have depleted slightly more energy and flux if given more time.
}
\end{figure}
Turning our attention to NTPA,
Fig.~\ref{fig:spectraVsLx2d} shows the electron energy spectra~$f(\gamma)$ at $t=6L_x/c$ for simulations with different~$L_x$ [the left panel shows $\gamma^2 f(\gamma)$, and the middle shows $\gamma^4 f(\gamma)$ to enhance small differences].
Aside from the smallest, all simulations have a
clear high-energy nonthermal power-law $f(\gamma)\sim \gamma^{-p}$;
for the largest simulation, it extends over almost a decade in energy.
At~$t=6L_x/c$, the power-law index~$p$ steepens slightly as $L_x/\sigma\rho_0$ increases from~160 to~2560, from~$p\approx 4.0$ to~$p\approx 4.5$ (Fig.~\ref{fig:spectraVsLx2d}, right---cf.~\S\ref{sec:NTPAfit}).
This steepening may be caused by a different acceleration mechanism that operates over long times (hence only in very large simulations), as we will discuss presently.
The time evolution of power-law index~$p$ and high-energy cutoff~$\gamma_c$ (cf.~\S\ref{sec:NTPAfit}), for all the~$L_x$, are shown in Fig.~\ref{fig:cutoffVsLx2d} (left and middle panels).
The initial development of the power law, as the first particles reach high energies, results in~$p(t)$ rapidly falling (i.e., the power law hardens/flattens) to approach an index between~3.9 and~4.5.
From its minimum value, the slope appears to steepen slowly over time, mostly notably for simulations with $L_x\geq 640\sigma\rho_0$, all of which show $p(t)$ rising by about~0.2 from its minimum.
Correspondingly, the cutoff~$\gamma_c(t)$ grows rapidly as the power law develops, attaining after only 1--2$L_x/c$ a value ranging from~$\gamma_c=6\sigma$ for $L_x=80\sigma\rho_0$ to~$\gamma_c=15\sigma$ for $L_x=2560\sigma\rho_0$.
For~$L_x\geq 1280\sigma\rho_0$, $\gamma_c(t)$ continues to grow over time, and even the case $L_x=640\sigma\rho_0$ experiences some limited growth after about~$t=7L_x/c$.
Thus at late time~$t=9L_x/c$, $\gamma_c$ ranges
from~$6\sigma$ to~$33\sigma$ as $L_x$ increases by a factor of 32, consistent with $\gamma_c \sim \sqrt{L_x}$;
importantly, although~$\gamma_c$ increases with~$L_x$, the increase is clearly sublinear.
This is consistent with very rapid (extreme) acceleration up to~$\gamma_c\sim 4\sigma$ \citep{Werner_etal-2016,Uzdensky-2020arxiv}, at which energy particles become trapped in plasmoids and cease extreme acceleration, but undergo much slower acceleration inside plasmoids \citep{Petropoulou_Sironi-2018,Hakobyan_etal-2021,Uzdensky-2020arxiv}.
Interestingly, the slow growth in~$\gamma_c(t)$ seems to correspond to the slow steepening in~$p(t)$, suggesting that the slower acceleration mechanism may yield a steeper power law.
The increase in~$p$ with~$L_x$ seen in Fig.~\ref{fig:spectraVsLx2d}(right) may simply reflect the increasing influence of the slower acceleration mechanism after longer times (accessible in large systems).
Very large reconnection simulations sometimes suffer at late times from slow-growing instabilities.
The largest case, $L_x=2560\sigma\rho_0$, exhibited potentially troubling energy nonconservation, with the total energy growing by almost~1~per~cent between $t=7L_x/c$ and $t=10L_x/c$.
We find that this typically results in unphysical heating of the lowest-energy particles but does not affect the high-energy spectrum.
To rule out this possibility, we compare the simulation with fiducial resolution $\Delta x=\sigma\rho_0/3$ to one with higher resolution, $\Delta x=\sigma\rho_0/4$ in Fig.~\ref{fig:cutoffVsLx2d}(right); both show essentially similar evolution of~$p(t)$ and~$\gamma_c(t)$, although the higher-resolution case suffered only a~0.05~per~cent increase in total energy.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.325\textwidth]{fig09a.pdf}%
\hfill
\includegraphics*[width=0.325\textwidth]{fig09b.pdf}
\hfill
\includegraphics*[width=0.325\textwidth]{fig09c.pdf}
}
\caption{ \label{fig:spectraVsLx2d}
For different system sizes $L_x$:
(left) electron energy spectra $f(\gamma)$, compensated by $\gamma^2$,
at time $t=6L_x/c$, with power-law indices of~4 and~4.4 indicated by dotted grey-on-yellow lines,
(middle) $\gamma^{4} f(\gamma)$ at $t=6L_x/c$, to enhance differences,
(right) the median power-law index during the time period
$5 < tc/L_x < 7$, with ``error'' bars containing 68~per~cent of the values fitted to spectra within this period.
For $L_x/\sigma\rho_0=$256 and~320, results from two simulations are shown (slightly offset in $L_x$ for clarity in the right panel).
}
\end{figure}
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.325\textwidth]{fig10a.pdf}%
\hfill
\includegraphics*[width=0.325\textwidth]{fig10b.pdf}%
\hfill
\includegraphics*[width=0.325\textwidth]{fig10c.pdf}%
}
\caption{ \label{fig:cutoffVsLx2d}
For different system sizes~$L_x$, the fitted values for the
power-law index~$p$ (left) and high-energy cutoff~$\gamma_c$ (middle) versus time.
(Right) For $L_x=2560\sigma\rho_0$, $p(t)$ (blue, left axis) and $\gamma_c(t)$ (red, right axis) are essentially the same for
two different resolutions, $\Delta x/\sigma\rho_0 = 1/3$ and $1/4$,
even though for this large system size, $\Delta x/\sigma\rho_0=1/3$
resulted in unphysical total energy growth of~1~per~cent, while $\Delta x/\sigma\rho_0 =1/4$ saw a more satisfactory~0.05~per~cent increase.
Error bars indicate uncertainty, encompassing the middle 68~per~cent of values measured over the associated time interval.
}
\end{figure}
In summary, the system-size dependence of 2D reconnection with $\sigma_h=1$ is weak but not entirely insignificant for $L_x \gtrsim 160\sigma\rho_0$.
The reconnection rate (normalized to $cB_0$) is $\beta_{\rm rec}\approx 0.03$ or $(c/v_A)\beta_{\rm rec}\approx 0.04$.
During reconnection, some particles are very rapidly accelerated into a power-law energy spectrum $\sim \gamma^{-p}$ with $p\approx 3.9$--$4.4$, extending up to a cutoff around $\gamma_c \gtrsim 6\sigma$.
This rapid acceleration presumably ends when particles become trapped in plasmoids.
Over time, a much slower acceleration mechanism continues to accelerate particles trapped in plasmoids \citep{Petropoulou_Sironi-2018,Hakobyan_etal-2021}, resulting in a slight steepening of~$p$ (up to~$\approx 4.6$ in our largest simulation) and slow but significant growth in $\gamma_c$ (up to~$\approx 33\sigma$ in our largest simulation).
\subsection{2D reconnection: guide magnetic field}
\label{sec:Bz2d}
We now investigate the effect of adding an initial
uniform guide magnetic field $B_{gz}\hat{\boldsymbol{z}}$, considering
$B_{gz}/B_0 \in \{$0, 0.25, 0.5, 0.75, 1, 1.5, 2, 3, 4$\}$.
All simulations presented in this subsection are 2D
with $L_x=1280\sigma\rho_0$,
starting with an initial current sheet half-thickness
$\delta=(2/3)\sigma\rho_0$ ($\eta=5$, $\beta_d=0.3$), and zero perturbation ($a=0$).
A weak guide field $B_{gz} \lesssim 0.25 B_0$ has very little effect on the magnetic energy evolution.
A strong guide field, however, slows reconnection and inhibits overall magnetic
energy conversion, as shown in Fig.~\ref{fig:energyFluxVsBgz2d}.
The left panel shows the transverse magnetic energy $U_{Bt}$ versus time, for simulations with $B_{gz}/B_0$ ranging from~0 to~4;
as $B_{gz}$ increases above $\simeq 0.5B_0$, $U_{Bt}(t)$ falls more
slowly, and ultimately decreases by a smaller amount.
Correspondingly (cf. Fig.~\ref{fig:energyFluxVsBgz2d}, right), the
unreconnected flux decreases more slowly, and less flux is reconnected
overall.
This slowing of reconnection has been previously attributed to
the decrease in the effective component ($x$-projection) of the Alfv\'{e}n velocity
\citep{Liu_etal-2014,Liu_etal-2015,Werner_Uzdensky-2017}.
The upstream Alfv\'{e}n velocity, including~$B_0$ and~$B_{gz}$, is
$v_A=c\sqrt{(B_0^2+B_{gz}^2)/(B_0^2+B_{gz}^2 + 4\upi h)}$,
where $h$ is the plasma enthalpy density ($h=4n_b\theta_b m_e c^2$ in the ultrarelativistic limit).
The projection of $v_A/c$ along~$B_0$ is therefore
\begin{eqnarray} \label{eq:vAx}
\frac{v_{A,x}}{c} &=&
\sqrt{ \frac{B_0^2}{B_0^2 + B_{gz}^2}} \frac{v_A}{c}
=
\sqrt{ \frac{B_0^2}{B_0^2 + B_{gz}^2 + 4\upi h}}
=
\sqrt{ \frac{B_0^2}{B_0^2 + 4\upi h_{\rm eff}}}
\end{eqnarray}
where, following \citet{Werner_Uzdensky-2017}, we define
$h_{\rm eff}\equiv h+B_{gz}^2/4\upi$ to be an effective enthalpy density including a contribution from the guide field.
Then, defining $\sigma_{h,\rm eff}\equiv B_0^2/(4\upi h_{\rm eff})$, we have $v_{A,x} = (\sigma_{h,\rm eff}^{-1}+1)^{-1/2}$, which reduces to
$v_{A}/c = (\sigma_h^{-1}+1)^{-1/2}$ for~$B_{gz}=0$.
Figure~\ref{fig:reconRateAndTotalDissipationVsBgz2d} quantifies the reconnection rates (left panel) and amount of magnetic energy and upstream flux depleted (right).
The left panel shows the
reconnection rates,
averaged over the time interval during which~$\psi(t)$
falls from~$0.9\psi_0$ to~$0.8\psi_0$.
The reconnection
rate, normalized to $B_0 c$ (blue circles) falls from roughly 0.03--0.04 for
$B_{gz}\lesssim 0.75B_0$ to around 0.01 for $B_{gz} = 4 B_0$.
The red squares show
reconnection rates normalized to $B_0 v_{A,x}$.
For $B_{gz}/B_0\gtrsim 1$, $(c/v_{A,x})\beta_{\rm rec}$ is relatively constant around 0.03; however, $(c/v_{A,x})\beta_{\rm rec}$ is closer to~0.04 or even~0.05 for $B_{gz}/B_0 < 1$.
As a compromise, Fig.~\ref{fig:reconRateAndTotalDissipationVsBgz2d} shows a dashed line at~$(c/v_{A,x})\beta_{\rm rec}=0.038$; the dotted line shows the corresponding $\beta_{\rm rec}$.
According to Eq.~(\ref{eq:vAx}), the guide field should not significantly
suppress reconnection for $B_{gz}/B_0 \ll \sigma_h^{-1/2} = 1$.
We note that a much stronger guide field would be needed to suppress nonrelativistic reconnection (where $\sigma_h \ll 1$).
On the other hand,
our previous work \citep{Werner_Uzdensky-2017} considered $\sigma_h \simeq 25$, and correspondingly
observed that even $B_{gz} = B_0/2$ significantly suppressed reconnection, slowing the reconnection rate and reducing the overall magnetic depletion
(although $B_{gz}=B_0/4$ had only a weak effect).
Figure~\ref{fig:reconRateAndTotalDissipationVsBgz2d}(right) shows that
the total fraction of $U_{Bt}$ converted to $U_{\rm plasma}$ decreases with stronger guide field---from around 27~per~cent for $B_{gz}=0$ to only about 6~per~cent for $B_{gz}=4 B_0$; similarly, the amount of flux reconnected decreases from 42~per~cent to around 21~per~cent.
We find that the fractional loss in~$\psi$ is fit fairly well by $0.86(L_x/L_y)(1+2B_{gz}^2/B_0^2)^{-1/5}$,
and the loss in~$U_{Bt}$ by $0.56(L_x/L_y)(1+4B_{gz}^2/B_0^2)^{-2/5}$;
these functional forms are meant for convenient comparison with future results and we offer no justification for them.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig11a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig11b.pdf}%
}
\caption{ \label{fig:energyFluxVsBgz2d}
For several guide field strengths $B_{gz}$:
(left) transverse magnetic energy versus time,
and (right) unreconnected flux versus time, with dot-dashed lines showing
the slopes corresponding to $\beta_{\rm rec}=$0.04 and~0.008 (cf.~\S\ref{sec:unreconnectedFlux}).
To estimate stochastic variability, two simulations are shown for
$B_{gz}=0$, two for $B_{gz}=0.25B_0$, and three for $B_{gz}=0.5B_0$.
}
\end{figure}
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig12a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig12b.pdf}%
}
\caption{ \label{fig:reconRateAndTotalDissipationVsBgz2d}
(Left) Reconnection rate versus $B_{gz}$:
blue circles show the reconnection
rate $\beta_{\rm rec}$ (normalized to $B_0 c$), and red squares show $(c/v_{A,x})\beta_{\rm rec}$---cf.~Eq.~(\ref{eq:vAx}).
The blue dotted line shows what~$\beta_{\rm rec}$ would be if $(c/v_{A,x})\beta_{\rm rec}$ were constant at~0.038 (the red dashed line).
The reconnection rates are averages over the period during which~$\psi(t)$
falls from $0.9\psi_0$ to $0.8\psi_0$ (i.e., between the horizontal
grey, dashed lines in Fig.~\ref{fig:energyFluxVsBgz2d}, right).
(Right) The fractional amount of transverse magnetic energy $U_{Bt}$ (blue x's) and upstream magnetic flux (red plusses) lost during reconnection, versus $B_{gz}$, along with empirical fits for both.
}
\end{figure}
Even a weak guide appears to enhance the reconnection rate slightly.
However, because stochastic variability can yield different reconnection rates, more work is needed to determine whether this trend is
statistically significant.
With stronger guide field, reconnection ends (or reaches a stable configuration) with a smaller amount amount of magnetic energy and flux depletion.
This may be
caused by the increased pressure of the guide magnetic field in plasmoids.
The guide magnetic field pressure resists compression with adiabatic index~$\Gamma=2$, compared with only~$\Gamma=4/3$ for relativistic plasma alone [e.g., compressing a flux tube with uniform magnetic field from volume~$V_i$ to~$V_f$ while preserving the magnetic flux inside requires work~$\propto (V_i/V_f)^{\Gamma-1}-1$ where~$\Gamma=2$; whereas compressing the plasma in the tube requires work~$\propto (V_i/V_f)^{4/3-1}-1$].
As a result, for the same amount of (transverse magnetic field) flux reconnected, the major plasmoid will be larger when the guide field is stronger.
Therefore, the growth of the major plasmoid reaches the system-size scale earlier (in terms of amount of flux reconnected, not time) and shuts down reconnection.
Perhaps not surprisingly, the slowing of reconnection inhibits NTPA,
as shown in Fig.~\ref{fig:ntpaVsBgz2d},
which displays the final particle energy spectra (left panel)
and the fitted power-law indices (right).
Overall, less energy is transferred from the magnetic field to the plasma,
and at any given well-above-average energy, stronger guide field results in
a smaller fraction of particles at that energy, hence NTPA is less
efficient.
As $B_{gz}$ increases, the power law
steepens from $p\simeq 4$ to $p\gtrsim 10$ at $B_{gz}=4B_0$.
The line $p\approx 4.4 + 2B_{gz}/B_0$ very roughly captures this trend, although it is important to remember that the uncertainty in measuring steep power laws ($p\gtrsim 4$) can be quite high (our fitting method returns consistent values characterizing the spectra, but fundamentally one should question what it means to measure a steep power law, e.g.,~$\sim \gamma^{-8}$ as $\gamma$ varies over just a half decade).
This general trend was also observed for $\sigma_h=25$ reconnection in \citet{Werner_Uzdensky-2017}, where the steepening was fairly well described by
$p=1.9+0.7\sigma_{h,\rm eff}^{-1/2}$; however, this formula provides a very poor fit for the $\sigma_h=1$ simulations, perhaps because we are in the regime of very steep power laws that cannot be reliably measured at higher energies without orders of magnitude more simulated particles.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig13a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig13b.pdf}%
}
\caption{ \label{fig:ntpaVsBgz2d}
(Left) Particle energy spectra $f(\gamma)$ at $t=12L_x/c$, for different
imposed guide magnetic fields, show that increasing guide field
inhibits NTPA, resulting in steeper power laws. Slopes with power-law indices of 4.4, 4.8, and 7 are shown for comparison. (Right) The steepening
is quantified by power-law indices fit to the high-energy
part of $f(\gamma)$, with error bars encompassing the middle 68~per~cent
of fitted indices over times from 4 to~10$L_x/c$. The line $4.4+2B_{gz}/B_0$ is shown as a guide, but it is important to remember that precise measurement of steep power laws is very difficult, and in fact the cases with $B_{gz}/B_0 \geq 1.5$ are not very different from the initial Maxwellian at $t=0$ (shown in dashed grey).
}
\end{figure}
\section{3D reconnection with moderate magnetization}
\label{sec:3d}
Despite many similarities between 2D and 3D in the $\sigma_h=1$ regime (especially in NTPA), current sheet evolution in 3D sometimes shows substantially different behaviour.
Reconnection still begins fairly rapidly in 3D, perhaps urged on (at very early times) by the thin tearing-unstable current sheet of the initial state, but the rate of magnetic energy conversion at early times is only about half that in 2D.
However, while 2D systems approach a non-reconnecting, relaxed steady state after 5--10$L_x/c$, 3D systems continue to convert magnetic energy at a slower rate (by an order of magnitude or more) for many tens of~$L_x/c$.
During 2D reconnection, part of the upstream magnetic energy is transferred to the plasma, and part is pumped into ``reconnected fields''---i.e., into plasmoids---where it remains in magnetic form \citep[see, e.g.,][]{Sironi_etal-2015}.
In 3D reconnection, plasma is also energized, but upstream magnetic energy pumped into plasmoids (or flux ropes) is subsequently converted to plasma energy \citep[rather than building up as in 2D; the decay of 3D flux ropes was also observed for $\sigma_h\gg 1$ in][]{Guo_etal-2020arxiv}.
Thus, while 2D reconnection simulations (in a closed periodic box) end up in a relaxed steady state with substantial magnetic field stored in one large plasmoid (for each initial current sheet), 3D reconnection results in a thickened, roughly uniform, much less structured, turbulent current layer with greatly diminished magnetic field. This thickened layer continues slowly to convert magnetic energy to plasma energy; ultimately more magnetic energy can be converted in 3D than in 2D.
Despite a lower reconnection rate in 3D (hence lower reconnection electric field), NTPA remains robust.
While 2D reconnection can be bursty due to stochastic creation, movement, and merging of plasmoids, these events tend to be random blips in a predetermined course of evolution toward a unique final state. In contrast, 3D reconnection exhibits greater variability, with stochastic events leading to a thicker turbulent layer in different ways.
In particular, we have often observed the following behaviour.
Because of RDKI, the current layer sometimes develops large-amplitude kinking in $z$, resulting in the layer dramatically folding over on itself like a breaking wave. This tends to deplete very rapidly almost all the magnetic energy within the original amplitude of oscillation, resulting in a thick turbulent layer.
However, this behaviour is not inevitable; sometimes the kink amplitude grows but stops short of folding over on itself, resulting in much slower energy conversion and layer growth.
It remains unclear---because it might require running simulations for hundreds or thousands of $L_x/c$---whether a simulation that does not undergo ``layer-folding'' will ever convert as much energy or thicken the layer as much as one that does.
In the following subsection, we present an overview of differences between 2D and 3D current sheet evolution, using our largest simulations. To study the effect of ``3D-ness'' we consider four configurations that are identical but for different values $L_z$. The two smallest $L_z$ behave similarly (i.e., like 2D reconnection), whereas the two largest exhibit 3D effects.
Following the overview (\S\ref{sec:overview3d}), which will describe magnetic energy conversion and NTPA as well as plasma and field evolution,
we include a brief discussion of increased stochastic variability, or sensitivity to initial conditions, in 3D (\S\ref{sec:variability3d}).
Then, using an extensive set of slightly smaller simulations, we will specifically explore the dependence of energy conversion and NTPA on the initial current sheet configuration
(\S\ref{sec:pertAndEta3d} and~\S\ref{sec:RDKIamp}), on the aspect ratio $L_z/L_x$ (i.e., the ``3D-ness,'' \S\ref{sec:Lz3d}), and finally on guide magnetic field (\S\ref{sec:Bz3d}).
We note that the 3D generalization of a plasmoid or magnetic island that forms in 2D reconnection is a flux rope. However, we continue to use the term ``plasmoid'' to refer to flux ropes or any concentrations of plasma in 3D reconnection, both for simplicity and because we have not yet thoroughly investigated these objects to determine whether they truly have the magnetic structure of flux ropes.
\subsection{Overview of 3D reconnection with $\sigma_h=1$, and comparison with 2D}
\label{sec:overview3d}
In this section we begin with an overview of
basic 3D current sheet behaviour
for one specific configuration, namely:
$\sigma_h=1$, zero guide field ($B_{gz}=0$), zero initial perturbation ($a=0$), an initially-thin current sheet $\delta=(2/3)\sigma \rho_0$ ($\eta=5$), and $L_x=512\sigma \rho_0 = L_y/2$ (the largest $L_x$ of any 3D simulations in this paper).
We will show results from simulations with $L_z/L_x=0$ (a 2D simulation),
1/32, 1/8, and 1; we will show in particular
that $L_z=L_x/32$ behaves essentially like the 2D simulation, while
the two largest simulations are similar to each other but quite different from the smaller two, presumably because of 3D effects that appear for sufficiently large $L_z$.
Before looking at the evolution of 2D and 3D current sheets,
it is helpful to see how global energies and unreconnected flux vary in time.
We begin by looking at global field energies versus time, normalized to initial magnetic energy $U_{B0}$, in Fig.~\ref{fig:energyVsTime2D3D}(a)-(c).
In these simulations without guide field,
almost all the field energy is contained in $U_{Bt}$, the energy in the transverse magnetic field
$\boldsymbol{B}_t = B_x \hat{\boldsymbol{x}} + B_y \hat{\boldsymbol{y}}$.
In all cases $U_{Bz}(t) < 0.003U_{B0}$,
and the energy in the electric field $\boldsymbol{E}$ (not shown)
also remains below $0.003 U_{B0}$.
The $B_z$ and $E$ field components can nevertheless be substantial within the (the small volume of) the layer.
Further investigation of $B_z$ is left for future work, but
we note that there should be no $B_z$ at all in a 2D MHD description \citep[for pair plasma with no initial guide field,][]{Zenitani_etal-2009}, while 2D PIC simulations do develop non-zero $B_z$ \citep{Zenitani_Hesse-2008}.
Figure~\ref{fig:energyVsTime2D3D}(a) shows $U_{Bt}$ decreasing over time in all cases as it is converted to plasma energy.
The simulations with larger $L_z/L_x=1/8,1$ show similar magnetic depletion, and those with smaller $L_z/L_x=0,1/32$ closely resemble each other; in this and other diagnostics we will see that the former ($L_z/L_x=1/8,1$) exhibit 3D behaviour, while the latter are essentially~2D.
We find that magnetic energy depletes
roughly half as fast in 3D as in 2D, at early times
up to $\sim 4L_x/c$ (Fig.~\ref{fig:energyVsTime2D3D}a; see also table~\ref{tab:finalEnergyAndFlux}, which compares energy changes and reconnection rates).
While the 2D simulations have mostly finished reconnection by $t\sim 4L_x/c$ (approaching a final state a few $L_x/c$ later), the 3D simulations embark upon a second stage of prolonged slow energy conversion \citep[cf.][]{Liu_etal-2011}, converting magnetic energy to plasma energy an order of magnitude still more slowly for many tens of $L_x/c$.
Were these simulations to run many times longer (which they did not, because of prohibitive computational cost),
the 3D simulations would eventually convert more magnetic energy than the essentially 2D simulations; we will see in~\S\ref{sec:Lz3d}, using smaller simulations ($L_x=341\sigma\rho_0$), that 3D simulations convert more magnetic energy over $\sim 50 L_x/c$ than 2D simulations.
\begin{figure}
\centering
\fullplot{
\raisebox{0.315\textwidth}{\large (a)}%
\includegraphics*[width=0.455\textwidth]{fig14a.pdf}%
\hfill
\raisebox{0.315\textwidth}{\large (b)}%
\includegraphics*[width=0.455\textwidth]{fig14b.pdf}\\
\raisebox{0.315\textwidth}{\large (c)}%
\includegraphics*[width=0.455\textwidth]{fig14c.pdf}%
\hfill
\raisebox{0.315\textwidth}{\large (d)}%
\includegraphics*[width=0.455\textwidth]{fig14d.pdf}\\
}
\caption{ \label{fig:energyVsTime2D3D}
Energy in various magnetic field components versus time for 4 different $L_z/L_x$, normalized to
$U_{B0} \approx \int dV \, B_0^2/8\upi $:
(a) $U_{Bt} = \int dV \, (B_x^2+B_y^2)/8\upi$,
(b) $U_{Bz} = \int dV \, B_z^2/8\upi$,
and (c) $U_{Bt,\rm layer}$, the magnetic energy in the layer (cf.~\S\ref{sec:terminology}).
The electric field energy (not shown) is always less than $0.003 U_{B0}$.
Finally (d), the unreconnected magnetic flux $\psi(t)$;
dotted grey lines indicate dimensionless reconnection rates,
$(cB_0L_z)^{-1} (d\psi /dt)$, of 0.045 and 0.015, which are roughly
the rates of nearly 2D and 3D simulations, respectively, between the grey
dashed lines, and also 0.002, roughly the 3D rate at later times.
}
\end{figure}
\begin{table}
\centering
\begin{tabular}{lccccc}
& $-\frac{\Delta U_{Bt,\rm up}}{U_{B0}}$
& $\frac{\Delta U_{Bt,\rm layer}}{U_{B0}}$
& $-\frac{\Delta U_{Bt}}{U_{B0}} \approx
\frac{\Delta U_{\rm plasma}}{U_{B0}}$
& $-\frac{\Delta \psi}{\psi_0}$
& $\beta_{\rm rec,0.95-0.90}$
\\
nearly 2D $\left(\frac{L_z}{L_x}=0,\frac{1}{32}\right)$: &
0.44--0.47 & 0.17--0.20 & 0.27 & 0.37--0.41 & 0.043--0.047
\\
fully 3D $\left(\frac{L_z}{L_x}=\frac{1}{8}, 1\right)$: &
0.18--0.21 & 0.01 & 0.18--0.20 & 0.14-0.16 & 0.01--0.02
\end{tabular}
\caption{\label{tab:finalEnergyAndFlux}
The fractional loss
in transverse magnetic
field energy $U_{Bt}$ in the entire simulation
(nearly equal to the gain in plasma energy
$U_{\rm plasma}$), divided into upstream and layer
regions ($U_{Bt,\rm up}+U_{Bt,\rm layer}=U_{Bt}$), and
the loss in unreconnected magnetic flux $\psi$,
over the first $10L_x/c$,
for simulations with $L_x=512\sigma\rho_0$ and $B_{gz}=0$.
Changes are shown normalized to initial magnetic energy $U_{B0}$ and
initial unreconnected flux $\psi_0$.
The fully 3D simulations would continue to
deplete magnetic energy if run longer (cf.~\S\ref{sec:Lz3d}).
Also shown is $\beta_{\rm rec}$ (cf.~\S\ref{sec:unreconnectedFlux}),
averaged over the time during which~$\psi(t)$
drops from~0.95$\psi_0$ to~0.90$\psi_0$.
}
\end{table}
More magnetic energy conversion is possible in 3D because the magnetic energy pumped into plasmoids can be depleted in 3D, whereas in 2D plasmoids remain stable.
To demonstrate how little magnetic energy remains stored in plasmoids in 3D,
we separately measure the transverse magnetic energy in the upstream and in the layer regions (cf.~\S\ref{sec:terminology}):
$U_{Bt} = U_{Bt,\rm up} + U_{Bt,\rm layer}$.
With no guide field and no initial perturbation, the initial magnetic energy
$U_{B0}=U_{Bt0}$ is entirely in unreconnected transverse field.
During magnetic reconnection in 2D, part of this gets converted into plasma energy $U_{\rm plasma}$, and part into $U_{Bt,\rm layer}$ in plasmoids.
Over the first $10L_x/c$, the (nearly) 2D simulations take the initial upstream magnetic energy $U_{B0}$ and convert about $\Delta U_{\rm plasma}\approx 0.27U_{B0}$ to plasma energy while building up a comparable amount, $\Delta U_{Bt,\rm layer}\approx 0.20_{B0}$, in plasmoids, where it remains trapped.
Since the energized particles are also trapped in plasmoids, we estimate that the magnetic energy in plasmoids is about 2/3 of the plasma energy in plasmoids.
Over the same time, fully 3D simulations convert $\Delta U_{\rm plasma}\approx 0.2 U_{B0}$ to plasma energy, but magnetic energy in the layer quickly decays, leaving a strikingly smaller upper bound on the amount of energy, $\Delta U_{Bt,\rm layer}\approx 0.01 U_{B0}$, that could be in plasmoids
(see Fig.~\ref{fig:energyVsTime2D3D}c and table~\ref{tab:finalEnergyAndFlux}).
Although 3D simulations deplete upstream energy $U_{Bt,\rm up}$ more slowly, they thus convert it more completely to plasma energy.
Along with magnetic energy, we can also measure the unreconnected magnetic flux, which decreases over time, faster in 2D than in 3D.
In 3D, as in 2D, some of the upstream flux is promptly reconnected, ending up in plasmoids, as in 2D.
Unlike in 2D, however, in 3D some upstream flux may be promptly annihilated, and also (only in 3D) reconnected flux is subsequently annihilated as plasmoids or flux rope structures become unstable and decay.
It is nontrivial to measure how much upstream flux is annihilated promptly without first undergoing reconnection, but we can conclude from the relatively small amount of magnetic energy in the layer that ultimately, in 3D, most of the lost upstream flux is annihilated (versus being stored permanently in plasmoids).
Figure~\ref{fig:energyVsTime2D3D}(d) shows the unreconnected flux $\psi(t)$ diminishing as reconnection occurs for the four different values of $L_z/L_x$.
Table~\ref{tab:finalEnergyAndFlux} shows that up to~$t=10L_x/c$, the
nearly 2D simulations lose 40 per~cent of the initial unreconnected flux, and fully 3D simulations lose only about 15 per~cent (though, over time, they would reconnect and/or annihilate more upstream flux, whereas the 2D simulations have essentially finished).
The reconnection rate $\beta_{\rm rec}$ at early times, measured between $\psi(t)=0.95\psi_0$ and $\psi(t)=0.90\psi_0$ (the horizontal dashed lines in Fig.~\ref{fig:energyVsTime2D3D}d) and normalized to~$B_0 c$, is about~0.045 for the nearly~2D simulations, and~0.015 for 3D (see table~\ref{tab:finalEnergyAndFlux}); later, around~$t=8L_x/c$, $\beta_{\rm rec}$ is on the order of $\sim 0.002$ for the 3D simulations (and fluctuating around zero for~2D).
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=\textwidth]{fig15a.pdf}%
}
\caption{ \label{fig:layerOffsetVsXz}
The displacement $\Delta y_c(x,z,t)$ of the current sheet central surface
from the original midplane (cf.~\S\ref{sec:sheetCenter}), for
$L_z=L_x$ at two different times, $t=0.3L_x/c$ (left) and $0.6 L_x/c$ (right),
shows increasing rippling due to the RDKI.
}
\end{figure}
Having shown significant differences (as well as qualitative similarities) between 2D and 3D energy and upstream flux depletion, we will describe the evolution of the current sheet. First, we will describe the kinking of the sheet, which cannot occur in 2D at all; then we will spend some time describing 3D behaviour that resembles 2D reconnection.
Perhaps the most expected (by now) new effect in 3D
is the kinking of the current sheet in the $z$ direction due to the RDKI \citep{Zenitani_Hoshino-2007,Zenitani_Hoshino-2008}.
Simulating the $z$ dimension enables the RDKI, which can occur because of the relative movement of electrons and positrons in the $z$-direction
and causes
rippling of the sheet.
Figure~\ref{fig:layerOffsetVsXz} shows (for the case with $L_z=L_x$) a current sheet rippling due to RDKI with an amplitude $\Delta y \approx \sigma\rho_0$ at $t=0.3L_x/c$, growing to $3\sigma\rho_0$ at $t=0.6L_x/c$ [for comparison, the initial current sheet thickness was $\delta=(2/3)\sigma \rho_0$].
It had originally been thought that RDKI---at least in cases with weak guide field---might out-compete the tearing instability, thereby suppressing reconnection \citep{Zenitani_Hoshino-2008}.
However, later, larger 3D simulations (in the high-$\sigma_h$ regime) found that, despite active RDKI, reconnection ultimately does not seem to be inhibited, even with zero
guide field \citep{Sironi_Spitkovsky-2014,Guo_etal-2014,Guo_etal-2015,Werner_Uzdensky-2017}.
Indeed, 3D simulations do exhibit behaviour that, at least locally, shares many characteristics with 2D reconnection.
Although energy is clearly transferred from magnetic fields to plasma,
determining whether and where reconnection occurs is nontrivial in 3D.
We tentatively suggest, based on our 3D simulations, that reconnection occurs in patches within the layer---patches that may be the analogues of elementary current sheets in 2D reconnection (i.e., the smallest inter-plasmoid current sheets, which are not further broken up by secondary tearing).
It is possible that reconnection is slower (overall) in 3D because the area of actively-reconnecting regions (patches) is smaller, while the local reconnection rate remains as high as in 2D \citep[such patchiness with locally high reconnection rates has been observed for $\sigma_h\sim 1$ by][]{Yin_etal-2008}.
Our results are consistent with this hypothesis, but so far we have been unable to measure local reconnection rates with enough certainty to conclude this definitively; we leave a dedicated investigation of this problem to future studies.
However, we offer evidence that reconnection does in fact occur (though in patches), and that the patches evolve over time, and continue to exhibit signatures of reconnection even at late times during the ``slow 3D reconnection stage'' (e.g., $t>4L_x/c$).
Specifically, we will see that actively-reconnecting regions are areas
where the current sheet is thin, that they exhibit outflows (roughly) in the $\pm x$ directions, and that they are the areas with strongest parallel (to $\boldsymbol{B}$) electric fields and $\boldsymbol{E}\bcdot\boldsymbol{J}>0$.
Figure~\ref{fig:neVsXy2D3D} shows the electron density $n_e$, normalized to $n_{be0}=n_{b0}/2$, at five different times ($tc/L_x=0.3$, 0.6, 1.1, 2.3, and 4.6) in $x$-$y$ slices,
at $z=0$ for the nearly 2D simulation with $L_z=L_x/32$, and at $z=100\sigma\rho_0$ and $z=300\sigma\rho_0$ for $L_z=L_x$.
A sample of magnetic field lines (traced while ignoring $B_z$) shows that $n_{e} \lesssim n_{be0}$ in the upstream region, where flux is unreconnected (cf.~\S\ref{sec:unreconnectedFlux}), while $n_e > n_{be0}$ in most of the layer.
Therefore, looking at density is a good if rough way to track the gross evolution of the shape of the layer.
For $L_z=L_x/32$, the picture is practically the same as in 2D:
the initial current sheet tears, small plasmoids form and merge into larger plasmoids, until there is just one plasmoid (and a monster one at that) in the layer. Between plasmoids, the layer is thin at X-points (really, $X$-lines extended in the $z$ direction) where reconnection takes place.
For $L_z=L_x$, the picture is initially very similar to that in 2D, when viewed in an $x$-$y$ slice.
For example, at $t=0.3L_x/c$, a similar number of plasmoids is visible
(although a slice at any value of $z$ looks qualitatively similar, these structures do not appear to extend uniformly in $z$; i.e., at early times we do not observe long flux ropes of length $L_z$).
Soon (by $t=0.6L_x/c$), however, the 3D layer shows general thickening without the clear plasmoid structures so familiar in 2D.
At some rare places (e.g., $x\approx z \approx 100 \sigma\rho_0$, $tc/L_x=1.1$--$2.3$), the layer remains very thin, as at X-points in 2D.
In the vicinity of these thin places we still see some ballooning of the layer, resembling a larger plasmoid, but smaller (in $y$) than in 2D, and with much less distinct density structure.
After $2.3L_x/c$, most of the layer is thick, but not as thick as plasmoids in 2D.
The $L_z=L_x/32$ simulation is fairly uniform in $z$, consistent with its
behaviour resembling 2D reconnection.
The 3D simulation, however, is very nonuniform, as we will see in a series
of plots showing various fields at time $t=2.3L_x/c$, at the centre of the layer, i.e., at $y_c(x,z,t)$ such that $B_x(x,y_c,z,t)=0$.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=\textwidth]{fig16a.pdf}%
}
\caption{ \label{fig:neVsXy2D3D}
A time sequence of electron density $n_e$ in the $x$-$y$ reconnection plane, in nearly 2D ($L_z=L_x/32$, left column) and 3D ($L_z=L_x$, middle and right) simulations for $tc/L_x=0.3, 0.6, 1.1, 2.3, 4.6$.
Blue lines follow magnetic field lines, ignoring $B_z$ (with a higher density of lines shown closer to the midplane).
The left column shows $z=0$ (which looks fairly similar to all other $z$ values for this nearly 2D case); the middle column shows $z=100 \sigma \rho_0\approx 0.2 L_z$ and the right column shows $z=300 \sigma \rho_0 \approx 0.6L_z$.
}
\end{figure}
\begin{figure}
\centering
\fullplot{
\raisebox{0.285\textwidth}{a.}%
\includegraphics*[width=0.30\textwidth]{fig17a.pdf}\hfill
\raisebox{0.285\textwidth}{c.}%
\includegraphics*[width=0.30\textwidth]{fig17c.pdf}\hfill
\raisebox{0.285\textwidth}{e.}%
\includegraphics*[width=0.30\textwidth]{fig17e.pdf}\\
\raisebox{0.285\textwidth}{b.}%
\includegraphics*[width=0.30\textwidth]{fig17b.pdf}\hfill
\raisebox{0.285\textwidth}{d.}%
\includegraphics*[width=0.30\textwidth]{fig17d.pdf}\hfill
\raisebox{0.285\textwidth}{f.}%
\includegraphics*[width=0.30\textwidth]{fig17f.pdf}%
}
\caption{ \label{fig:layerXz3dn6144}
Field quantities versus $x$ and $z$ at the layer central surface, i.e., at $y=y_c(x,z,t)$ such that $B_x(x,y_c,z,t)=0$ (cf.~\S\ref{sec:sheetCenter}).
$A$ and $B$ indicate---roughly---examples of actively-reconnecting regions.
(a) The current density in the $z$-direction, $J_z$, is highest in reconnecting regions, where the layer is thin.
(b) $\overline{v_x}=(v_{e,x} + v_{i,x})/c$, the average of electron and positron $x$-velocities, shows strong outflows from reconnection regions.
(c) The electric field, $\|\boldsymbol{E}\|$ is largest in reconnecting regions and also in strong outflows.
(d) Although small, $E_\|$, the magnitude of the component of $\boldsymbol{E}$ parallel to the local $\boldsymbol{B}$, is stronger in reconnecting regions.
(e) $\boldsymbol{E}\bcdot \boldsymbol{J}$ indicates dominant power transfer from fields to plasma, mostly in reconnecting and outflow regions.
(f) $\boldsymbol{E}_\|\bcdot\boldsymbol{J}_\|$ is positive and strongest in reconnecting regions, indicating significant plasma energization by parallel electric fields (or electric fields where $B\approx 0$).
All plotted quantities were averaged over $y_c-\sigma\rho_0 < y < y_c+\sigma\rho_0$.
}
\end{figure}
Figure~\ref{fig:layerXz3dn6144} shows various quantities at the layer central surface, $y_c(x,z,t)$ (cf.~\S\ref{sec:sheetCenter}), versus $x$ and $z$, for the $L_z=L_x$ case, at $t=2.3L_x/c$.
We focus particularly on region $A$, a patch of size (at this time) $\simeq 100\sigma\rho_0$ located around $x\approx z \approx 100\sigma\rho_0$, which we believe to be an actively-reconnecting region; we saw in Fig.~\ref{fig:neVsXy2D3D} that it is a site where the layer is near its thinnest.
Panel (a) shows the current density $J_z$ in the $z$-direction, which is indeed most prominent in this reconnecting region, consistent with the layer being thin. (We note that current sheets
are locally not necessarily parallel to $z$, but are nearly parallel to $z$ in many places.)
Panel (b) shows strong outflows from this region in the $\pm x$ directions,
as one would expect from reconnection; similarly, region~$B$, around $x\approx 320\sigma\rho_0$ and $z\approx 400\sigma\rho_0$ with size (at this time) similar to patch $A$, has a high current density and strong outflows.
There are also smaller reconnecting regions---e.g., at $x\approx 260\sigma\rho_0$ and $z\approx 280\sigma\rho_0$.
The ``line'' between the outflows thus appears to be a stagnation line, which is commonly associated with an X-line in the classic 2D picture of (symmetric) reconnection.
Panel (c) shows $\|\boldsymbol{E}\|$, which is sizeable ($\sim 0.1 B_0$) over a significant region, especially in the neighbourhood of large, strong outflows.
However, (panel d) the parallel electric field $|E_\||$ (parallel to the local
$\boldsymbol{B}$) is largest only in a region much closer to the stagnation line, as expected.
Panel (e) shows that $\boldsymbol{E}\bcdot \boldsymbol{J}$ has more strongly-positive areas than negative, indicating energy transfer from electromagnetic fields to the plasma;
this transfer occurs around X-lines and extended outflows, but as panel (f) shows, the region where $E_\| J_\|$ is most positive is limited to areas closer to the largest stagnation lines (in regions $A$ and $B$).
Figure~\ref{fig:layerXz3dn6144} thus suggests that reconnection is
occurring in patches (such as regions $A$ and $B$), perhaps not with such neat geometry as in 2D, but exhibiting basic signatures, such as a thin layer with high current density, strong outflows, particle energization, and indeed particle energization by parallel electric fields or by electric fields in a region of weak magnetic field.
Very early in time, the layer exhibits many small patches of active reconnection. At any point in time, a patch tends to have roughly the same size in $x$ and $z$ (the $x$-extent can be seen in Fig.~\ref{fig:neVsXy2D3D}).
For example, patches have size $\sim 20\sigma\rho_0$ at $t=0.3L_x/c$.
It is perhaps therefore not surprising that the simulation with $L_z=L_x/32=16\sigma\rho_0$---which is smaller than the patch size even at this early time---behaves like 2D reconnection.
Over time (for $L_z=L_x$), the number of reconnecting patches decreases rapidly, but the surviving patches grow in size, in both $x$ and $z$.
By $2.3L_x/c$, the 3D simulation has 2 major patches (regions $A$ and $B$), with perhaps a few much smaller, probably less active ones.
By $4.6L_x/c$, only region $A$ remains, although it seems to have shrunk in the $z$ direction, perhaps spreading its current density a little wider in $x$.
Importantly, region $A$ does continue to exhibit signs of active reconnection at these late-stage times, when the global reconnection rate is relatively slow in 3D.
We leave a systematic investigation of the behaviour and statistics of these patches to future work.
Although the above-described patches in 3D locally resemble elementary reconnecting current sheets in 2D, important differences emerge away from these patches.
As previously mentioned, we do not see the long-term storage of magnetic energy in highly-structured plasmoids as in 2D;
for example, in 3D reconnection we do not see clear signatures of persistent, growing structures extended across the simulation in the $z$-direction (as we would expect flux ropes to be), e.g., in Fig.~\ref{fig:layerXz3dn6144}.
In (nearly) 2D, the magnetic field is somewhat diminished far upstream of major X-lines, e.g., to $\approx 0.7 B_0$ more than $100\sigma\rho_0$ upstream at $t=4.6L_x/c$; at the same time, the magnetic field rarely
falls below $0.3B_0$ anywhere (Fig.~\ref{fig:BvsXy2D3D}, left).
In 2D, plasmoids continue to grow roughly up to the size of the simulation,
when the magnetic configuration prevents further reconnection.
In 3D, however, the current layer becomes a somewhat thickened turbulent region with relatively sharp boundaries (where most current flows) inside which the magnetic field is generally diminished.
The magnetic field is practically full strength outside of the sharp boundaries, while being less than $0.3 B_0$ in a large fraction of the volume within the boundaries (Fig.~\ref{fig:BvsXy2D3D}, middle and right).
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=\textwidth]{fig18a.pdf}%
}
\caption{ \label{fig:BvsXy2D3D}
The magnetic field strength $|\boldsymbol{B}|$ at $t=4.6L_x/c$ in the same $z$-planes shown in Fig.~\ref{fig:neVsXy2D3D}; (left) nearly 2D, $z=0$, (middle) 3D, $z=100\sigma\rho_0$, and (right) 3D, $z=300\sigma\rho_0$.
Contours are drawn for $B/B_0=\sqrt{0.1}\approx 0.32$ (dotted white), $\sqrt{0.3}\approx 0.55$ (dashed green), and $\sqrt{0.5}\approx 0.71$ (solid blue).
}
\end{figure}
The evolution of current sheets in 3D is complicated, and demands much more study.
For now, however, we will move on to consider one of the most important consequences of current sheet evolution and resulting energy conversion, namely NTPA.
Reconnection-driven NTPA is a promising candidate to explain, e.g., high-energy X-ray and gamma-ray emission from a variety of astrophysical sources.
It has been previously suggested that NTPA in 3D might be suppressed by the RDKI interfering with reconnection \citep{Zenitani_Hoshino-2007, Zenitani_Hoshino-2008}.
And since the reconnection electric field responsible for accelerating particles is proportional to the reconnection rate, and we have shown here that the global reconnection rate is slower in 3D than in 2D, we have another reason to suspect that NTPA might be considerably less efficient in 3D.
However, that is emphatically not the case here.
Figure~\ref{fig:ntpa2D3D} shows electron energy spectra~$f(\gamma)$ (compensated by~$\gamma^2$)---compared both at the same time ($t=9 L_x/c$) and after the same amount of magnetic energy has been given to particles ($\Delta U_{\rm plasma} \approx 0.17 \: U_{B0}$).
Remarkably, all the spectra are pretty similar, exhibiting high-energy nonthermal behaviour with a power-law slope $\approx -4$ [i.e., $f(\gamma)\sim \gamma^{-4}$].
However, though the difference may be small, 3D reconnection yields more particles with very high energies than 2D; in addition, since ultimately 3D reconnection can convert more magnetic energy to plasma energy than 2D, 3D reconnection may accelerate more particles overall (not just at the very highest energies).
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig19a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig19b.pdf}%
}
\caption{ \label{fig:ntpa2D3D}
Electron energy spectra $f(\gamma )$, compensated by $\gamma^2$, where particle energy is $\gamma m_e c^2$, for $L_z/L_x=0$, 1/32, 1/8, 1, (left) at the same time, $t=9 L_x/c$ and (right) when the same amount of magnetic energy ($\simeq 17$ per~cent) has been depleted. Dotted grey-on-yellow
lines indicate a range of power-law slopes $f(\gamma)\sim \gamma^{-3.4}$, $\gamma^{-4}$, and $\gamma^{-4.4}$.
}
\end{figure}
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig20a.pdf}%
}
\caption{ \label{fig:ntpaVsTime}
Electron energy spectra $f(\gamma)$, compensated by $\gamma^2$, at different times for the 3D case $L_z/L_x=1$: the spectra are shown at times $tc/L_x=1.1$, 2.2, 4.3, and 8.6, when fractions 0.05, 0.09, 0.13, and 0.17 of the initial magnetic energy $U_{B0}$ have been depleted.
Power-law indices of 3.4, 4, and 4.4 are shown for reference, as in
Fig.~\ref{fig:ntpa2D3D}.
Even though reconnection is much slower after $4L_x/c$ than before, significant NTPA occurs between $tc/L_x=$4.3 and 8.6.
}
\end{figure}
NTPA continues to occur (in 3D) even during the second stage of reconnection, when the global rates of reconnection and magnetic energy conversion are smaller by an order of magnitude compared with the first stage.
Figure~\ref{fig:ntpaVsTime} shows the spectra for $L_z=L_x$ at four different times, $tc/L_x=1.1, 2.3, 4.3, 8.6$, chosen so that between successive times, $U_{\rm plasma}$ increases by equal amounts, namely $0.04 U_{B0}$---i.e., $\Delta U_{\rm plasma}/U_{B0} = 0.05, 0.09, 0.13, 0.17$.
Even after $t\simeq 3L_x/c$, when the magnetic energy is at 89~per~cent and after which the reconnection rate drops significantly, we see that $f(\gamma)$ continues to increase at high energies.
Indeed, the increase in $f(\gamma)$ preserves the power-law slope at high energies; the high-energy slope of $f(\gamma)$ does not decrease after $3L_x/c$, nor does the high-energy cutoff decrease (if anything, the opposite occurs, possibly indicating re-acceleration of already-energetic particles that have returned coincidentally to actively reconnecting regions).
It is possible that NTPA in 3D is at least as efficient as in 2D, despite the slower (global) reconnection rate in 3D, because in 3D the local reconnection rate is as high as in 2D within actively-reconnecting patches.
Thus NTPA may be occurring in a similar manner in 2D and in 3D, but in a relatively smaller region in 3D.
This result is consistent with \citet{Yin_etal-2008};
the global reconnection rate drops because there are fewer/smaller patches, not because local reconnection rates drop.
Furthermore, the stark contrast in plasmoid evolution in 2D and 3D might promote more efficient NTPA in 3D.
In 2D, ballooning plasmoids expand deeply into the upstream, and can contain higher magnetic field than the upstream region; the magnetic fields of large plasmoids can trap high-energy particles \citep[with Larmor radii a large fraction of the plasmoid size;][]{Sironi_etal-2016,Uzdensky-2020arxiv}.
In 3D, however, any structures (flux ropes) formed from reconnected magnetic field decay \citep[see also][]{Zhou_etal-2020};
high-energy particles may be confined to the layer by the upstream magnetic fields, but probably not by distinct sub-structures within the current layer.
The lack of trapping may result in more particle
acceleration in 3D; whereas a particle trapped in a plasmoid does not get accelerated by reconnection \citep[but could be accelerated by a different, slower mechanism;][]{Petropoulou_Sironi-2018}, untrapped particles could potentially pass multiple times through reconnecting layers, thus experiencing a sort of multi-stage acceleration \citep{Dahlin_etal-2015,Zhang_etal-2021arxiv}.
It is worth estimating, from a Hillas criterion perspective \citep{Hillas-1984}, the
energy gain of a particle in a typical reconnection electric field
$E_{\rm rec}\sim 0.04B_0$, traversing an entire patch of active reconnection.
By $t=2.3L_x/c$, there is a patch (region $A$ in Fig.~\ref{fig:layerXz3dn6144}) in the simulation with $L_z=L_x$ that is about
$\ell\sim 100\sigma\rho_0$ in size.
Thus we estimate an energy gain $eE_{\rm rec}\ell\sim 4 \sigma m_e c^2$, which is roughly consistent with the high-energy cutoff
\citep[a similar result was given for elementary current sheets in 2D, for high-$\sigma_h$ reconnection;][]{Werner_etal-2016}.
We do not suggest this proves that all particles are being accelerated
directly by the reconnection electric field; the acceleration mechanism
is beyond the scope of this paper and remains under debate \citep[see, e.g.,][]{Drake_etal-2006,Nalewajko_etal-2015,Ball_etal-2018,Ball_etal-2019,Guo_etal-2019,Sironi_Beloborodov-2020,Kilian_etal-2020,Guo_etal-2020arxiv,Uzdensky-2020arxiv,Zhang_etal-2021arxiv}, but at least for 2D and 3D reconnection with weak guide field, above-average-energy particles probably have gyroradii larger than the layer thickness and will experience more acceleration by the motional electric field of plasmoids,
$\boldsymbol{E} = -(1/c)\boldsymbol{v}\times \boldsymbol{B}$
\citep{Guo_etal-2019,Kilian_etal-2020,Guo_etal-2020arxiv,Uzdensky-2020arxiv}.
We offer this estimate simply as a way to compare system size to
particle energies.
Thus we have seen both striking similarities and striking differences between 2D and 3D reconnection. Both convert significant amounts of magnetic energy to plasma energy, but the conversion rate is slower in 3D. Many of the signatures of reconnection at 2D X-points (or X-lines) are apparent in 3D simulations, but what happens in reconnection outflows is very different. While 2D simulations beget a growing, highly-structured hierarchy of distinct plasmoids containing energized particles and reconnected magnetic field, 3D simulations only form small plasmoid-like features, which subsequently decay, allowing magnetic field energy in the layer to be converted to plasma energy. As a result, 3D simulations ultimately convert more magnetic energy than in 2D.
The decay of plasmoids in 3D means that the reconnected flux in those plasmoids is annihilated, while in 2D, flux is almost perfectly conserved, such that all the lost upstream flux ends up as reconnected flux inside plasmoids. In 3D, the lost upstream flux is mostly annihilated (and it is yet to be determined what fraction is directly annihilated and what fraction is first reconnected and later annihilated).
Interestingly, perhaps the strongest similarity between 2D and 3D reconnection is NTPA; the spectrum of high-energy particles is very similar in both cases, especially if compared at times when the depleted magnetic energy is the same.
With this overview finished, we will continue to examine differences between 2D and 3D reconnection in more depth.
\subsection{3D reconnection: stochastic variability}
\label{sec:variability3d}
An important difference between 2D and 3D current sheet simulations is the greater stochastic variability in 3D.
When initializing the simulations, the initial electric and magnetic fields are precisely determined. However, the particle positions and velocities representing the plasma distribution in Monte Carlo fashion are chosen randomly (according to a precisely-given distribution). Thus two ``identical'' simulations will not be exactly the same, and the initial microscopic randomness can lead to variability at macroscopic scales.
In 2D reconnection, two ``identical'' simulations will yield different plasmoid formation, motion, and merging; this can affect the global reconnection and energy conversion rates.
Smaller simulations often show greater stochastic variation in global quantities such as energy versus time, because there are fewer plasmoids overall, increasing the influence of erratic individual plasmoids.
We have often observed that the process of plasmoid merging (especially in simulations without initial perturbation; cf.~\S\ref{sec:pertAndEta2d}) results in two monstrously large plasmoids in a layer spaced apart by nearly $L_x/2$. These two plasmoids will ultimately merge; but by symmetry due to periodic boundary conditions, one plasmoid can either move left or right to merge with the other.
This symmetry is always eventually broken to yield a lower-energy magnetic configuration, but its breaking can be ultimately traced to the initial microscopic randomness.
As these plasmoids decide which way to go for the final merger, the magnetic energy often stagnates at a roughly constant level for a short time before rapidly declining during the final plasmoid merger (see, e.g., Fig.~\ref{fig:energyFluxVsLx2d}, $L_x/\sigma\rho_0=640$, 1280, and 2560, around 3--5$L_x/c$).
However, due to randomness, sometimes the final two plasmoids are not created so symmetrically, and the magnetic energy decline continues smoothly.
Thus we often describe reconnection as ``bursty''; in radiative reconnection, where high-energy particles cool rapidly, e.g. via inverse Compton radiation, this burstiness is visible even in the spectrum of high-energy particles \citep{Werner_etal-2019}.
Thus initial random particle positions and velocities lead to macroscopic variability in 2D at system-size scales.
However, although significant, this stochastic variability is a sequence of random short detours on a path to essentially the same final state. In fact, \S\ref{sec:pertAndEta2d} showed that varying the initial current sheet configuration (varying more than just different random velocities from the same distribution) still had little effect on the overall evolution of reconnection.
Looking at coarse time scales of $O(L_x/c)$, the overall rate and amount of energy depletion are fairly independent of initial conditions (for the same upstream conditions).
In contrast, in a 3D reconnection simulation, random initialization can lead to significant macroscopic differences that persist for long times---at least as long as we have been able to simulate.
While a detailed study of random variation in large ensembles of ``identical'' simulations is beyond the scope of this paper, we present here an anecdote of high variability in 3D in a case with $L_x=256 \sigma\rho_0$ and $L_z=L_x/4$, which we ran three times with ``identical'' set-ups ($B_{gz}=0$, $\eta=5$, $\beta_d=0.3$, $a=0$).
The three simulations ultimately exhibit very different magnetic energy and flux evolution (Fig.~\ref{fig:variability3d}) even over scales of tens of $L_x/c$; this
variation is much greater than we see in 2D (although not perfectly comparable, Figs.~\ref{fig:perturb2dLx320}, \ref{fig:eta2dLx1280}, and \ref{fig:energyFluxVsLx2d} place a rough upper bound on variability in 2D;
also see Fig.~\ref{fig:energyAndFluxVsPert3d}).
Nevertheless, despite these differences, the particle energy spectra are similar despite different energy evolution (Fig.~\ref{fig:variability3d}, right), with the exception of one case exhibiting a bump above $\gamma \gtrsim 10\sigma$.
At early times, the evolutions of magnetic energy and flux (Fig.~\ref{fig:variability3d}, left and middle) are reasonably similar, but they diverge after $\simeq 5L_x/c$.
Simulation (A) starts off with the highest magnetic conversion rate but ends up converting the least amount (30~per~cent) of magnetic energy after $30L_x/c$.
Simulation (B) rapidly converts about 40~per~cent of the magnetic energy within $10L_x/c$, and then continues to convert energy at a much slower rate for at least the next $20L_x/c$.
And simulation (C) exhibits a medium-slow stage of energy conversion between~5 and~15$L_x/c$, followed by a stagnant or very slow final stage similar to the others.
One mechanism that can result in large random variation in 3D is the following. We will show in~\S\ref{sec:RDKIamp} that in some cases RDKI causes the layer to kink to very large amplitudes (comparable to the kink wavelength).
At some amplitude, the instability becomes highly non-linear, resulting in severe distortion of the current sheet, somewhat resembling breaking waves; when this happens, magnetic energy is rapidly depleted, leaving a much-thickened layer that is much more stable against tearing and RDKI.
This process rapidly transforms the layer, putting the simulation in a very different state.
Some initial configurations (e.g, with higher $\beta_d$, as we will see in~\S\ref{sec:RDKIamp}) seem to make this transformation more likely, but there appears to be a large range of conditions under which it is possible but not inevitable.
Thus two ``identical'' simulations, one that suffers this transformation and one that does not, may evolve very differently over long times (e.g., $50L_x/c$).
In light of this, we try to be careful in this study about not drawing conclusions from small differences between individual 3D simulations.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.325\textwidth]{fig21a.pdf}%
\hfill
\includegraphics*[width=0.325\textwidth]{fig21b.pdf}%
\hfill
\includegraphics*[width=0.325\textwidth]{fig21c.pdf}%
}
\caption{ \label{fig:variability3d}
3D reconnection exhibits significant stochastic variability, as
shown by three simulations (A, B, and~C) with $L_x=256\sigma\rho_0$ and $L_z=L_x/4$, with
otherwise identical parameters, but
different initial random particle positions and velocities.
Panels show:
transverse magnetic energy (left) and unreconnected flux (middle) versus time, and
(right) particle energy spectra at $20L_x/c$, compensated by
$\gamma^4$.
}
\end{figure}
\subsection{3D reconnection: dependence on initial current sheet configuration}
\label{sec:pertAndEta3d}
We showed in~\S\ref{sec:pertAndEta2d} that an initial magnetic perturbation barely affects energy conversion rates and NTPA in 2D reconnection, despite some differences in the evolution of the plasmoid hierarchy \citep[but see][and our comparison to it at the end of~\S\ref{sec:pertAndEta2d}]{Ball_etal-2018}.
Similarly, other parameters describing the initial current sheet can vary over a wide range without much affecting 2D reconnection.
Here, we explore a range of different current sheet configurations, as in~\S\ref{sec:pertAndEta2d}, but now in 3D, all the while keeping the upstream/background plasma the same.
In the substudies (3D-a,b,c) below, we will vary the same parameters as in~\S\ref{sec:pertAndEta2d}, (2D-a,b,c).
However, the range of varied parameters will be much more limited because of the increased cost of simulation in 3D; for instance, we examine only two different perturbation strengths in (3D-a) and two different values of $\eta$ in (3D-b).
As in~\S\ref{sec:pertAndEta2d}, we set $B_{gz}=0$.
We will see that the initial current sheet has more long-term influence in 3D than in 2D.
\emph{(3D-a) Varying the initial perturbation.}
In this substudy we will see that, in 3D,
an initial magnetic field perturbation (uniform in $z$) significantly alters the subsequent evolution---unlike in 2D, where the initial perturbation is relatively unimportant, as shown in~\S\ref{sec:pertAndEta2d}(2D-a).
Perhaps the easiest way to describe the effect of a perturbation in 3D is that (in simulations with no guide magnetic field) the perturbation tends to make reconnection more like 2D reconnection.
Figure~\ref{fig:energyAndFluxVsPert3d} shows energies and
flux versus time, for perturbed and unperturbed initial current sheets, in a fully
3D ($L_z=L_x$) simulation and---for comparison---in a nearly 2D ($L_z=L_x/16$) simulation---all with $L_x=341\sigma\rho_0$, $\eta=5$, $\beta_d=0.3$, and $B_{gz}=0$.
We consider only one non-zero perturbation strength, $a=12$, for which the initial magnetic separatrix height extends to about $s=6\delta = 4\sigma\rho_0\approx L_x/85$, beyond the current sheet but still very far from $L_y/4$.
To give an idea of stochastic variability, we show two simulation runs for the nearly 2D case without perturbation, and two for the 3D case with perturbation.
As expected from the 2D simulations in~\S\ref{sec:pertAndEta2d}(2D-a), the nearly 2D cases are almost identical in all three panels---the total transverse magnetic energy $U_{Bt}(t)$, the magnetic energy in the layer $U_{Bt,\rm layer}(t)$, and the upstream flux $\psi(t)$; the perturbation has no significant effect on energy or flux evolution in 2D.
In contrast, the 3D cases with $s=0$ and $s/\delta=6$ differ markedly; the $s=0$ case initially converts magnetic energy more slowly, with a correspondingly slower decline in unreconnected flux.
Interestingly, the early energy evolution in 3D
\emph{with} perturbation ($s/\delta=6$)
is very similar to the nearly 2D cases (\emph{with} and \emph{without} perturbation).
After$\:\sim 3L_x/c$, the initially-perturbed 3D simulation starts to exhibit slower energy conversion (compared with earlier times or with the 2D case at times up to $t\lesssim 6L_x/c$), as shown in Fig.~\ref{fig:energyAndFluxVsPert3d}(left).
Whereas the \emph{unperturbed} 3D case never has much magnetic energy $U_{Bt,\rm layer}$ in the layer,
the \emph{perturbed} 3D simulation builds up energy in $U_{Bt,\rm layer}$ until~$3L_x/c$, just as in 2D (Fig.~\ref{fig:energyAndFluxVsPert3d}, middle).
After~$3L_x/c$, the perturbed 3D case continues to deplete magnetic energy by converting \emph{upstream} magnetic field energy to plasma energy and also by converting~$U_{Bt,\rm layer}$ to plasma energy.
The amount of unreconnected flux $\psi(t)$ in the perturbed 3D case also follows the 2D cases up until~$3L_x/c$; after that the decline of $\psi$ slows substantially in 3D but continues apace in 2D for another $~3L_x/c$.
Although not shown here, the total magnetic energy continues to fall after~$10L_x/c$ in both 3D cases; for example it falls to~67~per~cent of its initial value after~$25L_x/c$ for the perturbed simulation and after~$50L_x/c$ for the unperturbed 3D case (and continues to drop further).
The uniform-in-$z$ initial perturbation thus appears to kick-start reconnection in 2D mode, creating long flux ropes that, being more uniform in $z$, are more stable; eventually 3D effects appear, however, and the plasmoid structures decay along with their associated magnetic fields.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.325\textwidth]{fig22a.pdf}%
\hspace{0.05in}%
\includegraphics*[width=0.325\textwidth]{fig22b.pdf}%
\hspace{0.05in}%
\includegraphics*[width=0.325\textwidth]{fig22c.pdf}%
}
\caption{ \label{fig:energyAndFluxVsPert3d}
In 3D a perturbation (which is uniform in $z$) makes simulations behave more like 2D simulations, especially at early times.
These plots show energies and flux versus time for four simulation configurations:
with perturbation ($s/\delta=6$, dashed lines) and without ($s=0$, solid), for 3D ($L_z=L_x$, blue) and nearly 2D ($L_z/L_x=1/16$, red) simulations.
All have $L_x=341\sigma\rho_0$, $\eta=5$, $\beta_d=0.3$, $B_{gz}=0$.
For two of the configurations---$L_z/L_x=1$, $s=0$ and $L_z=1/16,s/\delta=6$---two runs are shown (identical except for initially randomized particles).
The transverse magnetic energy, $U_{Bt}(t)/U_{B0}$ (left), is very similar for the 3D simulation with perturbation and both 2D simulations, but the 3D simulation without perturbation converts magnetic energy much more slowly.
The magnetic energy in the layer, $U_{Bt,\rm layer}(t)/U_{B0}$ (middle), however, shows that the 3D case with perturbation initially resembles the 2D simulations, but after~$3L_x/c$ depletes the ``reconnected'' magnetic field energy in the layer.
Similarly, the unreconnected magnetic flux $\psi(t)/\psi_0$ (right) shows that the 3D case with perturbation lies between the 2D cases and the 3D case without perturbation.
}
\end{figure}
The electron energy spectra are fairly similar for all cases,
as shown in Fig.~\ref{fig:ntpaVsPert3d}, especially when comparing spectra at the same amount of total energy depleted ($U_{Bt}=0.75U_{B0}$).
The 3D simulations have perhaps slightly more efficient NTPA than the nearly 2D simulations (e.g., they have more particles with $\gamma \approx 10\sigma$), but this enhancement might be attributable to stochastic variation.
The high-energy spectrum for the unperturbed 3D simulation grows more
slowly (than either the perturbed 3D simulation or 2D simulations), consistent with the slower energy conversion, and this is reflected in less NTPA at~$t=9L_x/c$ relative to the other cases.
Similarly, the perturbed 3D simulations have converted more magnetic energy to plasma energy at $t=9L_x/c$ than the other cases, and they correspondingly show more particles at high energies.
Differences in the high-energy spectra are magnified by plotting $\gamma^4f(\gamma)$; although the perturbed 3D cases appear to show significantly more NTPA than the 2D or unperturbed 3D cases, the difference is less than a 10~per~cent change in power-law index $p\approx 4$.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig23a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig23b.pdf}%
}
\caption{ \label{fig:ntpaVsPert3d}
Electron energy spectra (compensated by $\gamma^4$)
for the same (perturbed/unperturbed, 2D/3D) simulations as in Fig.~\ref{fig:energyAndFluxVsPert3d},
are all fairly similar at the time
when~$U_{Bt}/U_{B0}=0.75$ (left)---the differences at high energy would hardly be remarkable were it not for the $\gamma^4$.
Spectra are also shown (right)
at $t= 9L_x/c$, at which time simulations that have converted more magnetic field energy have more high-energy particles.
For comparison, the grey/dashed line shows the initial spectrum at $t=0$ in all simulations;
and, a power-law segment with index $p=3.6$ (green/dotted) demonstrates a ten~per~cent (i.e., small) difference from $p=4$ (which would be horizontal).
We note, since Fig.~\ref{fig:energyAndFluxVsPert3d} shows only the first $10L_x/c$, that the 3D case with $s=0$ does not reach $U_{Bt}(t)=0.75U_{B0}$ until
about $t=20L_x/c$, while all the other simulations convert $0.25U_{B0}$ in about $5L_x/c$.
}
\end{figure}
To avoid the early similarities between 2D and 3D with an initial perturbation, simulations discussed in the rest of~\S\ref{sec:3d} will have zero initial perturbation.
\emph{(3D-b) Varying current sheet density $\eta$ with fixed $\beta_d=0.3$; i.e., varying $\delta/\sigma\rho_0$ for fixed $\delta/\rho_d=2.1$}.
Varying the overdensity $\eta=n_{d0}/n_{b0}$, and fixing $\beta_d=0.3$ while maintaining the Harris equilibrium,
was seen to have relatively little effect on 2D reconnection in large systems, as long as $\eta \gg 0.1$ (see~\S\ref{sec:pertAndEta2d})---with the notable exception that decreasing $\eta$ (increasing $\delta/\sigma\rho_0$) merely delayed reconnection onset.
As noted in~\S\ref{sec:pertAndEta2d} (also see table~\ref{tab:etaEffect}), the initial gyroradius $\rho_d$ and skin depth $d_{ed}$ of the current sheet plasma, as well as its half-thickness $\delta$, all vary as $\eta^{-1}$ when the upstream $n_{b0}$ and $\theta_{b}$ are kept constant (hence associated time scales also vary as $\eta^{-1}$, explaining the onset delay).
As $\eta$ varies, the thickness $\delta$ thus varies relative to upstream plasma scales, but not relative to the drifting plasma scales.
Here we examine overdensities $\eta=1$ and $\eta=5$, in 3D ($L_z=L_x$) simulations with $L_x=256 \sigma\rho_0$, $\beta_d=0.3$, $B_{gz}=0$, and zero initial perturbation.
The choice of zero initial perturbation differs from~\S\ref{sec:pertAndEta2d}(2D-b), and is motivated by (3D-a), which showed that an initial perturbation can suppress 3D effects, at least temporarily.
Figures~\ref{fig:energyAndFluxVsEta3d} and~\ref{fig:ntpaVsEta3d} compare
energy and flux versus time, and resulting NTPA, for 3D simulations with
dense [$\eta=5$, $\delta=(2/3)\sigma\rho_0$, $\theta_d=0.10\sigma$]
and less dense [$\eta=1$, $\delta=(10/3)\sigma\rho_0$, $\theta_d=0.52\sigma$] initial
current sheets; for comparison, nearly 2D simulations are shown for the same parameters as the 3D simulations, except $L_z=L_x/16$.
To estimate stochastic variability in 3D, we show the results of three simulations with $\eta=5$ and $L_z=L_x$ (but just one run each for $\eta=1$ and both nearly-2D cases).
The total transverse magnetic energy (Fig.~\ref{fig:energyAndFluxVsEta3d}, left) shows that energy conversion is generally slower in 3D than in 2D, as expected (e.g., averaged over the first 10$L_x/c$). Also, energy conversion is slower for $\eta=1$ than for $\eta=5$.
However, the difference between the $\eta=1$ and $\eta=5$ cases is roughly the same in 3D as it is in 2D; and we saw in~\S\ref{sec:pertAndEta2d} that in 2D this difference vanishes for larger systems.
Therefore, any effect of $\eta$ could be entirely caused by the small system size; there is no indication that $\eta$ has a different effect in 3D than in 2D.
Most telling, the magnetic energy in the layer
(Fig.~\ref{fig:energyAndFluxVsEta3d}, middle) is nearly independent
of $\eta$ in 3D; i.e., varying $\eta$ does not appear to enhance or inhibit 3D effects significantly.
Although we did not explore the effect of $\eta$
on 3D simulations with an initial perturbation, we expect that a perturbation would make 3D simulations more closely resemble 2D simulations, which have at most a weak dependence on $\eta$.
The decay of upstream flux $\psi(t)$ (Fig.~\ref{fig:energyAndFluxVsEta3d}, right) also indicates no $\eta$-dependence in 3D beyond what exists in 2D.
The 2D cases have very similar $\psi(t)$; $\psi(t)$ is also similar for the 3D cases, which exhibit slower decay than the 2D cases (e.g., averaged over the first~$10L_x/c$).
As with total magnetic energy, the decay of upstream flux continues over much longer times than in 2D, at a slower pace than at early times.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.325\textwidth]{fig24a.pdf}%
\hspace{0.05in}%
\includegraphics*[width=0.325\textwidth]{fig24b.pdf}%
\hspace{0.05in}%
\includegraphics*[width=0.325\textwidth]{fig24c.pdf}%
}
\caption{ \label{fig:energyAndFluxVsEta3d}
As in 2D, varying overdensity $\eta$ in 3D does not dramatically affect the overall energy and flux evolution (differences shown here are attributable to finite system size).
These plots show energies and flux versus time for four simulation configurations:
$\eta=1$ (solid lines) and $\eta=5$ (dashed),
3D ($L_z/L_x=1$, blue) and nearly 2D ($L_z/L_x=1/16$, red).
All have $L_x=256\sigma\rho_0$, $\beta_d=0.3$, $B_{gz}=0$, and $a=0$.
Three different runs are shown for the 3D, $\eta=5$ case.
The magnetic energy evolution $U_{Bt}(t)$ (left) shows slower energy conversion in 3D than in 2D as expected; but the difference between $\eta=1$ and $\eta=5$ in 3D is comparable to the difference in 2D (also it is comparable to the stochastic variation among the three 3D, $\eta=5$ cases).
The difference in 2D vanishes in larger systems (see~\S\ref{sec:pertAndEta2d}, 2D-b).
The magnetic energy in the layer, $U_{Bt,\rm layer}(t)$ (middle), evolves similarly in all the 3D cases, very unlike the 2D cases, which store significant energy in ``reconnected'' field.
Similarly, the unreconnected flux $\psi(t)$ (right) is similar for all 3D cases, which differ greatly from the nearly identical 2D cases.
}
\end{figure}
Although it appears that $\eta=1$ current sheets result in more
efficient NTPA (Fig.~\ref{fig:ntpaVsEta3d}), we note again that
the difference in 3D between $\eta=1$ and $\eta=5$ is almost the same as the difference in 2D (a factor of~$\sim 2$ in particle energy).
This is especially clear when comparing particle energy spectra at the same amount of magnetic energy converted (namely, when $U_{Bt}/U_{B0}=0.85$).
In fact, for each $\eta$, the spectra from corresponding 2D and 3D simulations are nearly the same, aside from a small burst of particles accelerated up to $\gamma\simeq 10\sigma$ for the 2D, $\eta=1$ case; however, this burst is probably the result of some randomness that would be averaged out in a larger simulation---moreover, the ``burst'' is less than a factor of 2 in energy, and $f(\gamma)$ is already very low by such high energies.
Thus we conclude that the effects of $\eta$ are no different in 3D than in 2D; and we note that the effects of $\eta$ diminish in 2D for larger systems (\S\ref{sec:pertAndEta2d}, 2D-b).
To summarize: considering energy and particle acceleration, we detect no
differences between $\eta=1$ and $\eta=5$ that are clearly linked to 3D effects. However, an ensemble of larger 3D simulations may be needed to rule out significant 3D effects of $\eta$.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig25a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig25b.pdf}%
}
\caption{ \label{fig:ntpaVsEta3d}
Varying overdensity $\eta$ does not have any more effect on NTPA in 3D than it has in 2D; and in 2D, the effect vanishes for larger system sizes (see~\S\ref{sec:pertAndEta2d}, 2D-b).
These plots show the electron energy spectra (compensated by $\gamma^4$)
for the same simulations ($\eta\in \{1,5\}$, 2D/3D) as in Fig.~\ref{fig:energyAndFluxVsEta3d}, at the time
when~$U_{Bt}(t)/U_{B0}=0.85$ (left), and
at $t=10L_x/c$ (right). The grey/dashed line shows the initial spectrum at $t=0$ in all simulations.
}
\end{figure}
\emph{(3D-c) Varying $\beta_d\propto \eta^{-1}$; i.e., varying $\delta/\rho_d$ for fixed $\delta/\sigma\rho_0=(10/3)\sigma\rho_0$.}
When we vary $\eta\propto \beta_d^{-1}$,
we see dramatic differences in the evolution of 3D current sheets, although NTPA will turn out to be similar in all cases.
As described in \S\ref{sec:pertAndEta2d}(2D-c), changing $\beta_d$ and $\eta$ together, to maintain a constant product $\eta \beta_d=0.3$, keeps $\delta=\sigma\rho_0/(\eta \beta_d)$ constant (with respect to upstream plasma scales), while $\delta/\rho_d=(2/3)(\beta_d^{-2}-1)^{1/2}$ varies.
This allows us to study the evolution of different current sheet configurations while ensuring that the initial current sheet is always well resolved ($\delta/\Delta x = 10$) and of constant size relative to the simulation box [$(L_y/4)/\delta = 39$].
We consider cases that, in large 2D simulations in~\S\ref{sec:pertAndEta2d}(2D-c), yielded essentially similar results;
in 3D we use smaller systems, $L_z=L_x=256 \sigma\rho_0$,
but as in (2D-c), $B_{gz}=0$ and there is no initial perturbation.
We scan over
$\beta_d \in \{0.075$, 0.15, 0.3, 0.6$\}$, which implies $\eta \in \{4$, 2, 1, 0.5$\}$, respectively;
corresponding parameters (see table~\ref{tab:etaEffect}) are
$\delta/\sigma\rho_0=10/3$,
$\delta/\rho_d \in \{$8.9, 4.4, 2.1, 0.89$\}$, and
$\theta_d/\theta_b \in\{$0.50, 1.0, 2.1, 5.0$\}$.
For comparison, we also show 2D simulations with identical set-ups except for $L_z=0$.
Looking at overall characteristics, we notice first (in Fig.~\ref{fig:energyAndFluxVsBetad3d}, left) that the amount and rate of magnetic energy conversion is almost identical for the 2D simulations, but varies drastically with $\beta_d$ in 3D.
Correcting for the difference in reconnection onset times, we see that over the first few~$L_x/c$ after onset, the rate of magnetic energy conversion generally increases with~$\beta_d$.
For $\beta_d=0.6$, the rate is even faster than in 2D; for $\beta_d=0.3$ and $\beta_d=0.15$, it is a little slower than in 2D; and for $\beta_d=0.075$, the initial rate is the slowest of all.
After a few~$L_x/c$, all 3D cases exhibit a slower stage of energy conversion;
however,
the behaviour becomes more complicated, depending non-monotonically on~$\beta_d$.
(It might not be useful to try to ascribe a direct $\beta_d$-dependence to the later time evolution. The dependence could involve an element of randomness, and it could depend on the state of the system after early evolution that did strongly depend on $\beta_d$. For example, it often happens that a simulation that converts more energy at early times will convert energy more slowly at later times---for example, comparing $\beta_d=0.6$ and $\beta_d=0.15$, or comparing $\beta_d=0.3$ and $\beta_d=0.075$.)
The~$\beta_d=0.6$ (3D) case drastically slows down after~$2L_x/c$, almost as in 2D, except that, being 3D, it can convert somewhat more magnetic energy than in 2D.
The $\beta_d=0.3$ case also slows fairly drastically (though not quite as much as $\beta_d=0.6$) and as a result converts much less energy over~$20L_x/c$.
In contrast, the $\beta_d=0.15$ case slows a bit, but maintains a more moderate magnetic energy conversion rate, and catches up to the $\beta_d=0.6$ case in terms of total magnetic energy conversion after~$\simeq 15L_x/c$.
The $\beta_d=0.075$ case goes from slow to slower, and, like $\beta_d=0.3$, does not convert as much energy over~$20L_x/c$ as the other cases.
The magnetic energy in the layer (Fig.~\ref{fig:energyAndFluxVsBetad3d}, middle) is again almost identical for all the 2D cases.
In 3D, however, we see some differences.
The $\beta_d=0.6$ case, which has such rapid energy conversion at very early times, shows a 2D-like increase for~$\simeq 1L_x/c$, before decaying in a characteristically 3D manner.
The $\beta_d=0.15$ case shows a smaller, slower rise before decaying, while the other cases ($\beta_d=$0.075, 0.3) never have much magnetic energy in the layer.
Despite the differences, all 3D cases strongly contrast with the 2D cases, which store significant amounts of magnetic energy in the layer. Thus 3D cases can convert more magnetic energy overall even while the upstream flux is less depleted.
The upstream magnetic flux evolution (Fig.~\ref{fig:energyAndFluxVsBetad3d}, right) is consistent with the overall magnetic energy depletion. As usual, more upstream flux must be reconnected in 2D simulations to convert the same amount of magnetic energy, because in 2D much more of the upstream magnetic energy ends up stored in plasmoids. As we discussed above, in 2D, flux is very nearly conserved, with an increase in reconnected flux corresponding precisely to the decrease in upstream flux~$\psi$. In 3D, however, the lost upstream flux is eventually mostly annihilated (although some of it may first be reconnected before being subsequently annihilated).
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.325\textwidth]{fig26a.pdf}%
\hspace{0.05in}%
\includegraphics*[width=0.325\textwidth]{fig26b.pdf}%
\hspace{0.05in}%
\includegraphics*[width=0.325\textwidth]{fig26c.pdf}%
}
\caption{ \label{fig:energyAndFluxVsBetad3d}
Varying $\beta_d$ while varying $\eta$ (to hold $\eta\beta_d=0.3$ constant) dramatically alters reconnection in 3D, but not in 2D.
These plots show energies and flux versus time, for eight simulations:
$\beta_d\in\{$0.075,0.15,0.3,0.6$\}$ (hence $\eta\in\{4$,2,1,0.5$\}$), for both $L_z=L_x$ (3D) and $L_z=0$ (2D).
All simulations have $L_x=256\sigma\rho_0$, $B_{gz}=0$, and $a=0$.
The magnetic energy, $U_{Bt}(t)$ (left), is very similar for all 2D cases, but the 3D cases differ substantially in terms of rates and amounts of magnetic energy conversion.
The magnetic energy in the layer, $U_{Bt,\rm layer}(t)$ (middle), is again similar for all 2D cases; the 3D cases all exhibit the characteristic depletion of reconnected field energy, but differ by small amounts at early times.
The unreconnected flux $\psi(t)$ (right) qualitatively resembles the magnetic energy, except that the 2D simulations reconnect relatively more flux.
}
\end{figure}
Particle energy spectra are shown in Fig.~\ref{fig:ntpaVsBetad3d}---both
at times when 15~per~cent of the initial magnetic energy has been lost (left panel),
and at the same time, $t=10\: L_x/c$ (right).
All cases exhibit fairly similar NTPA, although when the spectra
are compensated by~$\gamma^4$, differences are quite apparent.
The trend is non-monotonic with the input parameters $\beta_d$ and $\eta$, which is not surprising given the non-monotonic dependence of energy conversion on these parameters.
There is evidently a good deal of stochastic variation in these cases, and
more work needs to be done to settle this issue.
However, given that the time over which 15~per~cent of the magnetic energy is converted differs by a factor of~$\simeq 10$ for 3D simulations with $\beta_d=0.6$ and $\beta_d=0.075$, the energy spectra are remarkably similar.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig27a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig27b.pdf}%
}
\caption{ \label{fig:ntpaVsBetad3d}
Although varying $\beta_d$ with $\eta$ (holding $\eta\beta_d$ constant) dramatically affects energy conversion versus time, NTPA is less affected, especially when compared at the same $U_{Bt}/U_{B0}$.
These plots show the electron energy spectra (compensated by $\gamma^4$)
for the same simulations ($\beta_d\in \{$0.075, 0.15, 0.3, 0.6$\}$, 2D/3D) as in Fig.~\ref{fig:energyAndFluxVsBetad3d} at the time
when~$U_{Bt}/U_{B0}=0.85$ (left), and
at $t=10L_x/c$ (right). For comparison, the grey/dashed line shows the initial spectrum at $t=0$; and, a power-law segment with $p=4.6$ (green/dotted) demonstrates a 15~per~cent (i.e., small) difference from $p=4$ (which would be horizontal).
}
\end{figure}
The substantial differences in energy and flux depletion (in 3D simulations with $\beta_d=0.6$ and $\beta_d=0.075$) are very intriguing, and worthy of discussion in a separate section,~\S\ref{sec:RDKIamp}.
Before that, we briefly summarize the results of this subsection.
\emph{Summary of 3D-a,b,c.}
In contrast to 2D, we have seen here and in~\S\ref{sec:variability3d} (and will see in more detail in~\S\ref{sec:RDKIamp}) that current sheets in 3D display a greater variety of behaviours, and that the overall evolution is less inevitable, than in 2D. Rather, the current sheet evolution can take very different paths, depending on initial conditions and stochastic variability. For example, simulations with identical upstream conditions can convert rather different amounts of magnetic energy over~$\sim 20L_x/c$, whereas 2D simulations with a wide variety of initial current sheets convert nearly the same magnetic energy when averaged over~$\sim 2L_x/c$ time windows, despite some random variability on shorter timescales (e.g., Figs.~\ref{fig:perturb2dLx320}, \ref{fig:eta2dLx1280}, \ref{fig:betad2dLx1280}, \ref{fig:energyFluxVsLx2d}).
An initial magnetic field perturbation, uniform in the $z$ dimension, triggers time evolution resembling 2D reconnection at early times; however, at later times when 2D reconnection would slow and approach a final stable magnetic configuration with substantial magnetic energy in plasmoids, 3D simulations generally deplete the magnetic energy in the layer (as well as in the upstream).
The effect of varying the overdensity $\eta=n_{d0}/n_{b0}$ for fixed $\beta_d$ (hence varying $\delta/\sigma\rho_0$ but constant $\delta/\rho_d=2.1$) is less clear, hard to distinguish amid stochastic variability and moderate system size, but our simulations did not show any effect of $\eta$ in 3D beyond what is observed in 2D (which is very minimal in large systems).
In contrast, varying $\beta_d\propto \eta^{-1}$ (hence varying $\delta/\rho_d$ while keeping $\delta/\sigma\rho_0=10/3$ fixed) had a pronounced effect on current sheet evolution, with larger $\beta_d$ leading to faster initial magnetic energy conversion and more conversion over $20L_x/c$---and we will explore this next.
Overall, global NTPA is surprisingly consistent given the variety of behaviours evolving from different initial current sheet configurations. This is especially true if particle spectra are compared at times when the same amount of magnetic energy has been converted to plasma energy.
\subsection{3D reconnection: the effect of varying $\beta_d$ on RDKI amplitude}
\label{sec:RDKIamp}
The linear and nonlinear stages of RDKI growth were investigated in detail in \cite{Zenitani_Hoshino-2007,Zenitani_Hoshino-2008}, both in theory and in 2D and 3D simulation, including the effect of RDKI on NTPA. However, by exploring a range of initial parameters, specifically a range of initial drift speeds, $\beta_d c$, we find that the evolution from the linear to the nonlinear stage can have varied and often important consequences for later evolution.
Varying $\beta_d\propto \eta^{-1}$, as in~\S\ref{sec:pertAndEta3d}(3D-c), to keep $\eta\beta_d=0.3$ constant, hence fixing $\delta/\sigma\rho_0=10/3$, leads to dramatic differences in current sheet evolution, including differences in the evolution of global quantities such as magnetic energy.
Here, we show these differences in more detail, and attribute them to the maximum amplitude to which RDKI can grow before becoming nonlinearly saturated. In the following, $\Delta y_c$ is the amplitude of the rippling of the current sheet central surface, $y_c(x,z)$ (see~\S\ref{sec:sheetCenter}).
We start by looking at the same 3D simulations run for~\S\ref{sec:pertAndEta3d}(3D-c), with common parameters $L_z=L_x=256\sigma\rho_0$, $\eta\beta_d=0.3$, $\delta/\sigma\rho_0=10/3$, $B_{gz}=0$, and zero initial perturbation.
We focus on the two extreme cases: $\beta_d=0.6$ ($\eta=0.5$, $\delta/\rho_d=0.89$), which shows very rapid, substantial magnetic energy conversion, followed by very slow, limited energy conversion; and $\beta_d=0.075$ ($\eta=4$, $\delta/\rho_d=8.9$), which shows slower energy conversion and less conversion overall after $20L_x/c$ (see table~\ref{tab:etaEffect} for values of other relevant parameters).
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=\textwidth]{fig28a.pdf}%
}
\caption{ \label{fig:kinkAmpVsBetad3d}
$B_x(y,z)$ in the $x=0$ plane for 3D ($L_z=L_x$) for two cases, (top row) $\beta_d=0.6$, $\eta=0.5$ and (bottom row) $\beta_d=0.075$, $\eta=4$, each at three times, when the total magnetic energy $U_B$ has fallen to fractions~0.998, 0.97, and 0.89 of its initial value $U_{B0}$.
For large~$\beta_d$ (top) the kink wavelength is much longer than the current sheet thickness, and the sheet ripples with large amplitude $\Delta y_c$ while converting very little magnetic energy to plasma energy, until nonlinear development causes the sheet to fold over on itself, rapidly converting magnetic energy~$0.08U_{B0}$ over a time~$0.4L_x/c$.
For small~$\beta_d$ (bottom) the kink wavelength is short, and~$\Delta y_c$ does not increase much beyond the original sheet thickness, resulting in slower magnetic energy conversion ($0.08U_{B0}$ over $5.4L_x/c$).
}
\end{figure}
The reason for these differences is illustrated in Fig.~\ref{fig:kinkAmpVsBetad3d}, which shows $B_x(0,y,z)$ in the $x=0$ plane at the three different times when 0.2, 3, and 11~per~cent of the initial magnetic energy has been converted to plasma energy, for the $\beta_d=0.6$ case (top) and $\beta_d=0.075$ (bottom).
For $\beta_d=0.6$, the layer kinks with long wavelength ($\lambda_z \gg \delta$) and the RDKI amplitude $\Delta y_c$ can grow to a relatively large value, $\Delta y_{c*} \sim\lambda_z$, before becoming highly nonlinear;
the nonlinear development causes the highly-distorted layer to fold over on itself, rapidly depleting magnetic energy in a layer of thickness~$\sim \lambda_z \gg \delta$.
It is this process, and not 2D-like reconnection (perhaps surprisingly, given that the energy conversion versus time resembles that in 2D reconnection), that results in rapid energy dissipation at early times.
In contrast, for $\beta_d=0.075$ (Fig.~\ref{fig:kinkAmpVsBetad3d}, bottom), the initial sheet kinks with short wavelength ($\lambda_z \lesssim \delta$); since $\Delta y_c$ cannot grow much beyond $\lambda_z$, the instability saturates before it can distort the sheet much (compared with $\delta$).
In this case, the nonlinear RDKI can deplete magnetic energy only within a layer of thickness $\sim \lambda_z \lesssim \delta$.
\citet{Zenitani_Hoshino-2007} derived the most unstable wavelength for RDKI in relativistic plasma \citep[using two-fluid theory with an assumption that the current-carrying particles execute Speiser orbits within a sublayer of half-thickness $\sqrt{\rho_d \delta}$, and ignoring the background plasma; an improved approximation for $\beta_d\gtrsim 0.8$ can be found in][]{Hoshino-2020arxiv}: $\lambda_{z,\rm RDKI} = 16\upi \gamma_d \beta_d^2 \delta = (32 \upi/3)\beta_d \rho_d$, which can be significantly larger than $\rho_d$.
For $\beta_d=0.6$, $\lambda_{z,\rm RDKI}\approx 23\delta = 75 \sigma\rho_0$ (in Fig.~\ref{fig:kinkAmpVsBetad3d}, we see $\lambda_z\approx 50 \sigma\rho_0$).
For $\beta_d=0.075$, $\lambda_{z,\rm RDKI}\approx 0.3 \delta = \sigma\rho_0$ (in Fig.~\ref{fig:kinkAmpVsBetad3d}, we estimate very roughly, $\lambda_z\sim 7 \sigma \rho_0$).
While the measured instability wavelength does not precisely match the
theory, the qualitative behaviour is in agreement, and in any case, there
is a strong correlation between $\lambda_z/\delta$ and the rate and amount of magnetic energy conversion.
In both these simulations we find that the maximum $\Delta y_c$ is on the order of half the wavelength in the $z$-direction.
For the $\beta_d=0.6$ case, the kink grows roughly to a maximum $\Delta y_{c*} \approx 20\sigma\rho_0 \approx 0.4 \lambda_{z}$; for the $\beta_d=0.075$ case, $\Delta y_{c*} \approx 3\sigma\rho_0 \approx 0.4 \lambda_z$.
Given the uncertainty in these measurements, we might as well say $\Delta y_{c*} \approx \lambda_z/2$ for simplicity.
We note that the largest kink amplitudes (for $\beta_d=0.6$) are a fraction $\Delta y_{c*}/(L_y/4)\lesssim 0.2$ of the global system; this fraction is sizeable, but not so close to 1 that we need to worry about the lower and upper current layers interacting.
We hypothesize that in 3D, RDKI at some wavelength $\lambda_{z}$ grows linearly, without substantial magnetic energy conversion or NTPA due to RDKI, until it nears a critical amplitude, $\Delta y_{c*} \sim 0.5 \lambda_{z}$.
If the kink mode grows beyond $\Delta y_{c*}$, RDKI becomes highly nonlinear and the current layer severely distorts, folding over on itself and rapidly depleting magnetic energy within $|y| < \Delta y_{c*}$.
This yields a turbulent layer of thickness $\simeq 2\Delta y_{c*}$ with greatly diminished magnetic field.
After this time, magnetic energy continues to be converted to plasma energy, but at a much slower rate because of the much thicker layer.
Importantly, RDKI does not inevitably grow to $\Delta y_{c*}$; depending on the initial layer parameters and also on stochastic behaviour, it may not reach the nonlinear stage that severely distorts and rapidly transforms the initially-thin current sheet.
This phenomenon of extreme nonlinear kinking can occur even in a 2D simulation in the $y$-$z$ plane, where tearing (at least in the $x$-direction) is forbidden and reconnection cannot occur \citep{Zenitani_Hoshino-2007}.
Although a systematic examination of RDKI without reconnection in 2D$yz$ simulations is beyond the scope of this paper \citep[but see, e.g.,][]{Zenitani_Hoshino-2005a,Zenitani_Hoshino-2007,Zenitani_Hoshino-2008,Cerutti_etal-2014b},
we compare to one such 2D simulation, identical in set-up with $\beta_d=0.6$ and $\eta=0.5$ (but lacking any $x$-dependence).
There we see clear development of the kink instability growing to
large amplitudes and developing highly nonlinear behaviour in Fig.~\ref{fig:kinkAmp2dyz} (which looks similar to Fig.~\ref{fig:kinkAmpVsBetad3d}, for $\beta_d=0.6$).
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=\textwidth]{fig29a.pdf}%
}
\caption{ \label{fig:kinkAmp2dyz}
$B_x(y,z)$ for 2D ($y$-$z$ plane) simulations at two different times, $1.5 L_z/c$ and $2.3 L_z/c$, for an initial current sheet of thickness $\delta=(10/3)\sigma\rho_0$ with $\beta_d=0.6$ and $\eta=0.5$.
Without $x$-dependence, tearing and reconnection are forbidden, but RDKI is allowed.
The kink instability grows to large amplitudes, where the current layer becomes highly deformed (from the small-amplitude sinusoidal perturbation), folding over onto itself and depleting magnetic energy
(compare to Fig.~\ref{fig:kinkAmpVsBetad3d}).
}
\end{figure}
Indeed, this 2D$yz$ configuration rapidly depletes magnetic energy, at least when it reaches the extreme nonlinear stage where the current layer is massively distorted.
For comparison, in Fig.~\ref{fig:energyAndNTPAbetadp6}
we overplot results for the same set-up
[$\beta_d=0.6$, $\eta=0.5$, $\delta/\rho_d=0.89$, $\delta=(10/3)\sigma\rho_0$, $a=0$, $L_y=512\sigma\rho_0$]
but different dimensionalities: (1) 2D$xy$ with $L_x=L_y/2$ and $L_z=0$,
(2) 3D with $L_x=L_z=L_y/2$, and (3) 2D$yz$ with $L_z=L_y/2$ and $L_x=0$.
First, in panel (a), we see that magnetic energy is converted to particle energy more rapidly (presumably triggered by RDKI) in the 2D$yz$ and 3D simulations than by reconnection in 2D$xy$, but that all cases exhibit rapid magnetic energy conversion (a decline in $U_B$) of roughly similar amounts. E.g., they all convert about 20--25 per~cent of $U_{B0}$ in less than 5$\:L_y/2c$ after onset and ultimately convert 20--27 per~cent of $U_{B0}$ within 20$\:L_y/2c$ of onset; the initial energy conversion rate is somewhat slower for the 2D$xy$ case than for the others, while the total energy converted to plasma energy is slightly higher for the 3D simulation (most notably for $t\gtrsim 5L_y/2c$).
Looking at the energy $U_{By}$ stored in $B_y$ field components ($U_{By}$ is a proxy for $U_{Bt,\rm layer}$ in the 2D$xy$ and 3D cases), we see (panel b) that the $2Dyz$ simulation has almost no $U_{By}$ (as expected), while the 3D simulation develops a small $U_{By}$ that decays, and the 2D$xy$ simulation stores $U_{By}$ permanently.
Thus the 2D$xy$ simulation sees a larger decrease in upstream magnetic energy than other cases, but some of that energy is transformed to ``reconnected field'' energy rather than being converted to plasma energy.
The similarities among these three cases are striking, especially considering that the 2D$xy$ simulation exhibits classic reconnection, which is forbidden in~2D$yz$.
The 2D$xy$ case conserves flux, while the 2D$yz$ simulation directly annihilates upstream flux. We see clear suggestions of both these 2D behaviours in 3D---plasmoid formation and the growth of $U_{Bt,\rm layer}$ (as in the 2D$xy$ case) and rapid magnetic energy depletion due to nonlinear kinking of the current sheet (as in the 2D$yz$ case; see~\S\ref{sec:Lz3d} for more detail). Thus in 3D, we expect that classic reconnection occurs to some extent, converting upstream flux to flux around plasmoids (as in 2D$xy$), and that some upstream flux is directly annihilated (as in 2D$yz$); furthermore, there is a unique 3D effect, namely the dissipation of plasmoids and annihilation of the associated ``reconnected'' flux. Without much more extensive diagnostics, it is difficult to determine how much upstream flux is directly annihilated as opposed to reconnected and then annihilated, i.e., to determine precisely the extents to which the different 2D-like processes occur in 3D.
Interestingly, all these cases exhibit similar NTPA, despite the fact that very different mechanisms come into play in (at least) the 2D$xy$ and 2D$yz$ simulations.
Compensating $f(\gamma)$ by $\gamma^4$, however, we do see some differences.
If we compare the spectra at times when the magnetic energy has fallen by 16~per~cent (Fig.~\ref{fig:energyAndNTPAbetadp6}c), we find that NTPA is strongest for 2D$xy$, followed by 2D$yz$, and weakest for 3D; here, differences in $f(\gamma)$ reach an order of magnitude, but this is not as significant as it might seem, considering the steepness $f(\gamma)\sim \gamma^{-4}$---the spectra are separated by less than a factor of~2 in $\gamma$ (or energy).
Over time, the 3D case develops stronger NTPA, and the three cases look fairly similar; differences in $f(\gamma)$ are less than a factor of a few, which, again, is relatively little considering the steep slope.
At late times [e.g., around $11L_y/2c$, Fig.~\ref{fig:energyAndNTPAbetadp6}d], the 3D case has the hardest spectrum (shallowest slope).
\begin{figure}
\centering
\fullplot{
\begin{tabular}{cc}
\includegraphics*[width=0.48\textwidth]{fig30a.pdf}
\hfill &
\includegraphics*[width=0.48\textwidth]{fig30b.pdf}%
\\
\includegraphics*[width=0.48\textwidth]{fig30c.pdf}%
\hfill &
\includegraphics*[width=0.48\textwidth]{fig30d.pdf}%
\end{tabular}
}
\caption{ \label{fig:energyAndNTPAbetadp6}
A comparison of energy conversion and NTPA for 2D$xy$, 3D, and 2D$yz$ simulations with the same set-up and size, except for unsimulated dimensions. Upper left: (transverse) magnetic energy $U_{Bt}(t)$; upper right: energy $U_{By}(t)$, stored in $B_y$ field components, a rough indicator of the amount of magnetic energy stored in the layer; lower left: particle spectra (compensated by $\gamma^4$) when $U_{Bt}$ has fallen to $0.84U_{B0}$ (cf. the grey dash-dotted line in the upper left); lower right: spectra at $t=11.4\:(L_y/2)/c$. The 2D$xy$ simulations exhibit familiar 2D reconnection but not RDKI; the 2D$yz$ simulations can undergo RDKI but not reconnection.
}
\end{figure}
Whereas \citet{Zenitani_Hoshino-2007,Zenitani_Hoshino-2008} also observed RDKI becoming nonlinear, those works observed little NTPA due to RDKI (for zero guide field)---in the linear stage, but also specifically in the nonlinear stage.
In contrast, when a long-wavelength RDKI mode ($\lambda_{z} \gg \delta$) reaches the nonlinear stage, we observe much more rapid magnetic energy declines than in the linear stage, and the nonlinear stage generates significant NTPA.
Although we must leave a systematic exploration of 2D$yz$ simulations to future work, we note that 2D$yz$ simulations with lower $\beta_d$ often do not show the violent nonlinear growth of the kink instability at large amplitudes (as in 3D; see Fig.~\ref{fig:kinkAmpVsBetad3d}).
In these cases, less magnetic energy is converted overall, although slow magnetic energy conversion can continue for tens of light-crossing times (as in 3D), and there is often no significant NTPA at all.
When looking at 3D simulations with different $\beta_d$, and comparing them to 2D reconnection, the similarly rapid magnetic energy conversion in 2D reconnection and 3D simulations with large $\beta_d$ (small $\delta/\rho_d$) might seem to suggest that these 3D simulations exhibit 2D-like reconnection.
This is (mostly) not the case here. The 3D simulation with large $\beta_d=0.6$ has relatively long $\lambda_{z,\rm RDKI}$ and exhibits extreme nonlinear kinking that much more closely resembles the 2D$yz$ simulation;
it bears little resemblance to the 2D$xy$ simulation in terms of magnetic field structure.
While magnetic reconnection could be playing a role at small scales in the 3D simulation with $\beta_d=0.6$, its strong resemblance to the 2D$yz$ simulation suggests that reconnection may be unimportant at early times when magnetic energy conversion is dominated by the nonlinear RDKI.
If long-wavelength kink modes grow to large amplitudes and distort the current layer enough to suppress reconnection, in the process converting magnetic energy and producing NTPA, then we will need to understand how this extreme-kinking mechanism compares with large-scale reconnection (considering both 2D and 3D reconnection models).
If, as the $\beta_d=0.6$ simulation in this section suggests, extreme kinking results in similar NTPA and overall energy conversion (in terms of both rate and amount) as 2D reconnection, then this mechanism becomes a competing candidate to explain phenomena for which reconnection has been invoked as an explanation.
One important difference is that guide magnetic field is believed to curtail RDKI rather severely \citep{Zenitani_Hoshino-2008}.
Another important difference is that the layer distortion appears to be self-limiting in the following sense:
it depletes magnetic energy within a thickness $2\Delta y_c \sim \lambda_{z,RDKI}$, resulting in a much thicker layer no longer highly unstable to RDKI.
(The same might be said of the tearing instability, which would saturate when the plasmoids fed by an elementary reconnecting current sheet grow to a height in $y$ comparable to the elementary current sheet length in $x$; however, in a large reconnecting system, the plasmoids detach from the elementary current sheet before growing so large and coalesce with other plasmoids, thus always maintaining thin current sheets. Consequently tearing and reconnection do not saturate until the plasmoids become comparable to the system size.)
For our 3D simulations, $\lambda_{z,\rm RDKI}$ happens to be a significant fraction of the system size, and so this mechanism yields dramatic conversion of magnetic to plasma energy.
However, if the most unstable wavelength $\lambda_{z,\rm RDKI}$ is independent of system size, then in the limit of astronomically-large systems, this mechanism may convert a negligible fraction of available magnetic energy, while suppressing reconnection.
In contrast, 2D reconnection is self-limiting (in a closed system) only when plasmoids grow to a size $\sim L_x$ that scales with the system size. We believe reconnection would continue for an arbitrary time, given a large-enough system.
(This is perhaps a remarkable property of 2D reconnection---that the macroscopically-large plasmoids do not destroy the kinetic-scale current sheets where reconnection occurs.)
On the other hand, because the maximum kink amplitude depends on the wavelength, $\Delta y_{c*} \sim \lambda_{z}/2$, it might be a mistake to assume that the most prominent or influential wavelength corresponds to the RDKI mode with the highest growth rate.
The most unstable mode might saturate without thickening the layer much, while a more slowly-growing mode with a longer wavelength might eventually reach a much larger amplitude.
This possibility needs further study; however, we will show in the following section that some simulations (with $\beta_d=0.3$) initially kink on a scale $\lambda_z \ll L_z$, but nonetheless over time develop large-amplitude kinking on the scale of the system, $\lambda_z = L_z$.
Whether such behaviour can scale with system size remains an outstanding question.
If large-amplitude kink modes are not inevitable, as some of our simulations suggest, it may be that short-wavelength RDKI modes develop and rapidly saturate, before their amplitudes grow much past the current sheet thickness.
In this case, RDKI may have a negligible effect on reconnection;
although
it appears that reconnection in this regime is slow (though perhaps locally fast, as suggested in~\S\ref{sec:overview3d}), this slowness may not have anything to do with RDKI.
\subsection{3D reconnection: $L_z$-dependence}
\label{sec:Lz3d}
In this subsection, we systematically
vary $L_z$
to explore when and how 3D effects arise with increasing $L_z$.
We consider $L_z/L_x\in \{$0, 1/32, 1/16, 3/32, 1/8, 1/4, 1/2, 1, 3/2$\}$
(where $L_z=0$ means 2D) for system size $L_x=341\sigma \rho_0$,
$\eta=5$, $\beta_d=0.3$, $\delta=(2/3) \sigma \rho_0$, $B_{gz}=0$, and
zero initial perturbation.
We begin by looking at magnetic energy versus time in Fig.~\ref{fig:energyAndFluxVsLz} (we note that the total and transverse magnetic energy, $U_{B}$ and $U_{Bt}$, are nearly the same; in all these cases, $U_{Bz} < 0.004U_{B0}$).
In the left panel, we see that in 2D and nearly-2D (i.e., $L_z/L_x \ll 1$), $U_{Bt}(t)$ decays rapidly over $\sim 8 L_x/c$;
after this time, reconnection releases very little additional magnetic energy.
As $L_z$ increases, the initial rate of magnetic energy depletion
slows, but ultimately more magnetic energy is converted to
particle energy---the case $L_z=L_x/2$ is an outlier, transferring
almost half the initial magnetic energy to particles (compared with
less than $0.3U_{B0}$ for 2D).
In these large-$L_z$ cases, we often see a stage of
relatively fast magnetic energy conversion ($\lesssim 5 L_x/c$),
followed by a later stage of much slower energy conversion.
Even after $50L_x/c$, slow energy conversion continues.
It is important to remember that, in the large-$L_z$ (3D) regime, stochastic variability can be considerable (cf.~\S\ref{sec:variability3d}).
Because of high computational costs, we were unable to run multiple simulations for each value of $L_z$; e.g., determining whether the dramatic energy release is caused by the special value of $L_z/L_x=1/2$ or is merely a statistical outlier, will have to be left to future studies. In the meantime, we can have confidence only in clear, monotonic trends with $L_z$.
A possible reason for the increased stochastic variability in 3D was discussed in~\S\ref{sec:RDKIamp}, and we will explore this further at the end of this subsection.
The clearest signature of the effect of $L_z$
is the ``reconnected magnetic field energy''
(Fig.~\ref{fig:energyAndFluxVsLz}, middle), which (at any fixed time~$t$)
decreases nearly monotonically with $L_z/L_x$.
We interpret this as follows, as discussed previously
in~\S\ref{sec:overview3d}.
In 2D, reconnection results in magnetic energy being stored in plasmoids; in the fully 2D simulation, the magnetic energy in the layer (predominantly in plasmoids) thus grows monotonically.
In 3D, reconnection can pump magnetic energy into plasmoids, but these magnetic structures are unstable (at least in the absence of appreciable guide field): they decay, converting their magnetic energy over time into plasma energy.
The almost-2D simulation with $L_z=L_x/32$ pumps nearly as much energy into plasmoids as the fully 2D simulation, but after $10L_x/c$ this energy
starts to decay slowly.
As $L_z/L_x$ becomes larger, the peak energy stored in plasmoids decreases.
We remind the reader that our measure of ``reconnected-field energy'' is the magnetic energy that is not in unreconnected field lines. It is thus an upper bound on the energy stored in structures formed from reconnected magnetic field; it could also contain, e.g., turbulent magnetic energy in the current layer.
The unreconnected magnetic \emph{flux} (Fig.~\ref{fig:energyAndFluxVsLz}, right) shows the expected overall decay over time as flux is reconnected or annihilated.
At any fixed time~$t$, the remaining upstream flux tends to increase with $L_z/L_x$ (i.e., larger $L_z/L_x$ implies a slower reconnection rate)---with the striking exception of $L_z/L_x=1/2$, and the less striking exception of $L_z/L_x=3/2$. Also, at very late time ($\simeq 40L_x/c$) we see the $L_z/L_x=1/4$ case overtaking $L_z/L_x=1/8$ in the amount of flux decay.
This may represent a trend with increasing $L_z/L_x$ up to some minimally-3D value $L_z/L_x \sim 1/4$, above which results are independent of $L_z/L_x$ (but with high stochastic variability);
a statistical ensemble of simulations with large $L_z/L_x$ will be needed to distinguish the real trend.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.325\textwidth]{fig31a.pdf}%
\hspace{0.05in}%
\includegraphics*[width=0.325\textwidth]{fig31b.pdf}%
\hspace{0.05in}%
\includegraphics*[width=0.325\textwidth]{fig31c.pdf}%
}
\caption{ \label{fig:energyAndFluxVsLz}
Total transverse magnetic energy $U_{Bt}(t)$ (left), magnetic energy in the layer $U_{Bt,\rm layer}(t)$ (middle), and unreconnected flux $\psi(t)$ (right),
for simulations with the same $L_x=341\sigma\rho_0$, $B_{gz}=0$, $\eta=5$, $\beta_d=0.3$, $a=0$, but varying $L_z/L_x$.
Cases $L_z/L_x=$1/16 and~1/8 each show two
simulations identical except for random initialization of particles.
In the plot of $U_{Bt}$ (left), the horizontal grey dashed line indicates $U_{Bt}=0.75U_{B0}$---NTPA spectra are shown in Fig.~\ref{fig:ntpaVsLz} when $U_{Bt}(t)$ crosses this line.
In the flux plot (right), the intersections of the two horizontal grey, dashed lines with $\psi(t)$ indicate where the average reconnection rate is measured for Fig.~\ref{fig:reconRateVsLz}; three grey, dotted lines show the slopes for constant reconnection rates~$\beta_{\rm rec}\in\{$0.04,0.004,0.0004$\}$ normalized to $cB_0$.
}
\end{figure}
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=3in]{fig32a.pdf}%
}
\caption{ \label{fig:reconRateVsLz}
Reconnection rates, normalized to $B_0 c$ on the left axis and to~$B_0v_A$ on the right (cf.~\S\ref{sec:unreconnectedFlux}), versus $L_z/L_x$, averaged over the period during which the upstream flux falls from~$0.9\psi_0$ to~$0.8\psi_0$ (between the horizontal grey dashed lines in Fig.~\ref{fig:energyAndFluxVsLz}, right),
for simulations with $L_x=341\sigma\rho_0$.
N.B. For 3D cases, the values of the instantaneous reconnection rate may vary significantly over the averaging time, and stochastic variability may also introduce significant uncertainty.
}
\end{figure}
Normalized to $B_0 c$, the reconnection rates $\beta_{\rm rec}$ (cf.~\S\ref{sec:unreconnectedFlux}) for all cases start out between 0.025--0.05, with larger $L_z/L_x$ already exhibiting slightly lower $\beta_{\rm rec}$; this is at very early times, probably dominated by the initial thin Harris sheet.
Within $10L_x/c$, the reconnection rates have all fallen below 0.01---because reconnection has finished in nearly-2D cases, and because 3D cases enter a stage of much slower reconnection.
Figure~\ref{fig:reconRateVsLz} shows reconnection rates averaged
over the time for the unreconnected flux
to fall from $0.9\psi_0$ to $0.8\psi_0$, as a function of $L_z/L_x$ (we note that, for $L_z/L_x \leq 1/8$, $\psi(t_{0.8})=0.8\psi_0$ for $t_{0.8}<4L_x/c$).
For 3D simulations, the average rates should be used with caution, because the reconnection rate can change substantially during this time (cf. Fig.~\ref{fig:energyAndFluxVsLz}, right).
In any case, it shows clearly how much slower 3D reconnection can be.
At later times, the cases $L_z/L_x\geq 1/4$ eventually exhibit reconnection rates $\beta_{\rm rec}\sim 10^{-3}$ (this measurement does not necessarily guarantee that 2D reconnection is actually taking place; flux could be depleted by direct annihilation as well as 2D-like reconnection---cf.~\S\ref{sec:terminology},~\ref{sec:overview3d}).
It is notable that by $\gtrsim 30L_x/c$, e.g., the $L_z/L_x=1$ simulation has ``used up'' only about half the upstream flux as the 2D simulation, but has converted more magnetic energy to particle energy, because in 2D more magnetic flux and energy remains in plasmoids.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig33a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig33b.pdf}%
}
\caption{ \label{fig:ntpaVsLz}
Particle energy spectra (compensated by $\gamma^4$) for different
$L_z/L_x$, for $L_x=341\sigma\rho_0$, at the time when 25~per~cent of
the initial magnetic energy has been depleted (left---and cf. Fig.~\ref{fig:energyAndFluxVsLz}, left), and at
time $t=20L_x/c$ (right).
The spectra are rather similar (especially on the left), except possibly for the $L_z=L_x/3$ case, but
there is a rough trend of increasing numbers of high-energy particles as
$L_z/L_x$ increases.
Dotted grey lines show the slope of the power law $f(\gamma)\sim \gamma^{-3.4}$; local power-law indices in the high-energy spectrum range from roughly~4.2 to~3.4.
}
\end{figure}
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig34a.pdf}%
}
\caption{ \label{fig:ntpaFracVsLz}
The fraction of particles with Lorentz factor $\gamma>10\sigma$ for the spectra shown in Fig.~\ref{fig:ntpaVsLz}---blue circles for spectra at time $t$ such that $U_{Bt}(t)=0.75U_{B0}$, and
red x's for spectra at $t=20L_x/c$.
Although the trend with~$L_z/L_x$ appears fairly robust, the significance of these results must be carefully considered due to the steep power law.
For example, increasing the threshold from~$10\sigma$ to~$12\sigma$ lowers the fraction of threshold-exceeding particles by a factor of~4--6 (for all $L_z/L_x$).
}
\end{figure}
Figure~\ref{fig:ntpaVsLz}(left) shows
particle energy spectra at times when 25~per~cent of the initial
magnetic energy has been depleted---i.e., when $U_{Bt}(t)=0.75U_{B0}$; the times can be estimated from the intersection of $U_{Bt}(t)$ and the horizontal grey dashed line in Fig.~\ref{fig:energyAndFluxVsLz}(left), and range from $t=$5--8$L_x/c$ for $L_z/L_x\leq 0.5$, and $t=15$--$18L_x/c$ for $L_z/L_x\in\{1,1.5\}$.
We see first that all these
spectra are fairly similar, especially for $\gamma \lesssim 10\sigma$;
and even above $\gamma \gtrsim 10\sigma$, spectra vary by less than a
factor of 2 in energy, with the exception of $L_z/L_x=1/4$, which
accelerates some particles to unusually high energies (we note that
a different case,
$L_z/L_x=1/2$, is the outlier in total magnetic energy depletion).
Second, however, there is a trend of increasing
numbers of high-energy particles (say with $\gamma\simeq 10\sigma$)
as $L_z/L_x$ increases, at least for small $L_z/L_x$.
Looking at spectra at the same time, $20L_x/c$, we see differences
enhanced somewhat, with competing effects: larger $L_z/L_x$ yields
slightly more high-energy particles, but larger amounts of magnetic
energy conversion also yield more high-energy particles.
Although the somewhat short range of the high-energy power-law section prevents us from measuring precise power-law indices in a useful way, we can roughly characterize the local power-law indices as varying between~4.2 and~3.4 (in Fig.~\ref{fig:ntpaVsLz}, horizontal represents a power law $\gamma^{-4}$, and dotted grey lines show $\gamma^{-3.4}$). Whereas simulations with $L_z/L_x\leq 1/8$ have power-law indices closer to~4, those with larger $L_z/L_x\geq 1/4$ appear to exhibit harder power laws with indices around 3.4 in the range
$3 \lesssim \gamma/\sigma \lesssim 8$---however these same simulations also have slightly steeper power-law indices around 4.2 in $2 \lesssim \gamma \lesssim 3$.
An alternative measure of NTPA efficiency is the fraction of particles accelerated beyond some high energy, e.g., $\gamma=10\sigma$. Figure~\ref{fig:ntpaFracVsLz} shows that this fraction increases significantly with $L_z/L_x$ until saturating around $L_z/L_x\approx 0.1$; thus 3D simulations accelerate almost an order of magnitude more particles beyond $\gamma=10\sigma$.
Enhancement of electron NTPA in 3D, relative to 2D reconnection, has been previously observed in
subrelativistic electron-ion plasma with $B_{gz}=0$ as well as with guide field $B_{gz}/B_0\leq 1.5$
\citep{Dahlin_etal-2015,Dahlin_etal-2017,Li_etal-2019}, and has just recently been seen in pair plasma with $\sigma_h=10$, $B_{gz}/B_0=0.1$ \citep{Zhang_etal-2021arxiv};
high-energy electrons were trapped by plasmoids in 2D, but not in 3D, thus allowing untrapped electrons to experience more acceleration in 3D.
In the subrelativistic case, there is some disagreement over whether the high-energy spectra in question are power laws or exponentials; a steep power law with index $p\gtrsim 4$ over a relatively small range in~$\gamma$ can look very much like an exponential decay in~$\gamma$.
An apparently dramatic $f_{3D}(\gamma)/f_{2D}(\gamma)=10$ \citep[measured at very high energies by][]{Dahlin_etal-2017} can be interpreted as particles gaining a less dramatic 20 or 30 per~cent more energy in 3D than in 2D.
In the more relativistic $\sigma_h=10$ case, the shallower/harder power law around $p\sim 2$ shows NTPA more clearly, and the highest-energy particles in~3D gain about 2--3 times more energy than in~2D \citep{Zhang_etal-2021arxiv}.
In our simulations we find, whether comparing 2D and 3D simulations at the same time or at the same $U_{Bt}/U_{Bt0}$, that usually $0.1 < f_{3D}(\gamma)/f_{2D}(\gamma) < 10$, and the ratio is often quite close to 1, especially when comparing the spectra at the same $U_{Bt}/U_{Bt0}$.
Although these spectra are power laws, they are steep power laws
[$f(\gamma)\sim \gamma^{-4}$] and so again even a substantial $f_{3D}/f_{2D}$ may imply only a rather modest increase in particle energy.
E.g., if $f_{2D}(\gamma)=a \gamma^{-4}$ for some constant $a$, and
$f_{3D}(\gamma) = 10 f_{2D}(\gamma)$, we can write
$f_{3D} \approx a (\gamma/1.78)^{-4}$---particles in 3D need only gain 78 per~cent more energy in 3D than in 2D to explain $f_{3D}/f_{2D}=10$.
Our results are not inconsistent with increased NTPA in 3D, but
from our perspective, the differences in NTPA between 2D and 3D, compared at the same $U_{Bt}/U_{Bt0}$, are fairly insignificant, on the order of stochastic variation (regardless of guide field, as we shall see in~\S\ref{sec:Bz3d}).
3D simulations convert magnetic energy less rapidly than 2D simulations, and so the building-up of the high-energy part of $f(\gamma)$ occurs more slowly;
however, as 3D simulations convert more magnetic energy to particle energy over long times (compared with 2D), they ultimately accelerate more particles than 2D simulations.
Therefore, NTPA is enhanced in 3D relative to 2D, but the shape of the high-energy power-law spectrum and cut-off (up to an overall normalization factor) does not differ significantly.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.65\textwidth]{fig35a.pdf}%
\hfill
\includegraphics*[width=0.33\textwidth]{fig35b.pdf}
}
\caption{ \label{fig:layerOffsetY}
(Left) The layer central surface displacement from the original midplane, $\Delta y_c(x,z)$ (cf.~\S\ref{sec:sheetCenter}), at four times, $tc/L_x\in \{$0.86, 2.6, 7.7, 23$\}$, shows low-amplitude short-wavelength RDKI modes dominating at early times, but at later times the dominant mode has $\lambda_z=L_z$. At $7.7L_x/c$ (and thereafter), abrupt spatial variation in $\Delta y_c$ indicates the layer folding over on itself.
(Right) Fourier spectra magnitudes (cf.~\S\ref{sec:sheetCenter}) of the data shown on the left, averaged over all $x$ (after taking the magnitudes), show that over time, longer-wavelength RDKI modes grow to larger amplitudes.
}
\end{figure}
We have remarked several times on the stochastic variability in 3D reconnection (cf.~\S\ref{sec:variability3d},~\ref{sec:pertAndEta3d}) and have discussed a mechanism for large variability in~\S\ref{sec:RDKIamp}.
When an RDKI mode grows beyond an amplitude $\Delta y_{c*}$ on the order of its wavelength $\lambda_z$, it becomes extremely nonlinear, and the current layer folds over on itself and rapidly depletes magnetic energy within $|y|\lesssim \Delta y_{c*}$.
In Fig.~\ref{fig:energyAndFluxVsLz}, the simulation with $L_z/L_x=1/2$ experienced particularly large energy depletion, converting significantly more magnetic energy to plasma energy in~$20L_x/c$ than any other simulation did in~$50L_x/c$, including simulations with $L_z/L_x=1/4$ and~$1$.
Indeed, this simulation ($L_z=L_x/2$) developed a large-amplitude RDKI mode with the largest-possible wavelength, $\lambda_z=L_z$, and it became highly nonlinear, rapidly depleting magnetic energy in a thick layer.
Figure~\ref{fig:layerOffsetY}(left) shows the offset of the layer central surface (\S\ref{sec:sheetCenter}),
$\Delta y_c(x,z,t) \equiv y_c(x,z,t)-y_c(x,z,0)$,
at four times: $tc/L_x\in\{$0.9, 2.6, 7.7, 23$\}$;
then Fig.~\ref{fig:layerOffsetY}(right) shows the Fourier spectrum (in $z$, averaged over $x$; cf.~\S\ref{sec:sheetCenter}) of this quantity for each time.
At early times, $t=0.9L_x/c$, RDKI develops prominently at $\lambda_z\approx L_z/10 = 17 \sigma\rho_0$.
At later times, the most prominent wavelength gets longer; at $t=7.7L_x/c$, rippling with $\lambda_z=L_z$ is the dominant component.
Indeed, by this time, there are clear signs of nonlinear development and the layer folding over on itself (causing rapid magnetic energy conversion).
By $t=23L_x/c$, the layer has become generally thick and turbulent with weak magnetic field, and the long-wavelength rippling with $\lambda_z=L_z$ is no longer so apparent.
According to \citet{Zenitani_Hoshino-2007}, the wavelength of the fastest-growing (linear) RDKI mode is $\lambda_{z,\rm RDKI} = 16\upi \gamma_d \beta_d^2 \delta = 4.7\delta = 3\sigma\rho_0$.
However, this fastest-growing mode should saturate at an amplitude $\Delta y_{c*}\sim \lambda_{z,\rm RDKI}/2$ (cf.~\S\ref{sec:RDKIamp}), whereas longer-wavelength modes can grow (though possibly at a slower rate) to larger amplitudes.
Indeed, Fig.~\ref{fig:layerOffsetY} is very roughly consistent with longer-wavelength modes growing slower, but to larger amplitudes, $\tilde{y}_c(k_z) \sim k_z^{-1}$ (where $k_z=2\upi/\lambda_z$).
For some reason, in the simulation with $L_z=L_x/2=170\sigma\rho_0$, the mode with $\lambda_z=L_z$ grew sufficiently to enter highly-nonlinear development; presumably in simulations with other $L_z/L_x$, this did not happen.
Of course, for $L_z<L_x/2$, no mode can develop to such large amplitude; but the simulations with $L_z=L_x$ and $L_z=3L_x/2$ could have developed $\lambda_z=L_z/2$ as well as even longer-wavelength modes (and indeed, they did, to some extent---but for some reason the resulting magnetic energy conversion was not as dramatic).
We speculate that beyond some wavelength (in sufficiently large systems), long-wavelength RDKI modes will simply not have time to grow before being disrupted by reconnection or faster-growing, shorter-wavelength RDKI modes.
We leave it to future work to investigate what determines the nonlinear development of long-wavelength RDKI modes; we expect this to be a difficult task because we have already seen that stochastic variability can play a large role.
In summary, in 3D (i.e., for large $L_z/L_x$), with zero guide field and zero initial perturbation, reconnection and magnetic energy
conversion occur more slowly (than in 2D) during the first active reconnection stage; for large $L_z/L_x$ almost all the
released upstream/unreconnected magnetic energy is converted to particle energy
(whereas in 2D, roughly half of it is trapped in reconnected magnetic
field).
Whereas 2D simulations progress to a final state limited by the magnetic field configuration, with magnetic energy and flux trapped in plasmoids, 3D current sheet evolution does not have a clear endpoint (at least not before $50L_x/c$), but enters a long-lasting second stage in which gradual magnetic energy and upstream flux depletion continue at a slower pace.
RDKI modes can play a dramatic role in 3D reconnection---and perhaps it is even misleading to refer to ``3D reconnection'' in some cases where RDKI drives the most rapid magnetic energy conversion.
The most rapid and largest magnetic energy releases seem to be related to a long-wavelength RDKI mode developing nonlinearly when its amplitude reaches the order of its wavelength, causing extreme distortion of the current layer that quickly depletes magnetic energy within the mode amplitude.
Whether this happens at a given wavelength depends on $L_z$, but also has a random element, which explains the stochastic variability of 3D ``reconnection.''
If nonlinear RDKI does cause dramatic magnetic energy conversion, the resulting layer becomes thick and turbulent, and allows only much slower magnetic conversion thereafter, presumably because of its increased thickness.
Importantly, despite all the complications in 3D, with potentially very different mechanisms driving magnetic energy conversion to plasma energy, NTPA efficiency does not suffer in 3D and actually appears to
be slightly enhanced, although it may occur over somewhat longer times
because of the slower reconnection rate.
The increase in the number of high-energy particles does depend on the rate of magnetic energy conversion, but for a fixed amount of energy gained by the plasma, $f(\gamma)$ is remarkably insensitive to $L_z$ (as well as to the initial current sheet configuration, as we saw previously).
However, in 3D, more magnetic energy can be converted to plasma energy than in 2D, and ultimately this leads to more particles at the highest (nonthermal) energies.
\subsection{3D reconnection: guide magnetic field}
\label{sec:Bz3d}
In this subsection we investigate 3D reconnection with different initial guide magnetic fields, $B_{gz}/B_0\in \{$0, 0.1, 0.25, 0.5, 0.75, 1$\}$,
using system size $L_z=L_x=256\sigma\rho_0$.
Other parameters are as usual: zero initial perturbation, $\eta=5$, $\beta_d=0.3$, $\delta=(2/3)\sigma\rho_0$.
The dependence of 3D reconnection on the guide magnetic field $B_{gz}$ is nontrivial and non-monotonic as a result of competing effects.
We find that the effect of guide magnetic field in 3D can be roughly summarized thus:
stronger $B_{gz}$ suppresses 3D effects, so that reconnection becomes more 2D-like;
this may be expected, because guide field suppresses 3D instabilities like RDKI \citep{Zenitani_Hoshino-2008}.
Stronger guide field promotes particle transport in $z$ (relative to perpendicular directions) by aligning $\boldsymbol{B}$ increasingly parallel to $z$, and this tends to even out variations in $z$ (relative to perpendicular directions).
Again, however, as with $B_{gz}=0$, we find that NTPA continues to be very similar in 2D and 3D for $B_{gz} > 0$.
In 2D, the added inertia of the guide field, being dragged along with the plasma, slows reconnection significantly when the guide field enthalpy density, $B_{gz}^2/4\upi$ becomes comparable to or larger than the upstream plasma enthalpy density, $h=4 \theta_b n_b m_e c^2$ (expressed here in the ultrarelativistic limit), i.e., when
$B_{gz}^2/B_0^2 \gtrsim 1/\sigma_h$ \citep[cf.~\S\ref{sec:Bz2d} and][]{Liu_etal-2015}.
The guide field enthalpy appears to slow reconnection in 3D as well.
However, because other 3D effects also slow reconnection, the fundamentally 2D effect of guide field enthalpy is not noticeable at low guide fields;
at low guide field, $B_{gz}/B_0 \lesssim 1/\sigma_h$, the guide field actually promotes reconnection by suppressing 3D effects that slow reconnection.
Figure~\ref{fig:energyVsBgz3d}(left) shows the time evolution of transverse magnetic energy $U_{Bt}(t)$ in 3D, for all six guide-field values (2D results are shown in fainter colours on the same graph, for comparison).
For $B_{gz}=0$, results from three 3D simulations are shown to give a rough estimate of stochastic variability, and all three exhibit relatively slow magnetic energy conversion.
Increasing $B_{gz}$ even to $B_{gz}=0.1B_0$ speeds up reconnection (presumably by suppressing 3D instabilities) and results in significantly more magnetic energy conversion over the first 20--30$L_x/c$; by $B_{gz}=B_0/4$, this trend saturates as the guide field enthalpy starts to suppress reconnection (this is more noticeable at later times). Somewhere around $B_{gz}\gtrsim 0.75B_0$ (and clearly for $B_{gz}=B_0$), the guide field is suppressing reconnection as in 2D, so that for $B_{gz}=B_0$, 3D reconnection is slower than for $B_{gz}=0$ at all times.
We should keep in mind, however, that without guide field, stochastic variability can significantly affect system evolution, including energy dissipation (cf.~\S\ref{sec:variability3d}), and further work should investigate ensembles of simulations with different guide fields.
We argue that the enhancement of reconnection by moderate $B_{gz}$ is a consequence of reconnection becoming more 2D-like, stabilizing flux ropes and preventing 3D instabilities from interfering.
This is supported by Fig.~\ref{fig:energyVsBgz3d}(right),
which shows the magnetic energy in the layer versus time, $U_{Bt,\rm layer}(t)$.
In~\S\ref{sec:Lz3d} we showed (for $B_{gz}=0$) that for 2D simulations, $U_{Bt,\rm layer}(t)$
increases monotonically until it saturates at a significant fraction
of the initial transverse magnetic energy, $U_{Bt0}$; whereas, as
$L_z/L_x$ is increased and 3D effects become important, $U_{Bt,\rm layer}$ increases only a little before decaying away.
Here we see the same thing in 3D: for $B_{gz}=0.1B_0$, $U_{Bt,\rm layer}(t)$ grows a little and then decays over time;
as $B_{gz}$ becomes stronger,
$U_{Bt,\rm layer}(t)$ becomes more 2D-like, increasing to a larger value and (for $B_{gz}/B_0=1$) saturating at a roughly constant value (because the guide field stabilizes flux ropes containing the reconnected field).
The decay of upstream flux $\psi(t)$ in~3D, shown in Fig.~\ref{fig:fluxAndReconRateVsBz2d3d}(left), is consistent with this picture.
$B_{gz}=0$ has the slowest reconnection rate ($\propto -d\psi/dt$), and, after $30L_x/c$, has the most remaining unreconnected flux upstream.
Intermediate guide fields yield faster reconnection and more ``reconnected'' flux, but by $B_{gz}/B_0=1$, reconnection is slowed as the guide field enthalpy becomes significant compared with the relativistic plasma enthalpy (as in 2D).
We now compare $U_{Bt}(t)$ and $\psi(t)$ for these 3D simulations with otherwise identical 2D simulations, for all six values of~$B_{gz}$.
Figure~\ref{fig:energyVsBgz3d}(left) shows~$U_{Bt}(t)$.
In 2D, increasing $B_{gz}$ monotonically slows magnetic energy conversion and results in less plasma energization overall.
In 3D, the initial depletion of $U_{Bt}$ is generally slower than in 2D, but plasma energization continues for a much longer time.
The $B_{gz}=0$ simulations reconnect so slowly in 3D that even after $30L_x/c$ they have converted less magnetic energy than their 2D counterparts.
For $0.1 \leq B_{gz}/B_0 \leq 0.5$, early-time conversion rates are slower in 3D, but after 20--30$L_x/c$, much more magnetic energy is converted in~3D than in~2D.
For $B_{gz}/B_0=0.75$, the 2D and 3D magnetic energy conversion rates are quite similar at early times, but again the 3D simulation eventually converts much more of $U_{Bt0}$ to $U_{\rm plasma}$.
Finally, for $B_{gz}/B_0=1$, the initial magnetic energy conversion rate in 3D is slower than in 2D (which, for reasons we do not yet understand, bucks the trend of becoming more 2D-like with strong guide field), but the 3D simulation ends up converting a roughly similar amount of energy as in 2D (because, like 2D simulations, it still stores a significant amount of energy in relatively stable flux ropes---cf.~Fig.~\ref{fig:energyVsBgz3d}, right).
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig36a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig36b.pdf}
}
\caption{ \label{fig:energyVsBgz3d}
Transverse magnetic energy $U_{Bt}(t)$ in the entire simulation (left) and $U_{Bt,\rm layer}(t)$, restricted to the layer (right), for both 2D (fainter, thinner lines) and 3D ($L_z=L_x$, thicker lines) simulations, for a range of
guide fields $B_{gz}$.
Three 3D simulations are shown for $B_{gz}=0$.
On the left plot, lines are individually labelled
with dimensionality (``2d'' or ``3D'') and $B_{gz}/B_0$.
These simulations have size~$L_x=256\sigma\rho_0$.
}
\end{figure}
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig37a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig37b.pdf}
}
\caption{ \label{fig:fluxAndReconRateVsBz2d3d}
(Left) Unreconnected flux $\psi(t)$ for the same simulations shown in Fig.~\ref{fig:energyVsBgz3d}, with horizontal grey, dotted lines indicating $0.9\psi_0$ and~$0.8\psi_0$.
(Right) Reconnection rates versus $B_{gz}/B_0$ for the simulations in the
left panel, averaged over the time for $\psi(t)$ to fall from~$0.9\psi_0$ to~$0.8\psi_0$ (except, since none of the three 3D, $B_{gz}=0$ cases reached~$0.8\psi_0$, we show~$\beta_{\rm rec}$ for just one of them, averaged between~$0.9\psi_0$ to~$0.81\psi_0$).
In 2D, increasing $B_{gz}$ reduces the reconnection rate; in 3D,
a small guide field increases the reconnection rate, but eventually, strong guide field will reduce the rate.
N.B. For 3D cases, the values of the instantaneous reconnection rate may vary significantly over the averaging time, and stochastic variability may also introduce significant uncertainty.
}
\end{figure}
The left panel of Fig.~\ref{fig:fluxAndReconRateVsBz2d3d} shows $\psi(t)$, again for 2D and 3D, for all six $B_{gz}$ values.
As $B_{gz}/B_0$ increases to $B_{gz}/B_0 \sim 0.75$, 2D reconnection rates ($\propto -d\psi/dt$) decrease, and 3D rates increase; correspondingly, the total amount of flux reconnected (over 30$L_x/c$) decreases in 2D and increases in 3D, until $\psi(t)$ becomes relatively similar in 2D and 3D for $B_{gz}/B_0\approx 0.75$.
This trend suggests that for $B_{gz}/B_0=1$, the 2D and 3D cases should be even more similar, but in fact (compared with $B_{gz}/B_0=0.75$) reconnection for $B_{gz}/B_0=1$ slows dramatically in 3D while it slows only a little in 2D.
We leave a full exploration of this to future work investigating 3D reconnection with stronger guide fields,
but we speculate that 3D reconnection may not reach the 2D limit until the purely 2D suppression (which begins for $B_{gz}/B_0 \gtrsim 1/\sigma_h$ and becomes stronger with higher $B_{gz}$) completely dominates over the suppression due to 3D effects.
At $B_{gz}/B_0=1$, we may be seeing the combined effects of guide-field suppression and 3D suppression of reconnection, while at some larger $B_{gz}$, we expect 3D effects to become weaker.
This is exemplified quantitatively in Fig.~\ref{fig:fluxAndReconRateVsBz2d3d}(right), where we graph the normalized reconnection rates, averaged over the period during which $\psi(t)$ falls from
$0.9\psi(0)$ to $0.8\psi(0)$.
The 2D reconnection rates fall with stronger $B_{gz}$, and the rate for $B_{gz}/B_0=1$ is roughly half of the maximum rate, which is realized for $B_{gz}/B_0 \lesssim 0.25$.
The 3D rate at $B_{gz}/B_0=0$ is an order of magnitude below the maximum 2D rate, but the 3D rate speeds up for intermediate $B_{gz}$ so that at $B_{gz}/B_0=0.75$, the 2D and 3D rates differ by less than a factor of 2.
As $B_{gz}/B_0$ increases from 0.75 to 1, the 2D rate drops smoothly, while the 3D rate drops more dramatically.
It appears likely that magnetic energy and flux would continue to decrease if we ran the 3D simulations for longer times.
However, we quantify the ``final'' amounts of remaining magnetic energy and flux after a long time, $t=30L_x/c$,
in table~\ref{tab:finalEnergyAndFluxVsBgz3d}---for both 2D and 3D simulations, for all $B_{gz}$.
In terms of these ``final'' values, we see that 2D and 3D reconnection
are most similar for $B_{gz}/B_0=1$, with $U_{Bt}$ differing by only 3 percentage points, and $\psi(t)$ by about 8 percentage points.
The table also shows that 3D simulations generally lose less upstream flux but convert more magnetic energy than 2D simulations,
because in 2D, energy in plasmoids is stable, whereas in 3D it is depleted (unless the guide field is sufficiently strong), as shown in Fig.~\ref{fig:energyVsBgz3d}(right).
We have seen that both guide field and 3D effects can suppress reconnection, slowing the overall rates of reconnection and release of magnetic energy.
We find that an initial guide field suppresses NTPA---to the same extent in both 2D and 3D---whereas 3D effects do not suppress NTPA, despite slowing reconnection.
Figure~\ref{fig:ntpaVsBgz3d} shows the electron energy spectra $f(\gamma)$ (compensated by $\gamma^4$) for each $B_{gz}$, both at
the same amount of depleted magnetic energy, $U_{Bt}=0.83U_{Bt0}$ (left),
and at the same time, $t=20L_x/c$ (right), for the 3D simulations as well as for comparable 2D simulations.
Keeping in mind that graphing $\gamma^4 f(\gamma)$ enhances small differences in $f(\gamma)$, we consider the 2D and 3D spectra (for the same $B_{gz}$) to be fairly similar in magnitude [separated by less than a factor of 2 in either $\gamma$ or in $f(\gamma)$], especially at $U_{Bt}(t)=0.83U_{Bt0}$.
As the guide field increases, the spectra in 2D and 3D become even closer, until for $B_{gz}=B_0$ they are nearly identical.
We cannot make a general statement about whether NTPA is more efficient in 2D or 3D without going into much more detail regarding the exact times and energies at which spectra are compared.
However, it does appear that, because 3D simulations can convert more magnetic energy to plasma energy, over long times 3D simulations accelerate more particles.
\citet{Dahlin_etal-2017} observed enhancement of NTPA in 3D, relative to 2D, in subrelativistic electron-ion reconnection at $B_{gz}=0$, and the 3D enhancement became stronger with increasing guide field, until this effect saturated around $B_{gz}/B_0 \simeq 1$ [\citet{Li_etal-2019} also observed enhancement in 3D for $B_{gz}/B_0=0.2$].
As we discussed in~\S\ref{sec:Lz3d},
our results are not inconsistent with these, but the differences between 2D and 3D spectra do not seem very significant if compared at times with the same amount of magnetic energy depletion; however, because more magnetic energy can be converted to plasma energy in 3D, 3D simulation can accelerate more particles to high energies.
However, as the guide field becomes stronger, we expect 3D simulations to become essentially identical to 2D simulations, and then we expect NTPA to be identical as well.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig38a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig38b.pdf}
}
\caption{ \label{fig:ntpaVsBgz3d}
Particle energy spectra compensated by $\gamma^4$ at a time (left) when $U_{Bt}=0.83U_{Bt0}$ (i.e., 17~per~cent of magnetic energy has been converted to plasma energy), and (right) at $t=20L_x/c$, for a range of guide fields $B_{gz}$, for 3D simulations with $L_z=L_x$ (thicker, darker lines) and, for comparison, for 2D (thinner, fainter).
Dotted grey-on-yellow lines show the slopes of $f(\gamma)\propto \gamma^{-p}$ for $p=5$, 5.6, and 6.5 ($p=4$ would be horizontal).
These simulations are the same as shown in Figs.~\ref{fig:energyVsBgz3d}, \ref{fig:fluxAndReconRateVsBz2d3d}.
The 2D simulation with $B_{gz}=B_0$ did not reach $U_{Bt}=0.83U_{Bt0}$, but remained around $U_{Bt}=0.85U_{Bt0}$ from $t=7L_x/c$ to $t=30L_x/c$.
}
\end{figure}
\begin{table}
\centering
\begin{tabular}{lcccccccc}
$B_{gz}/B_0$ & 0 & 0 & 0 & 0.1 & 0.25 & 0.5 & 0.75 & 1 \\
\hline
loss in $U_{Bt}$ (2D) & 27\% &&& 27\% & 25\% & 21\% & 18\% & 15\% \\
loss in flux (2D) & 44\% &&& 43\% & 43\% & 40\% & 36\% & 33\% \\
\hline
loss in $U_{Bt}$ ($L_z=L_x$) & 26\% & 23\% & 24\% & 41\% & 38\% & 36\% & 32\% & 18\% \\
loss in flux ($L_z=L_x$) & 20\% & 16\% & 17\% & 31\% & 33\% & 31\% & 29\% & 25\% \\
\end{tabular}
\caption{\label{tab:finalEnergyAndFluxVsBgz3d}
The fractional loss in transverse magnetic field energy $U_{Bt}$ and
unreconnected magnetic flux $\psi$ over $30L_x/c$ of reconnection, for various guide fields,
for simulations with $L_x=256\sigma\rho_0$, both 2D and $L_z=L_x$ (for $L_z=L_x$, three simulations are shown
with $B_{gz}=0$
to give some idea of stochastic
variation).
}
\end{table}
We have not yet investigated the strong guide field regime, $B_{gz}/B_0\gg 1$, because of the high computational cost associated with larger $L_z/L_x$ and longer simulation times;
it will be important in future work to explore this regime and in particular to confirm or deny conclusively that reconnection with very strong guide field rigorously approaches the ideal 2D limit with perfect uniformity in $z$.
\section{Comparison with large magnetization}
\label{sec:highSigmah}
In this paper we have found that, for $\sigma_h=1$, 3D reconnection is significantly slower than 2D reconnection, and lasts for much longer times over which it eventually converts more magnetic energy to plasma energy
(but that NTPA is quite similar in 2D and 3D).
However, relativistic reconnection with large magnetization, $\sigma_h\gg 1$, has been found to convert similar amounts of magnetic energy to plasma energy at similar rates in 2D and 3D \citep{Guo_etal-2014,Guo_etal-2015,Werner_Uzdensky-2017,Guo_etal-2020arxiv,Zhang_etal-2021arxiv}.\footnote{
For reasons that are not clear, \citet{Sironi_Spitkovsky-2014} measured a slower reconnection rate in 3D, about $1/4$ of the rate in 2D, but in spite of that
observed otherwise substantially similar reconnection in 2D and 3D.
We note that recent work with $\sigma_h=10$ and $B_{gz}/B_0=0.1$ found 3D
reconnection rates $(c/v_A)\beta_{\rm rec}\sim 0.075$, comparable to but less than $(c/v_A)\beta_{\rm rec}\sim 0.12$ found in~2D
\citep{Zhang_etal-2021arxiv}.
}
It thus may appear that 3D effects play less of a role for large-$\sigma_h$ reconnection.
However, as we will discuss in the following, there is a possibility that the 2D-like behaviour for 3D reconnection might be a result of the initial simulation configuration, and not an inevitable, intrinsic consequence of larger~$\sigma_h$.
The recent work by \citet{Guo_etal-2020arxiv} should be considered when comparing large and small $\sigma_h$, since it studied 3D relativistic pair reconnection with $\sigma_h$ ranging from around~1 to~200 (though starting from a transrelativistic temperature $\theta_b \sim 1$).
There, high reconnection rates and magnetic energy conversion rates (similar to rates in 2D) were observed for the entire range of~$\sigma_h$.
However, that study focused on very early times; those rates were measured within the first~$1\,L_x/c$.
Moreover, those simulations began with an initial field perturbation, uniform in the third dimension, and used a force-free configuration---i.e., an initial magnetic field with a substantial $B_z$ component localized to the layer.
We have shown that differences in magnetic energy conversion rates between 2D and 3D are suppressed by an initial perturbation (\S\ref{sec:pertAndEta3d}) and by an initially globally-uniform guide field $B_{gz}$ (\S\ref{sec:Bz3d}), particularly at early times.
Therefore, \citet{Guo_etal-2020arxiv} does not address our particular question about whether lower energy conversion rates might be observed in 3D if simulations ran for long times, especially starting with no initial perturbation and no guide field.
The very recent work by \citet{Zhang_etal-2021arxiv} studied reconnection in 3D pair plasma with $\sigma_h=10$ (and $B_{gz}/B_0=0.1$).
They observed slightly slower reconnection in 3D (compared with 2D), but not substantially slower. This adds to the evidence suggesting that for larger~$\sigma_h$, 3D reconnection is more similar to 2D reconnection.
However, their use of an initial perturbation and outflow boundary conditions led to an X-line running down the centre of the box the entire length in~$z$, which could conceivably be the dominant driver of 2D-like reconnection (i.e., and not the larger~$\sigma_h$).
The guide field, though weak, might also favour 2D-like reconnection, but the paper states that simulations with $B_{gz}=0$ did not differ significantly.
Thus we cannot confidently attribute the 2D-like speed of 3D reconnection purely to the difference in~$\sigma_h$.
\cite{Zhang_etal-2021arxiv} also found NTPA to be very similar in 2D and 3D---except that, similar to the earlier finding in subrelativistic electron-ion reconnection \citep{Dahlin_etal-2017}, the highest-energy particles in 3D were accelerated to higher energies (than in 2D) because they escaped from flux ropes (with finite extent in~$z$) and thus experienced additional acceleration in the reconnection electric field. Although a full comparison of this acceleration behaviour is beyond the scope of our work, preliminary analysis suggests that our $\sigma_h=1$ simulations show similar behaviour of high-energy particles experiencing (in 3D) additional direct electric field acceleration.
Our own simulations of large-$\sigma_h$ relativistic reconnection in 3D pair plasma from a previous work \citep{Werner_Uzdensky-2017}, with $\sigma_h=25$, also started with a magnetic field perturbation ($a=2.7$, $s/\delta=1.9$), but used a Harris sheet set-up (as in this paper) and ran for about~$5L_x/c$.
We also ran some simulations without perturbation, but mentioned them only in passing.
One of those simulations---3D ($L_z=L_x$), without guide field ($B_{gz}=0$) and without initial perturbation ($a=0$)---did show exceptionally slow reconnection (perhaps like our $\sigma_h=1$ simulations in this paper) and much-suppressed NTPA (unlike our $\sigma_h=1$ simulations).
At the time we attributed the slow reconnection to the influence of the initial current sheet; i.e., we believed this effect would have disappeared if we could have run simulations with larger $L_x/\delta$.
In light of our $\sigma_h=1$ results presented above, we revisit these $\sigma_h=25$ simulations.
Although the larger $\sigma_h$ is the most fundamental physical difference between the $\sigma_h=1$ simulations in this paper and those
in \citet{Werner_Uzdensky-2017} with $\sigma_h=25$,
practical considerations necessitated other differences.
Large $\sigma_h$ required higher grid resolution ($\Delta x = \sigma\rho_0/12$, to resolve the Debye length of the colder upstream plasma), so the $\sigma_h=25$
simulations were significantly smaller, with $L_x=80\sigma\rho_0 = 120\delta$;
whereas, our $\sigma_h=1$ simulations with $\Delta x=\sigma\rho_0/3$ are 4--6 times larger, with some parameter scans using $L_x=341 \sigma\rho_0=512\delta$ (e.g.,~\S\ref{sec:Lz3d}) and the largest 3D simulation reaching $L_x=512\sigma\rho_0=768\delta$ (cf.~\S\ref{sec:overview3d}).
Because the same overdensity $\eta=5$ was used in the large-$\sigma_h$ regime,
the $\sigma_h=25$ simulations required a much hotter initial current sheet, $\theta_d=10.5\theta_b$, to balance the relatively strong upstream magnetic pressure;
whereas for $\sigma_h=1$, the initial current sheet is cooler than the upstream plasma, $\theta_d=0.4\theta_b$.
The initial current sheet for $\sigma_h=25$ thus had, at its peak density, about~50 times the energy and inertia density of the upstream plasma; for $\sigma_h=1$, this ratio is only 2.
It is therefore reasonable to expect the initial current sheet to exert greater influence for $\sigma_h=25$ than for $\sigma_h=1$, especially considering that $L_x/\delta$ was several times smaller for $\sigma_h=25$.
The relatively high enthalpy of the initial current sheet plasma for $\sigma_h=25$, relative to the background plasma, may play a critical role in slowing down reconnection in 3D (with $a=0$ and $B_{gz}=0$).
Here is what we believe may be happening in that case.
When reconnection starts, the initial current sheet plasma is not trapped and swept away from reconnecting X-points, which ultimately control the reconnection rate \citep{Uzdensky_etal-2010}, but rather is dispersed widely in the vicinity of the layer.
Some of that hot, dense plasma recirculates into X-point inflows, where its high inertia---50 times higher than the background plasma's inertia---significantly lowers the Alfv\'{e}n velocity and slows reconnection.
Either an initial perturbation, or a weak guide field, or 2D-ness will facilitate the trapping of the initial current sheet plasma in plasmoids, sweeping it away from X-points and preventing this scenario.
Specifically, the 2D/3D difference disappeared either with a perturbation of $a=2.7$ or with a guide field of $B_{gz}\geq B_0/4$---where, importantly, $B_{gz}=B_0/4$ was strong enough to suppress 2D/3D differences but weak enough that it did not substantially affect reconnection rates or NTPA in either 2D or in 3D.
We now review the evidence that
the 3D simulation with $B_{gz}=0$ and $a=0$ was an outlier with slow reconnection and limited NTPA.
When we compare the transverse magnetic energy versus time, $U_{Bt}(t)$, for eight configurations with $\sigma_h=25$---all combinations of (1)~dimensionality: 2D and~3D ($L_z=L_x$), (2)~guide field: $B_{gz}=0$ and~$B_{gz}=B_0/4$, (3)~perturbation: $a=2.7$ and~$a=0$---we see three rough groups of similar $U_{Bt}(t)$ (Fig.~\ref{fig:highSigmahEnergy2d3d}, left).
First, there is a group of all four simulations with $B_{gz}=B_0/4$, which have similar $U_{Bt}(t)$, regardless of dimensionality or perturbation.
Second, there is a group of all simulations with $B_{gz}=0$---except for the outlier (i.e., 3D, $a=0$). Comparing these two groups, we see, as expected, that the small guide field slows/suppresses reconnection by a modest amount.
Within each group, we see that the 3D simulations can ultimately convert a bit more magnetic energy to plasma energy than the 2D simulations (cf.~\S\ref{sec:Lz3d}), but this effect is not very significant, at least in the first~$5L_x/c$.
The third group contains just the single outlier---3D, $B_{gz}=0$, $a=0$---with much slower plasma energization.
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig39a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig39b.pdf}
}
\caption{ \label{fig:highSigmahEnergy2d3d}
Magnetic energy versus time (left) in $B_x$ and $B_y$ components, and (right) in $B_y$ components only (as a proxy for $U_{Bt,\rm layer}$), for eight different configurations of $\sigma_h=25$ simulations from
\citet{Werner_Uzdensky-2017}.
Shown are all combinations of 2D (thin, red/yellow lines) or 3D (thicker, blue-ish), $B_{gz}/B_0=0$ (no symbols) or~$1/4$ (symbols, `$B_g$'), without initial perturbation (solid lines) or with a perturbation such that the separatrix initially extends to $s=1.8\delta$ (dashed lines, `pert').
Results for three different 2D simulations with $B_{gz}=0$ and no perturbation are shown. A dashed grey line marks the energy at which spectra are shown in Fig.~\ref{fig:highSigmahNTPA2d3d}.
}
\end{figure}
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=0.49\textwidth]{fig40a.pdf}%
\hfill
\includegraphics*[width=0.49\textwidth]{fig40b.pdf}
}
\caption{ \label{fig:highSigmahNTPA2d3d}
Electron energy spectra for the simulations in Fig.~\ref{fig:highSigmahEnergy2d3d}: (left)
at a time when approximately 15~per~cent of the magnetic energy has been depleted (see dashed grey line in Fig.~\ref{fig:highSigmahEnergy2d3d}), and (right) at time $t=4L_x/c$. A dashed grey line shows the spectra at $t=0$.
}
\end{figure}
\begin{figure}
\centering
\fullplot{
\includegraphics*[width=\textwidth]{fig41a.pdf}%
}
\caption{ \label{fig:highSigmahDensity3d}
In 3D simulations with $\sigma_h=25$, \emph{either} an initial perturbation \emph{or} a guide field can cause the drift particles to be trapped and cleared out of large areas, as shown by
the density $n_{de}$ of initially-drifting electrons in the original reconnection midplane (left column), roughly~$0.4L_x/c$ after reconnection onset when the magnetic energy has declined by about 10~per~cent, for:
(top row)
$B_{gz}=0$ and perturbation, $a=2.7$ or $s/\delta=1.8$;
(middle)
$B_{gz}=B_0/4$ and zero perturbation;
(bottom)
the outlier case with $B_{gz}=0$ and zero perturbation.
The right column shows the density $n_{be}$ of background electrons for reference.
Green contours outline areas where $n_{de} < 0.2 n_{be0}$.
The time $t'$ since onset and the fraction of remaining magnetic energy are noted on the right side.
}
\end{figure}
Interestingly, this grouping falls apart when we look at $U_{By}(t)$, the energy in magnetic field components $B_y$ (Fig.~\ref{fig:highSigmahEnergy2d3d}, right). Here, $U_{By}$ is a proxy for the magnetic energy $U_{Bt,\rm layer}$ in the layer, and we have seen (for $\sigma_h=1$, e.g.,~\S\ref{sec:Lz3d}) that 2D reconnection stores magnetic energy in plasmoids, whereas plasmoids tend to decay in 3D.
Here, there is a distinct difference between 2D and 3D, although the 3D case with $B_{gz}=0$ and $a=0$ is still an outlier with the lowest overall $U_{By}$. Both guide field and initial perturbation make 3D simulations behave more---but not entirely---like 2D simulations, something that we have seen for $\sigma_h=1$ (cf.~\S\ref{sec:pertAndEta3d}, \S\ref{sec:Bz3d}).
This suggests, even for $\sigma_h=25$, 3D simulations may be more different from 2D simulations than previously suspected.
However, it is important to remember that 2D and 3D simulations with $\sigma_h=25$ have rather similar total magnetic energy evolution (unlike with $\sigma_h=1$), apart from the one outlier case.
Looking at NTPA for these same simulations, we again see that the 3D case with $B_{gz}=0$, $a=0$ is a clear outlier.
Particle energy spectra are shown in Fig.~\ref{fig:highSigmahNTPA2d3d}---at times when 15~per~cent of the initial magnetic energy has been depleted, i.e., $U_{Bt}(t)=0.85U_{B0}$ (left), and also at $t\approx 4L_x/c$ (right).
All the simulations, except the outlier, have similar energy spectra, especially for $U_{Bt}(t)=0.85 U_{B0}$. (When compared at the same, $t=4L_x/c$, one sees that, since even a weak guide field slows reconnection a little, NTPA for $B_{gz}=B_0/4$ has not quite caught up to NTPA for $B_{gz}=0$.)
We further observe that that this outlier---in reconnection rate and NTPA, as shown above---is also an outlier in the behaviour of the initially-drifting plasma (i.e., the initial current sheet plasma).
In all simulations but the outlier---e.g., 2D or $B_{gz} > 0$ or $a>0$---the initially-drifting particles are rapidly swept away from X-points and trapped within plasmoids \citep[see our discussion of plasmoid formation with and without perturbation in~\S\ref{sec:pertAndEta2d}, and also][]{Ball_etal-2018}.
This can be clearly seen in Fig.~\ref{fig:highSigmahDensity3d}, which shows $n_{de}(x,y_c,z)$, the density of initially-drifting electrons, as well as the density $n_{be}(x,y_c,z)$ of background electrons, in the original midplane, $y=y_c$.
With $a=2.7$ ($s=1.8\delta$, top row) or with $B_{gz}=B_0/4$ (middle row), it takes only a short time to evacuate most of the initially-drifting plasma from large areas (e.g., large areas with $n_{de}<0.2n_{be0}$ are outlined---$0.2n_{be0}$ is a very small density compared with the initial peak density $n_{de}=5n_{be0}$).
However, with no perturbation and no guide field (bottom row), the initially-drifting particles remain scattered roughly uniformly about the layer, while spreading in $y$, including in the vicinity of X-points.
With 50 times the enthalpy density of the background plasma, the initially-drifting plasma can thus significantly load the inflows upstream of X-points, in this case lowering the effective~$\sigma_h$ and hence Alfv\'{e}n velocity, and thereby slowing reconnection.
Let us quickly review. In ($\sigma_h=25$) simulations that have $B_{gz}>0$ or have $a>0$ or are 2D,
\begin{itemize}[leftmargin=3mm, itemindent=0mm, labelsep=1mm]
\item the initial current sheet plasma is quickly swept away from the vicinity of X-points,
\item so that the plasma near X-points flows in from the upstream plasma with $\sigma_h \gg 1$,
\item and fast reconnection and NTPA ensue as in 2D simulations.
\end{itemize}
In the (outlier) 3D simulation with $B_{gz}=0$ and $a=0$,
\begin{itemize}[leftmargin=3mm, itemindent=0mm, labelsep=1mm]
\item the initial current sheet plasma is dispersed about the layer and can flow back into X-points,
\item reducing the~$\sigma_{h,X}$ of the plasma immediately upstream of X-points (since, for large~$\sigma_h$, the initial current sheet plasma has much higher enthalpy and inertia than the upstream plasma),
\item and reconnection is slow and NTPA is substantially diminished.
\end{itemize}
When the initial current sheet plasma can flow back into X-points, we estimate the hot magnetization~$\sigma_{h,X}$, which controls reconnection rate and NTPA, as follows.
From pressure balance, the pressure of the initial current sheet plasma is $p_d\equiv n_{d0}\theta_d m_e c^2/\gamma_d = B_0^2/8\upi$; since the enthalpy density is $h\sim 4 p_d$,
$\sigma_{h,d} \equiv B_0^2/(4\upi h) \sim O(1)$.
Even if the upstream has~$\sigma_h \gg 1$, $\sigma_{h,X}\sim [\sigma_{h}^{-1} + \sigma_{h,d}^{-1}]^{-1}$ may be~$O(1)$, roughly independent of the upstream~$\sigma_h$.
Lowering~$\sigma_h$ is known both to slow reconnection and to suppress NTPA \citep{Sironi_Spitkovsky-2014,Guo_etal-2015,Werner_etal-2016,Werner_Uzdensky-2017,Werner_etal-2018,Guo_etal-2020arxiv}.
In contrast, our 3D simulations with (asymptotic upstream) $\sigma_h=1$,
in which the initial current sheet plasma had only twice the enthalpy density of the background plasma, showed slow reconnection but \emph{no reduction} in NTPA (compared with 2D; cf.~\S\ref{sec:Lz3d}).
This suggests that the suppression of reconnection (and NTPA)
in the outlier $\sigma_h=25$ simulation may not be caused directly by 3D effects (as manifested for~$\sigma_h=1$), but rather may be a consequence of the initial current sheet plasma behaviour, enabled indirectly by 3D effects (as long as $B_{gz}=0$ and~$a=0$).
This raises the important question: in a much larger simulation, or in a simulation with open outflow boundary conditions, would the initial current sheet plasma continue to load the upstream plasma at X-points for arbitrarily long times?
I.e., is this an artificial effect of the initial and boundary conditions of an idealized simulation that might not be important in realistic astrophysical systems, or is it an inevitable consequence of 3D reconnection in some regimes?
In a much larger simulation, or if particles could escape the simulation, the loading of the upstream plasma (by hot, dense plasma from the initial current sheet) might conceivably be a brief, transient phenomenon.
Eventually, the initially-drifting particles might escape or be trapped or become so diluted by background plasma that they cannot effectively continue to increase the plasma inertia at X-points.
Indeed, it was this reasoning that led us to present only results from simulations with $a>0$ in \citet{Werner_Uzdensky-2017}; the perturbation appeared to minimize the influence of the initial current sheet, allowing us to access the large-system regime, $L_x\gg \delta$, with computationally-feasible simulations.%
\footnote{
Usually we refer to the large-system regime with respect to NTPA,
with $L_x \gg 40\sigma\rho_0$ \citep{Werner_etal-2016},
but here we use the
term in a different sense, with ``large'' implying $L_x\gg \delta$, to
reduce the influence of the initial current sheet.
A typical astrophysical system would almost
certainly be very large in both senses.
}
On the other hand, the upstream plasma is heated and compressed as it flows through a reconnecting X-point, and---if that plasma can recirculate back into the X-point inflows---that could provide long-term suppression or self-regulation of reconnection.
Unfortunately we cannot conclusively answer this question with our previously-run $\sigma_h=25$ simulations and must leave it to future research.
Even if upstream loading by the initial current sheet plasma explains the suppressed reconnection and NTPA in 3D, and even if this effect vanishes for larger systems, this second look at $\sigma_h=25$ simulations suggests that we may still need to investigate whether and how 3D reconnection differs from 2D reconnection for much larger systems with~$\sigma_h\gg 1$. Figure~\ref{fig:highSigmahEnergy2d3d} shows that, despite similar $U_{Bt}(t)$ and NTPA in 2D and 3D (for~$L_x=80\sigma\rho_0$), 3D reconnection with $\sigma_h=25$ nevertheless exhibits one very clear signature of fundamentally 3D effects---namely that magnetic energy stored in plasmoids (flux ropes) decays in 3D, but not in 2D.
This could significantly affect the long-term evolution of a reconnection current sheet; for example, if plasmoids disintegrate in 3D (even for~$\sigma_h\gg 1$), that might potentially kill the acceleration mechanism (due to conservation of magnetic moment in a compressing plasmoid) for particles trapped in plasmoids \citep{Petropoulou_Sironi-2018,Hakobyan_etal-2021}.
In summary, a definitive answer to the question of whether 3D reconnection with $\sigma_h\gg 1$ and weak or zero guide field can be slow compared with 2D reconnection (as is the case for reconnection with $\sigma_h=1$) will require more investigation.
We can at least say that $\sigma_h\gg 1$ reconnection shows some distinctive 3D effects, most notably the decay of flux-rope structures, resulting in more magnetic energy release in 3D than in 2D.
However, in some regimes with $\sigma_h \gg 1$, simulations have nevertheless shown substantial similarities between 2D and 3D reconnection in terms of magnetic energy evolution and NTPA.
It is possible that, in sufficiently large systems (larger than has been explored to date), $\sigma_h\gg 1$ reconnection is inevitably slow in 3D (compared with 2D), just as for $\sigma_h=1$.
Alternatively, it is possible that the slowing of reconnection and the less efficient NTPA for $\sigma_h=25$ (with zero guide field and no perturbation) is the result of a transient influence of the initial current sheet and that in more realistic astrophysical systems (at long times after transients effects have died away) 3D reconnection would eventually and inevitably resemble 2D reconnection.
\section{Summary}
\label{sec:summary}
Before a final discussion of the significance and some impacts of this study, we briefly summarize the most important results for magnetic reconnection in 2D and 3D ultrarelativistically-hot collisionless pair plasma with $\sigma_h=1$ (i.e., $\beta_{\rm plasma}=1/2$).
We remind the reader that we define $\sigma_h\equiv B_0^2/(16\upi n_b \theta_b m_e c^2)$ with respect to the upstream (ambient) plasma, excluding any guide magnetic field (here, $\theta_b m_e c^2$ is the temperature, and $\theta_b\gg 1$).
Thus the magnetic energy and thermal energy are comparable, and both greatly exceed the plasma rest-mass energy.
The results are organized in three subsections: (\ref{sec:sumQualitative}) major qualitative differences between 2D and 3D, (\ref{sec:sum2D}) 2D-specific results, and (\ref{sec:sum3D}) 3D-specific results.
\subsection{Qualitative similarities and differences in 2D and 3D}
\label{sec:sumQualitative}
Our observations of general, \emph{qualitative} aspects of current sheet evolution for (upstream) $\sigma_h=1$ are summarized below, emphasizing similarities and differences in 2D and 3D.
(We describe 2D behaviour here primarily to facilitate comparison with 3D; the qualitative 2D behaviour is familiar from previous works studying this and $\sigma_h\gg 1$ regimes.)
\begin{itemize}[leftmargin=3mm, itemindent=0mm, labelsep=1mm]
\item In 2D and 3D, the initial current sheet breaks up into smaller current sheets because of the tearing instability and starts reconnecting. Small plasmoids---structures of higher plasma density contained by magnetic field (possibly magnetic islands or flux ropes)---grow, fed by reconnection outflows from elementary reconnecting current sheets.
(\S\ref{sec:overview2d},~\ref{sec:overview3d})
\item In 3D, the current sheet ripples
due to RDKI---but this can be suppressed by guide magnetic field.
(\S\ref{sec:overview3d},~\ref{sec:Bz3d})
\item In 2D and 3D, reconnection pumps plasma and magnetic energy from the upstream region through X-points (or X-lines) into plasmoids; in the process, some upstream magnetic energy is converted to plasma energy and some remains in magnetic form in plasmoids.
(\S\ref{sec:overview2d},~\ref{sec:overview3d})
\item In 2D, plasmoids (magnetic islands) are distinct structural units; they have closed (eventually circular), reconnected magnetic field lines that permanently trap plasma and magnetic energy from reconnection outflows. They are essentially stable in 2D and can move about along the layer. Two colliding plasmoids will merge into one (while mostly conserving the trapped flux and energy). As reconnection continues (in a closed system), a single large plasmoid eventually engulfs (i.e., merges with) all other plasmoids, growing (and never shrinking) as long as reconnection continues.
(\S\ref{sec:overview2d},~\ref{sec:overview3d})
\item In 3D, plasmoids (flux ropes) can decay, not only losing their individual integrity but also converting their magnetic energy to plasma energy. Importantly, this means that upstream magnetic energy is more completely converted to plasma energy in 3D (because 2D reconnection hoards some magnetic energy in plasmoids). Strong guide field inhibits flux rope disintegration.
(\S\ref{sec:overview3d},~\ref{sec:Lz3d})
\item An indication of plasmoid decay---and the clearest signature of novel 3D behaviour that we have observed---is the magnetic energy $U_{Bt,\rm layer}$ in the plasmoid-containing current layer (defined in~\S\ref{sec:terminology}). In 2D, $U_{Bt,\rm layer}$ rises until reconnection ceases; in 3D, it rises at the onset of reconnection, but then declines as plasmoids decay faster than reconnection can inflate them. $U_{Bt,\rm layer}$ is more sensitive to 3D effects than other measures, and it also exhibits less stochastic variation from simulation to simulation and more closely correlates with~$L_z/L_x$ (Fig.~\ref{fig:energyAndFluxVsLz}).
As a rough indicator of 3D effects, the (much easier to calculate) magnetic energy in~$B_y$ components can be substituted for $U_{Bt,\rm layer}$.
(\S\ref{sec:overview3d},~\ref{sec:Lz3d})
\item In 2D reconnection, flux is conserved in a precise sense: the total upstream flux plus the flux around the largest plasmoid is almost constant, equal to the initial upstream flux. In 3D reconnection, flux may be outright annihilated. The small amount of magnetic energy $U_{Bt,\rm layer}$ left in the layer in 3D shows that upstream flux is eventually, if not directly, annihilated. It is unclear how much upstream flux is annihilated directly without undergoing any sort of reconnection, and how much first undergoes reconnection to be annihilated later as plasmoids decay. The resemblance between 3D simulations and 2D simulations in both the $x$-$z$ and $y$-$z$ planes suggests that both these processes occur in 3D, but the precise extents cannot be estimated without additional diagnostics. (\S\ref{sec:unreconnectedFlux},~\ref{sec:overview3d},~\ref{sec:RDKIamp},~\ref{sec:Lz3d})
\item We speculate that, in 3D, for weak guide field, small plasmoids may decay faster than large, long plasmoids, and that the decay of plasmoids
may disrupt reconnection in nearby X-lines in thin current sheets. Thus, if something (such as guide magnetic field) can cause reconnection to build up plasmoids faster than they decay, that may prevent interference with X-points, maintaining reconnection in a more 2D-like fashion.
\item Whereas 2D reconnection yields a highly-structured plasmoid hierarchy, in which large (system-size-scale), highly-magnetized plasmoids do not obliterate nearby thin (kinetic-scale) elementary current sheets, 3D reconnection tends to develop a thick, turbulent layer, throughout which the magnetic field is greatly diminished.
(\S\ref{sec:overview2d},~\ref{sec:overview3d})
\item In 3D, with weak guide field, the depletion of upstream magnetic energy and flux may
start rapidly (as in 2D, or perhaps a little more slowly), but then
continue at an order-of-magnitude slower rate for an
order-of-magnitude longer time than in 2D.
This may be due to the thickened layer, and/or to reconnection
occurring only in sparse, small areas of the layer.
Alternatively, true reconnection may operate to a lesser extent in 3D,
with other, slower mechanisms (perhaps turbulent diffusion)
depleting magnetic field.
(\S\ref{sec:pertAndEta3d},~\ref{sec:Lz3d})
\item In 2D, the current sheet evolution from initial to final state
(in a closed system) is in a broad sense
inevitable, despite stochastic plasmoid behaviour.
In 3D, however, stochastic variability in the
early current sheet evolution can lead to substantially different
states at much later times
(e.g., different layer thicknesses with different amounts
of magnetic energy converted to plasma energy and different
continuing rates of energy conversion).
(\S\ref{sec:variability3d}, ~\ref{sec:RDKIamp})
\item 2D reconnection (in a closed system) ceases when the major plasmoid become large (of order of the system size) and a stable magnetic configuration is realized. In 3D, we have not observed the cessation of magnetic energy depletion, even after 30--50~light-crossing times;
the final 3D state may depend strongly on the early evolution.
(\S\ref{sec:overview2d},~\ref{sec:Lx2d},~\ref{sec:Lz3d})
\item In 3D, increasing the initial guide magnetic field suppresses
3D effects, and
3D reconnection becomes increasingly similar to 2D reconnection.
(\S\ref{sec:Bz3d})
\item In 2D, guide magnetic field suppresses both reconnection and
NTPA. This is also true in 3D; however,
because guide field also suppresses 3D effects (which suppress
reconnection), increasing the guide field enhances reconnection
up to a point where 3D effects are effectively suppressed;
beyond that point, stronger guide field has the same suppression
effect as in~2D.
(\S\ref{sec:Bz2d},~\ref{sec:Bz3d})
\item Despite differences in 2D and 3D current sheet evolution, {\bf NTPA is robust} and remarkably similar in 2D and 3D as well as in initially-similar 3D simulations with manifestly different current sheet evolution. This is the case regardless of guide field.
(\S\ref{sec:Lz3d}, \ref{sec:Bz3d})
\item Moreover, NTPA is robust in a way that depends on the upstream
$\sigma_h=1$ and $B_{gz}/B_0$;
particle energy spectra are steeper than
those in simulations with larger~$\sigma_h$, and stronger guide field results in steeper power-law particle spectra.
(\S\ref{sec:highSigmah})
\end{itemize}
\subsection{Current sheet evolution in 2D}
\label{sec:sum2D}
Besides providing a baseline against which to compare
the~3D simulations, our~2D simulations constitute the first systematic study across a broad swath of the multidimensional parameter space describing~2D, $\sigma_h=1$ (ultrarelativistic pair-plasma) reconnection, which is of interest
in its own right. Here we summarize the most important 2D results.
\begin{itemize}[leftmargin=3mm, itemindent=0mm, labelsep=1mm]
\item Current sheet evolution depends significantly on the ambient~$\sigma_h$ and guide field strength~$B_{gz}/B_0$.
(\S\ref{sec:Bz2d}, \ref{sec:highSigmah})
\item The evolution of magnetic and plasma energy, and resulting NTPA,
is relatively (but not completely)
insensitive to the initial current sheet configuration,
including the initial magnetic field perturbation, the
density, temperature, and thickness of the initial current sheet,
and the drift speed of the current sheet plasma.
Thus reconnection is governed mostly by the ambient background plasma as described below.
(\S\ref{sec:pertAndEta2d})
\item For zero guide field (and, importantly, $\sigma_h=1$), reconnection rates (normalized to $B_0c$) are typically around $\beta_{\rm rec}\approx 0.02$--$0.03$, or (normalized to $B_0v_A$), $(c/v_A)\beta_{\rm rec}\approx 0.03$--$0.04$, the same order of magnitude but still significantly less than the often-assumed value of~0.1 \citep{Cassak_etal-2017}. [Values of $(c/v_A)\beta_{\rm rec}\approx 0.1$ are actually realized in 2D reconnection in relativistic pair plasma with $\sigma_h \gg 1$ \citep[e.g.,][]{Sironi_Spitkovsky-2014,Guo_etal-2015,Werner_etal-2018,Sironi_Beloborodov-2020}.]
(\S\ref{sec:Lx2d})
\item The magnetic energy evolution in systems with $L_x \gtrsim 160\sigma\rho_0$ already appears to
depend only very weakly on system size $L_x$ (for $B_{gz}=0$, up to $L_x=2560\sigma\rho_0$), although larger simulations experience slightly
slower reconnection rates than smaller simulations.
(\S\ref{sec:Lx2d})
\item For $B_{gz}=0$ (and, importantly, $\sigma_h=1$),
we observe NTPA (for $\sigma_h=1$) with a steep power-law slope
of roughly $f(\gamma)\sim \gamma^{-4}$.
(\S\ref{sec:Lx2d})
\item NTPA also has a fairly weak dependence on $L_x$
for $L_x\gtrsim 160\sigma\rho_0$;
although precise measurement of steep power laws is difficult,
our results (for $B_{gz}=0$)
suggest that the power-law index varies within only about
10~per~cent for $80 \lesssim L_x/\sigma\rho_0 \lesssim 2560$.
The maximum particle energy (or the high-energy cutoff of the
particle power-law energy spectrum) clearly increases
sublinearly with~$L_x$,
from around~$7\sigma$ for $L_x=80\sigma\rho_0$ to
perhaps~$33\sigma$ for $L_x=2560\sigma\rho_0$.
I.e., for a 32X increase in system size, the particle cutoff energy
increases by~$\sim$5X,
consistent with a~$\sim\sqrt{L_x}$ scaling
\citep[cf. ][]{Petropoulou_Sironi-2018,Hakobyan_etal-2021}.
(\S\ref{sec:Lx2d})
\item Guide magnetic field slows reconnection.
Increasing the guide field from zero to $B_{gz}=4B_0$ reduces the
reconnection rate
(normalized to $B_0c$; cf.~\S\ref{sec:unreconnectedFlux}) from
$\beta_{\rm rec}\approx 0.03$ to~$0.007$.
This is roughly (within 30~per~cent) consistent with
$\beta_{\rm rec} \approx 0.04 v_{A,x}/c$ where
$v_{A,x}^2/c^2=\sigma_{h,\rm eff}/(1+\sigma_{h,\rm eff})$ and
$\sigma_{h,\rm eff}\equiv (1/\sigma_h + B_{gz}^2/B_0^2)^{-1}$.
(\S\ref{sec:Bz2d})
\item Guide magnetic field inhibits NTPA,
yielding steeper power-law spectra.
The power-law index
steepens from~$\sim \gamma^{-4}$ at~$B_{gz}=0$ to~$\sim \gamma^{-6}$
at~$B_{gz}=B_0$, and it continues to be steeper for higher~$B_{gz}$.
Keeping in mind the difficulty of measuring steep power laws,
we find that the power-law index is roughly
$p\approx 4.4 + 2B_{gz}/B_0$; this should be taken as a rough guide
and not confirmation of a linear dependence on~$B_{gz}$.
For $B_{gz}/B_0 \gtrsim 1.5$, we measure $p\gtrsim 8$
and it is debatable whether the spectra are significantly nonthermal.
(\S\ref{sec:Bz2d})
\end{itemize}
\subsection{Current sheet evolution in 3D}
\label{sec:sum3D}
Our most important results for 3D simulations with $\sigma_h=1$
(with system sizes in the range $256 \leq L_x/\sigma\rho_0\leq 512$):
\begin{itemize}[leftmargin=3mm, itemindent=0mm, labelsep=1mm]
\item 3D effects can be introduced gradually by increasing the length
$L_z$ of the system in the third dimension. We find that,
for $B_{gz}=0$ and the range of system sizes considered, simulations
with $L_z\gtrsim L_x/8$ can exhibit fairly
obvious departures from 2D reconnection,
while those with $L_z\lesssim L_x/16$ almost always closely resemble 2D.
(\S\ref{sec:overview3d},~\ref{sec:Lz3d})
\item As $L_z/L_x$ is increased, 3D effects first appear in
the time evolution of the magnetic energy in the layer,
$U_{Bt,\rm layer}(t)$.
(The rest of this point applies to~$B_{gz}=0$ and~$a=0$.)
In 2D or very small $L_z/L_x \ll 1/32$,
$U_{Bt,\rm layer}(t)$ increases as long as reconnection
continues (up to around 25~per~cent of $U_{B0}$),
and does not decrease.
In 3D, $U_{Bt,\rm layer}(t)$
increases at very early times, but then starts to decrease because
plasmoids can disintegrate in~3D.
Even when~$L_z/L_x$ is small enough to avoid
other departures from 2D reconnection, $U_{Bt,\rm layer}$ may be
significantly less than in~2D.
The larger~$L_z/L_x$ is, the smaller~$U_{Bt,\rm layer}(t)$ is
(at any given~$t$), up to around $L_z/L_x\lesssim 1$.
For $L_z/L_x \gtrsim 1$, $U_{Bt,\rm layer}$ may rise to just a couple
per~cent of~$U_{B0}$ before falling to lower values.
(\S\ref{sec:overview3d},~\ref{sec:Lz3d})
\item
Random variability is substantially greater in 3D than in 2D;
identical simulations, aside from different random particle
initialization, can yield very different behaviour
(e.g., in terms of magnetic energy versus time) and very different
states even after long times. This contrasts strongly with the
variability in 2D, where similarly identical simulations will still
yield the same magnetic energy release
versus time (when measured over long timescales)
and end up in nearly the same final state after the same amount
of time. Although more familiar, it is the 2D case that is more
remarkable here---the stable plasmoids trap energetic plasma from reconnection outflows and prevent it from interfering with regions upstream of thin elementary current sheets even though the plasmoids are orders of magnitude larger.
(\S\ref{sec:variability3d})
\item
3D effects increase the sensitivity to the details of the initial current sheet configuration.
(\S\ref{sec:pertAndEta3d})
\item An initial magnetic field perturbation, uniform in $z$, can
suppress 3D effects on reconnection (at least for $B_{gz}=0$), especially at earlier times.
For this reason, our 3D simulations were
initialized with no perturbation (i.e., $a=0$) except when specifically studying the effect of~$a$.
(\S\ref{sec:pertAndEta3d})
\item For $B_{gz}=0$, increasing the initial current sheet overdensity $\eta$
(while maintaining pressure balance) might weakly increase the early-time reconnection rate and make reconnection more 2D-like, but a larger
ensemble of simulations will be needed to measure this effect beyond
stochastic variation.
(\S\ref{sec:pertAndEta3d})
\item For weak guide field, RDKI modes cause the current sheet to ripple. In the linear phase of the instability, a mode grows to a rippling amplitude comparable to its wavelength, without releasing much magnetic energy. The mode may subsequently enter the nonlinear phase of the instability, during which the current sheet becomes highly distorted, folding over on itself and rapidly releasing magnetic energy within the rippling amplitude.
This can occur even in a 2D~$y$-$z$-plane simulation (but not in the 2D~$x$-$y$ geometry used to study reconnection), and
the magnetic energy depletion can occur as fast as in
2D reconnection, even though driven by a manifestly
different large-scale mechanism.
(\S\ref{sec:RDKIamp}, especially Fig.~\ref{fig:kinkAmpVsBetad3d})
\item For $B_{gz}=0$, varying the initial current sheet drift speed, $\beta_d c$ (while keeping~$\eta \beta_d$ hence~$\delta/\sigma\rho_0$ fixed), has
a dramatic effect on current sheet evolution due to its influence
on the RDKI. Increasing $\beta_d$ increases
the growth rate of longer-wavelength RDKI modes, which can release more magnetic energy (than shorter-wavelength modes) if they grow fast enough to enter the nonlinear stage---simply because they can grow to larger amplitude.
(\S\ref{sec:pertAndEta3d},~\ref{sec:RDKIamp})
\item
The influence of~$\beta_d$ is only partly explained by its effect on
the linear-phase growth rate.
A slower-growing,
longer-wavelength mode may overtake a faster-growing,
shorter-wavelength mode because the latter saturates nonlinearly at smaller
amplitude.
The most influential mode may have much longer wavelength than the
(linearly) most unstable RDKI mode.
(\S\ref{sec:RDKIamp},~\ref{sec:Lz3d})
\item The nonlinear development of RDKI may play a key role in the
stochastic variability of 3D current sheet evolution (for weak guide field).
Although in some cases an RDKI mode with long wavelength
$\lambda_{z,\rm RDKI}\gg \delta$
will grow to a large amplitude and enter nonlinear development, rapidly
and dramatically changing the layer structure
while converting magnetic energy to plasma energy,
growth to the nonlinear stage is not inevitable.
Nonlinear RDKI development may be triggered, e.g., in only a fraction of an ensemble of
identically-initialized simulations
(up to random particle initialization),
thus putting them on very different evolutionary tracks.
(\S\ref{sec:RDKIamp},~\ref{sec:Lz3d}, Fig.~\ref{fig:layerOffsetY})
\item For $B_{gz}=0$, the reconnection rate (normalized to $B_0 v_A$; cf.~\S\ref{sec:unreconnectedFlux}) in the early stage
of 3D development (typically $\lesssim 10L_x/c$) varies from
at most $(c/v_A)\beta_{\rm rec}\sim 0.05$ (at small $L_z/L_x$, the same as
2D) to as little
as~$(c/v_A)\beta_{\rm rec}\sim 0.005$ for large $L_z/L_x$;
however, stochastic evolution results in large variability, and
even for large~$L_z/L_x$, higher rates may sometimes be observed.
(\S\ref{sec:Lz3d})
\item (For $B_{gz}=0$) For $L_z/L_x\gtrsim 1/8$, most simulations exhibit an early stage
of fast magnetic energy release
(but nonetheless slower than in 2D) lasting
tens of~$L_x/c$, followed by a slower stage with reconnection
rates (or more precisely, upstream flux depletion rates)
an order of magnitude lower, i.e., $(c/v_A)\beta_{\rm rec}\sim 0.0005$.
The slower stage can last at least up to~$50L_x/c$, and perhaps
much longer.
(\S\ref{sec:Lz3d})
\item Our results are consistent with the suggestion
of \citet{Yin_etal-2008} that local reconnection rates are as high as
in 2D, but in 3D only a relatively small area of the layer undergoes
active reconnection, resulting in a low global reconnection rate.
(However, we did not measure local reconnection rates.)
(\S\ref{sec:overview3d})
\item NTPA is nearly the same
in 3D as in 2D, if compared at times when the same amount of
magnetic energy has been converted to plasma energy.
It is equally remarkable that NTPA is nearly the same in
different 3D simulations that, despite being macroscopically identical at~$t=0$, undergo very different current sheet
evolution (and 2D $y$-$z$ simulations that suffer
large-wavelength nonlinear RDKI also yield similar NTPA).
(\S\ref{sec:RDKIamp}, \ref{sec:Lz3d}, \ref{sec:Bz3d})
\item For $B_{gz}=0$ (and, importantly, $\sigma_h=1$),
particle energy spectra show power laws
$f(\gamma)\sim \gamma^{-p}$ with $p \in [3.4,4.6]$.
NTPA actually appears to be slightly
more efficient in 3D than in 2D, sometimes yielding slightly harder power
laws (e.g., $\gamma^{-3.4}$), and with the fraction of particles
attaining $\gamma>10\sigma$ increasing by almost an order of magnitude
from 2D (or $L_z/L_x \ll 1/8$) to 3D ($L_z/L_x \gtrsim 1/8$).
Enhanced NTPA in~3D has been previously supposed to result from the weaker trapping of particles in plasmoids which, in 2D, ends particle acceleration \citep{Dahlin_etal-2015,Dahlin_etal-2017,Li_etal-2019,Zhang_etal-2021arxiv}.
However, because of the steep power law, these differences in NTPA
between~2D and~3D may be
ultimately unimportant relative to other effects and uncertainties.
(\S\ref{sec:Lz3d})
\item Weak guide field can enhance 3D reconnection because it suppresses 3D effects (that slow reconnection); this is noticeable in reconnection rates and the evolution of magnetic energy even for weak guide field, $B_{gz}/B_0\gtrsim 0.1$.
However, as the guide field becomes strong enough to suppress 3D effects substantially, reconnection behaves as in 2D, where increasing the guide field slows reconnection and inhibits NTPA. We believe that, for $\sigma_h=1$, guide field causes 3D reconnection to resemble 2D reconnection strongly around $B_{gz}/B_0 \gtrsim 1$ (but see the parenthetical caveat in the next bullet).
(\S\ref{sec:Bz3d})
\item
Stronger guide field therefore raises the
reconnection rate as long as 3D effects are more suppressive than
guide field effects. The reconnection rate rises from
$\beta_{\rm rec}\approx 0.002$ at~$B_{gz}=0$ to a maximum
around $\beta_{\rm rec}\approx 0.02$ for
$0.25 \lesssim B_{gz}/B_0 \lesssim 0.75$.
For even stronger guide field
the reconnection rate falls (however, because of system-size
limitations, we explored only up to $B_{gz}/B_0=1$ in~3D, with just a
single simulation for each value of~$B_{gz}/B_0$, and so this
conclusion needs to be confirmed by future studies with better
statistics and higher~$B_{gz}$).
(\S\ref{sec:Bz3d})
\item For each value of $B_{gz}/B_0$ investigated ($0\leq B_{gz}/B_0 \leq 1$), NTPA is practically the same in 3D as in 2D.
(\S\ref{sec:Bz3d})
\end{itemize}
\section{Discussion}
\label{sec:discussion}
\subsection{Differences between 2D and 3D: comparison with previous works}
\label{sec:discussionPrevWork}
Perhaps the most important general observation of this work is that, in the $\sigma_h=1$ regime, thin current sheets can evolve very differently in 2D and in 3D. However, they do not always evolve differently. In particular, a guide field---possibly even $B_{gz}/B_0\gtrsim 0.25$---can suppress 3D effects, and so can a small initial perturbation that is uniform in the third dimension.
Especially in cases with very weak guide field, 3D systems might not be well-modelled by 2D simulations. Important consequences or signatures of 3D effects include slower rates of magnetic-to-plasma energy conversion and disintegration of plasmoid or flux rope structures (which releases additional magnetic energy). Moreover, magnetic energy can be converted to plasma energy via a completely different mechanism---the nonlinear development of the RDKI (see~\S\ref{sec:discussionRDKI}). Nevertheless, despite these 3D effects, NTPA remains essentially the same in 3D as in 2D.
As pointed out in~\S\ref{sec:highSigmah}, the dramatic differences in~3D might be a bit of a surprise because they were much less evident in reconnection in the~$\sigma_h\gg 1$ regime, despite clear presence of RDKI.
However, some of these effects have been previously observed in~$\sigma_h\sim 1$ pair-plasma PIC simulations.
Of particular interest for their close relevance and identification of similar 3D effects, are: \citet{Yin_etal-2008}---hereafter, Yin08; \citet{Liu_etal-2011}---Liu11; and \citet{Kagan_etal-2013}---K13.
Yin08, Liu11, and K13 all used PIC simulation to study 3D reconnection in pair plasma with~$\sigma_h\sim 1$; simulation parameters are compared in table~\ref{tab:simParams}. Yin08 initialized subrelativistic plasma with $\theta_b=0.016$ and $\sigma_h\approx 0.2$ in simulations of size $L_x=240\sigma\rho_0$ and $L_z\leq L_x$, while Liu11 and K13 both used transrelativistic plasma with $\theta_b=1$ and $\sigma_h \in [0.5, 1.5]$; Liu11 had $L_x=60\sigma\rho_0$ and $L_z\leq L_x$, and K13 studied two configurations, one with $L_x=52\sigma\rho_0$ and $L_z=1.6L_x$ and another with $L_x=110\sigma\rho_0$ and $L_z=1.1L_x$.
None of these simulations used any initial perturbation. K13 was the only one that did not use a standard Harris-sheet set-up; it was also the only one to study the effect of guide field.
\begin{table}
\caption{\label{tab:simParams}
Approximate (ranges of) parameters for the largest 3D
simulations of Yin08, Liu11, and K13 (with two series of simulations, S1 and S2), translated to the terminology of this paper; values for 3D simulations of this paper are also shown. Parameters not defined in~\S\ref{sec:setup} or table~\ref{tab:etaEffect} are as follows: $v_A$ is the upstream Alfv\'{e}n velocity;
$\beta_{\rm plasma} \equiv 8\upi n_b \theta_b m_e c^2 / B_0^2$ is the upstream plasma beta;
$d_{e}^{\rm NR}\equiv (m_e c^2/4\upi n_b e^2)^{1/2}$ and $d_{e,d}^{\rm NR}\equiv (m_e c^2/4\upi n_d e^2)^{1/2}$ are the nonrelativistic collisionless background and drifting plasma skin depths (these nonrelativistic scales are not applicable to this paper);
$\ell_y$ is $L_y$ divided by the number of layers simulated;
$t_{\rm run}$ is the final simulation time.
}
\begin{tabular}{llllll}
& Yin08 & Liu11 & K13(S1) & K13(S2) & this work \\
\hline
$\sigma$ & 0.21 & 6.7 & 2 & 4 & $10^4$ \\
$\theta_b$ & 0.016 & 1 & 1 & 1 & $2.5\times 10^3$ \\
$\sigma_h$ & 0.20 & 1.5 & 0.46 & 0.92 & 1.0 \\
$v_A/c$ & 0.41 & 0.78 & 0.56 & 0.69 & 0.71 \\
$\beta_{\rm plasma}$
& 0.15 & 0.30 & 1 & 0.5 & 0.5 \\
$B_{gz}/B_0$ & 0 & 0 & 0--1 & 0--1 & 0--1 \\
\hline
$\eta=n_d/n_b$ & 3.3 & 3.3 & $^* 1.5$ & $^* 2.3$ & 0.5--5 \\
$\theta_d/\theta_b $ & 1 & 1 & $^*$ ? & $^*$ ? & 0.4--5\\
$\beta_d$ & 0.36 & 0.82 & $^*\!\lesssim 0.36$ & $^*\!\lesssim 0.36$ & 0.075--0.6 \\
\hline
$\delta/\sigma\rho_0$ & 0.18 & 0.21 & 3.7 & 3.9 & 0.67--3.3 \\
$\rho_b/\sigma\rho_0$ & 0.96 & 0.48 & 1.6 & 0.79 & 0.75 \\
$d_{e}/\sigma\rho_0 \phantom{|}^\dagger$
& 2.2 & 0.72 & 1.3 & 0.92 & 0.87 \\
$d_{e}^{\rm NR}/\sigma\rho_0$
& 2.2 & 0.39 & 0.71 & 0.50 & N/A
\\
$d_{e,d}^{\rm NR}/\sigma\rho_0 \phantom{}^\ddag$
& 1.2 & 0.21 & $^*$0.58 & $^*$0.33 & N/A
\\
\hline
$L_x/\sigma\rho_0$ & 240 & 60 & 52 & 110 & 256--512 \\
$\ell_y/L_x$ & 1 & 1 & 0.5 & 0.5 & 1 \\
$L_z/L_x$ & 0--1 & 0--1 & 1.6 & 1.1 & 0--1.5 \\
$t_{\rm run} v_A/L_x$ & 4.1 & 5.0 & $> 4.2$ & 2.6 & 7--35\\
\hline
\end{tabular} \\
$^*$K13 used a uniform density~$n_b$ with varying $\beta_d=\beta_d(y)$ to satisfy Ampere's law (without a separate drifting component), but pressure balance was not initially satisfied, resulting in an immediate compression of the current sheet by a factor~$\simeq\eta$ (and unknown temperature increase).
\\
$^\dagger$K13 referenced lengths to $\lambda_p \equiv d_{e}$.
\\
$^\ddag$Yin08 referenced lengths to $d_i\equiv d_{e,d}^{\rm NR}$.
\\
$^\ddag$Liu11 referenced lengths to $d_i\equiv \sqrt{2} d_{e,d}^{\rm NR}$.
\end{table}
All these previous $\sigma_h\sim 1$ studies observed tearing and reconnection as well as an initial linear growth and later nonlinear growth of RDKI modes.
Yin08 and Liu11 report similar reconnection rates
with no clear/significant difference between~2D and~3D, measuring
$(c/v_A)\beta_{\rm rec}\approx 0.06$--0.08 (with respect to the magnetic field~$B$ averaged over the ``inflow surface'' rather than the asymptotic upstream~$B_0$);
K13 reports 3D rates of $(c/v_A)\beta_{\rm rec}\approx 0.05$--0.08.
These reconnection rates for~$\sigma_h\sim 1$ are all roughly half of $(c/v_A)\beta_{\rm rec}$ measured for~$\sigma_h\gg 1$,
and they are consistent with our results for~$L_z\ll L_x$.
However, for $L_z\sim L_x$ we measured significantly lower values with $(c/v_A)\beta_{\rm rec} \lesssim 0.02$ (see Fig.~\ref{fig:reconRateVsLz}).
This may be because the previous works measured the rates in smaller systems at earlier times, or because (for Yin08 and Liu11) they normalize to a local magnetic field smaller than~$B_0$.
In addition to calculating similar reconnection rates in~2D and~3D, Yin08 shows very similar magnetic energy evolution in~2D and~3D
while hinting that the energy depletion appeared to be ``somewhat slower'' in the simulation with largest~$L_z$ because of the limited size of the reconnecting patch.
Of these three, Liu11 is the only one that (like us) specifically reports an early stage of faster magnetic energy depletion (attributed to tearing/reconnection) \emph{followed by a stage of slower depletion} (attributed to nonlinear RDKI development). In~Liu11, the slower stage results in significantly more release of magnetic energy in~3D than in~2D---a signature of flux rope disintegration (in contrast, Yin08 sees similar magnetic depletion in~2D and~3D; K13 shows $U_{B}(t)$ only for $B_{gz}=B_0/4$, and not for a very long time, but if anything, the rate of magnetic depletion speeds up at later times).
Both Yin08 and Liu11 (and possibly K13) appear to observe more growth and merging of flux ropes in~3D than we do; however, Liu11 and K13 run long enough to see the end of the ``period'' of flux rope merging, leaving a much less structured, more turbulent layer (which we have also observed).
Apparent qualitative differences among the various simulations may have more to do with simulation size than any fundamental difference; in fact, K13 reports (for $B_{gz}=B_0/4$) that in a smaller simulation flux ropes merge into a final, single flux rope (as in~2D), while in a larger simulation this merging is disrupted before reaching a single, large flux rope.
Although these simulations all ran in somewhat different regimes with different initial current sheet configurations (which, we have shown, can make a difference---see~\S\ref{sec:pertAndEta3d}--\ref{sec:RDKIamp}), the combined results appear consistent with the notion that as systems become increasingly larger, the stage of 2D-like reconnection becomes increasingly transient.
This may suggest that the later, slower stage---with significant~3D effects---is more relevant for astrophysically-large systems.
All these papers also observe nonlinear RDKI in some form, but always following a stage of tearing and reconnection (whereas we have shown that for some initial current sheet configurations, RDKI may drive the earliest stage of rapid magnetic energy release; see~\S\ref{sec:RDKIamp}). Interestingly, in Yin08, a long (in~$z$) current sheet forms between growing flux ropes, undergoes extreme distortion and folding via nonlinear RDKI (which Yin08 calls ``secondary'' kinking), \emph{and subsequently reforms} as a thin, reconnecting current sheet, suggesting such behaviour may be cyclic.
Liu11 also observes ``self-organization'' of initially-patchy reconnecting regions into a ``highly-elongated'' current sheet between growing flux ropes, although this sheet appears to be permanently disrupted by nonlinear RDKI (i.e., it does not reform as in Yin08).
Both Liu11 and K13 investigated particle acceleration and report nonthermal particle energy spectra.
The spectra are consistent with power laws~$\gamma^{-p}$ with (very roughly), for Liu11, $p\approx 2.5$ up to a cutoff $\gamma_c\sim 4\sigma$, and for K13, $p\approx 3.5$ up to $\gamma_c\sim 7\sigma$ (for the largest simulation, labelled S2K025L, with $\sigma=4$, $B_{gz}/B_0=0.25$, and $L_x=110\sigma\rho_0$; for a smaller simulation, $L_x=55\sigma\rho_0$, the spectra were nearly identical for $B_{gz}/B_0=0$ and~$0.25$).
Because the cutoffs of power laws can be misleading to compare for different power-law slopes, we also note a more straightforward measure of the (high) energy at which $f(\gamma)$ becomes small, namely the value~$\gamma_{-4}$ for which $f(\gamma_{-4})\equiv 10^{-4} \textrm{max}_\gamma f(\gamma)$; for Liu11, $\gamma_{-4}=7.1\sigma$ and for K13 (S2K025L), $\gamma_{-4}=7.8\sigma$; for our 3D simulations with $L_x=512\sigma\rho_0$, $\gamma_{-4}=10\sigma$.
Considering the somewhat different parameters, including system size and run time, these results seem reasonably consistent.
Liu11 and (to a lesser extent) K13 might measure less steep spectra than we do, but the cause of this could easily be the larger~$\sigma_h$ in Liu11 and/or the sub- or trans-relativistic upstream plasma, and/or the smaller system sizes and different initial current sheet configurations.
K13 investigated the effect of guide field, running simulations with $B_{gz}/B_0=0,$ 0.25, and~1. Although K13 does not report on relative reconnection or magnetic energy depletion rates, there is a suppression of NTPA for $B_{gz}/B_0=1$ relative to the weaker guide fields, also consistent with our results.
Thus it appears that the most important 3D effects that we have observed in an evolving current sheet are consistent with previous observations in similar parameter regimes, although the 3D effects have generally become clearer with our larger system size and scan over a broader range of simulation parameters.
Importantly, we have studied 3D effects with respect to different current sheet configurations, which can dramatically alter the strength of these effects, especially at early times.
In particular, we have highlighted the importance of (sufficiently fast-growing) long-wavelength RDKI modes---namely that they can grow to large amplitudes before entering the nonlinear stage in which the current layer can be dramatically transformed while rapidly releasing magnetic energy (see~\S\ref{sec:RDKIamp} and~\S\ref{sec:discussionRDKI}).
In addition, we have shown that NTPA can be driven by 2D-like reconnection, or the slower stage of reconnection (and possibly RDKI), or by rapid nonlinear RDKI, yielding very similar particle energy spectra in all cases.
Having discussed the $\sigma_h\sim 1$ regime, we briefly comment on differences between 3D reconnection in the~$\sigma_h=1$ regime compared with the~$\sigma_h \gg 1$ regime, where 3D reconnection has been observed to be quite similar to 2D reconnection \citep[e.g.,][]{Sironi_Spitkovsky-2014,Guo_etal-2014,Guo_etal-2015,Werner_Uzdensky-2017,Guo_etal-2020arxiv}---more details were given in~\S\ref{sec:highSigmah}.
Based on these $\sigma_h\gg 1$ studies, it is likely that large~$\sigma_h$ tends to enhance reconnection and/or suppress 3D reconnection-disrupting effects.
However, most of the large-$\sigma_h$ studies have used an initial perturbation of some kind, which we have shown can suppress 3D effects, especially at early times.
The extent of the role of the initial perturbation requires further study; regardless of the results, however, what is really needed is a better understanding of system-size dependence in 3D.
In the~$\sigma_h\sim 1$ regime, a 2D-like reconnection stage may be very transient in larger simulations; and due to resolution requirements, feasible simulations with $\sigma_h\sim 1$ can be relatively larger than those with $\sigma_h \gg 1$. It may be, therefore, that larger 3D simulations with $\sigma_h\gg 1$ begin to show distinctly 3D effects---such as the slowing of reconnection and disruption by RDKI, as well as the possibility of magnetic energy release through RDKI---at much larger sizes than anyone has yet simulated.
\subsection{Another magnetic energy release mechanism: nonlinear RDKI}
\label{sec:discussionRDKI}
We have found that, in at least some cases, a thin~3D current sheet can release magnetic energy by a mechanism that may have nothing to do with magnetic reconnection---namely the \emph{nonlinear} RDKI, which causes a highly-kinked (or rippled) current sheet to fold over on itself \citep[see~\S\ref{sec:RDKIamp};][]{Zenitani_Hoshino-2007}.
This mechanism can operate in 2D $y$-$z$ geometry (perpendicular to the magnetic field), where large-scale magnetic reconnection is impossible and upstream flux is directly annihilated, and it can release magnetic energy as fast as 2D reconnection while generating similar NTPA.
Importantly, the earlier \emph{linear} development of RDKI releases little magnetic energy and does not yield significant NTPA (and it might not even interfere with reconnection).
Because the nonlinear RDKI causes the rippling to saturate at an amplitude of the order of the instability wavelength, longer-wavelength RDKI modes can reach larger amplitudes and thus have greater impact, even if they grow more slowly; but of course they still have to grow enough to reach the nonlinear stage to trigger rapid magnetic energy release.
While it is possible that, in the nonlinear RDKI development, small-scale reconnection (or, alternatively, turbulence-enhanced magnetic diffusion) plays a role in releasing magnetic energy, it is clearly RDKI at large scales that drives the process of energy conversion.
As an alternate channel for releasing magnetic energy, RDKI may offer a neat solution to the ``triggering'' or ``onset'' problem faced by reconnection \citep[e.g.,][]{JiAstro2020Decadal-2019etal}.
This is the problem of reconciling the slow timescale of current sheet formation and magnetic energy build-up with the fast timescale of magnetic energy release; the current sheet must be relatively stable as it slowly forms, until at some point it suddenly becomes unstable---e.g., reconnection is triggered and releases energy rapidly \citep[e.g.,][]{Pucci_Velli-2014,Tenerani_etal-2016,Uzdensky_Loureiro-2016,Comisso_etal-2017,Huang_etal-2017,Huang_etal-2019}.
With reconnection, solving this problem likely requires understanding the mechanism behind current sheet formation.
With RDKI, there is a simpler explanation: the current sheet may form and RDKI (at sufficiently long wavelength) may grow slowly, because the linear stage of RDKI does not release much magnetic energy.
The trigger occurs when the RDKI amplitude becomes comparable to its wavelength, and nonlinear development begins.
However, to understand its potential astrophysical importance, it will be critical to determine the dependence of the nonlinear RDKI on system size.
For any given wavelength, the nonlinear RDKI cannot continue forever; it dissipates magnetic energy within a layer of thickness (at most) comparable to the wavelength. It might not continue to release upstream magnetic energy in the way that 2D reconnection does (as discussed in~\S\ref{sec:RDKIamp}, one might suppose 2D reconnection would saturate after plasmoids grow to the size of the elementary current sheet; however, simulations show that plasmoids detach and move away without disrupting the elementary current sheets, so that reconnection does not saturate until plasmoids grow to the system-size scale). However, RDKI could potentially continue to develop at different, longer wavelengths.
For instance, an RDKI mode with wavelength~$\lambda_{z,\rm RDKI}$ may grow in a current sheet of thickness~$\sim \delta$; if $\lambda_{z,\rm RDKI}\gg \delta$ and the mode reaches an amplitude of order~$\lambda_{z,\rm RDKI}$, nonlinear development may release the magnetic energy within a layer of thickness~$\delta' \sim \lambda_{z,\rm RDKI}$, resulting in a new current sheet of greater thickness,~$\delta' \gg \delta$.
Because shorter-wavelength modes are limited to smaller amplitudes, longer-wavelength modes tend to be more influential.
Indeed, in some of our simulations, current sheet evolution was eventually dominated by the RDKI mode with the longest-possible wavelength, $\lambda_{z,\rm RDKI}\sim L_z$; if this behaviour continues up to systems of arbitrary size, nonlinear RDKI could potentially release astrophysically-large amounts of energy.
This could occur in one fell swoop, or in a series of stages (bursts) cascading up to the system size: the thickened layer with $\delta'$ might subsequently be destabilized by a new mode with $\lambda_{z,\rm RDKI}' \gg \delta'$, which might grow enough to enter the nonlinear phase, suddenly releasing energy in a layer of thickness $\delta'' \sim \lambda_{z,\rm RDKI}'$, etc., until the entire system is substantially depleted of magnetic energy.
Alternatively, this process could slow to an effective halt well before reaching the system size, with further energy release requiring other processes to thin out the layer (e.g., the processes responsible for the current sheet formation in the first place).
Whether and how fast nonlinear RDKI, operating in sufficiently large systems, can release astrophysically-large amounts of energy, are important questions that we leave to future research.
\subsection{Observable astrophysical consequences}
In this paper we have investigated the plasma dynamics of current sheet evolution as well as resulting NTPA, which are unlikely to be directly observable in any astrophysical source; however, synchrotron and inverse Compton emission from high-energy particles may be observable.
Although we must leave a detailed study of radiation to future work, we can infer some observable consequences under simplifying assumptions.
Synchrotron and inverse Compton emission from high-energy electrons and positrons will depend on the shape and normalization of the distribution of accelerated particles.
Importantly, we have shown that the magnetization~$\sigma_h$ and guide field strength~$B_{gz}/B_0$ strongly affect both energy conversion rates and the NTPA spectrum; on the other hand, the dimensionality (2D vs.~3D) and initial current sheet configuration can affect energy conversion rates but, at most, weakly affect the NTPA spectrum.
This is a double-edged sword.
On one hand, we may be able to infer~$\sigma_h$ and~$B_{gz}/B_0$ from observations conveniently without knowing other details.
On the other hand, we may not be able to determine less influential details of the current sheet or even the effective dimensionality of astrophysical current sheets from observations. Here we say ``effective dimensionality'' because we have seen that, under some circumstances---e.g., strong guide field, small $L_z/L_x$, or perhaps some sort of perturbation that encourages uniformity in~$z$---a 3D current sheet can evolve as if it were in~2D.
The most basic observable signatures are simply the radiation intensity and (for a flaring event) duration.
For this reason the study of astrophysical reconnection has been justifiably obsessed with measuring and understanding the reconnection rate to determine whether reconnection can explain rapid magnetic energy releases.
We have observed that the reconnection rate in~2D simulations is about what would be expected for ``fast'' reconnection, i.e., $(c/v_A)\beta_{\rm rec}\sim 0.1$.
If more precision is desired, we have shown that the actual value is somewhat less than~0.1 for $\sigma_h=1$ and $B_{gz}=0$, and guide field can slow the rate even more, up to about a factor of~3 for strong guide field.
The reconnection rate in~3D can be substantially lower, leading to lower radiated power and longer duration (for a given source size); depending on precisely how 3D reconnection is triggered, it may yield two stages (as in our simulations), an initial flare somewhat less bright than one would expect from 2D simulations, following by a much longer, much dimmer afterglow.
Observed intensity could thus be an important diagnostic of effective dimensionality.
Unfortunately, it is complicated by the sensitivity to initial current sheet configurations, and also the difficulty of relating the total radiated power to the observed intensity at a particular viewing angle (discussed below).
The next most basic signature is probably the spectral energy distribution of observable radiation.
The photon spectra are a consequence of the particle spectra and should therefore depend on~$\sigma_h$ and~$B_{gz}/B_0$, but may be fairly insensitive to dimensionality and initial current sheet details.
For instance, for~$\sigma_h=1$ and $B_{gz}/B_0= 0$, we have seen that the particle energy power-law index is around~$p\approx 4$, and we can infer (assuming a steady-state, uniform, isotropic, weakly-radiative system) that the emitted photon energy power-law index would be around~$\alpha = (p-1)/2 \approx 1.5$.
For stronger~$B_{gz}$, the photon spectrum would steepen accordingly.
However, more detailed computation of emitted radiation from self-consistent PIC simulations is needed to determine whether these conclusions are valid outside of a much-simplified radiation emission model.
In particular, we need to consider the spatial- and angular-dependence of accelerated particles.
In addition, if radiation is very strong/efficient, the radiation reaction force will alter the high-energy particle distributions and possibly even the current sheet evolution.
Because ultrarelativistic particles emit synchrotron and inverse Compton emission narrowly beamed around their directions of motion, the angular dependence of particle spectra may have important consequences for observation of radiation along a particular line of sight \citep[see, e.g.,][]{Jaroschek_Hoshino-2009,Cerutti_etal-2012a,Cerutti_etal-2012b,Cerutti_etal-2013,Cerutti_etal-2014a,Cerutti_etal-2014b,Kagan_etal-2013,Christie_etal-2018,Werner_etal-2019,Mehlhaff_etal-2020,Sironi_Beloborodov-2020}.
For example, kinetic beaming---i.e., a strong energy-dependent anisotropy of particle spectra and hence radiation---observed in high-$\sigma_h$, 2D reconnection \citep{Cerutti_etal-2012b,Mehlhaff_etal-2020} may be sensitive to 3D effects and/or the current sheet configuration, even if the isotropically-averaged spectra are not.
Indeed, kinetic beaming has previously been observed to be present but weaker in 3D \citep[][]{Kagan_etal-2013,Cerutti_etal-2014b}; however, this issue needs systematic study over a range of $\sigma_h$, current sheet configurations, guide fields, and radiation strengths.
Since we have seen that the spatial distribution of magnetic field can be quite different in 2D and 3D (cf.~Fig.~\ref{fig:BvsXy2D3D}),
the spatial dependence of particle distributions may also be important in determining synchrotron emission.
For example, in 3D the nonlinear RDKI can result in a thickened current layer---a relatively large volume with greatly-diminished magnetic field. As a result, particles accelerated by the nonlinear RDKI may radiate very little within the layer, emitting and cooling only as they exit the layer, when they suddenly experience nearly the full ambient magnetic field. This could affect both the average radiated power (because radiation is suppressed in the layer) and the spectrum (because the magnetic field is mostly either full strength or low strength).
In contrast, there is less reason to believe that overall inverse Compton radiation would be different in 2D and 3D, unless the spectrum of soft photons (to be upscattered by high-energy electrons and positrons) is different---e.g., for synchrotron self-Compton, where the ``soft'' photons may be a result of synchrotron radiation.
If radiative cooling of particles is strong (compared with accelerating forces), the radiation reaction force can alter the spectra of accelerated particles.
This is especially interesting in the case of synchrotron radiation, because particles accelerated by 2D reconnection may experience significant acceleration in regions of low magnetic field, allowing them to exceed the synchrotron burnoff limit \citep[i.e., to exceed the energy at which the radiation reaction force in the ambient/upstream magnetic field would cancel electric-field acceleration,][]{Uzdensky_etal-2011,Cerutti_etal-2012a,Cerutti_etal-2012b}.
The different geometry of 3D current sheet evolution, not to mention different large-scale driving mechanisms (e.g., nonlinear RDKI vs. 2D reconnection), might affect whether and how particles can exceed the burnoff limit.
In \citet{Cerutti_etal-2014b}, particles accelerated by RDKI in 2D$yz$ geometry did not exceed the burnoff limit, although particles accelerated by a 3D evolving current sheet did (although not nearly as much as in 2D$xy$ reconnection); however, that was for one particular current sheet configuration and~$\sigma_h$ (though multiple guide field strengths were investigated).
It is possible that this conclusion would change for other systems, especially considering that we have shown that nonlinear RDKI is dramatically affected by the initial current sheet configuration~(see~\S\ref{sec:pertAndEta3d},~\ref{sec:RDKIamp}).
By showing that 3D effects as well as the initial configuration can substantially affect current sheet evolution, we have demonstrated the need to simulate radiation from these systems to determine observable signatures.
Although the insensitivity of NTPA spectra to these details may result in photon spectra that depend primarily on~$\sigma_h$ and~$B_{gz}/B_0$, this cannot be determined until we also understand the spatial and angular dependence of NTPA spectra and the emitted photon spectra.
\section{Conclusion}
\label{sec:conclusion}
Magnetic reconnection is a plasma process that is important in large part because it converts magnetic energy to plasma (particle) energy.
It may play a key role in relativistically-hot plasmas (including electron-positron pair plasmas) in a variety of astrophysical sources, accelerating particles to very high energies where they can emit observable radiation.
In general, but especially for astrophysical applications, we seek to understand the rate and amount of magnetic energy conversion as well as the development of nonthermal particle energy distributions that can be correlated with emitted radiation.
The primary way in which we can infer the role of reconnection in astrophysical sources, which cannot be directly probed, is to connect observable consequences of reconnection (i.e., radiation) with the source plasma conditions (such as magnetic field and plasma density), which can be inferred therefrom.
Most of the systematic studies of the effects of varying reconnection parameters in astrophysically-relevant relativistically-hot plasma have focused on highly-magnetized regimes.
We add to the literature this study focusing on the moderately-magnetized regime, $\sigma_h=1$, in which the ambient magnetic energy is roughly in equipartition with the ambient plasma energy.
It is especially important to understand the effect of
different values of the ``hot'' magnetization~$\sigma_h$ because it
exerts a substantial influence on reconnection: not only does it place an upper bound on the relative energy gain of particles, but it also determines the Alfv\'{e}n velocity, which controls the rate of reconnection.
This paper presents the first study that systemically investigates energy conversion and NTPA in reconnection across a wide variety of parameters in the ultrarelativistic $\sigma_h=1$ pair plasma regime.
Specifically, we explore the effects of different
initial current sheet configurations, system-sizes (in 2D), guide
magnetic fields, and (in 3D) aspect ratios $L_z/L_x$.
To compare simulations we look particularly at the rates at which
magnetic energy is converted to plasma energy, as well as the resulting
NTPA.
An extensive summary of briefly-stated important results can be found in~\S\ref{sec:summary}.
Our 2D simulations show that reconnection with $\sigma_h=1$ \emph{qualitatively} resembles relativistic reconnection in the $\sigma_h\gg 1$ regime;
however, they also measure significant quantitative differences---differences that could ultimately allow one to constrain~$\sigma_h$ in an astrophysical source based on observed radiation.
In contrast, our 3D simulations reveal substantially different qualitative behaviours, as well as (in many cases) significantly slower conversion of magnetic to plasma energy.
We find that in 3D there are are other mechanisms besides large-scale (2D-like) reconnection that can drive current sheet evolution and plasma energization.
It would therefore be more precise to say that we are studying the evolution of a thin current sheet in 3D rather than 3D reconnection.
Fascinatingly, regardless of the large-scale driver, NTPA remains robust, very similar in 2D and 3D---when compared at equivalent fractions of converted magnetic energy.
Results characterizing reconnection and energy conversion rates and NTPA in~2D for $\sigma_h=1$ are described in~\S\ref{sec:2d} and compared with the~$\sigma_h\gg 1$ regime in~\S\ref{sec:highSigmah}. By comparing to other works studying $\sigma_h\gg 1$, we see that $\sigma_h$ strongly influences these key outcomes. In 2D, for $\sigma_h=1$ and zero guide field, we find that the dimensionless reconnection rate (i.e., normalized to $B_0 v_A$) is around~0.03, which is somewhat less than the~$\sim 0.1$ observed for $\sigma_h \gg 1$.
Reconnection with $\sigma_h=1$ clearly yields NTPA, but generates a high-energy power-law energy spectrum $f(\gamma)\sim \gamma^{-4}$, significantly steeper than for~$\sigma_h\gg 1$.
Increasing the guide field $B_{gz}$ slows reconnection and suppresses NTPA, yielding still steeper power-law spectra.
We believe the guide field slows down reconnection by reducing the effective Alfv\'{e}n speed in the outflow direction, while the reduced compressibility of the guide-field-threaded plasma causes reconnection to end with more unreconnected flux remaining (in a closed system).
We expect these guide field effects to become significant for $B_{gz}/B_0 \gtrsim 1/\sigma_h$ (thus they should become evident only for much stronger guide field in nonrelativistic reconnection, where $\sigma_h\ll 1$ because $\sigma_h$ includes rest-mass enthalpy).
Moreover, we find that these outcomes in 2D are relatively insensitive to the details of the initial current sheet, indicating that the primary results of 2D reconnection (energy conversion and NTPA) may be robust functions of the ambient plasma, as hoped (to connect observations with astrophysical source conditions).
In 3D, the story is much different. Unlike relativistic reconnection with $\sigma_h\gg 1$, 3D effects substantially alter reconnection with $\sigma_h=1$.
In fact, true reconnection may have a much diminished role in the 3D evolution of thin current sheets, with other processes (likely driven by the RDKI) competing to convert magnetic energy to plasma energy.
As a result, the morphology of the current sheet does not feature the intricately hierarchical plasmoid structure familiar from 2D reconnection simulations.
In 3D, plasmoids can decay, and instead of storing magnetic energy as in 2D, they release their magnetic energy to the plasma.
Thus 3D ``reconnection'' can convert upstream magnetic energy more completely to plasma energy.
However, the rate of energy conversion may be significantly
(an order of magnitude) slower in 3D, even at early stages, and still
slower at later stages.
Reconnection simulations in both 2D and 3D evolve stochastically or chaotically, following different paths from the initially random particle distributions.
However, the competition among different magnetic-energy-releasing mechanisms leads to a variety of long-term behaviours in 3D, in contrast to the relative inevitability with which 2D reconnection evolves to a common final state
(despite unpredictable chaotic plasmoid behaviour along the way).
In~3D, current sheet evolution is significantly more sensitive to the initial current sheet configuration than in~2D. For example, an initial magnetic field perturbation (relatively inconsequential in 2D simulations) tends to force 3D evolution to resemble 2D-like reconnection.
The main competition to reconnection appears to come from the RDKI; importantly, we find that it might not be the fastest-growing RDKI mode that presents the stiffest competition, but rather the longest-wavelength, ``sufficiently fast-growing'' mode.
While the linear stage of the RDKI does not deplete much magnetic energy,
its nonlinear development---especially in large-amplitude modes---competes with reconnection to convert significant amounts of magnetic energy to plasma energy.
In particular, we have observed the RDKI-induced rippling of the current sheet grow to an amplitude comparable to its wavelength (without depleting much magnetic energy) and subsequently fold over on itself, rapidly converting most of the magnetic energy within the rippling amplitude to plasma energy.
Because longer-wavelength modes reach larger amplitudes, they can have a larger impact; whereas short-wavelength modes, though they may grow faster, saturate at small amplitudes without obstructing reconnection.
We find that a completely different mechanism can thus deplete magnetic energy as fast as reconnection. However, whereas reconnection results in a highly-structured chain of plasmoids that---despite monstrously large plasmoids---preserves thin (kinetic-scale) inter-plasmoid current sheets, this RDKI-triggered process results in a turbulent, generally-thickened current layer, which seems to disrupt reconnection.
Despite sometimes very large differences between 2D and 3D current sheet evolution---and even between different 3D evolutions---the resulting NTPA is remarkably similar in all cases (with the same guide magnetic field).
The presence of a guide magnetic field tends to suppress 3D effects, causing current sheet evolution to behave more like 2D reconnection.
This leads to an interesting non-monotonic trend in 3D with increasing guide field, since both 3D effects and guide field suppress reconnection.
A weak guide field tends to enhance the 3D reconnection rate by diminishing 3D effects;
but once the guide field is strong enough to suppress 3D effects nearly completely, further increase suppresses reconnection in 3D as in 2D.
Again, even with guide field, NTPA is very similar in 2D and 3D.
Increased guide field (unlike 3D effects) suppresses NTPA, leading to steeper power-law energy distributions.
This investigation shows that 3D current sheet evolution is not necessarily a perturbation or modification of 2D reconnection, but rather involves a complicated interaction of linear and nonlinear stages of multiple instabilities. This leads to a range of possible behaviours and an accompanying sensitivity to initial conditions that motivates further exploration of initial configurations and construction of parameter-space phase diagrams; these possibilities and sensitivities must then be considered in astrophysical modelling. It also increases the urgency of understanding the current sheet formation process in the first place. On the other hand, the diversity of behaviour highlights universalities in magnetic energy dissipation and resulting NTPA that could find important use in astrophysical models. While the rate of magnetic energy dissipation can vary significantly depending on the details of the current sheet, all configurations yield, though by different means, ``fast'' magnetic energy release in the sense that it is tremendously faster than naive magnetic diffusion; and the nonthermal particle spectra are even more universal, determined primarily by the ambient~$\sigma_h$ and~$B_{gz}/B_0$. These universalities will be invaluable for astrophysical modelling and may be important clues to a deeper understanding of magnetic energy dissipation and particle acceleration in plasmas.
\vspace{\baselineskip}
We would like to thank NSF, DOE, and NASA for supporting this work---in particular, grants NSF AST-1806084, NSF AST-1903335,
NASA ATP NNX16AB28G,
NASA ATP NNX17AK57G, NASA ATP 80NSSC20K0545.
In addition, this work would not have been possible without substantial supercomputing time.
A award of computer time was provided by the Innovative and Novel Computational Impact on Theory and Experiment (INCITE) program; for the 3D simulations, this research used resources of the Argonne Leadership Computing Facility, which is a DOE Office of Science User Facility supported under contract DE-AC02-06CH11357.
This work also used the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by National Science Foundation grant number ACI-1548562 \citep{XSEDE2014}; in particular, the large 2D simulations were run at the Texas Advanced Computing Center (TACC) at The University of Texas at Austin.
|
\section{Introduction}
\label{sec:intro}
Near term quantum devices have a small number of noisy qubits that can support execution of shallow depth circuits (i.e., those with few operational cycles) only. Variational Quantum Algorithms (VQA) aim to leverage the power as well as the limitations imposed by these devices to solve problems of interest such as combinatorial optimization \cite{farhi2014quantum,wang2018quantum,hadfield2019quantum,cook2020quantum}, quantum chemistry \cite{mcclean2016theory, grimsley2019adaptive}, and quantum machine learning \cite{10.1007/978-3-030-50433-5_45,biamonte2017quantum,torlai2020machine}. VQA divides the entire computation into functional modules, and outsources some of these modules to classical computers. The general framework of VQA can be divided into four steps: (i) encode the problem into a parameterized quantum state $\ket{\psi(\theta)}$ (called the ansatz), where $\theta = \{\theta_1,\theta_2, \hdots, \theta_k\}$ are $k$ parameters; (ii) prepare and measure the ansatz in a quantum computer, and determine the value of some objective function $C(\theta)$ (which depends on the problem at hand) from the measurement outcome; (iii) in a classical computer, optimize the set of parameters to find a better set $\theta' = \{\theta'_1,\theta'_2, \hdots, \theta'_k\}$ such that it minimizes (or maximizes) the objective function; (iv) repeat steps (ii) and (iii) with the new set of parameters until convergence.
Quantum Approximate Optimization Algorithm (QAOA) is a type of VQA that focuses on finding good approximate solutions to combinatorial optimization problems. It has been studied most widely for finding the maximum cut of a (weighted or unweighted) graph (called the Max-Cut problem) \cite{farhi2014quantum}. For this problem, given a graph $G = (V,E)$ where $V$ is the set of vertices and $E$ is the set of edges, the objective is to partition $V = V_1 \cup V_2$, such that $V_1 \cap V_2 = \phi$, and the number of edges crossing the partition is maximized. Throughout this paper, we shall consider {\it connected graphs} with $|V| = n$ and $|E| = m$, but the results can be easily extended to disconnected graphs as well.
In the initial algorithm proposed by Farhi \cite{farhi2014quantum} for the Max-Cut problem, a depth-$p$ QAOA consists of $p \geq 1$ layers of alternating operators on the initial state $\ket{\psi_0}$
\begin{equation}
\label{eq:ansatz}
\ket{\psi(\gamma,\beta)} = ( \displaystyle \Pi_{l = 1}^{p} e^{(-i\beta_l H_M)} e^{(-i\gamma_l H_P)}) \ket{\psi_0}
\end{equation}
where $H_P$ and $H_M$ are called the Problem and Mixer Hamiltonian respectively, and $\gamma = \{\gamma_1, \gamma_2, \hdots, \gamma_p\}$ and $\beta = \{\beta_1, \beta_2, \hdots, \beta_p\}$ are the parameters. It is to be noted that the depth $p$ of the QAOA is not related to the depth of the quantum circuit realizing the algorithm. The problem Hamiltonian describing the Max-Cut can be represented as in Eq.~(\ref{eq:max_cut}), where $w_{jk}$ is the weight associated with the edge $(j,k)$.
\begin{equation}
\label{eq:max_cut}
H_P = \frac{1}{2}\displaystyle \sum_{(j,k) \in E} w_{jk} (I - Z_j Z_k)
\end{equation}
Furthermore, the mixer Hamiltonian should be an operator that does not commute with the Problem Hamiltonian. In the traditional QAOA, the mixer Hamiltonian is $H_M = \displaystyle \sum_{i} X_i$.
Variations to this have been studied to improve the performance of the algorithm --- such as using other mixers \cite{bartschi2020grover,zhu2020adaptive, yu2021quantum}, training the parameters to reduce the classical optimization cost \cite{larkin2020evaluation}, and modifying the cost function for faster convergence \cite{barkoutsos2020improving}. In this paper we stick to the original problem and mixer hamiltonians proposed in the algorithm by Farhi et al. \cite{farhi2014quantum}. The applicability and effectiveness of our proposed method on the modifications of this algorithm can be looked at as a follow-up work. However, our proposed methods optimize the circuit corresponding to the problem hamiltonian. Since most of the modifications suggested in the literature aim to design more efficient mixers, our proposed optimization should be applicable on those as well.
The realization of the QAOA circuit for Max-cut requires two CNOT gates for each edge (details given in Sec.~\ref{sec:ansatz}). Hardware realization of a CNOT gate is, in general, significantly more erroneous than a single qubit gate. Even in the higher end devices of IBM Quantum, such as \textit{ibmq\_montreal}, \textit{ibmq\_manhattan}, the probability of error for a single qubit gate and a CNOT gate are $\mathcal{O}(10^{-4})$ and $\mathcal{O}(10^{-2})$, respectively~\cite{ibmquantum}. In other words, CNOT gates are $100$ times more likely to be erroneous than single qubit gates. Therefore, we focus primarily on reducing the number of CNOT gates in the design of QAOA ansatz for Max-cut.
\subsubsection*{Contributions of this paper}
In this paper, we
\begin{enumerate}[(i)]
\item propose two optimization methods for reducing the number of CNOT gates in the first layer of the QAOA ansatz based on (1) an Edge Coloring that can reduce upto $\lfloor \frac{n}{2} \rfloor$ CNOT gates, and (2) a Depth First Search (DFS) that can reduce $n-1$ CNOT gates.
\item prove that there exists no method that can reduce more than $n-1$ CNOT gates while still maintaining a fidelity of 1 with the original QAOA ansatz \cite{farhi2014quantum}.
\item show that while the Edge Coloring based optimization reduces the depth of the circuit, the DFS based method may increase the depth. We further analytically derive the criteria (involving the increase in the depth and the reduction in the number of CNOT gates) for which the DFS based optimization method still leads to a lower probability of error in the circuit, and show that the IBM Quantum Hardwares \cite{ibmquantum} conform to that criteria.
\item simulate our proposed optimization methods in Qiskit~\cite{Qiskit} with the \textit{ibmq\_manhattan} noise model and show that for graphs of different sparsity (Erdos-Renyi graphs with the probability of edge varying from 0.4 - 1)
\begin{enumerate}
\item the proposed reduction in the CNOT gate is still retained post transpilation
\item the DFS based method has lower error probability than the Edge Coloring method, which in its turn has lower error probability than the traditional QAOA ansatz.
\end{enumerate}
\end{enumerate}
Therefore, for any graph $G = (V,E)$, our proposed method provides reduction in the number of CNOT gates, and hence lowers the error probability of the circuit. Although the DFS method provably surpasses the Edge Coloring method, both in terms of reduction in CNOT gates and lowering the error probability, the latter reduces the depth of the QAOA circuit, and is also used in the DFS based method to arrange the edges which do not form a part of the DFS tree.
For the rest of this paper, we consider \emph{unweighted and connected graphs}, $i.e.$, $w_{jk} = 1$, $\forall$ $(j,k) \in E$. However, the circuit corresponding to the ansatz does not change if we have a weighted graph \cite{hadfield2019quantum}. Therefore, every analysis in this paper holds for a weighted graph as well. Furthermore, the analysis of this paper will hold as it is, or with some minimal modification, for disconnected graphs as well.
The rest of the paper is organized as follows - Section~\ref{sec:ansatz} briefly discusses the traditional QAOA ansatz design. In Section~\ref{sec:thm} we provide the proposed optimization and the criteria for it. Section~\ref{sec:edge_col} and ~\ref{sec:dfs} describe two methods of optimization based on Edge Coloring and DFS respectively. We provide the respective algorithms and analyze the conditions under which each one reduces the probability of error. We present the results of our simulation in section~\ref{sec:sim} and conclude in Section~\ref{sec:con}.
\section{Traditional ansatz design for QAOA}
\label{sec:ansatz}
The objective function of a depth-$p$ QAOA for Max-Cut~\cite{farhi2014quantum} can be expressed as
\begin{equation}
\label{eq:objective}
\max_{\psi(\gamma,\beta)} \bra{\psi(\gamma,\beta)} H_P \ket{\psi(\gamma,\beta)}
\end{equation}
where $\gamma = \{\gamma_1, \gamma_2, \hdots, \gamma_p\}$ and $\beta = \{\beta_1, \beta_2, \hdots, \beta_p\}$ are the parameters. The trial wavefunction $\ket{\psi(\gamma,\beta)}$ is called the ansatz. The QAOA ansatz has a fixed form as described in Eq.~(\ref{eq:ansatz}). The initial state $\ket{\psi_0}$ is usually the equal superposition of $n$ qubits, where $n = |V|$. Note that the depth of the circuit required to prepare $\ket{\psi_0}$ is 1 (Hadamard gates acting simultaneously on all the qubits). Similarly, for each layer of QAOA, the operator $exp(-i \beta_l H_M)$ can be realized by a depth one circuit of $R_x(\beta_l)$ gates acting simultaneously on all the qubits.
The operator $exp(-i \gamma_l H_P)$ has a more costly implementation. Note that
\begin{eqnarray*}
exp(-i \gamma_l H_P) = \displaystyle \Pi_{(i,j) \in E} exp\left(-i \gamma_l \left(\frac{I-Z_j Z_k}{2}\right)\right).
\end{eqnarray*}
The operator $exp\left(-i \gamma_l \left(\frac{I-Z_j Z_k}{2}\right)\right)$ acts on each edge $(j,k)$, and is realized as shown below:
\begin{figure}[H]
\centering
\begin{quantikz}
{q_{j}}&&\ctrl{1} & \qw & \ctrl{1} & \qw \\
{q_{k}}&&\targ{} & \gate{R_z(2\gamma_l)} & \targ{} & \qw
\end{quantikz}
\label{fig:z_jz_k}
\end{figure}
Here, $q_j$ and $q_k$ represent qubit indices $j$ and $k$, respectively. Note that Max-Cut is a symmetric problem, and therefore, the selection of control and target from qubits $q_j$ and $q_k$ for the CNOT gate corresponding to the edge $(j,k)$ is irrelevant, $i.e.$ the operator $exp\left(-i \gamma_l \left(\frac{I-Z_j Z_k}{2}\right)\right)$ can be equivalently realized as $CNOT_{kj} (I_k \otimes R_z(2\gamma_l)_{j})$ $CNOT_{kj}$. In Fig.~\ref{fig:depth_qaoa}(a) and (b), we show a 2-regular graph with four vertices and its corresponding QAOA circuit for $p = 1$ respectively.
\begin{figure}[htb]
\centering
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[scale=0.35]{graph.png}
\caption{A 2-regular graph with four vertices}
\label{col}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.4\textwidth}
\centering
\includegraphics[scale=0.4]{qaoa_circ.png}
\caption{Max-Cut QAOA circuit for $p=1$ corresponding to the graph}
\label{dfs}
\end{subfigure}
\caption{The Max-Cut QAOA circuit for $p=1$ corresponding to the 2-regular graph with four vertices; the values of $\gamma$ and $\beta$ can be arbitrary but those in this figure are the optimum values for this graph when $p = 1$}
\label{fig:depth_qaoa}
\end{figure}
\section{Methods for Optimized ansatz design}
\label{sec:thm}
Some recent studies have proposed optimization methods for the circuit of the QAOA ansatz with respect to the underlying hardware architecture \cite{alam2020circuit}. In this paper we propose two \textit{hardware independent} methods to reduce the number of CNOT gates in the traditional QAOA ansatz. The intuition is that in the circuit realization of the operator $exp\left(-i \gamma_l \left(\frac{I-Z_j Z_k}{2}\right)\right)$ as $CNOT_{jk} (I_j \otimes R_z(2\gamma_l)_{k})$ $CNOT_{jk}$, the first CNOT gate can be removed whenever it does not make any contribution to the overall effect of the operator. Our proposed method reduces the number of CNOT gates in the circuit irrespective of the hardware architecture, and hence is applicable for any quantum device.
In Theorem~\ref{thm:equiv} we prescribe the condition where the first CNOT gate is irrelevant to the effect of the said operator, and hence may be removed.
\begin{theorem}
\label{thm:equiv}
Let $\ket{\psi}$ be an $n$-qubit state prepared in a uniform superposition (upto relative phase) over all basis states $\ket{x_1, \hdots, x_n}$ such that the relative phase on each basis state is a function of a subset $S \subset$ $\{1,2,...,n\}$ of the $n$ qubits (and independent of remaining qubits) $i.e.$
\begin{center}
$\ket{\psi} = \frac{1}{\sqrt{2^n}}\displaystyle \sum_{x_1,...,x_n} e^{(i \phi(x_S))}\ket{x_1,...,x_n}$
\end{center}
where $x_S = \{x_i : i \in S\}$ and $\phi(x_S)$ depicts the relative phase of each superposition state. For any two qubits $\ket{j}$ and $\ket{k}$, where $\ket{k} \notin S$, and for the two operators $U_1 = CNOT_{jk} (I_j \otimes R_z(2\gamma_l)_{k}) CNOT_{jk}$ and $U_2 = (I_j \otimes R_z(2\gamma_l)_{k}) CNOT_{jk}$, we have
\begin{center}
$U_1\ket{\psi} = U_2\ket{\psi}$.
\end{center}
\end{theorem}
\begin{proof}
Let us consider the action of the operators $U_1$ and $U_2$ on any edge $(j,k)$.
\begin{equation}
U_1\ket{\psi} = CNOT_{jk} (I_j \otimes R_z(2\gamma_l)_{k}) (CNOT_{jk}) \ket{\psi} \nonumber
\end{equation}
\begin{align}
=&\displaystyle \sum_{x_1,...,x_n} CNOT_{jk} (I_j \otimes R_z(2\gamma_l)_{k}) (CNOT_{jk})\nonumber \\
& e^{i\phi(x_S)} \ket{x_1,...,x_n} \\
=&\displaystyle \sum_{x_1,...,x_n} CNOT_{jk} (I_j \otimes R_z(2\gamma_l)_{k}) \nonumber \\
& e^{i \phi(x_S)} \ket{x_1,..,x_k'=x_j \oplus x_k,.,x_n} \\
=&\displaystyle \sum_{x_1,...,x_n} e^{i(\phi(x_S)- \gamma_l (x_j \oplus x_k))} CNOT_{jk} \nonumber \\
& \ket{x_1,..,x_k'=x_j \oplus x_k,.,x_n} \\
\label{eq:u1}
=& \displaystyle \sum_{x_1,...,x_n} e^{i(\phi(x_S) - \gamma_l (x_j \oplus x_k))}\ket{x_1,...,x_n}
\end{align}
where $e^{i \phi(x_S)}$ is the cumulative effect of operators acting on previous edges (= 0 if $(j,k)$ is the first). We have dropped the normalization constant for brevity.
Similarly,
\begin{equation}
U_2\ket{\psi} = CNOT_{jk} (I_j \otimes R_z(2\gamma_l)_{x_k}) \ket{\psi} \nonumber
\end{equation}
\begin{equation}
= CNOT_{jk} \displaystyle \sum_{x_1,...,x_n} e^{i((\phi(x_S)) - \gamma_l x_k)}\ket{x_1,...,x_n} \nonumber
\end{equation}
\begin{equation}
\label{eq:u2_mid}
= \displaystyle \sum_{x_1,...,x_n} e^{i((\phi(x_S)) - \gamma_l x_k)} \ket{x_1,..,x_j \oplus x_k,..,x_n}
\end{equation}
where the qubit in $k^{\text{th}}$ position changes to $x_j \oplus x_k$ due to the $CNOT_{jk}$ operation. Now, substituting $x_k' = x_j \oplus x_k$ in the above equation, we get
\begin{equation}
U_2\ket{\psi} = \displaystyle \sum_{x_1,...,x_n} e^{i((\phi(x_S)) - \gamma_l x_k)} \ket{x_1,..,x_j \oplus x_k,..,x_n} \nonumber
\end{equation}
\begin{equation}
= \displaystyle \sum_{x_1,..,x_k',..,x_n} e^{i((\phi(x_S)) - \gamma_l (x_j \oplus x_k'))} \ket{x_1,..,x_k',..,x_n} \nonumber
\end{equation}
\begin{equation}
\label{eq:u2}
= \displaystyle \sum_{x_1,..,x_k,..,x_n} e^{i((\phi(x_S)) - \gamma_l (x_j \oplus x_k))} \ket{x_1,..,x_k,..,x_n}
\end{equation}
Here since $k \notin S$, the substitution in second last step, does not change the phase $e^{i \phi(x_S)}$. The last step is valid since $x_k'$ is a running index and hence can be changed to $x_k$. Thus Eq.~(\ref{eq:u1}) and Eq.~(\ref{eq:u2}) are identical.
\end{proof}
\begin{corollary}
\label{cor:cond}
For a graph $G$, we can optimize the circuit for the operator $exp\left(-i \gamma_l \left(\frac{I-Z_j Z_k}{2}\right)\right)$ corresponding to an edge $(j,k)$ replacing $U_1$ by $U_2$, provided that the target vertex does not occur in any of the edge operators run earlier. In other words, the following conditions are sufficient to optimize an edge:-
\begin{enumerate}
\item if the vertex $j$ is being operated on for the first time, then it acts either as a control or a target for the CNOT gate corresponding to the operator;
\item the vertex $j$ does not act as a target of the CNOT gate if it occurs as a part of any other edge operators run earlier.
\end{enumerate}
\end{corollary}
\begin{proof}
The first time we consider an edge adjacent to a vertex $j$, where $j \notin x_S$, (see Theorem~\ref{thm:equiv}) the relative phase $\phi(x_S)$ does not depend on $j$. Thus it satisfies the condition of Theorem~\ref{thm:equiv} and allows optimization of the operator.
On the other hand, if the vertex $j$ occurs as part of an edge operator already run, the phase on the basis state $\phi$ can potentially depend on $S$, $i.e.$ $j \in S$. By not allowing it to act as target, we satisfy the conditions of Theorem~\ref{thm:equiv}.
\end{proof}
From the above discussion, it follows that if we arbitrarily choose edges for applying the operator $exp\left(-i \gamma_l \left(\frac{I-Z_j Z_k}{2}\right)\right)$, then it cannot be guaranteed that a large number of edges will conform to Corollary~\ref{cor:cond}. The requirement, in fact, imposes a precedence ordering among the edges. In Section~\ref{sec:edge_col} and ~\ref{sec:dfs}, we provide two algorithmic procedures for maximizing the number of edges that satisfy the requirement in order to reduce the number of CNOT gates in the ansatz.
For the rest of the paper, we say that {\it an edge is optimized} if the operator $U_2$ can be operated on that edge instead of $U_1$.
\section{Edge Coloring based Ansatz Optimization }
\label{sec:edge_col}
The total error one incurs in a circuit depends on the number of operators (since a larger number of operators tend to incur more error) and the depth of the circuit (corresponding to relaxation error). In this section, we discuss how one can minimize the depth of the circuit. We also discuss the possibility of reduction in CNOT gates in the depth optimized circuit.
The operators $H_M$ act on distinct qubits and hence can be run in parallel contributing to a depth of 1 (for each step of the QAOA). On the other hand, the operators in $H_P$ can potentially contribute a lot to depth since the edge operators do not act on disjoint vertices.
At a given level of the circuit, we can only apply edge operators corresponding to a vertex disjoint set of edges. Thus the minimum depth of the circuit will correspond to the minimum value $k$ such that we can partition the set of edges $E$ as a disjoint union $\cup_i E_i$ where each subset $E_i$ consists of a vertex disjoint collection of edges. This in turn corresponds to the edge coloring problem in a graph.
Given a graph $G = (V,E)$ and a set of colors $\chi' = \{\chi'_1, \chi'_2, \hdots, \chi'_k\}$, an edge coloring \cite{west2001introduction} assigns a color to each edge $e \in E$, such that any two adjacent edges ($i.e.$, edges incident on a common vertex) must be assigned distinct colors. The edge coloring problem comprises of coloring the edges using the minimum number of colors $k$. Note that the operators corresponding to edges having the same color can therefore be executed in parallel.
Moreover,
\begin{enumerate}
\item the number of colors in optimal coloring, called the chromatic index, corresponds to the minimum depth of the circuit;
\item edges having the same color corresponds to the operators $exp\left(-i \gamma_l \left(\frac{I-Z_j Z_k}{2}\right)\right)$ that can be executed simultaneously.
\end{enumerate}
Optimal edge coloring is an NP-complete problem \cite{west2001introduction}. But it is not practical to allocate exponential time to find the optimal edge-coloring as a pre-processing step for QAOA. Vizing's Theorem states that every simple undirected graph can be edge-colored using at most $\Delta + 1$ colors, where $\Delta$ is the maximum degree of the graph \cite{vizing1964estimate}. This is within an additive factor of 1 since any edge-coloring must use at least $\Delta$ colors. Misra and Gries algorithm \cite{misra1992constructive} achieves the above bound constructively in $\mathcal{O}(n\cdot m)$ time. Therefore, we use the Misra and Gries edge coloring algorithm. Algorithm~\ref{alg:edgecol} below computes the sets of edges having the same color using Misra and Gries algorithm as a subroutine. It returns the largest set $S_{max}$ of edges having the same color in the coloring computed by Misra and Gries algorithm.
\begin{algorithm}[H]
\caption{Edge Coloring based Ansatz Optimization}
\label{alg:edgecol}
\begin{algorithmic}[1]
\REQUIRE A graph $G = (V,E)$.
\ENSURE Largest set $S_{max}$ of edges having the same color.
\STATE Use the Misra and Gries algorithm to color the edges of the graph $G$.
\STATE $S_i \leftarrow$ set of edges having the same color $i$, $1 \leq i \leq \chi'$.
\STATE $S_{max} \leftarrow$ $max\{S_1, S_2, \hdots, S_{\chi'}\}$.
\STATE Return $S_{max}$.
\end{algorithmic}
\end{algorithm}
This edge coloring approach provides the minimum depth achievable for QAOA ansatz using a polynomial time pre-processing. After reducing the depth, we now try to further reduce errors by decreasing the number of CNOT gates.
Recall that the operators corresponding to edges with the same color can be executed in parallel. We use the operators corresponding to the edges of $S_{max}$ as the first layer of operators. The other layers can be used in any order.
\begin{lemma}
Every edge in the first layer can be optimized according to Corollary~\ref{cor:cond}.
\end{lemma}
\begin{proof}
For every edge $(u,v)$ in the first layer, both the vertices are adjacent to an edge for the first time, $i.e.$, both $u, v \notin S$. Therefore, it satisfies the criteria of Corollary~\ref{cor:cond}, and hence can be optimized. In fact, any one of the qubits corresponding to the two vertices can be selected as the control for the CNOT operation.
\end{proof}
Some edges in the corresponding layers may be optimized as well. Nevertheless, it is trivial to come up with examples where this is not the case (e.g., a complete graph of 4-vertices). Therefore, in the worst case scenario, only the edges in the first layer can be optimized. However, since this method does not increase the depth of the circuit, it always leads to a more efficient circuit design than the traditional QAOA circuit with lower depth (by 1 since the first layer of CNOT is absent) and fewer CNOT gates.
For general graphs, the worst case scenario is, therefore, that only the edges in the first layer can be optimized. In the following subsection we provide an analysis on the number of optimized edges using this method.
\subsection{Lower and upper bound on the number of optimized edges}
Let us assume that the chromatic index of a graph $G = (V,E)$ is $\chi'$. Using the Misra and Gries Theorem \cite{misra1992constructive} we can find a polynomial time coloring using at most $\Delta + 1$ colors, where $\Delta$ is the maximum degree of the graph. Therefore, on an average, $\lceil \frac{m}{\Delta + 1} \rceil$ edges have the same color.
More precisely, two extreme cases arise: (i) the colors may be uniformly distributed, and the maximum number of edges having the same color is $\lceil \frac{m}{\Delta + 1} \rceil$; or (ii) one of the colors is used dominantly for most of the edges. Nevertheless, note that for all the edges adjacent to the same vertex, a particular color can be assigned to one of the edges only. Therefore, the dominant color can be used at most on $\lfloor \frac{n}{2} \rfloor$ edges, where $n = |V|$. Therefore, the possible number of optimized edges that can be obtained via the Edge Coloring method is as shown in Eq.~(\ref{eq:edge_col}).
\begin{equation}
\label{eq:edge_col}
\lceil \frac{m}{\Delta + 1} \rceil \leq ~\# ~Optimized ~Edges \leq \lfloor \frac{n}{2} \rfloor.
\end{equation}
\section{Depth First Search based Ansatz Optimization}
\label{sec:dfs}
As the edge coloring based algorithm can optimize at most $\lfloor \frac{n}{2} \rfloor$ edges, in this section, we present a Depth First Search (DFS) based optimization procedure which can optimize $n-1$ edges. Algorithm~\ref{alg:dfs}, for obtaining the optimized QAOA ansatz, uses the standard DFS algorithm \cite{cormen2009introduction}, by returning the tree edges or discovery edges forming the DFS tree.
In this method, we start from the first vertex of the DFS tree. For every edge $e = (u,v)$ in the DFS tree, the vertex $u$ is made the control and $v$ is made the target for the CNOT gate corresponding to that edge. The edges are operated on sequentially one after another, as in the set $E_{dfs}$ (the tree edges). Once every edge in the DFS tree has been operated on, the remaining edges can be executed in any order. In fact, one may opt to use the Edge Coloring method on the remaining edges to obtain the minimum depth of the circuit corresponding to these edges, although CNOT gates cannot be reduced any further.
\begin{algorithm}[H]
\caption{DFS based Ansatz Optimization }
\label{alg:dfs}
\begin{algorithmic}[1]
\REQUIRE A graph $G = (V,E)$.
\ENSURE A list $E_{dfs}$ of $n-1$ edges.
\STATE $E_{dfs} = \{\}$
\STATE $u \leftarrow$ randomly selected vertex from $V$.
\STATE Start DFS from the vertex $u$. For every vertex $v$ discovered from its predecessor $v'$, $E_{dfs} = E_{dfs} \cup (v',v)$.
\STATE Return $E_{dfs}$.
\end{algorithmic}
\end{algorithm}
\begin{theorem}
Each edge in the DFS tree can be optimized according to Corollary~\ref{cor:cond}.
\label{thm:dfs}
\end{theorem}
\begin{proof}
We prove this by the method of induction. Let $u$ be the vertex from which the DFS tree starts. Then $u$ is being operated on for the first time, and, hence, can act both as a control/target for the CNOT operation corresponding to the first edge (Corollary~\ref{cor:cond}). Choose $u$ to be the control.
\textbf{Base case}: If $v$ is the vertex that is discovered from $u$ via the edge $(u,v)$, then choosing $u$ as the control and $v$ as the target satisfies Corollary~\ref{cor:cond}. Therefore, the edge $(u,v)$ can be optimized.
\textbf{Induction hypothesis}: Let the DFS tree has been constructed upto some vertex $j$, and every edge $(e_1, e_2)$ in this DFS tree so far can be optimized, $i.e.$ $e_1$ acts as the control and $e_2$ as the target.
\textbf{Induction step}: Let the next vertex in the DFS tree, that is discovered from some vertex $i$, is $k$. From DFS algorithm, the vertex $i$ must have been discovered in some previous step. Since vertex $k$ was not previously discovered, so $k \notin x_S$ and hence the edge $(i,k)$ can be optimized if we select $i$ to be the control and $k$ as the target.
\end{proof}
Therefore, the DFS based optimization method provides $n-1$ optimized edges, $i.e.$, a reduction in the number of CNOT gates by $n-1$. We now show in Theorem~\ref{thm:optimal} that this is the maximum number of edges that can be optimized.
\begin{figure}[H]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[scale=0.4]{col.png}
\caption{Edge Coloring Based Optimization}
\label{col}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[scale=0.4]{dfs.png}
\caption{Depth First Search Based Optimization}
\label{dfs}
\end{subfigure}
\caption{Depth of the ansatz circuit when using (a) Edge Coloring and (b) DFS based method; edges having same color can be executed simultaneously. The depth of the spanning tree in the DFS based method is 4, compared to depth 2 for the Edge Coloring based method. However, the number of optimized edges in the Edge Coloring based method is 2, while that by the DFS based method is 3.}
\label{fig:depth}
\end{figure}
\begin{theorem}
Optimization of ansatz for Max-Cut QAOA with p=1, by deletion of the CNOT gate in the first level for an edge of the graph, can be done for no more than $n-1$ edges.
\label{thm:optimal}
\end{theorem}
\begin{proof}
Let us assume that there is some method by which at least $n$ edges can be optimized. Now, the connected subgraph which contains all the $n$ vertices and at least $n$ optimized edges must contain a cycle. Let $(u,v)$ be an edge of this cycle, $i.e.$, if $(u,v)$ is removed then the residual graph is a tree (in case there are $> n$ edges, the removal of edges can be performed recursively till such an edge $(u,v)$ is obtained whose removal makes the residual graph a tree). For this edge $(u,v)$, both the vertices $u$ and $v$ are endpoints of some other optimized edges as well. Therefore, from Corollary~\ref{cor:cond} both $u$ and $v$ must act as the control for the CNOT gate corresponding to the edge $(u,v)$ in order for this edge to be optimized, which is not possible. Therefore, it is not possible to optimize more than $n-1$ edges.
\end{proof}
Therefore, the DFS method is optimal in the number of optimized edges. However, we note that the DFS based method associates an ordering of the edges, $i.e.$, some of the edges which could have been operated on simultaneously, cannot be done so now.
This, in turn, can lead to an increase in the depth of the circuit. Hence, a penalty for this method producing optimal reduction in CNOT gates, is that it increases the depth of the circuit.
\begin{figure}[htb]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[scale=0.37]{qaoa_col.png}
\caption{Optimized circuit by Edge Coloring based method }
\label{col}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[scale=0.37]{qaoa_dfs.png}
\caption{Optimized circuit by DFS based method}
\label{dfs}
\end{subfigure}
\caption{Max-Cut QAOA ansatz with $p=1$ corresponding to (a) Edge Coloring and (b) DFS based optimization. In (a), the first CNOT gates of the operators have been deleted. The operators corresponding to $(q_1,q_2)$ and $(q_3,q_0)$ act in parallel. In (b), the first CNOT gates of three operators have been deleted, but the depth has increased.}
\label{fig:opt}
\end{figure}
In Fig.~\ref{fig:depth}, we show a 2-regular graph with four vertices. In Fig.~\ref{fig:depth}(a), the depth of the circuit corresponding to the operator $exp(-i\gamma_l H_P)$ is 2; the edges of the same color can be operated on simultaneously. If the red (or blue) edges form the first layer, then those two edges are optimized. However, if we use the DFS method, with the DFS tree starting from, say, vertex 1, then the edges $(1,2),(2,3)$ and $(3,4)$ can be optimized (Fig.~\ref{fig:depth}(b)). Now these three edges must be operated on one after another, followed by the fourth edge. Thus the depth of the circuit corresponding to the operator $exp(-i\gamma_l H_P)$ becomes 4. The circuits corresponding to these two scenarios are depicted in Fig~\ref{fig:opt}(a) and (b) respectively.
The question, therefore, is whether this increase in depth is always acceptable, even with the increased reduction in the number of CNOT gates as, with increased depth, the circuit becomes more prone to relaxation error.
Numerical analysis and simulation (Section~\ref{sec:sim}) establises that although the depth of the circuit is increased, the overall error probability of the circuit is reduced further.
\subsection{When is the DFS based method useful?}
In this subsection, we formulate a relation for which the increase in the depth still leads to a lower probability of error for the reduction in the number of CNOT gates. For this analysis, we make an assumption that the error in the circuit arises only from noisy CNOT gates and the depth of the circuit ($i.e.$, the $T_1$ time). Although this assumption is idealistic, the ansatz primarily consists of layers of CNOT gates. Furthermore, in superconductors, $R_z$ gates are executed virtually \cite{mckay2017efficient}, and hence does not lead to any gate error. Therefore, CNOT is the primary source of gate error and with increasing depth, the qubits become more prone to relaxation error. Therefore, this assumption allows for a simple but powerful model for analyzing the query at hand.
Let us assume that the time duration and the error probability of each CNOT gate is $t_{cx}$ and $p_{cx}$ respectively. Let there be $N$ layers of CNOT operations. Note that although there can be multiple CNOT gates in each layer, the time duration of each layer is $t_{cx}$ only. Therefore, the probability of no error ($i.e.$, the probability that the circuit remains error free) after $N$ layers of operations, considering only relaxation error, is
$exp(-\frac{N t_{cx}}{T_1})$.
Let there be $k$ CNOT gates in the original circuit. Therefore, the probability of no error after the operation of the CNOT gates, considering only CNOT gate error, is
$(1 - p_{cx})^k$.
Combining both the sources of the errors, Eq.~(\ref{eq:no_err}) gives the probability of success ($i.e.$, the probability of no error) after a single cycle of computation of the QAOA ansatz.
\begin{equation}
\label{eq:no_err}
P_{success} = (1 - p_{cx})^k \cdot exp(-\frac{N t_{cx}}{T_1})
\end{equation}
Henceforth, $P_{success}$ will refer to the probability of success ($i.e.$, how close the noisy outcome is to the noise-free ideal outcome) of the ansatz circuit execution for a single run of the algorithm. Note that in QAOA, the ansatz is computed multiple times for multiple cycles, and the objective is to maximize the expectation value of the outcome.
We further assume that after the optimization using DFS based method, $k_1$ CNOT gates have been reduced leading to an increase in $N_1$ layers of operations. The probability that this optimized circuit remains error-free is given in Eq.~(\ref{eq:opt_err}).
\begin{equation}
\label{eq:opt_err}
P^{opt}_{success} = (1 - p_{cx})^{(k-k_1)} \cdot exp(-\frac{(N+N_1) t_{cx}}{T_1})
\end{equation}
The optimization is fruitful only when $P^{opt}_{success} \geq P_{success}$. Note that
\begin{center}
$P^{opt}_{success} = P_{success} \cdot exp(-\frac{N_1 t_{cx}}{T_1}) / (1 - p_{cx})^{k_1}$
\end{center}
Since both $P^{opt}_{success}$ and $P_{success} \leq 1$, the required inequality holds only if $exp(- \frac{N_1 t_{cx}}{T_1}) / (1 - p_{cx})^{k_1} \geq 1$. In other words,
\begin{eqnarray}
\label{eq:cond1}
exp(-\frac{N_1 t_{cx}}{T_1}) &\geq& (1 - p_{cx})^{k_1} \nonumber\\
\Rightarrow N_1 &\leq& \lambda \times k_1, \nonumber\\
~where && \lambda = \frac{-ln(1 - p_{cx}) \cdot T_1}{t_{cx}}.
\end{eqnarray}
The constant $\lambda$ is defined in terms of parameters specific to the quantum device.
\subsubsection{Effect of varying $\lambda$}
Given that $\lambda = f(t_{cx},p_{cx},T_1)$, we expect the $T_1$ value to increase, and the $t_{cx}$ and $p_{cx}$ values to decrease as technology improves. The value of $\lambda$ increases for increasing $T_1$ and/or decreasing $t_{cx}$, whereas it decreases for decreasing $p_{cx}$. Therefore,
\begin{itemize}
\item If $p_{cx}$, the probability of error for CNOT gates decreases, the optimization becomes less useful since we are increasing the probability of relaxation error, but the reduction in error probability becomes less. As per this observation, for smaller $\lambda$, Eq.~(\ref{eq:cond1}) is satisfied when the increase in depth is reduced as well.
\item Similarly, if (i) $T_1$ increases, then the qubit can retain its conherence for a longer period of time, or (ii) $t_{cx}$ decreases, then the overall computation time of the circuit decreases as well, and the circuit can allow some relaxation in the depth even if $T_1$ remains unchanged. We observe that for both of these cases, by Eq.~(\ref{eq:cond1}), $\lambda$ increases, thus allowing more increase in depth for a given reduction in the number of CNOT gates.
\end{itemize}
\subsection{Trade-off between depth and reduction in CNOT gates}
If the DFS based method is not applied, then the number of layers of CNOT gates is equal to the number of color classes (as in Edge Coloring method). The maximum number of color classes is $\Delta + 1$ (as discussed in the previous section), and hence the maximum depth of the circuit is $\Delta + 1$ as well. Now, when the DFS based method is applied, the circuit can be divided into two disjoint sets of edges:
\begin{enumerate}
\item The set of edges belonging to the DFS tree which can be optimized. The depth of this portion of the circuit is at most $n-1$ ($i.e.$, the depth of the DFS tree). Each of the operators corresponding to these edges contains a single CNOT gate only, and hence the number of CNOT gate layers is $n-1$ as well.
\item The set of edges that do not belong to the DFS tree and hence are not optimized. The operators corresponding to these edges can be applied in any order, but after all the optimized edges. When removing the edges of the DFS tree, the degree of each vertex is reduced by at least 1. Therefore, the maximum degree of the remaining subgraph is at most $\Delta - 1$. Therefore, the depth of this portion of the circuit will be at most $\Delta$ (From Misra and Gries Algorithm). Each of the layer in this portion contains 2 CNOT gates, and hence the number of CNOT gate layers is $2\Delta$.
\end{enumerate}
Therefore, the maximum depth of the circuit after applying the DFS based optimization is $n-1 + \Delta$. In other words, the increase in depth due to this method is given by Eq.~(\ref{eq:cond2}).
\begin{eqnarray}
\label{eq:cond2}
n-1 + \Delta - (\Delta + 1) = n - 2
\end{eqnarray}
Recall that the number of CNOT gates reduced due to the DFS method is always $n-1$. Therefore, from Eq.~(\ref{eq:cond1}) and $~(\ref{eq:cond2})$, we get
\begin{eqnarray}
\label{eq:condition}
n - 2 &\leq& \lambda \cdot (n-1) \nonumber \\
\Rightarrow \lambda &\geq& \frac{n-2}{n-1}
\end{eqnarray}
In Table~\ref{tab:lambda} we show the average value of $\lambda$ for some IBM Quantum~\cite{ibmquantum} devices, ranging from the comparatively more noisy \textit{ibmq\_melbourne} to the comparatively less noisy \textit{ibmq\_manhattan}.
\begin{table}[htb]
\centering
\caption{Average value of $\lambda$ for four IBM Quantum machines ~\cite{ibmquantum}}
\begin{tabular}{|c|c|}
\hline
IBM Quantum devices & Avg value of $\lambda$\\
\hline
\textit{ibmq\_manhattan} & 3.6\\
\hline
\textit{ibmq\_montreal} & 2.47\\
\hline
\textit{ibmq\_sydney} & 3.35\\
\hline
\textit{ibmq\_melbourne} & 2.03\\
\hline
\end{tabular}
\label{tab:lambda}
\end{table}
Note that the lower bound on $\lambda$, $\frac{n-2}{n-1}$ (Eq.~(\ref{eq:condition}), is always less than $1$ for all $n$. In the asymptotic limit, $\frac{n-2}{n-1} \rightarrow 1$. Thus, the proposed DFS based optimization method leads to a lower error probability on any quantum device for which $\lambda \geq 1$. Table~\ref{tab:lambda} readily shows that the IBM Quantum hardwares conform to this requirement.
\section{Results of simulation}
\label{sec:sim}
\begin{table*}[t]
\centering
\caption{Comparison of Max-Cut QAOA ansatz circuits post transpilation on \textit{ibmq\_manhattan}: (i) Traditional, (ii) Edge coloring and (iii) DFS based optimization}
\begin{tabular}{|c|c|c|c|c|}
\hline
\multirow{2}{*}{Graph Family} & \multirow{2}{*}{\# qubits} & \multicolumn{3}{c|}{\# CNOT gates in Max-Cut QAOA ansatz circuit}\\
\cline{3-5}
& & Traditional & Edge coloring & DFS\\
\hline
\multirow{6}{*}{Complete graph} & 10 & 90 & 85 & 81\\
\cline{2-5}
& 20 & 380 & 370 & 361\\
\cline{2-5}
& 30 & 870 & 855 & 841\\
\cline{2-5}
& 40 & 1560 & 1540 & 1521\\
\cline{2-5}
& 50 & 2450 & 2425 & 2401\\
\cline{2-5}
& 60 & 3540 & 3510 & 3481\\
\hline
\multirow{6}{*}{Erdos-Renyi ($p_{edge}$ = 0.8)} & 10 & 70 & 66 & 61\\
\cline{2-5}
& 20 & 302 & 292 & 283\\
\cline{2-5}
& 30 & 698 & 683 & 669\\
\cline{2-5}
& 40 & 1216 & 1197 & 1177\\
\cline{2-5}
& 50 & 1956 & 1931 & 1907\\
\cline{2-5}
& 60 & 2822 & 2792 & 2763 \\
\hline
\multirow{6}{*}{Erdos-Renyi ($p_{edge}$ = 0.6)} & 10 & 50 & 46 & 41\\
\cline{2-5}
& 20 & 234 & 225 & 215\\
\cline{2-5}
& 30 & 504 & 491 & 475\\
\cline{2-5}
& 40 & 960 & 940 & 921\\
\cline{2-5}
& 50 & 1504 & 1479 & 1455\\
\cline{2-5}
& 60 & 2114 & 2085 & 2055 \\
\hline
\multirow{6}{*}{Erdos-Renyi ($p_{edge}$ = 0.4)} & 10 & 36 & 31 & 27\\
\cline{2-5}
& 20 & 164 & 154 & 145\\
\cline{2-5}
& 30 & 362 & 348 & 333\\
\cline{2-5}
& 40 & 586 & 566 & 547\\
\cline{2-5}
& 50 & 950 & 925 & 901\\
\cline{2-5}
& 60 & 1468 & 1440 & 1409 \\
\hline
\end{tabular}
\label{tab:cx_count}
\end{table*}
In this section we show the effect of our optimization methods on reducing the probability of error and the CNOT count of QAOA for Max-Cut. We first show that our proposed reduction is retained in the post transpilation circuit, which is executed on the quantum hardware. Next, we run our simulation with the noise model for \textit{ibmq\_manhattan} from IBM Quantum; this noise model corresponds to the actual noise in the IBM Quantum Manhattan device which has $65$ qubits and a Quantum Volume of $32$~\cite{ibmquantum}. For our simulation purpose, we have considered Erdos-Renyi graphs, where the probability that an edge exists between two vertices, $p_{edge}$, varies respectively from 0.4 to 1 (complete graph). The choice of Erdos-Renyi graph allows us to study the performance of these proposed methods for various sparsity of graphs.
The circuit that we construct is usually not executed as it is in the IBM Quantum hardware. It undergoes a process called transpilation in which
\begin{enumerate}[(i)]
\item the gates of the circuit are replaced with one, or a sequence of, basis gates which are actually executed in the quantum hardware. The basis gates of the IBM Quantum devices are \{$CNOT$, $SX$, $X$, $R_z$ and Identity\} \cite{ibmquantum},
\item the circuit is mapped to the underlying connectivity (called the coupling map) of the hardware \cite{bhattacharjee2020survey},
\item the number of gates in the circuit is reduced using logical equivalence \cite{burgholzer2020advanced}.
\end{enumerate}
A natural question, therefore, is whether the reduction in CNOT gates is retained post transpilation. In Table~\ref{tab:cx_count} we show the number of CNOT gates in the post transpilation circuit for the \textit{ibmq\_manhattan} device as the number of vertices is varied from $10 - 60$ for each of the graph family considered. Our results readily show that the proposed optimization in the number of CNOT gates still hold good even in the transpiled circuit. Since the \textit{ibmq\_manhattan} device is a 65-qubits device, we show the results upto 60 qubits, but the results show that the trend will continue for higher qubit devices as well, when they become available.
\begin{figure*}[t]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[scale=0.3]{erdos04.png}
\caption{Erdos-Renyi graphs with $p_{edge} = 0.4$}
\label{col}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[scale=0.3]{erdos06.png}
\caption{Erdos-Renyi graphs with $p_{edge} = 0.6$}
\label{dfs}
\end{subfigure}
\newline
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[scale=0.3]{erdos08.png}
\caption{Erdos-Renyi graphs with $p_{edge} = 0.8$}
\label{col}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[scale=0.3]{complete.png}
\caption{Complete graphs}
\label{dfs}
\end{subfigure}
\caption{$|\braket{\psi|\psi_e}|^2$ for graphs of various sparsity: Erdos Renyi graphs ($p_{edge} = 0.4,~0.6,~ 0.8$) and complete graphs}
\label{fig:prob_succ}
\end{figure*}
Let $\ket{\psi}$ be the state obtained from the noise-free (ideal) computation of the QAOA circuit, and the state obtained from noisy computation be $\ket{\psi_e}$. The probability of success of the noisy computation, then, is defined as
\begin{equation}
\label{eq:graph_succ}
P_{succ} = |\braket{\psi|\psi_e}|^2
\end{equation}
In Fig.~\ref{fig:prob_succ}(a) - (d) we plot $P_{succ}$ of the traditional QAOA ansatz, Edge Coloring based and the DFS based optimization method for Erdos-Renyi graphs, where $p_{edge}$, the probability that an edge exists between two vertices, varies from 0.4 to 1 (complete graph). The choice of Erdos-Renyi graph allows us to study the performance of these proposed methods for various sparsity of graphs. For each case we vary the number of vertices $n$ from 4 to 12. For each value of $n$ and $p_{edge}$, the results are averaged over 20 input graph instances, and each instance is an average of 100 randomly generated noisy circuits by the simulator model for \textit{ibmq\_manhattan} with noise. Our results readily show that the DFS based method outperforms both the Edge Coloring based method and the traditional QAOA in terms of lower error probability.
From Table~\ref{tab:cx_count}, and our simulation results in Fig.~\ref{fig:prob_succ}(a)-(d), we can infer that the DFS based optimization outperforms the Edge Coloring based optimization, which again, outperforms the traditional QAOA in the reduction in CNOT count, and the probability of error in the circuit in (i) the actual transpiled circuit that is executed on the quantum devices, as well as (ii) in realistic noisy scenario of quantum devices.
\section{Conclusion}
\label{sec:con}
In this paper we have proposed two methods to reduce the number of CNOT gates in the traditional QAOA ansatz. The Edge Coloring based method can reduce upto $\lfloor \frac{n}{2} \rfloor$ CNOT gates whereas the DFS based method can reduce $n - 1$ CNOT gates. While the former method provides a depth-optimized circuit, the latter method can increase the depth of the circuit. We analytically derive the constraint for which a particular increase in depth is acceptable given the number of CNOT gates reduced, and show that every graph satisfies this constraint. Therefore, these methods can reduce the number of CNOT gates in the QAOA ansatz for any graph. Finally, we show via simulation, with the \textit{ibmq\_manhattan} noise model, that the DFS based method outperforms the Edge Coloring based method, which in its turn, outperforms the traditional QAOA in terms of lower error probability in the circuit. The transpiler procedure of Qiskit maps a circuit to the underlying hardware connectivity graph, and some gates are reduced in this process. This transpiled circuit is executed on the real hardware. We show, with the \textit{ibmq\_manhattan} coupling map, that the reduction in the number of CNOT gates still holds post transpilation. Therefore, our proposed methods provide a universal way to an improved QAOA ansatz design. On a final note, all the calculations in this paper considers connected graph, but these carry over easily to disconnected graphs as well.
\section*{Acknowledgement}
We acknowledge the use of IBM Quantum services for this work. The views expressed are those of the authors, and do not reflect the official policy or position of IBM or the IBM Quantum team. In this paper we have used the noise model of \textit{ibmq\_manhattan}, which is one of IBM Quantum Hummingbird r2 Processors.
\section*{Code Availability}
A notebook providing the code to generate the plots of Fig.~\ref{fig:prob_succ}(a)-(d) is available open source at \href{https://github.com/RitajitMajumdar/Optimizing-Ansatz-Design-in-QAOA-for-Max-cut}{https://github.com/RitajitMajumdar/Optimizing-Ansatz-Design-in-QAOA-for-Max-cut}.
\bibliographystyle{unsrt}
|
\subsection{Motion Planning Transformer (MPT)}
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{figs/transformer_fig5}
\caption{Overview of MPT Module. (Start from the top left and move clockwise) The input map, and the start and goal encoded map are concatenated together and passed through a FCN (Fully Connected Networks). The output of which is reshaped and passed through the standard Transformer module. The classifier predicts the probability of each of the token from the encoder output to create the mask superimposed on the output map. The light green patches throughout the pipeline represents how a patch $p\times p$ remains connected with output of the classifier.}
\label{fig:NetArchitecture}
\end{figure}
The MPT module, a region proposal network, is similar to that proposed in \cite{NIPS2015_14bfa6bb}, but apart from having CNN and linear layers, we also include a Transformer model. The Transformer block we use is similar to \cite{vit}, which is derived from \cite{nips_attention}. An overview of the model is shown in Figure \ref{fig:NetArchitecture}. For a given input costmap $\bm{m_c}\in \mathbb{R}^{H\times W}$, start position $\bm{x_s}\in \mathbb{R}^2$, and goal position $\bm{x_g}\in\mathbb{R}^2$ we encode the start and goal position onto an array of size $H\times W$. The goal and start location are represented by patches of size $p\times p$ with values -1 and 1 respectively. The encoded map, $\bm{m_e}\in \mathbb{R}^{H\times W}$, is concatenated with $\bm{m_c}$ to form the input, $\bm{m}\in \mathbb{R}^{H\times W\times 2}$, to the feature extractor consisting of fully convolutional networks (FCN).
To extract the features from the encoder, we use a series of convolution, ReLU, and MaxPool layers, which have a combined receptive area of $p \times p$. We refer to this module as the feature extractor. The feature extractor is similar to the CNN used by the hybrid architecture proposed in \cite{vit}, but instead of using a pre-trained network, we learn the filter parameters from scratch. The feature extractor is similar to the convolutional layers used in Faster R-CNN's \citep{NIPS2015_14bfa6bb}.
The output of the feature extractor is of size $H_l\times W_l \times d$, where $H_l$ and $W_l$ is determined by the size of $\bm{m_c}$ and the convolution layers, and $d$ is the latent dimension of the Transformer. The output is reshaped to size $(H_l\cdot W_l) \times d$ in order to be compatible with the Transformer model. Each row vector is analogous to the word token's in NLP.
Each row of the flattened output of the feature extractor can be mapped to a patch of size $p \times p$ on $\bm{m_c}$ centered at position $\hat{x}\in\mathbb{R}^2$. We call all such points as anchor points. There are $H_l \times W_l$ anchor points for a given input map. To encode the positions of each patch, we add fixed sinusoidal position embeddings to the 1D features similar to \citep{vit, nips_attention}. Since the positional encoding used is of a fixed size, there is an upper bound on the size of the input map.
The modified patch embedding is passed through a series of multi-headed self-attention (MSA), and multi-layer perceptron (MLP) blocks, identical to the architecture used by \cite{nips_attention}. The MLP block consists of two linear layers with RELU non-linearity. We also apply Dropout, Layer Norm and residual connections after every MSA and MLP block. To save on memory, we add gradient checkpoint's \citep{chen2016training} to all MSA blocks. Since the feature extractor and the Transformer block are both invariant to the size of the input map, this architecture is scalable to maps of any size.
Each of the output token from the transformer encoder is passed through a classifier to identify regions that might represent a path. Similar to \cite{NIPS2015_14bfa6bb}, this is implemented using a $1\times 1$ convolution layer. Once all the tokens have been identified, a mask of dimension $H\times W$ is generated which highlights the region with a potential path.
\subsection{Path Planning}
The masked map is used as the reference map for finding a path and checking its feasibility. Any traditional or learning-based planners can be used to solve the given problem. In this work, we use variations of the Rapidly Exploring Random Trees (RRT) algorithm that guarantee optimality \citep{karaman2011sampling,doi:10.1177/0278364915614386,6942976} to find the path under different settings. Note that MPT can enhance any existing path planner. It is robust to different map sizes, making it ideal for real-world scenarios, especially for robot navigation indoors (e.g., hospitals and homes) and outdoors (e.g., streets, highways) environments, where there are areas of variable sizes. MPT selects a subset in a larger map with fixed local receptive fields. The selected local patches become the new input map for the path planner, which would otherwise explore the entire space for pathfinding.
\subsection{Setup}
\subsubsection{Environments}
To test the planning capabilities of our method, we evaluated the model on two different kinds of synthetic environments. The first environment is called the Random Forest, where 100 circular and square objects are randomly placed in different positions and orientations on the map. The second environment is called the Maze environment. A perfect maze is generated using the randomized depth-first search. One characteristic of a perfect maze is that any start and goal pairs on this map are reachable by a collision-free path. The distance resolution of each map was set at 5cm per pixel.
\subsubsection{Robot Models and SMP}
We test our algorithm on two robotic systems. The first is a simple Point Robot Model that can move in any direction in $\mathbb{R}^2$. Planning for this robot is done in the $\mathbb{R}^2$ space using the RRT* \citep{7989581} and Informed-RRT* (IRRT*) \citep{6942976} algorithms. The other robot system is a Dubins Car Model, controlled using wheel velocity and steering angle. The planning for this robot is done in the control space using the Stable Sparse RRT (SST) algorithm \citep{doi:10.1177/0278364915614386}. We used the implementation from the Open Motion Planning Library (OMPL) \citep{sucan2012the-open-motion-planning-library} for all SMPs.
\subsubsection{Dataset}
We created a simulated dataset for the environments and robot models mentioned above. We collected 25 paths for random start and goal pairs for 1750 environments using RRT* for both the Random Forest and Maze environments for the point robot. For the Dubins Car Model, we collected 50 paths for random start and goal pairs for 1000 environments using SST. All maps used for training were of size 480$\times$480.
\subsubsection{Training}
The anchor points are classified as positive if they represent a path patch and negative otherwise. All anchor points near the given path up to 0.7 meters are considered positive, and the rest are considered background. For the Dubins Car Model, we used the robot's position alone to pick the positive anchor points. Similar to \cite{NIPS2015_14bfa6bb}, RPN was trained with each mini-batch containing a single planning problem, with a 1:2 ratio of positive and negative samples. The negative samples are randomly sampled from the set of negative anchor points.
We trained the network by minimizing the cross-entropy loss for the select few anchor points using the Adam optimizer \citep{DBLP:journals/corr/KingmaB14} with $\beta_1 = 0.9$, $\beta_2=0.98$, and $\epsilon=1e^{-9}$. We varied the learning rate as proposed in \cite{nips_attention} with warm-up steps of 3200. Each model was trained for 70k epochs with a batch size of 400. The models were trained on one machine with 4 NVIDIA 2080GTX graphics card and took 30hrs to train.
\subsubsection{Metrics}
To evaluate our planning algorithm, we tested it across different environments for random start and goal pairs. Since all the planners have some form of optimality criteria, it is unclear when to stop searching. A common practice in the motion planning community is to set a cost threshold for each path. The planners stopped searching when the solution path length is below this threshold or a time limit is reached. The cost threshold was the path length from an RRT* planner that searched the space till a solution was found and was collected as part of the validation set. The metrics we are interested in are the planning accuracy, the time it takes to find a solution, and the number of vertices on the planning tree. All the statistics we report, apart from accuracy, consider successful solutions which terminated before the time limit.
\subsection{Point Robot Model}
\begin{figure}[t]
\centering
\begin{subfigure}[c]{\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/forest_model13_epoch24_91_0.png}
\end{subfigure}
\begin{subfigure}[c]{\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/maze_model38_epoch69_991_15.png}
\end{subfigure}
\caption{Planned paths for two different planning problems. Top Row: Random Forest environment. Bottom Row: Maze environment. Left Column: The ground truth path from the validation data. Center Column: The prediction probability from the MPT. Right Column: The masked map and the planned trajectory using RRT*. }
\label{fig:pointRobotPlannedPath}
\end{figure}
To evaluate the performance of our method, we compared it with RRT* and IRRT* without the aid of MPT. The aided versions of these planners are called MPT-RRT* and MPT-IRRT*. The first set of experiments examined the network's ability to generalize to environments of the same size as the training dataset. Hence we evaluated it across 2500 different environments with a single start and goal pair for both the Maze and Random Forest environment. The summary statistics of the experiment are reported in Table \ref{tab:pointRobot_480}. The MPT planners reduce the planning time and vertex count of the planning tree substantially. In Fig. \ref{fig:pointRobotPlannedPath}, we show two examples of a planned path from this experiment.
As the start and goal points are sampled randomly, the statistic's median values only provide a partial picture of the benefits of MPT. To better understand the advantages of MPT, we visualize the distribution of the planning time and vertices on the planning tree in Fig. \ref{fig:pointRobotStats} for the Random Forest and Maze environment using Letter-value plots \citep{letter-value-plot}. These plots help to observe the tail of the distribution of the metrics. Naive RRT* has a heavier tail distribution than MPT-RRT*, which is as expected because, for start and goal pairs further away, the planner needs to generate a denser graph, requiring more time. The MPT-RRT* only searches on the region proposed by the MPT, and as a result, needs fewer tree vertices to reach the goal.
We also observe that IRRT* also performs considerably better than RRT* in the Random Forest environment and achieves similar planning time and planning tree vertices compared to the aided planners. This is because IRRT*, like MPT's, reduces the planning search space once an initial solution is found by bounding the initial path with an Ellipse. Although this works for the Random Forest environments, for more complex maps such as the Maze, we see that MPT-RRT* and MPT-IRRT* perform considerably better than IRRT*.
\begin{figure}[t]
\centering
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width=\textwidth]{figs/pointTimeCombined.png}
\end{subfigure}
\hfill
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width=\textwidth]{figs/pointVerticesCombined.png}
\end{subfigure}
\caption{Planning statistics for the Point Robot Model. Left: The planning time for different SMP and MPT aided SMPs. Right: Number of vertices in the planning tree for SMP and MPT aided SMPs. MPT-RRT* and MPT-IRRT* consistently reduce the planning time and the vertices in the planning tree, resulting in a lower variance of these statistics for these planners.}
\label{fig:pointRobotStats}
\end{figure}
\begin{table}[t]
\caption{Comparing planning accuracy, median planning time, and median number of vertices in the planning tree for the Point Robot Model on environments of the same size as the training data.}
\label{tab:pointRobot_480}
\vspace{2mm}
\centering
\begin{tabular}{lcccccc}
\toprule
\multirow{2}{*}{Environment} & \multicolumn{3}{c}{Random Forest} & \multicolumn{3}{c}{Maze} \\
\cmidrule(lr){2-4}\cmidrule(lr){5-7}
& Accuracy & Time (sec) & Vertices & Accuracy & Time (sec) & Vertices \\
\midrule
RRT* & 99.88\% & 5.44 & 3227.5 & 100\% & 5.36& 2042 \\
IRRT* & 99.88\% & 0.42 & 267 & 100\% & 3.13& 1393.5 \\
MPT-RRT* & 97.68\% & 0.20 & 251 & 98.96\% & 0.83& 615 \\
MPT-IRRT* & 97.68\% & 0.07 & 133 & 98.96\% & 0.74& 557 \\
\bottomrule
\end{tabular}
\end{table}
The next experiment we conducted evaluated the planner's ability to generalize to map sizes that were not part of the training data. We tested the model on four different map sizes, 240$\times$240, 360$\times$240, 480$\times$240, and 560$\times$560, of the Random Forest environment on 1000 randomly generated maps with an increasing number of obstacles. The same MPT model as before was used without any retraining or fine-tuning. Three successful planned paths are shown in Fig. \ref{fig:different_mapSize}. The planning statistics are summarized in Table \ref{tab:different_mapSize}.
We notice that the MPT-RRT* and MPT-IRRT* can achieve lower planning time and tree vertices count compared to RRT* without any additional training or fine-tuning. Thus MPT generalizes well to maps of different sizes. Similar to the previous experiment with fixed map sizes, IRRT* achieves similar performance to MPT-RRT*, although MPT-IRRT* achieves the best performance of the bunch. Better performance of MPT-IRRT* could be attributed to the fact that unlike IRRT*, which has to search the entire space to find an initial solution, MPT-IRRT* samples only within the proposed region, thus arriving at an initial solution faster with fewer vertices on the tree.
For the map size of 560x560, we do see a significant drop in accuracy. This drop in performance might be because the current depth of MSA cannot capture the global information for tokens that are far away. With larger map sizes, the number of input tokens increases quadratically, and as noted in \cite{nips_attention}, a deeper MSA model performs better for longer input sequences. The problem requires further investigation and is an interesting future question to look at.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{figs/forest_point_diffsize.png}
\caption{Plot of paths for Random Forest environments of different size. The architecture of the MPT Model allows flexibility in planning for environments of different sizes. Left: 240$\times$240. Center: 360$\times$240. Right: 480$\times$240.}
\label{fig:different_mapSize}
\vspace{-1.5em}
\end{figure}
\begin{table}[t]
\caption{Comparing planning accuracy, planning time, and number of vertices in the tree for Point Robot on maps of the different sizes of the Random Forest environment.}
\centering
\vspace{2mm}
\begin{tabular}{l|c|ccccc}
\toprule
\multicolumn{1}{l}{Map Sizes}& \multicolumn{1}{c}{\begin{tabular}{c}Number of\\Obstacles\end{tabular}} & &RRT* & IRRT* & MPT-RRT* & MPT-IRRT* \\
\midrule
\multirow{3}{*}{240$\times$240} &\multirow{3}{*}{20} & Accuracy & 100\% & 100\% & 95.4\% & 95.4\% \\
& & Time (sec) & 4.862 & 0.25 & 0.27 & 0.015 \\
& & Vertices & 3199 & 229 & 402.5 & 52 \\
\midrule
\multirow{3}{*}{360$\times$240} &\multirow{3}{*}{35} & Accuracy & 98.9\% & 98.9\% & 97.4\% & 97.4\% \\
& & Time (sec) & 5.93 & 0.29 & 0.27 & 0.06 \\
& & Vertices & 3660 & 257 & 377 & 118 \\
\midrule
\multirow{3}{*}{480$\times$240} &\multirow{3}{*}{50} & Accuracy & 99.4\% & 99.4\% & 96.3\% & 96.3\% \\
& & Time (sec) & 6.31 & 0.39 & 0.27 & 0.07 \\
& & Vertices & 3480 & 291 & 348 & 129.5 \\
\midrule
\multirow{3}{*}{560$\times$560} &\multirow{3}{*}{100} & Accuracy & 100\% & 100\% & 75.6\% & 75.6\%\\
& & Time (sec) & 6.71 & 0.28 & 0.22 & 0.051 \\
& & Vertices & 3854 & 203 & 262 & 101 \\
\bottomrule
\end{tabular}
\label{tab:different_mapSize}
\end{table}
\begin{wraptable}[9]{r}{0.4\textwidth}
\vspace{-1.5em}
\caption{Planning accuracy and time for Dubins Car Model for the Random Forest Environment.}
\label{tab:dubinsCar}
\centering
\begin{tabular}{lcc}
\toprule
Planner & SST & MPT-SST \\
\midrule
Accuracy & 92.2\% & 84.9\% \\
Time (sec)& 28 & 15.8 \\
\bottomrule
\end{tabular}
\end{wraptable}
\subsection{Dubins Car Model}
To examine if the MPT can aid in the planning of non-holonomic systems, we trained a new MPT
model to plan for a Dubins Car Model. We tested the trained model on 1000 Random Forest environments and compared the metrics with those from the naive SST planner. The SST aided by the MPT is called MPT-SST. Two trajectories from this experiment is plotted in Fig. \ref{fig:dubins_car_map} (Left and Center). We can observe that the MPT highlights regions so that SST can generate a kinematically feasible path.
Planners such as SST, which sample in the control space, do not scale well to larger planning spaces, and MPT can help with reducing the search space. The timing results reported in Table \ref{tab:dubinsCar} support this claim. MPT-SST can plan nearly twice as fast as naive SST. The distribution of the planning time for MPT-SST and SST is given in Fig. \ref{fig:dubins_car_map} (Right).
\begin{figure}[t]
\centering
\includegraphics[width=0.325\textwidth]{figs/car_1.png}
\includegraphics[width=.325\textwidth]{figs/car_2.png}
\includegraphics[width=.325\textwidth]{figs/car_time.png}
\caption{MPT can also be trained to aid SMP planners for non-holonomic robots. Left and Center: Planned paths on Random Forest environment using MPT-SST. Right: The Letter-value plot comparing the planning time between SST and MPT-SST. MPT is able to reduce the planning time for SST.}
\label{fig:dubins_car_map}
\end{figure}
\begin{figure}[t]
\centering
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width=\textwidth]{figs/fg79_path_0.png}
\end{subfigure}
\begin{subfigure}[c]{0.49\textwidth}
\includegraphics[width=\textwidth]{figs/fg79_path_15.png}
\end{subfigure}
\caption{Path generated by MPT-RRT* on two random start and goal pairs on a map of building 079 at University of Freiburg. The model used to generate the patches were trained only on synthetic environments.}
\label{fig:real_world_map}
\end{figure}
\subsection{Real World Maps}
The final experiment we performed was to test the model trained on Point Robot Model on a real-world map. We obtained a map of building 079 University of Freiburg from publicly available 2D lidar scan data. We used the MPT model, trained on the synthetic Random Forest and Maze environment for the Point Robot, to plan on 20 randomly sampled start and goal pairs. MPT-RRT* was able to solve 17 of the 20 planning problems, and we plot two of these trajectories in Fig. \ref{fig:real_world_map}. This result is encouraging since it shows MPT's can solve planning problems in real-world maps by training on synthetic maps. Robots that have such generalization capabilities will be able to adapt to dynamic environments easily.
\section{Network Architecture}
\begin{wraptable}[9]{r}{0.4\textwidth}
\vspace{-2.5em}
\centering
\caption{Network architecture of the Feature Extractor}
\begin{tabular}{lc}
\toprule
Layer & Dimension \\
\midrule
2D Convolution & [2, 6, 5, 0] \\
2D Maxpool & [2] \\
ReLU & \\
2D Convolution & [6, 16, 5, 0] \\
2D Maxpool & [2] \\
ReLU & \\
Convolution & [16, 512, 5, 5]\\
\bottomrule
\end{tabular}
\label{tab:feature_extractor}
\end{wraptable}
In this section, we detail the network architecture used in our experiments. The Transformer architecture is similar to the ones proposed in \cite{nips_attention}. We used 6 layers of encoder block, each consisting of 3 heads. The dimension of the keys and queries was set at 512, and the dimension of the value was set at 256. The architecture of our feature extractor is given in Table \ref{tab:feature_extractor}. For the convolution layer, the dimensions in the brackets represent [Input Channel Size, Output Channel Size, Kernel Size, Stride], and for the Maxpool layer, it represents the Kernel Size.
\section{Successful Trajectories}
In this section we plot few more successful trajectories of fixed map sizes for the Random Forest and Maze environments in Fig. \ref{fig:forest_mult} and \ref{fig:maze_mult} respectively.
\begin{figure}[h]
\centering
\begin{subfigure}[c]{\textwidth}
\includegraphics[width=\textwidth]{figs/forest_model13_epoch24_57_0.png}
\end{subfigure}
\begin{subfigure}[c]{\textwidth}
\includegraphics[width=\textwidth]{figs/forest_model13_epoch24_757_0.png}
\end{subfigure}
\begin{subfigure}[c]{\textwidth}
\includegraphics[width=\textwidth]{figs/forest_model13_epoch24_847_0.png}
\end{subfigure}
\caption{Three different trajectories planned successfully using MPT on the Random Forest environment. Left Column: The ground truth path in the validation dataset. Middle Column: The probability estimate made by the MPT. Right Column: The planned path using RRT* on the region proposed by the MPT.}
\label{fig:forest_mult}
\end{figure}
\begin{figure}[h]
\centering
\begin{subfigure}[c]{\textwidth}
\includegraphics[width=\textwidth]{figs/model38_epoch69_300_0.png}
\end{subfigure}
\begin{subfigure}[c]{\textwidth}
\includegraphics[width=\textwidth]{figs/model38_epoch69_346_0.png}
\end{subfigure}
\begin{subfigure}[c]{\textwidth}
\includegraphics[width=\textwidth]{figs/model38_epoch69_900_0.png}
\end{subfigure}
\caption{Three different trajectories planned successfully using MPT on the Maze environment. Left Column: The ground truth path in the validation dataset. Middle Column: The probability estimate made by the MPT. Right Column: The planned path using RRT* on the region proposed by the MPT.}
\label{fig:maze_mult}
\end{figure}
\clearpage
\section{Distribution of Metrics for Maps of Different Sizes}
The distribution of metrics reported in Table 2 is given in Fig. \ref{fig:dist_multi_map}.
\begin{figure}[h]
\centering
\begin{subfigure}[c]{\textwidth}
\includegraphics[width=\textwidth]{figs/comb_diff_size_time.png}
\end{subfigure}
\begin{subfigure}[c]{\textwidth}
\includegraphics[width=\textwidth]{figs/comb_diff_size_vertices.png}
\end{subfigure}
\caption{The distribution of metrics for maps of different sizes. Top: The distribution of planning time for different algorithms for maps of different sizes. Bottom: The distribution of the number of vertices in the planning tree for different algorithms for maps of different sizes. MPT aided planners can achieve faster planning times and lower variance in the metrics across maps of different sizes, which were not part of the training data.}
\label{fig:dist_multi_map}
\end{figure}
\clearpage
\section{Failed Trajectories}
In Fig. \ref{fig:failed_forest} and \ref{fig:failed_maze}, we visualize some of the trajectories that MPT could not solve for the Random Forest and Maze environment, respectively. These failures were mainly caused by a single patch that was not classified as part of the path, resulting in a map with no valid paths.
\begin{figure}[h]
\centering
\begin{subfigure}[c]{\textwidth}
\includegraphics[width=\textwidth]{figs/failed_forest_model13_epoch24_40_0.png}
\end{subfigure}
\begin{subfigure}[c]{\textwidth}
\includegraphics[width=\textwidth]{figs/failed_forest_model13_epoch24_49_0.png}
\end{subfigure}
\begin{subfigure}[c]{\textwidth}
\includegraphics[width=\textwidth]{figs/failed_forest_model13_epoch24_6_0.png}
\end{subfigure}
\caption{Three different trajectories planned unsuccessfully using MPT on the Random Forest environment. Left Column: The ground truth path in the validation dataset. Middle Column: The probability estimate made by the MPT. Right Column: The planned path using RRT* on the region proposed by the MPT. We observe that in all three cases the failure to predict a few patches resulted in a scenario with no valid paths.}
\label{fig:failed_forest}
\end{figure}
\begin{figure}[h]
\centering
\begin{subfigure}[c]{\textwidth}
\includegraphics[width=\textwidth]{figs/failed_model38_epoch69_1713_0.png}
\end{subfigure}
\begin{subfigure}[c]{\textwidth}
\includegraphics[width=\textwidth]{figs/failed_model38_epoch69_43_0.png}
\end{subfigure}
\begin{subfigure}[c]{\textwidth}
\includegraphics[width=\textwidth]{figs/failed_model38_epoch69_99_0.png}
\end{subfigure}
\caption{Three different trajectories planned unsuccessfully using MPT on the Maze environment. Left Column: The ground truth path in the validation dataset. Middle Column: The probability estimate made by the MPT. Right Column: The planned path using RRT* on the region proposed by the MPT. We observe that in all three cases the failure to predict a few patches resulted in a scenario with no valid paths.}
\label{fig:failed_maze}
\end{figure}
\section{Introduction}
\input{2.Introduction}
\section{Related Work}
\input{3.RelatedWorks}
\section{Method}
\input{4.Method}
\section{Experiments}
\input{5.Experiments}
\section{Conclusion}
\input{6.Conclusion}
\medskip
|
\section{Introduction}
Network data, consisting of nodes (objects) and edges (objects' relationships), is ubiquitous in many real-world problems, such as social networks, protein-protein interaction networks, citation networks and so on. Recently, network embedding \cite{cai_comprehensive_2017,wu_comprehensive_2019,chen2019exploiting} techniques, which map the nodes of the original networks into the dense and low-dimensional vectors (called node embeddings) and preserve the network structure information as much as possible, have shown promising performance on many network data analysis tasks, such as node classification \cite{kipf_semi-supervised_2016,sun2021heterogeneous}, link prediction \cite{grover_node2vec:_2016,10.1145/3366423.3380073}, community detection \cite{fortunato2010community} and so on.
Typical network embedding methods could be roughly divided into two categories, unsupervised network embedding methods \cite{9185532} and semi-supervised network embedding methods \cite{8790139}. The former obtains the node embeddings by preserving the network structure information. Representative method like DeepWalk \cite{perozzi_deepwalk:_2014} utilizes the truncated random walks strategy to preserve network local information. The latter, semi-supervised network embedding methods, utilizes not only network structure information but also nodes' label information. Representative method like GCN \cite{kipf_semi-supervised_2016} obtains the target node embeddings by aggregating the neighbor nodes' feature information.
\begin{figure}[t]
\centering
\includegraphics[width=0.45\textwidth]{Figure_1.pdf}
\caption{The 2-dimensional network embedding for the imbalanced network (Zachary's Karate Network \cite{zachary_information_1977}) using: (a) GCN \cite{kipf_semi-supervised_2016} (b) the proposed ImGAGN is capable of discriminating between the real nodes (i.e., the blue circles and red circles) and the generated fake nodes (i.e., the green circles), and also between the minority nodes (i.e., the blue nodes) and the majority nodes (i.e., the red nodes).}
\label{fig2}
\end{figure}
However, the extensive existing network embedding methods assume that the nodes' labels are balanced, i.e., every class has roughly equal number of examples. Generally, these methods could not obtain good performance on the imbalanced networks in which the number of examples of one class (minority) is far less than that of other classes (majority), and the minority usually plays an essential role in the real-world problems. For example, for the fraudulent node detection in the online social networks, the number of fraudsters is far less than that of the normal users, and the fraudsters often try to disguise their identities as the normal users. Therefore, two key challenges of imbalanced network analysis are that: (1) The number of one class examples (minority nodes) is far less than that of other classes (majority nodes) in the network, and the labeling for minority nodes is extremely expensive. (2) The minority nodes are non-separability from the majority nodes, that is, it is difficult to find the support regions of majority and minority nodes in the networks (as shown in Figure 1(a)).
To address the above challenges,
we propose a semi-supervised generative adversarial graph network model, called ImGAGN. It introduces a GraphGenerator which can simulate both the minority class node's attribute distribution and network topological structure distribution by generating a set of minority class nodes linking to the real minority nodes to balance the original network classes distribution, then a GCN discriminator is trained to discriminate between real nodes and fake nodes, and also between minority nodes and majority nodes on the synthetic balanced network. Specifically, as shown in the Figure 1(b), the GraphGenerator iteratively learns to generate a set of minority nodes (green circles in Figure 1(b)) to make the original network classes balanced. The topological structure features of the generated nodes are obtained by linking the fake nodes to the real minority nodes (blue circles in Figure 1(b)) of the original network, and the attribute features of the generated nodes are obtained by averaging their neighbor nodes' (i.e., the real minority nodes) attribute features. Then the discriminator (GCN) is trained to discriminate whether the node is generated by generator and whether the node is minority class. From Figure 1, we can find that ImGAGN could generate a set of appropriate minority nodes to make the original minority nodes separate from the majority nodes, and the generated fake nodes separate from real nodes. We evaluate our proposed method on four publicly available real-world imbalanced datasets on semi-supervised imbalanced node classification task. Experimental results demonstrate that ImGAGN outperforms the state-of-the-art algorithms including both balanced network embedding methods and imbalanced network embedding methods. It is worth emphasizing that the GraphGenerator to generate the new balanced network is done after training/testing split, that is, the generated fake nodes would only be linked to the training minority nodes, but not the testing minority nodes.
The main contributions of this paper are summarized as follows:
\begin{itemize}
\item In this paper, we propose a novel semi-supervised generative adversarial graph network model, called ImGAGN, which utilizes a generator to simulate the minority class node distribution and generates a set of minority nodes to make original network classes balanced. Then GCN is trained to discriminate between the majority and minority nodes, and also between the fake nodes and real nodes on the synthetic balanced network classes.
\item Based on ImGAGN, we propose a novel generator for graph structure data, called GraphGenerator, which can effectively learn not only the nodes' attribute feature distribution but also the network topological structure distribution.
\item The proposed method is validated on four real-world imbalanced network datasets for imbalanced binary node classification and network layouts tasks. Experimental results demonstrate that the proposed method is superior to the state-of-the-art both balanced network embedding and imbalanced network embedding techniques. In addition, we released our codes to facilitate further researchers by others.\footnote{https://github.com/Leo-Q-316/ImGAGN.}
\end{itemize}
The rest of the paper is organized as follows. Section 2 will introduce some main related works. Section 3 will formulate the problem and provide a detailed introduction to the proposed method. In Section 4, we will introduce the experimental setups and results followed by the conclusions in Section 5.
\section{Related Works}
In this section, we introduce two main related research fields including imbalanced learning and imbalanced network embedding.
\subsection{Imbalanced learning}
Imbalanced learning techniques \cite{5128907,johnson2019survey} aim at solving the problem with imbalanced data in which at least the number of one class data (minority) is far less than that of other classes (majority). Generally speaking, the minority class is often high-impact on many real-world problems, such as the cancer detection in medical diagnosis and fraud detection in financial system.
Existing methods for imbalanced learning mainly include: (1) sampling based methods, which learn the imbalanced classification by oversampling \cite{han2005borderline} the minority class or undersampling \cite{liu2008exploratory} the majority class. Representative method like SMOTE \cite{chawla_smote_2002} generates artificial data from existing minority class. (2) cost-sensitive learning based methods \cite{elkan2001foundations,ting2002instance}, which utilize different cost matrices for calculating the cost of any particular data examples misclassified. (3) kernel-based methods \cite{akbani2004applying}, which employ classifier like support vector machines (SVMs) \cite{suykens1999least} to maximize the separation margin. and (4) GANs based methods \cite{shamsolmoali_imbalanced_2020,montahaei_adversarial_2018,douzas2018effective}, which are similar to our proposed method using the generator to create the minority class for balancing the data classes distribution. However, to our best knowledge, little work has employed these GANs based methods to the imbalanced network data.
\subsection{Imbalanced network embedding}
Imbalanced network embedding methods aim at solving the imbalanced learning problems on graph structure data.
GRADE \cite{he_graph-based_nodate} is the classic method for imbalanced network embedding. It utilizes the global similarity matrix to obtain the compact minority class clusters, and learns the decision boundary between majority and minority classes by selecting the examples from the regions where the density changes the most. Wu et al. \cite{wu_imverde:_2018} propose a novel random walk strategy, called vertex-diminished random walk (VDRW), which discourages the random particle to the nodes visited. Based on VDRW, they introduce the semi-supervised network embedding method ImVerde which consists of the context sampling and the balanced-batch sampling strategies to improve the quality of the node-context pairs. SPARC \cite{zhou_sparc:_2018} obtains the imbalanced node embedding in a mutually way, which can jointly predict the minority class and the neighbor context in the networks. RSDNE \cite{wang_rsdne_nodate} explores the network embedding with completely-imbalanced labels. It learns the imbalanced node embedding by allowing the intra-class nodes on the same manifold in the embedding space and removing the known connections between the inter-class nodes. DR-GCN \cite{shi_multi-class_2020} proposes two types of regularization to tackle imbalanced network embedding. It utilizes a conditional adversarial training to discriminate the nodes from different classes, and a distribution alignment training is employed to balance the majority and minority classes learning.
\begin{figure*}[]
\centerline{\includegraphics[width=0.8\textwidth]{Figure_2.pdf}}
\caption{The architecture of ImGAGN. The minority and majority nodes of original imbalanced network are represented by red and blue solid circles respectively, and the synthetic minority nodes generated by GraphGenerator are represented by red hollow circles in artificial synthetic classes balanced network. In addition, The links between real nodes are represented by solid lines, and the links between synthetic minority nodes and real minority nodes are represented by dashed lines.}
\label{fig}
\end{figure*}
\section{Proposed Method}
In this section, we first provide several needed concepts related to the proposed method. Then, we present our proposed method ImGAGN in detail. Finally, we analyze the time complexity of the proposed method.
\subsection{Preliminary}
Before presenting our proposed ImGAGN, we provide a brief introduction to the needed concepts for proposing our method.
\begin{itemize}
\item \textbf{Imbalanced network:} given an imbalanced network $\mathcal{G}_{im}=(V,E,A,X,C)$, where $V$ is the set of $n$ nodes, $E$ is the set of edges, $A$ is the adjacency matrix, $X \in R^{n \times f}$ is the node feature matrix with feature dimension $f$, and $C = \{c_{min},c_{maj}\}$ is the set of node classes. $|c_{min}|$ and $|c_{maj}|$ represent the number of nodes in their classes. The network $\mathcal{G}_{im}=(V,E,A,X,C)$ is an imbalanced network if $|c_{min}|$ is far less than $|c_{maj}|$ (i.e., $|c_{min}| \ll |c_{maj}|$).
\item \textbf{Imbalanced network embedding:} imbalanced network embedding aims at mapping the node $v_{i} \in V$ of an imbalanced network $\mathcal{G}_{im}=(V,E,A,X,C)$ into a continuous low-dimensional vector $\vec{h_{i}} \in R^{d}$ ($d \ll n$), such that the nodes with the same class label are closer than the nodes with the different class labels in the embedding space.
\item \textbf{GANs:} GANs \cite{goodfellow2014generative,gui2020review,wang2019enhancing,wang2018neural,yu2019generating} are a class of neural networks which consist of a generator and a discriminator. The key idea of generator $G$ is that it aims at generating the fake data to simulate the real data distribution to confuse discriminator. The goal of discriminator $D$ is to correctly classify both the real training data and fake data generated from generator $G$. The GANs methods can be formulated as follows \cite{goodfellow2014generative}:
\begin{equation}\begin{array}{l}
\min _{G} \max _{D} V(D, G)=E_{x \sim p_{\text {data }}(x)}[\log D(x)] \\
\quad+E_{z \sim p_{z}(z)}[\log (1-D(G(z)))]
\end{array}\end{equation}
where $x$ is the real data obeying the distribution $p_{data}$, and $z$ is the noise variable obeying the distribution $p_{z}$.
\end{itemize}
\subsection{ImGAGN}
To address the imbalanced classification problems on graph, we propose a GANs based imbalanced learning method, called ImGAGN, which incorporates GCN with a novel generator named GraphGenerator for graph structure data. It generates a set of synthetic minority nodes such that the number of nodes in different classes can be balanced. In addition, GraphGenerator can effectively learn not only the nodes' attribute distribution but also the network topological structure distribution. Then GCN discriminator is trained to discriminate between real nodes and fake (i.e., generated) nodes, and also between minority nodes and majority nodes on the synthetic balanced network. The architecture of ImGAGN is shown in Figure 2.
\subsubsection{\textbf{GraphGenerator (G)}}
Unlike traditional GAN processing regular Euclidean data (e.g., images and text) which data is dependent on each other, the generator only need to learn the data feature distribution. For graph structure data, because the data (i.e., nodes) is independent of each other, the generator needs to learn not only nodes' attribute features distribution (e.g., the node features) but also network topological structure distribution (e.g., the node link relationships). In this paper, we propose a novel generator for graph data, called GraphGenerator, which can generate the node link relationships between the synthetic minority nodes and the real minority nodes, and the features of the synthetic minority nodes are obtained by averaging the features of the linked real minority nodes.
GraphGenerator $G_{graph}: \mathcal{Z} \to \mathcal{F} \times \mathcal{T} $ is a fully connected neural network, where $\mathcal{Z}$ is the noise space with $d_{z}$ dimension, and $\mathcal{F},\mathcal{T}$ are network feature space and network structure space respectively. Specifically, for an imbalanced network $\mathcal{G}_{im}=(V,E,A,X,C)$, let $n_{maj}$ and $n_{min}$ represent the number of majority nodes and the number of minority nodes respectively with $n = n_{maj}+n_{min}$. Let $n_{g} = n_{maj}-n_{min}$ represents the number of nodes needing to be generated for balancing the network classes distribution. Thus, the number of units in input layer is $d_{z}$, and the number of units in output layer is $d_{o} = n_{g} \times n_{min}$. For better understanding, we convert the output vector $\vec{o} \in R^{d_{o}}$ into the matrix form $O \in R^{n_{g} \times n_{min}}$, and then we apply $softmax(O_{i})$ function to normalize each row in $O$ as equation (2):
\begin{equation}
T_{i} = softmax(O_{i}) = \frac{e^{O_{ij}}}{\sum_{j=1}^{n_{min}}e^{O_{ij}}}, for i=1,...,n_{g}
\end{equation}
where each row $O_{i}$ represents the link relationship between each generated minority node to all real minority nodes. In addition, each element $T_{ij}$ represents the link normalized weight between the generate node $u_{i} \in U$ and original minority node $v_{j} \in V$
, where $U$ is the set of generated minority nodes. Thus, $T$ represents the networks topological structure information between generated minority nodes and original minority nodes.
To generate the nodes' attribute features $X_{g} \in R^{n_{min} \times f}$ of the generated minority nodes, we aggregate the neighbor nodes' attribute features of each generated minority nodes as equation (3):
\begin{equation}
X_{g} = TX_{min}
\end{equation}
where $X_{min} \in R^{n_{min} \times f} \subset X$ is the real minority nodes' features matrix of the original imbalanced network $\mathcal{G}_{im}$. And $f$ is the dimension of original minority node features.
\subsubsection{\textbf{Discriminator (D)}}
In this paper, we utilize the two-layer GCN \cite{kipf_semi-supervised_2016} as our discriminator, and the input of GCN is the new network $\mathcal{G}_{bal}=(V^{'},E^{'},A^{'},X^{'},C^{'})$ with balanced classes distribution by incorporating the generated minority nodes from GraphGenerator into the original imbalanced networks $\mathcal{G}_{im}$, where $V^{'}$ represents the new nodes set which consists of the nodes in $\mathcal{G}_{im}$ and the generated minority nodes by GraphGenerator, $E^{'}$ represents the new edges set which consists of the all edges in $\mathcal{G}_{im}$ and the generated edges by GraphGenerator, $A^{'}$ and $X^{'}$ are the new adjacency matrix and feature matrix associated to $V^{'}$ respectively.
$C^{'}=\{(real,minority),(real,majority),(fake,minority)\}$ represents the node labels set. It is worth noting that the GraphGenerator do not generate the majority nodes, thus the label $(fake, majority)$ is not included to $C^{'}$.
The goal of discriminator is to discriminate whether the nodes are generated by generator (i.e., fake) and also whether the node is minority class. Therefore, we can utilize the GCN as a node multi-class classification classifier, and the output $Y$ of GCN is calculated by equation (4) \cite{kipf_semi-supervised_2016} as follows:
\begin{equation}
Y = softmax(\widehat{A^{'}}ReLU(\widehat{A^{'}}X^{'}\Omega^{0})\Omega^{1})
\end{equation}
where $\widehat{A^{'}}=\widehat{D^{-\frac{1}{2}}}(\widehat{A^{'}}+I_{N})\widehat{D^{-\frac{1}{2}}}$ is the pre-processing step following \cite{kipf_semi-supervised_2016} with identity matrix $I_{N}$ and $D_{ij}=\sum_{j}A_{ij}$. $\Omega^{0}$ and $\Omega^{1}$ are input-to-hidden and hidden-to-out weight matrices respectively.
\subsubsection{\textbf{Model Optimization}}
The loss function of the GraphGenerator is as equation (5).
\begin{equation}\begin{array}{l}
\mathcal{L}_{gen} = \mathcal{L}_{rf} + \mathcal{L}_{mi} + \mathcal{L}_{di} + \mathcal{L}_{re} \\
\quad = \sum_{i=1}^{n_{g}}-q_{i}log Pr(\widehat{y_{i}}=real | \vec{x_{i}}) \\
\quad + \sum_{i=1}^{n_{g}}-q_{i}log Pr(\widehat{y_{i}}=minority | \vec{x_{i}}) \\
\quad + \frac{1}{|n_{g}|} \sum_{i=1}^{n_{g}}\sum_{j=1}^{n_{min}} ||\vec{x_{i}}-\vec{x_{j}}||_{2}^{2} \\
\quad + \alpha||\Theta||_{2}^{2}
\end{array}
\end{equation}
where this loss function consists of four terms. The first $\mathcal{L}_{rf}$ and second terms $\mathcal{L}_{mi}$ are the confusing discriminator loss over the generated minority data, in which $q_{i} \in C^{'}$ and $\widehat{y_{i}} \in Y$ denotes the ground-truth labels and the output (prediction probability) of the discriminator respectively, and $\vec{x_{i}}$ is the node embedding vector. The third term $\mathcal{L}_{di}$ aims at making the generated minority nodes close to the real minority nodes. The last term $\mathcal{L}_{re}$ is regularizer, in which $\Theta$ is the set of training weights of GraphGenerator with regularization coefficient $\alpha$.
The loss function of discriminator is as equation (6):
\begin{equation}\begin{array}{l}
\mathcal{L}_{dis} = \mathcal{L}_{fa} + \mathcal{L}_{cl} + \mathcal{L}_{mm} + \mathcal{L}_{ree} \\
\quad = \sum_{i=1}^{n_{g}+n_{min}+n_{maj}} -[q_{i}log(Pr(\widehat{y_{i}}=fake | \vec{x_{i}})) \\ +(1-q_{i})log(1-Pr(\widehat{y_{i}}=fake | \vec{x_{i}}))] \\
\quad + \sum_{i=1}^{n_{g}+n_{min}+n_{maj}} -[q_{i}log(Pr(\widehat{y_{i}}=minority | \vec{x_{i}})) \\ +(1-q_{i})log(1-Pr(\widehat{y_{i}}=minority | \vec{x_{i}}))] \\
\quad - \sum_{i=1}^{n_{min}}\sum_{j=1}^{n_{maj}}||\vec{h_{i}}-\vec{h_{j}}||_{2}^{2} \\
\quad + \beta||\Omega||_{2}^{2}
\end{array}
\end{equation}
where this loss function consists of four terms. The first term $\mathcal{L}_{fa}$ is the cross entropy loss to discriminate that the node is generated by generator or real node of the original network. The second term $\mathcal{L}_{cl}$ is also the cross entropy loss to discriminate that the node is minority class or majority class. The third term $\mathcal{L}_{mm}$ aims at making the embeddings of the different class nodes are far away from each other. The last term $\mathcal{L}_{ree}$ is regularizer, in which $\Omega$ is the set of training weights of the discriminator with regularization coefficient $\beta$.
Finally, the adversarial training objective function of ImGAGN is given as equation (7):
\begin{equation}\begin{array}{l}
\min _{G} \max _{D} V(D, G)=E_{x \sim p_{\text {data }}(x)}[\log D(x)+\mathcal{L}_{cl}+\mathcal{L}_{mm} + \mathcal{L}_{ree}] \\
\quad+E_{z \sim p_{z}(z)}[\log (1-D(G(z)))+\mathcal{L}_{mi}+\mathcal{L}_{di} + \mathcal{L}_{re}]
\end{array}\end{equation}
The goal of GraphGenerator is to generate the fake minority nodes to simulate the real minority nodes distribution to confuse discriminator. The goal of discriminator is to correctly classify between the real training nodes and the fake nodes generated from GraphGenerator, and also between the minority nodes and the majority nodes.
\subsection{Time Complexity}
The time complexity of the proposed ImGAGN is as follows. The complexity for updating generator is $O((L-1)n_{g}H^{2}+n_{g}n_{min}^{2})$, where $L$ is the number of fully connected layers of generator,
and $H$ is the hidden layer dimension size of generator. The complexity for updating discriminator is $O(K|E|d+Knd^{2})$, where $K$ is the number of layers of GCN, $|E|$ is the number of edges, and $d$ is the hidden layer dimension size of GCN. Therefore, the total time complexity of ImGAGN is $O((L-1)n_{g}H^{2}+n_{g}n_{min}^{2})+\lambda_{2}(K|E|d+Knd^{2}))$, where $\lambda_{2}$ is the number of discriminator training steps for once generator training. Furthermore, the time complexity for GCN is $O(K|E|d+Knd^2)$ and the time complexity for GraphSAGE is $O(r^K nd^2)$, where $r$ is GraphSAGE’s batch. Our method is $O((L-1)n_gH^2+n_g n_{min}^2+\lambda_{2} (K|E|d+Knd^2))$, so it means that we only have little time cost $O((L-1)n_gH^2+n_g n_{min}^2)$ which can be simplified to $O(nH^2)$ on the generator in comparison to GCN to deal with imbalanced networks because $\lambda_{2}$ is a constant number (less than 100) and it can be removed for time complexity calculation.
\section{Experiment}
In this section, we conduct the experiments on four real-world datasets to validate the effectiveness of the proposed method. Include the imbalanced node classification task, network layouts task and parameters sensitivity analysis task, aiming to answer the following research questions (RQ):
\begin{itemize}
\item \textbf{RQ1:} How does ImGAGN perform compared with both the state-of-the-art balanced network embedding methods and imbalanced network embedding methods on imbalanced node classification task?
\item \textbf{RQ2:} Can ImGAGN learn the node embeddings such that the representation of minority class nodes can separate from the majority class nodes?
\item \textbf{RQ3:} How do different hyper-parameters (e.g., generated nodes ratio) influence the performance of ImGAGN?
\end{itemize}
\subsection{Experimental setup}
\subsubsection{\textbf{Datasets:}}
We conduct experiments on four publicly real-world datasets including Cora \cite{mccallum2000automating}, Citeseer \cite{giles1998citeseer}, Pubmed \cite{sen2008collective}, and DBLP \cite{tang2008arnetminer} datasets. The statistic information of the datasets is summarized in Table 1.
\begin{table}[htbp]
\centering
\caption{The statistic information of the network datasets.}
\begin{tabular}{c|c|c|c|c}
\toprule
Datasets & Cora & Citeseer & Pubmed & DBLP \\
\midrule
The number of nodes & 2708 & 3312 & 16452 & 20783 \\
The number of edges & 5429 & 4715 & 39308 & 58188 \\
The number of classes & 7 & 6 & 3 & 10 \\
Feature dimension & 1433 & 3703 & 500 & 1000 \\
Ratio of the minority class & 6.65\% & 7.52\% & 5.25\% & 1.31\% \\
\bottomrule
\end{tabular}%
\label{tab:addlabel}%
\end{table}%
\begin{itemize}
\item Cora \cite{mccallum2000automating}, Citeseer \cite{giles1998citeseer}, Pubmed \cite{sen2008collective}, and DBLP \cite{tang2008arnetminer} are the citation network datasets which consist of the nodes representing papers and the edges representing citation relationship between two papers. For each paper, a sparse bag-of-words vector is utilized as the node feature vector. For these four original datasets, the node classes (labels) are defined according to the several research topics, and each class has the roughly equal number of nodes. In our experiments, for validating the effectiveness of the proposed method on the imbalanced networks, following \cite{zhou2018sparc}, all these four balanced networks are reconstructed as the binary imbalanced networks by setting the smallest class as the minority class and the residual classes as the majority class. Specifically, taking Cora dataset for an example, there are seven classes\footnote{Neural Networks: $30.21\%$, Rule Learning: $6.65\%$, Reinforcement Learning: $8.01\%$, Probabilistic Method: $15.73\%$, Theory: $12.96\%$, Genetic Algorithm: $15.44\%$, and Case Based: $11.00\%$.} in total. Thus, the smallest class Rule Learning ($6.65\%$) is used as the minority class, and the residual classes ($93.35\%$,) are used as majority class.
\item For each dataset, the training, validation and testing are randomly split as ratio 7:1:2. It is worth emphasizing that the ImGAGN to generate the new balanced network is done after training/testing split, that is, the generated fake nodes would only be linked to the training minority nodes, but not the testing minority nodes.
\end{itemize}
\begin{table*}[htbp]
\centering
\caption{The imbalanced binary node classification results on Cora, Citeseer, Pubmed and DBLP datasets. The best results are marked in bold.}
\begin{tabular}{c|ccc|ccc|ccc|ccc}
\toprule
Datasets & \multicolumn{3}{c|}{Cora} & \multicolumn{3}{c|}{Citeseer} & \multicolumn{3}{c|}{Pubmed} & \multicolumn{3}{c}{DBLP} \\
\midrule
Metrics & Recall & Precision & AUC & Recall & Precision & AUC & Recall & Precision & AUC & Recall & Precision & AUC \\
\midrule
GCN & 0.7222 & 0.8394 & 0.8973 & 0.32 & 0.5882 & 0.6388 & 0.0014 & 0.0077 & 0.8657 & 0.0363 & 0.6667 & 0.8013 \\
GraphSAGE & 0.8056 & 0.8667 & 0.8926 & 0.32 & 0.4074 & 0.776 & 0.0025 & 0.0054 & 0.8792 & 0.0075 & 0.01 & 0.6125 \\
GCN-SMOTE & 0.8611 & 0.6279 & 0.867 & 0.36 & 0.36 & 0.586 & 0.5376 & 0.0022 & 0.8772 & 0.5273 & 0.8947 & 0.8619 \\
DeepWalk & 0.75 & 0.5676 & 0.883 & 0.18 & 0.012 & 0.572 & 0.3006 & 0.5327 & 0.7835 & 0.3091 & 0.0021 & 0.8648 \\
Node2vec & 0.5833 & 0.4706 & 0.6971 & 0.107 & 0.0031 & 0.5227 & 0.3294 & 0.4474 & 0.8292 & 0.0016 & 0.0056 & 0.4893 \\
LINE & 0.2222 & 0.6327 & 0.8917 & 0.101 & 0.4086 & 0.8071 & 0.0982 & 0.2527 & 0.8639 & 0.0032 & 0.0734 & 0.7701 \\
SPARC & 0.6944 & 0.8333 & 0.8822 & 0.24 & 0.6 & 0.785 & 0.0026 & 0.0039 & 0.1272 & 0.0098 & 0.0017 & 0.4525 \\
DR-GCN & 0.7123 & 0.7899 & 0.8776 & 0.481 & 0.5614 & 0.6102 & 0.4876 & 0.2275 & 0.6714 & 0.519 & 0.7698 & 0.8122 \\
RECT & 0.8944 & 0.8714 & 0.8912 & 0.78 & 0.5455 & 0.7237 & 0.7624 & \textbf{0.6154} & 0.7232 & 0.8182 & \textbf{0.9016} & 0.9227 \\
\textbf{ImGAGN} & \textbf{0.9187} & \textbf{0.893} & \textbf{0.9223} & \textbf{0.84} & \textbf{0.7121} & \textbf{0.8837} & \textbf{0.8768} & 0.5807 & \textbf{0.9086} & \textbf{0.9055} & 0.8525 & \textbf{0.9382} \\
\bottomrule
\end{tabular}%
\label{tab:addlabel}%
\end{table*}%
\subsubsection{\textbf{Comparison Algorithms:}} To validate the effectiveness of the proposed method ImGAGN, we compare it with following nine state-of-the-art methods which can be grouped into two categories: balanced network embedding methods (i.e., GCN, GraphSAGE, DeepWalk, Node2vec and LINE) and imbalanced network embedding methods (i.e., GCN-SOMTE, SPARC, DR-GCN and RECT).
\begin{itemize}
\item \textbf{GCN}: Graph convolutional network (GCN) \cite{kipf_semi-supervised_2016} is the most representative balanced network embedding method which obtains the node embedding by aggregating the neighbor nodes' features.
\item \textbf{GraphSAGE}: GraphSAGE \cite{hamilton2017inductive} is also a representative GNN method. Unlike GCN taking the full-size neighbor nodes to obtain the node embedding, GraphSAGE adopts a fixed number of neighbor nodes for each target node to save the memory. In addition, it learns three different aggretators including Mean-aggregator, LSTM-aggregator and Pooling aggregator. We use the best performances of these three aggregator as the final results of GraphSAGE.
\item \textbf{GCN-SMOTE}: Synthetic minority oversampling technique (SMOTE) \cite{chawla_smote_2002} is the most frequently used method to address the imbalanced classification problem by generating synthetic samples from existing minority samples. In this paper, in order to fully show the performance of the GNN methods, we incorporate the SMOTE technique into GCN for improving its performance on imbalanced network embedding problem. Specifically, we utilize the SMOTE as the data preprocessing technique only on the training set. SOMTE is used to oversample the minority class nodes to balance the classes distribution, and then the GCN is trained on the training set with balanced class distribution.
\item \textbf{DeepWalk}: DeepWalk \cite{perozzi_deepwalk:_2014} is the most representative unsupervised network embedding method which adopts the random walk over the network to sample a set of network paths, and the neural language model (SkipGram) is applied to these network paths to obtain the node embedding.
\item \textbf{Node2vec}: Node2vec \cite{grover_node2vec:_2016} is also an unsupervised network embedding method which obtains the node embedding by using a biased random walk strategy to preserve the homophily and structure equivalence relationships in the networks.
\item \textbf{LINE}: LINE \cite{tang_line:_2015} obtains the network embedding by simultaneously optimizing the first-order and second-order proximities of the networks.
\item \textbf{SPARC}: SPARC \cite{zhou_sparc:_2018} is an imbalanced network embedding method. It obtains the imbalanced embedding in a mutually way, which can jointly predict the minority class and the neighbor context in the networks.
\item \textbf{DR-GCN}: DR-GCN \cite{shi_multi-class_2020} is also a GCN based imbalanced network embedding method, and it proposes to use conditioned adversarial training to enhance the separation of different classes. In addition, the distribution alignment training is applied to balance between the majority nodes and the minority nodes.
\item \textbf{RECT}: RECT \cite{wang_network_2020} is the state-of-the-art imbalanced network embedding method which is a variant of GNN. It obtains the imbalanced network embedding by learning the knowledge of class-semantic information in the networks.
\end{itemize}
\begin{figure*}[t]
\centerline{\includegraphics[width=0.85\textwidth]{Figure_3.pdf}}
\caption{The 2-dimensional imbalanced network layout with t-SNE on Cora dataset. The red circles represent the majority nodes of the original networks. The blue circles represent the minority nodes of the original networks. The yellow circles represent the minority nodes generated by SMOTE. The green circles represent the minority nodes generated by the proposed ImGAGN. The proposed ImGAGN is capable of discriminating between the real nodes (i.e., the blue circles and red circles) and the generated fake nodes (i.e., the green circles), and between the minority nodes (i.e., the blue nodes) and the majority nodes (i.e., the red nodes).}
\label{fig}
\end{figure*}
\subsubsection{\textbf{Parameters:}} All the codes we used are provided by authors. For GCN, following \cite{kipf_semi-supervised_2016}, the number of layers of the networks is set $K=2$. For GraphSAGE, we set $K=2, S_{1}=5, S_{2}=5$ according to the author suggesting. For GCN-SMOTE, the number of generated minority samples by SMOTE is equal to the difference between the majority and minority nodes of the training set. For DeepWalk, we adopt the default hyperparameters (i.e., window size $win=10$, walk length $len = 40$ and the number of walks $t=90$). For Node2vec, we optimize its hyperparameters by a grid search over $p,q \in \{0.25,0.50,1,2,4\}$. For LINE, the hyperparameter negative samples $ns=5$. For SPARC, the length of random walk sequences $\mu =10$. Moreover, the embedding dimension of unsupervised network embedding methods (i.e., DeepWalk, Node2vec and LINE) are set as $d=128$, and the logistic regression classifier is employed to evaluate the node embedding. For semi-supervised network embedding methods (i.e., GCN, GCN-SMOTE, GraphSAGE, SPARC and RECT), we use the outputs of their last hidden layer as the node embedding (the embedding dimension is also $128$).
The hyperparameters of our proposed method ImGAGN are set as follows. For generator, it consists of 3 fully connected layers with $100$ units in input layer and $200$ units in hidden layer. The number of units of output layer is equal to the difference between the majority class and minority class of the training set. $Tanh()$ is utilized as the activation function. For discriminator, it consists of the two-layer GCN followed by a softmax function, and ReLU \cite{glorot_deep_nodate} is utilized as the activation function. In addition, we perform generator and discriminator updates in $1:100$ ratio, and Adam SGD optimizer \cite{kingma_adam_2017} is utilized as the optimizer throughout the experiments.
\subsubsection{\textbf{Repeatability:}}
All the methods are run on a single machine with 14 CPU cores at 2.60GHZ and 2 Tesla P100 GPU with 32G memory using 1 thread.
\subsection{Imbalanced binary node classification (RQ1)}
To answer the RQ1 (i.e., how does ImGAGN perform compared with both the state-of-the-art balanced and imbalanced network embedding), we first conduct imbalanced binary node classification experiment on the four real-world network datasets. Three common classification metrics are used to evaluate the performance for all algorithms. Include: (1) recall, which measures the ratio of correctly classified nodes of all minority test nodes. (2) precision, which measures the ratio of correctly classified nodes of all predicted minority nodes (3) AUC scores, which measures model performance at all classification thresholds. We run experiments 10 times and use average scores for each metric. The experimental results are shown in Table 2.
From experimental results, in general, we can observe that:
\begin{itemize}
\item The proposed method ImGAGN substantially outperforms all comparison methods with respect to recall and AUC on all datasets and is comparable with RECT in terms of precision on Pubmed and DBLP datasets, nonetheless, ImGAGN improves significantly by 11.44\% and 8.73\% on these two datasets respectively in terms of recall which is usually more important than precision in many minority class classification problems, such as rare disease prediction \cite{schubach2017imbalance}. Thus, the overall performance could validate the effectiveness of the proposed method.
\item As expected, the imbalanced network embedding methods (i.e., GCN-SOMTE, SPARC, RECT, DR-GCN and ImGAGN) achieve better performance than the balanced network embedding methods (i.e., GCN, GraphSAGE, DeepWalk, Node2vec and LINE) in most case. It is reasonable since the former methods focus more on label learning of the minority class samples.
\item The GCN-SMOTE achieves better performance than original GCN, which shows that simple oversampling technique is capable of improving original GCN performance on imbalanced network data. The proposed method ImGAGN can also be thought of as an oversampling technique due to the operations of GraphGenerator. However, it obtains better performance than GCN-SMOTE, the improvements could be attributed to the GraphGenerator could well capture both implicit topological structure distribution and nodes' attribute distribution of the minority nodes.
\end{itemize}
\subsection{Network layout (RQ2)}
To answer the RQ2 (i.e., can ImGAGN learn the node embeddings such that the representation of minority class nodes can separate from the majority class nodes? ), we visualize the network layout in the embedding space, and we take Cora dataset for an example. Specifically, we firstly learn the nodes embedding in a 128-dimensional vector space for different network embedding methods, and then employ the t-SNE \cite{maaten2008visualizing} to map the 128-dimensional into the 2-dimensional space for visualization. The experimental results are shown in Figure 3. From the experimental results, in general, we can observe that:
\begin{itemize}
\item The proposed ImGAGN is well capable of discriminating between the real nodes (i.e., the blue circles and red circles) and the generated fake nodes (i.e., the green circles), and also between the minority nodes (i.e., the blue nodes) and the majority nodes(i.e., the red nodes), which validates ImGAGN is able to capture the latent representation of the minority nodes and majority nodes. We attribute such performance to the architecture of the GAN-based methods, that is, the loss functions of GraphGenerator (i.e., equation (4)) and discriminator (i.e., equation (6)) explicitly learn the discrimination between the real nodes and fake nodes, and also between minority nodes and majority nodes.
\item Generally speaking, the end-to-end semi-supervised network embedding methods (i.e., GCN, GCN-SMOTE, GraphSAGE, SPARC, RECT and ImGAGN) can better discriminate the majority and the minority classes than the unsupervised network embedding methods (i.e., DeepWalk, Node2vec and LINE). One explanation is that semi-supervised methods leverage both the nodes' features and label information to learn separable representation of the minority and majority nodes.
\end{itemize}
\subsection{Parameters sensitivity analysis (RQ3)}
To answer the RQ3 (i.e., How do different hyper-parameters influence the performance of ImGAGN?). We conduct the imbalanced node classification experiments on Cora and DBLP datasets and report the performance changes with respect to two crucial hyperparameters of the ImGAGN. One is $\lambda_{1}$ which is the ratio of the number of all training minority nodes (i.e., the original minority nodes and the generated minority nodes in training set) to the number of majority nodes, and we vary it from $0$ to $1$ with step size $0.1$. Another is $\lambda_{2}$ which is the number of discriminator training steps to for once generator training), and we vary it from $10$ to $100$ with step size $10$. The experimental results are shown in Figure 4 and Figure 5. From experimental results, in general, we can observe that:
\begin{itemize}
\item The imbalanced node classification performance, especially for recall, increases with the increase of training minority nodes ratio $\lambda_{1}$ and then tend to flat. One explanation is that when $\lambda_{1}$ is small, the training network is still classes imbalanced, which leads to bad classification performance. It is worth nothing that $\lambda_{1}=0$ degenerates our method to GCN, thus the performance with $\lambda_{1}=0$ also demonstrates that the proposed GraphGenerator could improve the GCN performance on imbalanced networks.
\item The performance increases with the increase of discriminator training steps $\lambda_{2}$ and then tend to flat. It is reasonable since the discriminator needs a certain number of training steps to learn the node embeddings.
\item Particular speaking, we found the proposed method ImGAGN could achieve high performance with $\lambda_{1} > 0.7$ and $\lambda_{2}>50$.
\end{itemize}
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth]{Figure_4.pdf}
\caption{Hyper-parameter sensitivity analysis of the training minority nodes ratio $\lambda_{1}$.}
\label{fig2}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth]{Figure_5.pdf}
\caption{Hyper-parameter sensitivity analysis of the discriminator training steps $\lambda_{2}$.}
\label{fig2}
\end{figure}
\section{Conclusion}
In this paper, to address the imbalanced network embedding problem, we proposed a semi-supervised network embedding method ImGAGN, which utilized a GraphGenerator to simulate both the minority class nodes’ attribute distribution and network topological structure distribution. It generated a set of synthetic minority nodes such that the number of nodes in different classes can be balanced. Then GCN discriminator was trained to discriminate between real nodes and fake nodes, and also between minority nodes and majority nodes. The extensive comparative studies, including the imbalanced node classification, network layouts and hyper-parameters sensitivity analysis, are conducted to validate the effectiveness of the proposed method. The empirical evaluation on four real-world datasets demonstrated that the proposed ImGAGN could outperform the state-of-the-art imbalanced network embedding algorithms on imbalanced node classification task in most cases with respect to recall, precision and AUC. In addition, the visualization results showed that the ImGAGN is capable of learning the node embedding such that the minority class nodes can separate from the majority class nodes.
\section{Acknowledgement}
This work is partially supported by the Science and Technology Innovation Committee Foundation of Shenzhen under the Grant No. JCYJ20200109141235597 and ZDSYS201703031748284, National Science Foundation of China under grant number 61761136008, Shenzhen Peacock Plan under Grant No. KQTD2016112514355531, Program for Guangdong Introducing Innovative and Entrepreneurial Teams under grant number 2017ZT07X386, ARC Discovery Project under the Grant No. DP190101985 and ARC Training Centre for Information Resilience under the Grant No. IC200100022.
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
Network data, consisting of nodes (objects) and edges (objects' relationships), is ubiquitous in many real-world problems, such as social networks, protein-protein interaction networks, citation networks and so on. Recently, network embedding \cite{cai_comprehensive_2017,wu_comprehensive_2019,chen2019exploiting} techniques, which map the nodes of the original networks into the dense and low-dimensional vectors (called node embeddings) and preserve the network structure information as much as possible, have shown promising performance on many network data analysis tasks, such as node classification \cite{kipf_semi-supervised_2016,sun2021heterogeneous}, link prediction \cite{grover_node2vec:_2016,10.1145/3366423.3380073}, community detection \cite{fortunato2010community} and so on.
Typical network embedding methods could be roughly divided into two categories, unsupervised network embedding methods \cite{9185532} and semi-supervised network embedding methods \cite{8790139}. The former obtains the node embeddings by preserving the network structure information. Representative method like DeepWalk \cite{perozzi_deepwalk:_2014} utilizes the truncated random walks strategy to preserve network local information. The latter, semi-supervised network embedding methods, utilizes not only network structure information but also nodes' label information. Representative method like GCN \cite{kipf_semi-supervised_2016} obtains the target node embeddings by aggregating the neighbor nodes' feature information.
\begin{figure}[t]
\centering
\includegraphics[width=0.45\textwidth]{Figure_1.pdf}
\caption{The 2-dimensional network embedding for the imbalanced network (Zachary's Karate Network \cite{zachary_information_1977}) using: (a) GCN \cite{kipf_semi-supervised_2016} (b) the proposed ImGAGN is capable of discriminating between the real nodes (i.e., the blue circles and red circles) and the generated fake nodes (i.e., the green circles), and also between the minority nodes (i.e., the blue nodes) and the majority nodes (i.e., the red nodes).}
\label{fig2}
\end{figure}
However, the extensive existing network embedding methods assume that the nodes' labels are balanced, i.e., every class has roughly equal number of examples. Generally, these methods could not obtain good performance on the imbalanced networks in which the number of examples of one class (minority) is far less than that of other classes (majority), and the minority usually plays an essential role in the real-world problems. For example, for the fraudulent node detection in the online social networks, the number of fraudsters is far less than that of the normal users, and the fraudsters often try to disguise their identities as the normal users. Therefore, two key challenges of imbalanced network analysis are that: (1) The number of one class examples (minority nodes) is far less than that of other classes (majority nodes) in the network, and the labeling for minority nodes is extremely expensive. (2) The minority nodes are non-separability from the majority nodes, that is, it is difficult to find the support regions of majority and minority nodes in the networks (as shown in Figure 1(a)).
To address the above challenges,
we propose a semi-supervised generative adversarial graph network model, called ImGAGN. It introduces a GraphGenerator which can simulate both the minority class node's attribute distribution and network topological structure distribution by generating a set of minority class nodes linking to the real minority nodes to balance the original network classes distribution, then a GCN discriminator is trained to discriminate between real nodes and fake nodes, and also between minority nodes and majority nodes on the synthetic balanced network. Specifically, as shown in the Figure 1(b), the GraphGenerator iteratively learns to generate a set of minority nodes (green circles in Figure 1(b)) to make the original network classes balanced. The topological structure features of the generated nodes are obtained by linking the fake nodes to the real minority nodes (blue circles in Figure 1(b)) of the original network, and the attribute features of the generated nodes are obtained by averaging their neighbor nodes' (i.e., the real minority nodes) attribute features. Then the discriminator (GCN) is trained to discriminate whether the node is generated by generator and whether the node is minority class. From Figure 1, we can find that ImGAGN could generate a set of appropriate minority nodes to make the original minority nodes separate from the majority nodes, and the generated fake nodes separate from real nodes. We evaluate our proposed method on four publicly available real-world imbalanced datasets on semi-supervised imbalanced node classification task. Experimental results demonstrate that ImGAGN outperforms the state-of-the-art algorithms including both balanced network embedding methods and imbalanced network embedding methods. It is worth emphasizing that the GraphGenerator to generate the new balanced network is done after training/testing split, that is, the generated fake nodes would only be linked to the training minority nodes, but not the testing minority nodes.
The main contributions of this paper are summarized as follows:
\begin{itemize}
\item In this paper, we propose a novel semi-supervised generative adversarial graph network model, called ImGAGN, which utilizes a generator to simulate the minority class node distribution and generates a set of minority nodes to make original network classes balanced. Then GCN is trained to discriminate between the majority and minority nodes, and also between the fake nodes and real nodes on the synthetic balanced network classes.
\item Based on ImGAGN, we propose a novel generator for graph structure data, called GraphGenerator, which can effectively learn not only the nodes' attribute feature distribution but also the network topological structure distribution.
\item The proposed method is validated on four real-world imbalanced network datasets for imbalanced binary node classification and network layouts tasks. Experimental results demonstrate that the proposed method is superior to the state-of-the-art both balanced network embedding and imbalanced network embedding techniques. In addition, we released our codes to facilitate further researchers by others.\footnote{https://github.com/Leo-Q-316/ImGAGN.}
\end{itemize}
The rest of the paper is organized as follows. Section 2 will introduce some main related works. Section 3 will formulate the problem and provide a detailed introduction to the proposed method. In Section 4, we will introduce the experimental setups and results followed by the conclusions in Section 5.
\section{Related Works}
In this section, we introduce two main related research fields including imbalanced learning and imbalanced network embedding.
\subsection{Imbalanced learning}
Imbalanced learning techniques \cite{5128907,johnson2019survey} aim at solving the problem with imbalanced data in which at least the number of one class data (minority) is far less than that of other classes (majority). Generally speaking, the minority class is often high-impact on many real-world problems, such as the cancer detection in medical diagnosis and fraud detection in financial system.
Existing methods for imbalanced learning mainly include: (1) sampling based methods, which learn the imbalanced classification by oversampling \cite{han2005borderline} the minority class or undersampling \cite{liu2008exploratory} the majority class. Representative method like SMOTE \cite{chawla_smote_2002} generates artificial data from existing minority class. (2) cost-sensitive learning based methods \cite{elkan2001foundations,ting2002instance}, which utilize different cost matrices for calculating the cost of any particular data examples misclassified. (3) kernel-based methods \cite{akbani2004applying}, which employ classifier like support vector machines (SVMs) \cite{suykens1999least} to maximize the separation margin. and (4) GANs based methods \cite{shamsolmoali_imbalanced_2020,montahaei_adversarial_2018,douzas2018effective}, which are similar to our proposed method using the generator to create the minority class for balancing the data classes distribution. However, to our best knowledge, little work has employed these GANs based methods to the imbalanced network data.
\subsection{Imbalanced network embedding}
Imbalanced network embedding methods aim at solving the imbalanced learning problems on graph structure data.
GRADE \cite{he_graph-based_nodate} is the classic method for imbalanced network embedding. It utilizes the global similarity matrix to obtain the compact minority class clusters, and learns the decision boundary between majority and minority classes by selecting the examples from the regions where the density changes the most. Wu et al. \cite{wu_imverde:_2018} propose a novel random walk strategy, called vertex-diminished random walk (VDRW), which discourages the random particle to the nodes visited. Based on VDRW, they introduce the semi-supervised network embedding method ImVerde which consists of the context sampling and the balanced-batch sampling strategies to improve the quality of the node-context pairs. SPARC \cite{zhou_sparc:_2018} obtains the imbalanced node embedding in a mutually way, which can jointly predict the minority class and the neighbor context in the networks. RSDNE \cite{wang_rsdne_nodate} explores the network embedding with completely-imbalanced labels. It learns the imbalanced node embedding by allowing the intra-class nodes on the same manifold in the embedding space and removing the known connections between the inter-class nodes. DR-GCN \cite{shi_multi-class_2020} proposes two types of regularization to tackle imbalanced network embedding. It utilizes a conditional adversarial training to discriminate the nodes from different classes, and a distribution alignment training is employed to balance the majority and minority classes learning.
\begin{figure*}[]
\centerline{\includegraphics[width=0.8\textwidth]{Figure_2.pdf}}
\caption{The architecture of ImGAGN. The minority and majority nodes of original imbalanced network are represented by red and blue solid circles respectively, and the synthetic minority nodes generated by GraphGenerator are represented by red hollow circles in artificial synthetic classes balanced network. In addition, The links between real nodes are represented by solid lines, and the links between synthetic minority nodes and real minority nodes are represented by dashed lines.}
\label{fig}
\end{figure*}
\section{Proposed Method}
In this section, we first provide several needed concepts related to the proposed method. Then, we present our proposed method ImGAGN in detail. Finally, we analyze the time complexity of the proposed method.
\subsection{Preliminary}
Before presenting our proposed ImGAGN, we provide a brief introduction to the needed concepts for proposing our method.
\begin{itemize}
\item \textbf{Imbalanced network:} given an imbalanced network $\mathcal{G}_{im}=(V,E,A,X,C)$, where $V$ is the set of $n$ nodes, $E$ is the set of edges, $A$ is the adjacency matrix, $X \in R^{n \times f}$ is the node feature matrix with feature dimension $f$, and $C = \{c_{min},c_{maj}\}$ is the set of node classes. $|c_{min}|$ and $|c_{maj}|$ represent the number of nodes in their classes. The network $\mathcal{G}_{im}=(V,E,A,X,C)$ is an imbalanced network if $|c_{min}|$ is far less than $|c_{maj}|$ (i.e., $|c_{min}| \ll |c_{maj}|$).
\item \textbf{Imbalanced network embedding:} imbalanced network embedding aims at mapping the node $v_{i} \in V$ of an imbalanced network $\mathcal{G}_{im}=(V,E,A,X,C)$ into a continuous low-dimensional vector $\vec{h_{i}} \in R^{d}$ ($d \ll n$), such that the nodes with the same class label are closer than the nodes with the different class labels in the embedding space.
\item \textbf{GANs:} GANs \cite{goodfellow2014generative,gui2020review,wang2019enhancing,wang2018neural,yu2019generating} are a class of neural networks which consist of a generator and a discriminator. The key idea of generator $G$ is that it aims at generating the fake data to simulate the real data distribution to confuse discriminator. The goal of discriminator $D$ is to correctly classify both the real training data and fake data generated from generator $G$. The GANs methods can be formulated as follows \cite{goodfellow2014generative}:
\begin{equation}\begin{array}{l}
\min _{G} \max _{D} V(D, G)=E_{x \sim p_{\text {data }}(x)}[\log D(x)] \\
\quad+E_{z \sim p_{z}(z)}[\log (1-D(G(z)))]
\end{array}\end{equation}
where $x$ is the real data obeying the distribution $p_{data}$, and $z$ is the noise variable obeying the distribution $p_{z}$.
\end{itemize}
\subsection{ImGAGN}
To address the imbalanced classification problems on graph, we propose a GANs based imbalanced learning method, called ImGAGN, which incorporates GCN with a novel generator named GraphGenerator for graph structure data. It generates a set of synthetic minority nodes such that the number of nodes in different classes can be balanced. In addition, GraphGenerator can effectively learn not only the nodes' attribute distribution but also the network topological structure distribution. Then GCN discriminator is trained to discriminate between real nodes and fake (i.e., generated) nodes, and also between minority nodes and majority nodes on the synthetic balanced network. The architecture of ImGAGN is shown in Figure 2.
\subsubsection{\textbf{GraphGenerator (G)}}
Unlike traditional GAN processing regular Euclidean data (e.g., images and text) which data is dependent on each other, the generator only need to learn the data feature distribution. For graph structure data, because the data (i.e., nodes) is independent of each other, the generator needs to learn not only nodes' attribute features distribution (e.g., the node features) but also network topological structure distribution (e.g., the node link relationships). In this paper, we propose a novel generator for graph data, called GraphGenerator, which can generate the node link relationships between the synthetic minority nodes and the real minority nodes, and the features of the synthetic minority nodes are obtained by averaging the features of the linked real minority nodes.
GraphGenerator $G_{graph}: \mathcal{Z} \to \mathcal{F} \times \mathcal{T} $ is a fully connected neural network, where $\mathcal{Z}$ is the noise space with $d_{z}$ dimension, and $\mathcal{F},\mathcal{T}$ are network feature space and network structure space respectively. Specifically, for an imbalanced network $\mathcal{G}_{im}=(V,E,A,X,C)$, let $n_{maj}$ and $n_{min}$ represent the number of majority nodes and the number of minority nodes respectively with $n = n_{maj}+n_{min}$. Let $n_{g} = n_{maj}-n_{min}$ represents the number of nodes needing to be generated for balancing the network classes distribution. Thus, the number of units in input layer is $d_{z}$, and the number of units in output layer is $d_{o} = n_{g} \times n_{min}$. For better understanding, we convert the output vector $\vec{o} \in R^{d_{o}}$ into the matrix form $O \in R^{n_{g} \times n_{min}}$, and then we apply $softmax(O_{i})$ function to normalize each row in $O$ as equation (2):
\begin{equation}
T_{i} = softmax(O_{i}) = \frac{e^{O_{ij}}}{\sum_{j=1}^{n_{min}}e^{O_{ij}}}, for i=1,...,n_{g}
\end{equation}
where each row $O_{i}$ represents the link relationship between each generated minority node to all real minority nodes. In addition, each element $T_{ij}$ represents the link normalized weight between the generate node $u_{i} \in U$ and original minority node $v_{j} \in V$
, where $U$ is the set of generated minority nodes. Thus, $T$ represents the networks topological structure information between generated minority nodes and original minority nodes.
To generate the nodes' attribute features $X_{g} \in R^{n_{min} \times f}$ of the generated minority nodes, we aggregate the neighbor nodes' attribute features of each generated minority nodes as equation (3):
\begin{equation}
X_{g} = TX_{min}
\end{equation}
where $X_{min} \in R^{n_{min} \times f} \subset X$ is the real minority nodes' features matrix of the original imbalanced network $\mathcal{G}_{im}$. And $f$ is the dimension of original minority node features.
\subsubsection{\textbf{Discriminator (D)}}
In this paper, we utilize the two-layer GCN \cite{kipf_semi-supervised_2016} as our discriminator, and the input of GCN is the new network $\mathcal{G}_{bal}=(V^{'},E^{'},A^{'},X^{'},C^{'})$ with balanced classes distribution by incorporating the generated minority nodes from GraphGenerator into the original imbalanced networks $\mathcal{G}_{im}$, where $V^{'}$ represents the new nodes set which consists of the nodes in $\mathcal{G}_{im}$ and the generated minority nodes by GraphGenerator, $E^{'}$ represents the new edges set which consists of the all edges in $\mathcal{G}_{im}$ and the generated edges by GraphGenerator, $A^{'}$ and $X^{'}$ are the new adjacency matrix and feature matrix associated to $V^{'}$ respectively.
$C^{'}=\{(real,minority),(real,majority),(fake,minority)\}$ represents the node labels set. It is worth noting that the GraphGenerator do not generate the majority nodes, thus the label $(fake, majority)$ is not included to $C^{'}$.
The goal of discriminator is to discriminate whether the nodes are generated by generator (i.e., fake) and also whether the node is minority class. Therefore, we can utilize the GCN as a node multi-class classification classifier, and the output $Y$ of GCN is calculated by equation (4) \cite{kipf_semi-supervised_2016} as follows:
\begin{equation}
Y = softmax(\widehat{A^{'}}ReLU(\widehat{A^{'}}X^{'}\Omega^{0})\Omega^{1})
\end{equation}
where $\widehat{A^{'}}=\widehat{D^{-\frac{1}{2}}}(\widehat{A^{'}}+I_{N})\widehat{D^{-\frac{1}{2}}}$ is the pre-processing step following \cite{kipf_semi-supervised_2016} with identity matrix $I_{N}$ and $D_{ij}=\sum_{j}A_{ij}$. $\Omega^{0}$ and $\Omega^{1}$ are input-to-hidden and hidden-to-out weight matrices respectively.
\subsubsection{\textbf{Model Optimization}}
The loss function of the GraphGenerator is as equation (5).
\begin{equation}\begin{array}{l}
\mathcal{L}_{gen} = \mathcal{L}_{rf} + \mathcal{L}_{mi} + \mathcal{L}_{di} + \mathcal{L}_{re} \\
\quad = \sum_{i=1}^{n_{g}}-q_{i}log Pr(\widehat{y_{i}}=real | \vec{x_{i}}) \\
\quad + \sum_{i=1}^{n_{g}}-q_{i}log Pr(\widehat{y_{i}}=minority | \vec{x_{i}}) \\
\quad + \frac{1}{|n_{g}|} \sum_{i=1}^{n_{g}}\sum_{j=1}^{n_{min}} ||\vec{x_{i}}-\vec{x_{j}}||_{2}^{2} \\
\quad + \alpha||\Theta||_{2}^{2}
\end{array}
\end{equation}
where this loss function consists of four terms. The first $\mathcal{L}_{rf}$ and second terms $\mathcal{L}_{mi}$ are the confusing discriminator loss over the generated minority data, in which $q_{i} \in C^{'}$ and $\widehat{y_{i}} \in Y$ denotes the ground-truth labels and the output (prediction probability) of the discriminator respectively, and $\vec{x_{i}}$ is the node embedding vector. The third term $\mathcal{L}_{di}$ aims at making the generated minority nodes close to the real minority nodes. The last term $\mathcal{L}_{re}$ is regularizer, in which $\Theta$ is the set of training weights of GraphGenerator with regularization coefficient $\alpha$.
The loss function of discriminator is as equation (6):
\begin{equation}\begin{array}{l}
\mathcal{L}_{dis} = \mathcal{L}_{fa} + \mathcal{L}_{cl} + \mathcal{L}_{mm} + \mathcal{L}_{ree} \\
\quad = \sum_{i=1}^{n_{g}+n_{min}+n_{maj}} -[q_{i}log(Pr(\widehat{y_{i}}=fake | \vec{x_{i}})) \\ +(1-q_{i})log(1-Pr(\widehat{y_{i}}=fake | \vec{x_{i}}))] \\
\quad + \sum_{i=1}^{n_{g}+n_{min}+n_{maj}} -[q_{i}log(Pr(\widehat{y_{i}}=minority | \vec{x_{i}})) \\ +(1-q_{i})log(1-Pr(\widehat{y_{i}}=minority | \vec{x_{i}}))] \\
\quad - \sum_{i=1}^{n_{min}}\sum_{j=1}^{n_{maj}}||\vec{h_{i}}-\vec{h_{j}}||_{2}^{2} \\
\quad + \beta||\Omega||_{2}^{2}
\end{array}
\end{equation}
where this loss function consists of four terms. The first term $\mathcal{L}_{fa}$ is the cross entropy loss to discriminate that the node is generated by generator or real node of the original network. The second term $\mathcal{L}_{cl}$ is also the cross entropy loss to discriminate that the node is minority class or majority class. The third term $\mathcal{L}_{mm}$ aims at making the embeddings of the different class nodes are far away from each other. The last term $\mathcal{L}_{ree}$ is regularizer, in which $\Omega$ is the set of training weights of the discriminator with regularization coefficient $\beta$.
Finally, the adversarial training objective function of ImGAGN is given as equation (7):
\begin{equation}\begin{array}{l}
\min _{G} \max _{D} V(D, G)=E_{x \sim p_{\text {data }}(x)}[\log D(x)+\mathcal{L}_{cl}+\mathcal{L}_{mm} + \mathcal{L}_{ree}] \\
\quad+E_{z \sim p_{z}(z)}[\log (1-D(G(z)))+\mathcal{L}_{mi}+\mathcal{L}_{di} + \mathcal{L}_{re}]
\end{array}\end{equation}
The goal of GraphGenerator is to generate the fake minority nodes to simulate the real minority nodes distribution to confuse discriminator. The goal of discriminator is to correctly classify between the real training nodes and the fake nodes generated from GraphGenerator, and also between the minority nodes and the majority nodes.
\subsection{Time Complexity}
The time complexity of the proposed ImGAGN is as follows. The complexity for updating generator is $O((L-1)n_{g}H^{2}+n_{g}n_{min}^{2})$, where $L$ is the number of fully connected layers of generator,
and $H$ is the hidden layer dimension size of generator. The complexity for updating discriminator is $O(K|E|d+Knd^{2})$, where $K$ is the number of layers of GCN, $|E|$ is the number of edges, and $d$ is the hidden layer dimension size of GCN. Therefore, the total time complexity of ImGAGN is $O((L-1)n_{g}H^{2}+n_{g}n_{min}^{2})+\lambda_{2}(K|E|d+Knd^{2}))$, where $\lambda_{2}$ is the number of discriminator training steps for once generator training. Furthermore, the time complexity for GCN is $O(K|E|d+Knd^2)$ and the time complexity for GraphSAGE is $O(r^K nd^2)$, where $r$ is GraphSAGE’s batch. Our method is $O((L-1)n_gH^2+n_g n_{min}^2+\lambda_{2} (K|E|d+Knd^2))$, so it means that we only have little time cost $O((L-1)n_gH^2+n_g n_{min}^2)$ which can be simplified to $O(nH^2)$ on the generator in comparison to GCN to deal with imbalanced networks because $\lambda_{2}$ is a constant number (less than 100) and it can be removed for time complexity calculation.
\section{Experiment}
In this section, we conduct the experiments on four real-world datasets to validate the effectiveness of the proposed method. Include the imbalanced node classification task, network layouts task and parameters sensitivity analysis task, aiming to answer the following research questions (RQ):
\begin{itemize}
\item \textbf{RQ1:} How does ImGAGN perform compared with both the state-of-the-art balanced network embedding methods and imbalanced network embedding methods on imbalanced node classification task?
\item \textbf{RQ2:} Can ImGAGN learn the node embeddings such that the representation of minority class nodes can separate from the majority class nodes?
\item \textbf{RQ3:} How do different hyper-parameters (e.g., generated nodes ratio) influence the performance of ImGAGN?
\end{itemize}
\subsection{Experimental setup}
\subsubsection{\textbf{Datasets:}}
We conduct experiments on four publicly real-world datasets including Cora \cite{mccallum2000automating}, Citeseer \cite{giles1998citeseer}, Pubmed \cite{sen2008collective}, and DBLP \cite{tang2008arnetminer} datasets. The statistic information of the datasets is summarized in Table 1.
\begin{table}[htbp]
\centering
\caption{The statistic information of the network datasets.}
\begin{tabular}{c|c|c|c|c}
\toprule
Datasets & Cora & Citeseer & Pubmed & DBLP \\
\midrule
The number of nodes & 2708 & 3312 & 16452 & 20783 \\
The number of edges & 5429 & 4715 & 39308 & 58188 \\
The number of classes & 7 & 6 & 3 & 10 \\
Feature dimension & 1433 & 3703 & 500 & 1000 \\
Ratio of the minority class & 6.65\% & 7.52\% & 5.25\% & 1.31\% \\
\bottomrule
\end{tabular}%
\label{tab:addlabel}%
\end{table}%
\begin{itemize}
\item Cora \cite{mccallum2000automating}, Citeseer \cite{giles1998citeseer}, Pubmed \cite{sen2008collective}, and DBLP \cite{tang2008arnetminer} are the citation network datasets which consist of the nodes representing papers and the edges representing citation relationship between two papers. For each paper, a sparse bag-of-words vector is utilized as the node feature vector. For these four original datasets, the node classes (labels) are defined according to the several research topics, and each class has the roughly equal number of nodes. In our experiments, for validating the effectiveness of the proposed method on the imbalanced networks, following \cite{zhou2018sparc}, all these four balanced networks are reconstructed as the binary imbalanced networks by setting the smallest class as the minority class and the residual classes as the majority class. Specifically, taking Cora dataset for an example, there are seven classes\footnote{Neural Networks: $30.21\%$, Rule Learning: $6.65\%$, Reinforcement Learning: $8.01\%$, Probabilistic Method: $15.73\%$, Theory: $12.96\%$, Genetic Algorithm: $15.44\%$, and Case Based: $11.00\%$.} in total. Thus, the smallest class Rule Learning ($6.65\%$) is used as the minority class, and the residual classes ($93.35\%$,) are used as majority class.
\item For each dataset, the training, validation and testing are randomly split as ratio 7:1:2. It is worth emphasizing that the ImGAGN to generate the new balanced network is done after training/testing split, that is, the generated fake nodes would only be linked to the training minority nodes, but not the testing minority nodes.
\end{itemize}
\begin{table*}[htbp]
\centering
\caption{The imbalanced binary node classification results on Cora, Citeseer, Pubmed and DBLP datasets. The best results are marked in bold.}
\begin{tabular}{c|ccc|ccc|ccc|ccc}
\toprule
Datasets & \multicolumn{3}{c|}{Cora} & \multicolumn{3}{c|}{Citeseer} & \multicolumn{3}{c|}{Pubmed} & \multicolumn{3}{c}{DBLP} \\
\midrule
Metrics & Recall & Precision & AUC & Recall & Precision & AUC & Recall & Precision & AUC & Recall & Precision & AUC \\
\midrule
GCN & 0.7222 & 0.8394 & 0.8973 & 0.32 & 0.5882 & 0.6388 & 0.0014 & 0.0077 & 0.8657 & 0.0363 & 0.6667 & 0.8013 \\
GraphSAGE & 0.8056 & 0.8667 & 0.8926 & 0.32 & 0.4074 & 0.776 & 0.0025 & 0.0054 & 0.8792 & 0.0075 & 0.01 & 0.6125 \\
GCN-SMOTE & 0.8611 & 0.6279 & 0.867 & 0.36 & 0.36 & 0.586 & 0.5376 & 0.0022 & 0.8772 & 0.5273 & 0.8947 & 0.8619 \\
DeepWalk & 0.75 & 0.5676 & 0.883 & 0.18 & 0.012 & 0.572 & 0.3006 & 0.5327 & 0.7835 & 0.3091 & 0.0021 & 0.8648 \\
Node2vec & 0.5833 & 0.4706 & 0.6971 & 0.107 & 0.0031 & 0.5227 & 0.3294 & 0.4474 & 0.8292 & 0.0016 & 0.0056 & 0.4893 \\
LINE & 0.2222 & 0.6327 & 0.8917 & 0.101 & 0.4086 & 0.8071 & 0.0982 & 0.2527 & 0.8639 & 0.0032 & 0.0734 & 0.7701 \\
SPARC & 0.6944 & 0.8333 & 0.8822 & 0.24 & 0.6 & 0.785 & 0.0026 & 0.0039 & 0.1272 & 0.0098 & 0.0017 & 0.4525 \\
DR-GCN & 0.7123 & 0.7899 & 0.8776 & 0.481 & 0.5614 & 0.6102 & 0.4876 & 0.2275 & 0.6714 & 0.519 & 0.7698 & 0.8122 \\
RECT & 0.8944 & 0.8714 & 0.8912 & 0.78 & 0.5455 & 0.7237 & 0.7624 & \textbf{0.6154} & 0.7232 & 0.8182 & \textbf{0.9016} & 0.9227 \\
\textbf{ImGAGN} & \textbf{0.9187} & \textbf{0.893} & \textbf{0.9223} & \textbf{0.84} & \textbf{0.7121} & \textbf{0.8837} & \textbf{0.8768} & 0.5807 & \textbf{0.9086} & \textbf{0.9055} & 0.8525 & \textbf{0.9382} \\
\bottomrule
\end{tabular}%
\label{tab:addlabel}%
\end{table*}%
\subsubsection{\textbf{Comparison Algorithms:}} To validate the effectiveness of the proposed method ImGAGN, we compare it with following nine state-of-the-art methods which can be grouped into two categories: balanced network embedding methods (i.e., GCN, GraphSAGE, DeepWalk, Node2vec and LINE) and imbalanced network embedding methods (i.e., GCN-SOMTE, SPARC, DR-GCN and RECT).
\begin{itemize}
\item \textbf{GCN}: Graph convolutional network (GCN) \cite{kipf_semi-supervised_2016} is the most representative balanced network embedding method which obtains the node embedding by aggregating the neighbor nodes' features.
\item \textbf{GraphSAGE}: GraphSAGE \cite{hamilton2017inductive} is also a representative GNN method. Unlike GCN taking the full-size neighbor nodes to obtain the node embedding, GraphSAGE adopts a fixed number of neighbor nodes for each target node to save the memory. In addition, it learns three different aggretators including Mean-aggregator, LSTM-aggregator and Pooling aggregator. We use the best performances of these three aggregator as the final results of GraphSAGE.
\item \textbf{GCN-SMOTE}: Synthetic minority oversampling technique (SMOTE) \cite{chawla_smote_2002} is the most frequently used method to address the imbalanced classification problem by generating synthetic samples from existing minority samples. In this paper, in order to fully show the performance of the GNN methods, we incorporate the SMOTE technique into GCN for improving its performance on imbalanced network embedding problem. Specifically, we utilize the SMOTE as the data preprocessing technique only on the training set. SOMTE is used to oversample the minority class nodes to balance the classes distribution, and then the GCN is trained on the training set with balanced class distribution.
\item \textbf{DeepWalk}: DeepWalk \cite{perozzi_deepwalk:_2014} is the most representative unsupervised network embedding method which adopts the random walk over the network to sample a set of network paths, and the neural language model (SkipGram) is applied to these network paths to obtain the node embedding.
\item \textbf{Node2vec}: Node2vec \cite{grover_node2vec:_2016} is also an unsupervised network embedding method which obtains the node embedding by using a biased random walk strategy to preserve the homophily and structure equivalence relationships in the networks.
\item \textbf{LINE}: LINE \cite{tang_line:_2015} obtains the network embedding by simultaneously optimizing the first-order and second-order proximities of the networks.
\item \textbf{SPARC}: SPARC \cite{zhou_sparc:_2018} is an imbalanced network embedding method. It obtains the imbalanced embedding in a mutually way, which can jointly predict the minority class and the neighbor context in the networks.
\item \textbf{DR-GCN}: DR-GCN \cite{shi_multi-class_2020} is also a GCN based imbalanced network embedding method, and it proposes to use conditioned adversarial training to enhance the separation of different classes. In addition, the distribution alignment training is applied to balance between the majority nodes and the minority nodes.
\item \textbf{RECT}: RECT \cite{wang_network_2020} is the state-of-the-art imbalanced network embedding method which is a variant of GNN. It obtains the imbalanced network embedding by learning the knowledge of class-semantic information in the networks.
\end{itemize}
\begin{figure*}[t]
\centerline{\includegraphics[width=0.85\textwidth]{Figure_3.pdf}}
\caption{The 2-dimensional imbalanced network layout with t-SNE on Cora dataset. The red circles represent the majority nodes of the original networks. The blue circles represent the minority nodes of the original networks. The yellow circles represent the minority nodes generated by SMOTE. The green circles represent the minority nodes generated by the proposed ImGAGN. The proposed ImGAGN is capable of discriminating between the real nodes (i.e., the blue circles and red circles) and the generated fake nodes (i.e., the green circles), and between the minority nodes (i.e., the blue nodes) and the majority nodes (i.e., the red nodes).}
\label{fig}
\end{figure*}
\subsubsection{\textbf{Parameters:}} All the codes we used are provided by authors. For GCN, following \cite{kipf_semi-supervised_2016}, the number of layers of the networks is set $K=2$. For GraphSAGE, we set $K=2, S_{1}=5, S_{2}=5$ according to the author suggesting. For GCN-SMOTE, the number of generated minority samples by SMOTE is equal to the difference between the majority and minority nodes of the training set. For DeepWalk, we adopt the default hyperparameters (i.e., window size $win=10$, walk length $len = 40$ and the number of walks $t=90$). For Node2vec, we optimize its hyperparameters by a grid search over $p,q \in \{0.25,0.50,1,2,4\}$. For LINE, the hyperparameter negative samples $ns=5$. For SPARC, the length of random walk sequences $\mu =10$. Moreover, the embedding dimension of unsupervised network embedding methods (i.e., DeepWalk, Node2vec and LINE) are set as $d=128$, and the logistic regression classifier is employed to evaluate the node embedding. For semi-supervised network embedding methods (i.e., GCN, GCN-SMOTE, GraphSAGE, SPARC and RECT), we use the outputs of their last hidden layer as the node embedding (the embedding dimension is also $128$).
The hyperparameters of our proposed method ImGAGN are set as follows. For generator, it consists of 3 fully connected layers with $100$ units in input layer and $200$ units in hidden layer. The number of units of output layer is equal to the difference between the majority class and minority class of the training set. $Tanh()$ is utilized as the activation function. For discriminator, it consists of the two-layer GCN followed by a softmax function, and ReLU \cite{glorot_deep_nodate} is utilized as the activation function. In addition, we perform generator and discriminator updates in $1:100$ ratio, and Adam SGD optimizer \cite{kingma_adam_2017} is utilized as the optimizer throughout the experiments.
\subsubsection{\textbf{Repeatability:}}
All the methods are run on a single machine with 14 CPU cores at 2.60GHZ and 2 Tesla P100 GPU with 32G memory using 1 thread.
\subsection{Imbalanced binary node classification (RQ1)}
To answer the RQ1 (i.e., how does ImGAGN perform compared with both the state-of-the-art balanced and imbalanced network embedding), we first conduct imbalanced binary node classification experiment on the four real-world network datasets. Three common classification metrics are used to evaluate the performance for all algorithms. Include: (1) recall, which measures the ratio of correctly classified nodes of all minority test nodes. (2) precision, which measures the ratio of correctly classified nodes of all predicted minority nodes (3) AUC scores, which measures model performance at all classification thresholds. We run experiments 10 times and use average scores for each metric. The experimental results are shown in Table 2.
From experimental results, in general, we can observe that:
\begin{itemize}
\item The proposed method ImGAGN substantially outperforms all comparison methods with respect to recall and AUC on all datasets and is comparable with RECT in terms of precision on Pubmed and DBLP datasets, nonetheless, ImGAGN improves significantly by 11.44\% and 8.73\% on these two datasets respectively in terms of recall which is usually more important than precision in many minority class classification problems, such as rare disease prediction \cite{schubach2017imbalance}. Thus, the overall performance could validate the effectiveness of the proposed method.
\item As expected, the imbalanced network embedding methods (i.e., GCN-SOMTE, SPARC, RECT, DR-GCN and ImGAGN) achieve better performance than the balanced network embedding methods (i.e., GCN, GraphSAGE, DeepWalk, Node2vec and LINE) in most case. It is reasonable since the former methods focus more on label learning of the minority class samples.
\item The GCN-SMOTE achieves better performance than original GCN, which shows that simple oversampling technique is capable of improving original GCN performance on imbalanced network data. The proposed method ImGAGN can also be thought of as an oversampling technique due to the operations of GraphGenerator. However, it obtains better performance than GCN-SMOTE, the improvements could be attributed to the GraphGenerator could well capture both implicit topological structure distribution and nodes' attribute distribution of the minority nodes.
\end{itemize}
\subsection{Network layout (RQ2)}
To answer the RQ2 (i.e., can ImGAGN learn the node embeddings such that the representation of minority class nodes can separate from the majority class nodes? ), we visualize the network layout in the embedding space, and we take Cora dataset for an example. Specifically, we firstly learn the nodes embedding in a 128-dimensional vector space for different network embedding methods, and then employ the t-SNE \cite{maaten2008visualizing} to map the 128-dimensional into the 2-dimensional space for visualization. The experimental results are shown in Figure 3. From the experimental results, in general, we can observe that:
\begin{itemize}
\item The proposed ImGAGN is well capable of discriminating between the real nodes (i.e., the blue circles and red circles) and the generated fake nodes (i.e., the green circles), and also between the minority nodes (i.e., the blue nodes) and the majority nodes(i.e., the red nodes), which validates ImGAGN is able to capture the latent representation of the minority nodes and majority nodes. We attribute such performance to the architecture of the GAN-based methods, that is, the loss functions of GraphGenerator (i.e., equation (4)) and discriminator (i.e., equation (6)) explicitly learn the discrimination between the real nodes and fake nodes, and also between minority nodes and majority nodes.
\item Generally speaking, the end-to-end semi-supervised network embedding methods (i.e., GCN, GCN-SMOTE, GraphSAGE, SPARC, RECT and ImGAGN) can better discriminate the majority and the minority classes than the unsupervised network embedding methods (i.e., DeepWalk, Node2vec and LINE). One explanation is that semi-supervised methods leverage both the nodes' features and label information to learn separable representation of the minority and majority nodes.
\end{itemize}
\subsection{Parameters sensitivity analysis (RQ3)}
To answer the RQ3 (i.e., How do different hyper-parameters influence the performance of ImGAGN?). We conduct the imbalanced node classification experiments on Cora and DBLP datasets and report the performance changes with respect to two crucial hyperparameters of the ImGAGN. One is $\lambda_{1}$ which is the ratio of the number of all training minority nodes (i.e., the original minority nodes and the generated minority nodes in training set) to the number of majority nodes, and we vary it from $0$ to $1$ with step size $0.1$. Another is $\lambda_{2}$ which is the number of discriminator training steps to for once generator training), and we vary it from $10$ to $100$ with step size $10$. The experimental results are shown in Figure 4 and Figure 5. From experimental results, in general, we can observe that:
\begin{itemize}
\item The imbalanced node classification performance, especially for recall, increases with the increase of training minority nodes ratio $\lambda_{1}$ and then tend to flat. One explanation is that when $\lambda_{1}$ is small, the training network is still classes imbalanced, which leads to bad classification performance. It is worth nothing that $\lambda_{1}=0$ degenerates our method to GCN, thus the performance with $\lambda_{1}=0$ also demonstrates that the proposed GraphGenerator could improve the GCN performance on imbalanced networks.
\item The performance increases with the increase of discriminator training steps $\lambda_{2}$ and then tend to flat. It is reasonable since the discriminator needs a certain number of training steps to learn the node embeddings.
\item Particular speaking, we found the proposed method ImGAGN could achieve high performance with $\lambda_{1} > 0.7$ and $\lambda_{2}>50$.
\end{itemize}
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth]{Figure_4.pdf}
\caption{Hyper-parameter sensitivity analysis of the training minority nodes ratio $\lambda_{1}$.}
\label{fig2}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth]{Figure_5.pdf}
\caption{Hyper-parameter sensitivity analysis of the discriminator training steps $\lambda_{2}$.}
\label{fig2}
\end{figure}
\section{Conclusion}
In this paper, to address the imbalanced network embedding problem, we proposed a semi-supervised network embedding method ImGAGN, which utilized a GraphGenerator to simulate both the minority class nodes’ attribute distribution and network topological structure distribution. It generated a set of synthetic minority nodes such that the number of nodes in different classes can be balanced. Then GCN discriminator was trained to discriminate between real nodes and fake nodes, and also between minority nodes and majority nodes. The extensive comparative studies, including the imbalanced node classification, network layouts and hyper-parameters sensitivity analysis, are conducted to validate the effectiveness of the proposed method. The empirical evaluation on four real-world datasets demonstrated that the proposed ImGAGN could outperform the state-of-the-art imbalanced network embedding algorithms on imbalanced node classification task in most cases with respect to recall, precision and AUC. In addition, the visualization results showed that the ImGAGN is capable of learning the node embedding such that the minority class nodes can separate from the majority class nodes.
\section{Acknowledgement}
This work is partially supported by the Science and Technology Innovation Committee Foundation of Shenzhen under the Grant No. JCYJ20200109141235597 and ZDSYS201703031748284, National Science Foundation of China under grant number 61761136008, Shenzhen Peacock Plan under Grant No. KQTD2016112514355531, Program for Guangdong Introducing Innovative and Entrepreneurial Teams under grant number 2017ZT07X386, ARC Discovery Project under the Grant No. DP190101985 and ARC Training Centre for Information Resilience under the Grant No. IC200100022.
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
Experimental control and data acquisition systems are widespread in many fields of scientific and industrial research where test and measurement systems need to be controlled and experimental data have to be gathered.
For the application of controlling experiments in the field of atomic, molecular and optical (AMO) physics digital pulses, analog, radio and microwave frequency signals need to be generated at well-defined times.
For instance, laser cooling and trapping of atomic gases down to ultralow temperatures typically require a temporal resolution of one microsecond. For this task, field programmable gate arrays (FPGAs) are very well suited. These can generate arbitrary digital pulses which can be used to program digital-to-analog converters (DAC), direct-digital synthesizers (DDS), and other devices with the requested timing resolution. As a result, FPGAs are already successfully employed in both commercial \cite{NI-FPGA} and open source \cite{Artiq2007} control systems.
Owing to their flexibility, FPGAs also find application for a wide range of different tasks, encompassing clock signal generation \cite{Ketterle2013}, DDS programming \cite{Meschede2015,Katori2015,Lu2017}, arbitrary waveform generation \cite{Hobson2019}, lock-in demodulation \cite{LIF2018}, high-speed data acquisition (DAQ) \cite{Tokamak2018}, digital feedback servo system \cite{Sias2018,Madison2018}. Moreover, FPGAs are increasingly used for the control of quantum systems and processors and as feedback devices for quantum measurements, and can be even used within cryogenic environments \cite{QuantumFeedback2013, cryogenic2016, cryogenic2017, SpinControl2020, QubiC2021}. Applications of FPGAs in space are becoming of growing interest \cite{FPGA_in_space}.
Despite of all of these applications, the development of a custom FPGA-based system is time consuming and commercial solutions tend to be expensive. Nonetheless, the advent of cheap, multi-purpose FPGA development boards targeted for hobbyists and students, offers a solution with low-cost and short development time, from which also experimental research can benefit thanks to the impressive capabilities of these boards.
Here we present a control system with a novel approach based on a commercial, low-cost system-on-a-chip (SoC) board, consisting of a central processing unit (CPU) which is tightly connected to an FPGA and to a set of hardware interfaces used to communicate with external devices.
A Linux operating system, executed on the CPU, gives the flexibility to use high-level programming languages, which can be quickly adapted to any specific request, such as interfacing with external devices like USB, Secure Digital (SD) memory card or Ethernet with no need of additional hardware or specifically designed micro-controllers.
Furthermore, the presence of an electrically isolated Gigabit Ethernet interface, allows fast data transfer and easy connection to remote locations.
All these features represent a clear advantage of FPGA-SoC systems with respect to previously realized FPGA-based solutions \cite{Prevedelli2020}, not only in terms of the superior data rates offered by the Ethernet interface, but also by the additional flexibility given by the presence of the easy programmable CPU and the fact that these are stand-alone systems which can be utilised independently on the hardware and software environment.
As a powerful simple application of such extended capabilities, here we demonstrate a novel scheme to auto-synchronize several boards using only two coaxial cables and the Ethernet communication. Without user interaction or dedicated real-time networking hardware \cite{WhiteRabbit}, the propagation delays of the signals among distant boards are measured by the boards and are corrected automatically with a residual timing error approaching 1\,ns.
The paper is organized as follows: First, we present the board architecture in Sec. \ref{sec:hardware}, and the developed software in Sec. \ref{sec:software}. We then present the measured performance and the auto-synchronization scheme in Sec. \ref{sec:results}, and discuss the results in Sec. \ref{sec:conclusions}.
\section{Hardware architecture} \label{sec:hardware}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{figures_pdf/figure_1}
\caption{a) Schematics of the control hardware. The experiment control sequence is sent from the control computer over an Ethernet network (yellow connections) to the FPGA-SoC boards (red). Each board, hosted in separated racks, where digital, analog and DDS devices can be freely inserted, is connect via buffer cards to a bus (gray ribbon cable). All FPGA-SoC boards are clocked (green connections) either by an external clock source or by the primary board clock signal. All the boards are synchronized via the clock and the trigger (blue connections) signals. b) Image of the FPGA-SoC board (red; back side visible), mounted on the buffer card (green; 100\,mm $\times$ 160\,mm Eurocard size). Backplane and power connectors are located on the right and bottom side. The trigger and clock I/Os are on the left-top and left-bottom side respectively. c) Image of the FPGA-SoC board (front side). The SoC is located in the center, the Ethernet connector is on the top side, and the two rows of pin sockets on the left and right side are used to connect with the buffer card. The external clock input is on one of the connectors on the bottom.}
\label{fig:setup}
\end{center}
\end{figure}
An overview of our setup is presented in Fig. \ref{fig:setup}a. A control computer generates the experiment control sequence (represented by a list of actions to be executed at a precise time) which is sent over Ethernet to one or several FPGA-SoC boards (distinguished by their IP address). Each FPGA-SoC board, hosted within a 19'' rack, drives via a buffer card a parallel bus over which digital and analog output devices and DDS are programmed at the specified time. These devices ultimately control the experiment and all physical parameters.
The system is compatible with the well-established architecture in use at LENS, consisting of
digital output devices with 16 TTL channels, analog output devices with two channels with 16-bit DACs with maximum $\pm$10\,V output, and DDS devices with two channels with up to 200\,MHz output frequency, which can be modulated in frequency and amplitude.
After the user has uploaded the control sequence, the experiment starts and the FPGA-SoC consecutively puts the data on the bus at the time defined in the time-stamp part of the control sequence. Once all samples are generated, the entire sequence can be repeated several times. For better timing accuracy, the clock source of the FPGA-SoC can be switched from the internal crystal oscillator to an externally provided clock signal.
The heart of our control system is the Cora-Z7 board from Digilent \cite{Cora}, which hosts the Zynq-7007S (Zynq-7010) FPGA-SoC from Xilinx with a single (dual) core CPU (ARM Cortex A9) clocked at 650\,MHz. This represents the smallest FPGA-SoC from the Xilinx Zynq-7000 series. The board is provided with 512\,MB of DDR3 SDRAM (16\,bits data clocked at 525\,MHz) with Gigabit Ethernet and USB host and device ports. The FPGA part is nearly the same for the two variants and is similar to the low-end Artix-7 FPGA series, aiming for low-cost, low-power consumption and less demanding applications. It should be noted, that, while we choose a particular FPGA-SoC board with Gigabit Ethernet to implement our control system, the system and the methods presented in this paper can be implemented with any other FPGA-SoC boards with similar performance. For example, the DE10-Nano from Terasinc Inc. is a possible alternative\cite{TerasicDE10Nano}.
A custom designed buffer card \cite{github} is used to buffer the FPGA-SoC board signals and to shift the voltage level from the internal 3.3\,V logic level to the 5\,V (TTL) level of the bus. The buffer card also provides the needed buffers for the clock and trigger line used for the synchronization of different boards, as described below. An image of the FPGA-SoC board mounted on the buffer card is shown in Fig. \ref{fig:setup}b, and in Fig. \ref{fig:setup}c an image of the FPGA-SoC board (front side) is shown.
\subsection{The FPGA logic}\label{sub:logic}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{figures_pdf/figure_2}
\caption{Simplified block diagram of the Zynq-7000 SoC with the user data flow on the chip (thick lines). The processing system (PS, green) with 32-bit dual-core CPU allows the server and driver to access periphery like Gigabit Ethernet (GigE) and DDR memory using high-level programming languages and Linux system services. The programmable logic (PL, yellow/orange/red) holds the custom implementation of hardware. Interfaces efficiently transfer data between the two parts. Two phase-locked loop (PLL) modules are generating three different clocks (clock out, bus clock and detection clock) from an external clock source or from the PL system clock (yellow; selected by the multiplexer ``MUX''). An overall dynamic phase shift $\phi_{ext}$ can be applied, as well as an individual phase shift $\phi_{det}$ on the detection clock. The user data is received over GigE by the server and is written via the driver into a memory region, reserved for direct-memory-access (DMA). The timing module reads the data via DMA from memory and uses one FIFO (TX) to buffer and transmit the data into the bus clocking region (orange). Data is read back into memory with the same DMA interface and another FIFO (RX). The auto-synchronization module generates a pulse on the trigger line and waits for its reception and a programmed number of cycles $N_w$ before it gives the start signal for the timing module to generate the data on the bus. In combination with a phase-shifted detection clock (red), the pulse round-trip number of cycles $N_{RT}$ between two boards can be measured. All control and status registers in the PL part can be accessed by the driver via the AXI Light interface and are transmitted with clock-domain-crossing (CDC \cite{CummingsCDC2008}) modules between the different clocking regions. The DMA and timing modules send interrupts (IRQ) to notify the driver of important events.}
\label{fig:logic}
\end{center}
\end{figure}
Here we give an overview on the logic used in the FPGA to generate the experiment control data on the bus and all the signals necessary for the synchronization of several boards. A simplified block diagram is shown in Fig. \ref{fig:logic}.
The board is basically composed of two parts: the processing system (PS, top, green), consisting of a CPU on which a Linux operating system is running, and the programmable logic (PL, bottom, yellow), where our custom hardware is implemented. The two parts of the FPGA-SoC are tightly bound via interfaces and buses, enabling mutual data exchange at high speed. In such a way, the two main tasks of the board are effectively separated among the two independent parts of the FPGA-SoC system itself. While the processing system handles the communication via Ethernet with an external control computer, the logic part produces the signals on the bus. The driver mediates between both parts and coordinates the access to the external memory. The source code for programming the FPGA is written in Verilog. It is synthesized and implemented with the Vivado 2017.4 software from Xilinx running on Ubuntu 18.04 LTS, and is available online \cite{github}. Detailed information on the FPGA resources used for this application is reported in Tab. \ref{tab:resources} in Appendix \ref{sub:resources}.
In brief, we use one general purpose I/O (GPIO) port for the reading and writing of memory mapped registers (via AMBA AXI-4 Lite interface \cite{AMBA-AXI}), and one high performance (HP) port to efficiently transfer the experiment control sequence from the memory to the PL part and vice-versa (using direct memory access DMA \cite{AXI-DMA} via an AXI stream bus). The clock frequency for the PL part, CPU and the DDR memory are set to their default values, corresponding to 50\,MHz, 650\,MHz and 525\,MHz respectively.
The experiment control sequence (represented by thick lines in Fig. \ref{fig:logic}) is sent via Ethernet from the control computer to a TCP/IP server application running on the CPU. The server application interacts with a Linux kernel driver module \cite{github}, which writes the data into DDR memory and programs the FPGA registers using the AXI Light bus. The data are transferred via DMA from the memory into a transmit (TX) first-in-first-out (FIFO) buffer \cite{CummingsFIFO2002a,CummingsFIFO2002b} which holds a maximum of 8192 samples of 128\,bits each.
The FIFO serves to buffer gaps in the DMA data transmission, and allows efficient transfer of data between regions using different clocks (clock domains).
In addition, we have implemented a receive DMA channel (RX), which can be used, for example, to read data from an analog input device that sends data on the bus.
In our case, the experiment control sequence consists of 64\,bits per sample: 32\,bits are used for the time-stamp, 7 address bits select which device on the bus to be updated, and 16 device specific bits define the new state of the device \footnote{\label{note96bits}An optional extended version uses 12 instead of 8 bytes per sample. This allows to have two independent buses driven by a single FPGA-SoC board with a modified buffer card.}.
The time-stamp defines at which time the bus should be updated with the specific data and address of the corresponding device. After the bus has been updated, a pseudo-clock pulse (strobe) is generated by the FPGA on another pin of the bus, to initiate the state change of the selected device \footnote{The strobe signal is generated by the FPGA. For $\Gamma_{sample}$ = 1\,MHz it is a 500\,ns long pulse starting 240\,ns after the bus has been updated. The bus clock frequency must be at least twice the bus output rate to generate the strobe signal.}. The time-stamp is defined in units of 1/$\Gamma_{sample}$ with $\Gamma_{sample}$ being the output sampling rate of the bus, typically set to 1\,MHz or 10\,MHz.
The timing module is responsible to output the data on the bus. It first takes out one 64\,bits-wide sample from the 128\,bits of the TX FIFO, and it compares the time-stamp with an internal counter running at $\Gamma_{sample}$. When they are equal, the module outputs the 16+7 data and address bits, and it generates the previously mentioned strobe signal.
%
The timing module internally uses a dedicated 50\,MHz bus clock, which can be either the PL system clock (i.e. the internal oscillator of the FPGA-SoC board), or it can be generated from an external clock signal using a phase locked loop (PLL) of the FPGA-SoC. In the latter case, the frequency allowed for the external clock signal ranges from a minimum value of 10\,MHz, limited by the PLL, to a maximum of 300\,MHz, limited by the input buffer on the buffer card. A second PLL is used as a software controlled multiplexer (MUX) to switch between the two clock sources\footnote{Cascading two PLL's is not advised, but in our case, we need both for dynamic phase shifting. In addition, this allows to use an external clock input pin in a different clocking region which would be otherwise inaccessible.}. Both PLLs enable to dynamically change the phase of the generated clock signals. The auto-synchronization module, discussed in Sec. \ref{sub:auto-sync}, is using these signals to synchronize several boards. The timing module can also trigger the output of the experimental sequence, which alternatively can be started by an external hardware trigger or via software.
Finally, both the DMA TX/RX channels and the timing module communicate with the driver via interrupts. The DMA channels generate interrupts when buffers need to be updated. The timing module generates one interrupt when the experimental control sequence has been completed. Further interrupts are generated at a configurable frequency, typically 16\,Hz, and are used to update the board status in the control software.
\subsection{Auto-synchronization}\label{sub:auto-sync}
In order to synchronize several FPGA-SoC boards, all boards need to start the experimental control sequence simultaneously and they need to use the same clock source to execute each command at the same time. The common clock can be either generated by one board, or
provided externally. In both cases, a suitable amplification and distribution system to all boards is needed, which might introduce unknown phase shifts. Additionally, a starting (trigger) signal needs to be distributed from one board to all the others, and can accumulate an unknown delay. As discussed in the following, our scheme takes into account and corrects for both these effects.
To compensate the delay on the start trigger signal, we adopt a scalable scheme, where one trigger line is connected with high impedance to all participating boards, see Fig. \ref{fig:auto-sync}a. The trigger line is a coaxial cable with 50\,$\Omega$ termination on both ends to avoid unwanted reflections. One board, called the primary board, receives the start signal from the control computer (or from an external hardware trigger), and generates a pulse in the trigger line which is detected by the other ``secondary'' boards. In order to compensate for the pulse propagation time between the boards, the propagation time is automatically measured in advance, such that each board can delay its execution accordingly and all boards can start at the same time.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.95\columnwidth]{figures_pdf/figure_3}
\caption{Triggering and auto-synchronization scheme for multiple boards. a) In the simplest configuration all boards are connected with a common clock (period $T$) provided by the primary board and daisy-chained from one board to the next using splitters. Additionally, a common trigger coaxial cable directly connects all boards and is terminated by $50\,\Omega$. The primary board generates a pulse in the trigger cable which all secondary boards detect with individual delay.
The primary and secondary boards wait until all secondary boards have received the trigger pulse and start generating output simultaneously. %
The delays between the primary and secondary boards for the clock $\tau_c^i$ and the trigger $\tau_p^i$ are indicated ($i\ \in\ 0 \ldots N$), with $N$ the number of secondary boards.
b) The trigger delay $\tau_p^i$ of each secondary board $i$ is measured during the auto-synchronization by determining the round-trip time $t_{RT}^i$ of the pulse (orange)
from the difference of the number of cycles from the generation of the pulse ($N_{gen}^i$, blue) and its detection ($N_{det}^i$, red). The time correction $\Delta t_{RT}^i < T$ is obtained by repeating the measurement and detecting the reflected pulse with a phase-shifted detection clock with increasing detection phase (black, seven phases shown) with respect to the bus clock (green) which is used to generate the pulse. At the phase $\phi_{-}^{p,i}$
the measured $N_{RT}^i$ reduces by one cycle and $\Delta t_{RT}^i$ is obtained. For board $i$ the trigger delay is calculated from $\tau_p^i = t_{RT}^i/2$. For the determination of $\Delta \tau_c^i$ a similar measurement is done on each secondary board where $\Delta t_s^i$, $N_{bus}^i$ and $\phi_{-}^{s,i}$ replace the roles of $\Delta t_{RT}^i$, $N_{gen}^i$ and $\phi_{-}^{p,i}$ in the figure. The clock delay $\Delta \tau_c^i$ at board $i$ is calculated from Eq. \eqref{eq:clock-delay-simple}. See text and Appendix \ref{sub:auto-sync-model} for more details\cite{note_cdc
and figures \ref{fig:auto-sync-data}a and \ref{fig:auto-sync-det} for example detection signal for varying detection phase.}
\label{fig:auto-sync}
\end{center}
\end{figure}
To measure the propagation delay, the primary board instructs via Ethernet one of the secondary boards to introduce a short circuit in the trigger line using a bipolar or a field-effect transistor. Then the primary board generates a pulse in the trigger line, and it measures the round-trip time $t_{RT}^i = N_{RT}^i\,T + \Delta t_{RT}^i$ needed by the pulse to propagate to the secondary board $i$, be reflected at the short circuit, and travel back (see Fig. \ref{fig:auto-sync}b).
Here $N_{RT}^i = N_{det}^i - N_{gen}^i$ is the number of cycles between the generation ($N_{gen}^i$, blue) and the detection ($N_{det}^i$, red) of the pulse, and $\Delta t_{RT}^i < T$ is a fraction of the period $T$ of the bus clock of the primary board.
While $N_{RT}^i$ can be measured directly, $\Delta t_{RT}^i$ cannot. This limits the resolution to the period $T$, which is 20\,ns for the chosen 50\,MHz bus clock frequency, and would not be satisfactory for bus output rates above 10\,MHz.
To measure the total delay with higher accuracy, the reflected pulse is sampled with a phase shifted replica (detection clock) of the bus clock signal. A train of trigger pulses is generated, and the phase shift of the detection clock is varied between pulses.
For a linear increase of the detection clock phase, at:
\begin{equation}\label{eq:time-corr-prim}
\phi_{-}^{p,i} = \Delta t_{RT}^i \frac{\ensuremath{360^{\circ}}}{T}\ ,
\end{equation}
the measured $N_{RT}^i$ reduces by one. This change in $N_{RT}^i$ is detected, and $\Delta t_{RT}^i$ can be obtained \footnote{The actual algorithm to find the phase jump is similar to the Bisection method of finding the root of a function.}.
In principle, this method would allow one to achieve a time resolution of about 20\,ps, given the 0.3$^{\circ}$ phase resolution of the PLL at the used clock frequency. However, noise in the generation and detection of the pulse actually limits the resolution to larger values.
This measurement is repeated for each secondary board $i = 0 \ldots N$.
With the measured round-trip time $t_{RT}^i$, the propagation time of the pulse from the primary board to the $i$-th secondary board is calculated as:
\begin{equation}\label{eq:propagation-time-simple}
\tau_p^i = t_{RT}^i/2\ .
\end{equation}
It is important to notice that in this simplified treatment we neglect all additional (but constant) delays, both internal to the FPGA and due to the electronics needed for the generation and detection of the pulse. Details of the full model accounting for these additional delays are given in Appendix \ref{sub:auto-sync-model}.
In order to achieve a perfect synchronization among all boards, the measurement of $\tau_p^i$ for each board discussed above is not sufficient, since the clocks of the secondary boards must be corrected for the delays $\tau_c^i$
introduced along the clock distribution line (see Fig. \ref{fig:auto-sync}a). In this case however, one needs to know only the introduced clock delay $\Delta \tau_c^i = \tau_c^i \% T$, where $\%$ is the modulus.
To this end, a second set of measurements is carried out, where the primary board generates a train of pulses similarly to the previous scheme, but the measurement is now taken on the secondary boards. Since the pulses do not need to be reflected, all the secondary boards can measure the respective clock delay simultaneously. Each secondary board determines the time $\Delta t_s^i$ between the arrival of the pulse and the previous rising bus clock edge, local to the secondary board.
Similarly to the delay $\Delta t_{RT}^i$, here the quantity $\Delta t_s^i$ is obtained by detecting the arrival of the pulse with both the detection and the bus clock simultaneously, giving $N_{det}^i$ and $N_{bus}^i$ (blue dashed line in Fig. \ref{fig:auto-sync}b), respectively. The difference between the two signals $N_{det}^i - N_{bus}^i$ is monitored for a reduction of one cycle at the phase:
\begin{equation}\label{eq:time-corr-sec}
\phi_{-}^{s,i} = \Delta t_s^i \frac{\ensuremath{360^{\circ}}}{T}\ ,
\end{equation}
and $\Delta t_s^i$ can be obtained. At the secondary board location, the calculated pulse delay with respect to the primary bus clock is $\Delta t_{RT}^i/2$ and the difference to the observed delay $\Delta t_s^i$ gives the unknown clock delay:
\begin{equation}\label{eq:clock-delay-simple}
\Delta \tau_c^i = \Delta t_{RT}^i/2 -\Delta t_s^i\ .
\end{equation}
Once $\Delta t_{RT}^i$ and $\Delta t_s^i$ are determined for each secondary board $i = 0 \ldots N$, the external clock PLL phases $\phi_{ext}^i$ of each secondary board can be set to $\phi_{ext}^i = -\Delta \tau_c^i \frac{\ensuremath{360^{\circ}}}{T}$. In this way, the clocks of all secondary boards are synchronized with that of the primary one and the auto-synchronization measurement is completed and all parameters are set.
In order to simultaneously generate data on all boards, the primary board sends a pulse in the trigger line. It then waits until all secondary boards have detected the trigger pulse, i.e. it waits the largest propagation time $\tau_p^i$. Each secondary board $i$ waits $\tau_p^i$ less time than the primary board. After these waiting times, all boards synchronously start generating output of data on their bus.
While we refer the reader to Appendix \ref{sub:auto-sync-model} for more details, we emphasize that our auto-synchronization scheme allows for the synchronization of many boards on time scales of order of nanoseconds with a relatively simple scheme and few external components. A first experimental demonstration of this scheme together with measurements of the residual synchronization timing error are presented in Sec. \ref{sec:results}.
\section{Software} \label{sec:software}
In this section we summarize the software implementation on the PS/CPU part of the SoC, on which a Linux operating system is running\footnote{Petalinux 2017.4 from Xilinx which is built on Linux kernel version 4.9 and is compiled on Ubuntu LTS 18.04.}. This is a fully featured operating system which provides system services and interfaces to external devices, and that can be configured for our specific needs. The PL part appears for the operating system like an external device, and our device driver can communicate with it via registers \cite{github}.
\subsection{Control computer software}\label{sub:control}
Many research laboratories, including ours, typically employ either Labview or LabWindows/CVI \cite{NI_Labview_CVI} as user application programs. While our setup is currently adapted to work with this software, we emphasize that any other user application can be easily implemented on our hardware, provided that the data need to be sent via Ethernet to our TCP/IP server running on the FPGA-SoC. No additional driver nor hardware is required, and no constraints on the operating system are given for the control computer. For example, the freely-available, Python-based control software ``labscript suite'' \cite{Starkey2013} might be a viable alternative to the above mentioned commercial solutions. We provide the necessary files in Ref. \onlinecite{github} to use our FPGA-SoC board together with the suite.
In our specific case, we upgraded an existing control system based on a digital I/O card \cite{DIO64} installed on the experiment control computer, driving the bus via a 2\,m long cable and a buffer card. The FPGA-SoC system replaces completely the former system, maintaining the compatibility with previous hardware and software. For this, a new Windows dynamic link library (DLL) has been written, which communicates via Ethernet with the FPGA-SoC while keeping the same functions of the previous I/O card.
\subsection{TCP/IP server and Linux device driver}\label{sub:driver}
We have designed a simple TCP/IP server application, running on the FPGA-SoC, which receives commands and the user data from the control computer, and which communicates with our device driver that mediates with the two FPGA-Soc parts, see Fig. \ref{fig:logic}.
Our server application can control, via the device driver, the FPGA PL part, write the user data into reserved DMA (coherent) memory, and receive status information from it. The driver allows a user application to read back data from the PL part, wait for interrupts or for the end of the sequence.
The driver maintains the ring buffers for the DMA transfer, and responds to the corresponding interrupts.
We have reserved 128\,MiB of memory for coherent DMA transfer. This size corresponds to $10^7$ samples and 10 seconds of contiguous data output at $\Gamma_{sample}$ = 1\,MHz. However, most applications typically do not require such a large number of samples and dense output of data. If needed, data could be uploaded via Ethernet during the experimental run as well. The reserved size
is sufficiently large to store all user data directly into coherent memory, which keeps the server and driver simple, and it avoids additional copying for repeated runs.
A timer interrupt, generated by the PL part, and transmitted by the driver, allows the server application to send status information at regular intervals to the control computer.
\subsection{Startup script}\label{sub:linux-init}
When the board is powered up, a bootloader reads from a SD card the binary data to program the PL part and to load the required Linux image into memory, and to start the operating system. After this is completed, our startup script reads a configuration file from the SD card which contains the IP address and other information, with which it configures the Linux system and launches our TCP/IP server application. The server may either initiate the auto-synchronization procedure on startup, or wait for instructions from the control computer. A startup script and a text configuration file are used to change the configuration of the board without the need of recompiling the binary code from the sources.
\section{Measurements and results}\label{sec:results}
In this section we present and discuss measurements done on the FPGA-SoC board. For these measurements, specific code running on the FPGA-SoC system has been written, and the data has been acquired directly on the board and stored on a micro-SD card \footnote{As permanent storage medium the board uses a micro-SD (Secure Digital) card which primarily contains the Linux boot loader and boot image but can contain additional files and folders and can be used as a hard drive. The Linux image is unpacked by the bootloader in a RAM drive, but if needed it can also be expanded into a partition of the SD card. Additionally, a USB flash drive can be attached to the board for external storage.} for further analysis. Except for the verification of the synchronization error, no external measurement was needed. All the data presented in the paper is available in Ref. \onlinecite{Zenodo}.
In the first part, Sec. \ref{sub:DMA}, measurements of the DMA transmission rates are shown, defining how fast data can be transmitted from the external memory into the PL part and back. This represents a direct measure of the maximum sampling rate at which the board can contiguously output and input data.
In the second part, Sec. \ref{sub:uploading}, we present measurements on the data uploading rates over Gigabit Ethernet for both the Cora-Z7-10 and Cora-Z7-07S boards.
This measurement confirms that Gigabit Ethernet is a good choice for experiments where a fast cycle time is required.
In the last part, Sec. \ref{sub:auto-sync-data}, we present first measurements of the proposed auto-synchronization scheme outlined in Sec. \ref{sub:auto-sync}, tested on a simple two-board configuration.
An additional measurement presented in Appendix \ref{sub:start-stop} demonstrates the start- and stop trigger option in cycling mode.
\subsection{DMA transmission rates}\label{sub:DMA}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{figures_pdf/figure_4}
\caption{Measurement of DMA data transmission rates of the Cora-Z7-10 board as a function of number of samples $N$. Each measurement point is the mean value of at least 20 measurements and the error bar corresponds to the standard deviation. The curves are fits to the data as explained in Appendix \ref{sub:DMA-fit}. The vertical dotted line at 8192 samples corresponds to the TX and RX FIFO buffer size. The horizontal dotted line at 600\,MB/s corresponds to 1 sample/cyle for the 50\,MHz PL clock frequency and the horizontal red dotted line is the fitted $\Gamma_{DMA}$ = 341(1)\,MB/s for large number of samples.}
\label{fig:DMA}
\end{center}
\end{figure}
In order to measure the DMA data transmission rates of the FPGA-SoC board we have temporarily added a module in the PL part which allows one to transmit data without delay in a ``loop-back'' configuration between the TX and the RX FIFO buffers (see Fig. \ref{fig:logic}), and to measure the time interval required to transmit a certain number of samples. From the measured time $t$ and the number of samples $N$ we calculate the average data rate $\Gamma$ in MB/s using: $\Gamma = \beta\,N/t$, with $\beta = 12$ bytes per sample for this measurement.
In particular, we measure three distinct rates, shown in Fig. \ref{fig:DMA} for the Cora-Z7-10 board, as a function of the number of samples $N$: the transmission rate from the memory to the PL part (TX DMA, red circles), the transmission rate from the PL part to the memory (RX DMA, orange squares) and the transmission rate through the RX FIFO (green diamonds).
Each experimental point (error bar) shown in the figure represents the mean value (standard deviation) of at least 20 repeated measurements for each $N$.
The data are well fitted to a simple model (solid curves in Fig. \ref{fig:DMA}) that has one delay and two rates as free parameters. For details about the fitting function, and the fit results, we refer the reader to Appendix \ref{sub:DMA-fit} and Tab. \ref{tab:DMA-fit} therein.
For the measurement of the TX transmission rate (red circles in Fig. \ref{fig:DMA}) we measure the time interval from the first sample received out of the TX FIFO until the $N$-th sample is received. The first four samples are transmitted with the maximum possible rate of one sample per cycle, i.e. $\Gamma_{max} = \beta \times f_{PL}$ = 600\,MB/s (horizontal black dotted line) for the PL clock frequency of $f_{PL} =$ 50\,MHz. This is because the TX FIFO already contains three to four samples when the measurement starts (in agreement with the simulated latency of the used FIFO). As $N$ is increased, the rate reduces rapidly until it reaches a constant rate $\Gamma_{DMA}$ (horizontal red dotted line), corresponding to the transmission rate from memory to the PL part. We remark, that this characterization does not allow to measure a possible delay between the start of the DMA transmission, initiated by the CPU and the arrival of the first sample.
The second measurement (orange squares in Fig. \ref{fig:DMA}) shows the RX transmission rate obtained from the time interval between the first sample written into the RX FIFO and the RX DMA interrupt \footnote{The interrupts are generated in the PL part and are thus directly accessible during the transmission rate measurement without involving the CPU.}, which indicates that all N samples have been transmitted from the PL part to the external memory. This second rate increases for increasing $N$, from very small to the same $\Gamma_{DMA}$ as observed for the TX measurement. This initial increase is consistent with a constant delay of 202(8) PL cycles, required for the RX DMA channel to start or finish the transmission. This delay is larger than expected \footnote{On the TX DMA side we observe a delay of about 30 cycles between the arrival of the last data out of the FIFO and the TX interrupt.}, and it points to a significant latency in the RX channel. Nonetheless, the large RX FIFO can easily compensate for such a latency.
The third measurement, shown in Fig. \ref{fig:DMA} as green diamonds, was taken simultaneously with the RX transmission rate, and it shows the data rate through the RX FIFO: namely, the rate obtained from the time N samples need to pass through the RX FIFO during active RX transmission. As long as the RX FIFO is not full, one sample per cycle is transmitted, corresponding to $\Gamma_{max}$. When the RX FIFO becomes full with $N_{FIFO} = 8192$ samples (dotted vertical line in Fig. \ref{fig:DMA}), the rate reduces to the RX and TX data transmission rate $\Gamma_{DMA}$. Since the RX FIFO is simultaneously loaded with $\Gamma_{max}$, and unloaded with $\Gamma_{DMA}$, we expect this rate to drop once the number of transmitted samples reaches $N_{FIFO} \frac{\Gamma_{max}}{\Gamma_{max}-\Gamma_{DMA}} \approx 19\times 10^3$ samples, a value close to the observed one of $20(1)\times10^3$ samples.
All three measurements give for large number of samples a consistent DMA transmission rate of $\Gamma_{DMA} = 341(1) MB/s$ (averaged over all measurements).
This rate deviates with the specified rates from Xilinx \cite{AXI-DMA} for the default settings. In particular, the TX rate is lower while the RX rate is higher than specified. However, their measured sum is 684(2)\,MB/s, which is only 2\% lower than the value expected from the specification of 700\,MB/s.
Although the exact reason for this discrepancy is not clear (the ratio between the TX and RX rates can be adjusted \cite{Zynq-SDK-performance, SoC-performance}), the observed overall performance allows us to conclude that our DMA transmission rates are indeed close to the maximum possible ones for a single HP port. Finally, from the measured DMA transmission rate we can also directly deduce the maximum contiguous bus data rate of $\Gamma_{DMA}/\beta$ = 30 - 40\,MHz~\footnote{The measured $\Gamma_{DMA}$ corresponds to a maximum $\Gamma_{sample}$ of 42\,MHz (28\,MHz) for the 8 (12) bytes per sample versions. The given rates apply independently for data output and input on the bus and for simultaneous output and input (if the bus supports).}.
We note that, the FPGA-SoC has 4 HP ports, and in our design there should be enough free resources to use at least an additional one to increase the DMA rate even further, if higher bus rates are needed. Short ``bursts'' of data output (input) of up to 8192 samples at higher frequencies are already possible with the present setup as long as there is sufficient time before the ``burst'' to fill (empty) the TX (RX) FIFO and the rate afterwards is slow enough to prevent the TX (RX) FIFO from becoming empty (full). Although not shown here, we have performed the same measurement for the Cora-Z7-07S board, finding no significant deviation from the results presented in Fig. \ref{fig:DMA}.
\subsection{Ethernet uploading rates}\label{sub:uploading}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{figures_pdf/figure_5}\\
\caption{a) Measured rates for uploading and writing to reserved DMA memory (solid symbols) and uploading only (open symbols) as a function of number of samples $N$ for the Cora-Z7-10 (blue circles) and Cora-C7-07S board (red squares). The horizontal dotted line indicates the theoretical maximum rate of 118.7\,MB/s for Gigabit Ethernet \cite{GigE} and the vertical dotted line indicates the size of the receive buffer of the server. The numbers are the measured uploading and writing rates for $10^7$ samples. Each data point is the mean of at least 15 measurements and the error represents the standard deviation. The dotted curves are fits with Eq. \eqref{eq:fit-model} with a delay and single rate and the fit results are summarized in Tab. \ref{tab:DMA-fit}. b) Same data as in panel a but time for uploading or uploading and writing to memory is shown. Numbers give the \chg{fitted} time needed for uploading and writing to DMA memory for 4 samples and \chg{$10^7$} samples for the Cora-Z7-10 (blue) and Cora-Z7-07S (orange) boards.}
\label{fig:uploading}
\end{center}
\end{figure}
The uploading rate from the control computer to the FPGA-SoC board over Gigabit Ethernet is another measure of the performance of our system. It can be a limitation for experiments where short cycle times are need, like experiments with optical tweezers \cite{Lukin2016} or with ions \cite{Ott2017}.
Fig. \ref{fig:uploading}a shows the uploading rate measured for the Cora-Z7-10 (solid blue circle) and Cora-Z7-07S (solid orange square) board. This measurement includes the total time of uploading and writing into reserved DMA memory. For each board the fastest strategy is used depending if a dual-core CPU is present (Cora-Z7-10) or only a single-core CPU (Cora-Z7-07S): for the dual-core CPU the server uses one thread to receive the uploaded data and a second thread to write the data into reserved DMA memory in parallel. For the single-core CPU it is fastest to immediately write the uploaded data into reserved DMA memory using a single thread\footnote{The change in the rate between using a single or two threads on both boards is only about 10\%.}. Fig. \ref{fig:uploading}b shows the corresponding times for the same data as in Fig. \ref{fig:uploading}a.
The rates are calculated from $\Gamma = N \beta / ( t_{tot} - t_{ACK} - t_{RT}^{net}/2 )$ where $N$ is the number of transmitted samples and $\beta$ = 12 bytes per samples used for the measurement. The time $t_{tot}$ is when uploading and writing to memory is finished, and $t_{ACK}$ is the time when the server acknowledged to receive the data from the user application. The network round-trip time $t_{RT}^{net}$ is obtained during each individual measurement as the time from the acknowledge of the server ($t_{ACK}$) until the arrival of the first data at the server. We take half of $t_{RT}^{net}$ under the assumption that sending and receiving involves the same delays, which is not necessarily the case. For each datapoint we have taken at least 15 measurements and plot the mean value and standard deviation (error bar).
For small number of samples the observed uploading rate is small. This can be interpreted as a fixed delay (of order of a few 100$\,\mu$s, see Fig. \ref{fig:uploading}b), which the user application or the server needs to start sending or receiving the data. For increasing number of samples, this delay becomes less important and the rate reaches a peak of about 70 - 80\,MB/s at $32 \times 10^3$ samples (vertical dotted line) and decreases for number of samples beyond this. At $10^7$ samples the uploading and writing rate is 56.5(3)\,MB/s (47.2(4)\,MB/s) for the Cora-Z7-10 (Cora-Z7-07S) board, which corresponds to a time of \chg{2.13(1)\,s (2.54(2)\,s)}. This time is even faster than the typical calculation time the user application needs (about 7\,s with labscript-suite) to generate this number of samples.
The peak in the rate is correlated with the receive buffer size (512\,kiB) of the server. If chosen too small the decrease in the rate at higher $N$ becomes much worse. This indicates that the overhead in handling large lists of small buffers can become significant. In this respect the Cora-Z7-10 board performs slightly better than the Cora-Z7-07S board, which is limited by a single-core CPU.
For comparison, we present another measurement where only data are uploaded, but no writing to the reserved DMA memory is done. The resulting rates for the Cora-Z7-10 (open blue circle) and Cora-Z7-07S (open orange square) board are shown in Fig. \ref{fig:uploading}a and b. For the calculation of the rate, $t_{tot}$ is now the time until all data is uploaded without writing to reserved DMA memory. For the Cora-Z7-10 board the peak uploading rate reaches about 110\,MB/s which is very close to the theoretical maximum of 118.7\,MB/s for Gigabit Ethernet \cite{GigE}. The Cora-Z7-07S board is with about 90\,MB/s only slightly slower. In this measurement the CPU is still copying data into temporary buffers which explains the difference of the boards, and the observed decrease of the rate after the peak.
With Eq. \eqref{eq:fit-model} in Appendix \ref{sub:DMA-fit} we fit the measurements with a delay time and a single transmission rate (dotted curves in Fig. \ref{fig:uploading}). We use the standard deviation of each data point to get more weight on the large number of samples with less noise. See Tab. \ref{tab:DMA-fit} for the fit results. The numbers in the figure are the fitted rates and times for both boards when uploading and writing 10.5 $\times 10^6$ samples to reserved DMA memory.
The observed fast uploading and writing rates confirm that the FPGA-SoC board is indeed the right choice for applications where fast cycle times are requested.
\subsection{Auto-synchronization}\label{sub:auto-sync-data}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{figures_pdf/figure_6}\\
\caption{Auto-synchronization result for two boards at different trigger cable lengths. a) Round-trip cycle time $N_{RT}$ for the reflected pulse leading edge vs. detector phase shows jumps of one cycle at specific phases ($\phi_{+}$ and $\phi_{-}$, see Sec. \ref{sub:auto-sync} and Appendix \ref{sub:auto-sync-model} for details). Data is shown for selected cable lengths.
b) Pulse round-trip time $t_{RT}$ calculated with Eq. \eqref{eq:propagation-time} for the trailing edge of the pulse for 12 cable lengths. The slope of the linear fit gives a propagation delay per unit cable length of $\frac{d\tau_p}{d L}$ = 4.9(4)\,ns/m, when averaged over leading and trailing edges of the pulse. c) Synchronization error as a function of cable length. Each point and error bar is the mean and standard deviation of five repetitions with external clock phase 0, 90, 180 and 270$^{\circ}$. The red shaded area represents the 68\% confidence interval of the average error over all data giving (-0.5 $\pm$ 1.3)\,ns.
The insert shows all signal traces of the primary (blue) and secondary board (red) used to measure the synchronization error.}
\label{fig:auto-sync-data}
\end{center}
\end{figure}
Here we present the first realization of the auto-synchronization scheme proposed in Sec. \ref{sub:auto-sync}. In particular, first tests have been done utilizing two boards connected with different trigger cable lengths and using different external clock phases.
Without loss of generality, we present the synchronization of the two boards that are directly connected with the trigger line, terminated with 50\,$\Omega$ on the primary board side and switchable on the secondary board side from 50\,$\Omega$ to high impedance to reflect the pulse.
In the following we omit the index $i = 0$ since here only one secondary board is used.
For details on the theoretical analysis and the measurement of the secondary board external clock PLL phase we refer the reader to Appendix \ref{sub:auto-sync-model} and \ref{sub:auto-sync-phase}.
On the primary board we measure the round-trip cycle time $N_{RT}$ of the reflected pulse, and the phase $\phi_{-}^p$ at which $N_{RT}$ is reduced by one, see Fig. \ref{fig:auto-sync-data}a for different lengths of the trigger coaxial cable\footnote{For cable lengths $< 3$\,m the actual setup cannot detect the round-trip time since the reflected pulse is too close to the generated one. However, this situation is automatically detected and with the proposed scheme and further technical improvements shorter cables should be detectable.}. Combining both measured values of $N_{RT}$ and $\phi_{-}^p$ we obtain, from Eq. \eqref{eq:propagation-time} in Appendix \ref{sub:auto-sync-model}, the round-trip time $t_{RT}$ shown in Fig. \ref{fig:auto-sync-data}b.
From a linear fit to the data (green line) we obtain the propagation delay per unit of cable length $L$ of $\frac{d\tau_p}{d L}$ = 4.9(4)\,ns/m, when averaged over leading and trailing edges of the pulse. This value is consistent with the expected one \cite{RG58}.
Based on a similar measurement protocol\footnote{For the measurement on the secondary board the pulse is not reflected to avoid interference of the incoming with the reflected pulse. However, we have not observed a difference in the measurement result.}, the secondary board determines the phase $\phi_-^{s}$ of the negative jump in $N_{det}-N_{bus}$ for the received pulse. The local clock of the second board is locked to the external clock provided by the primary one, where a short (ca. 20\,cm long) cable is employed to ensure no additional phase shifts. To simulate different delays $\Delta \tau_c$ of the external clock, four different auto-synchronization measurements are performed, where the external clock PLL phase of the secondary board is set to 0, 90, 180 or 270$^{\circ}$, corresponding to $\Delta \tau_c$ = 0, 5, 10 or 15\,ns respectively.
The resulting synchronization error is verified in a final measurement for each cable length and $\Delta \tau_c$ after the auto-synchronization is finished, see Fig. \ref{fig:auto-sync-data}c. For this measurement,
the resulting phase $\phi_{ext}$, obtained from Eq. \eqref{eq:clock-phase} in Appendix \ref{sub:auto-sync-model}, is added to the previously set external PLL clock phase $\Delta \tau_c \frac{\ensuremath{360^{\circ}}}{T}$, which, for perfect synchronization, should be compensated by $\phi_{ext}$.
Then the primary board generates a trigger pulse and waits $N_w^{prim} = \tau_p \mathbin{/\mkern-6mu/} T$ cycles (see Eq. \eqref{eq:detection-time} and \eqref{eq:waiting-time} in Appendix \ref{sub:auto-sync-model}; the symbol $\mathbin{/\mkern-6mu/}$ represents integer division), before it starts generating data on the bus. The secondary board starts generating data on the bus as soon as the trigger signal is detected.
The synchronization error corresponds to the difference between the times at which secondary and primary boards start generating data on their own buses. The corresponding traces are recorded with an oscilloscope, see the inset of Fig. \ref{fig:auto-sync-data}, and are fitted with a sigmoid function to obtain the synchronization error. See Appendix \ref{sub:trace-fit} for further details.
In Fig. \ref{fig:auto-sync-data}c each data point (error bar) represents the mean (standard deviation) of the synchronization error, measured at least five times for each of the four external clock phases ($\Delta \tau_c$). Averaging over all cable lengths, we obtain a synchronization error of (-0.5 $\pm$ 1.3)\,ns (red shaded area in Fig. \ref{fig:auto-sync-data}c) which is much smaller than the 25\,ns time resolution for the maximum possible bus output rate of 40\,MHz of the board.
Finally we remark that, although the basic principle of our auto-synchronization scheme is very simple, being based on a round-trip time measurement, the details can be involved. Developing such a scheme on a FPGA-only platform is feasible, but it might be challenging and time-consuming. In turn, our FPGA-SoC board allows one to implement a simple pulse generation and detection in hardware, but to analyze the data and calculate the ideal settings to minimize the error, via the CPU, by software. In this way, the system could be quickly developed, errors corrected and the formulas implemented in software with no need to change the hardware every time. We believe that, the auto-synchronization is not only a useful feature, but it is also a perfect example of the flexibility which the FPGA-SoC approach offers.
\section{conclusions and outlook}\label{sec:conclusions}
In conclusion, we have successfully implemented a versatile experimental control system based on a commercial, low-cost, and stand-alone FPGA-SoC board. %
We have demonstrated that the board can sustain bus output and input rates of up to 40\,MHz and we have
shown how the board can automatically synchronize with a timing error approaching 1\,ns.
Furthermore, we have proven the extreme flexibility, easy Ethernet connectivity, and computational power of the FPGA-SoC system, showing several examples in which the operating system, running on the board itself, is used not only to control the FPGA hardware, but also for data acquisition and analysis. Finally, we stress that no specific device driver or proprietary software, or operating system is needed to use our device, and that the whole source code to program the FPGA-SoC is freely available \cite{github}.
Although not discussed in the present work, our system can be easily extended to include the control of additional devices through the on-board USB host controller \cite{USBTMC}, or via adapter with the older GPIB standard \cite{GPIB}, widespread in many laboratories, or to directly read data with analog-to-digital converters (ADC).
We also emphasize that our design is stand-alone and lightweight, and the power consumption of less than 2\,W, makes it compatible for the operation in remote locations, and even for experiments in space \cite{ClockSpace2018,BECinspace2020,InterferometerSpace2021}.
We believe that the auto-synchronization feature, devised and implemented in this work, will also help several experimental setups on ground with growing complexity: for instance, setups which must bridge large distances to challenge relativity \cite{Hensen2015}, to detect gravitational waves with large-scale atom interferometers \cite{Kasevich2013,Bouyer2018}, and to measure difference of gravitational red-shift between two separated atomic lattice clocks \cite{Katori2020}.
Finally, our architecture, thanks to the rich features and flexibility offered by the new FPGA-SoC board, may find application in various research fields, extending well beyond our original purpose of controlling AMO physics experiments.
\begin{acknowledgments}
We thank Jacopo Catani for fruitful discussions, borrowing equipment and careful reading of the manuscript, Roberto Concas and Fabio Corti for machining and soldering a prototype buffer card, Giacomo Mazzamuto for help with github, and all members of the Quantum Gases Group at LENS, in particular Leonardo Fallani and Daniele Tusi and the Yb team for testing the boards in their experiment. This work was supported by the ERC through grant No. 637738 PoLiChroM and by the Italian MIUR through the FARE grant No. R168HMHFYM P-HELiCS. N.P. acknowledges support from European Research Council, Grant No. 772126 (TICTOCGRAV).
\end{acknowledgments}
The authors declare that they have no competing interests.
\section*{Data Availability Statement}
\begin{center}
\renewcommand\arraystretch{1.2}
\begin{tabular}{| >{\raggedright\arraybackslash}p{0.3\linewidth} | >{\raggedright\arraybackslash}p{0.65\linewidth} |}
\hline
\textbf{AVAILABILITY OF DATA} & \textbf{STATEMENT OF DATA AVAILABILITY}\\
\hline
Data openly available in a public repository that issues datasets with DOIs
&
The data that support the findings of this study are openly available at https://doi.org/10.5281/zenodo.4893285
\\\hline
\end{tabular}
\end{center}
|
\section*{Introduction}
This paper is the second, after \cite{dkss:1}, part of our project devoted to a systematic
development of the theory of perverse schobers which are (enhanced, triangulated) categorical
analogs of perverse sheaves, see \cite {kapranov-schechtman:schobers} for a general discussion. It
can be read independently and provides results which may be of interest in their own way. Its
position in the general program can be explained as follows.
Let $(X,N)$ be a {\em stratified surface}, i.e., a topological surface $X$, possibly with boundary,
together with a finite set $N$ of interior points. It gives rise to the abelian category
${\on{Perv}}(X,N)$ of perverse sheaves on $X$ with singularities in $N$. In \cite{dkss:1}, we
established a ``purely abelian'' description of ${\on{Perv}}(X,N)$ in terms of so-called {\em Milnor
sheaves}. The latter are certain functors ${ \EuScript F}$ which associate a vector space ${ \EuScript F}(A,A')$ to any
pair (Milnor disk) consisting of a closed disk $A\subset X$ meeting $N$ in at most one point and a
nonempty union of closed arcs $A'\subset {\partial} A$. Explicitly, for a perverse sheaf $F\in{\on{Perv}}(X,N)$
the corresponding Milnor sheaf has the form ${ \EuScript F}(A,A') = \mathbb{H}^0(A,A'; F)$, the $0$th relative
hypercohomology (the only nontrivial hypercohomology!) with coefficients in $F$.
This description, being purely abelian, is suitable for categorification. In this paper we study the
features of such categorification for the case when a disk $A$ is fixed, and only $A'\subset {\partial} A$
varies. In this case the categorical analog of ${ \EuScript F}$ reduces to data of (enhanced triangulated)
categories ${ \EuScript F}_n$, one for each $n\geq 0$ (corresponding to the case when $A'$ consists of $(n+1)$
arcs) which are connected by a {\em paracyclic structure} , i.e., by an action of the category
$\Lambda_{\infty}$ covering the cyclic category $\Lambda$ of Connes, see \cite{dkss:1}. The
categorical analog of $F$ itself is a perverse schober on the disk $A$, and ${ \EuScript F}_n$ can be thought
as the Fukaya category of $A$ with coefficients in the schober and with supports (``stops'') at
$A'$.
As proposed in \cite {kapranov-schechtman:schobers}, a schober on a disk should be described by
a spherical adjunction, i.e., an adjoint pair of functors
\[
\xymatrix{
{\EuScript C} \ar@<.4ex>[r]^F&\Di\ar@<.4ex>[l]^G
}
\]
such that the cones of the unit and the counit of the adjunction are equivalences of categories,
see \cite{AL:spherical} and compare with the description of perverse sheaves on the disk
\cite{beil:gluing, GGM}. Our main discovery is that from the point of view of a spherical functor
$F$, the collection of categories ${ \EuScript F}_n$ turns out to be nothing but $S_\bullet(F)$, the {\em
relative Waldhausen $S$-construction} of $F$, cf. \cite{waldhausen}. More precisely,
$S_\bullet(F)$ can be defined for any exact functor $F$ of stable ${\infty}$-categories and is a
simplicial object (in the category of stable ${\infty}$-categories). We prove that for a spherical
functor $F$, this simplicial structure is naturally extended to a paracyclic one, matching the
intuition related to a disk on a surface with a bunch of arcs in the boundary. The action of the
center of $\Lambda_{\infty}$ corresponds, geometrically, to a categorical analog of the monodromy of a
perverse sheaf.
Thus, from the point of view of perverse schobers on stratified surfaces, the structures studied in
this paper constitute the local data from which a general schober is glued. From the standpoint of
spherical functors however, these structures are absolutely remarkable features which have no
conceivable explanation outside of the analogy with perverse sheaves. In a paper to follow, we plan
to glue the structures defined here (associated naturally to disks), into global data on a
stratified surface.
In this paper we adopt the language of stable ${\infty}$-categories as our preferred way of enhancing
triangulated categories. The main reason for this is the following.
From the very beginning, the theory of spherical functors has been burdened by technical
difficulties. Namely, as pointed out in \cite{AL:spherical}, the context of classical
triangulated categories is not sufficient since we need to take the cones of unit and counit natural
transformations. We need therefore some enhancement of the triangulated structure. The most
immediate such enhancement (and the most frequently used in this context) is that of
pre-triangulated dg-categories \cite{BK:enhanced}.
However, the concept of adjoint functors in the dg-enhanced triangulated context was always
considered rather subtle as it appears to require a tower of explicit higher coherence data, see
\cite{AL:Fourier, AL:spherical, AL:bar}. For example, for adjoint dg-functors $F$ and $G$ the
``identification'' of the complexes ${\operatorname{Hom}}^\bullet(F(x), y)$ and ${\operatorname{Hom}}^\bullet(x, G(y))$ should have
the form of a {\em quasi}-isomorphism and one immediately hits the questions like ``which
direction?" or ``what are the relations among such quasi-isomorphisms?" etc.
The approach with ${\infty}$-categories allows one to avoid answering such questions explicitly while
retaining a sufficient pool of possible answers. This is done via encoding ${\infty}$-functors between
${\infty}$-categories in terms of (co)Cartesian fibrations over the interval $\Delta^1$, see
\cite[\S5.2]{lurie:htt} and a discussion below in \S
\ref{subset:stab-triang}\ref{par:cocart-implicit} More precisely, an ${\infty}$-functor $F: {\EuScript C} \to\Di$ of
${\infty}$-categories gives rise to a coCartesian fibration $\Gamma(F)\to\Delta^1$ (covariant
Grothendieck construction). The functor $F$ {\em has a right adjoint}, if this fibration is also
Cartesian and so {\em can be identified} with the contravariant Grothendieck construction of some
${\infty}$-functor $G: \Di\to{\EuScript C}$. Such $G$ ({\em an adjoint}, one among many) is defined uniquely up to a
contractible space of choices but we can delay making this choice until really necessary (or avoid
it altogether) working with $\Gamma(F)$ instead. This ``print-on-demand'' approach saves a lot of
effort in the end.
For this reason we spend some time explaining, once for all future occasions, the details of the
concept of ${\infty}$-categorical adjunctions and its compatibility with other approaches to adjunctions,
esp. in the dg-setting. The outcome is that our approach (i.e., that of \cite{lurie:htt}) covers all
the other ones used previously, and so various earlier results retain their value in our context.
However, while it is known that (pre-triangulated) dg-categories give rise to (stable)
${\infty}$-categories \cite{lurie:ha, {faonte:nerve}}, the simplifications mentioned above do not seem to
simplify further by specializing to the dg-case.
Another advantage of ${\infty}$-categories is that various auxiliary totalities (e.g. the collection of
all spherical adjunctions) readily make sense as ${\infty}$-categories but in the dg-setting they do not
form dg-categories, so some form of ${\infty}$-categories is needed anyway.
In fact, the Grothendieck construction of a functor plays quite a prominent role in our study. In
addition to the above use of this concept, the procedure of {\em gluing a semiorthogonal
decomposition (SOD) from a gluing functor}, introduced in the dg-context in \cite{kuznetsov-lunts},
amounts to nothing but (the category of sections of) the Grothendieck construction. Further, the
first level $S_1(F)$ of the relative Waldhausen S-construction of a functor $F$, is again nothing
but the category of sections of the Grothendieck construction (i.e., the category with an SOD glued
via $F$).
\iffalse
\vskip .5cm
................
\vskip .5cm
General question: should we consider triangulated categories as analogs of single vector spaces or of complexes of vector spaces?
Our point of view: as analogs of single vector spaces! Viewing them as analogs of complexes leads to
motivic cohomology/motivic invariants point of view which loses more information
\vskip .5cm
................
\vskip .5cm
The theory of spherical adjunctions can be seen as a deep conceptualization of the
number $2$, comparable to the ``internal'' incarnations of $2$ as subobject
classifiers in topos theory. Many examples of spherical functors involve one category
being approximately ``one half'' of the other. The simplest sphere $S^0$ is a $2$-element
discrete set.
\fi
\vskip .2cm
The paper consists of three chapters.
In Chapter 1 we develop the theory of spherical functors and spherical adjunctions in the framework
of stable ${\infty}$-categories (the existing treatments in the literature use dg-categories). \S
\ref{subset:stab-triang} is a general reminder on (stable) ${\infty}$-categories. We pay particular
attention to the ``implicit'' way of defining ${\infty}$-functors in terms of (co)Cartesian fibrations
over $\Delta^1$. In \S \ref{subsec:adj-oo} we discuss, following \cite[\S 5.2]{lurie:htt}, the
${\infty}$-categorical approach to adjunctions. In \S \ref{subsec:comparison}, we give a comparison of
this approach with other approaches that are used in the literature: the one involving Fourier-Mukai
kernels in various dg-contexts \cite{AL:Fourier, AL:bar} and the one involving higher coherence data
\cite{{riehl-verity:2-cat}, {riehl-verity:yoneda}, {riehl-verity}}. In \ref{subsec:sphad} we
define spherical adjunctions as those for which the cones (fibers) of the unit and the counit are
equivalences. We give a few elementary examples related, in one way or another, to the concept of
sheaves on spheres. They can be used to motivate a more sophisticated example in \S
\ref{sec:massey} which provides a family of spherical functors, one for each $n\geq 0$. The
importance of this family is that it can be seen as the representing object of the spherical
S-construction discussed later in \S \ref {subsec:rel-S-sph}.
\vskip .2cm
Chapter \ref{sec:SOD} is devoted to the theory of semi-orthogonal decompositions (SOD) in the
framework of stable ${\infty}$-categories. While the concept of an SOD itself makes sense in the context
of classical triangulated categories, and we recall it in \S \ref {subsec:SOD-triang}, deeper
aspects such as gluing SODs via gluing functors, require an enhancement. In \S \ref{sec:semi} we
develop these aspects when the enhancement is ${\infty}$-categorical. In line with our emphasis on
Cartesian and coCartesian fibrations as encoding data of ${\infty}$-functors, we introduce the concepts
of Cartesian and coCartesian SOD's which appear very naturally in the ${\infty}$-categorical context.
They lead to not one but two concepts of gluing functors which, when they exist, go in opposite
directions and it is convenient to distinguish them. In \S \ref{subsec:admissibility}, we relate
the concept of admissibilty of an SOD (traditionally defined as the condition that various
orthogonals form SODs as well) with the existence of (${\infty}$-categorical) adjoints of the gluing
functors. In \S \ref{subsec:mutation}, we study the mutation functor (identification of the left and
right orthogonal of an admissible subcategory) and upgrade it to a ``coordinate change functor'', a
categorical analog of a $2\times 2$ matrix, see \S \ref{subsec:mutation} \ref{par:mut-change} In \S
\ref{subsec:4-period} we adapt to the stable ${\infty}$-categorical context the interpretation, due
originally to Halpern-Leistner and Shipman \cite{halpern-shipman}, of spherical functors as gluing
functors of $4$-periodic SODs.
\vskip .2cm
In Chapter \ref{sec:spher-S} we study $S_\bullet(F)$, the Waldhausen S-construction of a spherical
functor $F$ and prove that it has a natural structure a paracyclic $2$-Segal object in ${\EuScript Cat}_{\infty}$
with a certain prescribed pattern of adjunctions between face and degeneracy functors. In \S
\ref{subsec:rel-S} we recall the paracyclic category $\Lambda_{\infty}$ and the original construction of
$S_\bullet(F)$ by Waldhausen \cite{waldhausen}. In \S \ref{subsec:rel-S-sph} we interpret
$S_\bullet(F)$ in the particular case of spherical $F$ in terms of the corresponding $4$-periodic
SOD. This definition (we call it the {\em spherical S-construction}) makes the paracyclic symmetry
manifest. Its equivalence with the original $S_\bullet(F)$, proved in Theorem \ref{thm:sphericalS},
is one of the forms of the main result of this paper. A less technical formulation, Theorem
\ref{thm:main}, says that $S_\bullet(F)$ admits a natural paracyclic structure. In the final \S
\ref{subsec:sph-admis} we complement this result with additional properties holding for any
$S_\bullet(F)$ with $F$ spherical or not: the {\em $2$-Segal property} \cite{DK:HSS} and {\em
admissibility} (a certain pattern of adjunctions between faces and degeneracies)
\cite{dyckerhoff:DK}. We conjecture that all these properties taken together, characterize spherical
functors. More precisely, Conjecture \ref{conj:sphad=adm2Segpar} says that the ${\infty}$-category of
spherical adjunctions is equivalent to that of paracyclic $2$-Segal admissible stable
${\infty}$-categories.
\vskip .2cm
We are grateful to A. Bodzenta and A. Bondal for useful discussions.
T.D. acknowledges the support
of the VolkswagenStiftung through the Lichtenberg Professorship Programme. The research of T.D. is
further supported by the Deutsche Forschungsgemeinschaft under Germany's Excellence Strategy -- EXC
2121 ``Quantum Universe'' -- 390833306.
The research of M.K. was supported by World Premier
International Research Center Initiative (WPI Initiative), MEXT, Japan. The research of Y.S. was
partially supported by Munson-Simu award of KSU.
\section{Spherical adjunctions}
\subsection{(Stable) ${\infty}$-categories and triangulated
categories}\label{subset:stab-triang}
As in \cite{dkss:1}, we use the language of ${\infty}$-categories \cite{lurie:htt}. In this section, we
review some basic notions for convenience of the reader and to fix the notation.
\paragraph{Basic reminder.}
We denote by $ {\operatorname{Set}}_\Delta$ the category of simplicial sets. A simplicial set $X$ will be sometimes denoted
$X_\bullet$ or $(X_n)_{n\geq 0}$ with $X_n$ being the set of $n$-simplices. We denote by $X^{\operatorname{op}}$
the opposite simplicial set to $X$.
If $\mathbf k$ is a field, we denote by ${\operatorname{dgVect}}_\mathbf k$
the category of cochain complexes (dg-vector spaces) over $\mathbf k$. A cochain complex $V$ will be sometimes
denoted $V^\bullet$ to indicate the grading.
\vskip .2cm
By an ${\infty}$-{\em category} we mean a simplicial set ${\EuScript C}$ satisfying the weak Kan condition
(i.e., an $({\infty},1)$-category, in a more precise terminology).
An ${\infty}$-functor between ${\infty}$-categories ${\EuScript C},\Di$ is simply a morphism of simplicial sets.
Similarly, the {\em functor ${\infty}$-category} $\operatorname{Fun}({\EuScript C},\Di)$ is simply the simplicial mapping space between ${\EuScript C}$ and $\Di$
as simplicial sets.
\vskip .2cm
By an {\em object} of ${\EuScript C}$ we mean a vertex, i.e.,
an element of ${\EuScript C}_0$. By a {\em morphism}
of ${\EuScript C}$ we mean an edge, i.e., an element of ${\EuScript C}_1$, and we use the notation
${\operatorname{Hom}}_{\EuScript C}(x,y)$ for the set of edges of ${\EuScript C}$ beginning at $x$ and ending at $y$.
An ordinary small category $C$ can be considered as an ${\infty}$-category by forming the nerve of $C$, denoted $\operatorname{N} C$.
In the opposite direction, any ${\infty}$-category ${\EuScript C}$ gives an ordinary category $\operatorname{h}\!{\EuScript C}$ with
the same objects known as the {\em homotopy category} of ${\EuScript C}$. Its morphisms are certain
equivalence classes of morphisms of ${\EuScript C}$. A morphism in ${\EuScript C}$ is called an {\em equivalence},
if its image in $\operatorname{h}\!{\EuScript C}$ is an isomorphism.
\vskip .2cm
For any two objects $x,y$ of an ${\infty}$-category ${\EuScript C}$ we have a Kan complex $\operatorname{Map}_{\EuScript C}(x,y)\in {\operatorname{Set}}_\Delta$
(the {\em mapping space})
whose set of vertices is ${\operatorname{Hom}}_{\EuScript C}(x,y)$. In fact, there are several different versions of this spaces,
all homotopy equivalent to each other. They are functorial with respect to morphisms, i.e., we have
morphisms of simplicial sets (with ${\operatorname{Hom}}_{\EuScript C}$ being a discrete simplicial set).
\begin{equation}\label{eq:map-functor}
{\operatorname{Hom}}_{\EuScript C}(y,z) \times \operatorname{Map}_{\EuScript C} (x,y) \longrightarrow \operatorname{Map}_{\EuScript C}(x,z), \quad \operatorname{Map}_{\EuScript C}(y,z) \times{\operatorname{Hom}}_{\EuScript C}(x,y) \longrightarrow
\operatorname{Map}_{\EuScript C}(x,z).
\end{equation}
However, there is no preferred composition law
\[
\operatorname{Map}_{\EuScript C}(y,z) \times \operatorname{Map}_{\EuScript C} (x,y) \longrightarrow \operatorname{Map}_{\EuScript C}(x,z)
\]
for
such mapping spaces although one can define it up to contractible space of choices, see \cite[1.2.2]{lurie:htt}.
The mapping spaces can be used to define ${\infty}$-categorical initial and final objects (e.g., $x$ is initial, if each $\operatorname{Map}_{\EuScript C}(x,y)$ is contractible) and therefore other types of universal constructions such as
limits $\varprojlim$ and colimits $\varinjlim$.
\vskip .2cm
In the opposite direction, let $\mathcal{C}$ be a category enriched in Kan complexes. Then we can
associate to $\mathcal{C}$ an ${\infty}$-category $N^\Delta \mathcal{C}$ by means of the {\em simplicial nerve
construction}, see \cite[1.1.5]{lurie:htt} and references therein.
\vskip .2cm
By a {\em dg-category} we will mean a category enriched in ${\operatorname{dgVect}}_\mathbf k$. Given a dg-category $\mathcal {A}$,
one can associate to it an ${\infty}$-category $N^{\operatorname{dg}} \mathcal {A}$ by means of the {\em dg-nerve construction}
(the simplicial set formed by ``Sugawara simplices'' \cite[App.2]{HS}) \cite[1.3.1]{lurie:ha} \cite{faonte:nerve}. The homotopy category $\operatorname{h}\! N^{\operatorname{dg}} A$ is identified with the $0$th
cohomology category $H^0 \mathcal {A}$.
More generally, any $A_{\infty}$-category gives an ${\infty}$-category by means of the {\em $A_{\infty}$-nerve construction}
\cite{faonte:nerve}. In this way various Fukaya $A_{\infty}$-categories appearing in symplectic geometry
\cite{seidel} can be included in the ${\infty}$-categorical formalism.
\paragraph{The ${\infty}$-category of ${\infty}$-categories} \label{par:oo-2-level}
The simplicial nerve construction provides the fundamental object: ${\EuScript Cat}_{\infty}$, the {\em ${\infty}$-category of
all (small) ${\infty}$-categories}. It is defined as the simplicial nerve of the simplicially enriched category
${\EuScript Cat}_{\infty}^\Delta$ whose objects are ${\infty}$-categories as above and $\operatorname{Map}({\EuScript C},\Di)$ is
the largest Kan subcomplex in $\operatorname{Fun}({\EuScript C},\Di)$, see \cite[Def. 2.0.0.2]{lurie:htt}.
\vskip .2cm
This approach is very convenient and will be used later in this paper to speak about
${\infty}$-categories formed by related data such as, ${\infty}$-functors, adjunctions of ${\infty}$-categories
etc. It will be eventually used to speak about ${\infty}$-categories of perverse schobers.
\vskip .2cm
Sometimes, a more general approach is necessary. Indeed, the term ``${\infty}$-category'' adopted in
\cite{lurie:htt} and here, stands for ``$({\infty},1)$-category'', with the ``$1$'' signifying that all
$p$-morphisms with $p>1$ are weakly invertible. So this concept does not include all classical, even
strict, $2$-categories. In particular, it cannot describe ${\EuScript Cat}$, the classical {\em $2$-category
of all usual categories} (objects: categories, morphisms: functors, $2$-morphisms: natural
transformations). One can form the corresponding part of ${\EuScript Cat}_{\infty}$ as defined above, but it
captures only {\em invertible} natural transformations.
\vskip .2cm
The more general structure possessed by ``the category of all ${\infty}$-categories''
(and capturing that on ${\EuScript Cat}$) is that of $({\infty},2)$-category. Let us give
a simplified (partially strict) definition, following \cite{riehl-verity}.
See \cite{lurie:ha} for a ``fully lax'' definition.
\begin{defi}\label{def:oo-2}
An {\em $({\infty},2)$-category} is a simplicially enriched category $\mathcal{C}$ such that
for any objects $c,d\in \mathcal{C}$ the simplicial set ${\operatorname{Hom}}_\mathcal{C}(c,d)$ is an ${\infty}$-category.
\end{defi}
This is analogous to defining a strict classical $2$-category as a category
enriched in the category of categories. In fact, any strict $2$-category $C$
can be considered as an $({\infty},2)$-category $\operatorname{N}^{\operatorname{Hom}} C$ by taking the nerves of all the
categories ${\operatorname{Hom}}_C(c,d)$.
\vskip .2cm
Further, we have the $({\infty},2)$-category ${\EuScript Cat}_{\infty}^{(2)}$ of all
${\infty}$-categories. Its objects are all ${\infty}$-categories, and the
simplicial Hom-sets are the functor ${\infty}$-categories $\operatorname{Fun}({\EuScript C},\Di)$.
This $({\infty},2)$-category refines the ${\infty}$-category ${\EuScript Cat}_{\infty}$ defined above.
\paragraph{Stable ${\infty}$-categories.}\label{par:stable-oo}
We refer to \cite{lurie:stable, lurie:ha} for a general treatment. It is convenient to take,
as the definition, the following reformulation
\cite[Prop. 1.1.3.4]{lurie:ha}:
\begin{defi}
An ${\infty}$-category ${\EuScript C}$ is called {\em stable}, if:
\begin{itemize}
\item[(S1)] ${\EuScript C}$ has finite limits and colimits, in particular, an initial and and a final object.
\item[(S2)] ${\EuScript C}$ is pointed, i.e., initial and final objects coincide. Such objects are called
{\em zero objects} and denoted $0$.
\item [(S3)] A commutative square (i.e., an ${\infty}$-functor $\Delta^1\times\Delta^1\to{\EuScript C}$)
\[
\xymatrix{
X\ar[d] \ar[r]&Y\ar[d]
\\
T\ar[r]& Z
}
\]
is Cartesian (i.e., $X$ is equivalent to $\varprojlim\, \{ T\to Z \leftarrow Y\}$) if an only if it is coCartesian
(i.e., $\varinjlim\, \{ T\leftarrow X\to Y\}$ is equivalent to $Z$). Such squares are called
{\em biCartesian}.
\end{itemize}
\end{defi}
If ${\EuScript C}$ is stable, then the category $\operatorname{h}\!{\EuScript C}$ is triangulated in the classical sense of
Verdier, see \cite[Thm. 3.11]{lurie:stable} \cite[Thm. 1.1.2.14]{lurie:ha}.
Distinguished triangles in $\operatorname{h}\!{\EuScript C}$ come
from biCartesian squares as in (S3) with $T=0$. Such squares are called {\em triangles} in ${\EuScript C}$.
Let $f: {\EuScript C}\to\Di$ be an ${\infty}$-functor of stable ${\infty}$-categories. It is known \cite[Prop. 1.1.4.1]{lurie:ha}
that $f$ being {\em left exact} (i.e., commuting with finite limits) is equivalent to $f$ being
{\em right exact} (i.e., commuting with finite colimits). If this is the case, we call $f$ {\em exact}.
If $f$ is exact, then the induced functor $\operatorname{h}\! f: \operatorname{h}\!{\EuScript C}\to\operatorname{h}\!\Di$ is triangulated,
i.e., preserves the shift and the class of distinguished triangles.
In this way stable ${\infty}$-categories provide ``enhancements" of classical triangulated categories.
\vskip .2cm
Unlike classical triangulated categories, in a stable ${\infty}$-category ${\EuScript C}$ we can speak about
(homotopy) canonical cones of morphisms by fitting them into triangles in ${\EuScript C}$.
In fact, it is convenient to separate this concept into
two: the {\em fiber} of a morphism $f: X\to Y$, denoted $ \operatorname{Fib}(f)$ and the {\em cofiber}, or
{\em cone}, denoted ${\operatorname{Cof}}(f)$ as well as $\operatorname{Cone}(f)$. They are defined as
\begin{equation}\label{eq:(co)fiber}
\begin{gathered}
\operatorname{Fib}(f) \,=\, \varprojlim \, \{ X\buildrel f\over\to Y \leftarrow 0\},
\quad \operatorname{Cone}(f)={\operatorname{Cof}}(f) \,=\, \varinjlim \, \{ 0\leftarrow X \buildrel f\over\to Y\}.
\end{gathered}
\end{equation}
As any limits and colimits in an ${\infty}$-category, the fiber and cofiber are defined uniquely up
to a contractible choice.
The suspension and desuspension functors in ${\EuScript C}$ (corresponding to the shift
by $\pm 1$ in the triangulated category $\operatorname{h}\!{\EuScript C}$) are defined as
\[
X[1] \,=\,{\operatorname{Cof}}\{X\to 0\}, \quad X[-1] \,=\, \operatorname{Fib}\{ 0\to X\}.
\]
We further have an equivalence $ \operatorname{Fib}(f) \simeq {\operatorname{Cof}} (f)[-1]$.
Another advantage of stable ${\infty}$-categories is the following:
\begin{prop}
(a) If $ {\EuScript A}, {\EuScript C}$ are ${\infty}$-categories and ${\EuScript C}$ is stable, then $\operatorname{Fun}( {\EuScript A}, {\EuScript C})$ is stable.
\vskip .2cm
(b) Let $ {\EuScript A}$ and ${\EuScript C}$ be stable ${\infty}$-categories. The fiber and cofiber of any morphism
of exact ${\infty}$-functors $ {\EuScript A}\to {\EuScript C}$ is exact.
\end{prop}
\noindent{\sl Proof:} (a) is a particular case of \cite[Prop. 4.1]{lurie:stable} (in fact, one can take
for $ {\EuScript A}$ any simplicial set, not necessarily an ${\infty}$-category). To show (b), note that
projective limit is a left exact operation, so $ \operatorname{Fib}(f)$, being a projective limit of left
exact functors, is left exact and hence is exact since $ {\EuScript A}$ and ${\EuScript C}$ are stable.
Similarly, ${\operatorname{Cof}}(f)$, being an inductive limit of right exact functors, is right exact hence exact. \qed
The {\em ${\infty}$-category of stable ${\infty}$-categories} ${\EuScript S}t$ is defined as the simplicial subset
in ${\EuScript Cat}_{\infty}$ formed by simplices whose vertices correspond to stable ${\infty}$-categories and
edges to exact ${\infty}$-functors.
\vskip .2cm
Another type of enhancements used in the literature is that of {\em dg-enhancements},
given by pre-triangulated dg-categories \cite{BK:enhanced}. It can be included
into the ${\infty}$-categorical picture: as shown in \cite{faonte:nerve}, for a pre-triangulated
dg-category $C$, its dg-nerve $N^{\operatorname{dg}} C$ is a stable ${\infty}$-category, which gives the same
triangulated category $\operatorname{h}\! N^{\operatorname{dg}} C = H^0 C$. Note, however, an interesting difference
with
the ${\infty}$-categorical case:
for any dg-functor $f: C\to D$ between pre-triangulated dg-categories, the induced functor
$H^0 f: H^0 C\to H^0 D$ is triangulated, there is no need to impose any further conditions on $f$. This is
because the concept of an distinguished triangle, at the dg-level, can be encoded by
a dg-functor from some particular dg-category with three objects, see \cite[\S I.3]{seidel}
\paragraph{Cartesian and coCartesian fibrations.}
Let $C$ be an ordinary category. According to the classical theory of Grothendieck \cite[Ch.3]{fantechi},
fibered categories $X\to C$ over $C$ provide an alternative, implicit way
of encoding contravariant pseudo-functors $F: C\to{\EuScript Cat}$
from $C$ to the $2$-category of categories. Covariant
pseudo-functors are similarly encoded by cofibered categories.
An ${\infty}$-categorical generalization of this point of view is given by Lurie's theory of
Cartesian and coCartesian fibrations of ${\infty}$-categories
\cite[\S 2.4]{lurie:htt}. Let us recall
the necessary concepts at the level of generality we need.
\begin{defi}\label{def:cart-edge}
(a) Let $p: {\EuScript X}\to{\EuScript C}$ be an ${\infty}$-functor of ${\infty}$-categories.
An edge (morphism) $f: x\to y$ in ${\EuScript X}$ is called {\em $p$-Cartesian}, if the natural map
\[
{\EuScript X}_{/f} \longrightarrow {\EuScript X}_{/y} \times_{{\EuScript C}_{/p(y)}} {\EuScript C}_{/p(f)}
\]
is a trivial Kan fibration of simplicial sets.
\vskip .2cm
(b) An edge $f$ is called $p$-{\em coCartesian}, if it corresponds to a Cartesian edge for
the opposite ${\infty}$-functor $p^{\operatorname{op}}: {\EuScript X}^{\operatorname{op}}\to{\EuScript C}^{\operatorname{op}}$.
\end{defi}
Here ${\EuScript X}_{/y}$ is the over(${\infty}$-)category whose objects are objects in ${\EuScript X}$ over $y$, and similarly
for ${\EuScript C}_{/p(y)}$. The notation ${\EuScript X}_{/f}$ means the over(${\infty}$-)category whose objects are morphisms
in ${\EuScript X}$ over $f$, and similarly for ${\EuScript C}_{/p(f)}$, see \cite[Def. 1.2.9.2]{lurie:htt} for the general concept.
\vskip .2cm
Explicitly, $f$ being $p$-coCartesian can be expressed by a condition involving
undercategories instead of overcategories
and the source rather than the target of $f$.
Notationally, for an edge $f: a \to b$ in ${\EuScript X}$, we will write $a \overset{!}{\to} b$ if $f$ is coCartesian and $a \overset{*}{\to} b$ if $f$ is
Cartesian.
\begin{defi}(\cite{lurie:htt}, Def. 2.4.2.1 + Prop. 2.3.1.5) \label{def:cart-fib}
Let ${\EuScript C}$ be (the nerve of) an ordinary category and $p: {\EuScript X}\to{\EuScript C}$ be an ${\infty}$-functor of ${\infty}$-categories.
\begin{itemize}
\item[(a)]
We say that $p$ is a {\em Cartesian fibration}, if for every edge $g: c\to d$ of ${\EuScript C}$ and any vertex (object)
$\widetilde d$ of ${\EuScript X}$ with $p(\widetilde d)=d$, there is a $p$-Cartesian edge $\widetilde g: \widetilde c\to\widetilde d$ with
$p(\widetilde g)=g$.
\item[(b)] We say that $p$ is a {\em coCartesian fibration}, if $p^{\operatorname{op}}: {\EuScript X}^{\operatorname{op}}\to {\EuScript C}^{\operatorname{op}}$ is a Cartesian fibration.
\end{itemize}
\end{defi}
\begin{rem}\label{rem:marked}
An important technical tool for study of Cartesian (say) fibrations over ${\EuScript C}$ is Lurie's model category
$( {\operatorname{Set}}^+_\Delta)_{/{\EuScript C}}$ whose objects are {\em marked simplicial sets over ${\EuScript C}$}, i.e.,
simplicial sets $S\to{\EuScript C}$ over ${\EuScript C}$ with a distinguished set of edges called {\em marked edges}
\cite[\S 3.1]{lurie:htt}.
Cartesian fibrations (with Cartesian edges as marked edges) are characterized as fibrant objects
in $( {\operatorname{Set}}^+_\Delta)_{/{\EuScript C}}$, so one can use fibrant resolutions etc. We do not go into further details
in this reminder but will use this technique occasionally in the proofs.
\end{rem}
Examples of (co)Cartesian fibrations are provided by the ${\infty}$-categorical version of the
Grothendieck construction (relative nerve) \cite[Def. 3.2.5.2]{lurie:htt}.
Let $C$ be an ordinary small category with nerve ${\EuScript C}=\operatorname{N}{\EuScript C}$ and $q: C \to {\operatorname{Set}}_{\Delta}$ be a functor such that, for every
object $c$ of $C$, the simplicial set $q(c)$ is an $\infty$-category. Let also $[n]$ be the poset
${0<1\cdots <n}$ considered as a category, with the nerve $\Delta^n$, the $n$-simplex.
\begin{defi}\label{def:groth}
(a) The {\em covariant Grothendieck construction} of $q$ is the simplicial set $\Gamma(q)$
defined as follows. An $n$-simplex in $\Gamma(q)$ consists of
\begin{enumerate}
\item a functor $\sigma: [n] \to C$,
\item for every $I \subset [n]$, a map $\Delta^I \to q(\sigma(\max(I)))$,
\end{enumerate}
such that, for every $J \subset I \subset [n]$, the diagram
\[
\begin{tikzcd}
\Delta^J \ar{r}\ar{d} & q(\sigma(\max(J))) \ar{d}\\
\Delta^I \ar{r} & q(\sigma(\max(I)))
\end{tikzcd}
\]
commutes. There is a natural forgetful map (${\infty}$-functor) $\pi: \Gamma(q) \to {\EuScript C}=\operatorname{N}(C)$.
\vskip .2cm
(b) The {\em contravariant Grothendieck construction} $\chi(q)$ of $q$ is the simplicial set
$\Gamma((-)^{\operatorname{op}} \circ q)^{{\operatorname{op}}}$. It comes equipped with a natural forgetful map $\pi: \chi(q) \to
{\EuScript C}^{\operatorname{op}}$.
\end{defi}
As shown in \cite[\S 3.2.5]{lurie:htt}, the functor $\Gamma(q) \to {\EuScript C}$ is a coCartesian fibration
while the functor $\chi(q) \to {\EuScript C}^{{\operatorname{op}}}$ is a Cartesian fibration.
Conversely, see \cite[Th.3.2.0.1]{lurie:htt} any coCartesian fibration $p: {\EuScript X}\to{\EuScript C}$ gives rise to a {\em classifying ${\infty}$-functor}
$Q: C \to{\EuScript Cat}_{\infty}$ defined uniquely up to a contractible space of
choices. For an object $c$ of $C$, the value $ {\EuScript Q}(c)$
is the preimage $p^{-1}(c)\subset{\EuScript X}$ in the sense of simplicial sets.
If ${\EuScript X}=\Gamma(q) \to {\EuScript C}^{\operatorname{op}}$
is the covariant Grothendieck construction of a functor $q$ as in Definition \ref{def:groth},
then $\operatorname{N}(q)$ is a classifying ${\infty}$-functor.
Dually, any Cartesian fibration ${\EuScript X}$ over ${\EuScript C}$ gives rise to a classifying functor $Q: {\EuScript C}^{\operatorname{op}}\to {\EuScript Cat}_{\infty}$.
Again, the value $Q(c)$ on an object $c\in C$ is the preimage $p^{-1}(c)\subset {\EuScript X}$.
If ${\EuScript X}=\chi(q)\to{\EuScript C}^{\operatorname{op}}$ is the contravariant Grothendieck construction as above, then $\operatorname{N}(q)$
can be taken as a classifying ${\infty}$-functor.
\begin{rem}\label{rem:subtlety}
If ${\EuScript C}=\operatorname{N} C$ is the nerve of an ordinary category (as we assume),
then an ${\infty}$-functor $Q: {\EuScript C}\to{\EuScript Cat}_{\infty}$
is a more general datum than an input of Definition \ref{def:groth}, i.e., an ordinary functor $q: C\to {\operatorname{Set}}_\Delta$ with each $q(c)$ an ${\infty}$-category. More precisely, $Q$ does give for each object $c\in C$ an ${\infty}$-category
$Q(c)$ and for each morphsim $u: c\to c'$ an ${\infty}$-functor $Q(u): Q(c)\to Q(c')$,
but the compatibility of these data with composition holds not strictly but up to higher homotopies
which are hidden in the definition of ${\EuScript Cat}_{\infty}$ as a simplicial nerve. For example,
for a composable pair of morphisms $u,v$ in $C$ we are given a $2$-simplex with edges $Q(uv), Q(u), Q(v)$ and so on.
So $Q$ is an analog of a pseudo-functor $C\to{\EuScript Cat}$ in the classical sense.
\end{rem}
\paragraph{(Co)Cartesian fibrations over $\Delta^1$: implicit ${\infty}$-functors.}
\label{par:cocart-implicit}
In general, the construction of a classifying ${\infty}$-functor of a (co)Cartesian fibration is quite involved
and uses the model category of marked simplicial sets over ${\EuScript C}$ (see Remark \ref{rem:marked}).
Let us work out explicitly the simplest example, which will be of particular importance for
the discussion of adjoint functors later in the paper.
\begin{ex}[(Grothendieck construction for a single ${\infty}$-functor)]\label{ex:groth-single}
Let $C=[1]=\{0\to 1\}$ and ${\EuScript C}=\operatorname{N} C=\Delta^1$ be the (${\infty}$-) category with two objects $0,1$ and one
non-trivial morphism.
A covariant ${\infty}$-functor $q: {\EuScript C}\to{\EuScript Cat}_{\infty}$ reduces to a single ${\infty}$-functor
$F: {\EuScript A}=q(0) \to {\EuScript B}=q(1)$ of ${\infty}$-categories, the subtlety
from Remark \ref{rem:subtlety} not appearing here. In this case we denote the covariant Grothendieck construction of
$q$ by $\Gamma(F)$.
Similarly, a contravariant functor $q: {\EuScript C}\to{\EuScript Cat}_{\infty}$
is the same as a single ${\infty}$-functor $G: {\EuScript B} \to {\EuScript A} $. In this case we denote the contravariant Grothendieck
construction of $q$ by $\chi(G)$.
Let us describe the covariant Grothendieck construction $\Gamma(F)$ associated
to an ${\infty}$-functor $F: {\EuScript A}\to {\EuScript B}$ explicitly.
For any integers $0\leq m\leq n$ we have the standard embedding $[m]\hookrightarrow [n]$.
For any simplicial set $X$ we will denote ${\partial}_{[m]}: X_n \to X_m$
the iterated face map associated to this embedding.
Specializing Definition \ref{def:groth} to our case, we see that an $n$-simplex $s\in \Gamma(F)_n$
is a datum of:
\begin{itemize}
\item[(1)] An integer $-1\leq m\leq n$.
\item[(2)] An $n$-simplex $s_ {\EuScript B}\in {\EuScript B}_n$ and, in case $m\geq 0$,
an $m$-simplex $s_ {\EuScript A}\in {\EuScript A}_m$ such that
${\partial}_{[m]} s_ {\EuScript B} = F(s_ {\EuScript A})$.
\end{itemize}
The projection $p: \Gamma(F)\to\Delta^1$ sends an $n$-simplex $(m, s_ {\EuScript A}, s_ {\EuScript B})$
to (the degeneration of) $0$, if $m=n$, to (the degeneration of) $1$, if $m=-1$ and to
(the degeneration of) the $1$-dimensional simplex of $\Delta^1$ otherwise.
In particular, a {\em section} of $p$, i.e., an edge (morphism) projecting to the edge of $\Delta^1$,
is a datum of an object $a=s_ {\EuScript A} \in {\EuScript A}$, $b\in {\EuScript B}$ and a morphism $s_ {\EuScript B}: F(a)\to b$ in $ {\EuScript B}$.
Among these, coCartesian edges of $\Gamma(F)$
are distinguished by the condition that $s_ {\EuScript B}$ is an equivalence. In particular, we can take
for $a$ any object of $ {\EuScript A}$ and put $s_ {\EuScript B}=\operatorname{Id}: F(a)\to F(a)$.
We see that any object in $ {\EuScript A}= p^{-1}(0)$ is a source of a coCartesian edge, so $p$ is a coCartesian fibration,
by Definition \ref{def:cart-fib}(b).
\end{ex}
\begin{ex}[(Implicit ${\infty}$-functors)]\label{ex:implicit}
Conversely,
given a Cartesian (resp. coCartesian) fibration $p: {\EuScript X}\to\Delta^1$, we have two ${\infty}$-categories
$ {\EuScript A}=p^{-1}(0)$, $ {\EuScript B}=p^{-1}(1)$ (the preimages as simplicial sets). A classifying ${\infty}$-functor for $p$
reduces to a single ${\infty}$-functor $G: {\EuScript B}\to {\EuScript A}$ (resp. $F: {\EuScript A}\to {\EuScript B}$).
We can say that $G$ (resp. $F$)
is an {\em implicit ${\infty}$-functor} defined by a Cartesian (resp. coCartesian) fibration $p$.
Let us explain how to construct $G$ for a Cartesian fibration $p$ (the case of a coCartesian fibration beingn dual).
We define $G$ (a morphism of simplicial sets)
inductively on simplices of dimensions $0,1,2,\cdots$.
\begin{itemize}
\item[(0)] Let $y$ be an object (vertex) of $ {\EuScript B}$, so $p(y)=1$.
Let $g: 0\to 1$ be the nontrivial edge of $\Delta^1$.
By Definition \ref{def:cart-edge}(a),
there is a Cartesian edge $f_y: x\to y$ in ${\EuScript X}$ covering $g$, so that $x$ is a vertex of $ {\EuScript A}$.
Choosing such an edge, we put $G(y)=x$, thus defining $G$ on $0$-simplices.
\item[(1)] Let $v: y_0\to y_1$ be a morphism (edge) in $ {\EuScript B}$. Let $x_i=G(y_i)$ be as constructed before.
Because ${\EuScript X}$ is an ${\infty}$-category
(weak Kan complex) we can form a pseudo-composition of $v$ and $f_{y_0}$, filling a horn of type
$\Lambda^2_1\subset \Delta^2$ , see \cite[Def. 1.1.2.1]{lurie:htt} for notation. This gives a
$2$-simplex (triangle) denoted \circled{1} in Fig. \ref{fig:implicit1}. The third side of this
triangle is a morphism which we denote $c: x_0= G(y_0) \to y_1$.
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=0.3]
\node at (-6,0){\small$\bullet$};
\node at ( 6,0){\small$\bullet$};
\draw [decoration={markings,mark=at position 0.7 with
{\arrow[scale=1.5,>=stealth]{>}}},postaction={decorate},
line width = .3mm] (-6,0) -- (6,0) ;
\node at (-7,0){$0$};
\node at (7,0){$1$};
\node at (0,-1) {$g$};
\node at (6,4) {\small$\bullet$};
\node at (6,9) {\small$\bullet$};
\node at (-6,4) {\small$\bullet$};
\node at (-6,9) {\small$\bullet$};
\drarr (6,9) -- (6,4);
\drarr (-6,9) -- (-6,4);
\drarr (-6,9) -- (6,9);
\drarr (-6,4) -- (6,4);
\drarr (-6,9) -- (6,4);
\node at (2,7.5){\circled{1}};
\node at (-2, 5.5){\circled{2}};
\node at (7,9){$y_0$};
\node at (7,4){$y_1$};
\node at (-7,9){$x_0$};
\node at (-7,4){$x_1$};
\node at (7, 6.5){$v$};
\node at (-7, 6.5){$u$};
\node at (0,10){$f_{y_0}$};
\node at (0,3){$f_{y_1}$};
\node at (0.5, 7){$c$};
\draw (6, 6.5) ellipse (2.5cm and 5cm);
\draw (-6, 6.5) ellipse (2.5cm and 5cm);
\node at (-8.5, 11){$ {\EuScript A}$};
\node at (8.5, 11){$ {\EuScript B}$};
\end{tikzpicture}
\caption{An implicit ${\infty}$-functor from a Cartesian fibration: values on morphisms.}\label{fig:implicit1}
\end{figure}
Next, the condition for $f_{y_1}$ to be a Cartesian edge reduces in our particular case to saying that
${\EuScript X}_{/f_{y_1}} \to {\EuScript X}_{/y_1}$ is a trivial Kan fibration, in particular, it is surjective on $0$-simplices.
Now, $c$ is an object ($0$-simplex) of ${\EuScript X}_{y_1}$, while a $0$-simplex of ${\EuScript X}_{f_{y_1}}$ covering $c$
is, by definition, a triangle of the form \circled{2} in Fig. \ref{fig:implicit1}.
Denote $u: x_0\to x_1$ its third side. We put $G(v)=u$, thus defining $G$ on $1$-simplices.
\item[(2)] Next, let $\tau$ be a $2$-simplex in $ {\EuScript B}$ with vertices $y_0,y_1,y_2$
and edges $v_{ij}: y_i\to y_j$, $i<j$.
Let $x_i= G(y_i)$ and $u_{ij}= G(v_{ij})$, already constructed.
The previous choices give us triangles filling a part of the boundary of the $3$-dimensional prism
$\Delta^1\times\Delta^2$, see Fig. \ref{fig:implicit2}. More precisely, $\tau$ fills one (the right) foundation of the prism and the
triangles of the form \circled{1} and \circled{2} for the $v_{ij}$ fill the walls of the prism (which are
therefore triangulated), but the left foundation is not filled.
Now, we have the standard triangulation of the prism, compatible with the triangulations of the walls.
It has 3 tetrahedra which we denote by the sets of their vertices:
$[x_0y_0y_1y_2]$, $[x_0 x_1 y_1 y_2]$ and $[x_0 x_1 x_2 y_2]$. We fill these tetrahedra (by $3$-simplices
of ${\EuScript X}$) as follows.
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=0.4]
\node at (10,0){\small$\bullet$};
\node at (10,9){\small$\bullet$};
\node at (6,4){\small$\bullet$};
\node at (-10,0){\small$\bullet$};
\node at (-10,9){\small$\bullet$};
\node at (-14,4){\small$\bullet$};
\draw [decoration={markings,mark=at position 0.7 with
{\arrow[scale=1.5,>=stealth]{>}}},postaction={decorate},
line width = .3mm] (-10,0) -- (10,0) ;
\draw [decoration={markings,mark=at position 0.7 with
{\arrow[scale=1.5,>=stealth]{>}}},postaction={decorate},
line width = .3mm] (-10,9) -- (10,9) ;
\draw [decoration={markings,mark=at position 0.7 with
{\arrow[scale=1.5,>=stealth]{>}}},postaction={decorate},
line width = .3mm] (10,9) -- (6,4) ;
\draw [decoration={markings,mark=at position 0.7 with
{\arrow[scale=1.5,>=stealth]{>}}},postaction={decorate},
line width = .3mm] (6,4)-- (10,0) ;
\draw [decoration={markings,mark=at position 0.7 with
{\arrow[scale=1.5,>=stealth]{>}}},postaction={decorate},
line width = .3mm] (10,9) -- (10,0) ;
\draw [decoration={markings,mark=at position 0.7 with
{\arrow[scale=1.5,>=stealth]{>}}},postaction={decorate},
line width = .3mm] (-14,4) -- (6,4) ;
\draw [decoration={markings,mark=at position 0.7 with
{\arrow[scale=1.5,>=stealth]{>}}},postaction={decorate},
line width = .3mm] (-10,9) -- (-14,4) ;
\draw [decoration={markings,mark=at position 0.7 with
{\arrow[scale=1.5,>=stealth]{>}}},postaction={decorate},
line width = .3mm] (-14,4) -- (-10,0) ;
\draw [dashed, decoration={markings,mark=at position 0.7 with
{\arrow[scale=1.5,>=stealth]{>}}},postaction={decorate},
line width = .35mm] (-10,9) -- (-10,0) ;
\drarr (-14,4) -- (10,0);
\drarr (-10,9) -- (6,4);
\draw [dashed, decoration={markings,mark=at position 0.7 with
{\arrow[scale=1.5,>=stealth]{>}}},postaction={decorate},
line width = .15mm] (-10,9) -- (10,0) ;
\node at (11,9){$y_0$};
\node at (7,4){$y_1$};
\node at (11,0){$y_2$};
\node at (-11,9){$x_0$};
\node at (-15,4){$x_1$};
\node at (-11,0){$x_2$};
\node at (11, 4.5){$v_{02}$};
\node at (7,7){$v_{01}$};
\node at (8,3){$v_{12}$};
\node at (0, 10){$f_{y_0}$};
\node at (0,-1){$f_{y_2}$};
\node at (-3,4.8){$f_{y_1}$};
\node at (-13,7){$u_{01}$};
\node at (-13,2){$u_{12}$};
\node at (-9,5){$u_{02}$};
\node at (8.5,5){\circled{$\tau$}};
\end{tikzpicture}
\caption{An implicit ${\infty}$-functor from a Cartesian fibration: values on $2$-simplices.}\label{fig:implicit2}
\end{figure}
First, out of the desired $[x_0y_0y_1y_2]$ we already have the horn $\Lambda^3_1$, so we can fill it by
the weak Kan property of ${\EuScript M}$. After this, we have a horn $\Lambda^3_2$ out of the desired $[x_0 x_1 y_1 y_2]$,
so we fill it as well. After this, the triangle $[x_0 x_1 y_2]$ represents a $1$-simplex in ${\EuScript X}_{/y_2}$
whose vertices $[x_0y_1]$ and $[x_1y_2]$ are lifted to ${\EuScript X}_{/f_{y_2}}$. Now, as before, the fact
that $f_{y_2}$ is a Cartesian edge, means that ${\EuScript X}_{/f_{y_2}} \to {\EuScript X}_{/y_2}$ is a trivial Kan fibration.
Therefore any map of $\Delta^1$ to the $ {\EuScript X}_{/y_2}$ with a lift over ${\partial} \Delta^1$ to ${\EuScript X}_{/f_{y_2}}$
can be lifted to a map of $\Delta^1$ to ${\EuScript X}_{/f_{y_2}}$, i.e., to a $1$-simplex of ${\EuScript X}_{/f_{y_2}}$.
Such a $1$-simplex is, by definition \cite[Prop. 1.2.9.2]{lurie:htt} nothing but a tetrahedron ($3$-simplex of ${\EuScript X}$)
of the desired shape $[x_0 x_1 x_2 y_2]$.
The $4$th side of this tetrahedron is a $2$-simplex $\sigma$ in $ {\EuScript A}$ with sides $u_{01}, u_{12}, u_{02}$.
We put $G(\tau)=\sigma$, thus defining $G$ on $2$-simplices.
\item[(...)]
\item[($n$)] Suppose that $G$ is defined on simplices of dimension $<n$ and $\tau$ is an $n$-simplex
of $ {\EuScript B}$, with vertices $y_0,\cdots, y_n$.
Then we have a filling of the part of the boundary of the prism $\Delta^1\times\Delta^n$
obtained by omitting one of the two foundations. This prism has a standard triangulation into $n+1$ simplices
of dimension $n+1$. We fill them inductively. The first $n$ fillings are possible by the weak Kan property
of ${\EuScript X}$, filling horns of shapes $\Lambda^{n+1}_i$, $i=1,\cdots, n$. After this, we obtain an $(n-1)$-simplex
in ${\EuScript X}_{/y_n}$ whose boundary is lifted into ${\EuScript X}_{/f_{y_n}}$, so the condition that $f_{y_n}$ is a Cartesian edge
allows us to extend the lifting from the boundary to an entire simplex, thus obtaining an $(n-1)$-simplex
in ${\EuScript X}_{/f_{y_n}}$. This $(n-1)$-simplex is interpreted as an $(n+1)$-simplex in ${\EuScript X}$ which completes
the triangulation of the prism. The last face of this $(n+1)$-simplex is an $n$-simplex $\sigma$
filling the other foundation of the prism.
We define $G(\tau)=\sigma$, this defining $G$ on $n$-simplices.
\end{itemize}
\noindent Comparing with (the contravariant modification of) Example \ref{ex:groth-single},
we note that the all the simplices constructed in the above inductive process, form in fact a copy of the
contravariant Grothendieck construction of $G$. Therefore we have an ${\infty}$-functor $\chi(G)\to {\EuScript X}$
compatible with the projection to $\Delta^1$.
One can further show that it is an equivalence of ${\infty}$-categories.
\end{ex}
\paragraph{Kan extensions.} Our frequent technical tool will be the theory of ${\infty}$-categorical Kan
extensions see \cite[\S 4.3] {lurie:htt} or a short summary in \cite[\S A.5]{dkss:1}. Since it will
be used only in the proofs, we do not give a summary here. We will frequently be using the
following fundamental result about Kan extensions (see \cite[Prop. 4.3.2.15]{lurie:htt}):
\begin{prop}\label{prop:kan-ext-equiv}
(a) Let $ {\EuScript J} $ be an ${\infty}$-category and ${\EuScript {I}}\subset {\EuScript J}$ a full subcategory.
Let $\Di$ be a category with colimits and ${\EuScript E}\subset\operatorname{Fun}({\EuScript {I}},\Di)$
be a full subcategory. Let ${\EuScript E}^!\subset \operatorname{Fun}( {\EuScript J},\Di)$
be the full subcategory formed by ${\infty}$-functors which are
left Kan extensions of functors from ${\EuScript E}$. Then the
restriction functor ${\EuScript E}^!\to{\EuScript E}$ is a trivial Kan fibration, in particular, it is
an equivalence of ${\infty}$-categories.
(b) A similar statement for right Kan extensions, if $\Di$ has limits. \qed
\end{prop}
In fact, we only need the existence of those (co)limits that enter in the standard pointwise
formulas for Kan extensions \cite[Eq. (A.5.2)]{dkss:1}.
We will also be using a variant of Proposition \ref{prop:kan-ext-equiv} for relative Kan extension
which is also included in \cite[Prop. 4.3.2.15]{lurie:htt}.
\subsection {Adjunctions for ${\infty}$-categories}\label{subsec:adj-oo}
\paragraph{Grothendieck constructions and adjunctions.}
The Grothendieck constructions provide a very effective means to study
adjunctions of $\infty$-categories \cite[\S 5.2.2]{lurie:htt}.
\begin{defi}\label{def:bicart}
Let ${\EuScript C} = \operatorname{N} C$ be the nerve of an ordinary category. An ${\infty}$-functor $p: {\EuScript X}\to{\EuScript C}$ is called
a {\em biCartesian fibration}, if it is both a Cartesian and coCartesian fibration.
\end{defi}
A biCartesian fibration gives rise to two ${\infty}$-functors ${\EuScript C}\to{\EuScript Cat}_{\infty}$: a covariant
${\infty}$-functor $Q^{\operatorname{coCart}}$ classifying $p$ as a coCartesian fibration and a contravariant
${\infty}$-functor $Q^{\operatorname{Cart}}$ classifying $p$ as a Cartesian fibration. The values of $Q^{\operatorname{coCart}}$
and $Q^{\operatorname{Cart}}$ on an object $c\in{\EuScript C}$ are the same: the preimage $p^{-1}(c)\subset {\EuScript X}$, so for a morphism
$g: c\to d$ in $C$ we have a pair of ${\infty}$-functors in the opposite directions
\[
\xymatrix{
p^{-1}(c) \ar@<.4ex>[rr]^{Q^{\operatorname{coCart}}(g) } &&p^{-1}(d) \ar@<.4ex>[ll]^{Q^{\operatorname{Cart}}(g) }.
}
\]
By analogy with the classical category theory, it is natural to regard such ${\infty}$-functors as adjoint.
By focusing on pairs of single functors as in Example \ref{ex:groth-single}, this leads to the following definition.
\begin{defi} \begin{enumerate}
\item An {\em adjunction} of $\infty$-categories is a biCartesian
fibration
\[
p: {\EuScript X} \longrightarrow \Delta^1.
\]
\item Let $F: {\EuScript A} \to {\EuScript B}$ be an ${\infty}$-functor of $\infty$-categories. We say
that
\begin{enumerate}
\item $F$ admits a right adjoint if the covariant Grothendieck construction
$\Gamma(F) \to \Delta^1$ (always a coCartesian fibration) is also a Cartesian fibration.
\item $F$ admits a left adjoint if the contravariant Grothendieck construction (always a Cartesian fibration)
$\chi(F) \to \Delta^1$ is also a coCartesian fibration.
\end{enumerate}
\end{enumerate}
\end{defi}
Given an adjunction $p: {\EuScript X} \to \Delta^1$, we define $\infty$-categories $ {\EuScript A} = p^{-1}(0)$ and
$ {\EuScript B}= p^{-1}(1)$ as in Example \ref{ex:groth-single}. Further, as in Example \ref{ex:implicit},
we construct implicit ${\infty}$-functors
\begin{equation}\label{eq:adj}
\xymatrix{
{\EuScript A} \ar@<.4ex>[r]^{F } & {\EuScript B} \ar@<.4ex>[l]^{G },
}
\end{equation}
defined by $p$ as a Cartesian ($G$) and coCartesian ($F$) fibration.
\begin{defi}
Let $F,G$ be ${\infty}$-functors in opposite directions as in \eqref{eq:adj}. We say that $G$
{\em is right adjoint to} $F$ and $F$ {\em is left adjoint to} $G$
(or equivalently, $(F,G)$ {\em form an adjoint pair}), if there exists a biCartesian
fibration $p: {\EuScript X}\to\Delta^1$ such that $ {\EuScript A}=p^{-1}(0)$, $ {\EuScript B}=p^{-1}(1)$ and moreover, $F$ is an
implicit ${\infty}$-functor defined by $p$ as a coCartesian fibration while $G$ is an implicit
${\infty}$-functor defined by $p$ as a Cartesian fibration. We will use the notation $F\dashv G$ to
signify that $(F,G)$ form an adjoint pair, as well as to (implicitly) signify the adjunction itself.
\end{defi}
Thus being an adjoint pair is a property and not an additional structure.
\paragraph{Units and counits.}
A natural expectation for a workable concept of adjoint ${\infty}$-functors $F,G$ would be data of weak equivalences
of
mapping spaces
\begin{equation}
\operatorname{Map}_ {\EuScript B}(F(a), b) \,\simeq \, \operatorname{Map}_{\EuScript C}(a,G(b))
\end{equation}
or, the data of unit and counit morphisms of ${\infty}$-functors
\begin{equation}
u: \operatorname{Id}_ {\EuScript A} \longrightarrow GF, \quad c: FG\to\operatorname{Id}_ {\EuScript B}
\end{equation}
inducing such weak equivalences via functorialities \eqref{eq:map-functor}.
This is achieved by the following result.
\begin{prop}\label{prop:oo-unit}
For a pair of ${\infty}$-functors
$\xymatrix{
{\EuScript A} \ar@<.4ex>[r]^{F } & {\EuScript B} \ar@<.4ex>[l]^{G }
}$
the following are equivalent:
\begin{itemize}
\item[(i)] $(F,G)$ is an adjoint pair.
\item [(ii)] There exists a morphism $u: \operatorname{Id}_ {\EuScript A}\to GF$ (a {\em unit}) such that for any objects $a\in {\EuScript A}$, $b\in {\EuScript B}$ the composition
\[
\operatorname{Map}_ {\EuScript B}(F(a), b) \buildrel G_*\over\longrightarrow \operatorname{Map}_ {\EuScript A}(GF(a), G(b)) \buildrel u^*\over\longrightarrow \operatorname{Map}_ {\EuScript A}(a, G(b))
\]
is a weak equivalence.
\item[(iii)] There exists a morphism $c: FG\to\operatorname{Id}_ {\EuScript B}$ (a {\em counit}) such that for any objects
$a\in {\EuScript A}$, $b\in {\EuScript B}$ the
composition
\[
\operatorname{Map}_ {\EuScript A}(a, G(b)) \buildrel F_* \over\longrightarrow \operatorname{Map}_ {\EuScript B}(F(a), FG(B)) \buildrel c_*\over\longrightarrow \operatorname{Map}_ {\EuScript B}(F(a), b)
\]
is a weak equivalence.
\end{itemize}
\end{prop}
\noindent{\sl Proof:} The equivalence of (i) and (ii) is \cite[Prop. 5.2.2.8]{lurie:htt}. The equivalence of
(i) and (iii) is dual. \qed
\begin{cor}
If $(F,G)$ is an adjoint pair of ${\infty}$-functors as above, then at the level of homotopy categories
we get an adjoint paiir of functors $\xymatrix{
\operatorname{h}\! {\EuScript A} \ar@<.4ex>[r]^{\operatorname{h}\! F } &\operatorname{h}\! {\EuScript B} \ar@<.4ex>[l]^{\operatorname{h}\! G }
}$ in the usual sense. \qed
\end{cor}
\vskip .2cm
We will now provide a description of the unit and counit, based on the universal properties of
Cartesian and coCartesian edges as effectively captured within Lurie's model structures on marked
simplicial sets. The diagramatic manipulations that we will use to construct and relate functors of
$\infty$-categories are very efficiently controlled by Lurie's theory of relative Kan extensions
\cite{lurie:htt}. Later in the text, we will omit details about analogous arguments.
\begin{lem}[(Units and counits implicitly)] \label{lem:unit}
Let $p: {\EuScript X} \to \Delta^1$ be an adjunction of $\infty$-categories,
with $ {\EuScript A}=p^{-1}(0)$, $ {\EuScript B}=p^{-1}(1)$.
\vskip .2cm
(a) Let
$
{\EuScript U} \subset \operatorname{Fun}(\Delta^2, {\EuScript X})
$
be the full subcategory spanned by $2$-simplices of the form
\begin{equation}\label{eq:unitdiag}
\begin{tikzcd}
a \ar{rd}{!}\ar{d} & \\
a' \ar{r}{*} & b
\end{tikzcd}
\end{equation}
with $a,a'$ objects of $ {\EuScript A}$, $b$ an object of $ {\EuScript B}$, the edge $a \to b$ coCartesian, and the
edge $a' \to b$ Cartesian. Then the functor
$
{\EuScript U} \to {\EuScript A}
$
given by evaluation on the vertex $\{0\}$ (so a triangle \eqref{eq:unitdiag}
is sent into $a$), is a trivial Kan fibration (in particular, an equivalence of
${\infty}$-categories).
\vskip .2cm
(b) Let ${\EuScript C}\subset \operatorname{Fun}(\Delta^2, {\EuScript X})$ be the full subcategory spanned by $2$-simplices of the form
\begin{equation}\label{eq:counitdiag}
\begin{tikzcd}
b' \ar{rd} & \\
a \ar{r}{*} \ar{u}{!}& b
\end{tikzcd}
\end{equation}
with $a$ an object of $ {\EuScript A}$, $b,b'$ objects of $ {\EuScript B}$, the edge $a\to b$
Cartesian and the egde $a\to b'$ coCartesian.
Then the functor ${\EuScript C}\to {\EuScript B}$ given by evaluation on the vertex $\{2\}$
(so a triangle \eqref{eq:counitdiag} is set into $b$) is
a trivial Kan fibration (in particular, an equivalence of
${\infty}$-categories).
\end{lem}
\begin{proof} We prove (a), since (b) is similar.
First consider the $\infty$-category $ {\EuScript V} \subset \operatorname{Fun}(\Delta^1, {\EuScript X})$ spanned by the
coCartesian edges $a \to b$ that cover the edge $0 \to 1$. By \cite[Ex. 4.3.1.4]{lurie:htt},
the edges in $ {\EuScript V}$ are precisely the left $p$-Kan extensions of their restriction to the
vertex $a$ along the inclusion $\{0\} \subset \Delta^1$. By
\cite[Prop. 4.3.2.15]{lurie:htt} we therefore
obtain that the functor $ {\EuScript V} \to {\EuScript A}$ given by evaluation at $\{0\}$ is a trivial Kan
fibration. Similarly, let $ {\EuScript V}' \subset \operatorname{Fun}(\Delta^1, {\EuScript X})$ denote the full subcategory
spanned by the Cartesian edges $a' \to b$ in ${\EuScript X}$ covering $0 \to 1$. The projection functor
$ {\EuScript V}' \to {\EuScript B}$ given by evaluation at $\{1\}$ is a trivial Kan fibration, since the edges in
$ {\EuScript V}'$ are precisely the right $p$-Kan extensions of their restriction to $\{1\}$. Since
trivial Kan fibrations are stable under pullback and composition, we obtain that the
composite of
\[
{\EuScript V} \times_{ {\EuScript B}} {\EuScript V}' \to {\EuScript V} \to {\EuScript A}
\]
is a trivial Kan fibration. Further, the horn inclusion of marked simplicial sets
\[
\Lambda^2_2 \subset \Delta^2
\]
with degenerate edges and the edge $\{1,2\}$ marked. This inclusion is marked anodyne
\cite[Def. 3.1.1.1]{lurie:htt}. Further, the simplicial set ${\EuScript X}$, being a Cartesian
fibration, defines (by marking Cartesian edges)
a
fibrant object in the Cartesian model structure on
$( {\operatorname{Set}}_{\Delta}^+)_{/\Delta^1}$ (in $\Delta^1$ every edge is marked).
Therefore, the induced pullback map on mapping $\infty$-categories
\[
\operatorname{Map}^{\flat}_{\Delta^1}(\Delta^2, X) \longrightarrow \operatorname{Map}^{\flat}_{\Delta^1}(\Lambda^2_2, X),
\]
is a trivial Kan
fibration. Here the superscript $\flat$ means that we consider mapping ${\infty}$-categories with respect to the above mentioned markings, see \cite[\S 3.1.3]{lurie:htt}.
We conclude the proof by noting that the map $ {\EuScript U} \to {\EuScript V} \times_{ {\EuScript B}} {\EuScript V}'$ is
obtained as a pullback of the latter trivial Kan fibration so that it is a trivial Kan
fibration as well.
\end{proof}
\begin{rem}
Any morphism $a \to a'$ appearing as a part of a diagram \eqref{eq:unitdiag} is a potential choice of a
unit morphism $a \to G(F(a))$. Lemma \ref{lem:unit}(a) expresses the fact that, for every object $a$ of
$ {\EuScript A}$, there is an essentially unique unit morphism $a \to a'$.
Similarly, any morphism $b' \to b$ appearing as a part of a diagram
\eqref{eq:counitdiag} is a potential choice of a
counit morphism $F(G(b))\to b$. Lemma \ref{lem:unit}(b) expresses the fact that, for every object $b$ of
$ {\EuScript B}$, there is an essentially unique counit morphism $b'\to b$.
\end{rem}
\paragraph{Exactness; adjoints of the composition.}
For future use we record the analogs of some familiar properties of adjoints known in the classical
category theory.
\begin{prop}\label{prop:adj-exact} \cite[Prop. 5.2.4.3]{lurie:htt}
If an ${\infty}$-functor $F: {\EuScript A}\to {\EuScript B}$ has a right adjoint, then it is left exact (i.e., preserves all colimits which exist in $ {\EuScript A}$).
If $F$ has a left adjoint, then it is right exact (i.e., preserves all limits that exist in $ {\EuScript A}$).
\qed
\end{prop}
\begin{prop}\cite[Prop. 5.2.2.6]{lurie:htt}
Let $\xymatrix{
{\EuScript A} \ar@<.4ex>[r]^{F } & {\EuScript B} \ar@<.4ex>[l]^{G }
\ar@<.4ex>[r]^{F' } &{\EuScript C} \ar@<.4ex>[l]^{G'}
}$
be a diagram of ${\infty}$-categories and
${\infty}$-functors such that $(F,G)$ and $(F', G')$ are adjoint pairs.
Then $(F'\circ F, G\circ G')$ is also an adjoint pair. \qed
\end{prop}
\subsection{Comparison with other approaches to adjunctions}\label{subsec:comparison}
\paragraph{Fourier-Mukai kernels.}
As mentioned in \S\ref{subset:stab-triang}, any dg-category $\mathcal {A}$ gives an ${\infty}$-category
$\operatorname{N}^{\operatorname{dg}}\mathcal {A}$; any dg-functor $f: \mathcal {A}\to\mathcal{B}$ gives an ${\infty}$-functor
$\operatorname{N}^{\operatorname{dg}} f: \operatorname{N}^{\operatorname{dg}} \mathcal {A} \to \operatorname{N}^{\operatorname{dg}} \mathcal{B}$. Therefore the ${\infty}$-categorical theory of adjoints
provides an approach to defining adjointness for dg-functors. In this and the next
paragraph we compare it with other approaches to this question that have been
used in the literature.
\vskip .2cm
Let $X$ be a smooth projective algebraic variety over a field $\mathbf k$.
We denote by ${\operatorname{Coh}}_X$ the abelian category of coherent sheaves on $X$,
and by $D^b{\operatorname{Coh}}_X$ its bounded derived category. We can think of
objects of $D^b{\operatorname{Coh}}_X$ as complexes of quasicoherent sheaves with
bounded coherent cohomology.
We have eqiivalences of
(usual) categories
\[
D^b{\operatorname{Coh}}_X \,\simeq\, H^0 D^b_{\operatorname{dg}} {\operatorname{Coh}}_X\,\simeq\, \operatorname{h}\! D^b_{({\infty})} {\operatorname{Coh}}_X,
\]
where:
\begin{itemize}
\item $D^b_{\operatorname{dg}} {\operatorname{Coh}}_X$ is the dg-enhancement of $D^b{\operatorname{Coh}}_X $ obtained
by taking an injective resolution of each object and
forming the Hom-complexes between such resolutions \cite{BK:enhanced}.
\item $D^b_{({\infty})} {\operatorname{Coh}}_X = \operatorname{N}^{\operatorname{dg}} D^b_{\operatorname{dg}} {\operatorname{Coh}}_X$ is the dg-nerve
of the dg-enhancement.
\end{itemize}
Let $X,Y$ be two smooth projective varieties.
Denote $X\buildrel p_X\over\leftarrow X\times Y \buildrel p_Y\over\to Y$
the projections. For an object ${\mathcal K}\in D^b{\operatorname{Coh}}_{X\times Y}$
we have the {\em Fourier-Mukai functor}
\[
F_{\mathcal K} : D^b{\operatorname{Coh}}_X \longrightarrow D^b{\operatorname{Coh}}_Y, \quad \mathcal{F}\mapsto Rp_{Y*} (p_X^*\mathcal{F} \otimes {\mathcal K}).
\]
The complex ${\mathcal K}$ is referred to as the {\em Fourier-Mukai kernel} (of $F_{\mathcal K}$), see
\cite{huybrechts}.
By passing to appropriate resolutions, $F_{\mathcal K}$ can be lifted to a dg-functor and
then to
to an ${\infty}$-functor
\[
F_{\mathcal K}^{\operatorname{dg}}: D^b_{\operatorname{dg}} {\operatorname{Coh}}_X\longrightarrow D^b_{\operatorname{dg}} {\operatorname{Coh}}_Y,\quad
F_{\mathcal K}^{({\infty})} =\operatorname{N}^{\operatorname{dg}} F_{\mathcal K}^{\operatorname{dg}}: D^b_{({\infty})} {\operatorname{Coh}}_X\longrightarrow D^b_{({\infty})} {\operatorname{Coh}}_Y
\]
respectively.
If $W$ is another smooth projective variety and ${\mathcal{L}}\in D^b{\operatorname{Coh}}_{W\times X}$ is
another Fourier-Mukai kernel, then the composition
$F_{\mathcal K}\circ F_{\mathcal{L}}: D^b{\operatorname{Coh}}_W\to D^b{\operatorname{Coh}}_X$ is identified with $F_{{\mathcal K}*{\mathcal{L}}}$,
where the complex ${\mathcal K}*{\mathcal{L}}$, known as the {\em convolution} of ${\mathcal K}$ and ${\mathcal{L}}$, is defined as
\[
{\mathcal K}*{\mathcal{L}}\,=\, Rp_{WX*} (p_{XY}^*{\mathcal K} \otimes p_{WY}^*{\mathcal{L}}),
\]
with $p_{WX}, p_{XY}, p_{WY}$ are the projections of $W\times X\times Y$ to the pairwise
products. The classical Serre-Grothendieck duality implies
\cite[Lem. 1.2]{bondal-orlov}
that a left adjoint
of $F_{\mathcal K}$ in the sense of usual categories is provided by $F_{^\bigstar\hskip -0.05cm {\mathcal K}}$,
where
\[
^\bigstar \hskip -0.05cm {\mathcal K} \,=\, (\sigma^*{\mathcal K})\otimes p_X^* \Omega^n_X[n] \,\in\, D^b{\operatorname{Coh}}_{Y\times X},
\quad n=\dim(X), \,\, \sigma:Y\times X\to X\times Y,
\]
with $\sigma$ being the permutation. Further
\cite[Th. 3.1]{AL:Fourier} this theory provides a natural trace morphism
\[
{^\bigstar\hskip -0.05cm {\mathcal K}}*{\mathcal K} \longrightarrow {\mathcal O}_{\Delta_X},
\]
which induces the counit
for the adjoint pair $(F_{{^\bigstar\hskip -0.05cm {\mathcal K}}}, F_{\mathcal K})$.
Proposition \ref{prop:oo-unit} implies therefore the following
corollary which can be seen as an instance of \cite[Prop. 5.2.2.12]{lurie:htt}.
\begin{cor}
$(F_{{^\bigstar\hskip -0.05cm {\mathcal K}}}^{({\infty})}, F_{\mathcal K}^{({\infty})})$
form an adjoint pair of ${\infty}$-functors. \qed
\end{cor}
Therefore the theory of dg-enhanced adjoints for Fourier-Mukai functors
as developed in \cite{AL:Fourier}
and used in \cite{AL:spherical}, can be embedded into the
${\infty}$-categorical approach adopted in this paper.
\paragraph{ Bimodule dg-functors.} In a similar spirit,
let $\mathcal {A}$ be a small dg-category, and $\operatorname{Mod}_\mathcal {A}$ be the dg-category of right
$\mathcal {A}$-modules, i.e., of contravariant dg-functors $\mathcal {A}\to{\operatorname{dgVect}}_\mathbf k$.
The {\em derived category} of $\mathcal {A}$, denoted $D(\mathcal {A})$, is the localization of
$H^0 \operatorname{Mod}_\mathcal {A}$ (the homotopy category) by quasi-isomorphisms.
It is a triangulated category which can be represented as
\[
D(\mathcal {A}) \, \simeq \, H^0 D_{\operatorname{dg}}(\mathcal {A}) \,\=\, \operatorname{h}\! D_{({\infty})}(\mathcal {A}),
\]
where:
\begin{itemize}
\item $D_{\operatorname{dg}}(\mathcal {A})$ is the dg-enhancement of $D(\mathcal {A})$ constructed in
\cite{AL:bar} and denoted $\overline{\bf Mod}\text{-} \mathcal {A}$. It is defined
using the standard bar-resolution $\operatorname{Bar}(E)$ for
any $E\in\operatorname{Mod}_ {\EuScript A}$, that is
\[
{\operatorname{Hom}}^\bullet_{D_{\operatorname{dg}}(\mathcal {A})}(E,F) \, := \, {\operatorname{Hom}}^\bullet_{\operatorname{Mod}_\mathcal {A})}(\operatorname{Bar}(E), F).
\]
Thus closed degree $0$ morphisms in $D_{\operatorname{dg}}(\mathcal {A})$ are $A_{\infty}$-morphisms
of dg-modules.
\item $ D_{({\infty})}(\mathcal {A}) = N^{\operatorname{dg}} D_{\operatorname{dg}}(\mathcal {A})$ is the ${\infty}$-category obtained
as the dg-nerve of the dg-enhancement.
\end{itemize}
Let $\mathcal {A}, \mathcal{B}$ be two dg-categories. An $(\mathcal {A}, \mathcal{B})$-{\em bimodule}
is, by definition, a covariant dg-functor $\mathcal {A}\times \mathcal{B}^{\operatorname{op}} \to{\operatorname{dgVect}}_\mathbf k$. Any
$(\mathcal {A}, \mathcal{B})$-bimodule $M$
defines a dg-functor $F_M^{\operatorname{dg}}: D_{\operatorname{dg}}(\mathcal {A})\to D_{\operatorname{dg}}(\mathcal{B})$ by the derived
tensor product over $\mathcal {A}$, see \cite[Def. 3.9]{AL:bar}. It induces an
${\infty}$-functor $F_M^{({\infty})}: D_{({\infty})}(\mathcal {A})\to D_{({\infty})}(\mathcal{B})$ by passing
to dg--nerves and an exact functor of triangulated categories
$F_M: D(\mathcal {A})\to D(\mathcal{B})$ by passing to $H^0$-categories.
We define the {\em left and right transposes} of $M$ to be the
$(\mathcal{B}, \mathcal {A})$-bimodules whose values on the objects are
given by
\[
\begin{gathered}
^\bigstar \hskip -0.05cm M(b,a) \,=\,{\operatorname{Hom}}^\bullet_{D_{{\operatorname{dg}}}(\mathcal {A}^{\operatorname{op}})}
\bigl( M(-,b),\, {\operatorname{Hom}}^\bullet_\mathcal {A}(a,-)\bigr),\\
M^\bigstar(b,a)\,=\, {\operatorname{Hom}}^\bullet_{D_{{\operatorname{dg}}}(\mathcal{B})}
\bigl(M(a,-),\, {\operatorname{Hom}}^\bullet_\mathcal{B}(-,b)\bigr),
\end{gathered}
\]
cf. \cite[Def. 3.31]{AL:bar}. Theorem 1.1 of \cite{AL:bar} together with
Proposition \ref{prop:oo-unit} above, give the following fact.
\begin{prop}
(a) If $M$ is $\mathcal{B}$-perfect, then $F^{({\infty})}_{M^\bigstar}$ is a right
${\infty}$-categorical adjoint of $F^{({\infty})}_M$, and $F_{M^\bigstar}$
is a right adjoint of $F_M$ in the sense of usual categories.
\vskip .2cm
(b) If $M$ is $\mathcal {A}$-perfect, then $F^{({\infty})}_{^\bigstar \hskip -0.05cm M}$
is a left ${\infty}$-categorical adjoint of $F^{({\infty})}_M$, and
$F_{^\bigstar \hskip -0.05cm M}$ is a left adjoint of $F_M$ in the sense of
usual categories. \qed
\end{prop}
Therefore the theory of dg-enhanced adjoints for bimodule functors
developed in \cite{AL:bar}
can be embedded into the
${\infty}$-categorical approach adopted in this paper.
\paragraph{Homotopy coherent adjunctions. }
The concept of adjunction can give rise to a certain ambiguity
which is present already in the classical
case of functors
$\xymatrix{
A \ar@<.4ex>[r]^{F } &B \ar@<.4ex>[l]^{G }
}$
between usual categories. In this case the strongest and most natural adjunction
data for $(F,G)$ (a {\em coherent adjunction}) consists
\cite{maclane} of two transformations $u: \operatorname{Id}_A\to GF$ (unit) and $c: FG\to\operatorname{Id}_B$ (counit)
satisfying the following compatibility condition:
\begin{itemize}
\item[(AC)] The compositions
\[
F=F\circ \operatorname{Id}_A \buildrel F\circ u \over\longrightarrow F\circ G\circ F \buildrel c\circ F\over\longrightarrow \operatorname{Id}_B\circ F =F,
\quad
G=\operatorname{Id}_A\circ G \buildrel u\circ G\over\longrightarrow G\circ F\circ G \buildrel G\circ c\over\longrightarrow G\circ
\operatorname{Id}_B=G
\]
are equal to the identity transformations of $F$ and $G$ respectively.
\end{itemize}
This condition implies that for any objects $a\in A$, $b\in B$ the maps
\[
\xymatrix{
{\operatorname{Hom}}_B(F(a),b) \ar@<.4ex>[r]^{u^*_{a,b} } &{\operatorname{Hom}}_A(a,G(b))
\ar@<.4ex>[l]^{c_*^{a,b} }
}
\]
defined as the compositions (cf.
Proposition \ref{prop:oo-unit})
\[
\begin{gathered}
{\operatorname{Hom}}_B(F(a), b) \buildrel G_*\over\longrightarrow {\operatorname{Hom}}_A (GF(a), G(b)) \buildrel u^*\over\longrightarrow
{\operatorname{Hom}}_A(a, G(b)),
\\
{\operatorname{Hom}}_A(a, G(b)) \buildrel F_* \over\longrightarrow {\operatorname{Hom}}_B(F(a), FG(b)) \buildrel c_*\over\longrightarrow {\operatorname{Hom}}_B(F(a), b),
\end{gathered}
\]
are inverse to each other.
Now, a seemingly weaker datum would be that of a
single transformation, say $u$,
such that all the $u^*_{a,b}$ are bijections (or $c$ such that all the
$c_*^{a,b}$ are bijections). However, it is a classical
fact that any such partial datum can be always completed to
a pair $(u,c)$ satisfying (AC). So the possible ambiguity in the definition
of an adjunction is, in the case of usual categories, an inessential one.
\vskip .2cm
In the ${\infty}$-categorical case we have a similar ambiguity. To address it, one needs to
do two things:
\begin{itemize}
\item[(1)] Formulate an ${\infty}$-categorical analog of the condition (AC)
(i.e., define the stronger concept of a {\em homotopy coherent adjunction}).
\item[(2)] Compare this concept with the approach via biCartesian fibrations
over $\Delta^1$ discussed above.
\end{itemize}
\noindent This has been done in \cite{riehl-verity}. For completeness, let us
give some details of their analysis.
\vskip .2cm
Following \cite{auderset, schanuel-street}, let ${\bf{Adj}}^{\leq 2}$ be the {\em strict (classical)
$2$-category generated by the universal adjunction}. That is, ${\bf{Adj}}^{\leq 2}$ has
two objects $a$ and $b$ corresponding to indeterminate categories $A$ and $B$ above.
As a $2$-category, it is generated by
two $1$-morphisms $f: a\to b$ and $g: b\to a$ corresponding to indeterminate
functors $F,G$ above and two $2$-morphisms $u: \operatorname{Id}_a\to gf$ and $c: fg\to\operatorname{Id}_b$.
which are subject only to the two relations mimicking (AC). Geometrically, the
objects, $1$- and $2$-morphisms of ${\bf{Adj}}^{\leq 2}$ can be depicted as $0$- $1$-
and $2$-dimensional cells of the standard ``equatorial'' cell
decomposition of the $2$-sphere $S^2$, and the two relations in (AC) as two
$3$-cells which extend $S^2$ to a $3$-sphere $S^3$ for which $S^2$ is
the equator, see Fig. \ref{fig:2-sphere}.
By construction, a (coherent) adjunction of usual categories is the same
as a strict $2$-functor of strict $2$-categories ${\bf{Adj}}^{\leq 2} \to{\EuScript Cat}$.
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=0.4]
\node at (0,0){$\bullet$};
\node at (-1,0){$a$};
\node at (12,0){$\bullet$};
\node at (13,0){$b$};
\draw [dashed, decoration={markings,mark=at position 0.6 with
{\arrow[scale=1.5,>=stealth]{>}}},postaction={decorate},
line width=.2mm] (0,0) arc (180:0:6cm and 2cm);
\draw [ decoration={markings,mark=at position 0.6 with
{\arrow[scale=1.5,>=stealth]{>}}},postaction={decorate},
line width=.2mm] (12,0) arc (360:180:6cm and 2cm);
\draw (6,0) circle (6cm);
\node at (4.5,-2.7){$g$};
\node at (7.6, 2.8){$f$};
\node[rotate=45] at (1.4,1.7) {\huge $\Rightarrow$};
\node[rotate=45] at (1,2.3) {$u$};
\node[rotate=45] at (11, -2.2) {\huge $\Rightarrow$};
\node[rotate=45] at (10.4,-1.8) {$c$};
\end{tikzpicture}
\caption{Generating morphisms of ${\bf{Adj}}^{\leq 2}$ and cells of the $2$-sphere.}
\label{fig:2-sphere}
\end{figure}
\begin{ex}\label{ex:univ-monad}
The above definition of ${\bf{Adj}}^{\leq 2}$ is implicit, but one can describe it completely
explicitly \cite{auderset, schanuel-street}. For example,
the category
${\operatorname{Hom}}_{{\bf{Adj}}^{\leq 2}}(a,a)$ is isomorphic to $\Delta_+$, the augmented simplicial
category formed by all finite ordinals, including the empty one, so that the
ordinal $\{1,\cdots, n\}$ (taken to be $\emptyset$ for $n=0$)
corresponds to $(gf)^n$. Similarly,
${\operatorname{Hom}}_{{\bf{Adj}}^{\leq 2}}(b,b)$ is isomorphic to $\Delta^{\operatorname{op}}_+$, the opposite category,
with $\{1,\cdots, n\}$ corresponding to $(fg)^n$. This reflects the fact that
for an adjoint pair $(F,G)$ the composition $GF$ is a comonad, while $FG$
is a monad.
\end{ex}
As with any strict $2$-category, ${\bf{Adj}}^{\leq 2}$ gives rise to an $({\infty},2)$-category
$\operatorname{N}^{\operatorname{Hom}} {\bf{Adj}}^{\leq 2}$ by taking the nerves of all the ${\operatorname{Hom}}$-categories,
see Definition \ref{def:oo-2}.
\begin{defi}\cite{riehl-verity}
A {\em homotopy coherent adjunction of ${\infty}$-categories} is a functor of
simplicial categories $ N^{\operatorname{Hom}} {\bf{Adj}}^{\leq 2} \to{\EuScript Cat}_{\infty}^{(2)}$,
where ${\EuScript Cat}_{\infty}^{(2)}$ is the $({\infty},2)$-category of all ${\infty}$-categories,
see \S \ref{subset:stab-triang}\ref{par:oo-2-level}
\end{defi}
\begin{thm}\cite{riehl-verity}
Let $\xymatrix{
{\EuScript A} \ar@<.4ex>[r]^{F } & {\EuScript B} \ar@<.4ex>[l]^{G }
}$ be a pair of ${\infty}$-functors between ${\infty}$-categories. The following are equivalent:
\begin{itemize}
\item[(i)] $(F,G)$ is an adjoint pair.
\item[(ii)] There exists a homotopy coherent adjunction $N^{\operatorname{Hom}} {\bf{Adj}}^{\leq 2} \to{\EuScript Cat}_{\infty}^{(2)}$ sending $a,b$ to $ {\EuScript A}, {\EuScript B}$ and $f,g$ to $F,G$ respectively
\end{itemize}
\end{thm}
\noindent{\sl Proof:} (i)$\Rightarrow$(ii) is the content of \cite[Th. 4.3.9]{riehl-verity} taken
together with equivalence between the two definition of adjoints for ${\infty}$-functors:
that of \cite{lurie:htt} which we use and that of \cite{riehl-verity}. That equivalence
is established in \cite {riehl-verity:2-cat, riehl-verity:yoneda}, see the introduction
to \cite {riehl-verity}. The implication (ii)$\Rightarrow$(i) follows from Proposition
\ref{prop:oo-unit}.
\qed
\vskip .2cm
Therefore the possible ambiguity as to the notion of adjointness for ${\infty}$-categories
is also inessential.
\begin{rem}\label{rem:adj-sphere}
The geometric datum (Fig. \ref{fig:2-sphere}) used to define ${\bf{Adj}}^{\leq 2}$
has an analog for any $n$. This is the sphere $S^{n+1}$ with its equatorial
cell decomposition (two $p$-cells for each $0\leq p\leq n+1$). These cells
are oriented so that for any $p$ the two $p$-cells comprising $S^p\subset S^{n+1}$
induce the same orientation of $S^p$. This means that their ``directions''
are, intuitively, the opposite of each other, unlike the standard globe
diagram \cite{street:parity}, where these directions are the same.
It would be interesting to construct an $n$-category (in some sense)
${\bf{Adj}}^{\leq n}$ ``freely generated'' by the above $S^{n+1}$ so that the $p$-cells
for $p\leq n$ give generating $p$-morphisms, and the two $(n+1)$-cells
serve as relations. Intuitively, ${\bf{Adj}}^{\leq n}$ should describe adjunctions
for $(n-1)$-categories.
It is not clear whether one can construct ${\bf{Adj}}^{\leq n}$
as a classical strict $n$-category in the spirit of Street's orientals
\cite{street:orientals}, as this $S^{n+1}$ is not a loop-free pasting diagram
\cite{street:parity}.
A study of adjunctions for $2$-categories undertaken in
\cite{macdonald-stone}, leads to a construction of ${\bf{Adj}}^{\leq 3}$
as a Gray $3$-category (i.e., with $2$-dimensional associativity holding
up to a canonical ``braiding'' $3$-isomorphism). From the point of view of
Example \ref{ex:univ-monad}, the $2$-category ${\operatorname{Hom}}(a,a)$ in that ${\bf{Adj}}^{\leq 3}$
(the universal $2$-comonad) is a certain $2$-categorification of the category
$\Delta_+$ in which faces and degenerations become adjoint
$1$-morphisms \cite{macdonald-scull}.
A strict $2$-categorification of $\Delta_+$ with these
properties can be obtained \cite[\S 3.1.3]{dyckerhoff:DK} by viewing
each finite ordinal (a poset) as a category. It will be used below
in \S \ref {subsec:sph-admis}\ref{par:admis}
\end{rem}
\subsection{Spherical adjunctions of stable ${\infty}$-categories}\label{subsec:sphad}
\paragraph{Adjunctions of stable ${\infty}$-categories.}
By an {\em adjunction of stable $\infty$-categories}
we will mean an adjunction of ${\infty}$-categories, i.e., a biCartesian fibration
$p: X \to \Delta^1$,
such that both ${\infty}$-categories $ {\EuScript A}= p^{-1}(0)$ and $ {\EuScript B}= p^{-1}(1)$ are stable.
\begin{prop}
Let $\xymatrix{
{\EuScript A} \ar@<.4ex>[r]^{F } & {\EuScript B} \ar@<.4ex>[l]^{G },
}$ be an adjoint pair of ${\infty}$-functors with $ {\EuScript A}$ and $ {\EuScript B}$ being stable ${\infty}$-categories.
Then both $F$ and $G$ are exact functors.
\end{prop}
\noindent{\sl Proof:} By Proposition \ref{prop:adj-exact}, $F$ is left exact and $G$ is
right exact. But between stable ${\infty}$-categories, any left or right exact functor
is automatically exact, see \S\ref{subset:stab-triang}\ref{par:stable-oo} \qed
\paragraph{Twist and cotwist.}
Let $p: {\EuScript X}\to\Delta^1$ be an adjunction of stable ${\infty}$-categories and
$\xymatrix{
{\EuScript A} \ar@<.4ex>[r]^{F } & {\EuScript B} \ar@<.4ex>[l]^{G }
}$
be an adjoint pair of ${\infty}$-functors associated to $p$. By Proposition \ref {prop:oo-unit}
we can choose unit and counit transformations
\[
u: \operatorname{Id}_ {\EuScript A} \longrightarrow GF, \quad c: FG\longrightarrow \operatorname{Id}_ {\EuScript B}.
\]
By analogy with the pre-triangulated dg-setting \cite{AL:spherical}
we define the {\em twist} and {\em cotwist functors} associated
to the adjunction are as the cofiber of the unit and the fiber of the counit, see \eqref{eq:(co)fiber}:
\begin{equation}\label{eq:twist-as-cofiber}
T_ {\EuScript A}\,=\,{\operatorname{Cof}}(u): {\EuScript A} \longrightarrow {\EuScript A}, \quad T_ {\EuScript B} \,=\, \operatorname{Fib}(c): {\EuScript B}\longrightarrow {\EuScript B}.
\end{equation}
We will now provide a more implicit description of twist and cotwist based on universal properties
which will be used extensively lateron.
\begin{lem}\label{lem:twist} Let $p:X \to \Delta^1$ be an adjunction of stable $\infty$-categories. Let
\[
{\EuScript M} \subset \operatorname{Fun}(\Delta^1 \times \Delta^1 \amalg_{\Delta^1} \Delta^2, X)
\]
denote the full subcategory spanned by diagrams of the form
\[
\begin{tikzcd}
0 \ar{d} & \ar{l} a \ar{rd}{!}\ar{d} & \\
a'' & \ar{l} a' \ar{r}{*} & b
\end{tikzcd}
\]
where the right-hand triangle forms a diagram as in \eqref{eq:unitdiag} and the left-hand
square is a biCartesian square in $ {\EuScript A}$. Then the functor
\[
{\EuScript M} \longrightarrow {\EuScript A}
\]
given by restriction to $a$ is a trivial Kan fibration.
\end{lem}
\begin{proof} The diagrams in $ {\EuScript A}$ of the form
\begin{equation}\label{eq:rightkan}
\begin{tikzcd}
a \ar{d}\ar{r} & a'\\
0
\end{tikzcd}
\end{equation}
are precisely the right Kan extension of their restriction to $a \to a'$. Further, the
biCartesian squares in $ {\EuScript A}$ of the form
\begin{equation}\label{eq:bicart}
\begin{tikzcd}
a \ar{d}\ar{r} & a' \ar{d}\\
0 \ar{r} & a''
\end{tikzcd}
\end{equation}
are precisely the left Kan extensions of their restriction to \eqref{eq:rightkan}. Let ${\EuScript N}
\subset \operatorname{Fun}(\Delta^1 \times \Delta^1, {\EuScript A})$ denote the full subcategory spanned by the
biCartesian squares of the form \eqref{eq:bicart}. Applying \cite[4.3.2.15]{lurie:htt}
twice, we obtain that the restriction functor
\[
{\EuScript N} \longrightarrow \operatorname{Fun}(\Delta^1, {\EuScript A})
\]
given by restriction to the edge $a \to a'$ is a trivial Kan fibration. We have a pullback
square
\[
\begin{tikzcd}
{\EuScript M} \ar{r}\ar{d} & {\EuScript N} \ar{d} \\
{\EuScript V} \ar{r} & \operatorname{Fun}(\Delta^1, {\EuScript A})
\end{tikzcd}
\]
so that the map ${\EuScript M} \to {\EuScript V}$ is a trivial Kan fibration. We conclude the proof by Lemma
\ref{lem:unit}.
\end{proof}
Given an adjunction $p: {\EuScript X} \to \Delta^1$ of stable $\infty$-categories, we consider the
correspondence
\[
{\EuScript A} \overset{r}{\longleftarrow} {\EuScript M} \overset{s}{\longrightarrow} {\EuScript A}
\]
where $r$ is the trivial Kan fibration from Lemma \ref{lem:twist} and $s$ is the projection
to $a''$. We obtain a functor
\begin{equation}\label{eq:twist}
T_{ {\EuScript A}}: {\EuScript A} \longrightarrow {\EuScript A}
\end{equation}
as the composite of a section of $r$ with $s$. Comparing with \eqref{eq:twist-as-cofiber},
we see that
$T_{ {\EuScript A}}$ is the twist associated with the
adjunction (which justifies the same notation).
Indeed, it assigns to an object $a$ in $ {\EuScript A}$ a cofiber of a unit
morphism $a \to G(F(a))$ as constructed in Lemma \ref{lem:unit}.
Applying the same construction to $p^{{\operatorname{op}}}$, we obtain a functor
\begin{equation}\label{eq:cotwist}
T_{ {\EuScript B}}: {\EuScript B} \longrightarrow {\EuScript B},
\end{equation}
which is identified with the cotwist \eqref{eq:twist-as-cofiber},
as it associates to an object $b$ in $ {\EuScript B}$ a fiber of a counit morphism
$F(G(b)) \to b$.
\paragraph{Spherical adjunctions.}
\begin{defi}\label{def:sphad}
An adjunction of stable $\infty$-categories $p: X \to \Delta^1$ is called {\em
spherical} if its associated twist and cotwist are equivalences of $\infty$-categories. An adjoint pair $\xymatrix{
{\EuScript A} \ar@<.4ex>[r]^{F } & {\EuScript B} \ar@<.4ex>[l]^{G }
}$
associated to a spherical adjunction, will be called an {\em adjoint pair of
spherical ${\infty}$-functors}. An ${\infty}$-functor will be called {\em left spherical},
resp {\em right spherical}, if it is the respective part of an adjoint
pair $(F,G)$ of spherical ${\infty}$-functors.
\end{defi}
We will show later (Corollary \ref{cor:left-sph-right}) that the concepts
of a left and a right spherical functors are in fact equivalent, so the term
``spherical functor'' can be used unambiguously.
Further, the usual definition of
of a spherical functor, cf. \cite[Th. 1.1]{AL:spherical} includes
a requirement of the existence of one more adjoint, for example,
the right adjoint of $G$. We will show later (Corollary \ref{cor:sph-oo-adjoints})
that this requirement is also unnecessary.
\paragraph{Examples of spherical adjunctions.}
\begin{ex}[(``Classical'' dg-examples)]
Because of the comparison done in \S \ref{subsec:comparison}, all the
examples of spherical dg-functors between pre-triangulated dg-categories
that have been considered in the literature, translate into our framework.
Let us consider just two examples.
\vskip .2cm
(a) (Calabi-Yau fibrations) Let $X\buildrel p\over\to B$ be a smooth projective
morphism of algebric varieties over $\mathbf k$. Assume that all the fibers $X_b=p^{-1}(b)$,
$b\in B$
are $n$-dimensional Calabi-Yau varieties, i.e., $H^i(X_b, {\mathcal O})$ is
isomorphic to $\mathbf k_b$ (the residue field of the point $b$) for $i=0,n$ and is zero otherwise.
Then
\[
\xymatrix{
D^b_{({\infty})}{\operatorname{Coh}}_B \ar@<.4ex>[r]^{p^* } &D^b_{({\infty})}{\operatorname{Coh}}_X \ar@<.4ex>[l]^{Rp_* },
}
\]
form an adjoint pair of spherical ${\infty}$-functors.
\vskip .2cm
(b) (Sphere fibrations) For a CW-complex $Y$ we denote by $\on{Sh}_Y$ the
category of sheaves of $\mathbf k$-vector spaces on $Y$ and by $D^b_{({\infty})}\on{Sh}_Y$
the corresponding derived ${\infty}$-category. Let $S\buildrel p\over\to B$
be a locally trivial fibration of CW-complexes with fibers $S_b=p^{-1}(b)$
homeomorphic to the $n$-sphere $S^n$. Then
\[
\xymatrix{
D^b_{({\infty})}\on{Sh}_B \ar@<.4ex>[r]^{p^* } &D^b_{({\infty})}\on{Sh}_S \ar@<.4ex>[l]^{Rp_* },
}
\]
form an adjoint pair of spherical ${\infty}$-functors.
One can obtain a smaller example by restricting to full ${\infty}$-subcategories consisting of complexes
with locally constant cohomology sheaves, i.e., to local systems with values in the stable
${\infty}$-category $D^b_{({\infty})} {\on{Vect}}_\mathbf k$. This class of examples was generalized to local systems with
values in an arbitrary stable ${\infty}$-category in \cite{christ}.
\end{ex}
\begin{ex}[(Versions of the sphere example)]\label{ex:versions-sphere}
(a) (Version of the $S^0$-example) Let $ {\EuScript A}$ be any stable ${\infty}$-category.
Let $\delta: {\EuScript A}\to {\EuScript A}\times A$ be the diagonal embedding. Then
\[
\xymatrix{
{\EuScript A} \ar@<.4ex>[r]^{\hskip -0.5cm\delta } & {\EuScript A} \times {\EuScript A} \ar@<.4ex>[l]^{\hskip -0.5cm \oplus },
}
\]
form an adjoint pair of spherical ${\infty}$-functors. The twist $T_ {\EuScript A}$ is the identity,
while $T_{ {\EuScript A}\times {\EuScript A}}$ is the permutation functor.
\vskip .2cm
(b) (Version of the $S^1$-example) A local system of $\mathbf k$-vector spaces on $S^1$
is the same as a $\mathbf k[t,t^{-1}]$-module, and pullback of (trivial) local systems
under $S^1\to {\on{pt}}$ corresponds to the pullback, denote it ${\varepsilon}$,
of $\mathbf k$-modules under the
evaluation map $\mathbf k[t, t^{-1}]\to\mathbf k$, $t\mapsto 1$.
At the purely algebraic level, one can see directly that
\[
\xymatrix{
D^b_{({\infty})} {\on{Vect}}_\mathbf k \ar@<.4ex>[r]^{ \hskip -0.5cm {\varepsilon} } &D^b_{({\infty})}\operatorname{Mod}_{\mathbf k[t,t^{-1}]} \ar@<.4ex>[l]^{\hskip -0.5cm \eta },
}
\quad \eta (M) = R{\operatorname{Hom}}_{\mathbf k[t,t^{-1}]}(\mathbf k, M),
\]
form an adjoint pair of spherical ${\infty}$-functors. The functor $\eta$
can be calculated using the $2$-term free resolution $\{\mathbf k[t,t^{-1}]\buildrel t-1\over\to
\mathbf k[t,t^{-1}]\}$ of $\mathbf k$.
This gives
the twist on $D^b_{({\infty})} {\on{Vect}}_k$ to be identified with the shift
functor $V\mapsto V[-1]$, while the cotwist on $D^b_{({\infty})}\operatorname{Mod}_{\mathbf k[t,t^{-1}]}$ is identified with the identity.
We also have an essentially equivalent
version with the left adjoint $\delta$ of ${\varepsilon}$ (which corresponds to taking
homology instead of cohomology of local systems on $S^1$):
\[
\xymatrix{ D^b_{({\infty})}\operatorname{Mod}_{\mathbf k[t,t^{-1}]}
\ar@<.4ex>[r]^{ \hskip 0.5cm \delta } & D^b_{({\infty})} {\on{Vect}}_\mathbf k
\ar@<.4ex>[l]^{\hskip 0.5cm {\varepsilon} },
}
\quad \delta (M) = M\otimes^L_{\mathbf k[t,t^{-1}]}\mathbf k.
\]
Here the twist on $ D^b_{({\infty})}\operatorname{Mod}_{\mathbf k[t,t^{-1}]} $ is the identity while
the cotwist on $D^b_{({\infty})} {\on{Vect}}_\mathbf k$ is the functor $V\mapsto V[1]$.
\vskip .2cm
(c) (Global algebro-geometric version of (b)) Let $X$ be a smooth algebraic
variety over $\mathbf k$ and $i: Y\hookrightarrow X$ is an irreducible hypersurface, so that we have
the line bundle (sheaf of ideals)
${\mathcal O}_X(-Y)
\subset {\mathcal O}_X$ with ${\mathcal O}_X(-Y)|_Y = N^*_{Y/X}$, the conormal bundle.
Then
\[
\xymatrix{
D^b_{({\infty})}{\operatorname{Coh}}_X \ar@<.4ex>[r]^{\delta } & D^b_{({\infty})}{\operatorname{Coh}}_Y
\ar@<.4ex>[l]^{{\varepsilon}=i_* }
},
\quad \delta = (-)\otimes^L_{{\mathcal O}_X} {\mathcal O}_Y.
\]
form an adjoint pair of spherical ${\infty}$-functors.
The twist on $D^b_{({\infty})}{\operatorname{Coh}}_X$ is given by
tensoring with ${\mathcal O}_X(-Y)$,
while the cotwist on
$D^b_{({\infty})}{\operatorname{Coh}}_Y$ is given by tensoring with
$N^*_{Y/X}[1]$, see \cite{segal}.
\vskip .2cm
(d) (Graded $\mathbf k[t]$-modules) We will need yet another version of the above.
Let $ {\on{Vect}}_\mathbf k^\mathbb{Z}$ be the category of $\mathbb{Z}$-graded $\mathbf k$-vector spaces and
$\operatorname{Mod}_{\mathbf k[t]}^\mathbb{Z}$ be the category of $\mathbb{Z}$-graded $\mathbf k[t]$-modules, with $\deg(t)=1$.
Thus an object of $\operatorname{Mod}_{\mathbf k[t]}^\mathbb{Z}$ is a diagram of vector spaces and linear operators
\begin{equation}\label{eq:graded-k[t]}
M\,=\, \bigl\{ \cdots \buildrel t\over\longrightarrow M^{-1} \buildrel t\over\longrightarrow M^0 \buildrel t\over\longrightarrow M^1
\buildrel t\over\longrightarrow\cdots\bigr\}
\end{equation}
The embedding ${\varepsilon}: {\on{Vect}}_\mathbf k^\mathbb{Z}\hookrightarrow \operatorname{Mod}_{\mathbf k[t]}^\mathbb{Z}$ is the pullback
under the evaluation map $\mathbf k[t]\to \mathbf k$ at $t=0$. As in (b),
\[
\xymatrix{
{\EuScript A}=D^b_{({\infty})}\operatorname{Mod}_{\mathbf k[t]}^\mathbb{Z}
\ar@<.4ex>[r]^{ \delta } & D^b_{({\infty})} {\on{Vect}}^\mathbb{Z} _\mathbf k = {\EuScript B}
\ar@<.4ex>[l]^{ {\varepsilon} },
}
\quad \delta = (-) \otimes^L_{\mathbf k[t]}\mathbf k,
\]
form an adjoint pair of spherical ${\infty}$-functors.
Note that a graded $\mathbf k[t]$-module $M$ is $\delta$-flat, if $t$ is injective, i.e.,
\eqref{eq:graded-k[t]} represents a filtration. In this case $\delta (M)= {\operatorname{ gr}} (M)$
is the associated graded space of this filtration.
As before, calculating $\delta$ using the $2$-term graded
resolution $\{\mathbf k[t]\buildrel t\over\to \mathbf k[t]\}$ of $\mathbf k$ gives that
the twist $T_ {\EuScript A}$ is given by
the shift of grading of the graded $\mathbf k[t]$-modules $(T_ {\EuScript A} M)_i = M_{i-1}$,
The cotwist $T_ {\EuScript B}$
is given by $(T_ {\EuScript B}(V))_i=V_{i-1} [1]$.
\end{ex}
\subsection{Spherical adjunctions from Massey products in stable $\infty$-categories}
\label{sec:massey}
In this section we work out some particular examples of spherical adjunctions which
will be used later in the study of the relative Waldhausen $S$-construction.
Recall \cite{waldhausen:free} that the main idea behind the ``classical''
$S$-construction (for exact categories in the sense of Quillen) is to
rigidify the concept of a filtered object by including
all the quotients of the filtration into the set of data. Adapting this to
the context of stable ${\infty}$-categories and to filtrations of infinite length,
we arrive at the following concepts.
\paragraph{Filtered and graded objects.} Let $\Di$ be a stable ${\infty}$-category.
\begin{defi}
(a) A {\em filtered object} in $\Di$ is an ${\infty}$-functor $Z: (\mathbb{Z},\leq)\to\Di$, i.e.,
a diagram of objects and morphsims in $\Di$
\[
Z\,=\,\bigl\{ \cdots\to Z_i \longrightarrow Z_{i+1} \longrightarrow Z_{i+2} \to \cdots\bigr\}.
\]
(b) A {\em graded object} in $\Di$ is an ${\infty}$-functor from $\mathbb{Z}$, considered as
a discrete category, into $\Di$, i.e., simply a family of objects $A=(A_i)_{i\in\mathbb{Z}}$
in $\Di$.
\end{defi}
We have ${\infty}$-categories ${\operatorname{Filt}}(\Di)$ and ${\operatorname{Grad}}(\Di)$ formed by filtered and
graded objects in $\Di$. They are connected by a pair of ${\infty}$-functors
\[
\xymatrix{
{\operatorname{Filt}}(\Di) \ar@<.4ex>[r]^{ {\operatorname{ gr}}} & {\operatorname{Grad}}(\Di) \ar@<.4ex>[l]^{ {\varepsilon} },
}
\]
where ${\varepsilon}$ considers any graded object as a filtered object with zero
morphisms, and ${\operatorname{ gr}}$ is the functor of
``taking the associated graded":
\[
{\operatorname{ gr}}_i(Z) \,=\, {\operatorname{Cof}}\, \{Z_{i-1} \longrightarrow Z_i\}.
\]
The following is an analog of Example \ref{ex:versions-sphere}(d).
\begin{prop}\label{prop:filt-and-gr}
The pair $({\operatorname{ gr}}, {\varepsilon})$ is a spherical adjoint pair. The associated twist
$T_{{\operatorname{Filt}}(\Di)}$ is given by the shift of the filtration: $(T_{{\operatorname{Filt}}(\Di)}(Z))_i = Z_{i-1}$.
The cotwist
$T_{{\operatorname{Grad}}(\Di)}$ is given by $(T_{{\operatorname{Grad}}(\Di)}(A))_i = A_{i-1}[1]$.
\end{prop}
\noindent {\sl Proof:} The adjunction for $({\operatorname{ gr}}, {\varepsilon})$
is obtained by defining the unit and counit maps
\[
u_Z: Z \longrightarrow {\varepsilon} ({\operatorname{ gr}} (Z)), \quad c_A: {\operatorname{ gr}}({\varepsilon}(A))\longrightarrow A
\]
explicitly. That is,
$u_Z$ consists of the canonical maps $Z_i \to{\operatorname{Cof}} \{Z_{i-1}\to Z_i\}$,
while $c_A$ consists of the canonical maps ${\operatorname{Cof}}\{ A_{i-1}\buildrel 0\over\to A_i\} \to A_i$. The fact that the adjunction is spherical, follows by
straightforward calculation of the twist and cotwist and their identifications with what is
claimed in the proposition.
\qed
\paragraph{Filtered objects and chain complexes.}
In classical homological algebra, a filtration gives rise to
a kind of chain complex formed by the quotients of the filtration
and the $\operatorname{{Ext}}^1$-classes between successive quotients.
We adapt this point of view to our case as follows.
\vskip .2cm
Let $\mathbb{Z}^{\lhd}$ denote the poset obtained from $\mathbb{Z}$ by adjoining a minimal element which
we denote by $(-\infty)$. Let $P \subset \mathbb{Z}^\lhd \times \mathbb{Z}$ be the
subposet
consisting of those pairs $(i,j)$ which satisfy $i \le j$.
To avoid cluttering the
notation, in this section we will identify a poset with its nerve.
\begin{defi}\label{def:chain-cx}
Let $\Di$ be a stable $\infty$-category.
A {\em chain complex in $\Di$} is a functor
$X: P \to \Di$
satisfying the following conditions:
\begin{enumerate}
\item For every $i \in \mathbb{Z}^\lhd$, the object $X(i,i)$ is a zero object.
\item For every triple $(i,j,k)$ of elements of $\mathbb{Z}^\lhd$ with $i \leq k\leq j\leq l$, the square
\[
\begin{tikzcd}
X(i,j) \ar{r}\ar{d} & X(i,l) \ar{d}\\
X(k,j) \ar{r} & X(k,l)
\end{tikzcd}
\]
is biCartesian in $\Di$.
\end{enumerate}
We denote by $\operatorname{Ch}(\Di)$ the full subcategory of $\operatorname{Fun}(P,\Di)$ spanned by the chain
complexes.
\end{defi}
\begin{prop}\label{prop:filtered}
The ${\infty}$-functor
$\operatorname{Ch}(\Di)
\to {\operatorname{Filt}} (\Di)= \operatorname{Fun}(\mathbb{Z}, \Di)$
given by restriction along the inclusion of posets
\[
\alpha: \mathbb{Z} \hookrightarrow P, \; i \mapsto (-\infty, i),
\]
is an equivalence.
\end{prop}
\noindent{\sl Proof:}
Let us factor the embedding
$\alpha: \mathbb{Z}\hookrightarrow P$ into $\mathbb{Z}\buildrel\beta\over\hookrightarrow Q \buildrel
\gamma\over\hookrightarrow P$, where $Q$ is obtained by adding to
$\mathbb{Z}=\{(-{\infty}, i)\}$ elements of the form $(i,i)$, $i\in\mathbb{Z}$.
We can identify $\operatorname{Fun}(\mathbb{Z},\Di)=\operatorname{Fun}'(Q,\Di)$ where $\operatorname{Fun}'(Q,\Di)\subset
\operatorname{Fun}(Q,\Di)$ consists of functors equal to $0$ on each $(i,i)$.
One can seen this extension by $0$'s as an instance of $\beta_*$,
the right Kan
extension from $\mathbb{Z}$ to $Q$ (along $\beta$), since the undercategories appearing in the pointwise formulas, are empty.
Next consider $\gamma_!$, the left Kan extension from $Q$ to $P$
and its effect on functors from $\operatorname{Fun}'(Q,D)$. If $Z$ is such a functor,
then the pointwise formula gives the value of $(\gamma_!Z)(i,j)$
as the colimit over the overcategory
of $(i,j)$ in $Q$, and this colimit reduces
to
\[
\varinjlim \,\bigl\{ Z(i,i) \leftarrow Z(-{\infty},i) \to Z(-{\infty},j)\bigr\} \,=\,
{\operatorname{Cof}}\bigl\{ Z(-{\infty},i) \to Z(-{\infty},j)\bigr\},
\]
which is, naively, the ``quotient'' of the filtered object represented
by $Z$. It is clear that $\gamma_!Z$, i.e., the system of these ``quotients'', satisfies the
conditions of Definition \ref{def:chain-cx} so that $\operatorname{Ch}(\Di)$
consists precisely of the functors which are left Kan extensions
of those from $\operatorname{Fun}'(Q,\Di)$. Applying Proposition
\ref{prop:kan-ext-equiv} finishes the proof. \qed
\paragraph{Chain complexes as coherence data.}
To relate the data comprising a chain complex $X$ in a stable $\infty$-category to more traditional
terminology, we define the {\em $i$th term } of $X$ to be the object
$Y_i = X(i,i+1)[i]$.
The connecting morphisms of the exact triangles
\[
\begin{tikzcd}
X(i,i+1) \ar{r}\ar{d} & X(i,i+2) \ar{d} \ar{r} & 0 \ar{d}\\
X(i+1,i+1) \ar{r} & X(i+1,i+2) \ar{r} & X(i,i+1)[1]
\end{tikzcd}
\]
induce morphisms $d: Y_{i+1} \to Y_{i}$, called the {\em differentials} of $X$. To interpret the
remaining data, we introduce the subposet $P(n) \subset P$ consisting of those pairs $(i,j)$ of
integers satisfying $j - i \le n$. With this notation we have
\begin{enumerate}
\item The datum $X|P(1)$ is equivalent to the collection of terms $\{Y_i\}$ of the complex.
\item The datum $X|P(2)$ is equivalent to the sequence
\[
\dots \overset{d}{\to} Y_{i+1} \overset{d}{\to} Y_i \overset{d}{\to} Y_{i-1}
\overset{d}{\to} \dots
\]
of differentials.
\item The datum $X|P(3)$ is equivalent to the sequence of differentials together with a
homotopy $d^2 \simeq 0$ for every consecutive pair of differentials.
\item[($n$)] The datum $X|P(n)$ is equivalent to the sequence of differentials together with a
coherent system of homotopies $d^k \simeq 0$ for $2 \le k < n$. These
homotopies can be interpreted as trivializations of higher Massey products
associated to the sequence of differentials.
The triviality of these products is,
classically, the condition for the next level
Massey products being defined.
\end{enumerate}
\paragraph{Pseudo-complexes.}
We use the above analysis of chain complexes to
interpolate between filtered and graded objects.
\begin{defi}
Let $n \ge 0$. By a {\em pseudo-complex of order $n$} in
a stable ${\infty}$-category $\Di$ we mean a functor
$X: P(n) \to \Di$
satisfying:
\begin{enumerate}
\item For every $i \in \mathbb{Z}$, the object $X(i,i)$ is a zero object.
\item For every triple $(i,j,k)$ of integers with $i < j < k$ and $k-j \le n$, the square
\[
\begin{tikzcd}
X(i,j) \ar{r}\ar{d} & X(i,k) \ar{d}\\
X(j,j) \ar{r} & X(j,k)
\end{tikzcd}
\]
is biCartesian in $\Di$.
\end{enumerate}
We denote by
\[
\operatorname{Ch}^{(n)}(\Di) \subset \operatorname{Fun}(P(n), \Di)
\]
the full ${\infty}$-subcategory consisting of pseudo-complexes of order $n$.
\end{defi}
\begin{ex}
A pseudo-complex of order $1$ is the same as a graded object in $\Di$,
i.e., $\operatorname{Ch}^{(1)}(\Di)={\operatorname{Grad}}(\Di)$.
A chain complex can be formally considered as a pseudo-complex of order
${\infty}$.
\end{ex}
The following is a generalization of Proposition \ref{prop:filt-and-gr}.
\begin{prop}\label{prop:spherical}
Let $\Di$ be a stable $\infty$-category and $n \ge 0$.
Let $F: \operatorname{Ch}(\Di) \to \operatorname{Ch}^{(n)}(\Di)$ be the functor
given by restriction along $P(n) \subset P$. Then:
\begin{enumerate}
\item[(a)] After the identification of $\operatorname{Ch}(\Di)$ with ${\operatorname{Filt}}(\Di)$ given by Proposition
\ref{prop:filtered}, the functor $F$ associates to a filtered object
\[
\dots \to Z_{i-1} \to Z_i \to Z_{i+1} \to \dots
\]
the pseudo--complex $X: P(n) \to \Di$ with $X(i,j) = {\operatorname{Cof}}(Z_i \to Z_j)$.
\item[(b)] The functor $F$ has a right adjoint $G$ which associates to a pseudo--complex
$X : P(n) \to \Di$,
the (complex corresponding to the) filtered object
\[
\dots \to X(i-1,i-1+n) \to X(i,i+n) \to X(i+1,i+1+n) \to \dots.
\]
\item[(c)] The adjoint pair $(F,G)$ is spherical. The corresponding spherical twists
are given on objects by the formulas (in the first one we identify $\operatorname{Ch}(\Di)$ with ${\operatorname{Filt}}(\Di)$):
\[
\begin{gathered}
T_{\operatorname{Ch}(\Di)} = T_{{\operatorname{Filt}}(\Di)}: (Z_i)_{i \in \mathbb{Z}} \mapsto (Z_{i-n}[1])_{i \in \mathbb{Z}}
\\
T_{\operatorname{Ch}^{(n)}(\Di)}: (X_{(i,j)})_{(i,j) \in P(n)} \mapsto
(X_{(i-n,j-n)})_{(i,j) \in P(n)}.
\end{gathered}
\]
\end{enumerate}
\end{prop}
\noindent{\sl Proof:} Part (a) is immediate from the way the identification of Proposition
\ref{prop:filtered} is constructed.
\vskip .2cm
To prove (b),
let $p: {\EuScript X}=\Gamma(F) \to \Delta^1$ be the covariant Grothendieck construction of the functor $\Delta^1 \to
{\EuScript Cat}_{\infty}$ classifying $F$, see Example \ref{ex:groth-single}.
Let also ${\EuScript L}({\EuScript X}) = \operatorname{Fun}_{\Delta^1}(\Delta^1, {\EuScript X})$
be the $\infty$-category of sections of $p$. Thus an object of ${\EuScript L}({\EuScript X})$
is an edge $s$ of ${\EuScript X}$ covering the non-degenerate edge of $\Delta^1$.
\vskip .2cm
We use two lemmas. First,
we describe ${\EuScript L}({\EuScript X})$ directly,
using the description of $F$ in the form given in (a). More precisely,
denote by $Q(n) \subset P$ the subposet spanned by $P(n)$ together with the pairs $(-{\infty},i)$.
\begin{lem}\label{lem:groth-pseudocx}
${\EuScript L}({\EuScript X})$ is equivalent to the full subcategory of
$\operatorname{Fun}(Q(n),\Di)$ consisting of those diagrams
$X: Q(n) \to \Di$
whose restriction to $P(n)$ is a pseudo--complex.
\end{lem}
\noindent{\sl Proof of Lemma \ref{lem:groth-pseudocx}:}
By definition, see Example \ref{ex:groth-single}, an object of ${\EuScript L}({\EuScript X})$
is a datum of:
\begin{itemize}
\item[($1$)] A filtered object, denote it $Z$.
\item[($2$)] A pseudo-complex of order $n$, denote it $X'$.
\item[($3$)] A morphism ${\varphi}: F(Z)\to X'$ in $\operatorname{Ch}^{(n)}(\Di)$.
\end{itemize}
Let us compare this with the data contained in a functor $X:Q(n)\to\Di$ as in the lemma. Such an
$X$ gives
us, first of all:
\begin{itemize}
\item[($1'$)] A filtered object, denote it $Z$, represented by the restriction of $X$ to $\mathbb{Z}=\{(-{\infty},,i)\}$
\item[($2'$)] A pseudo-complex, denote it $X'$, represented by the restriction of $X$ to $P(n)$.
\end{itemize}
These data match (1) and (2) above. The remaning data contained in $X$, are the
morphisms $\xi_{ijk}: X(-{\infty},i) \to X(j,k)$ obtained as the values of $X$ on the morphisms
(inequalities) $(-{\infty},i) \leq (j,k)$ between $\mathbb{Z}$ and $Q(n)$. The compatibillity of the $\xi_{ijk}$
with the values of $X$ on $\Mor(Q(n))$ implies that they factor through the
\[
\xi_{ij}: X(\-{\infty},j) \longrightarrow X(i,j), \quad (i,j)\in Q(n).
\]
Since $X(i,i)=0$, we see that $\xi_{ij}$ factors through a morphism
\[
{\varphi}_{ij}: F(Z)_{ij} \,=\,{\operatorname{Cof}}\bigl\{ X(-{\infty},i) \to X(-{\infty},j)\bigr\} \longrightarrow X(i,j).
\]
These morphisms account for the data (3) above. In this way we identify ${\EuScript L}({\EuScript X})$ with
the category of functors $X$ as in the lemma, at the level of objects. The
remaining details are left to the reader. Lemma \ref{lem:groth-pseudocx} is proved.
\vskip .2cm
Next, we use the lemma just proved to describe (co)Cartesian sections (edges) of $p$.
\begin{lem}\label{lem:cart-edges-X}
Let $s$ be a section of $p$ and $X: Q(n)\to\Di$ be a functor corresponding to $s$
by Lemma \ref{lem:groth-pseudocx}. Then:
\begin{enumerate}
\item[(a)] $s$ is a coCartesian edge if and only if, for every $i < j \le i+n$,
the corresponding square
\[
\begin{tikzcd}
X(-{\infty},i ) \ar{r}\ar{d} & X(i, i) \ar{d} \\
X(-{\infty},j) \ar{r} & X(i,j)
\end{tikzcd}
\]
is biCartesian in $\Di$.
\item[(b)] $s$ is a Cartesian edge if and only if, for every $j \in \mathbb{Z}$, the
morphism $X(-{\infty} ,j) \to X(j-n,j)$
is an equivalence.
\end{enumerate}
\end{lem}
\noindent{\sl Proof of Lemma \ref{lem:cart-edges-X}:} (a) Recall the description of coCartesian
edges of $\Gamma(F)$ in Example
\ref{ex:groth-single}. Applying it to our case, we see that $s$ is coCartesian
if and only if the morphism $\alpha: F(Z)\to X'$
in the datum (3) in the proof of Lemma \ref{lem:groth-pseudocx}, is an equivalence.
This is just an alternative formulation of the condition in (a).
\vskip .2cm
(b) Let $s$ be represented by a filtered object $Z$, a pseudo-complex $X'$ and a
morphism $\alpha: F(Z)\to X'$, as above. Thus $s: Z\to X'$ as an edge of $\Gamma(F)$.
The condition that $s$ is Cartesian means that ${\EuScript X}_{/s} \to {\EuScript X}_{/X'}$ is a trivial Kan fibration, cf.
Example \ref{ex:implicit}(1). Let us consider this condition
at the level of objects. Objects of ${\EuScript X}_{/s}$ are triangles ($2$-simplices) in ${\EuScript X}$
\[
\xymatrix{
W\ar[d]_{\varphi} \ar[dr]^c&
\\
Z
\ar[r]_s&X'
}
\]
with $W\in {\operatorname{Filt}}(\Di)$, while objects of
${\EuScript X}_{/X'}$ are arrows $c$ which are a priori parts of such triangles. So the condition means,
in particular, that any $c$ can be completed to such a triangle. Now $c$, by definition of $\Gamma(F)$,
consists of a morphism, call it $\beta$, from $F(W)=({\operatorname{Cof}}\{W_i\to W_j\})_{(i,j)\in P(n)}$ to
$X'=(X_{ij})_{(i,j)\in P(n)}$. So the condition at the level of objects implies that for
any such $W,\beta$ we can find a morphism of filtered objects ${\varphi}: W\to Z$ fitting
into a triangle as above. Notice that we can take $W$ to consist of $W_j = X(j-n,j)$,
and this implies that the condition stated in part (b) of our lemma, is necessary.
The sufficiency follows easily. Lemma \ref{lem:cart-edges-X} is proved.
\vskip .2cm
The two above lemmas imply that $p$ is a biCartesian fibration and that the adjoint
functor $G$ of $F$ is as claimed in Proposition \ref{prop:spherical}(b), thus proving this part.
Finally, part (c) of Proposition \ref{prop:spherical}
now follows in a straightforward manner
by explicitly computing the spherical twists via the descriptions of the
Cartesian and coCartesian sections.
This finishes the proof of
Proposition \ref{prop:spherical}.
\section{Semiorthogonal decompositions}\label{sec:SOD}
In this chapter, we reformulate the notion of a spherical adjunction in terms of
semiorthogonal
decompositions and deduce various corollaries that arise from this beautiful point of
view, which was pioneered
in \cite{halpern-shipman}.
Our first goal will be to describe an approach to semiorthogonal decompositions
via stable
$\infty$-categories. We start by recalling the standard setup \cite{BK:SOD, bondal-orlov}.
\subsection{Semiorthogonal decompositions of triangulated categories}\label{subsec:SOD-triang}
Let $\mathcal{C}$ be a category.
A full subcategory $\mathcal{S}\subset\mathcal{C}$ will be called {\em strictly full}, if any object of $\mathcal{C}$
isomorphic to an object in $\mathcal{S}$, is in $\mathcal{S}$. Thus, every full subcategory $\mathcal {A}\subset\mathcal{C}$
has its {\em strictly full closure} $\mathcal {A}^{\operatorname{sf}}\supset \mathcal {A}$ which is the full subcategory consisting of all objects isomorphic to
objects of $\mathcal {A}$. Clearly, the embedding $\mathcal {A}\subset \mathcal {A}^{\operatorname{sf}}$ is an equivalence.
For this reason will sometimes not distinguish between full and stricly full subcategories.
\vskip .2cm
Let $\mathcal{C}$ be a triangulated category. Let $\mathcal{B}, \mathcal {A} \subset\mathcal{C}$ be full triangulated subcategories.
We say that $\mathcal{B}$ and $\mathcal {A}$ are {\em orthogonal}, if for any objects $b\in\mathcal{B}$
we have ${\operatorname{Hom}}_\mathcal{C}(b,a)=0$.
Suppose given one full triangulated subcategory $\mathcal {A}\subset \mathcal{C}$.
The {\em left} and {\em right orthogonals} of $\mathcal {A}$ in $\mathcal{C}$
are the full subcategories $^\perp \mathcal {A}, \mathcal {A}^\perp\subset\mathcal{C}$ whose objects
are defined as follows:
\begin{equation}\label{eq:orth-triang}
^\perp\mathcal {A}\,=\, \bigl\{ c\in\mathcal{C}\bigl| \, {\operatorname{Hom}}(c,a)=0,\,\,\forall \, a\in\mathcal {A}\bigr\},
\quad \mathcal {A}^\perp \,=\, \bigl\{ c\in\mathcal{C}\bigl| \, {\operatorname{Hom}}(a,c)=0,\,\,\forall \, a\in\mathcal {A}\bigr\}.
\end{equation}
They are strictly full
triangulated subcategories in $\mathcal {A}$.
\begin{defi}\label{def:SOD-triang}
An ordered pair $(\mathcal {A},\mathcal{B})$ of full triangulated subcategories of
$\mathcal{C}$ is called a {\em semiorthogonal decomposition} (or {\em SOD}, for short)
if
\begin{enumerate}
\item $\mathcal{B}$ and $\mathcal {A}$ are orthogonal.
\item for every object $x \in \mathcal{C}$, there exists a distinguished triangle
\begin{equation}\label{eq:extab}
b \longrightarrow x \longrightarrow a \longrightarrow b[1]
\end{equation}
with $a \in \mathcal {A}$ and $b \in \mathcal{B}$.
\end{enumerate}
\end{defi}
\noindent
Let $(\mathcal {A}, \mathcal{B})$ be an SOD of $\mathcal{T}$.
Denote by $\operatorname{EXT}(\mathcal {A}, \mathcal{B})$
the category of distinguished triangles in $\mathcal{C}$ of the form
\eqref{eq:extab}. We have the obvious
forgetful functor $\Xi: \operatorname{EXT}(\mathcal {A},\mathcal{B}) \to \mathcal{C}$. Condition (2)
implies that this functor is essentially surjective. Condition (1) further implies, that $\Xi$ induces
a bijection on isomorphism classes of objects. Therefore, $\Xi$ is very close to being an equivalence
of categories, but we cannot quite deduce that:
the attempt to construct an inverse to $\Xi$ fails due
to the non-functoriality of cones.
\vskip .2cm
It is straightforward to show that, if $(\mathcal {A},\mathcal{B})$ forms an SOD of $\mathcal{C}$, then
$\mathcal{B}$ is, essentially, uniquely determined by $ {\EuScript A}$:
the embedding $\mathcal{B}\subset {{}^\perp\mathcal {A}}$ is an
equivalence of categories, so $^\perp\mathcal {A}$ is the strictly full closure of $\mathcal{B}$.
Similarly, $\mathcal {A}\subset\mathcal{B}^\perp$ is an equivalence, so $\mathcal{B}^\perp$ is the strictly full closure
of $\mathcal {A}$.
\vskip .2cm
Vice versa, we can start with any full
triangulated subcategory $\mathcal {A} \subset \mathcal{C}$ and ask when the ordered pair $( {\EuScript A},{^\perp} {\EuScript A})$
is n SOD of $\mathcal{C}$. The condition for that is that
the inclusion functor $i: \mathcal {A} \to \mathcal{C}$ has a left
adjoint. The adjoint in that case associates to any $x\in\mathcal{C}$ the $a$-term in the triangle \eqref{eq:extab}, this term being (in that case) functorial in $x$.
We call such subcategories $\mathcal {A} \subset \mathcal{C}$ {\em left admissible}.
Dually, the condition that $i$ has a right adjoint, means that $(\mathcal {A}^\perp, \mathcal {A})$
is an SOD of $\mathcal{C}$. In this case we call $\mathcal {A}$
{\em right admissible}.
If $i$ has both a left and a
right adjoint, in which case $\mathcal {A} \subset \mathcal{C}$ is called {\em admissible}, then the restriction of a right
adjoint to ${^\perp} {\EuScript A}$ defines a functor
\[
g: {^\perp} {\EuScript A} \to {\EuScript A}.
\]
A calculation similar to the above analysis of $p$ shows that, morally, we ought to be able to
recover the triangulated category $\mathcal{C}$ from the triple $( {\EuScript A},{^\perp} {\EuScript A}, g)$. However, in trying to
prove this statement, we run into obstacles of the same kind as above.
\vskip .2cm
In conclusion, the axiomatic framework of triangulated categories does not allow for a fully
satisfying framework of semiorthogonal decompositions. It is well-known that the above issues can be resolved by passing to a suitable context of enhanced triangulated categories.
See \cite{kuznetsov-lunts} for a treatment using dg-enhancements.
Our choice here is to
use stable $\infty$-categories.
\subsection{Semiorthogonal decompositions of stable ${\infty}$-categories and gluing functors}
\label{sec:semi}
\paragraph{Orthogonal subcategories.}
Let ${\EuScript C}$ be an ${\infty}$-category. A full ${\infty}$-subcategory ${\EuScript S}\subset{\EuScript C}$ will be called
{\em strictly full}, if any object of ${\EuScript C}$ equivalent to an object of ${\EuScript S}$, is in ${\EuScript S}$.
As in the case of the usual categories in \S \ref{subsec:SOD-triang}, any full
${\infty}$-subcategory $ {\EuScript A}\subset {\EuScript C}$ has its strictly full closure $ {\EuScript A}^{\operatorname{sf}}\subset {\EuScript C}$ which is the
full ${\infty}$-subcategory
spanned by all objects equivalent to objects of $ {\EuScript A}$. Thus $ {\EuScript A} \subset {\EuScript A}^{\operatorname{sf}}$
is an equivalence of ${\infty}$-categories. For this reason we will sometimes not distinguish
between full and strictly full ${\infty}$-subcategories.
\vskip .2cm
A full subcategory $ {\EuScript A}$ of a stable $\infty$-category ${\EuScript C}$ is called {\em stable} if it contains a zero
object and is closed under forming fibers and cofibers. Such an $ {\EuScript A}$ is, in particular, a stable
${\infty}$-category in the intrinsic sense. In particular, $\operatorname{h}\! {\EuScript A}$ is a full triangulated subcategory
in the triangulated category $\operatorname{h}\!{\EuScript C}$.
\vskip .2cm
Let ${\EuScript C}$ be a stable ${\infty}$-category. Let $ {\EuScript B}, {\EuScript A}\subset{\EuScript C}$ be two stable subcategories.
We say that $ {\EuScript B}$ and $ {\EuScript A}$ are {\em orthogonal} if for any objects $b\in {\EuScript B}$, $a\in {\EuScript A}$,
the space $\operatorname{Map}_{\EuScript C}(b,a)$ is contractible. Note that this is equivalent to the condition
that the subcategories $\operatorname{h}\! {\EuScript B}$ and $\operatorname{h}\! {\EuScript A}$ in the triangulated category $\operatorname{h}\!{\EuScript C}$
are orthogonal in the sense of \S \ref{subsec:SOD-triang}.
\vskip .2cm
For a stable subcategory $ {\EuScript A}\subset {\EuScript C}$ we define its {\em left} and {\em right orthogonals} to be
the full ${\infty}$-subcategories $^\perp {\EuScript A}, {\EuScript A}^\perp\subset{\EuScript C}$ whose objects are specified by:
\[
\begin{gathered}
^\perp {\EuScript A}\,=\, \bigl\{ c\in {\EuScript C}\bigl| \, \operatorname{Map}_{\EuScript C}(c,a) \text{ is contractible, }\,\,\forall \, a\in\mathcal {A}\bigr\},
\\
{\EuScript A}^\perp \,=\, \bigl\{ c\in {\EuScript C}\bigl| \, \operatorname{Map}_{\EuScript C}(a,c) \text{ is contractible, }\,\,\forall \, a\in\mathcal {A}\bigr\}.
\end{gathered}
\]
They are strictly full stable subcategories in ${\EuScript C}$. Their associated triangulated subcategories
in $\operatorname{h}\!{\EuScript C}$ can be identified as follows:
\[
\operatorname{h}\!\, ({^\perp {\EuScript A}})\,=\, ^\perp(\operatorname{h}\! {\EuScript A}),\quad \operatorname{h}\!( {\EuScript A}^\perp) \,=\, (\operatorname{h}\! {\EuScript A})^\perp,
\]
where the orthogonals in the right hand sides are understood as in \eqref{eq:orth-triang}.
\paragraph{The ${\infty}$-categories $\co{ {\EuScript A}, {\EuScript B}}$ and $\arr{ {\EuScript A}, {\EuScript B}}$.}
Let ${\EuScript C}$ be a stable $\infty$-category and
let $ {\EuScript A}, {\EuScript B}$ be stable subcategories of ${\EuScript C}$. We denote by
\[
\co{ {\EuScript A}, {\EuScript B}} \subset \operatorname{Fun}(\Delta^1 \times \Delta^1, {\EuScript C})
\]
the full subcategory spanned by biCartesian squares in ${\EuScript C}$ of the form
\begin{equation}\label{eq:semidiag}
\begin{tikzcd}
b\arrow{r} \arrow{d} & x \arrow{d}\\
0 \arrow{r} &a
\end{tikzcd}
\end{equation}
such that $a$ is an object of $ {\EuScript A}$, $b$ is an object of $ {\EuScript B}$, and $0$ denotes a zero object in ${\EuScript C}$. We have the ${\infty}$-functor $\xi: \co{ {\EuScript A}, {\EuScript B}}\to {\EuScript C}$ given by the projection to the vertex $x$.
\vskip .2cm
Let further
\[
\arr{ {\EuScript A}, {\EuScript B}} \subset \operatorname{Fun}(\Delta^1, {\EuScript C})
\]
denote the full subcategory spanned by edges $a \to b$ in ${\EuScript C}$ with $a$ in $ {\EuScript A}$ and $b$ in
$ {\EuScript B}$. We have the ${\infty}$-functor $ \operatorname{Fib}: \arr{ {\EuScript A}, {\EuScript B}}\to{\EuScript C}$ which associated to a morphism
its fiber.
\begin{prop}\label{prop:<AB>}
(a) The ${\infty}$-categories $\co{A, {\EuScript B}}$ and $\arr{ {\EuScript A}, {\EuScript B}}$ are equivalent.
\vskip .2cm
(b) The equivalence can be chosen to as to send the ${\infty}$-functor $\xi$ to the ${\infty}$-functor
$ \operatorname{Fib}$.
\vskip .2cm
(c) Assume that $ {\EuScript B}, {\EuScript A}$ are orthogonal. Then the ${\infty}$-functor in (b) is fully faithful, i.e., it defines an equivalence of $\co{ {\EuScript A}, {\EuScript B}}\simeq
\arr{ {\EuScript A}, {\EuScript B}}$ with a full ${\infty}$-subcategory in ${\EuScript C}$.
\end{prop}
\noindent{\sl Proof:}
Consider the full subcategory
\[
{\EuScript M} \subset \operatorname{Fun}(\Delta^1 \times \Delta^2, {\EuScript C})
\]
spanned by diagrams
\[
\begin{tikzcd}
b\arrow{r} \arrow{d} & x \arrow{d} \arrow{r} & \arrow{d} 0\\
0 \arrow{r} & a \arrow{r} & b'
\end{tikzcd}
\]
with both squares biCartesian, $a$ in $ {\EuScript A}$, and $b, b'$ in $ {\EuScript B}$. The functors
\[
\co{ {\EuScript A}, {\EuScript B}} \longleftarrow {\EuScript M} \longrightarrow \arr{ {\EuScript A}, {\EuScript B}}
\]
given by projecting to $x$ and $a \to b'$, respectively, are equivalences. The functor
$ \operatorname{Fib}$ can be identified with the composite
\[
\arr{ {\EuScript A}, {\EuScript B}} \overset{\simeq}{\longrightarrow} \co{ {\EuScript A}, {\EuScript B}}\buildrel\xi\over \longrightarrow {\EuScript C}
\]
which proves (a) and (b).
\vskip .2cm
Let us prove (c) in the following form: the functor $ \operatorname{Fib}: \arr{ {\EuScript A}, {\EuScript B}}\to{\EuScript C}$
induces weak equivalences on the $\operatorname{Map}$-spaces, i.e., for any two objects
$a\buildrel f\over\to b$ and $a'\buildrel f'\over\to b'$ of $ \arr{ {\EuScript A}, {\EuScript B}}$ the map
\[
\operatorname{Map}_{\arr{ {\EuScript A}, {\EuScript B}}}(f, f') \longrightarrow \operatorname{Map}_{\EuScript C}( \operatorname{Fib}(f), \operatorname{Fib}(f'))
\]
is a weak equivalence. Note that we have a weak equivalence
\[
\operatorname{Fib}\{a\buildrel f\over\longrightarrow b\}\, \simeq\, {\operatorname{Cof}}\{a[-1]\buildrel f[-1]\over\longrightarrow b[-1]\}.
\]
Therefore we have
weak equivalences
\[
\begin{gathered}
\operatorname{Map}_{\EuScript C}( \operatorname{Fib}(f), \operatorname{Fib}(f')) \,\simeq\, \operatorname{Map}_{\EuScript C}({\operatorname{Cof}}(f[-1]), \operatorname{Fib}(f')) \,\simeq\,
\\
\operatorname{Fib}^T\biggl\{ \operatorname{Fib}^T\bigl\{\operatorname{Map}_{\EuScript C}(b[-1],a') \to \operatorname{Map}_{\EuScript C}(b[-1], b')\bigr\} \longrightarrow
\operatorname{Fib}^T\bigl\{ \operatorname{Map}_{\EuScript C}(a[-1], a') \to\operatorname{Map}_{\EuScript C} (a[-1], b')\bigr\}\biggr\}
\,\simeq \,
\\
\operatorname{Fib}^T\biggl\{ \operatorname{Fib}^T\bigl\{\operatorname{Map}_{\EuScript C}(b,a') \to \operatorname{Map}_{\EuScript C}(b, b')\bigr\} \longrightarrow
\operatorname{Fib}^T\bigl\{ \operatorname{Map}_{\EuScript C}(a, a') \to\operatorname{Map}_{\EuScript C} (a, b')\bigr\}\biggr\},
\end{gathered}
\]
where $ \operatorname{Fib}^T$ means the homotopy fiber in the category of spaces.
Now, since the $\operatorname{Map}_{\EuScript C}(b,a')$ is contractible, the last iterated homotopy fiber above
can be identified with the
homotopy fiber product
\[
\operatorname{Map}_{\EuScript C}(b,b') \times^h_{\operatorname{Map}_{\EuScript C}(a,b')} \operatorname{Map}_{\EuScript C}(a,a') \,\simeq \,
\operatorname{Map}_{\arr{ {\EuScript A}, {\EuScript B}}}(f, f').
\]
This chain of identifications is clearly compatible with the one induced by the functor $ \operatorname{Fib}$.
This proves the statement (c). \qed
\paragraph{Semiorthogonal decompositions.}
\begin{defi}\label{def:oo-SOD}
Let ${\EuScript C}$ be a stable $\infty$-category. We say that an ordered pair $( {\EuScript A}, {\EuScript B})$ of full
stable subcategories of ${\EuScript C}$ forms a {\em semiorthogonal decomposition}
(or an {\em SOD}, for short) of ${\EuScript C}$, if the
functor $\xi: \co{ {\EuScript A}, {\EuScript B}} \to {\EuScript C}$ (or, equivalently, the functor
$ \operatorname{Fib}: \arr{ {\EuScript A}, {\EuScript B}}\to{\EuScript C}$) is an equivalence of $\infty$-categories.
\end{defi}
\begin{prop}\label{prop:perp}
Suppose $ {\EuScript A}$ and $ {\EuScript B}$ are strictly full stable ${\infty}$-subcategories in ${\EuScript C}$.
Suppose further that $( {\EuScript A}, {\EuScript B})$ forms a semiorthogonal decomposition
of a stable $\infty$-category ${\EuScript C}$.
Then we have $ {\EuScript B} = {^\perp} {\EuScript A}$ and $ {\EuScript A} = {\EuScript B}^{\perp}$.
\end{prop}
\noindent{\sl Proof:} Under the equivalence $ \operatorname{Fib}: \arr{ {\EuScript A}, {\EuScript B}} \to {\EuScript C}$, the subcategory
$ {\EuScript A}$ (resp. $ {\EuScript B}$) of ${\EuScript C}$ is identified with the subcategory of $\arr{ {\EuScript A}, {\EuScript B}}$ given by
arrows of the form $a \to 0$ (resp. $0 \to b$). The mapping space $\operatorname{Map}_{\arr{ {\EuScript A}, {\EuScript B}}}(0 \to
b', a \to b)$ is equivalent to $\operatorname{Map}_{ {\EuScript B}}(b',b)$ which is contractible for every $b'$ in
$ {\EuScript B}$ if and only if $b$ is a zero object. This implies $ {\EuScript A} = {\EuScript B}^{\perp}$ and the first claim
follows analogously.
\qed
\begin{cor}\label{SOD:oo.vs.triang}
A pair of stable subcategories
$( {\EuScript A}, {\EuScript B})$ form an SOD of a stable ${\infty}$-category ${\EuScript C}$ if and only if the pair of full triangulated subcategories
$(\operatorname{h}\! {\EuScript A}, \operatorname{h}\! {\EuScript B})$ form an SOD of the triangulated category $\operatorname{h}\!{\EuScript C}$ in the sense of
\S \ref{subsec:SOD-triang}.
\end{cor}
The corollary means that we have a bijection between semiorthogonal decompositions of ${\EuScript C}$
as a stable ${\infty}$-category and of $\operatorname{h}\!{\EuScript C}$ as a triangulated category.
\vskip .2cm
\noindent{\sl Proof:} ``If'': Suppose that $(\operatorname{h}\! {\EuScript A}, \operatorname{h}\! {\EuScript B})$ form an SOD of $\operatorname{h}\!{\EuScript C}$. Then
$\operatorname{h}\! {\EuScript B}$ and $\operatorname{h}\! {\EuScript A}$ are orthogonal which implies that $ {\EuScript B}$ and $ {\EuScript A}$ are orthogonal.
Therefore, by Proposition \ref{prop:<AB>}(c)
the ${\infty}$-functor $\xi$ is fully faithful and its essential image consists of
objects $x\in{\EuScript C}$ which include into a biCartesian square of the form
\eqref{eq:semidiag}. But by definition of the triangulated structure in $\operatorname{h}\!{\EuScript C}$,
existence of such a square is equivalent to existence of a distinguished triangle in $\operatorname{h}\!{\EuScript C}$
of the form \eqref{eq:extab}. Since $(\operatorname{h}\! {\EuScript A}, \operatorname{h}\! {\EuScript B})$ form an SOD of $\operatorname{h}\!{\EuScript C}$, such a triangle
does exist for any $x\in \operatorname{Ob}(\operatorname{h}\!{\EuScript C}) = \operatorname{Ob}({\EuScript C})$. Therefore a square \eqref{eq:semidiag}
exists as well. This means that the essential image of $\xi$ is all of ${\EuScript C}$, i.e.,
that $\xi$ is an equivalence.
\vskip .2cm
``Only if'': Suppose that $\xi$ is an equivalence. Then by Proposition \ref{prop:perp},
$ {\EuScript B}$ and $ {\EuScript A}$ are orthogonal, which means that $\operatorname{h}\! {\EuScript B}$ and $\operatorname{h}\! {\EuScript A}$ are orthogonal
in $\operatorname{h}\!{\EuScript C}$, thus fulfilling condition (1) of Definition \ref{def:SOD-triang}.
Further, essential surjectivity of $\xi$ together with the definition of
triangulated structure in $\operatorname{h}\!{\EuScript C}$, imply that condition (2) of the same definition
holds as well. \qed
\paragraph{Cartesian and coCartesian SODs.}
Let ${\EuScript C}$ be a stable $\infty$-category, equipped with a semiorthogonal decomposition $( {\EuScript A}, {\EuScript B})$. We
introduce a simplicial set $\chi( {\EuScript A}, {\EuScript B})$ with a map $p: \chi( {\EuScript A}, {\EuScript B}) \to \Delta^1$ as follows: An
$n$-simplex of $\chi( {\EuScript A}, {\EuScript B})$ consists of
\begin{enumerate}
\item[(1)] an $n$-simplex $j: [n] \to [1]$ of $\Delta^1$,
\item[(2)] an $n$-simplex $\sigma: \Delta^n \to {\EuScript C}$ such that
$\sigma|\Delta^{j^{-1}(0)} \subset
{\EuScript A}$ and $\sigma|\Delta^{j^{-1}(1)} \subset {\EuScript B}$.
\end{enumerate}
The map $p$ is the forgetful map to part (1) of the data.
By construction,
$p^{-1}(0)= {\EuScript A}$ and $p^{-1}(1)= {\EuScript B}$.
It follows immediately from the inner horn filling property of
${\EuScript C}$ that $\chi( {\EuScript A}, {\EuScript B})$ is an ${\infty}$-category (and this implies that $p$ it is an
inner fibration \cite[Prop.2.3.1.5]{lurie:htt}).
\vskip .2cm
The the $\infty$-category of sections of $p$ is identified as follows.
\begin{prop}\label{prop:sections}
There exists a canonical equivalence
\[
\operatorname{Fun}_{\Delta^1}(\Delta^1, \chi( {\EuScript A}, {\EuScript B})) \,\simeq \, {\EuScript C}.
\]
\end{prop}
\noindent{\sl Proof:}
Notice that $\operatorname{Fun}_{\Delta^1}(\Delta^1, \chi( {\EuScript A}, {\EuScript B}))$ is isomorphic as a simplicial
set to the $\infty$-category $\arr{ {\EuScript A}, {\EuScript B}}$. See Proposition \ref{prop:<AB>}(a)
where it is proved that $\arr{ {\EuScript A}, {\EuScript B}}$ is equivalent to the ${\infty}$-category
$\co{ {\EuScript A}, {\EuScript B}}$ so ${\EuScript C}$ is equivalent to them both by the definition of an SOD. \qed
\begin{defi}\label{def:Cart-SOD}
A semiorthogonal decomposition $( {\EuScript A}, {\EuScript B})$ of a stable ${\infty}$-category ${\EuScript C}$
is called {\em Cartesian} (resp. {\em coCartesian}), if the ${\infty}$-functor
$p: \chi( {\EuScript A}, {\EuScript B})\to\Delta^1$ is a Cartesian (resp. coCartesian) fibration.
\end{defi}
Recall (\S \ref{def:cart-fib}) that the condition of $p$ being a (co)Cartesian fibration is
formulated in terms of $p$-(co)Cartesian edges of $\chi( {\EuScript A}, {\EuScript B})$. Let us reformulate
this condition using the specifics of our case.
\begin{defi}\label{def:(A,B)-cart}
Let ${\EuScript C}$ be a stable $\infty$-category equipped with a semiorthogonal decomposition
$( {\EuScript A}, {\EuScript B})$. An edge $e: a \to b$ in ${\EuScript C}$ with $a$ in $ {\EuScript A}$ and $b$ in $ {\EuScript B}$ is called
\begin{enumerate}
\item[(1)] $( {\EuScript A}, {\EuScript B})$-{\em Cartesian} if $e$ defines a final object of the $\infty$-category
${\EuScript C}_{/b} \times_{{\EuScript C}} {\EuScript A}$.
\item[(2)] $( {\EuScript A}, {\EuScript B})$-{\em coCartesian} if $e$ defines an initial object of the $\infty$-category
${\EuScript C}_{a/} \times_{{\EuScript C}} {\EuScript B}$.
\end{enumerate}
\end{defi}
Note that an edge $e: a \to b$ in ${\EuScript C}$ with $a$ in $ {\EuScript A}$ and $b$ in $ {\EuScript B}$ is the same as
as an edge of $\chi( {\EuScript A}, {\EuScript B})$ covering the nontrivial edge of $\Delta^1$.
\begin{prop}\label{prop:cartesian}
Let ${\EuScript C}$ be a stable $\infty$-category equipped with a semiorthogonal decomposition
$( {\EuScript A}, {\EuScript B})$. Let $e: a \to b$ in ${\EuScript C}$ be an edge with $a$ in $ {\EuScript A}$ and $b$ in $ {\EuScript B}$. Then the
following are equivalent:
\begin{itemize}
\item[(i)] $e$ is $( {\EuScript A}, {\EuScript B})$-Cartesian (resp. $( {\EuScript A}, {\EuScript B})$-coCartesian).
\item[(ii)] $e$, considered as an edge of $\chi( {\EuScript A}, {\EuScript B})$, is $p$-Cartesian (resp.
$p$-coCartesian).
\end{itemize}
\end{prop}
\noindent{\sl Proof:} We prove the equivalence of the two versions of ``Cartesian''.
The coCartesian case follows by duality.
We need to prove that $e$ is $p$-Cartesian if and only if the
corresponding object in ${\EuScript C}_{/b} \times_{{\EuScript C}} {\EuScript A}$ is a final object. The edge $e$ being
$p$-Cartesian is equivalent to the map
\[
\chi( {\EuScript A}, {\EuScript B})_{/e} \longrightarrow \chi( {\EuScript A}, {\EuScript B})_{/b} \times_{\Delta^1_{/1}} \Delta^1_{/\{0 \to 1\}}
\]
being a trivial Kan fibration. Unravelling the definition, this map idenitifies with the map
\[
{\EuScript C}_{/e} \times_{{\EuScript C}} {\EuScript A} \longrightarrow {\EuScript C}_{/b} \times_{{\EuScript C}} {\EuScript A}
\]
which is a trivial Kan fibration if and only if $e$ defines a final object in ${\EuScript C}_{/b}
\times_{{\EuScript C}} {\EuScript A}$. \qed
\begin{cor}
A semiorthogonal decomposition $( {\EuScript A}, {\EuScript B})$ is Cartesian
if and only if for every object $b$ of $ {\EuScript B}$, there exists an
$( {\EuScript A}, {\EuScript B})$-Cartesian edge $a \to b$. Dually, $( {\EuScript A}, {\EuScript B})$ is coCartesian
if and only if for every object $a$ of $ {\EuScript A}$, there exists an $( {\EuScript A}, {\EuScript B})$-coCartesian edge $a \to b$. \qed
\end{cor}
\paragraph{Gluing functors.}
Let ${\EuScript C}$ be a stable $\infty$-category with Cartesian semiorthogonal decomposition
$( {\EuScript A}, {\EuScript B})$. The Cartesian fibration $p: \chi( {\EuScript A}, {\EuScript B}) \to \Delta^1$ corresponds uniquely to an ${\infty}$-functor
$G: {\EuScript B} \to {\EuScript A}$ so that $\chi( {\EuScript A}, {\EuScript B})$ is identified with the contravariant Grothendieck
construction $\chi(G)$. See Example \ref{ex:implicit}.
We call $G$ a {\em Cartesian gluing functor} of $( {\EuScript A}, {\EuScript B})$.
Dually, given a coCartesian semiorthogonal
decomposition $( {\EuScript A}, {\EuScript B})$, we similarly obtain an ${\infty}$-functor $F: {\EuScript A} \to {\EuScript B}$ so that $\chi( {\EuScript A}, {\EuScript B})$
is identified with the covariant Grothendieck construction $\Gamma(F)$. We
will refer to $F$ as a {\em coCartesian gluing functor} of $( {\EuScript A}, {\EuScript B})$.
This analysis can be summarized as follows.
\begin{prop}\label{prop:gluing} Let ${\EuScript C}$ be a stable $\infty$-category equipped with a
semiorthogonal decomposition $( {\EuScript A}, {\EuScript B})$.
\begin{enumerate}
\item[(1)] Assume that $( {\EuScript A}, {\EuScript B})$ is Cartesian with Cartesian gluing functor $G: {\EuScript B} \to {\EuScript A}$. Then:
\begin{enumerate}
\item[(1a)] There is an equivalence of $\infty$-categories
\[
\operatorname{Fun}_{\Delta^1}(\Delta^1, \chi(G)) \overset{\simeq}{\longrightarrow} {\EuScript C}.
\]
In particular, ${\EuScript C}$ can be uniquely recovered, up to equivalence, from the triple $( {\EuScript A}, {\EuScript B},G)$.
\item[(1b)] For two objects $a\in {\EuScript A}, b\in {\EuScript B}$, a morphism $a\to b$ in ${\EuScript C}$
is $( {\EuScript A}, {\EuScript B})$-Cartesian, if and only if the morphism $a\to G(b)$
in $ {\EuScript A}$, corresponding to it by (1a), is an equivalence.
\end{enumerate}
\item[(2)] Assume that $( {\EuScript A}, {\EuScript B})$ is coCartesian with
coCartesian gluing functor $F: {\EuScript A} \to {\EuScript B}$. Then:
\begin{enumerate}
\item[(2a)] T here is an equivalence of $\infty$-categories
\[
\operatorname{Fun}_{\Delta^1}(\Delta^1, \Gamma(F)) \overset{\simeq}{\longrightarrow} {\EuScript C}.
\]
In particular, ${\EuScript C}$ can be uniquely recovered, up to equivalence, from the triple $( {\EuScript A}, {\EuScript B},F)$.
\item[(2b)] For two objects $a\in {\EuScript A}, b\in {\EuScript B}$, a morphism $a\to b$ in ${\EuScript C}$
is $( {\EuScript A}, {\EuScript B})$-coCartesian, if and only if the morphism $F(a)\to b$
in $ {\EuScript B}$, corresponding to it by (2a), is an equivalence.
\end{enumerate}
\end{enumerate}
\end{prop}
\begin{proof} Parts (1a) and (2a) are immediate from Proposition \ref{prop:sections} and Example \ref{ex:implicit}
(based on \cite[\S 3.3.2]{lurie:htt}). Parts (1b) and (2b) follow from the definition
of the Grothendieck constructions.
\end{proof}
\begin{rem}
Any ${\infty}$-functor $p: {\EuScript X}\to\Delta^1$, being an inner fibration, can be seen
as a {\em correspondence} between the $\infty$-categories $p^{-1}(0)$ and $p^{-1}(1)$, see
\cite[\S 2.3.1]{lurie:htt}. In these terms,
Proposition \ref{prop:sections} says that every semiorthogonal
decomposition $( {\EuScript A}, {\EuScript B})$ of ${\EuScript C}$, (co)Cartesian or not, arises from a correspondence between
$ {\EuScript A}$ and $ {\EuScript B}$. The Cartesian and coCartesian cases correspond to
the situation when this correspondence is given by a functor.
\end{rem}
\subsection{Admissiblity}\label{subsec:admissibility}
\paragraph{Admissibility and adjoints to the embedding.}
If $( {\EuScript A}, {\EuScript B})$ forms an
SOD of a stable ${\infty}$-category ${\EuScript C}$ then, by
Proposition \ref{prop:perp},
the subcategories $ {\EuScript A}$ and $ {\EuScript B}$ determine each other
uniquely. It is therefore natural to ask which conditions a full stable subcategory $ {\EuScript A} \subset
{\EuScript C}$ needs to satisfy so that $( {\EuScript A},{^\perp} {\EuScript A})$ forms an SOD.
\begin{defi} Let ${\EuScript C}$ be a stable $\infty$-category, and let $ {\EuScript A} \subset {\EuScript C}$ be a stable full
subcategory. Let $i: {\EuScript A} \to {\EuScript C}$ denote the corresponding embedding functor. We call $ {\EuScript A}
\subset {\EuScript C}$
\begin{itemize}
\item {\em left admissible} if $i$ has a left adjoint, denoted ${^*}i$;
\item {\em right admissible} if $i$ has a right adjoint, denoted $i^*$;
\item {\em admissible} if $i$ has both a left and a right adjoint.
\end{itemize}
\end{defi}
\begin{prop}\label{prop:adm} Let ${\EuScript C}$ be a stable $\infty$-category and let $ {\EuScript A}$ be a stable full
subcategory. Then:
\begin{enumerate}
\item[(a)] The pair $( {\EuScript A},{^\perp} {\EuScript A})$ forms an SOD of ${\EuScript C}$ if and only if
$ {\EuScript A} \subset {\EuScript C}$ is left admissible. If these conditions hold, then we further
have ${^\perp} {\EuScript A} = {\operatorname{Ker}}({^*}i)$ for any left adjoint ${^*}i$
of the inclusion $i: {\EuScript A} \to
{\EuScript C}$.
\item[(b)] The pair $( {\EuScript A}^\perp, {\EuScript A})$ forms an SOD of ${\EuScript C}$ if and only if
$ {\EuScript A} \subset {\EuScript C}$ is right admissible. If these conditions hold, then we further
have $ {\EuScript A}^\perp = {\operatorname{Ker}}(i^*)$ for any right adjoint $i^*$ of the inclusion $i: {\EuScript A} \to
{\EuScript C}$.
\end{enumerate}
\end{prop}
\noindent{\sl Proof:} We show (a), the statement of (b) is dual.
\vskip .2cm
Suppose that $i$ has a left adjoint ${^*i}: {\EuScript C} \to {\EuScript A}$.
Consider the triangulated category $\operatorname{h}\!{\EuScript C}$ and its full triangulated
subcategories $\operatorname{h}\! {\EuScript A}$ and $\operatorname{h}\!({^\perp} {\EuScript A})= {^\perp}\operatorname{h}\! {\EuScript A}$.
Note that
${\operatorname{Hom}}_{\operatorname{h}\!{\EuScript C}}(x,y) = \pi_0 \operatorname{Map}_{\EuScript C}(x,y)$. Therefore
the equivalence
\[
\operatorname{Map}_{{\EuScript C}}(x,i(a)) \simeq \operatorname{Map}_{ {\EuScript A}}({^*i}(x),a)
\]
implies that the functor $\operatorname{h}\!\, ({^*i}): \operatorname{h}\!{\EuScript C}\to\operatorname{h}\! {\EuScript A}$ is the left adjoint,
in the usual categorical sense, of the
embedding functor $\operatorname{h}\! i: \operatorname{h}\! {\EuScript A}\to\operatorname{h}\!{\EuScript C}$. This is a well known condition
for $(\operatorname{h}\! {\EuScript A}, \operatorname{h}\!({^\perp} {\EuScript A})= {^\perp} (\operatorname{h}\! {\EuScript A}))$ to form an SOD
decomposition of $\operatorname{h}\!{\EuScript C}$ in the sense of triangulated categories.
Therefore by Proposition \ref{SOD:oo.vs.triang}, $( {\EuScript A}, {^\perp {\EuScript A}})$
form an SOD of ${\EuScript C}$ in the sense of stable ${\infty}$-categories.
\vskip .2cm
Conversely, suppose that $( {\EuScript A}, {^\perp {\EuScript A}})$ is an SOD of ${\EuScript C}$.
By Definition \ref{def:oo-SOD}, cf. Proposition \ref{prop:<AB>}(a),
this means that the functor $ \operatorname{Fib}: \arr{ {\EuScript A}, {^\perp {\EuScript A}}}\to{\EuScript C}$ is
an equivalence. Under this equivalence, the embedding
$i: {\EuScript A}\to{\EuScript C}$ is identified with the embedding
$i': {\EuScript A} \to \arr{ {\EuScript A}, {^\perp {\EuScript A}}}$ sending an object $a$ to the
arrow $\{a\to 0\}$. But a left adjoint to $i'$ is immediately found to
be the projection functor sending an arrow $\{a\to b\}$ to $a$.
\qed
\iffalse
${^\perp} {\EuScript A}$ coincides with the full subcategory $\ker(p) \subset {\EuScript C}$ spanned
by objects $x$ such that $p(x)$ is a zero object. Let $\pi: \chi(i) \to \Delta^1$ denote the
contravariant Grothendieck construction of $i$, interpreted as an functor $i: [1] \to
{\operatorname{Set}}_{\Delta}$. The assumption that $i$ have a left adjoint is equivalent to $\pi$ being a
coCartesian fibration as well. Consider the full subcategory
\[
{\EuScript M} \subset \operatorname{Fun}(\Delta^1 \times \Delta^2, \chi(i))
\]
spanned by diagrams
\begin{equation}\label{eq:diam}
\begin{tikzcd}
x'' \arrow{r} \arrow{d} & x \arrow{d} \arrow{r}{!} & \arrow{d} a\\
0 \arrow{r} & x' \arrow{r}{!} & a'
\end{tikzcd}
\end{equation}
such that
\begin{enumerate}[label=(\roman{*})]
\item the left-hand square is a biCartesian square contained in $\pi^{-1}(\{0\}) =
{\EuScript C}$, and
\item $x'$ is an object of $ {\EuScript A} \subset {\EuScript C}$ and $x''$ is an object of ${^\perp} {\EuScript A}
\subset {\EuScript C}$,
\item the edges marked $!$ are $\pi$-coCartesian edges covering the edge $0 \to 1$,
so that $a \to a'$ is an edge in $\pi^{-1}(\{1\}) = {\EuScript A}$.
\end{enumerate}
It is clear that the apparent forgetful functor
\[
{\EuScript M} \to \co{ {\EuScript A},{^\perp} {\EuScript A}}
\]
is an equivalence. To show that $( {\EuScript A},{^\perp} {\EuScript A})$ forms a semiorthogonal decomposition of
${\EuScript C}$, we need to show that the forgetful functor $\co{ {\EuScript A},{^\perp} {\EuScript A}} \to {\EuScript C}$ is an
equivalence. By two-out-of-three, it suffices to show that the forgetful functor
\[
{\EuScript M} \to {\EuScript C}
\]
given by projection to the vertex $x$ is an equivalence. We claim that the condition on a
diagram in ${\EuScript M}$ that $x''$ lie in ${^\perp} {\EuScript A}$ may be replaced by the condition that the map
$a \to a'$ is an equivalence in $ {\EuScript A}$. This follows immediately from the above observation
${^\perp} {\EuScript A} = \ker(p)$: the universal properties of the coCartesian edges in \eqref{eq:diam}
imply that the arrow $a \to a'$ is equivalent to the image of $x \to x'$ under $p$. It
follows that we have an equivalence ${\EuScript M} \to {\EuScript M}'$ where ${\EuScript M}' \subset \operatorname{Fun}(\Delta^1 \times
\Delta^1, \chi(i))$ is the full subcategory spanned by diagrams
\begin{equation}\label{eq:diagmprime}
\begin{tikzcd}
x \arrow{d} \arrow{r}{!}\arrow{dr}{!} & \arrow{d}{\simeq} a\\
x' \arrow{r}{!} & a'
\end{tikzcd}
\end{equation}
with $x'$ in $ {\EuScript A} \subset \pi^{-1}(\{0\}) = {\EuScript C}$, the edges marked $!$ coCartesian, and $a \to a'$ an
equivalence in $\pi^{-1}(\{1\}) = {\EuScript A}$. We project further to the full subcategory ${\EuScript M}''$ spanned by
diagrams in $\operatorname{Fun}(\Delta^2, \chi(i))$ of the form
\begin{equation}\label{eq:diagmprimeprime}
\begin{tikzcd}
x \arrow{d} \arrow{dr}{!} &\\
x' \arrow{r}{!} & a'
\end{tikzcd}
\end{equation}
obtaining an equivalence ${\EuScript M}' \to {\EuScript M}''$. Finally, we note that the condition that $x'$ be in
$i( {\EuScript A})$ is equivalent to the coCartesian edge $x' \to a'$ being Cartesian as well. It then
follows from the universal properties of coCartesian and Cartesian edges (formulated in
terms of marked anodyne maps) that the further projection ${\EuScript M}'' \to {\EuScript C}$ to the vertex $x$ is
an equivalence, concluding the argument.
Vice versa, suppose that $( {\EuScript A}, {\EuScript B})$ forms a semiorthogonal decomposition of ${\EuScript C}$ and set $ {\EuScript B}
= {^\perp} {\EuScript A}$. We claim that this implies that the Cartesian
fibration $\pi: \chi(i) \to \Delta^1$ is coCartesian as well. More precisely, we claim that an
edge in $\chi$ covering $0 \to 1$, corresponding by construction of $\chi(i)$ to an edge $e: x \to
i(a)$ with $x \in {\EuScript C}$, $a \in {\EuScript A}$, is coCartesian if and only if the fiber of $e$ in ${\EuScript C}$ lies in
$ {\EuScript B}$. Assuming this claim, we have, by assumption, that the functor $\co{ {\EuScript A}, {\EuScript B}} \to {\EuScript C}$ is
an equivalence so that, for every vertex $x$ of ${\EuScript C}$, there exists a coCartesian edge $x \to
x'$ covering $0 \to 1$ verifying that $\pi$ is a coCartesian fibration.
To show that the above edges are indeed coCartesian, we argue as follows: Let $e$ be an edge
as above with fiber in $ {\EuScript B}$. Translating the definition of a coCartesian edge, we need to
show that, the map
\[
t: {\EuScript C}_{e/} \times_{{\EuScript C}} {\EuScript A} \to {\EuScript C}_{x/} \times_{{\EuScript C}} {\EuScript A}
\]
is a trivial Kan fibration. Since this map is a right fibration (\cite[2.1.2.1]{lurie:htt})
it suffices by \cite[2.1.3.4]{lurie:htt} to show that its fibers are contractible. To this
end, it suffices to show that, for every vertex $a'$ in $ {\EuScript A}$, the base change
\[
t_{a'}: {\EuScript C}_{e/} \times_{{\EuScript C}} \{a'\} \to {\EuScript C}_{x/} \times_{{\EuScript C}} \{a'\}
\]
is a trivial Kan fibration. It is a Kan fibration, since it is a left fibration of Kan
complexes so that it suffices to show that it is a homotopy equivalence. Completing $e$ to a
fiber sequence
\[
\begin{tikzcd}
x'' \ar{r}{g}\ar{d} & x \ar{d}{e}\\
0 \ar{r} & i(a)
\end{tikzcd}
\]
in ${\EuScript C}$, we may identify the map $t_{a'}$ with the forgetful map
\[
\operatorname{fib}(\operatorname{Map}_{{\EuScript C}}(x,a') \overset{g^*}{\to} \operatorname{Map}_{{\EuScript C}}(x'',a')) \to \operatorname{Map}_{{\EuScript C}}(x,a').
\]
Since $x''$ lies in $ {\EuScript B}$, we have $\operatorname{Map}_{{\EuScript C}}(x'',a') \simeq \{0\}$ so that the latter map is
a homotopy equivalence as required.
\fi
\paragraph{ Admissibility and (co)Cartesian SODs.}
Admissibility is not only related to the existence of semiorthogonal decompositions but also to
their properties of being Cartesian and coCartesian:
\begin{prop}\label{prop:cartright} Let ${\EuScript C}$ be a stable $\infty$-category equipped with a
semiorthogonal decomposition $( {\EuScript A}, {\EuScript B})$, so that $ {\EuScript A}$ is left
admissible and $ {\EuScript B}$ is right admissible. Then:
\begin{enumerate}
\item[(a)] The semiorthogonal decomposition $( {\EuScript A}, {\EuScript B})$ is Cartesian if and only if the
subcategory $ {\EuScript A} \buildrel i_ {\EuScript A}\over\hookrightarrow {\EuScript C}$ is also right admissible. If these equivalent conditions
hold, then the restriction to $ {\EuScript B}$ of any right adjoint $i_ {\EuScript A}^*: {\EuScript C} \to {\EuScript A}$ of $i_ {\EuScript A}$
is a Cartesian gluing functor.
\item[(b)] The semiorthogonal decomposition $( {\EuScript A}, {\EuScript B})$ is coCartesian if and only if the
subcategory $ {\EuScript B} \buildrel i_ {\EuScript B}\over\hookrightarrow {\EuScript C}$ is also left admissible. If these equivalent conditions
hold, then the restriction to $ {\EuScript A}$ of any left adjoint ${^*i}_ {\EuScript B}: {\EuScript C} \to {\EuScript B}$ of $i_ {\EuScript B}$
is a coCartesian gluing functor.
\end{enumerate}
\end{prop}
\noindent{\sl Proof:} We show (a), the statement of (b) is dual.
\vskip .2cm
First, suppose $( {\EuScript A}, {\EuScript B})$ is Cartesian. This means that
the ${\infty}$-category
$\chi( {\EuScript A}, {\EuScript B})$ is identified, as an ${\infty}$-category over $\Delta^1$,
with the contravariant Grothendieck construction
$\chi(G)$, where $G: {\EuScript B}\to {\EuScript A}$ is a coCartesian gluing functor.
Let ${\EuScript L} := \operatorname{Fun}_{\Delta^1}(\Delta^1, \chi(G))$ be the
$\infty$-category of sections of $\chi(G)\to\Delta^1$.
By definition, an object of ${\EuScript L}$ is a triple $(a,b,u)$,
where $a\in {\EuScript A}$, $b\in {\EuScript B}$ are objects and $u: a\to G(b)$
is a morphism in $ {\EuScript A}$. Sometimes for simplicity we will
write an object of ${\EuScript L}$ simply as $\{a\buildrel u\over\to G(b)\}$.
Since, by definition, ${\EuScript C}$ is identified with the ${\infty}$-category
of sections of $\chi( {\EuScript A}, {\EuScript B})\to\Delta^1$, we have an identification
${\EuScript C}\={\EuScript L}$. Note that this identification sends an object $a\in {\EuScript A}$
to the triple $(a,0,0)$ and an object $b\in {\EuScript B}$ to the triple
$(0,b[1],0)$. Let $ {\EuScript A}'$ and $ {\EuScript B}'$ be the full subcategories
of ${\EuScript L}$ spanned by triples of these types, so that $ {\EuScript A}'\= {\EuScript A}$ and
$ {\EuScript B}'\= {\EuScript B}$.
\vskip .2cm
Let $r: {\EuScript X}= \Gamma(i') \to \Delta^1$ be the covariant
Grothendieck construction of the inclusion $i': {\EuScript A}' \subset {\EuScript L}$.
We need to show that $i'$ admits a
right adjoint which is equivalent to the map $r$ being a Cartesian fibration. By definition, an edge $e$ of ${\EuScript X}$ covering the edge
$0 \to 1$ of $\Delta^1$ is a datum consisting of
objects $a,a'\in {\EuScript A}$, $b\in {\EuScript B}$ and a square
\begin{equation}\label{eq:psicart}
\begin{tikzcd}
a' \ar{r} \ar{d} & a \ar{d}\\
0 \ar{r} & G(b)
\end{tikzcd}
\end{equation}
in $ {\EuScript A}$. The following lemma is
straightforward and its proof is left to the reader.
\begin{lem}
An edge $e$ of ${\EuScript X}$ covering $0\to 1$ is Cartesian,
if and only if the square \eqref{eq:psicart} is Cartesian
(i.e., is biCartesian, since $ {\EuScript A}$ is stable). \qed
\end{lem}
We conclude that $r$ is Cartesian, as any object $\{a\buildrel u\over
\to G(b)\}$
of $r^{-1}(1)= {\EuScript L}$ is the target of a Cartesian edge as above
with $a'= \operatorname{Fib}(u)$.
\vskip .2cm
Further, let us determine the
restriction to $ {\EuScript B}'$ of any right adjoint, denote it $j$, of $i'$.
By construction, see Example \ref{ex:implicit}, the value
of $j$ on an object $L=\{ a\buildrel u\over
\to G(b)\} $ of ${\EuScript L}$ is obtained as the source of a Cartesian edge
terminating at $L$, i.e., as the fiber of $u$. If $L$ has $a=0$, then
$ \operatorname{Fib}(u)=G(b)[-1]$. By a similar analysis on all the simplices,
we find that $j= G[-1]$. To bring this to a statement about the
right adjoint of $i: {\EuScript A}\to{\EuScript C}$, we recall from the above
that the identification of
$ {\EuScript B}\subset{\EuScript C}$ with $ {\EuScript B}'\subset {\EuScript L}$ is given by $b\mapsto
(0, b[1], 0)$.
Therefore, the restriction of the adjoint of the embedding
$ {\EuScript A} \hookrightarrow {\EuScript L}$ to $ {\EuScript B}$ is
simply given by $G$ so that we have verified
the last claim of the proposition.
\vskip .2cm
Conversely, suppose that $i: {\EuScript A} \to {\EuScript C}$ admits a right adjoint,
i.e., that
the
covariant Grothendieck construction
$r: {\EuScript X}=\Gamma(i) \to \Delta^1$ is Cartesian.
Let ${\EuScript Y} \subset {\EuScript X}$ be the full
subcategory spanned by the objects of $ {\EuScript A} = r^{-1}(0)$ and the vertices of
${\EuScript C} =
r^{-1}(1)$ that are contained in $ {\EuScript B}$. Then it is immediate to verify that the
restriction $r|_{{\EuScript Y}}: {\EuScript Y} \to \Delta^1$ is a Cartesian fibration and
the $r|_{{\EuScript Y}}$-Cartesian edges are
precisely those whose image in ${\EuScript X}$ is $r$-Cartesian. Further, the simplicial set ${\EuScript Y}$ is
isomorphic to $\chi( {\EuScript A}, {\EuScript B})$ over $\Delta^1$. Thus $( {\EuScript A}, {\EuScript B})$ is Cartesian.
\qed
\begin{cor} Gluing functors are exact. \end{cor}
\begin{proof} By Proposition \ref{prop:cartright}, the gluing functor is given by the restriction of
a right or a left adjoint. So it is left or left exact and therefore exact since it acts between
stable ${\infty}$-categories.
\end{proof}
\begin{rem}\label{rem:sections} Let ${\EuScript C}$ be a stable $\infty$-category equipped with a Cartesian
semiorthogonal decomposition $( {\EuScript A}, {\EuScript B})$ with right gluing functor $G: {\EuScript B} \to {\EuScript A}$. By Proposition
\ref{prop:gluing}, we obtain a canonical equivalence
\[
{\varphi}: {\EuScript C}\buildrel \simeq\over\longrightarrow {\EuScript L} =
\operatorname{Fun}_{\Delta^1}(\Delta^1,\chi(G)).
\]
Under this equivalence, we have the following
identifications:
\begin{enumerate}
\item ${\varphi}( {\EuScript A}{^\perp}) \subset {\EuScript L}$ is the full subcategory
spanned by the Cartesian sections.
\item ${\varphi}( {\EuScript A}) \subset {\EuScript L}$ is the full subcategory
spanned by sections of the form $a \to 0$ with $a \in {\EuScript A}$.
\item ${\varphi}( {\EuScript B}) \subset {\EuScript L}$ is the full subcategory
spanned by sections of the form $0 \to G(b)$ with $b \in {\EuScript B}$.
\end{enumerate}
If we further assume that $( {\EuScript A}, {\EuScript B})$ is coCartesian, then
\begin{enumerate}[resume]
\item ${\varphi}({^\perp} {\EuScript B}) \subset {\EuScript L}$ is the full subcategory
spanned by the coCartesian sections.
\end{enumerate}
\end{rem}
\subsection{Mutation}\label{subsec:mutation}
\paragraph{Definition of the left and right mutation functors.}
Let ${\EuScript C}$ be a stable $\infty$-category and let $ {\EuScript A} \subset {\EuScript C}$ be an admissible subcategory.
This means that the embedding functor $i_ {\EuScript A}: {\EuScript A}\hookrightarrow{\EuScript C}$ has both a left adjoint
${^*}i_ {\EuScript A}: {\EuScript C}\to {\EuScript A}$ and a right adjoint $i_ {\EuScript A}^*:{\EuScript C}\to {\EuScript A}$.
By Proposition \ref{prop:adm}, we obtain that $( {\EuScript A}^\perp, {\EuScript A})$ and $( {\EuScript A}, {^\perp} {\EuScript A})$ form
semiorthogonal decompositions. This implies that:
\begin{itemize}
\item $ {\EuScript A}^\perp$ is left admissible, i.e., the embedding $i_{ {\EuScript A}^\perp}: {\EuScript A}^\perp\hookrightarrow{\EuScript C}$
has a left adjoint ${^*}i_{ {\EuScript A}^\perp}: {\EuScript C}\to {\EuScript A}^\perp$.
\item ${^\perp} {\EuScript A}$ is right adissible, i.e., the embeddibg
$i_{{^\perp} {\EuScript A}}: {^\perp} {\EuScript A}\hookrightarrow{\EuScript C}$ has a right adjoint $i_{{^\perp} {\EuScript A}}^*: {\EuScript C}\to {^\perp} {\EuScript A}$.
\end{itemize}
\begin{defi}\label{defi:mutation}
We define the {\em left} and {\em right mutation ${\infty}$-functors} around $ {\EuScript A}$ to be
the compositions
\[
\begin{gathered}
\lambda_{ {\EuScript A}}\, =\, ({^*}i_{A^\perp})\circ i_{{^\perp} {\EuScript A}}
: {^\perp} {\EuScript A} \longrightarrow {\EuScript A}^{\perp},
\\
\rho_{ {\EuScript A}}\,=\, (i_{{^\perp} {\EuScript A}}^*) \circ i_{ {\EuScript A}^\perp}
: {\EuScript A}^{\perp} \longrightarrow {^\perp} {\EuScript A}.
\end{gathered}
\]
We will also
refer to the subcategory
$ {\EuScript A}^{\perp}$ as the {\em left mutation} of ${^\perp} {\EuScript A}$ around $ {\EuScript A}$, and vice versa,
refer to
${^\perp} {\EuScript A}$ as the {\em right mutation} of $ {\EuScript A}^{\perp}$ around $ {\EuScript A}$.
\end{defi}
\paragraph{Mutation functors are equivalences.}
By construction, $(\lambda_ {\EuScript A}, \rho_ {\EuScript A})$ form an adjoint pair. In fact, we have a stronger
statement.
\begin{prop} \label{cor:mutation}
$\lambda_ {\EuScript A}$ and $\rho_ {\EuScript A}$ are mutually quasi-inverse equivalences of ${\infty}$-categories.
\end{prop}
Before giving the proof, we introduce some terminology. Let
\begin{equation}\label{eq:M-sq}
\xymatrix{
a'' \ar[r]^v \ar[d] _{ { \textstyle S\quad =\quad \quad\quad}} & a \ar[d]^u \\
0 \ar[r] & a'
}
\end{equation}
be a biCartesian square in ${\EuScript C}$. We will say that $S$ is an {\em M-square}
(short for {\em mutation square}), if
$a''$ is an object of $ {\EuScript A}^{\perp}$, while $a$ is an object of $ {\EuScript A}$, and $a'$ is an
object of ${^\perp} {\EuScript A}$.
\begin{lem}\label{lem:flip}
Let $S$ be a biCartesian square in ${\EuScript C}$. Then the following are equivalent:
\begin{enumerate}
\item[(i)] $S$ is an M-square.
\item[(ii)] The edge $a'' \to a$ is $( {\EuScript A}^{\perp}, {\EuScript A})$-coCartesian.
\item[(iii)] The edge $a \to a'$ is $( {\EuScript A},{^\perp} {\EuScript A})$-Cartesian.
\end{enumerate}
\end{lem}
\noindent{\sl Proof of the lemma:}
We show (i) $\Leftrightarrow$ (iii), the equivalence (i) $\Leftrightarrow$ (ii)
follows by passing to opposites.
\vskip .2cm
Note that
by Proposition \ref{prop:cartright},
$( {\EuScript A}^\perp, {\EuScript A})$ is a coCartesian SOD and $( {\EuScript A}, {^\perp} {\EuScript A})$ is a Cartesian SOD.
Set $ {\EuScript B} =
{^\perp} {\EuScript A}$ and write $i_ {\EuScript B}=i_{{^\perp} {\EuScript A}}$.
\vskip .2cm
Since $( {\EuScript A}, {\EuScript B})$ is Cartesian, we have a Cartesian gluing functor $G: {\EuScript B} \to {\EuScript A}$.
We recall that $G$ is obtained by analyzing the Cartesian fibration
$p: \chi( {\EuScript A}, {\EuScript B})\to\Delta^1$
and identifying it with the contravariant Grothendieck construction
$p: \chi(G)\to\Delta^1$.
Thus a Cartesian edge $u: a\to a'$ with $a\in {\EuScript A}$,
$a'\in {\EuScript B}$ can be, after composing with an equivalence in the source, identified
with the canonical edge $G(a')\to a'$ in $\chi(G)$.
At the same time $G$ is identified with
$i_ {\EuScript A}^* i_ {\EuScript B}$ by Proposition \ref{prop:cartright}. The functor $i_ {\EuScript A}^*$
can be obtained from the diagram
\[
{\EuScript C} \buildrel {\operatorname{Cof}}\over \longleftarrow \{ {\EuScript A}^\perp, {\EuScript A} \} \buildrel \operatorname{pr}_ {\EuScript A}\over\longrightarrow {\EuScript A}.
\]
Here ${\operatorname{Cof}}$ is an equivalence which we invert, and
$\operatorname{pr}_ {\EuScript A}$ is the projection functor sending an arrow
$c\to a$ with $c\in {\EuScript A}^\perp$, $a\in {\EuScript A}$, to $a$. This latter functor is obviously
right adjoint to the embedding $ {\EuScript A}\to \{ {\EuScript A}^\perp, {\EuScript A} \}$ sending
$a$ to $0\to a$.
Therefore $G(a')$, together with its canonical Cartesian edge to $a'$, is found by lifting $a'$ to $ \{ {\EuScript A}^\perp, {\EuScript A} \}$ via ${\operatorname{Cof}}$
and then projecting to $ {\EuScript A}$ which means precisely forming an
M-square. Lemma is proved. \qed
\vskip .2cm
\noindent {\sl Proof of Proposition \ref{cor:mutation}:}
Let ${\EuScript M} \subset \operatorname{Fun}(\Delta^1 \times \Delta^1, {\EuScript C})$ be the full subcategory formed
by M-squares. We have the projections
\[
{\EuScript A}^\perp \buildrel q''\over \longleftarrow {\EuScript M} \buildrel q'\over \longrightarrow {^\perp {\EuScript A}}.
\]
sending an M-square as in \eqref{eq:M-sq} to $a'$ and $a''$ respectively.
We claim that both $q'$ and $q''$ are trivial Kan fibrations. Let us show this for
$q'$, the case of $q''$ being similar. Indeed, we can decompose $q''$ as
the composition
\[
{\EuScript M} \buildrel q'_2 \over\longrightarrow \operatorname{Cart} \buildrel q'_1\over\longrightarrow {^\perp {\EuScript A}},
\]
where $\operatorname{Cart}\subset \operatorname{Fun}(\Delta^1, {\EuScript C})$ is the full ${\infty}$-subcategory
spanned by $( {\EuScript A}, {^\perp {\EuScript A}})$-Cartesian edges, $q'_2$ projects a square
\eqref{eq:M-sq} to $a\to a'$ and $q'_1$ projects $a\to a'$ to $a'$.
Now, $q'_1$ is a trivial Kan fibration since $( {\EuScript A}, {^\perp {\EuScript A}})$ is a Cartesian SOD,
and $q'_2$ is a trivial Kan fibration by the ``uniqueness of the triangle" property
\cite[Rem. 1.1.1.7]{lurie:ha}.
Thus composing $q''$ with a section of $q'$, we get an equivalence $ {\EuScript A}^\perp\to{^\perp {\EuScript A}}$
defined canonically up to a contractible choice. We finally note that
$\rho_{ {\EuScript A}}= (i_{{^\perp} {\EuScript A}}^*) \circ i_{ {\EuScript A}^\perp}$ can be identified with such
a composite equivalence. This can be seen by analyzing $i_{^\perp {\EuScript A}}^*$ in terms
of Cartesian edges of the covariant Grothendieck construction $\Gamma(i_{^\perp {\EuScript A}})$
in a way completely analogous to the proof of Proposition \ref{prop:cartright}.
We leave the details to the reader. Proposition \ref{cor:mutation} is proved. \qed
\paragraph{Mutation functors as coordinate changes.}\label{par:mut-change}
Let ${\EuScript C}$ be a stable $\infty$-category and $ {\EuScript A} \subset {\EuScript C}$ an admissible subcategory.
By Proposition \ref{prop:<AB>}(a), we have equivalences of $\infty$-categories
\[
\arr{ {\EuScript A}^{\perp}, {\EuScript A}} \overset{\operatorname{fib}}{\longrightarrow} {\EuScript C} \overset{\operatorname{fib}}{\longleftarrow} \arr{ {\EuScript A}, {^{\perp} {\EuScript A}}}.
\]
Let $\psi: \arr{ {\EuScript A}^{\perp}, {\EuScript A}} \buildrel \simeq\over\to \arr{ {\EuScript A}, {^{\perp} {\EuScript A}}}$
be the resulting coordinate change
equivalence.
Lemma \ref{lem:flip} allows for an explicit description of $\psi$ as follows. Let
\begin{equation}\label{eq:m}
\Di \subset \operatorname{Fun}(\Delta^2 \times \Delta^2, {\EuScript C})
\end{equation}
be the full ${\infty}$-subcategory
consisting of diagrams of the form
\begin{equation}\label{eq:mdiag}
\begin{tikzcd}
x \ar{r}\ar{d} & b \ar{d}{!} \ar{r} & 0 \ar{d}\\
a' \ar{r} \ar{d} & y \ar{r}{*}\ar{d} & b' \ar{d}\\
0 \ar{r} & a \ar{r} & x'
\end{tikzcd}
\end{equation}
such that
\begin{itemize}
\item $b$ in $ {\EuScript A}^{\perp}$, $b'$ in ${^\perp} {\EuScript A}$, $a,a',y$ in $ {\EuScript A}$,
\item all squares are biCartesian in ${\EuScript C}$,
\item the edge $!$ is $( {\EuScript A}^{\perp}, {\EuScript A})$-coCartesian, and the edge $*$ is $( {\EuScript A},
{^{\perp} {\EuScript A}})$-Cartesian.
\end{itemize}
We have the projections
\[
\arr{ {\EuScript A}^{\perp}, {\EuScript A}} \buildrel p\over\longleftarrow \Di \buildrel q\over\longrightarrow \arr{ {\EuScript A},
{^{\perp} {\EuScript A}}},
\]
$p$ being the projection to the edge $b \to a$ and $q$ the projection to the edge $a' \to b'$.
\begin{prop}\label{rem:basechange}
(a) Both $p$ and $q$ are trivial Kan fibrations.
\vskip .2cm
(b) The equivalence $\psi$ is given by
composing a section of $p$ with $q$.
\vskip .2cm
(c) The equivalence with ${\EuScript C}$ is given by
projecting to the vertex $x$.
\end{prop}
\noindent{\sl Proof:} Parts (a) and (b) follow immediately from Lemma \ref{lem:flip}. To
see (c), notice that
the diagram \eqref{eq:mdiag} exhibits $x$ simultaneously
as the fiber of the morphism $b \to a$ and the fiber of $a' \to b'$. \qed
\paragraph{Relation between Cartesian and coCartesian gluing functors.}
Let $( {\EuScript A}, {\EuScript B})$ be a semiorthogonal decomposition (SOD)
of ${\EuScript C}$.
Definition \ref{def:Cart-SOD} and Proposition \ref{prop:cartright} can be
summarized as follows.
\vskip .2cm
If $( {\EuScript A}, {\EuScript B})$ is Cartesian, then we have a Cartesian gluing functor $G: {\EuScript B}\to {\EuScript A}$, so that
$\chi( {\EuScript A}, {\EuScript B})$ is identified with the contravariant Grothendieck construction $\chi(G)$.
In this case the pair $( {\EuScript A}^\perp, {\EuScript A})$ also forms an SOD, which is coCartesian
and therefore has a coCartesian gluing functor $\Phi: {\EuScript A}^\perp\to {\EuScript A}$.
\vskip .2cm
If $( {\EuScript A}, {\EuScript B})$ is coCartesian, then we have a coCartesian gluing functor
$F: {\EuScript A}\to {\EuScript B}$ so that
$\chi( {\EuScript A}, {\EuScript B})$ is identified with the covariant Grothendieck construction $\Gamma(F)$.
In this case the pair $( {\EuScript B}, {^\perp {\EuScript B}})$ also forms an SOD, which is Cartesian
and therefore has a Cartesian gluing functor $\Psi: {^\perp {\EuScript B}}\to {\EuScript B}$.
\begin{prop}\label{prop:L-R-glue}
(a) If $( {\EuScript A}, {\EuScript B})$ is Cartesian, then we have an identification $\Phi \= G\circ\rho_ {\EuScript A}$.
(b) If $( {\EuScript A}, {\EuScript B})$ is coCartesian, then we have an identification $\Psi\= F\circ\lambda_ {\EuScript B}$.
\end{prop}
\noindent{\sl Proof:} We prove (a), since (b) is similar. For $a''\in {\EuScript A}^\perp$, the value
$\Phi(a'')$ is, by Proposition \ref{prop:cartright}, found by including $a''$ into an M-square
\eqref{eq:M-sq} and projecting to $a\in {\EuScript A}$. The value $\rho_ {\EuScript A} (a'')$ is found by including
$a''$ into (the same) M-square and projecting to $ {\EuScript B}={^\perp {\EuScript A}}$, getting
some $a'\in {\EuScript B}$. Finally,
for any $a'\in {\EuScript B}$ (in particular, for the one just obtained) the value of $G(a')$
is found by including $a''$ into an M-square and projecting to $a$.
Since we can take the same M-square to serve all three cases, the claim follows. \qed
\subsection{$4$-periodic semiorthogonal decompositions and spherical adjunctions}
\label{subsec:4-period}
In this section, we provide an alternative definition of spherical adjunctions in terms of
$4$-periodic semiorthogonal decompositions. This perspective is due to \cite{halpern-shipman}.
\iffalse
Examples of periodic semiorthogonal decompositions for triangulated categories can
also be found in
\cite{iyama-kato-miyachi}.
In our context, these latter examples arise from the spherical functor
$ {\EuScript A} \to 0$ where $ {\EuScript A}$ is any stable $\infty$-category.
\fi
\paragraph{ BiCartesian decompositions, orthogonals and adjoints.}
Let ${\EuScript C}$ be a stable $\infty$-category, and $( {\EuScript A}, {\EuScript B})$ be a semiorthogonal decomposition (SOD)
of ${\EuScript C}$. We assume that $ {\EuScript A}$ and $ {\EuScript B}$ are strictly full, so $ {\EuScript B}={^\perp {\EuScript A}}$ and $ {\EuScript A}= {\EuScript B}^\perp$.
As in other situations, we say that
$( {\EuScript A}, {\EuScript B})$ of ${\EuScript C}$ is a {\em biCartesian} SOD, if it is both Cartesian and coCartesian.
In this case
the ${\infty}$-functor $p: \chi( {\EuScript A}, {\EuScript B})\to\Delta^1$, see
Definition \ref{def:Cart-SOD}, is a biCartesian fibration. Further, we have the
Cartesian gluing
functor $G: {\EuScript B}\to {\EuScript A}$ (coming from the fact that $( {\EuScript A}, {\EuScript B})$ is Cartesian) and
the coCartesian gluing functor $F: {\EuScript A}\to {\EuScript B}$ (coming from the fact that $( {\EuScript A}, {\EuScript B})$ is coCartesian).
These functors are adjoint to each other: $(F,G)$ is an adjoint pair and $\chi( {\EuScript A}, {\EuScript B})\to \Delta^1$
provides the adjunction datum.
Moreover,
every consecutive pair in the
sequence
\[
{\EuScript A}^{\perp}\quad,\quad {\EuScript A}\quad,\quad {\EuScript B}={^\perp {\EuScript A}}\quad,\quad {^\perp} {\EuScript B}={^{\perp\perp}} {\EuScript A}
\]
forms an SOD, more precisely, $( {\EuScript A}^\perp, {\EuScript A})$ is coCartesian, $( {\EuScript A}, {\EuScript B})$ is biCartesian
and $( {\EuScript B}, {^\perp {\EuScript B}})$ is Cartesian.
These categories are related by the gluing functors:
\[
\xymatrix{
{\EuScript A}^\perp \ar[r]^\Phi&
{\EuScript A} \ar@<.4ex>[r]^{F } & {\EuScript B} \ar@<.4ex>[l]^{G }& ^{\perp} {\EuScript B},\ar[l]_\Psi
}
\]
with $\Phi, F$ being coCartesian gluing functors of coCartesian SODs and
$G, \Psi$ being
right Cartesian functors of Cartesian SODs.
\begin{prop}\label{prop:bicart-adj}
\begin{itemize}
\item[(a)] Let $( {\EuScript A}, {\EuScript B})$ being a biCartesian SOD. Then:
\begin{itemize}
\item[(a1)] The coCartesian gluing functor $\Phi$ of $( {\EuScript A}^\perp, {\EuScript A})$ is identified, by composing
with the mutation functor $\rho_ {\EuScript A}$, with the right adjoint of the coCartesian
gluing functor $F$
for $( {\EuScript A}, {\EuScript B})$, i.e., $\Phi=G\circ\rho_ {\EuScript A}$.
\item[(a2)] The Cartesian gluing functor $\Psi$ for $( {\EuScript B}, {^\perp {\EuScript B}})$ is identified, by
composing with the mutation functor $\lambda_ {\EuScript B}$, with the left adjoint of the
Cartesian gluing functor $G$ for $( {\EuScript A}, {\EuScript B})$, i.e., $\Psi=F\circ\lambda_ {\EuScript B}$.
\end{itemize}
\item[(b)] Let $( {\EuScript A}, {\EuScript B})$ be a coCartesian SOD of ${\EuScript C}$, with the corresponding gluing functors
$ {\EuScript A}\buildrel F\over\longrightarrow {\EuScript B} \buildrel \Psi\over\longleftarrow {^\perp {\EuScript B}}$. Then $( {\EuScript A}, {\EuScript B})$ is
biCartesian, if and only if $F$ has a right adjoint.
\item[(c)] Let $( {\EuScript A}, {\EuScript B})$ be a Cartesian SOD of ${\EuScript C}$, with the corresponding gluing functors
$ {\EuScript A}^\perp\buildrel\Phi\over\longrightarrow {\EuScript A} \buildrel G\over\longleftarrow {\EuScript B}$. Then $( {\EuScript A}, {\EuScript B})$ is
biCartesian, if and only if $G$ has a left adjoint.
\end{itemize}
\end{prop}
\noindent{\sl Proof:} Part (a) follows from Proposition \ref{prop:L-R-glue}. Let us prove
(b), part (c) being similar. The ``only if'' part follows since for a biCartesian SOD
$( {\EuScript A}, {\EuScript B})$, the biCartesian fibration
$\chi( {\EuScript A}, {\EuScript B})$ provides an adjoint pair of functors. So see the ``if'' part, we identify
$\chi( {\EuScript A}, {\EuScript B})$ with the covariant Grothendieck construction $\Gamma(F)$. So the existence
of a right adjoint to $F$ means that $\Gamma(F)\to\Delta^1$ is a biCartesian fibration and
so $( {\EuScript A}, {\EuScript B})$ is a bicartesian SOD. \qed
\vskip .2cm
One advantage of expressing an adjunction $p: {\EuScript X} \to \Delta^1$ of stable $\infty$-categories in
terms of a semiorthogonal decomposition of the stable $\infty$-category ${\EuScript C} =
\operatorname{Fun}_{\Delta^1}(\Delta^1, {\EuScript X})$ consists of the accessibility of {\em further} iterated adjoints by
mutating the semiorthogonal decomposition to the left and the right.
To illustrate this, we use the following concepts.
\begin{defi}
(a) An {\em admissible chain of orthogonals} in a stable ${\infty}$-category ${\EuScript C}$ is a sequence
$( {\EuScript A}_i)_{i\in \mathbb{Z}}$ formed by strictly full stable subcategories such that for each
$i$ we have $ {\EuScript A}_i = {^\perp {\EuScript A}}_{i-1}= {\EuScript A}_{i+1}^\perp$ and each pair $( {\EuScript A}_i, A_{i+1})$
forms a semiorthogonal decomposition of ${\EuScript C}$.
\vskip .2cm
(b) A semiorthogonal decomposition $( {\EuScript A}, {\EuScript B})$ of ${\EuScript C}$ is called ${\infty}$-{\em admissible},
if it can be included into a (necessarily unique)
admissible chain of orthogonals $( {\EuScript A}_i)_{i\in\mathbb{Z}}$
so that $ {\EuScript A}= {\EuScript A}_0$ and $ {\EuScript B}= {\EuScript A}_1$.
\end{defi}
Proposition \ref{prop:bicart-adj} implies at once the following.
\begin{cor}\label{cor:oo-adm-adj}
Let $( {\EuScript A}, {\EuScript B})$ be a semiorthogonal decomposition of ${\EuScript C}$. The following
are equivalent:
\begin{itemize}
\item[(i)] $( {\EuScript A}, {\EuScript B})$ is ${\infty}$-admissible.
\item[(ii)] $( {\EuScript A}, {\EuScript B})$ is coCartesian and the corresonding coCartesian
gluing functor $F: {\EuScript A}\to {\EuScript B}$ admits
iterated left and right adjoints of all orders.
\item[(iii)] $( {\EuScript A}, {\EuScript B})$ is Cartesian and the corresonding Cartesian
gluing functor $G: {\EuScript B}\to {\EuScript A}$ admits
iterated left and right adjoints of all orders.
\end{itemize} \qed
\end{cor}
\paragraph{$4$-periodic SODs.}
\begin{defi} Let ${\EuScript C}$ be a stable $\infty$-category. An admissible chain of orthogonals
$( {\EuScript A}_i)_{i\in \mathbb{Z}}$ is called {\em $4$-periodic}, if $ {\EuScript A}_i= {\EuScript A}_{i+4}$ for all $i$.
An SOD $( {\EuScript A}, {\EuScript B})$ of ${\EuScript C}$ is called {\em $4$-periodic}, if
it includes into a $4$-periodic admissible chain of orthogonals $( {\EuScript A}_i)$
so that $ {\EuScript A}= {\EuScript A}_0, {\EuScript B}= {\EuScript A}_1$.
\end{defi}
\begin{prop}
Let $( {\EuScript A}, {\EuScript B})$ be an SOD of a stable ${\infty}$-category ${\EuScript C}$.
The following are equivalent:
\begin{itemize}
\item[(i)] $( {\EuScript A}, {\EuScript B})$ is $4$-periodic.
\item[(ii)] $( {\EuScript A}, {\EuScript B})$ is biCartesian and the pair $({^\perp} {\EuScript B}, {\EuScript A}^{\perp})$ also forms
an SOD of ${\EuScript C}$. \qed
\end{itemize}
\end{prop}
In the next paragraph we give a criterion for a biCartesian SOD to be $4$-periodic.
\paragraph{The relative suspension functor of a biCartesian SOD.}
Let ${\EuScript C}$ be a stable $\infty$-category equipped with a biCartesian SOD
$( {\EuScript A}, {\EuScript B})$. Denote $ {\EuScript A}' = {^{\perp} {\EuScript B}}$ and $ {\EuScript B}' = { {\EuScript A}^{\perp}}$.
By a {\em bimutation cube} ({\em MM-cube} for short) we will mean a
cubical diagram, i.e., a functor $Q\in \operatorname{Fun}((\Delta^1)^3,{\EuScript C})$ of the form
\begin{equation}\label{eq:cubical}
\xymatrix@=0.4cm{ & x' \ar[rr]\ar[dd] & & y' \ar[dd]\\
x \ar[ur]\ar[rr]\ar[dd]_{\textstyle{Q\quad = \quad\quad}
}
& & y \ar[dd]\ar[ur] & \\
& 0 \ar[rr] & & z' \\
0 \ar[ur] \ar[rr]& & z \ar[ur] & }
\end{equation}
satisfying the following conditions:
\begin{enumerate}
\item[(MM1)] the front and back faces of $Q$ are biCartesian squares
in ${\EuScript C}$,
\item[(MM2)] the vertices $x$, $x'$ are objects of $ {\EuScript A}$, the vertices $y$, $y'$ are objects
of $ {\EuScript B}$,
\item[(MM3)] the morphism $x \to y$ is coCartesian (and therefore $z\in {\EuScript A}'$),
while the morphism $x' \to y'$ is Cartesian (and therefore $z'\in {\EuScript B}'$).
\end{enumerate}
Let $ {\EuScript K} \subset \operatorname{Fun}((\Delta^1)^3,{\EuScript C})$ be the full subcategory spanned by MM-cubes.
\begin{lem}\label{lem:cube} Consider the
diagram of $\infty$-categories
\[
{\EuScript C} \overset{ \operatorname{Fib}}{\longleftarrow} \arr{ {\EuScript A}, {\EuScript B}} \overset{p}{\longleftarrow} {\EuScript K}
\overset{q}{\longrightarrow} \arr{ {\EuScript A}', {\EuScript B}'} \overset{ \operatorname{Fib}}{\longrightarrow} {\EuScript C}
\]
where $p$ projects an MM-cube $Q$ to the diagonal arrow $x \to y'$,
while $q$ projects $Q$ to the arrow $z \to z'$.
As usual, $ \operatorname{Fib}$ denotes the functor associating to a morphism its fiber.
In this diagram, the functors $p$ and $q$ are trivial Kan fibrations.
\end{lem}
\noindent{\sl Proof:} Consider first $p: {\EuScript K}\to \arr{ {\EuScript A}, {\EuScript B}}$. Let
${\EuScript E}\subset \operatorname{Fun}((\Delta^1)^2,{\EuScript C})$ be the full subcategory formed by possible
upper squares of MM-cubes, i.e., by diagrams
\[
\xymatrix{
x'\ar[r]&y'
\\
x\ar[u] \ar[r]&y\ar[u]
}
\]
with $x,x'\in {\EuScript A}$, $y,y'\in {\EuScript B}$ the arrow $x \to y$ coCartesian and
the arrow $x' \to y'$ Cartesian. We factor $p$ as the composition
$ {\EuScript K} \buildrel p_2\over\to {\EuScript E} \buildrel p_1\over\to \arr{ {\EuScript A}, {\EuScript B}}$.
Any arrow $x\to y'$ can be represented, in an essentially unique way,
as a composition of a coCartesian arrow $x'\to y'$ and an arrow $x\to x'$,
as well as a composition of an arrow $y\to y'$ and a Cartesian arrow
$x\to y$. This fact upgrades, in a standard way, see \cite[4.3.2.15]{lurie:htt}
to the conclusion that $p_1$ is a trivial Kan fibration. Further,
$p_2$ is a trivial Kan fibration by a similar upgrade of the ``uniqueness of triangle'' property.
\vskip .2cm
Look now at $q: {\EuScript K}\to \arr{ {\EuScript A}', {\EuScript B}'}$. Using the fact the ${\operatorname{Cof}}: \arr{ {\EuScript A}, {\EuScript B}}\to{\EuScript C}$
is an equivalence, we get that ${\operatorname{Cof}}$ induces an equivalence
${\operatorname{Cof}}_1: \operatorname{Fun}(\Delta^1, \arr{ {\EuScript A}, {\EuScript B}})\to \operatorname{Fun}(\Delta^1, {\EuScript C})$. We consider
the preimage, under ${\operatorname{Cof}}_1$, of $\arr{ {\EuScript A}', {\EuScript B}'}\subset \operatorname{Fun}(\Delta^1, {\EuScript C})$.
This preimage is easily identified with the category ${\EuScript E}$ of possible
upper parts of MM-cubes. Therefore $q$ is a trivial Kan fibration as well.\qed
\vskip .5cm
Note that $ \operatorname{Fib}: \arr{ {\EuScript A}, {\EuScript B}} \to {\EuScript C}$ is an equivalence by definition of an SOD and
Proposition \ref{prop:<AB>}(a), but $ \operatorname{Fib}: \arr{ {\EuScript A}', {\EuScript B}'} \to {\EuScript C}$ is not necessarily an
equivalence, since $( {\EuScript A}', {\EuScript B}')$ may not be an SOD.
\begin{defi}\label{def:rel-susp}
The {\em relative suspension functor} of a biCartesian SOD is the
(essentially unique) functor $\tau: {\EuScript C}\to{\EuScript C}$ obtained as the composite
of an inverse of $ \operatorname{Fib} \circ p$ with $ \operatorname{Fib} \circ q$ in the diagram of
Lemma \ref{lem:cube}.
\end{defi}
\begin{exa}
Consider the trivial semiorthogonal
decomposition with $ {\EuScript A} = {\EuScript C}$ and $ {\EuScript B} \subset {\EuScript C}$ the full subcategory spanned by the zero
objects. Then
$\tau: {\EuScript C} \to {\EuScript C}$ is the usual suspension functor of the stable ${\infty}$-category ${\EuScript C}$.
\end{exa}
\begin{cor}\label{cor:sphericaltau} Let ${\EuScript C}$ be a stable $\infty$-category equipped with a
biCartesian SOD $( {\EuScript A}, {\EuScript B})$. Then the following conditions are
equivalent:
\begin{enumerate}
\item[(i)] $( {\EuScript A}, {\EuScript B})$ is $4$-periodic.
\item[(ii)] The relative suspension $\tau: {\EuScript C} \to {\EuScript C}$ is an equivalence.
\end{enumerate}
\end{cor}
\begin{proof}
Indeed, (i) is equivalent to the property that $( {\EuScript A}', {\EuScript B}')$ is an SOD. At the same
time, (ii) is equivalent to the property that $ \operatorname{Fib}: \arr{ {\EuScript A}', {\EuScript B}'}\to{\EuScript C}$ is an equivalence,
by the two out of three property of equivalences, since all the other functors
in the diagram of Lemma \ref{lem:cube} are equivalences. Now, the
property that $ \operatorname{Fib}: \arr{ {\EuScript A}', {\EuScript B}'}\to{\EuScript C}$ is an equivalence is equivalent to
$( {\EuScript A}', {\EuScript B}')$ being an SOD by Definition \ref{def:oo-SOD}.
\end{proof}
\paragraph{$4$-periodicity and spherical adjunctions.}
Let ${\EuScript C}$ be a stable $\infty$-category equipped with a biCartesian
semiorthogonal decomposition $( {\EuScript A}, {\EuScript B})$.
\begin{prop}\label{prop:taures}The relative suspension functor
$\tau$
restricts to the functors of $\infty$-categories given by the following formulas:
\[
\rho_{ {\EuScript B}}: {\EuScript A} \to {\EuScript A}',\; x \mapsto \operatorname{fib}(x \to f(x))[1],
\]
where $x \to f(x)$ is a coCartesian morphism in ${\EuScript C}$,
\[
\lambda_{ {\EuScript A}}[1]: {\EuScript B} \to {\EuScript B}',\; y \mapsto \operatorname{fib}(g(y) \to y)[1],
\]
where $g(y) \to y$ is a Cartesian morphism in ${\EuScript C}$,
\[
{\EuScript A}' \to {\EuScript A},\; \operatorname{fib}(x \to f(x)) \mapsto \operatorname{fib}(x \to g(f(x)))[1],
\]
where $x \to g(f(x))$ is a unit morphism,
\[
B' \to {\EuScript B},\; \operatorname{fib}(g(y) \to y) \mapsto \operatorname{fib}(f(g(y) \to y),
\]
where $g(f(y)) \to y$ is a counit morphism.
\end{prop}
\begin{proof} This can be directly read off from \eqref{eq:cubical}.
\end{proof}
\vskip .2cm
Further, let
$
F: {\EuScript A} \longleftrightarrow {\EuScript B}: G
$
be the left and right gluing functors for $( {\EuScript A}, {\EuScript B})$,
so $(F,G)$ form an adjoint pair, with $p: \chi( {\EuScript A}, {\EuScript B})\to\Delta^1$ providing the adjunction.
\begin{prop}\label{prop:sph=4per}
Let ${\EuScript C}$ be a stable $\infty$-category with a biCartesian semiorthogonal decomposition
$( {\EuScript A}, {\EuScript B})$. Then the following are equivalent:
\begin{enumerate}
\item[(i)] $( {\EuScript A}, {\EuScript B})$ is $4$-periodic.
\item[(ii)] The adjunction $p: \chi( {\EuScript A}, {\EuScript B}) \to \Delta^1$ is spherical.
\end{enumerate}
\end{prop}
\noindent{\sl Proof:}
By Proposition \ref{prop:taures}, the functor $\tau^2[-1]$ stabilizes the subcategories
$ {\EuScript A}$, $ {\EuScript B}$, as well as $ {\EuScript A}'$ and $ {\EuScript B}'$.
By Remark \ref{rem:sections}, this implies that
$\tau^2[-1]$ induces an endofunctor of the biCartesian fibration $\pi: \chi( {\EuScript A}, {\EuScript B}) \to
\Delta^1$ that preserves Cartesian and coCartesian edges. By \cite[3.3.1.5]{lurie:htt}, the
functor $\tau^2[-1]$ is an equivalence if and only if its restrictions to $ {\EuScript A}$ and $ {\EuScript B}$ are equivalences.
But the restriction of $\tau^2[-1]$ to $ {\EuScript A}$ and $ {\EuScript B}$
recovers the twist and cotwist functors from \eqref{eq:twist} and \eqref{eq:cotwist}, respectively,
defined as
\[
T_{ {\EuScript A}}: {\EuScript A} \longrightarrow {\EuScript A}, \; x \mapsto {\operatorname{Cof}}\{ x \to G(F(x))\},
\]
where $x \to G(F(x))$ is a unit morphism, and
\[
T_{ {\EuScript B}}: {\EuScript B} \longrightarrow {\EuScript B}, \; x \mapsto \operatorname{Fib}\{ F(G(y)) \to y\}
\]
where $F(G(y)) \to y$ is a counit morphism.
So $\tau^2[-1]$ is an equivalence if and only if i
the spherical twist and cotwist are equivalences. But $\tau^2[-1]$ is
an equivalence if and only if $\tau$ is an equivalence so that the claimed statement follows
from Corollary \ref{cor:sphericaltau}. \qed
\paragraph{The spherical functor package.}
\begin{cor}\label{cor:left-sph-right}
The concepts of left and right spherical ${\infty}$-functors (Definition \ref{def:sphad})
are equivalent.
\end{cor}
\noindent{\sl Proof:}
Indeed, let $(F: {\EuScript A} \leftrightarrow B: G)$ be an adjoint pair of spherical ${\infty}$-functors so $F$ is left spherical and $G$ is right spherical.
By Proposition \ref{prop:sph=4per},
it gives rise to a stable ${\infty}$-category ${\EuScript C}$ with a $4$-periodic
SOD $( {\EuScript A}, {\EuScript B})$ for which $F$ is a coCartesian gluing functor and $G$ is
a Cartesian gluing functor. Now $( {\EuScript B}, {\EuScript A}'={^\perp {\EuScript B}})$ is also a $4$-periodic
SOD of ${\EuScript C}$, a part of the same admissible chain of orthogonals, and
Proposition \ref{prop:bicart-adj} implies that its {\em Cartesian} gluing
functor is identified with $F$, so $F$ is right spherical. Similarly,
$(\mathcal{B}'= {\EuScript A}^\perp, {\EuScript A})$ is also a $4$-periodic SOD of ${\EuScript C}$ and its {\em coCartesian}
gluing functor is identified with $G$, so $G$ is left spherical. \qed
\vskip .2cm
We will therefore use the term {\em spherical ${\infty}$-functor}.
\begin{cor}\label{cor:sph-oo-adjoints}
Any spherical ${\infty}$-functor $F: {\EuScript A}\to {\EuScript B}$ has iterated
left and right adjoints of all orders, each of which
is a spherical ${\infty}$-functor itself.
\end{cor}
\noindent{\sl Proof:} This follows from Proposition \ref{cor:oo-adm-adj},
as a $4$-periodic SOD is ${\infty}$-admissible. \qed
\vskip .2cm
We will denote the iterated adjoints by
\[
\cdots {^{** }F}, {^*F}, F, F^* = G, F^{**} = G^*, \cdots
\]
\begin{lem}\label{lem:grid} Let $( {\EuScript A}, {\EuScript B})$ be a $4$-periodic semiorthogonal decomposition of a stable
$\infty$-category ${\EuScript C}$. Consider the $\infty$-category $\Di$ defined
in \eqref{eq:m}, so an object of $\Di$ is a diagram as in
\eqref{eq:mdiag}. For such a diagram,
the following are equivalent:
\begin{enumerate}
\item[(i)] The edge $b \to a$ is $( {\EuScript A}^{\perp}, {\EuScript A})$-Cartesian.
\item[(ii)] The edge $a' \to b'$ is $( {\EuScript A}, {\EuScript B})$-coCartesian.
\end{enumerate}
In other words, the base change equivalence
\[
\psi: \arr{ {\EuScript A}^{\perp}, {\EuScript A}} \simeq \arr{ {\EuScript A}, {\EuScript B}}
\]
identifies Cartesian with coCartesian edges.
\end{lem}
\begin{proof}
This is an immediate consequence of Lemma \ref{lem:flip}: The edge $b \to a$ being
$( {\EuScript A}^{\perp}, {\EuScript A})$-Cartesian is equivalent to $x$ being an object of $ {\EuScript A}^{\perp\perp}$. The
edge $a' \to b$ being $( {\EuScript A}, {\EuScript B})$-coCartesian is equivalent to $x$ being an object of
$^{\perp} {\EuScript B}$. But due to $4$-periodicity, we have $ {\EuScript A}^{\perp\perp} = {^{\perp} {\EuScript B}}$.
\end{proof}
\vskip .2cm
Usually, the definition of the spherical functor considered in the
literature includes various additional conditions, see \cite[Th. 1.1] {AL:spherical}.
The following corollary shows that these
additional conditions are actually consequences
of the requirement that the twist and cotwist are
equivalences.
\begin{cor}[(The spherical functor package)]
Let $(F: {\EuScript A} \leftrightarrow B: G)$ be an adjoint pair of spherical ${\infty}$-functors, so $G=F^*$ and $F={^*G}$. Then:
\begin{enumerate}
\item[(1)] The cotwist with respect to the adjoint pair
of spherical functors $(G, G^* = F^{**})$ is an inverse to the twist with
respect to $(F,G)$.
\item[(2)] The twist with respect to $({^*F} = {^{**}G} , F)$ is an inverse to the cotwist with
respect to $ (F,G)$.
\item[(3)] There is a canonical equivalence $F^{**} \simeq F T^{-1}_{ {\EuScript A}}$.
\item[(4)] There is a canonical equivalence ${^{**}G} \simeq T_{ {\EuScript A}}^{-1} G$.
\end{enumerate}
\end{cor}
\noindent{\sl Proof:}
To show (1), consider a diagram of the form \eqref{eq:mdiag} satisfying the equivalent conditions of Lemma
\ref{lem:grid}. The condition on $b \to a$ to be Cartesian implies that the morphism $y \to a$ is a
counit morphism for the adjoint pair $(G,G^*)$ modelled by the biCartesian semiorthogonal
decomposition $( {\EuScript A}^{\perp}, {\EuScript A})$. In particular, we have $y \simeq G(G^*(a))$, and the exact triangle
\begin{equation}\label{ex:triangle}
\begin{tikzcd}
a' \ar{r} \ar{d} & y\ar{d} \\
0 \ar{r} & a
\end{tikzcd}
\end{equation}
exhibits $a'$ as a cotwist of $a$, with respect to $(G, G^*)$. The condition that $a' \to b'$ be
coCartesian implies that the morphism $a' \to y$ is a unit for the adjoint pair
$(F,G)$ modelled
by $( {\EuScript A}, {\EuScript B})$. In particular, the exact triangle \eqref{ex:triangle} exhibits $a$ as a twist of $a'$
with respect to the $(F,G)$. We deduce that the cotwist with respect to
$(G,G^*)$ and the
twist with respect to $(F,G)$ are inverse to one another. This shows (1),
part (2) is similar.
Similarly, the statement $F \simeq G^* T_{ {\EuScript A}}$ follows from direct inspection of \eqref{eq:mdiag}
showing (3). Claim (4) then follows from (3) by passing to left adjoints.
\qed
We conclude by noting that, building on the above results, a version of the $2$-out-of-$4$ property,
describing equivalent characterizations of spherical adjunctions of stable $\infty$-categories, has
been established in \cite{christ}, generalizing the results of \cite{AL:spherical} for dg-categories.
\section{The spherical S-construction}\label{sec:spher-S}
In this chapter, we show that Waldhausen's relative S-construction admits a canonical
paracyclic structure when applied to a spherical functor of stable $\infty$-categories. The idea
that lies behind our proof is that the spherical adjunctions from Proposition \ref{prop:spherical}
organize into a coparacyclic object in the category $\on{Sph}$ of spherical adjunctions which
corepresents the relative S-construction. For technical convenience, we provide a more direct
combinatorial proof. We start with some reminders.
\subsection{The relative S-construction}\label{subsec:rel-S}
\paragraph{Cyclic and paracyclic objects.}
Along with the simplex category $\Delta$, we will use Connes' {\em cyclic category}
$\Lambda$ and its universal cover, the {\em paracyclic category} $\Lambda_{\infty}$
introduced by Elmendorf \cite{elmendorf} under a different name.
Here is a brief reminder, see
\cite[\S 2.1]{dkss:1} for a detailed survey.
Objects of $\Lambda$ and $\Lambda_{\infty}$ are formal symbols $\langle n \rangle$, $n\in\mathbb{Z}_+$
which can be thought of as finite cyclic ordinals $\mathbb{Z}/(n+1)\mathbb{Z}$, see
\cite{dk:crossed}. The set ${\operatorname{Hom}}_{\Lambda_{\infty}}(\langle m \rangle,\langle n \rangle)$ can be identified \cite{elmendorf}
with the set of monotone maps $f: \mathbb{Z}\to\mathbb{Z}$ such that
\begin{equation}\label{eq:Lambda-oo-monot}
f(i+m+1)=f(i)+n+1.
\end{equation}
The composition of morphisms is the composition of monotone maps.
The map $\tau_n: i\mapsto i+1$
is an invertible element of ${\operatorname{Hom}}_{\Lambda_{\infty}}(\langle n \rangle, \langle n \rangle)$ known as the
{\em paracyclic rotation}. The category $\Lambda$ can be obtained from
$\Lambda_{\infty}$ by imposing additional relations $\tau_n^{n+1}=\operatorname{Id}$.
Thus we have a functor $p: \Lambda_{\infty}\to\Lambda$ bijective on objects
and surjective on morphisms.
We have an embedding (functor bijective on objects and injective on morphsims)
${\varepsilon}: \Delta\hookrightarrow \Lambda_{\infty}$, $[n]\mapsto \langle n \rangle$
obtained by extending monotone maps $[m]\to[n]$ using \eqref{eq:Lambda-oo-monot}.
The set $\Mor(\Lambda_{\infty})$ of morphisms of $\Lambda_{\infty}$ is generated
by $\Mor(\Delta)$ and the automorphisms $\tau_n$.
The composition $p{\varepsilon}: \Delta\to\Lambda$ is also an embedding.
By a {\em paracyclic} (resp. {\em cyclic}) object in an ordinary category $\mathcal{C}$ we will mean
a contravariant functor $X$ from $\Lambda_{\infty}$ (resp. $\Lambda$) to $\mathcal{C}$ and write $X_n=X(\langle n \rangle)$.
Thus a (para)cyclic object can be seen as a simplicial object $X_\bullet$ together
with extra {\em (para)cyclic symmetry} given by the action of $\tau_n$ on $X_n$ for each $n$ and commuting,
in a certain definite way, cf. \cite[\S 2.1]{dkss:1} with the simplicial structure.
By a {\em (para)cyclic ${\infty}$-category} we will mean a (para)cyclic object ${\EuScript C}_\bullet$ in the category
of simplicial sets such that each ${\EuScript C}_n$ is an ${\infty}$-category.
\paragraph{The S-construction of a stable ${\infty}$-category.} The Waldhausen S-construction of a
stable (more generally, exact) ${\infty}$-category was originally defined in \cite[\S 7.3]{DK:HSS}.
In this paper we modify the approach of \cite{DK:HSS} to get this and related constructions take values in
${\infty}$-categories, not just in Kan complexes
(which appear as maximal ${\infty}$-subgroupoids in those
${\infty}$-categories).
For $n\geq 0$ let $J(n)$ be the poset formed by pairs $0\leq i<j\leq n$ with order $(i,j)\leq (k,l)$
iff $i\leq k$ and $j\leq l$. As usual, we consider $J(n)$ as a category.
Thus $J(n)=\Mor [n]$ is the category of morphisms of the poset $[n]$ considered as a category. The correspondence $[n]\mapsto J(n)$
defines a covariant functor $J: \Delta\to{\EuScript Cat}$.
\begin{defi} Let ${\EuScript C}$ be a stable ${\infty}$-category. We denote $S_n{\EuScript C}\subset\operatorname{Fun}(J(n),{\EuScript C})$
the full ${\infty}$-subcategory spanned by functors $Y: J(n)\to {\EuScript C}$ satisfying the conditions:
\begin{itemize}
\item[(1)] For each $0\leq i\leq n $ we have $Y(i,i)=0$ (is a zero object).
\item[(2)] For each $0\leq i\leq k\leq j\leq l\leq n$ the square
\[
\xymatrix{
Y(i,j)\ar[d] \ar[r]&Y(i,l)\ar[d]
\\
Y(k,j) \ar[r]& Y(k,l)
}
\]
\end{itemize}
is a biCartesian square in ${\EuScript C}$. When $n$ varies, the ${\infty}$-categories $S_n{\EuScript C}$ unite into a simplicial
${\infty}$-category $S_\bullet {\EuScript C}$ called the {\em Waldhausen S-construction} of ${\EuScript C}$.
\end{defi}
The following is well known \cite[Rem.4.1.2]{lurie:rotation}:
\begin{prop}\label{prop:A_n-quiv}
The projection $S_n({\EuScript C})\to\operatorname{Fun}(\Delta^{n-1}, {\EuScript C})$ given by restriction to
the subposet
\[
[n-1] \simeq \bigl\{ (0,1) <(0,2) <\cdots < (0,n)\bigr\} \,\subset \, J_n,
\]
is an equivalence of ${\infty}$-categories. In particular, $S_n({\EuScript C})$ is stable. \qed
\end{prop}
Note that $\operatorname{Fun}(\Delta^{n-1}, {\EuScript C})$ can be seen as the category of representations
of the $A_n$-quiver in ${\EuScript C}$.
If ${\EuScript C}=\operatorname{N}^{\operatorname{dg}}(\mathcal{V})$ is the dg-nerve of a $2$-periodic pre-triangulated dg-category $\mathcal{V}$, then
by \cite{DK:triangulated} $S_\bullet{\EuScript C}$ has a natural structure of a cyclic, not just a simplicial
${\infty}$-category. The cyclic symmetry $\tau_n$ is given by the
(derived) Coxeter functor of the $A_n$-quiver.
This was extended to ``$2$-periodic'' stable ${\infty}$-categories by Lurie \cite {lurie:rotation}. Lurie
further showed that for ${\EuScript C}$ an arbitrary stable ${\infty}$-category, $S_\bullet{\EuScript C}$ always has a
paracyclic structure, with $\tau_n^{n+1}=\Sigma^2$ being the functor of shift by $2$ (also see \cite{dyck:a1}).
We will obtain this latter statement as a consequence of our more general results below.
\paragraph{The relative $S$-construction of an exact functor.} Let $F: {\EuScript A}\to {\EuScript B}$ be an exact
functor of stable ${\infty}$-categories. For every $n\geq 0$ we have then the induced
${\infty}$-functor $F_*: S_n {\EuScript A}\to S_n {\EuScript B}$.
Following \cite[Def. 1.5.4] {waldhausen},
we define the ${\infty}$-category $S_n(F)$ as the pullback (in the category of simplicial sets)
\begin{equation}\label{eq:S-n-F}
\xymatrix{
S_n(F)\ar[d] \ar[r]& S_{n+1} {\EuScript B}\ar[d]^{{\partial}_{n+1}}
\\
S_n {\EuScript A} \ar[r]_{F_*} &S_n {\EuScript B}.
}
\end{equation}
When $n$ varies, the $S_n(F)$ unite into a simplicial ${\infty}$-category $S_\bullet(F)$
which we call the {\em relative Waldhausen S-construction} of $F$.
\begin{exas}\label{exas:S_n-F}
(a) We have $S_0(F)= {\EuScript B}$, while $S_1(F) = \operatorname{Fun}_{\Delta^1}(\Delta^1, \Gamma(F))$
is the category of sections of the covariant Grothendieck construction of $F$.
In other words, an object of $S_1(F)$ consists of objects $a\in {\EuScript A}$, $b\in {\EuScript B}$ and
a morphism $F(a)\to b$ in $ {\EuScript B}$. Thus $S_1(F)$ is the stable ${\infty}$-category
glued out of $ {\EuScript A}$ and $ {\EuScript B}$ using $F$ as a coCartesian gluing functor.
\vskip .2cm
(b) More generally, for any $n$ by applying Proposition
\ref{prop:A_n-quiv}
we identify $S_n(F)$ with the category of sections of the covariant
Grothendieck construction of the composite functor
\[
\operatorname{Fun}(\Delta^{n-1}, {\EuScript A}) \buildrel \operatorname{ev}_{n-1}\over\longrightarrow
{\EuScript A} \buildrel F\over \longrightarrow {\EuScript B},
\]
where $\operatorname{ev}_{n-1}$ is the functor of evaluation at the $(n-1)$st (last)
vertex of $\Delta^{n-1}$.
In particular, $S_n(F)$
is a stable ${\infty}$-category.
This also means that ``size-wise'',
$S_n(F)$ looks roughly like the sum of $n$ copies of $ {\EuScript A}$ and one copy of $ {\EuScript B}$.
For this reason $S_\bullet (F)$ can be seen as a categorical analog
of the nerve of the Picard groupoid $[\Psi\buildrel b\over \to\Phi]$ associated to a $2$-term
complex, i.e., a morphism of abelian groups $b: \Psi\to\Phi$. Indeed,
as an abelian group, $\operatorname{N}_n[\Psi\to\Phi] = \Psi^{\oplus n} \oplus\Phi$, cf.
\cite[Ex. 4.3.3]{dkss:1}. In fact, the relative S-construction is a particular case of the
{\em categorified Dold-Kan nerve} of \cite{dyckerhoff:DK} applied to the $2$-term complex
$ {\EuScript A}\buildrel F\over \to {\EuScript B}$ of stable ${\infty}$-categories.
\vskip .2cm
(c) Note that we could use any ${\partial}_i$, $i=0,\cdots, n+1$ instead of ${\partial}_0$
to form the pullback diagram above, and it would lead to an equivalent construction
of $S_n(F)$. This is because the paracyclic symmetry (Coxeter functor)
$\tau_{n+1}: S_{n+1} {\EuScript B}\to S_{n+1} {\EuScript B}$ rotates the ${\partial}_i$. This phenomenon will
be used later in the proof of Theorem \ref {thm:2-Seg}.
\end{exas}
\paragraph{Relative S-construction in terms of SOD.} Given $F: {\EuScript A}\to {\EuScript B}$ as above,
we can construct a stable ${\infty}$-category ${\EuScript C}$
with a coCartesian SOD $( {\EuScript A}, {\EuScript B})$
by using $F$ as a coCartesian gluing functor. Explicitly,
${\EuScript C} = \operatorname{Fun}_{\Delta^1}(\Delta^1, \Gamma(F))$. Let us describe the simplicial
${\infty}$-category $S_\bullet(F)$ in terms of ${\EuScript C}$. This can be compared with
the ``simplified'' description of the relative S-construction given by Waldhausen
in \cite[p. 344] {waldhausen}.
\begin{defi}\label{def:rel-pushf}
(a) Let ${\EuScript C}$ be a stable $\infty$-category equipped with a coCartesian SOD
$( {\EuScript A}, {\EuScript B})$. We say that a square
\[
\begin{tikzcd}
a \ar{r}\ar{d}{u} & b\ar{d}\\
a' \ar{r} & b'
\end{tikzcd}
\]
with $a,a'$ objects of $ {\EuScript A}$ and $b,b'$ objects of $ {\EuScript B}$, is a {\em relative pushout square} for $( {\EuScript A}, {\EuScript B})$, if the
right-hand square of the induced rectangle
\[
\begin{tikzcd}
a \ar{r}{!} \ar{d}{u} & t\ar{d}{v} \ar{r}& b\ar{d}\\
a' \ar{r}{!} & t' \ar{r} & b'
\end{tikzcd}
\]
is a biCartesian square in $ {\EuScript B}$.
(b) Dually, we define a {\em relative pullback square} with respect to a
Cartesian SOD as a relative pushout square in the opposite category.
\end{defi}
Let now $[n]^\rhd = \{0,1,\cdots, n, {\infty}\}$ be the poset obtained from $[n]$
by adjoining a maximal element ${\infty}$. As before,
let $W(n)= \Mor ([n]^\rhd)$ be
the poset formed by all pairs $(i\leq j)$ with $i,j\in [n]^\rhd$.
It is clear that the various posets $W(n)$ organize into a cosimplicial object
\begin{equation}\label{eq:simplicial}
\Delta \longrightarrow {\EuScript Cat},\; [n] \mapsto W(n).
\end{equation}
\begin{defi}\label{def:S-n-C-A-B}
Let ${\EuScript C}$ be a stable $\infty$-category equipped with a coCartesian semiorthogonal
decomposition $( {\EuScript A}, {\EuScript B})$ We define
$
S_n({\EuScript C}; ( {\EuScript A}, {\EuScript B})) \subset \operatorname{Fun}(W(n),{\EuScript C})
$
as the full subcategory spanned by those diagrams
$
X: W(n) \longrightarrow {\EuScript C}
$
satisfying:
\begin{enumerate}
\item The objects $X(i,{\infty})$, $0 \le i \le n$, lie in $ {\EuScript B}$.
\item The objects $X(i,j)$, $0 \le i \le j \le n$, lie in $ {\EuScript A}$.
\item For every $0 \le i \le {\infty}$, the object $X(i,i)$ is a zero object in
$ {\EuScript A}$.
\item For every $0 \le i \leq k \leq j \leq l \le n$, the square
\[
\begin{tikzcd}
X(i, j) \ar{r}\ar{d} & X(i,l) \ar{d}\\
X(k,j) \ar{r} & X(k,l)
\end{tikzcd}
\]
is a biCartesian square in $ {\EuScript A}$.
\item For every $0 \le i\le k\le j \le n$, the square
\[
\begin{tikzcd}
X(i,j) \ar{r}\ar{d} &X(i,{\infty}) \ar{d}\\
X(k,j) \ar{r} & X(k,{\infty})
\end{tikzcd}
\]
is a relative pushforward square with respect to $( {\EuScript A}, {\EuScript B})$.
\end{enumerate}
Via the functoriality from \eqref{eq:simplicial}, we obtain a simplicial $\infty$-category
\[
S_{\bullet}({\EuScript C};( {\EuScript A}, {\EuScript B})): \Delta^{{\operatorname{op}}} \longrightarrow {\EuScript Cat}_{\infty}
\]
which we call the {\em relative S-construction} of ${\EuScript C}$ with respect to
$( {\EuScript A}, {\EuScript B})$.
\end{defi}
\begin{prop}
The simplicial $\infty$-category
$S_{\bullet}({\EuScript C};( {\EuScript A}, {\EuScript B}))$ is equivalent to $S_\bullet(F)$, where $F: {\EuScript A}\to {\EuScript B}$
is a coCartesian gluing functor for $( {\EuScript A}, {\EuScript B})$.
\end{prop}
\noindent{\sl Proof:} To compare with the definition \eqref{eq:S-n-F}
of $S_\bullet(F)$, we
recall the interpretation of
the induced rectangle in part (a) of Definition \ref{def:rel-pushf} in terms of
$F$. Indeed, from the definition of $( {\EuScript A}, {\EuScript B})$-coCartesian arrows,
we have that the induced rectangle has
$t=F(a)$, $t'=F(a')$ and $v=F(u)$. After this identification,
Definition \ref{def:S-n-C-A-B} becomes identical to
the fiber product in
\eqref{eq:S-n-F}: introducing
the extra label ${\infty}$ corresponds to introducing the label $n+1$
in the definition of $S_n(F)$ by putting ${\partial}_{n+1}:S_{n+1} {\EuScript B}\to
S_n {\EuScript B}$
as the right vertical arrow of the fiber product diagram. \qed
\subsection{The relative S-construction of a spherical functor}
\label{subsec:rel-S-sph}
\paragraph{ The main result: paracyclic structure.}
We now formulate the main result of this paper whose proof will be finished in \S \ref{par:spher-vs-rel-S}
\begin{thm}\label{thm:main}
(i) Suppose that $F: {\EuScript A}\to {\EuScript B}$ is a spherical functor of stable ${\infty}$-categories.
The simplicial ${\infty}$-category $S_\bullet(F)$ has a natural lift
to a paracyclic ${\infty}$-category.
\vskip .2cm
(ii) Equivalently, suppose that $( {\EuScript A}, {\EuScript B})$ is a $4$-periodic SOD of a stable
${\infty}$-category ${\EuScript C}$. The simplicial ${\infty}$-category $S_{\bullet}({\EuScript C};( {\EuScript A}, {\EuScript B}))$
has a natural lift
to a paracyclic ${\infty}$-category.
\end{thm}
It suffices to prove the form (ii) of the theorem. To do this, we give another definition of
$S_{\bullet}({\EuScript C};( {\EuScript A}, {\EuScript B}))$ whose paracyclic nature will be clear from the outset.
This is analogous to the approach to the absolute S-construction in \cite[\S4.3] {lurie:rotation}.
\paragraph{The spherical S-construction.}
We start with extending each poset $W(n)$ as follows.
\begin{defi}\label{def:R(n)}
Let $\mathbb{Z}^{\rhd}$ denote the poset obtained from $\mathbb{Z}$ by adjoining a maximal element $\infty$. Similarly,
let $\mathbb{Z}^{\lhd}$ be $\mathbb{Z}$ with a minimal element adjoined,
denoted by $-\infty$. Let
$
R(n) \subset \mathbb{Z}^{\lhd} \times \mathbb{Z}^{\rhd}
$
be the poset consisting of:
\begin{itemize}
\item all pairs $(i,j)$ of integers such that $0 \le j-i \le n+1$,
\item all elements of the form $(-\infty,j)$ where $j \in \mathbb{Z}^{\rhd}$,
\item all elements of the form $(i,\infty)$ where $i \in \mathbb{Z}^{\lhd}$.
\end{itemize}
\end{defi}
\noindent It is evident that the various posets $R(n)$ organize into a coparacyclic object
\begin{equation}\label{eq:paracyclic}
R: \Lambda_{\infty} \longrightarrow {\EuScript Cat},\; \langle n \rangle \mapsto R(n).
\end{equation}
\iffalse
\begin{exa} For $n=1$, the poset $R(n)$ may be depicted as follows:
\begin{equation}\label{eq:rn}
\begin{tikzcd}
\dots & (-\infty,0) \ar{r}\ar{dd} & (-\infty,1)\ar{ddd} \ar{r} & (-\infty,2)\ar{dddd}
\ar{r} & (-\infty,3) \ar{ddddd}
\ar{r} & \dots \ar{r} & (-\infty,\infty) \ar{d} \\
& & & & & & \vdots \ar{d}\\
\ddots & (-2,0) \ar{rrrrr}\ar{d} & & & & &(-2,\infty) \ar{d} \\
& (-1,0) \ar{d}\ar{r} & (-1,1)\ar{d} \ar{rrrr} & & & &(-1,\infty) \ar{d} \\
& (0,0) \ar{r} & (0,1)\ar{d} \ar{r} & (0,2)\ar{d} \ar{rrr} & & &(0,\infty) \ar{d} \\
& & (1,1) \ar{r} & (1,2) \ar{r} & (1,3) \ar{rr} & &(1,\infty) \\
&&&&\ddots&& \vdots
\end{tikzcd}
\end{equation}
\end{exa}
\fi
\begin{defi}\label{defi:sphericalS} Let ${\EuScript C}$ be a stable $\infty$-category equipped with a $4$-periodic SOD $( {\EuScript A}, {\EuScript B})$ and set $ {\EuScript B}' = {\EuScript A}^{\perp}$, $ {\EuScript A}'={^\perp} {\EuScript B}$ so that every consecutive pair of the
sequence
\[
\cdots \quad , \quad {\EuScript B}'\quad,\quad {\EuScript A}\quad,\quad {\EuScript B}\quad , \quad {\EuScript A}' \quad , \quad {\EuScript B}' \quad , \quad \cdots
\]
forms an SOD. We define
\[
S_n^{\on{sph}}({\EuScript C}; ( {\EuScript A}, {\EuScript B}))\, \subset\, \operatorname{Fun}(R(n), {\EuScript C})
\]
to be the full subcategory spanned by those diagrams
$
X: R(n) \to {\EuScript C}
$
which satisfy the following conditions:
\begin{enumerate}
\item The objects $X(-\infty,j)$, $j \in \mathbb{Z}$ lie in $ {\EuScript B}'$.
\item The objects $X(i,j)$ for $-\infty < i \le j < \infty$ lie in $ {\EuScript A}$.
\item The objects $X(i,\infty)$, $i \in \mathbb{Z}$ lie in $ {\EuScript B}$.
\item For every $i \in \mathbb{Z}$, the object $X(i,i)$ is a zero object in ${\EuScript C}$.
\item The object $X(-{\infty}, {\infty})$ is also a zero object in ${\EuScript C}$.
\item For every $j \in \mathbb{Z}$, the morphism $X(-\infty,j) \to X(j-n-1,j)$ is $( {\EuScript B}', {\EuScript A})$-coCartesian (so its fiber lies in $ {\EuScript B}$).
\item For every $i \in \mathbb{Z}$, the morphism $X(i, i+n+1) \to X(i,\infty)$ is $( {\EuScript A}, {\EuScript B})$-Cartesian (so its fiber lies in $ {\EuScript B}'$).
\item For every $i \in \mathbb{Z}$, the square
\[
\begin{tikzcd}
X(-\infty, i+n+1) \ar{r}\ar{d}{!} & X(-\infty,\infty)=0 \ar{d}\\
X(i,i+n+1) \ar{r}{*} & X(i,\infty)
\end{tikzcd}
\]
is a biCartesian square in ${\EuScript C}$.
\item For every $-\infty < i\le k\le j \le l < \infty$ with $l-i \le n+1$, the square
\[
\begin{tikzcd}
X(i, j) \ar{r}\ar{d} & X(i,l) \ar{d}\\
X(k,j) \ar{r} & X(k,l)
\end{tikzcd}
\]
is a biCartesian square in $ {\EuScript A}$.
\item For every $-\infty < i \le k\le j < \infty$ with $j-i \le n+1$, the square
\[
\begin{tikzcd}
X(i, j) \ar{r}\ar{d} & X(i,\infty) \ar{d}\\
X(k,j) \ar{r} & X(k,\infty)
\end{tikzcd}
\]
is a relative pushout square with respect to $( {\EuScript A}, {\EuScript B})$.
\item For every $-\infty < i \leq j\leq l < \infty$ with $l-i \le n+1$, the square
\[
\begin{tikzcd}
X(-\infty, j) \ar{r}\ar{d} &X(i,j) \ar{d}\\
X(-\infty,l) \ar{r} & X(i,l)
\end{tikzcd}
\]
is a relative pullback square with respect to $( {\EuScript B}', {\EuScript A})$.
\end{enumerate}
Via the functoriality from \eqref{eq:paracyclic}, we obtain a paracyclic $\infty$-category
\[
S^{\operatorname{sph}}_{\bullet}({\EuScript C};( {\EuScript A}, {\EuScript B})): \Lambda_{\infty}^{{\operatorname{op}}} \longrightarrow {\EuScript Cat}_{\infty}
\]
which we call the {\em spherical $S$-construction} of ${\EuScript C}$ with respect to $( {\EuScript A}, {\EuScript B})$.
\end{defi}
\paragraph{Spherical vs. relative S-construction.}\label{par:spher-vs-rel-S}
Theorem \ref{thm:main} in the form (ii) would follow from the next result.
\begin{thm}\label{thm:sphericalS} Let ${\EuScript C}$ be a stable $\infty$-category equipped with a $4$-periodic SOD
$( {\EuScript A}, {\EuScript B})$. Then the functor given by restriction along the inclusions $W(n)
\subset R(n)$ induces an equivalence
\[
S^{\operatorname{sph}}_{\bullet}({\EuScript C};( {\EuScript A}, {\EuScript B}))|\Delta^{{\operatorname{op}}} \,\simeq \,
S_{\bullet}({\EuScript C};( {\EuScript A}, {\EuScript B}))
\]
between the simplicial object underlying the spherical $S$-construction and the relative
$S$-construction of ${\EuScript C}$ with respect to $( {\EuScript A}, {\EuScript B})$.
\end{thm}
\noindent{\sl Proof:}
To show this, we use an elaboration of the argument in the proof of
\cite[Prop.4.3.3]{lurie:rotation}, given by iterated application of \cite[4.3.2.15]{lurie:htt} using
the various universal properties and Proposition \ref{prop:cartesian} to control those properties
via relative Kan extensions. Most crucial will be Lemma \ref{lem:grid} which implies
that the mutation of an $( {\EuScript A}, {\EuScript B})$-relative pushout square is an $( {\EuScript B}', {\EuScript A})$-relative pullback square
and vice versa.
We prove that for each $n\geq 0$ the functor
\[
S^{\operatorname{sph}}_{n}({\EuScript C};( {\EuScript A}, {\EuScript B})) \longrightarrow S_{n}({\EuScript C};( {\EuScript A}, {\EuScript B}))
\]
given by restriction from $R(n)$ to
$W(n)$, is an equivalence.
To this end, we represent $R(n)$
as the union of an infinite chain of inclusions of subposets
\begin{equation}\label{eq:exhaustion}
W(n)=W_0(n) \subset W_1(n)\subset \cdots \subset R(n) = \bigcup_i W_i(n)
\end{equation}
and, for each $i\in\mathbb{Z}_{\geq 0}$, define a full subcategory $S_n^{(i)}\subset
\operatorname{Fun}(W_i(n), {\EuScript C})$ so that:
\begin{itemize}
\item[(1)] $S_n^{(0)}= S_n({\EuScript C}; ( {\EuScript A}, {\EuScript B}))$, and $S_n^{\on{sph}}({\EuScript C}; ( {\EuScript A}, {\EuScript B}))$ is
the limit of the $S_n^{(i)}$.
\item[(2)] Objects of $S_n^{(i)}$, $i\geq 1$, are precisely left or right
(depending on $i$) Kan extensions of those of $S_n^{(i-1)}$.
\end{itemize}
\noindent The statement about equivalence will then follow by iterated
application of \cite[4.3.2.15]{lurie:htt}, as all the
restrictions
\[
S_n({\EuScript C}; ( {\EuScript A}, {\EuScript B})) = S_n^{(0)} \longleftarrow S_n^{(1)} \longleftarrow \cdots\longleftarrow
S_n^{\on{sph}}({\EuScript C}; ( {\EuScript A}, {\EuScript B}))
\]
will be shown to be equivalences.
\vskip .2cm
We will define each $W_i(n)$ by adding a single element $(p_i,q_i)$ to $W_{i-1}(n)$,
and define $S_n^{(i)}$ as the category of diagrams $W_i(n)\to {\EuScript C}$ which
satisfy all those conditions of Definition \ref{defi:sphericalS} which make
sense within $W_i(n)$. The adding of the object $X(p_i,q_i)$ corresponding to
the new element will be referred to as {\em filling}. This object will
be determined from the previous part of the diagram either as some
pullback, in which case diagrams from $S_n^{(i)}$ will be right Kan extensions
of those from $S_n^{(i-1)}$, or as some pushout, in which case diagrams
from $S_n^{(i)}$ will be left Kan extensions.
To explain the order of adding elements, we subdivide the poset $R(n)$
into the following subsets which we visualize as geometric figures
in Fig. \ref{fig:domains}:
\begin{itemize}
\item ``Triangles'' $T_m$, $m\in\mathbb{Z}$ consisting of $(i,j)$ with $m(n+1) \leq i\leq j < (m+1)(n+1)$.
\item ``Triangles'' $T'_m$, $m\in \mathbb{Z}$, consisting of $(i\leq j)$ with $(m-1)(n+1) \leq i < m(n+1)$ and $m(n+1) \leq j < (m+1)(n+1)$.
\item ``Vertical intervals'' $V_m$, $m\in \mathbb{Z}$, consisting of $(i,{\infty})$ with $m(n+1) \leq i <
(m+1)(n+1)$.
\item ``Horizontal intervals'' $H_m$, $m\in\mathbb{Z}$, consisting of $(-{\infty},j)$
with $m(n+1) \leq j<
(m+1)(n+1)$.
\end{itemize}
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=0.4]
\draw (-1,0) -- (-6,5) --(-1,5) --(-1,0);
\draw (0,0) -- (0,5) -- (5,0) -- (0,0);
\draw (0,-1) --(5,-1) -- (5, -6) -- (0,-1);
\draw (6,-1) -- (6, -6) -- (11,-6) -- (6, -1);
\draw (15, 10) -- (15, -10);
\draw (-10, 10) -- (15, 10);
\draw[line width=2] (15,0) -- (15,5);
\draw[line width=2] (15,-1) -- (15,-6);
\draw[line width=2] (-6,10) -- (-1,10);
\draw[line width=2] (0,10) -- (5, 10);
\node at (9,-8) {\huge$\vdots$};
\node at (-3, 7) {\huge$\vdots$};
\node at (-3,3.5){$T_0$};
\node at (1.5,2){$T'_0$};
\node at (3, -2.5){$T_1$};
\node at (7.5, -4){$T'_1$};
\node at (16, 2.5){$V_0$};
\node at (16, -3.5){$V_1$};
\node at (-3.5, 11){$H_0$};
\node at (2.5, 11){$H_1$};
\node at (-8.5, 11){\huge$\cdots$};
\node at (7.5, 11){\huge$\cdots$};
\node at (16, -8){\huge$\vdots$};
\node at (16,7){\huge$\vdots$};
\node at (15,5){$\bullet$};
\node at (15,0){$\bullet$};
\node at (15,-1){$\bullet$};
\node at (15,-6){$\bullet$};
\node at (-6,10){$\bullet$};
\node at (-1,10){$\bullet$};
\node at (0,10){$\bullet$};
\node at (5,10){$\bullet$};
\node at (16.5,5){\small$(0,{\infty})$};
\node at (16.5,0){\small$(n,{\infty})$};
\node at (17.5,-1){\small$(n+1,{\infty})$};
\node at (17.5,-6){\small$(2n+1,{\infty})$};
\node[rotate=90] at (-6,12){\small$ (-{\infty}, 0)$};
\node[rotate=90] at (-1,12){\small$ (-{\infty}, n)$};
\node[rotate=90] at (0,13){\small$ (-{\infty}, n+1)$};
\node[rotate=90] at (5,13){\small$ (-{\infty}, 2n+1)$};
\node at (-6,5){$\bullet$};
\node at (-7, 5.3){\tiny$(0,0)$};
\node at (15,10){$\bullet$};
\node at (16, 11){\small$(-{\infty}, {\infty})$};
\node at (-1,5){$\bullet$};
\node[rotate=90] at (-1,6){\tiny$(0,n)$};
\node at (0,5){$\bullet$};
\node[rotate=45] at (1.5, 6.5) {\tiny$((0,n+1)$};
\node at (-1,0){$\bullet$};
\node at (-2.5,0){\tiny$(n,n)$};
\node at (5,0){$\bullet$};
\node[rotate=45] at (6.5,1.5){\tiny$(n, 2n+1)$};
\node at (0,0){$\bullet$};
\node at (6,-1){$\bullet$};
\node[rotate=45] at (8, 1){\tiny$(n+1, 2n+2)$};
\end{tikzpicture}
\caption{Subdivision of $R(n)$ into triangles and intervals.}\label{fig:domains}
\end{figure}
\noindent Our initial set $W_0(n) = W(n)$ is the union of $T_0$ and $V_0$.
We now start the filling procedure by filling the triangle $T'_0$ from top to bottom.
\vskip .2cm
So we take $(p_1, q_1)=(0, n+1)$, the top vertex of $T'_0$. The corresponding object
$X(0, n+1)$ is found by factoring $X(0,n)\to X(0,{\infty})$ as
$X(0,n)\to X(0, n+1)\buildrel *\over\to
X(0,{\infty})$ with the $*$-arrow being $( {\EuScript A}, {\EuScript B})$-Cartesian. This amounts to a
right Kan extension. Alternatively, in terms of the gluing functors $F: {\EuScript A}\leftrightarrow {\EuScript B}:G$
classifying the SOD $( {\EuScript A},
B)$, we
have $X(0, n+1)=G(X(0,{\infty})$ and the arrow $X(0,n+1)\to X(0,{\infty})$
involving is given by the counit of the
adjunction.
Next, we fill the object straight down, at $(p_2, q_2)=(1, n+1)$. This object $X(1, n+1)$
is found to fit into a biCartesian square
\[
\xymatrix{
X(0,n)\ar[d] \ar[r]& X(0, n+1)
\ar@{-->}[d]
\\
X(1,n) \ar@{-->}[r]& X(1, n+1),
}
\]
i.e., as a pushout. This amounts to a left Kan extension.
Next, we fill the object at $(p_3, q_3)=(1, n+2)$ by factoring $X(1, n+1)\to X(1,{\infty})$
as $X(1, n+1)\to X(1, n+2)\buildrel *\over\to X(1,{\infty})$.
Continuing like this, we will the entire triangle $T'_0$. We further fill $T_1$ in the same
way, row by row from the top by forming pushots and starting each row from
$X(i,i)=0$.
\vskip .2cm
After this, we fill $X(n+1,{\infty})$ in the vertical interval $V_1$ from the condition that
the left square below is an $( {\EuScript A}, {\EuScript B})$-relative pushout square in ${\EuScript C}$ or, equivalently
(in terms of the gluing functor $F$)
the right square below is biCartesian in $ {\EuScript B}$:
\[
\xymatrix{
X(n, 2n+1)\ar[d] \ar[r]& X(n,{\infty})\ar@{-->}[d]
\\
X(n+1, 2n+1) \ar@{-->}[r]& X(n+1, {\infty})
}
\quad\quad
\quad
\xymatrix{
F(X(n, 2n+1))\ar[d] \ar[r]& X(n,{\infty})\ar@{-->}[d]
\\
F(X(n+1, 2n+1)) \ar@{-->}[r]& X(n+1, {\infty}).
}
\]
This amounts to a left Kan extension.
Continuing like this we can fill, going down row by row, all the $T_m, T'_m, V_m$ with
$m\geq 0$. For this, we do not yet need the fact that our adjunction was spherical
or, equivalently, that $( {\EuScript A}, {\EuScript B})$ was $4$-periodic.
\vskip .2cm
To fill the entire diagram, we need to alternate the above ``downward moves'' with
the ``upward moves'' in the opposite direction which are as follows.
\vskip .2cm
Note that condition (8) of Definition \ref{defi:sphericalS} recovers the objects in $H_1$
as mutations of the corresponding objects in $V_0$, the objects in $H_2$ as
mutations of those in $V_1$ etc.
After doing this (and here we use $4$-periodicity), Lemma \ref{lem:grid} implies that the mutation of an $( {\EuScript A}, {\EuScript B})$-relative pushout square
is an $( {\EuScript B}', {\EuScript A})$-relative pullback square and vice versa. Thus the diagram with
the objects in $H_{\geq 1}$ filled in, will satisfy all the relevant conditions of
Definition \ref{defi:sphericalS}.
After this, already having $X(-{\infty}, n+1)$ (from $H_1$), we fill $X(-{\infty}, n)$
(from $H_0$ which is not yet filled) from the condition that it fits into $( {\EuScript B}', {\EuScript A})$-relative
pullback square
\[
\xymatrix{
X(-{\infty}, n) \ar@{-->}[r] \ar@{-->}[d]& X(-{\infty}, n+1)\ar[d]
\\
X(0,n) \ar[r]&X(0, n+1).
}
\]
This amounts to a right Kan extension.
\vskip .2cm
After this, we fill $H_0$ and the bottom row of $T'_{-1}$ from right to left.
That is, we find $X(-1,n)$, the object at the right end of the bottom row of
$T'_{-1}$ by factoring $X(-{\infty},n)\to X(0,n)$ as $X({\infty},n) \buildrel !\over\to X(-1,n)
\to X(0,n)$, which amounts to a left Kan extension.
Then we find $X(-1, n-1)$ to fit into a biCartesian square
\[
\xymatrix{
X(-1, n-1)\ar@{-->}[d] \ar@{-->}[r]& X(-1,n)\ar[d]
\\
X(0,n-1)\ar[r]& X(0,n),
}
\]
a right Kan extension, and so on.
\vskip .2cm
By alternating the two types of moves, downward and upward, we exhaust the set $R(n)$
by a single sequence to get a chain of inclusions as in
\eqref{eq:exhaustion}.
\qed
\paragraph{Examples: the spherical $S_0$ and $S_1$.}
Let us describe the low-dimensional cells of the spherical $S$-construction more explicitly and illustrate
the action of the paracyclic shift.
\begin{exa}
The $\infty$-category $S_0^{\on{sph}}({\EuScript C}, ( {\EuScript A}, {\EuScript B}))$ consists of diagrams of the form
\begin{equation}\label{eq:r0}
\adjustbox{scale=.8,center}{%
\begin{tikzcd}
{\mathbf{ \cdots}} \ar{r}& X(-\infty,0) \ar{r}\ar{dd}{!} & X(-\infty,1)\ar{ddd}{!} \ar{r} &
X(-\infty,2)\ar{dddd}{!}
\ar{r} & \dots \ar{r} & X(-\infty,\infty) \ar{d} \\
& & & & & \vdots \ar{d}\\
& X(-1,0) \ar{d}\ar{rrrr}{*} & & & &X(-1,\infty) \ar{d} \\
& X(0,0) \ar{r} & X(0,1)\ar{d} \ar{rrr}{*} & & &X(0,\infty) \ar{d} \\
& & X(1,1) \ar{r} & X(1,2) \ar{rr}{*} & &X(1,\infty) \ar{d}\\
&&&&&{\mathbf{ \vdots}}
\end{tikzcd}
}
\end{equation}
satisfying the conditions of Definition \ref{defi:sphericalS}. By the argument of the proof
of Theorem \ref{thm:sphericalS}, the projection map onto the subdiagram given by $X(0,\infty)$ is an equivalence,
showing that $S^{\operatorname{sph}}_0({\EuScript C};( {\EuScript A}, {\EuScript B})) \simeq {\EuScript B}$. To exhibit the action of
the paracyclic shift, which simply acts by shifting all indices by $+1$, we restrict attention to
the the subdiagram
\[
\begin{tikzcd}
X(0,1) \ar{r}{*}\ar{d} & X(0,\infty) \ar{d}\\
X(1,1)=0 \ar{r} & X(1,\infty)
\end{tikzcd}
\]
of \eqref{eq:r0} which, by the conditions of Definition \ref{defi:sphericalS}, is an
$( {\EuScript A}, {\EuScript B})$-relative pushout square. Refining this square to the rectangle
\[
\begin{tikzcd}
X(0,1) \ar{r}{!}\ar{d} & Y \ar{r}\ar{d}& X(0,\infty) \ar{d}\\
X(1,1)=0 \ar{r}{!} & Y'=0 \ar{r} & X(1,\infty)
\end{tikzcd}
\]
this means that the right-most square is biCartesian in $ {\EuScript B}$ (with $Y'=0$ since $X(1,1)=0$).
But this is simply saying that the
diagram exhibits $X(1,\infty)$ as $T_{ {\EuScript B}}(X(0,\infty))[1]$, since the map $Y \to X(0,\infty)$ is a
counit map for the adjunction classified by the biCartesian SOD $( {\EuScript A}, {\EuScript B})$.
Therefore, the paracyclic shift acts via the suspended spherical cotwist $T_{ {\EuScript B}}[1]$.
\end{exa}
\begin{exa}
The $\infty$-category $S_1^{\on{sph}}({\EuScript C}; ( {\EuScript A}, {\EuScript B}))$ consists of diagrams of the form
\[
\adjustbox{scale=.8,center}{%
\begin{tikzcd}
\cdots\ar{r} & X(-\infty,0) \ar{r}\ar{dd} & X(-\infty,1)\ar{ddd} \ar{r} & X(-\infty,2)\ar{dddd}
\ar{r} & X(-\infty,3) \ar{ddddd}
\ar{r} & \dots \ar{r} & X(-\infty,\infty) \ar{d} \\
& & & & & & \vdots \ar{d}\\
\ddots & X(-2,0) \ar{rrrrr}\ar{d} & & & & &X(-2,\infty) \ar{d} \\
& X(-1,0) \ar{d}\ar{r} & X(-1,1)\ar{d} \ar{rrrr} & & & &X(-1,\infty) \ar{d} \\
& X(0,0) \ar{r} & X(0,1)\ar{d} \ar{r} & X(0,2)\ar{d} \ar{rrr} & & &X(0,\infty) \ar{d} \\
& & X(1,1) \ar{r} & X(1,2) \ar{r} & X(1,3) \ar{rr} & &X(1,\infty)\ar{d} \\
&&&&{\mathbf{\ddots}}&&{\mathbf{ \vdots}}
\end{tikzcd}
}
\]
satisfying the conditions of Definition \ref{defi:sphericalS}. The projection functor to
the subdiagram
\[
X(0,1) \longrightarrow X(0,\infty)
\]
defines a trivial Kan fibration thus providing an equivalence $S^{\operatorname{sph}}_1({\EuScript C};( {\EuScript A}, {\EuScript B})) \simeq
\{ {\EuScript A}, {\EuScript B}\} \overset{\operatorname{fib}}{\simeq} {\EuScript C}$. To unravel the action of the paracyclic twist, consider the subdiagram
\[
\begin{tikzcd}
X(0,1) \ar{r}\ar{d} & X(0,2) \ar{r}{*}\ar{d} & X(0,\infty) \ar{d}\\
X(1,1) \ar{r} & X(1,2) \ar{r} & X(1,\infty)
\end{tikzcd}
\]
which we may refine to a cubical diagram
\[
\adjustbox{scale=.8,center}{%
\begin{tikzcd}
& Y \ar[rr]\ar[dd] & & X(0,\infty) \ar[dd]\\
X(0,1) \ar{ur}{!} \ar[rr]\ar[dd] & & X(0,2) \ar[dd]\ar{ur}{*} & \\
& Y' \ar[rr] & & X(1,\infty)\\
X(1,1) \ar{ur}{!} \ar[rr]& & X(1,2) \ar[ur] &
\end{tikzcd}
}
\]
Since the front and back faces of this cube are biCartesian squares, the cofiber of the
morphism $X(1,2) \to X(1,\infty)$ in ${\EuScript C}$ is equivalent to the totalization of top
commutative square
\[
\begin{tikzcd}
X(0,1) \ar{r}{!} \ar{d} & Y \ar{d}\\
X(0,2) \ar{r}{*} & X(0,\infty)
\end{tikzcd}
\]
in ${\EuScript C}$. But therefore, by inspection of \eqref{eq:cubical} and the subsequent definitions,
the fiber of $X(1,2) \to X(1,\infty)$ is precisely the image of $\operatorname{fib}(X(0,1) \to
X(0,\infty))$ under the relative suspension functor $\tau$ from Definition \ref{def:rel-susp}. In
conclusion, the action of the paracyclic shift on $S^{\operatorname{sph}}_1({\EuScript C};( {\EuScript A}, {\EuScript B})) \simeq {\EuScript C}$ is
given by the relative suspension functor $\tau$.
In the case when $ {\EuScript B} = 0$, so that the relative $S$-construction simply becomes the absolute
$S$-construction of $ {\EuScript A}$, we have $\tau = [1]$ in agreement with the known paracyclic
structure in this case.
\end{exa}
\subsection{Spherical functors and admissible paracyclic stable $\infty$-categories}\label{subsec:sph-admis}
In this section we note some important additional features of the relative S-construction $S_\bullet(F)$ (for any exact $F$, spherical or not).
We conjecture that these features, together with the paracyclic structure
which is specific to the spherical case, characterize the spherical S-constructions,
\paragraph{Admissibility: faces adjoint to degenerations.} \label{par:admis}
Suppose that ${\EuScript E}_{\bullet}: \Delta^{{\operatorname{op}}} \to {\EuScript Cat}_{\infty}$ is a simplicial $\infty$-category, i.e., a simplicial object in $ {\operatorname{Set}}_\Delta$ such that each ${\EuScript E}_n$ is
an ${\infty}$-category. Then the
simplicial identities imply the following relations between the face and degeneracy maps relating
${\EuScript E}_{n-1}$ and ${\EuScript E}_n$:
\begin{equation}\label{eq:face-deg-rels}
{\partial}_0s_0\buildrel =\over\to \operatorname{Id}_{{\EuScript E}_{n-1}}, \quad
\operatorname{Id}_{{\EuScript E}_{n-1}}\buildrel =\over\to {\partial}_1 s_0,\quad
{\partial}_1 s_1\buildrel =\over\to \operatorname{Id}_{{\EuScript E}_{n-1}},
\quad \cdots , \quad
\operatorname{Id}_{{\EuScript E}_{n-1}} \buildrel =\over\to {\partial}_n s_{n-1}.
\end{equation}
\begin{defi}\label{def:admis}
We say that ${\EuScript E}_{\bullet}$ is {\em admissible}, if the above identity maps form counit and
units, respectively, of adjunctions
\[
{\partial}_0 \dashv s_0 \dashv {\partial}_1 \dashv \dots \dashv s_{n-1} \dashv {\partial}_n.
\]
Analogously, we call a paracyclic $\infty$-category admissible if the chain of adjunctions continues
infintely on both sides. We denote $({\EuScript Cat}_{\infty})_\Delta^{\operatorname{adm}}$,
resp. $({\EuScript Cat}_{\infty})_{\Lambda_{\infty}}^{\operatorname{adm}}$ the full subcategory in
$({\EuScript Cat}_{\infty})_\Delta$, resp. $({\EuScript Cat}_{\infty})_{\Lambda_{\infty}}$
spanned by admissible simplicial, resp. paracyclic ${\infty}$-categories.
\end{defi}
\begin{prop}
(a) Let $F: {\EuScript A}\to {\EuScript B}$ be an exact functor of stable ${\infty}$-categories.
The relative S-construction $S_\bullet(F)$ is an admissible
simplicial stable ${\infty}$-category.
\vskip .2cm
(b) Let ${\EuScript C}$ be a stable $\infty$-category equipped with a $4$-periodic semiorthogonal
decomposition $( {\EuScript A}, {\EuScript B})$. The spherical S-construction
$S^{\on{sph}}_\bullet({\EuScript C}; ( {\EuScript A}, {\EuScript B}))$ is an admissible paracyclic stable ${\infty}$-category.
\end{prop}
\noindent {\sl Proof:} It is enough to show (a), since (b) then follows, in virtue of Theorem \ref{thm:main},
by applying an appropriate paracyclic rotation to bring any desired pair $({\partial}_i, s_i)$ or $(s_i, {\partial}_{i+1})$
of paracyclic face and degeneracy operators into the simplicial range.
So let us focus on (a) and consider the statement about the adjunction $({\partial}_i, s_i)$.
By Proposition \ref{prop:oo-unit}, it suffices
to prove that for any
two objects $X\in S_n(F)$, $Y\in S_{n-1}(F)$,
the map of the mapping spaces
\[
\eta: \operatorname{Map}_{S_n(F)}(X, s_i Y)
\buildrel \simeq\over\longrightarrow \operatorname{Map}_{S_{n-1}(F)}({\partial}_i X, Y).
\]
induced by the identity map ${\partial}_is_i\buildrel =\over \to\operatorname{Id}$ from \eqref{eq:face-deg-rels},
is an equivalence. But $s_i$ consists in adding zeroes to the diagram $Y$,
and we can think of this added $0$ as an initial and final object in the strict sense.
With this understanding,
both the source and the target of $\eta$ involve exactly the same data, so $\eta$ is an identification
``on the nose''. The statement about the adjunction $(s_i, {\partial}_{i+1})$ is similar. \qed
\vskip .2cm
The patterns of adjunctions of Definition \ref{def:admis}
can be implemented universally by
upgrading the categories $\Delta$ and $\Lambda_{\infty}$ to
(classical, strict) $2$-categories $\DDelta$, $\LLambda_{\infty}$
in which the corresponding coface and codegeneration maps become
$1$-morphisms equipped with canonical adjunction data (unit and counit
$2$-morphisms).
\vskip .2cm
Following \cite[\S3.1.3]{dyckerhoff:DK}, we define
$\DDelta$ to have the same objects (finite ordinals $[m]$) and $1$-morphisms
(monotone maps) as $\Delta$. For two
monotone maps $f, f': [m]\to [n]$, there is a unique $2$-morphism
$f\Rightarrow f'$ if and only if $f\leq f'$ pointwise.
In other words, $\DDelta$ is obtained by regarding each poset $[m]$
as a category and considering ${\operatorname{Hom}}_\Delta([m], [n])$, the set of
monotone maps, not just as a set but as a poset, therefore a category,
and this category is ${\operatorname{Hom}}_\DDelta([m], [n])$. In $\DDelta$,
we have ``internal'' adjunctions corresponding to those of
Definition \ref{def:admis}. In each unit-counit pair of these adjunctions,
one member, corresponding to
\eqref{eq:face-deg-rels}, is an identity $2$-morphism but the
other one is a $2$-morphism represented by a nontrivial inequality $f \leq f'$.
Passing to nerves, we may
interpret $\DDelta$ as enriched in ${\EuScript Cat}_{\infty}$.
\begin{rem}
A more ``relaxed'' (and augmented) version of the $2$-category $\DDelta$ can be obtained
as ${\operatorname{Hom}}_{{\bf{Adj}}^{\leq 3}}(a,a)$, where
${\bf{Adj}}^{\leq 3}$ is the Gray $3$-category generated by the universal
$2$-dimensional adjunction, see \cite { macdonald-scull, macdonald-stone}
and
Remark \ref{rem:adj-sphere} above. This $2$-category can be seen
as generated by the {\em universal $2$-comonad}, just as the augmented
simplex category $\Delta_+$ is generated by the universal comonad
in the classical sense, see \cite{{auderset}, macdonald-scull}.
\end{rem}
Similarly, to define $\LLambda_{\infty}$, we interpret ${\operatorname{Hom}}_{\Lambda_{\infty}}(\langle m \rangle, \langle n \rangle)$
as the set of quasi-periodic monotone maps $\mathbb{Z}\to\mathbb{Z}$, see
\eqref {eq:Lambda-oo-monot}, make it into a poset by pointwise
inequality and interpret it as a category ${\operatorname{Hom}}_{\LLambda_{\infty}}(\langle m \rangle, \langle n \rangle)$.
We regard $\LLambda_{\infty}$ as enriched in ${\EuScript Cat}_{\infty}$ as well.
Note that the category ${\EuScript Cat}_{\infty}$ can itself be considered as enriched in
${\EuScript Cat}_{\infty}$.
\begin{defi}
A {\em 2-simplicial} (resp. {\em $2$-paracyclic}) ${\infty}$-category
is a ${\EuScript Cat}_{\infty}$-enriched
functor from $\DDelta^{({\operatorname{op}},-)}$ (resp. from $\LLambda_{\infty}^{({\operatorname{op}}, -)}$) to ${\EuScript Cat}_{\infty}$. We denote by $({\EuScript Cat}_{\infty})_{\DDelta}$ (resp.
$({\EuScript Cat}_{\infty})_{\LLambda_{\infty}}$)
the ${\infty}$-category formed by $2$-simplicial, resp. $2$-paracyclic
${\infty}$-categories.
\end{defi}
Thus a $2$-simplicial ${\infty}$-category is equipped with the
``missing'' adjunction natural transformations completing each identity transformation in
\eqref{eq:face-deg-rels} to a unit-counit pair. Similarly for $2$-paracyclic ${\infty}$-categories.
\begin{prop} The forgetful functors
$
({\EuScript Cat}_{\infty})_{\DDelta} \to ({\EuScript Cat}_{\infty})_{\Delta}
$
and
$
({\EuScript Cat}_{\infty})_{\LLambda_{\infty}} \to ({\EuScript Cat}_{\infty})_{\Lambda_{\infty}}
$
take values in the full subcategories of admissible objects.
\end{prop}
\noindent {\sl Proof:} This follows from the property, discussed above,
that in $\DDelta$ as well as in $\LLambda_{\infty}$ the corresponding
members of the unit-counit pairs are identity $2$-morphisms. \qed
\begin{conj} The forgetful functors
$
({\EuScript Cat}_{\infty})_{\DDelta} \to ({\EuScript Cat}_{\infty})_{\Delta}^{{\operatorname{adm}}}
$
and
$
({\EuScript Cat}_{\infty})_{\LLambda_{\infty}} \to ({\EuScript Cat}_{\infty})_{\Lambda_{\infty}}^{{\operatorname{adm}}}
$
induce equivalences of $\infty$-categories.
\end{conj}
In other words, the datum of a $2$-simplicial (resp. $2$-paracyclic) lift of a given simplicial
(resp. paracyclic) $\infty$-category is a property and not an extra datum.
For example, as already pointed out in
Example \ref{exas:S_n-F}(b), the relative S-construction is a particular case of the categorified
Dold-Kan nerve of \cite{dyckerhoff:DK}. It was established in \cite{dyckerhoff:DK}
that the relative Dold-Kan nerve always comes with a natural $2$-simplicial structure.
\paragraph{$2$-Segal property.}
Recall from \cite {DK:HSS} the concept of a $2$-Segal simplicial object.
For $n\geq 2$ let
\[
P_n\,=\,{\operatorname{Conv}}\bigl\{ v_n = e^{2\pi i k/{(n+1)}}, \quad k=0,1,\cdots, n\bigr\} \,\subset
\mathbb{C}
\]
be the regular convex $(n+1)$-gon with vertices labelled by $0,1,\cdots, n$
counterclockwise. Each triangulation $\mathcal{T}$ of $P_n$ into straight triangles
lifts to a $2$-dimensional simplicial subset $\Delta^\mathcal{T}\subset\Delta^n$
by lifting each triangle $[v_i, v_j, v_k]$ into the $2$-simplex of $\Delta^n$
with vertices $i,j,k$.
Let ${\mathbf{C}}$ be a combninatorial model category (see \cite[\S4.1]{DK:HSS}
and references therein) and $Y_\bullet$ be a simplicial object in ${\mathbf{C}}$.
The simplicial subset $\Delta^\mathcal{T}\subset\Delta^n$ above gives rise to the
{\em derived membrane space} $Y_\mathcal{T} = (\Delta^\mathcal{T}, Y_\bullet)_R \in \operatorname{Ob}({\mathbf{C}})$
which is the homotopy limit (fiber product) in ${\mathbf{C}}$ of the diagram formed
by copies of $Y_i$ associated to each non-degenerate $i$-simplex of
$\Delta^\mathcal{T}$, $i=1,2$. It comes with a natural morphism $f_\mathcal{T}: Y_n\to Y_\mathcal{T}$.
The simplicial object $Y_\bullet$ is called {\em $2$-Segal}, if $f_\mathcal{T}$
is an equivalence for each $n\geq 2$ and each triangulation $\mathcal{T}$ of $P_n$.
We apply this to the {\em Joyal model structure} on the category $ {\operatorname{Set}}_\Delta$
of simplicial sets \cite{cisinski}. Fibrant objects in this structure are
${\infty}$-categories and, more generally, fibrations are so-called inner
fibrations of simplicial sets, see also \cite[\S2.3]{lurie:htt}. Equivalences between fibrant
objects are equivalences of ${\infty}$-categories.
Let $F: {\EuScript A}\to {\EuScript B}$ be an exact functor of stable ${\infty}$-categories. The relative
S-construction $S_\bullet(F)$, being a simplicial ${\infty}$-category is, in
particular, a simplicial object in $ {\operatorname{Set}}_\Delta$.
\begin{thm}\label{thm:2-Seg}
$S_\bullet(F)$ is $2$-Segal with respect to the Joyal model
structure.
\end{thm}
\begin{rems}
(a) This result generalizes and refines \cite[Th.7.3.3]{DK:HSS}
(specialized to the case of stable ${\infty}$-categories). That is, we consider the
relative and not just the absolute S-construction as in \cite{DK:HSS}.
Further, we consider each $S_n(F)$ as an ${\infty}$-category rather than
taking the maximal ${\infty}$-subgroupoid (i.e., the maximal Kan subcomplex)
there as in \cite{DK:HSS}.
\vskip .2cm
(b) In the interpretation of $S_n(F)$ as the Fukaya category of the disk
with coefficients in a schober given by $F$, see the Introduction,
the $2$-Segal property corresponds to the descent condition
such as holds, e. g., for cohomology with coefficients in a perverse sheaf,
cf. \cite{dkss:1}.
\end{rems}
\noindent{\sl Proof of Theorem \ref{thm:2-Seg}:}
We apply the path space criterion \cite[Th.6.3.2]{DK:HSS} reducing the
$2$-Segal property of a simplicial object $Y_\bullet$ to the much
simpler $1$-Segal (or simply Segal) property of the two {\em path space
objects}
\[
P^\lhd Y_\bullet \,=\, i_\lhd^*\, Y_\bullet, \quad P^\rhd Y_\bullet \,=\,
i_{\rhd}^*\, Y_\bullet.
\]
Here $i_\lhd, i_\rhd: \Delta\hookrightarrow\Delta$ are two self-embeddings of $\Delta$
(considered as the category of nonempty finite ordinals) given by adding an extra minimal, resp. maximal element to each ordinal.
Thus the $n$-simplices and the face maps of $ P^\lhd\, Y_\bullet,
P^\rhd Y_\bullet$ are given by
\[
(P^\lhd Y_\bullet)_n\,=\, (P^\rhd Y_\bullet)_n\,=\, Y_{n+1},\quad
{\partial}_i^{P^\lhd Y_\bullet}={\partial}_{i+1}^Y, \,\,\,
{\partial}_i^{P^\rhd Y_\bullet}={\partial}_{i}^Y,\,\,\, i=0,\cdots, n.
\]
Recall that a simplicial object $Z_\bullet$ in ${\mathbf{C}}$ is called $1$-Segal,
if for any $n\geq 1$ the natural morphism
\[
f_n: Z_n\longrightarrow Z_1\times^R_{Z_0} Z_1\times^R_{Z_0} \cdots
\times^R_{Z_0} Z_1
\]
to the $n$-fold homotopy fiber product is an equivalence.
We apply the criterion to $Y_\bullet=S_\bullet(F)$ and first look at $P^\rhd Y_\bullet$. The object $(P^\rhd S_\bullet(F))_n$ being, by definition,
nothing but $S_{n+1}(F)$,
Example \ref{exas:S_n-F}(b) identifies it with the category of sections
of the covariant Grothendieck construction of the
functor
\[
F(\operatorname{ev}_n): \operatorname{Fun}(\Delta^n, {\EuScript A})\longrightarrow {\EuScript B}.
\]
Further, in this identification the simplicial structure on $P^\rhd S_\bullet(F)$
is given by the functoriality of $\Delta^n$ in $[n]$. This means
that the target of the morphism $f_n$ for $Z_\bullet =P^\rhd S_\bullet(F)$
is identifed with the category of sections
of the covariant Grothendieck construction of the
functor
\[
F(\operatorname{ev}'_n): \operatorname{Fun}\bigl(\Delta^1 \amalg_{\Delta^0} \Delta^1 \amalg_{\Delta^0} \dots \amalg_{\Delta^0}
\Delta^1, \, {\EuScript A} \bigr) \longrightarrow {\EuScript B},
\]
where $\operatorname{ev}'_n$ is the functor of evaluation at the last vertex of the chain
of $(\Delta^1)$'s. The morphism $f_n$ itself is induced by restriction
with respect to the embedding
\[
\Delta^1 \amalg_{\Delta^0} \Delta^1 \amalg_{\Delta^0} \dots \amalg_{\Delta^0}
\Delta^1 \, \subset \, \Delta^n.
\]
This embedding being an inner-anodyne morphism \cite[Def. 2.0.0.3]{lurie:htt},
it follows that $f_n$ is equivalence. This shows that
$P^\rhd S_\bullet(F)$ is $1$-Segal.
\vskip .2cm
To analyze the other path space $P^\lhd$, we first use a seemingly different model
of the relative S-construction. For $n\geq 0$ define the ${\infty}$-category
$S'_n(F)$ by the fiber product as in \eqref{eq:S-n-F} but with ${\partial}_0$, not
${\partial}_{n+1}$ as the right vertical arrow. We unite the various $S'_n(F)$
into a simplicial ${\infty}$-category $S'_\bullet(F)$ by using the simplicial
structure on the $S_{\bullet+1}( {\EuScript B})$, the top right corners of the diagrams
\eqref{eq:S-n-F}, via the self-embedding $i_\lhd: \Delta\hookrightarrow\Delta$.
This is, in fact, the original definition of the relative S-construction in
\cite{waldhausen}. Now, Example \ref{exas:S_n-F}(b) is immediately
modified to this case. It identifies $(P^\lhd S'_\bullet(F))_n= S'_{n+1}(F)$
with the category of sections of the {\em contravariant} Grothendieck
construction of the functor
\[
F(\operatorname{ev}_0): \operatorname{Fun}(\Delta^n, {\EuScript A}) \longrightarrow B,
\]
where $\operatorname{ev}_0$ is the evaluation at the $0$th vertiex of $\Delta^n$.
The simplicial structure on $P^\lhd S'_\bullet(F)$ is again given by
the functoriality of $\Delta^n$ in $[n]$. So the same argument as before
shows that $P^\lhd S'_\bullet(F)$ is $1$-Segal.
\vskip .2cm
It remains to notice that the simplicial ${\infty}$-categories $S_\bullet(F)$
and $S'_\bullet(F)$ are in fact equivalent. The equivalence is obtained
by using the paracyclic structure on $S_\bullet( {\EuScript B})$
(given by \cite[Prop.4.3.3]{lurie:rotation} or applying our
Theorem \ref{thm:sphericalS} to the spherical functor $ {\EuScript B}\to 0$).
This structure gives, for each $n\geq 0$, the paracyclic rotation
$\tau_{n+1}: S_{n+1}( {\EuScript B})\to S_{n+1}( {\EuScript B})$ which sends ${\partial}_{n+1}$
to ${\partial}_0$ and shifts the other ${\partial}_i$.
This identification $S_\bullet(F)\simeq S'_\bullet(F)$ shows that both
$P^\rhd$ and $P^\lhd$ of this simplicial ${\infty}$-category are $1$-Segal, so
it is $2$-Segal. \qed
\paragraph{Combining the properties.}
Combining all the properties established in this and the previous sections,
we arrive at:
\begin{cor}
Let $F: {\EuScript A}\to {\EuScript B}$
be a spherical functor of stable ${\infty}$-categories.
Then the relative $S$-construction
$S_\bullet(F)$ is an admissible $2$-Segal paracyclic stable
$\infty$-category. \qed
\end{cor}
Let $\on{Sph}$ be the ${\infty}$-category formed by spherical adjunctions, and
${\EuScript S}t_{\Lambda_{\infty}}^{{\operatorname{adm}}, \text{$2$-Segal}}$ be
the ${\infty}$-category formed by stable $\infty$-categories and admissible $2$-Segal paracyclic
stable $\infty$-categories.
\begin{conj}\label{conj:sphad=adm2Segpar}
The relative $S$-construction induces an equivalence
$
\on{Sph} \longrightarrow {\EuScript S}t_{\Lambda_{\infty}}^{{\operatorname{adm}}, \text{$2$-Segal}}.
$
\end{conj}
|
\section{#1} \vspace{0mm}}
\newcommand{\SubSection}[1]{\vspace{0mm} \subsection{#1} \vspace{-1mm}}
\newcommand{\SubSubSection}[1]{\vspace{-1mm} \subsubsection{#1} \vspace{-1mm}}
\newcommand{\Paragraph}[1]{\vspace{1.25mm}\noindent\textbf{#1.}\hspace{0.5mm}}
\newcommand{{\color{green}\checkmark}}{{\color{green}\checkmark}}
\newcommand{{\color{red}\ding{55}}}{{\color{red}\ding{55}}}
\newcommand{\firstkey}[1]{\textcolor{red}{\textbf{#1}}}
\newcommand{\secondkey}[1]{\textcolor{blue}{\textbf{#1}}}
\newcommand{\marcos}[1]{\textcolor{cyan}{\parbox{\linewidth}{\ \\Marcos: #1}}}
\newcommand\Mark[1]{\textsuperscript#1}
\def\cvprPaperID{10612}
\defCVPR 2021{CVPR 2021}
\begin{document}
\begin{textblock*}{\textwidth}(0cm,0cm)
\large\noindent{\copyright~2021 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.}
\end{textblock*}
\thispagestyle{empty}
\title{Radar-Camera Pixel Depth Association for Depth Completion}
\author{
Yunfei Long\Mark{1}, Daniel Morris\Mark{1}, Xiaoming Liu\Mark{1}, \\
Marcos Castro\Mark{2},
Punarjay Chakravarty\Mark{2},
and Praveen Narayanan\Mark{2} \\
\Mark{1}Michigan State University, \Mark{2}Ford Motor Company \\
{\tt\small \{longyunf,dmorris,liuxm\}@msu.edu},
{\tt\small \{mgerard8,pchakra5,pnaray11\}@ford.com}
}
\twocolumn[{%
\renewcommand\twocolumn[1][]{#1}%
\maketitle
\pagenumbering{arabic}
\vspace{-1mm}
\noindent\begin{minipage}{\linewidth}
\begin{center}
\captionsetup{font=small}
\begin{tabular}{@{}c@{}c@{}c@{}}
\includegraphics[trim = 190 27 210 30, clip, width=0.32\linewidth]{./Figures/02_rd_raw} &
\includegraphics[trim = 190 27 210 30, clip, width=0.32\linewidth]{./Figures/02_expand} &
\includegraphics[trim = 190 27 140 30, clip, width=0.37\linewidth]{./Figures/02_prd_our} \vspace{-1mm}\\
{\small (a)} & {\small (b)} & {\small (c)} \\
\end{tabular}
\vspace{-3mm}
\captionof{figure}{\small Radar-camera depth completion: (a) an image with $0.3$ seconds ($5$ sweeps) of radar hits projected onto it, (b) enhanced radar depths at confidence level $0.9$ eliminate occluded pixels and expand visible hits, and (c) final predicted depth through depth completion. }
\label{fig:radar_augmentation}
\end{center} \vspace{4mm}
\end{minipage}
}]
\begin{abstract}
While radar and video data can be readily fused at the detection level, fusing them at the pixel level is potentially more beneficial.
This is also more challenging in part due to the sparsity of radar, but also because automotive radar beams are much wider than a typical pixel combined with a large baseline between camera and radar, which results in poor association between radar pixels and color pixel.
A consequence is that depth completion methods designed for LiDAR and video fare poorly for radar and video. Here we propose a radar-to-pixel association stage which learns a mapping from radar returns to pixels.
This mapping also serves to densify radar returns. Using this as a first stage, followed by a more traditional depth completion method, we are able to achieve image-guided depth completion with radar and video. We demonstrate performance superior to camera and radar alone on the nuScenes dataset.
Our source code is available at \url{https://github.com/longyunf/rc-pda}.
\end{abstract}
\vspace{-2mm}
\section{Introduction}
We seek to incorporate automotive radar as a contributing sensor to $3$D scene estimation.
While recent work fuses radar with video for the objective of achieving improved object detection~\cite{Chadwick:DistantVehicle:2019,Nabati:RadarRegionProposal:2019,lim:2019:radar,Meyer:2019:radar,Nobis:2019:radar}, here we aim for pixel-level fusion of depth estimates, and ask if {\it fusing video with radar can lead to improved dense depth estimation of a scene}.
Up to the present, outdoor depth estimation has been dominated by LiDAR, stereo, and monocular techniques. The fusion of LiDAR and video has lead to increasingly accurate dense depth completion~\cite{Imran2019}.
At the same time, radar has been relegated to the task of object detection in vehicle's Advanced Driver Assistance Systems (ADAS)~\cite{Marti:ADAS:2019}. However, phased array automotive radar technologies have been advancing in accuracy and discrimination~\cite{Hasch:AutomotiveRadar:2015}. Here we investigate the suitability of using radar instead of LiDAR for the task of dense depth estimation. Unlike LiDAR, automotive radars are already ubiquitous, being integrated in most vehicles for collision warning and similar tasks. If successfully fused with video, radar could provide an inexpensive alternative to LiDARs for $3$D scene modeling and perception.
However, to achieve this, attentive algorithm design is required in order to overcome some of the limitations of radar, including coarser, lower resolution, and sparser depth measurements than typical LiDARs.
This paper proposes a method to fuse radar returns with image data and achieve depth completion; namely a dense depth map over pixels in a camera. We develop a two-stage algorithm. The first stage builds an association between radar returns and image pixels, during which we resolve some of the uncertainty in projecting radar returns into a camera.
In addition, this stage is able to filter occluded radar returns and ``densify'' the projected radar depth map along with a confidence measure for these associations
(see Fig.~\ref{fig:radar_augmentation} (a,b)). Once a faithful association between radar hits and camera pixels is achieved, the second stage uses a more standard depth completion approach to combine radar and image data and estimate a dense depth map, as in Fig.~\ref{fig:radar_augmentation}(c).
A practical challenge to our fusion goal is the lack of public datasets with radar.
KITTI~\cite{Geiger2012CVPR}, the dataset used most extensively for LiDAR depth completion, does not include radar and nor do the Waymo~\cite{sun2020scalability} or ArgoVerse~\cite{chang2019argoverse} datasets.
The main exception is nuScenes~\cite{nuscenes2019} and the small Astyx~\cite{Meyer:RadarDataset:2019} which have radar, but unfortunately do not include a dense, pixel-aligned depth map as created by Uhrig~\emph{et al.}~\cite{Uhrig2017THREEDV}. Similarly, the Oxford Radar Robot Car dataset~\cite{RadarRobotCarDatasetICRA2020} includes camera, LiDAR and raw radar data, but no annotations are available for scene understanding.
As a result, all experiments of this work will use the nuScenes dataset along with its annotations. However, we find single LiDAR scans insufficient to train depth completion, and so accumulate scans to build semi-dense depth maps for training and evaluating depth completion.
The main contributions of this work include:
\begin{itemize}
\item Radar-camera pixel depth association that upgrades the projection of radar onto images and prepares a densified depth layer.
\item Enhanced radar depth that improves radar-camera depth completion over raw radar depth.
\item LiDAR ground truth accumulation that leverages optical flow for occluded pixel elimination, leading to higher quality dense depth images.
\end{itemize}
\section{Related Work}
\Paragraph{Radar for ADAS}
Frequency Modulated Continuous Wave radars are inexpensive and all-weather, and have served as the key sensor for modern ADAS. Ongoing advances are improving radar resolution and target discrimination~\cite{Hasch:AutomotiveRadar:2015}, while convolutional networks has been used to add discriminative power to radar data, moving beyond target detection and tracking to include classifying road environments~\cite{Lee:RadarRoadRecognition:2019,Sim:radar-for-road-environments:2020}, and seeing beyond-line-of-sight targets~\cite{Scheiner:DopplerSeeingAroundCorners:2020}. Nevertheless, the low spatial resolution of radar means that the $3$D environment, including object shape and classification, are only coarsely obtained. A key path to upgrading the capabilities of radar is through integration with additional sensor modalities~\cite{Marti:ADAS:2019}.
\Paragraph{Radar-camera fusion}
Early fusion of video with radar, such as~\cite{ji:radar:2008}, relied on radar for cueing image regions for object detection or road boundary estimation~\cite{Janda:RoadBoundary:2013}, or used optical flow to improve radar tracks~\cite{Garcia:RadarFlowTracks:2012}.
With the advent of deep learning, much more extensive multi-modal fusion has become possible~\cite{Feng:MultiModealFusion:2020}.
However, to the best of our knowledge, no prior work has conducted pixel-level dense depth fusion between radar and video.
\Paragraph{Radar-camera object detection}
Object detection is a key task in $3$D perception~\cite{kinematic-3d-object-detection-in-monocular-video}.
There has been significant recent interest in combining radar with video for improved object detection. In~\cite{Chadwick:DistantVehicle:2019}, ResNet blocks~\cite{He:2016:DeepResNet} are used to combine both color images and image-projected radar returns to improve longer-range vehicle detection.
In~\cite{lim:2019:radar}, an FFT applied to raw radar data generates a polar detection array which is merged with a bird's eye projection of the camera image, and targets are estimated with a single shot detector~\cite{liu:2016:ssd}. In~\cite{Meyer:2019:radar}, features from both images and a bird's-eye representation of radar enter a region proposal network that outputs bounding boxes~\cite{Simonyan:VGG:2015}. An alternative model for radar hits is a $3$m vertical line on the ground plane which is projected into the image plane by~\cite{Nobis:2019:radar}, and combined with VGG blocks to classify vehicle detections at multiple scales.
Our work differs fundamentally from these methods in that our goal is dense depth estimation, rather than object classification. But we do share similarity in radar representation: we project radar hits into an image plane. However, the key novelty in our work is that we learn a neighborhood pixel association model for radar hits, rather than relying on projected circles~\cite{Chadwick:DistantVehicle:2019} or lines~\cite{Nobis:2019:radar}.
\Paragraph{LiDAR-camera depth completion}
Our task of depth estimation has the same goal as LiDAR-camera depth completion~\cite{jaritz2018sparse,qiu2019deeplidar,xu2019depth,Imran2019,depth-completion-with-twin-surface-extrapolation-at-occlusion-boundaries}. However, radar is far sparser than LiDAR and has lower accuracy, which makes these methods unsuitable for this task.
Our radar enhancement stage densifies the projected radar depths, followed by a more traditional depth completion architecture.
\Paragraph{Monocular depth estimation}
Monocular depth inference may be supervised by LiDAR or self-supervised.
Self-supervised methods learn depth by minimizing photometric error between images captured by cameras with known relative positions.
Additional constrains such as semantics segmentation~\cite{zhu2020edge, ramirez2018geometry}, optical flow~\cite{janai2018unsupervised, vijayanarasimhan2017sfm}, surface normal~\cite{yang2017unsupervised} and proxy disparity labels~\cite{watson2019self} improve performance. Recently, self-supervised PackNet~\cite{Guizilini:packnet:2020} has achieved competitive results. Supervised methods~\cite{fu2018deep, diaz2019soft} include continuous depth regression and discrete depth classification. BTS~\cite{lee2019big} achieves state of the art by improving upsampling via additional plane constraints, and more recently~\cite{ren2020suw} combines supervised and self-supervised methods.
Our goal is not monocular depth estimation, but rather to improve what is achievable from monocular depth estimation through fusion with radar.
\begin{figure*}[t!]
\captionsetup{font=small}
\centering
\includegraphics[width=\textwidth]{Figures/diagram}
\vspace{-6mm}
\caption{\small Our two-stage architecture. Network $1$ learns $N$-channel \emph{radar-camera pixel depth association} (RC-PDA), here illustrated for two radar pixels (marked with white squares) on their neighboring pixels (white boxes). The RC-PDA is converted into a \emph{multi-channel enhanced radar} (MER), and input to Network $2$ which performs image-guided depth completion. \vspace{-3mm}}
\label{fig:network_arch}
\end{figure*}
\Section{Method}
While there are a variety of data-spaces in which radar can be fused with video, the most natural, given our objective of estimating a high resolution depth map, is in the image space. But this immediately presents a problem: to which pixel in an image does a {\it radar pixel} belong? By radar pixel we mean a simple point projection of the estimated $3$D radar hit into the camera. The nuScenes dataset~\cite{nuscenes2019} provides extrinsic and intrinsic calibration parameters needed to map the radar point clouds from the radar coordinates system to the egocentric and camera coordinate systems.
Assuming that the actual depth of the image pixel is the same as the radar pixel depth turns out to be fairly inaccurate. We describe some of the problems with this model, and then propose a new pixel association model. We present a method for building this new model and show its benefit by incorporating it into radar-camera depth completion. Fig.~\ref{fig:network_arch} shows the diagram of the proposed method.
\SubSection{Radar Hit Projection Model}
Single-row scanning automotive radars can be modeled as measuring points in a plane extending usually horizontally (relative to the vehicle platform) in front of the vehicle, as in~\cite{Chadwick:DistantVehicle:2019}. While radars can measure accurate depth, often the depth they give when projected into a camera is incorrect, as can be seen in the examples in Fig.~\ref{fig:radar_examples}. An important source of this error is the large width of radar beams which means that the hits extend well beyond the assumed horizontal plane. In other words, the height of measured radar hits is inaccurate~\cite{nabati2021centerfusion}.
\begin{figure}[t!]
\captionsetup{font=small}
\centering
\includegraphics[trim=6 0 0 0,clip,width=\linewidth]{Figures/radar_hits1}
\vspace{-7mm}
\caption{\small Examples of radar hits projected into a camera. While the hits project into the vicinity of the target that they hit, their image position can be quite different from their actual location. For example, radar depths in the yellow/red box are larger/smaller than corresponding image depths (meters).}
\label{fig:radar_examples}
\end{figure}
In addition to beam width, another source of projected point depth difference is occlusion caused by the significant baseline between radars on the grill, and cameras on the roof or driver mirror. Further, these depth differences only increase when radar hits are accumulated over a short interval, and thus more opportunity for occlusions.
In addition to pixel association errors, we are faced with the problem that automotive radars generate far sparser depth scans than LiDAR. There is typically a {\it single row} of returns, rather than anywhere up to $128$ rows in LiDAR, and the azimuth spacing of radar returns can be an order of magnitude greater than LiDAR. This sparsity significantly increases the difficulty in depth completion. One solution is to accumulate radar pixels over a short time interval, and to account for their $3$D position using both ego-motion and radial velocity. Nevertheless, this accumulation introduces additional pixel association errors (in part from not having tangential velocity) and more opportunities for occlusions.
\begin{figure*}[t!]
\captionsetup{font=small}
\centering
\begin{tabular}{@{}c@{}c@{}}
\includegraphics[trim=196 245 239 145,clip,width=0.54\linewidth]{Figures/radar_detections.pdf} &
\includegraphics[trim=160 300 270 50,clip,width=0.45\linewidth]{Figures/radar_detections_2.pdf} \vspace{-1mm}\\
{\small (a)} & {\small (b)} \vspace{-3mm}\\
\end{tabular}
\caption{\small Illustration of depth differences between camera and radar, and how our proposed association method (Pixel Depth Association: RC-PDA) can address this. (a) Radar hits are modeled on a ground-parallel plane (dashed black line). Actual returns may be outside this plane, as illustrated with orange stars $A,B,C$ at depths $D_A,D_B,D_C$ respectively. We project the corresponding in-plane points, $A_p,B_p,C_p$ (green diamonds), into the camera, and call these the \emph{radar pixels}. (b) The camera view showing the radar pixels $A_p,B_p,C_p$. Now the true image depth of these pixels is $D_A$, the front of the truck, which agrees only with $A_p$ which is visible, and not for $B_p$ and $C_p$ which are occluded. This illustrates why radar pixel depths are often incorrect from the camera perspective. Finding associations from radar pixels to the projected true points $A,B,C$ would solve this, but is difficult. Rather, we seek a neighborhood \emph{depth association} for each radar pixel, that specifies which pixels within a neighborhood (dashed blue regions) have the same depth as the radar pixel, shown here by the orange regions. For example, the orange pixels in the neighborhood of $A_p$ have a RC-PDA of $1$ while the remaining neighborhood pixels have a RC-PDA of $0$, all relative to $A_p$. See Sec.~\ref{sec:RCPDA} for details. }
\label{fig:radar_projection_model}
\end{figure*}
\SubSection{Radar-Camera Pixel Depth Association}
\label{sec:RCPDA}
In using radar to aid depth estimation we face the problem of determining which, if any, point in the image does a radar return correspond to? This radar pixel to camera pixel association is a difficult problem, and we do not have ground truth to determine this. Thus we reformulate this problem slightly to make it more tractable.
The new question we ask is: ``Which pixels in the vicinity of the projected radar pixel have the same depth of that radar return?" We call this \emph{Radar-Camera Pixel Depth Association}: RC-PDA or simply PDA.
It is a one-to-many mapping, rather than one-to-one mapping, and has four key advantages.
First, we do not need to distinguish between many good but ambiguous matches and rather can return many pixels with the same depth.
This simplifies the problem.
Second, by associating the radar return with multiple pixels, our method explicitly densifies the radar depth map, which facilitates the second stage of full-image depth estimation.
Third, our question simultaneously addresses the occlusion problem; if there are no nearby pixels with that depth, then the radar pixel is automatically inferred to be occluded.
Fourth, we are able to leverage a LiDAR-based ground-truth depth map as the supervision, rather than a difficult-to-define ``ground truth" pixel association.
Fig.~\ref{fig:radar_projection_model} illustrates image depths obtained from raw radar projections and RC-PDA around each radar pixel. It shows the height errors of measured radar points and how some hits visible to the radar are occluded from the camera.
\SubSubSection{RC-PDA Model}
We model RC-PDA over a neighborhood around the projected radar pixel in the color image.
At each radar pixel we define a patch around the radar location and seek to classify each pixel in this patch as having the same depth or not as the radar pixel, within a predetermined threshold. A similar connectivity model has been used for image segmentation~\cite{kampffmeyer2018connnet}.
Radar pixels and the patches around them are illustrated in Figs.~\ref{fig:PDA_MER}(a) and (b), respectively.
\begin{figure}[t!]
\captionsetup{font=small}
\centering
\begin{tabular}{@{}c@{}c@{}c@{}}
\includegraphics[trim=140 320 640 80,clip,width=0.32\linewidth]{Figures/PDA_1.pdf} &
\includegraphics[trim=140 320 640 80,clip,width=0.32\linewidth]{Figures/PDA_2.pdf} &
\includegraphics[trim=140 320 640 80,clip,width=0.32\linewidth]{Figures/PDA_3.pdf} \vspace{-1mm}\\
{\small (a)} & {\small (b)} & {\small (c)} \vspace{-2mm}\\
\end{tabular}
\caption{\small Overview of our radar depth representation. (a) Radar pixels indicate sparse depth in image space. (b) For each radar pixel, a Pixel Depth Association (RC-PDA) probability over neighboring pixels is calculated, indicated with shaded contours. (c) Radar pixel depths are propagated to neighboring pixels to create a Multi-channel Enhanced Radar (MER) image. Each channel is a densified depth at a given confidence level. \vspace{-4mm}}
\label{fig:PDA_MER}
\end{figure}
The connection to each pixel in a $h\times w$ neighborhood has $N=w h$ elements, and can be encoded as an $N$-channel RC-PDA which we label $\mathbf{A}(i,j,k)$, where $k=1,\cdots,N$.
Here $(i,j)$ is the radar pixel coordinate, and the $k$'th neighbor has offset $(i_k,j_k)$ from $(i,j)$. Now the label for $\mathbf{A}(i,j,k)$ is 1 if the neighboring pixel has the same depth as radar pixel and 0 otherwise. More precisely, if $E_{ijk}=d(i,j)-d_T(i+i_k,j+j_k)$ is the difference between radar pixel depth, $d(i,j)$, and the neighboring LiDAR pixel depth, $d_T(i+i_k,j+j_k)$, and $\tilde{E}_{ijk}=E_{ijk}/d(i,j)$ is the relative depth difference, then:
\begin{equation}
\label{eq:aff_label}
\mathbf{A}(i,j,k) =\left\{
\begin{aligned}
1 & , & \text{if } (|E_{ijk}| < T_a) \wedge (|\tilde{E}_{ijk}| < T_r) \\
0 & , & \text{otherwise}.
\end{aligned}
\right.
\end{equation}
We note that labels $\mathbf{A}(i,j,k)$ are only defined when there is both a radar pixel at $(i,j)$ and a LiDAR depth $d_T(i+i_k,j+j_k)$. We define a binary weight $w(i,j,k)\in\{0,1\}$ to be $1$ when both conditions are satisfied and $0$ otherwise. During training we minimize the weighted binary cross entropy loss~\cite{kampffmeyer2018connnet} between labels $\mathbf{A}(i,j,k)$ and predicted RC-PDA:
\begin{equation}
\begin{split}
\mathcal{L}_{CE} = \sum_{i,j,k} w(i,j,k)[&-\mathbf{A}(i,j,k) z(i,j,k) \\
& + \log(1 + \exp(z(i,j,k)))].
\end{split}
\label{eq:sce_loss}
\end{equation}
The network output, $z(i,j,k)$, is passed through a Sigmoid to obtain $\hat{\mathbf{A}}(i,j,k)$, the estimated RC-PDA.
Our network thus predicts a RC-PDA confidence in a range of $0$ to $1$ representing the probability that each pixel in this patch has the same depth as the radar pixel. This prediction also applies to the image pixel at the same coordinates as the radar pixel, $\it{i.e.}$ $(i,j)$, as like other pixels, the depth at this image pixel may differ from the radar depth for a variety of reasons, including those illustrated in Fig.~\ref{fig:radar_projection_model}.
\SubSection{From RC-PDA to MER}
\label{sec:pda_mer}
The RC-PDA gives the probability that neighboring pixels have the same depth as the measured radar pixel. We can convert the radar depths along with predicted RC-PDA into a partially filled depth image plus a corresponding confidence as follows.
Each of $N$ neighbors to a given radar pixel is given depth $d(i,j)$ and confidence $\hat{\mathbf{A}}(i,j,k)$.
If more than one radar depth is expanded to the same pixel, the radar depth with the maximum RC-PDA is kept.
The expanded depth is represented as $\mathbf{D}(i,j)$ with confidence $\hat{\mathbf{A}}(i,j)$.
Now many of the low-confidence pixels will have incorrect depth.
Instead of eliminating low-confidence depths, we convert this expanded depth image into a multi-channel image where each channel $l$ is given depth $\mathbf{D}(i,j)$ if its confidence $\hat{\mathbf{A}}(i,j)$ is greater than a channel threshold $T_l$, where $l=1,\cdots,N_e$ and $N_e$ is the total number of channels of the enhanced depth.
The result is a \emph{Multi-Channel Enhanced Radar} (MER) image with each channel representing radar-derived depth at a particular confidence level (see Fig.~\ref{fig:PDA_MER}(c)).
Our MER representation for depth can correctly encode many complex cases of radar-camera projection, a few of which are illustrated in Fig.~\ref{fig:radar_projection_model}. These cases include when radar hits are occluded and no nearby pixels have similar depth. They also include cases where the radar pixel is just inside or just outside the boundary of a target. In each case, those nearby pixels with the same depth as the radar can be given the radar depth with high confidence, while the remaining neighborhood pixels are given low confidence, and their depth are specified on separate channels of the MER.
The purpose of using multiple channels for depths with different confidences in MER is to facilitate the task of Network $2$ in Fig.~\ref{fig:network_arch} in performing the dense depth completion. High confidence channels give the greatest benefit, but low confidence channels may also provide useful data. In all cases they densify the depth beyond single radar pixels, easing the depth completion task.
\SubSection{Estimating RC-PDA}
We next select inputs to Network $1$ in Fig.~\ref{fig:network_arch} from which it can learn to infer the RC-PDA.
These are the image, the radar pixels with their depths as well as the image flow and the radar flow from current to a neighboring frame. Here we briefly explain the intuition for each of these.
The image provides scene context for each radar pixel, as well as object boundary information. The radar pixels provide depth for interpreting the context and a basis for predicting the depth of nearby pixels. As radar is very sparse, we accumulate radar from a short time history, $0.3$ seconds, and transform it into the current frame using both ego-motion and the radial velocity similar to that done in~\cite{Nobis:2019:radar}.
Now a pairing of image optical flow and radar scene flow provides an occlusion and depth difference cue. For static objects, the optical flow should exactly equal radar scene flow, when the pixel depth is the same as radar pixel depth. Conversely, radar pixels that are occluded from the camera view will have different scene flow from the optical flow of a static object occluding them (Fig.~\ref{Figure:flows}). Similarly, objects moving radially will have consistent flow. By providing flow, we expect that Network $1$ will learn to leverage flow similarity in predicting RC-PDA for each radar pixel.
\begin{figure}[t!]
\captionsetup{font=small}
\centering
\scalebox{1.1}{
\begin{tabular}{@{}c@{}c@{}c@{}}
\includegraphics[width=1.17in]{./Figures/radar_flow1} &
\includegraphics[width=0.89in]{./Figures/radar_flow3} &
\includegraphics[width=0.88in]{./Figures/radar_flow2} \vspace{-1mm}\\
\footnotesize{ (a) Radar depth} & \footnotesize{ (b) Optical flow} & \footnotesize{ (c) Radar flow} \vspace{-2mm}\\
\end{tabular} }
\caption{\small An example of how radar scene flow and optical flow differences are used to infer occlusions of radar pixels. The radar flows are plotted as yellow if the $L_2$ norm of radar/optical flow differences are larger than a threshold. Note that we do not explicitly filter radar, rather provide flow to Network $1$ in Fig.~\ref{fig:network_arch} so that it can implicitly filter radar while estimating RC-PDA. \vspace{-3mm}}
\label{Figure:flows}
\end{figure}
\begin{figure}[t!]
\captionsetup{font=small}
\centering
\begin{tabular}{@{}c@{}c@{}c@{}}
\includegraphics[trim=0 0 0 0,clip,width=2.8cm]{./Figures/artifact1} &
\includegraphics[trim=0 0 0 0,clip,width=2.8cm]{./Figures/artifact3} &
\includegraphics[trim=0 0 0 0,clip,width=2.8cm]{./Figures/artifact2} \vspace{-1mm}\\
\footnotesize{ (a) } & \footnotesize{ (b)} & \footnotesize{ (c) } \vspace{-3mm}\\
\end{tabular}
\caption{\small We noticed that when LiDAR with a regular scan pattern, as in (b) for image (a), is used to train depth completion, our network learns to predict the LiDAR points well, but not the remaining pixels. This leaves large artifacts, as in (c), and motivates us to create a semi-dense depth LiDAR training set. \vspace{-2mm}}
\label{fig:artifacts}
\end{figure}
\SubSection{LiDAR-based Supervision}
\label{sec:lidar-supervision}
To train both the RC-PDA and the final dense depth estimate, we use a dense ground truth depth. This is because, as illustrated in Fig.~\ref{fig:artifacts}, training with sparse LiDAR leads to significant artifacts. We now describe how we build a semi-dense depth image from LiDAR scans.
\SubSubSection{LiDAR Accumulation}
To our knowledge, there is no existing public dataset specially designed for depth completion with radar.
Thus we create a semi-dense ground truth depth from nuScenes dataset, a public dataset with radar data and designed for object detection and segmentation.
We use the $32$-ray LiDAR as depth label and notice that the sparse depth label generated from a single frame will lead to a biased model predicting depth with artifacts, {\it i.e.}, only predictions for pixels with ground truth are reasonable.
Thus, we use semi-dense LiDAR depth as label, which is created by accumulating multiple LiDAR frames.
With ego motion and calibration parameters, all static points can be transformed to destination image frame.
Moving points are compensated by bounding box poses at each frame, which are estimated by interpolating bounding boxes provided by nuScenes in key frames.
\SubSubSection{Occlusion Removal via Flow Consistency}
\label{sec:flowconsistency}
When a foreground object occludes some of the accumulated LiDAR points, the resulting dense depth may include depth artifacts as the occluded pixels appear in gaps in the foreground object. KITTI~\cite{Uhrig2017THREEDV} takes advantages of the depth from stereo images to filter out such occluded points.
As no stereo images are available in nuScenes, we propose detecting and removing occluded LiDAR points based on optical-scene flow consistency.
\begin{figure}[t!]
\captionsetup{font=small}
\centering
\scalebox{1.09}{
\begin{tabular}{@{}c@{}c@{}c@{}}
\includegraphics[width=1.17in]{./Figures/lidar_flow1} &
\includegraphics[width=0.9in]{./Figures/lidar_flow3} &
\includegraphics[width=0.9in]{./Figures/lidar_flow2} \vspace{-1mm}\\
\footnotesize{ (a) LiDAR depth} & \footnotesize{ (b) Optical flow} & \footnotesize{ (c) LiDAR flow} \vspace{-2mm}\\
\end{tabular} }
\caption{\small An example of how LiDAR scene flow and optical flow differences are used to infer occlusions of LiDAR pixels. LiDAR flows are plotted as yellow if the $L_2$ norm of LiDAR/optical flow differences are larger than a threshold. This is used in the accumulation of LiDAR for building ground-truth depth maps, see Fig.~\ref{Figure:flow_filter}. \vspace{-6mm}}
\label{Figure:flows_lidar}
\end{figure}
The scene flow of LiDAR points, termed \emph{LiDAR flow}, is computed by projecting LiDAR points into two neighboring images and measuring the change in their coordinates.
On moving objects, the point's positions are corrected with the object motion.
On static visible objects, LiDAR flow will equal optical flow, while on occluded surfaces LiDAR flow is usually different from the optical flow at the same pixel, see Fig.~\ref{Figure:flows_lidar}.
We calculate optical flow with~\cite{teed2020raft} pretrained on KITTI, and measure the difference between the two flows at the same pixel via the $L_2$ norm of their difference.
Points with flow difference larger than a threshold $T_f$ are discarded as occluded points. Fig.~\ref{Figure:flow_filter} shows an example of using flow consistency to filter out occluded LiDAR depths.
\begin{figure}[t!]
\captionsetup{font=small}
\centering
\begin{tabular}{@{}c@{}c@{}}
\includegraphics[trim=60 20 100 20,clip,width=1.53in]{./Figures/before_filter} &
\includegraphics[trim=60 20 60 20,clip,width=1.75in]{./Figures/after_filter} \vspace{-2mm} \\
{\small (a) } & {\small (b) }
\end{tabular}
\vspace{-3mm}
\caption{\small An example of using LiDAR flow and optical flow consistency to filter occluded pixels. (a) Accumulated LiDAR depth, and (b) Accumulated LiDAR depth with flow consistency filtering. \vspace{-6mm}}
\label{Figure:flow_filter}
\end{figure}
\SubSubSection{Occlusion Removal via Segmentation}
Flow-based occluded pixel removal may fail in two cases.
When there is little to no parallax, both optical and scene flow will be small, and their difference becomes not measurable.
This occurs mostly at long range or along the motion direction.
Further, LiDAR flow on moving objects can in some cases be identical to the occluded LiDAR flow behind it.
In both of these cases flow consistency is insufficient to remove occluded pixels from the final depth estimate.
\begin{figure}[t!]
\captionsetup{font=small}
\vspace{-1mm}
\centering
\scalebox{1}{
\begin{tabular}{@{}c@{\hspace{0.3em}}c@{}}
\includegraphics[trim=70 0 60 0,clip,width=1.6in]{./Figures/seg1} &
\includegraphics[trim=70 0 60 0,clip,width=1.6in]{./Figures/seg2} \vspace{-3mm}\\
{\small (a) Car image} & {\small (b) Semantic seg.~\& bound.~box} \vspace{-1mm}\\
\includegraphics[trim=70 0 35 0,clip,width=1.6in]{./Figures/seg5} &
\includegraphics[trim=70 0 35 0,clip,width=1.6in]{./Figures/seg6} \vspace{-4mm}\\
{\small (c) Depth before filtering} & {\small (d) Depth after filtering} \vspace{-3mm}\\
\end{tabular} }
\caption{\small For small flow instances and some movers, flow consistency is insufficient to remove accumulated but occluded LiDAR pixels, see (a,c). To remove these occluded pixels, we first find vehicle pixels as the intersection between semantic segmentation and $2$D bounding box, see (b). From the $3$D bounding box we know the maximum depth of the vehicle, and so can filter out all accumulated depths greater than this that are actually occluded, see (d). \vspace{-8mm}}
\label{Figure:occlusion_box_seg}
\end{figure}
To solve this problem, we use a combination of $3$D bounding boxes and semantic segmentation to remove occluded points appearing on top of objects. First, accurate pixel region of an instance is determined by the intersection of $3$D bounding box projection and semantic segmentation.
The maximum depth of bounding box corners is used to decide whether LiDAR points falling on the object are on it or behind it.
Points within the semantic segmentation and closer than this maximum distance are kept, while points in the segmentation and behind the bounding box are filtered out as occluded LiDAR points. Fig.~\ref{Figure:occlusion_box_seg} shows an example of removing occluded points appearing on vehicle instances. We use a semantic segmentation model~\cite{cheng2020panoptic} pre-trained with CityScape~\cite{cordts2016cityscapes} to segment vehicle pixels.
\begin{figure*}[t!]
\captionsetup{font=small}
\centering
\scalebox{0.7}
{
\begin{tabular}{@{}c@{}c@{}c@{}c@{}c@{}c@{}c@{}}
\includegraphics[width=1.3in]{./Figures/fig10/00_d_raw} &
\includegraphics[width=1.3in]{./Figures/fig10/00_seg} &
\includegraphics[width=1.3in]{./Figures/fig10/00_pda} &
\includegraphics[width=1.3in]{./Figures/fig10/00_exp} &
\includegraphics[width=1.3in]{./Figures/fig10/00_prd} &
\includegraphics[width=1.3in]{./Figures/fig10/00_prd_mono} &
\includegraphics[width=1.3in]{./Figures/fig10/00_prd_traditional} \vspace{-2mm}\\
\includegraphics[width=1.3in]{./Figures/fig10/23_d_raw} &
\includegraphics[width=1.3in]{./Figures/fig10/23_seg} &
\includegraphics[width=1.3in]{./Figures/fig10/23_pda} &
\includegraphics[width=1.3in]{./Figures/fig10/23_exp} &
\includegraphics[width=1.3in]{./Figures/fig10/23_prd} &
\includegraphics[width=1.3in]{./Figures/fig10/23_prd_mono} &
\includegraphics[width=1.3in]{./Figures/fig10/23_prd_traditional} \vspace{-2mm}\\
\includegraphics[width=1.3in]{./Figures/no_colorbar/09_d_raw} &
\includegraphics[ width=1.3in]{./Figures/no_colorbar/09_seg} &
\includegraphics[width=1.3in]{./Figures/no_colorbar/09_pda} &
\includegraphics[width=1.3in]{./Figures/no_colorbar/09_exp} &
\includegraphics[width=1.3in]{./Figures/no_colorbar/09_prd} &
\includegraphics[width=1.3in]{./Figures/no_colorbar/09_prd_mono} &
\includegraphics[width=1.3in]{./Figures/no_colorbar/09_prd_traditional} \vspace{-2mm}\\
\includegraphics[width=1.3in]{./Figures/fig10/21_d_raw} &
\includegraphics[width=1.3in]{./Figures/fig10/21_seg} &
\includegraphics[width=1.3in]{./Figures/fig10/21_pda} &
\includegraphics[width=1.3in]{./Figures/fig10/21_exp} &
\includegraphics[width=1.3in]{./Figures/fig10/21_prd} &
\includegraphics[width=1.3in]{./Figures/fig10/21_prd_mono} &
\includegraphics[width=1.3in]{./Figures/fig10/21_prd_traditional} \vspace{-2mm}\\
\includegraphics[trim = 0 0 0 440, clip, width=1.3in]{./Figures/fig10/09_d_raw} &
&
\includegraphics[trim = 0 0 0 440, clip, width=1.3in]{./Figures/fig10/09_pda} &
\includegraphics[trim = 0 0 0 440, clip, width=1.3in]{./Figures/fig10/09_exp} &
\includegraphics[trim = 0 0 0 440, clip, width=1.3in]{./Figures/fig10/09_prd} &
\includegraphics[trim = 0 0 0 440, clip, width=1.3in]{./Figures/fig10/09_prd_mono} &
\includegraphics[trim = 0 0 0 440, clip, width=1.3in]{./Figures/fig10/09_prd_traditional} \vspace{-5mm}\\
{\small (a) } & {\small (b) } & {\small (c)} & {\small (d)} & {\small (e)} & {\small (f)} & {\small (g)} \vspace{-3mm} \\
\end{tabular}}
\caption{\small (a) Raw radar depths (b) Each color pixel with a maximum RC-PDA $>0.6$ is marked with a color indicating which radar pixel it is associated with. (c) The RC-PDA score with values $>0.6$ for each pixel. (d) The MER channel with RC-PDA $>0.6$. (e) Our final predicted depth. (f) Depth from monocular input to Stage $2$. (g) Depth from monocular and raw radar input to Stage $2$.}
\label{Figure:vis_aff}
\end{figure*}
\SubSection{Algorithm Summary}
We propose a two-stage depth estimation process, as in Fig.~\ref{fig:network_arch}. The Stage $1$ estimates RC-PDA for each radar pixel, which is transformed into our MER representation as detailed in Sec.~\ref{sec:pda_mer} and fed into Stage $2$ which performs conventional depth completion. Both stages are supervised by the accumulated dense LiDAR, with pixels not having a LiDAR depth given zero weight.
Network $1$ uses an encoder-decoder network with skip connections similar to U-Net~\cite{Ronneberger:2015:unet} and~\cite{ma2019self} with details in supplementary material.
\Section{Experimental Results}
\noindent\textbf{Dataset} We train and test on a subset of images from the nuScenes dataset~\cite{nuscenes2019}, including $12,610$, $1,628$ and $1,623$ samples for training, validation and testing, respectively.
The data are collected with moving ego vehicle so flow calculation described in Sec.~\ref{sec:flowconsistency} can be applied. The depth range for training and testing is $0$-$50$ meters. Resolutions of inputs and outputs are $400\times192$.
As described in Sec.~\ref{sec:lidar-supervision}, we build semi-dense depth images by accumulating LiDAR pixels from $21$ subsequent frames and $4$ previous frames (sampled every other frame), and use these for supervision.
\noindent\textbf{Implementation details}
For parameters, we use $T_a=1$\,m and $T_r=0.05$ for Eq.~\ref{eq:aff_label}. In Sec.~\ref{sec:flowconsistency} we use $T_f=3$ to decide flow consistency. MER has $6$ channels with $T_1$ to $T_6$ set as $0.5$, $0.6$, $0.7$, $0.8$, $0.9$ and $0.95$, respectively.
At Stage $1$, we use a U-Net with $5$ levels of resolutions and $180$ output channels, corresponding to $180$ pixels in a rectangle neighborhood with size $w=5$ and $h=36$. As the radar points are typically on the lower part of image, to fully leveraging the neighborhood, the neighborhood center is below the rectangle center with $30$ pixel above, $5$ pixels below and $2$ pixels on left and right to provide more space for radar points to extend upwards. At Stage $2$, we employ two existing depth completion architectures,~\cite{ma2018sparse} and~\cite{li2020multi}, originally designed for LiDAR-camera pairs.
\begin{figure*}[t!]
\captionsetup{font=small}
\centering
\scalebox{0.85}
{
\begin{tabular}{@{}c@{}c@{}c@{}c@{}}
\includegraphics[width=2in]{./Figures/fig11/01_rd_raw}&
\includegraphics[width=2in]{./Figures/fig11/01_expand}&
\includegraphics[width=2in]{./Figures/fig11/01_prd_sp}&
\includegraphics[trim=0cm 0.4cm 0cm 0cm, clip=true, width=2.35in]{./Figures/fig11/01_prd_our} \vspace{-2.3mm}\\
\includegraphics[width=2in]{./Figures/fig11/00_rd_raw}&
\includegraphics[width=2in]{./Figures/fig11/00_expand}&
\includegraphics[width=2in]{./Figures/fig11/00_prd_sp}&
\includegraphics[trim=0cm 0.4cm 0cm 0.1cm, clip=true, width=2.35in]{./Figures/fig11/00_prd_our} \vspace{-2mm}\\
{\small (a) } & {\small (b) } & {\small (c)} & {\small (d)}\\
\end{tabular}}
\vspace{-2.5mm}
\caption{\small Qualitative depth completion comparison showing gains from using MER over raw radar. (a) Raw radar on top of image versus (b) A MER channel with RC-PDA $>0.8$ on top of image. Depth completion (c) without and (d) with using MER.\vspace{-2mm}}
\label{Figure:vis_cmp}
\end{figure*}
\SubSection{Visualization of Predicted RC-PDA}
The predicted RC-PDA and estimated depths from Stage $2$~\cite{ma2018sparse} are visualized in Fig.~\ref{Figure:vis_aff}. Column (a) shows the raw radar pixels plotted on images and often include occluded radar pixels. Column (b) shows how image pixels are associated with different radar pixels according to their maximum RC-PDA. Radar pixels and their associated neighboring pixels are marked with the same color. Notice in column (c) that RC-PDA is high within objects and decreases after crossing boundaries. Occluded radar depth are mostly discarded as their predicted RC-PDA is low. In column (e), the dense depths predicted from MER are improved over predictions from (g) raw radar and/or (f) monocular. For example, in Row $2$ of Fig.~\ref{Figure:vis_aff}, our predicted pole depth in (e) has better boundaries than monocular-only in (f), and monocular plus raw radar in (g).
How we achieve this can be intuitively understood by comparing raw radar in (a) with our MER in (d), the output of Stage $1$. While raw radar has many incorrect depths, MER selects correct radar depths and extends these depths along the pole and background, enabling improved final depth inference.
\SubSection{Accuracy of MER}
To be useful in improving radar-camera depth completion, the enhanced radar depth in the vicinity of radar points should be better than alternatives. We compare the depth error of the MER in regions where RC-PDA is $>0.9$, with a few baseline methods and results are shown in Tab.~\ref{tab:expand_error}.
The enhanced radar depth from Stage $1$ improves over not only raw radar depth but also depth estimates from Stage $2$ using monocular as well as monocular plus radar. In comparison, Stage $2$ keeps the accuracy of the enhanced radar depth when using it for depth completion. The depth error for raw radar depth is very large since many of them are occluded and far behind foreground. About $35\%$ of radar points in the test frames have a maximum RC-PDA smaller than $T_1$ in their neighborhood and are discarded as occluded points.
Further, Fig.~\ref{Figure:aff_curve} shows the depth error and per-image average area of expanded depth from $6$ MER channels, respectively.
It shows, as confidence increases, higher RC-PDA corresponds to higher accuracy and smaller expanded areas.
\begin{table}[t]
\captionsetup{font=small}
\begin{center}
\scalebox{0.7}{
\begin{tabular}{|c|c|cccc|}
\hline
Network $2$ & Input & MAE & Abs Rel & RMSE & RMSE log\\
\hline
& Image & $2.385$ & $0.110$ & $3.505$ & $0.150$ \\
Ma~\emph{et al.}~\cite{ma2018sparse}
& Image, radar & $1.609$ & $0.078$ & $2.865$ & $0.126$ \\
& Image, radar, MER & $\mathbf{1.229}$ & $\mathbf{0.058}$ & $\mathbf{2.651}$ & $\mathbf{0.114}$ \\
\hline
\hline
\multirow{2}{*}{Li~\emph{et al.}~\cite{li2020multi}}
& Image, radar & $1.759$ & $0.084$ & $3.039$ & $0.133$ \\
& Image, radar, MER & $\mathbf{1.274}$ & $\mathbf{0.061}$ & $\mathbf{2.670}$ & $\mathbf{0.116}$ \\
\hline
\hline
\multirow{2}{*}{None}
& MER & $1.251$ & $0.059$ & $2.701$ & $0.117$ \\
& Radar & $7.369$ & $0.475$ & $10.900$ & $0.448$ \\
\hline
\end{tabular} }
\end{center}
\vspace{-5mm}
\caption{\small Depth error (m) in image regions around non-occluded radar returns, defined as regions with RC-PDA $>0.9$.\vspace{-3mm}}
\label{tab:expand_error}
\vspace{-0.5mm}
\end{table}
\begin{figure}[t!]
\captionsetup{font=small}
\centering
\begin{tabular}{c}
\includegraphics[width=2.8in]{./Figures/mae_pda_curve3} \\
\end{tabular}
\vspace{-3mm}
\caption{\small Image area and depth error of enhanced radar in MER for regions with minimal RC-PDA at $6$ confidence levels. \vspace{-3mm}}
\label{Figure:aff_curve}
\end{figure}
\SubSection{Comparison of Depth Completion}
To evaluate effectiveness of the enhanced radar depth in depth completion, we compare the depth error with and without using MER as input for Network $2$, and show performance in Tabs.~\ref{tab:d_error} and~\ref{tab:d_error_low}. The results show that including radar improves depth completion over monocular, while using our proposed MER further improves the accuracy of depth completion for the same network.
Qualitative comparisons between depth completion~\cite{ma2018sparse} with and without using MER are shown in Fig.~\ref{Figure:vis_cmp}.
This shows improvement from MER in estimating object depth boundaries including close objects (such as the traffic sign on the bottom image) and far objects.
\begin{table}[t]
\captionsetup{font=small}
\begin{center}
\scalebox{0.7}{
\begin{tabular}{|c|c|cccc|}
\hline
Network $2$ & Input & MAE & Abs Rel & RMSE & RMSE log\\
\hline
& Image & $1.808$ & $0.102$ & $3.552$ & $0.160$ \\
Ma~\emph{et al.}~\cite{ma2018sparse}
& Image, radar & $1.569$ & $0.090$ & $3.327$ & $0.152$ \\
& Image, radar, MER & $\mathbf{1.472}$ & $\mathbf{0.085}$ & $\mathbf{3.179}$ & $\mathbf{0.144}$ \\
\hline
\hline
\multirow{2}{*}{Li~\emph{et al.}~\cite{li2020multi}}
& Image, radar & $1.821$ & $0.107$ & $3.650$ & $0.170$ \\
& Image, radar, MER & $\mathbf{1.655}$ & $\mathbf{0.094}$ & $\mathbf{3.463}$ & $\mathbf{0.159}$ \\
\hline
\end{tabular} }
\end{center}
\vspace{-5mm}
\caption{\small Full-image depth estimation/completion errors (m).\vspace{-2mm}}
\label{tab:d_error}
\end{table}
\begin{table}[t]
\captionsetup{font=small}
\begin{center}
\scalebox{0.7}{
\begin{tabular}{|c|c|cccc|}
\hline
Network $2$ & Input & MAE & Abs Rel & RMSE & RMSE log\\
\hline
& Image & $2.673$ & $0.153$ & $4.259$ & $0.202$ \\
Ma~\emph{et al.}~\cite{ma2018sparse}
& Image, radar & $2.263$ & $0.134$ & $4.028$ & $0.194$ \\
& Image, radar, MER & $\mathbf{2.078}$ & $\mathbf{0.124}$ & $\mathbf{3.864}$ & $\mathbf{0.183}$ \\
\hline
\hline
\multirow{2}{*}{Li~\emph{et al.}~\cite{li2020multi}}
& Image, radar & $2.515$ & $0.154$ & $4.266$ & $0.211$ \\
& Image, radar, MER & $\mathbf{2.189}$ & $\mathbf{0.132}$ & $\mathbf{3.943}$ & $\mathbf{0.193}$\\
\hline
\end{tabular} }
\end{center}
\vspace{-5mm}
\caption{\small Depth estimation/completion errors (m) in the low-height region ($0.3$-$2$ meters above ground).\vspace{-3mm}}
\label{tab:d_error_low}
\end{table}
\Section{Conclusion}
Radar-based depth completion introduces additional challenges and complexities beyond LiDAR-based depth completion. A significant difficulty is the large ambiguity in associating radar pixels with image pixels. We address this with RC-PDA, a learned measure that associates radar hits with nearby image pixels at the same depth. From RC-PDA we create an enhanced and densified radar image called MER. Our experiments show that depth completion using MER achieves improved accuracy over depth completion with raw radar. As part of this work we also create a semi-dense accumulated LiDAR depth dataset for training depth completion on nuScenes.
\vspace{2mm}
\noindent\textbf{Acknowledgement}
This work was supported by the Ford-MSU Alliance.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}\label{sec1}
On March 3 2021, LHC announced that 59 new hadrons were reported over the past decade \cite{LHCb59:2021}. The present situation of hadronic states is far beyond what Gell-Mann and Zweig thought possible \cite{GellMann:1964nj,Zweig:1981pd,Zweig:1964jf}. Among these observed states, these $P_c$ states existing in the $\Lambda_{b}{\to}{J/\psi}K^{-}p$ decay are good candidate of exotic molecular pentaquark \cite{Aaij:2015tga,Aaij:2016phn,Aaij:2019vzc,Chen:2015loa,He:2015cea,Chen:2015moa,Chen:2019asm,Wang:2019ato,He:2019rva,Liu:2019zvb,Burns:2019iih}. The name {\it pentaquark} was firstly proposed in Refs. \cite{Gignoux:1987cn,Lipkin:1987sk}, and there were many theoretical explorations of pentaquark \cite{Leandri:1989su,Genovese:1997tm,Gao:1999ar,Gerasyuta:2002kq,Lipkin:1998pb}.
In 2003, LEPS announced the observation of $\Theta^{+}(1540)$ with strangeness $S=+1$ \cite{Nakano:2003qx}, which has the $uudd\bar{s}$ component. Of course, $\Theta^{+}(1540)$ stimulated extensive studies of pentaquark \cite{Rosner:2003ia,Chen:2016qju,Karliner:2017qhf,Liu:2019zoy}. However, some high precision experiments such as BES \cite{Bai:2004gk}, BaBar \cite{Aubert:2005qi}, Belle \cite{Wang:2005fc} and CDF \cite{Litvintsev:2004yw} did not confirmed the existence of $\Theta^{+}(1540)$. Facing this situation, one realized again that our understanding of non-perturbative behavior of quantum chromodynamics (QCD) is still absent, which is a lesson of $\Theta^{+}(1540)$.
Obviously, it is not the end of the exploration of exotic hadronic matter. With the accumulation of experimental data, more and more charmoniumlike $XYZ$ states have been discovered since 2013, which again inspired theorist's extensive interest in investigating exotic hadronic states \cite{Guo:2014iya,Esposito:2014rxa,Ali:2017jda,Brambilla:2019esw,Hosaka:2016pey,Richard:2016eis,Lebed:2016hpi,Esposito:2016noz}. Especially, in 2015, the LHCb Collaboration measured the $\Lambda_{b}^{0} \rightarrow J/\psi K^{-}p$ decay and observed two hidden-charm pentaquark-like resonances $P_{c}(4380)$ and $P_{c}(4450)$ in the $J/\psi p$ invariant mass spectrum, which indicates that they have a minimal quark content of $uudc\bar{c}$ \cite{Aaij:2015tga,Aaij:2016phn}.
In 2019, LHCb found three narrow $P_c$ structures in the ${J/\psi}p$ invariant mass spectrum of $\Lambda_{b}{\to}{J/\psi}K^{-}p$ \cite{Aaij:2019vzc}, where this new measurement shows that
$P_{c}(4450)$ is actually composed of two substructures, $P_{c}(4440)$ and $P_{c}(4457)$ with $5.4\sigma$ significance. The characteristic mass spectrum of $P_c$
provides the strong evidence of existence of hidden-charm molecular pentaquarks \cite{Chen:2019asm,Wang:2019ato,He:2019rva,Liu:2019zvb,Burns:2019iih}.
At present, searching for exotic hadronic states is still full of challenge and opportunity. As theorists, we should provide valuable prediction, which requires us to
continue to find some crucial hint for exotic hadronic states.
We may borrow some idea of proposing stable tetraquark state with $QQ\bar{q}\bar{q}$ configuration.
Stimulated by the observation of double charm baryon $\Xi_{cc}^{++}(3620)$ \cite{Aaij:2017ueg}, Refs. \cite{Luo:2017eub,Karliner:2017qjm,Eichten:2017ffp} studied the possible stable tetraquark state with the $QQ\bar{q}\bar{q}$ configuration. Here, the $QQ\bar{q}\bar{q}$ configuration can be obtained by replacing light quark $q$ of double heavy baryon $QQq$ with $\bar q\bar{q}$ pair since the color structure of $q$ and $\bar q\bar{q}$ can be the same. Along this line, we may continue to replace $\bar{q}$ of $QQ\bar{q}\bar{q}$ with a $QQ$ pair and get the $QQQQ\bar{q}$ configuration, which is a typical pentaquark configuration (see Fig. \ref{aa}).
\begin{figure}[htpb]
\includegraphics[width=240pt]{aa.pdf}
\caption{Evolution of the $QQQQ\bar q$ pentaquark from double heavy tetraquark and baryon.}\label{aa}
\end{figure}
As a new type of pentaquark, the $QQQQ\bar{q}$ pentaquark
draw our attention to explore its mass spectrum and decay behavior. This information is valuable to future experimental search for the $QQQQ\bar{q}$ pentaquark. In this work,
we systematically present the mass spectrum of the S-wave $QQQQ\bar{q}$ system within the framework of Chromomagnetic Interaction (CMI) model \cite{DeRujula:1975qlm}, which has been widely adopted to study the mass spectra of multiquark systems
\cite{Luo:2017eub,Wu:2016gas,Wu:2018xdi,Chen:2016ont,Wu:2016vtq,Liu:2016ogz,Wu:2017weo,Zhou:2018pcv,
Li:2018vhp,An:2019idk,Cheng:2020irt,Cheng:2019obk,Hogaasen:2013nca,Weng:2018mmf,
Weng:2019ynva,Weng:2020jao,Zhao:2014qva,Cheng:2020nho,An:2020jix}.
Moreover, in this work we estimate the ratios of the possible decays of the $QQQQ\bar{q}$ pentaquarks, which are crucial for hunting this kind of pentaquark.
This paper is organized as follows. After introduction, the adopted CMI model will be introduced in In Sec. \ref{sec2}.
In Sec. \ref{sec4}, we present the mass spectra of S-wave $QQQQ\bar{q}$ pentaquarks and estimate the ratios of possible strong decay widths. Finally,
a short summary is followed in Sec. \ref{sec5}.
\section{The chromomagnetic Interaction Model}\label{sec2}
The effective Hamiltonian involved in the estimated mass spectrum of these discussed pentaquarks
\begin{eqnarray}\label{Eq1}
H&=&\sum_im_i+H_{\textrm{CMI}} \nonumber \\
&=&\sum_im_i-\sum_{i<j}C_{ij} \vec\lambda_i\cdot \vec\lambda_j \vec\sigma_i\cdot\vec\sigma_j,
\end{eqnarray}
where $m_i$ denotes the effective mass of the $i$-th constituent quark when considering these effects from kinetic energy, color confinement, and so on.
$H_{\rm CMI}$, as the chromomagnetic interaction Hamiltonian, is composed of the Pauli matrices $\sigma_{i}$ and the Gell-Mann matrices $\lambda_i$. Here,
$\lambda_i$ should be replaced by $-\lambda_i^{*}$ for antiquark, and $C_{ij}$ denotes the effective coupling constant between the $i$-th quark and $j$-th quark, which depends on the quark masses and the ground particle's spatial wave function. As input, $C_{ij}$ is fixed by the involved hadron masses.
Usually, there exists overestimate to the predicted hadron masses by Eq. (\ref{Eq1})
\cite{Wu:2016vtq, Wu:2016gas, Chen:2016ont, Wu:2017weo, Luo:2017eub, Zhou:2018pcv, Li:2018vhp, Wu:2018xdi,An:2019idk,Liu:2016ogz,Cheng:2020irt,Cheng:2019obk}. Additionally, the effective mass $m_{i}$ in Eq. \ref{Eq1} does not incorporate attraction sufficiently.
The mass of the pentaquark state reads as
\begin{eqnarray}
\label{Eq11}
M=M_{ref}-\langle H_{\rm CMI}\rangle_{ref}+\langle H_{\rm CMI}\rangle,
\end{eqnarray}
where $M_{ref}$, $\langle H_{\rm CMI}\rangle_{ref}$, and $\langle H_{\rm CMI}\rangle$ are the physical mass of the reference system, the corresponding CMI eigenvalue, and the CMI eigenvalue for the discussed multi-quark state, respectively.
For $M_{ref}$, we can use the threshold of a reference baryon-meson system whose quark content is the same as the considered pentaquark states.
For example, for the $nncc\bar{c}$ system, we can use $(M_{\Xi_{cc}}+M_{D})-(\langle H_{\rm CMI}\rangle_{\Xi_{cc}}+\langle H_{\rm CMI}\rangle_{D})+\langle H_{\rm CMI}\rangle_{nncc\bar{c}}$
or $(M_{\Sigma_{c}}+M_{J/\psi})-(\langle H_{\rm CMI}\rangle_{\Sigma_{c}}+\langle H_{\rm CMI}\rangle_{J/\psi})+\langle H_{\rm CMI}\rangle_{nncc\bar{c}}$ to obtain the mass spectra.
All possible used $M_{ref}$ values are shown in Table \ref{comp}
\footnote{Since some of the heavy flavor baryons are not-yet observed, we adopt the theoretical results in Ref. \cite{Weng:2018mmf} as input.
}
With this treatment, the dynamical effects that are not incorporated in the original approach can be compensated \cite{Zhou:2018pcv}. This approach of determining the masses of pentaquark states is named as the reference mass scheme here.
Alternatively, a color-electric term can be introduced here \cite{Hogaasen:2013nca,Liu:2019zoy,Weng:2018mmf,Weng:2019ynva,Weng:2020jao,An:2020jix}
\begin{eqnarray}\label{Eq2}
H_{\rm CEI}=-\sum_{i<j}A_{ij} \vec\lambda_i\cdot \vec\lambda_j.
\end{eqnarray}
With deduction
\begin{eqnarray}
&&\sum_{i<j}(m^{0}_{i}+m^{0}_{j})\vec\lambda_i\cdot \vec\lambda_j\nonumber\\
&&=\frac{1}{2}\sum_{i,j}(m^{0}_{i}+m^{0}_{j})\vec\lambda_i\cdot \vec\lambda_j-\sum_{i}m^{0}_{i}(\vec\lambda_i)^{2}\nonumber\\
&&=(\sum_{i}m^{0}_{i}\vec\lambda_i)\cdot\left(\sum_{i}\vec\lambda_i\right)-\frac{16}{3}\sum_{i}m^{0}_{i},
\end{eqnarray}
where the color operator $\sum_{i}\vec\lambda_i$ nullifies any colorless physical
state, we have the Hamiltonian of the modified CMI model
\begin{eqnarray}\label{Eq3}
H&=&\sum_im^{0}_i+H_{\rm CEI}+H_{\rm CMI}\nonumber\\
&=&\sum_im^{0}_i-\sum_{i<j}A_{ij} \vec\lambda_i\cdot \vec\lambda_j-\sum_{i<j}v_{ij} \vec\lambda_i\cdot \vec\lambda_j \vec\sigma_i\cdot\vec\sigma_j, \nonumber\\
&=&-\frac{3}{4}(-\frac{4}{3}\sum_im^{0}_i)-\sum_{i<j}A_{ij} \vec\lambda_i\cdot \vec\lambda_j\nonumber\\&&-\sum_{i<j}v_{ij} \vec\lambda_i\cdot \vec\lambda_j \vec\sigma_i\cdot\vec\sigma_j, \nonumber\\
&=&-\frac{3}{4}\sum_{i<j}[\frac{1}{4}(m^{0}_{i}+m^{0}_{j})+\frac{4}{3}A_{ij}] \vec\lambda_i\cdot \vec\lambda_j\nonumber\\&&-\sum_{i<j}v_{ij} \vec\lambda_i\cdot \vec\lambda_j \vec\sigma_i\cdot\vec\sigma_j, \nonumber\\
&=&-\frac{3}{4}\sum_{i<j}m_{ij}V^{\rm C}_{ij}-\sum_{i<j}v_{ij}V^{\rm CMI}_{ij}.
\end{eqnarray}
In the above expression, $V^{\rm C}_{ij}=\vec\lambda_i\cdot \vec\lambda_j$ and $ V^{\rm CMI}_{ij}=\vec\lambda_i\cdot \vec\lambda_j \vec\sigma_i\cdot\vec\sigma_j$
denote the color and color-magnetic interactions between quarks, respectively.
A new mass parameter of quark pair should be defined, i.e.,
\begin{eqnarray}
m_{ij}=\frac{1}{4}(m^{0}_{i}+m^{0}_{j})+\frac{4}{3}A_{ij}.
\end{eqnarray}
By these conventional hadron masses (see Table \ref{comp}), we may fix the parameters $m_{ij}$ and $v_{ij}$ (see Table \ref{parameter2}). Interested readers can refer to Refs. \cite{Weng:2019ynva,Weng:2018mmf,Liu:2019zoy,Weng:2020jao} for more details. With this modified CMI model scheme, we can give mass spectrum of the $QQQQ\bar{q}$ pentaquarks.
\renewcommand\arraystretch{1.5}
\begin{table}[htbp]
\caption{The adopted masses of conventional hadrons for determining parameters (in units of MeV) \cite{Tanabashi:2018oca,Weng:2018mmf}.
}\label{comp}
\begin{tabular}{crc|crc}
\bottomrule[1.5pt]
\bottomrule[0.5pt]
hadrons& $I(J^P)$ &Mass & hadrons& $I(J^P)$ &Mass \\
\bottomrule[0.7pt]
$D$&$1/2(0^{-})$&1869.6&$D^{*}$&$1/2(1^{-})$&2010.3\\
$D_{s}$&$0(0^{-})$&1968.3&$D^{*}_{s}$&$0(1^{-})$&2112.2\\
$B$&$1/2(0^{-})$&5279.3&$B^{*}$&$1/2(1^{-})$&5324.7\\
$B_{s}$&$0(0^{-})$&5366.8&$B^{*}_{s}$&$0(1^{-})$&5415.4\\
$\Xi_{cc}$&$1/2(1/2^{+})$&3621.4&$\Xi^{*}_{cc}$&$1/2(3/2^{+})$&(3696.1)\\
$\Omega_{cc}$&$0(1/2^{+})$&(3731.8)&$\Omega^{*}_{cc}$&$0(3/2^{+})$&(3802.4)\\
$\Xi_{cb}$&$1/2(1/2^{+})$&(6922.3)&$\Omega_{cb}$&$0(1/2^{+})$&(7010.7)\\
$\Xi'_{cb}$&$1/2(1/2^{+})$&(6947.9)&$\Omega'_{cb}$&$0(1/2^{+})$&(7047.0)\\
$\Xi^{*}_{cb}$&$1/2(3/2^{+})$&(6973.2)&$\Omega^{*}_{cb}$&$0(3/2^{+})$&(7065.7)\\
$\Xi_{bb}$&$1/2(1/2^{+})$&(10168.9)&$\Xi^{*}_{bb}$&$1/2(3/2^{+})$&(10188.8)\\
$\Omega_{bb}$&$0(1/2^{+})$&(10259.0)&$\Omega^{*}_{bb}$&$0(3/2^{+})$&(10267.5)\\
$\Omega_{ccc}$&$0(3/2^{+})$&(4785.6)&$\Omega_{bbb}$&$0(3/2^{+})$&(14309.7)\\
$\Omega_{ccb}$&$0(1/2^{+})$&(7990.3)&$\Omega_{ccb}^{*}$&$0(3/2^{+})$&(8021.8)\\
$\Omega_{bbc}$&$0(1/2^{+})$&(11165.0)&$\Omega_{bbc}^{*}$&$0(3/2^{+})$&(11196.4)\\
\bottomrule[0.5pt]
\midrule[1.5pt]
\end{tabular}
\end{table}
\begin{table}[t]
\centering \caption{Coupling parameters for the schemes in units of MeV. Here, $q=n,\,s$ ($n=u$, $d$) and $Q=c,\,b$.
}\label{parameter2}
\renewcommand\arraystretch{1.5}
\begin{tabular}{ccccccc}
\bottomrule[1.5pt]
\bottomrule[0.5pt]
\multicolumn{7}{c}{The reference mass scheme}\\
$C_{cc}$&$C_{bb}$&$C_{cb}$&$C_{c\bar{n}}$&$C_{c\bar{s}}$&$C_{b\bar{n}}$&$C_{b\bar{s}}$\\
3.3&1.8&2.0&6.6&6.7&2.1&2.3\\
\bottomrule[1.0pt]
\multicolumn{7}{c}{The modified CMI model scheme}\\
$m_{cc}$&$m_{bb}$&$m_{cb}$&$m_{c\bar{n}}$&$m_{c\bar{s}}$&$m_{b\bar{n}}$&$m_{b\bar{s}}$\\
792.9&2382.4&1604.0&493.3&519.0&1328.3&1350.8\\
\bottomrule[0.5pt]
$v_{cc}$&$v_{bb}$&$v_{cb}$&$v_{c\bar{n}}$&$v_{c\bar{s}}$&$v_{b\bar{n}}$&$v_{b\bar{s}}$\\
3.5&1.9&2.0&6.6&6.7&2.1&2.3\\
\bottomrule[0.5pt]
\bottomrule[1.5pt]
\end{tabular}
\end{table}
When calculating the mass of the pentaquark states, we need the information of the total wave function, which is composed of space, flavor, color, and spin wave functions, i.e.,
\begin{eqnarray}
\psi_{\textrm{tot}}=\psi_{space}\otimes\psi_{flavor}\otimes\psi_{color}\otimes\psi_{spin}.
\end{eqnarray}
Since we only focus on the low-lying $S$-wave pentaquark states, the symmetrical constraint from spatial pentaquark wave function is trivial. For the discussed $QQQQ\bar{q}$ pentaquarks, their $\psi_{flavor}\otimes\psi_{color}\otimes\psi_{spin}$ wave functions of pentaquark states should be fully antisymmetric when exchanging identical quarks.
All the possible flavor combinations for the $QQQQ\bar{q}$ pentaquark system are shown in Table \ref{flavor1}, by which we further determine $\psi_{flavor}\otimes\psi_{color}\otimes\psi_{spin}$ wave functions, which satisfy $\{1234\}$, $\{123\}$, and $\{12\}\{34\}$ symmetry. Here, we use the notation $\{1234\}$ to label that the 1st, 2nd, 3rd, and 4th quarks have antisymmetry property, which will be applied in the following discussions. For identifying the pentaquark configuration with certain exchanging symmetry, we use the approach of Young diagram and Young tableau, which represents the irreducible bases of the permutation group.
\begin{table}[t]
\centering \caption{All possible flavor combinations for the $QQQQ\bar{q}$ pentaquark system with $q=n,\,s$ ($n=u$, $d$) and $Q=c,\,b$.
}\label{flavor1}
\renewcommand\arraystretch{1.4}
\begin{tabular}{p{1.3cm}|p{1.3cm}p{1.3cm}p{1.3cm}p{1.3cm}p{1.3cm}}
\bottomrule[1.5pt]
\bottomrule[0.5pt]
System&\multicolumn{5}{c}{Flavor combinations}\\
\bottomrule[0.5pt]
\multirow{2}*{$QQQQ\bar{q}$}&$cccc\bar{n}$&$cccb\bar{n}$&$ccbb\bar{n}$&$bbbc\bar{n}$&$bbbb\bar{n}$\\
&$cccc\bar{s}$&$cccb\bar{s}$&$ccbb\bar{s}$&$bbbc\bar{s}$&$bbbb\bar{s}$\\
\bottomrule[0.5pt]
\bottomrule[1.5pt]
\end{tabular}
\end{table}
The color wave functions for $QQQQ\bar{q}$ pentaquark system are expressed a direct product
\begin{eqnarray}
\left[3_c\otimes3_c\otimes3_c\otimes3_c\right]\otimes\bar{3}_c.\label{colorproduct}
\end{eqnarray}
Due to the requirement of color confinement,
the color wave function must be a singlet.
Therefore, the four heavy quarks should be in the color triplet states, and the corresponding partition [211] reads as
\begin{align}
&\begin{tabular}{|c|c|}
\hline
1 & 2 \\
\cline{1-2}
\multicolumn{1}{|c|}{3} \\
\cline{1-1}
\multicolumn{1}{|c|}{4} \\
\cline{1-1}
\end{tabular}
=\left\{\left(12\right)_6\left(34\right)_{\bar{3}}\right\}_3,
\quad
\begin{tabular}{|c|c|}
\hline
1 & 3 \\
\cline{1-2}
\multicolumn{1}{|c|}{2} \\
\cline{1-1}
\multicolumn{1}{|c|}{4} \\
\cline{1-1}
\end{tabular}
=\left\{\left(12\right)_{\bar{3}} 34\right\}_3,
\quad
\nonumber \\
&\begin{tabular}{|c|c|}
\hline
1 & 4 \\
\cline{1-2}
\multicolumn{1}{|c|}{2} \\
\cline{1-1}
\multicolumn{1}{|c|}{3} \\
\cline{1-1}
\end{tabular}
=\left\{\left(123\right)_14\right\}_3.
\label{eq-color1}
\end{align}
Here, the subscript labels the irreducible representation of $\rm SU(3)$.
Then, by combining the antitriplet from light anti-quark with the deduced three color triplets in Eq. (\ref{eq-color1}), we obtain
three color singlets for all the studied pentquark systems
\begin{eqnarray}
\label{eq-color2}
\vert C_1 \rangle&=&\vert \left [\left(12\right)_{6}\left(34\right)_{\bar{3}}\right ]_{3} (\bar{5})_{\bar{3}}\rangle =
\begin{tabular}{|c|c|}
\hline
1 & 2 \\
\cline{1-2}
\multicolumn{1}{|c|}{3} \\
\cline{1-1}
\multicolumn{1}{|c|}{4} \\
\cline{1-1}
\end{tabular}_3
\otimes(\bar{5})_{\bar{3}}\nonumber \\
&=&\frac{1}{4\sqrt{3}}\Big[(2bbgr-2bbrg+gbrb-gbbr+bgrb-bgbr \nonumber\\
&&-rbgb+rbbg-brgb+brbg)\bar{b}+(2rrbg-2rrgb \nonumber\\
&&+rgrb-rgbr+grrb-grbr+rbgr-rbrg+brgr\nonumber\\
&&-brrg)\bar{r}+(2ggrb-2ggbr-rggb+rgbg-grgb \nonumber\\
&&+grbg+gbgr-gbrg+bggr-bgrg)\bar{g}\Big],
\end{eqnarray}
\begin{eqnarray}
\label{eq-color3}
\vert C_2 \rangle&=&\vert \left [\left(12\right)_{\bar{3}}34\right ]_{3} (\bar{5})_{\bar{3}}\rangle =
\begin{tabular}{|c|c|}
\hline
1 & 3 \\
\cline{1-2}
\multicolumn{1}{|c|}{2} \\
\cline{1-1}
\multicolumn{1}{|c|}{4} \\
\cline{1-1}
\end{tabular}_3
\otimes(\bar{5})_{\bar{3}} \nonumber\\
&=&\frac{1}{12}\Big[(3bgbr-3gbbr-3brbg+3rbbg-rbgb+gbrb \nonumber\\
&&+2grbb+brgb-bgrb-2rgbb)\bar{b}+(3grrb-3rgrb \nonumber\\
&&-3brrg+3rbrg-rbgr-2gbrr+2bgrr-grbr \nonumber \\
&&+rgbr+brgr)\bar{r}+(3grgb-3rggb+3bggr-3gbgr\nonumber \\
&&-grbg+rgbg+2rbgg-2brgg+gbrg-bgrg)\bar{g}\Big], \nonumber\\
\end{eqnarray}
and
\begin{eqnarray}
\label{eq-color4}
\vert C_3 \rangle&=&\vert \left [\left(123\right)_{1}4\right ]_{3} (\bar{5})_{\bar{3}}\rangle =
\begin{tabular}{|c|c|}
\hline
1 & 4 \\
\cline{1-2}
\multicolumn{1}{|c|}{2} \\
\cline{1-1}
\multicolumn{1}{|c|}{3} \\
\cline{1-1}
\end{tabular}_3
\otimes(\bar{5})_{\bar{3}}\nonumber\\
&=&\frac{1}{3\sqrt{2}}\Big[(grbb-rgbb+rbgb-brgb+bgrb-gbrb)\bar{b}\nonumber\\
&&+(grbr-rgbr+rbgr-brgr+bgrr-gbrr)\bar{r}\nonumber \\
&&+(grbg-rgbg+rbgg-brgg+bgrg-gbrg)\bar{g}\Big]. \nonumber\\
\end{eqnarray}
Next, we discuss spin wave function in spin space.
For the $QQQQ\bar q$ pentaquark system with total spin $J=5/2$, the spin state can be represented by partition [5], i.e.,
\begin{align}
\label{eq-spin1}
\begin{tabular}{|c|c|c|c|c|}
\hline
1&2&3&4&5 \\
\cline{1-5}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_1$}\end{tabular}}.\\ \nonumber
\end{align}
The spin wave functions for the $QQQQ\bar q$ pentaquark with $J=3/2$ are represented in partition [41] as
\begin{align}
\label{eq-spin2}
&\begin{tabular}{|c|c|c|c|}
\hline
1&2&3&4 \\
\cline{1-4}
\multicolumn{1}{|c|}{5} \\
\cline{1-1}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_1$}\end{tabular}},
\begin{tabular}{|c|c|c|c|}
\hline
1&2&3&5 \\
\cline{1-4}
\multicolumn{1}{|c|}{4} \\
\cline{1-1}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_2$}\end{tabular}},
\begin{tabular}{|c|c|c|c|}
\hline
1&2&4&5 \\
\cline{1-4}
\multicolumn{1}{|c|}{3} \\
\cline{1-1}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_3$}\end{tabular}},
\begin{tabular}{|c|c|c|c|}
\hline
1&3&4&5 \\
\cline{1-4}
\multicolumn{1}{|c|}{2} \\
\cline{1-1}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_4$}\end{tabular}}. \nonumber \\
\end{align}
With the similar method, one obtains the spin wave functions for the $QQQQ\bar q$ pentaquark states with $J=1/2$, which can be represented in partition [32] as
\begin{align}
\label{eq-spin3}
&\begin{tabular}{|c|c|c|}
\hline
1&2&3 \\
\cline{1-3}
4&5 \\
\cline{1-2}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_1$}\end{tabular}},
\begin{tabular}{|c|c|c|}
\hline
1&2&4 \\
\cline{1-3}
3&5 \\
\cline{1-2}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_2$}\end{tabular}},
\begin{tabular}{|c|c|c|}
\hline
1&3&4 \\
\cline{1-3}
2&5 \\
\cline{1-2}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_3$}\end{tabular}},
\begin{tabular}{|c|c|c|}
\hline
1&2&5 \\
\cline{1-3}
3&4 \\
\cline{1-2}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_4$}\end{tabular}},
\begin{tabular}{|c|c|c|}
\hline
1&3&5 \\
\cline{1-3}
2&4 \\
\cline{1-2}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_5$}\end{tabular}}.\\ \nonumber
\end{align}
Since the particle 5 corresponds to a light antiquark, we can isolate this antiquark and discuss the symmetry property of the first four heavy quarks 1, 2, 3, and 4 in $\psi_{color}\otimes\psi_{spin}$ space.
When the antiquark 5 is separated from the spin wave functions,
the spin states represented in Young tableaux without the antiquark 5 can be directly obtained from Eqs. (\ref{eq-spin1}-\ref{eq-spin3}) as
\begin{align}
\label{eq-spin4}
J=\frac52: \quad
&\begin{tabular}{|c|c|c|c|}
\hline
1&2&3&4 \\
\cline{1-4}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_1$}\end{tabular}} ,
\nonumber
\\
J=\frac32: \quad
&\begin{tabular}{|c|c|c|c|}
\hline
1&2&3&4 \\
\cline{1-4}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_1$}\end{tabular}} ,
\begin{tabular}{|c|c|c|}
\hline
1&2&3 \\
\cline{1-3}
\multicolumn{1}{|c|}{4} \\
\cline{1-1}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_2$}\end{tabular}},
\begin{tabular}{|c|c|c|c|}
\hline
1&2&4 \\
\cline{1-3}
\multicolumn{1}{|c|}{3} \\
\cline{1-1}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_3$}\end{tabular}},
\begin{tabular}{|c|c|c|c|}
\hline
1&3&4 \\
\cline{1-3}
\multicolumn{1}{|c|}{2} \\
\cline{1-1}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_4$}\end{tabular}},
\nonumber
\\
J=\frac12: \quad
&\begin{tabular}{|c|c|c|}
\hline
1&2&3 \\
\cline{1-3}
4 \\
\cline{1-1}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_1$}\end{tabular}},
\begin{tabular}{|c|c|c|}
\hline
1&2&4 \\
\cline{1-3}
3 \\
\cline{1-1}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_2$}\end{tabular}},
\begin{tabular}{|c|c|c|}
\hline
1&3&4 \\
\cline{1-3}
2 \\
\cline{1-1}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_3$}\end{tabular}},
\begin{tabular}{|c|c|}
\hline
1&2 \\
\cline{1-2}
3&4 \\
\cline{1-2}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_4$}\end{tabular}},
\begin{tabular}{|c|c|}
\hline
1&3 \\
\cline{1-2}
2&4 \\
\cline{1-2}
\end{tabular}_{\begin{tabular}{|c|} \multicolumn{1}{c}{$S_5$}\end{tabular}}.\nonumber \\
\end{align}
In Eq. (\ref{eq-spin4}), the spin states can be identified with the Young-Yamanouchi basis vectors for partitions [4], [31], and [22].
For constructing the $\psi_{flavor}\otimes\psi_{color}\otimes\psi_{spin}$ wave functions of $QQQQ\bar{q}$ pentaquark system, we should combine the partition [211] of the color singlets in Eq. (\ref{eq-color1}) with partitions [4], [31], [22] of the spin states in Eq. (\ref{eq-spin4}) by the inner product of the permutation group. Thus, the $\psi_{color}\otimes\psi_{spin}$ wave functions with a certain symmetry can be constructed.
We get the corresponding Young diagram representations of $\psi_{color}\otimes\psi_{spin}$ bases \cite{Itzykson:1965hk,Stancu:1999qr,Park:2017jbn}
\begin{flalign}\label{colorspin}
\includegraphics[width=\columnwidth]{CS_Young}
\end{flalign}
By using the Clebsch-Gordan (CG) coefficient of the permutation group $S_{n}$, one obtains the coupling scheme designed to construct the $\psi_{color}\otimes\psi_{spin}$ states.
Here, any CG coefficient of $S_{n}$ can be factorized into an isoscalar factor, called $K$ matrix, and the Clebsch-Gordan (CG) coefficient of $S_{n-1}$ \cite{Stancu:1999qr}.
The expression of Clebsch-Gordan (CG) coefficient of $S_{n}$ reads as
\begin{align}
&S([f^{\prime}]p^{\prime}q^{\prime}y^{\prime}[f^{\prime\prime}]p^{\prime\prime}q^{\prime\prime}y^{\prime\prime}\vert[f]pqy)=\nonumber\\
&K([f^{\prime}]p^{\prime}[f^{\prime\prime}]p^{\prime\prime}\vert[f]p)
S([f^{\prime}_{p^{\prime}}]q^{\prime}y^{\prime}[f^{\prime\prime}_{p^{\prime\prime}}]q^{\prime\prime}y^{\prime\prime}\vert[f_p]qy).
\label{K-matrix}
\end{align}
Here, $S$ in the left-hand (right-hand) side denotes a CG coefficient of $S_n$ ($S_{n-1}$) and
the $[f]$ ($[f_p]$) is a Young tableau of $S_{n}$ ($S_{n-1}$) with $[f]pqy$ as a specific Young-Yamanouchi basis vector. And, $p$ ($q$) is the row of the $n$ $(n-1)$-th particle in the Young-Yamanouchi basis vector, while $y$ is the distribution of the $n-2$ remaining particles. In fact, a similar application is also used in Refs. \cite{Park:2017jbn,Park:2016mez,Park:2016cmg,Park:2015nha,Park:2018oib}.
According to the isoscalar factors for $S_{3}$ and $S_{4}$ in Tables 6.2 and 6.3 of Ref. \cite{Stancu:1999qr}, the corresponding Clebsch-Gordan (CG) coefficient of $S_{4}$ can be obtained. The corresponding Young-Yamanouchi basis vector obtained from the $\psi_{color}\otimes\psi_{spin}$ coupling (see Eq.(\ref{colorspin})) are collected in Eq. (\ref{colorspin1}) of Appendix. \ref{sec10}.
We can combine the flavor wave functions with the $\psi_{color}\otimes\psi_{spin}$ wave functions (see Eq. \ref{colorspin1}) of $QQQQ\bar q$ pentaquark states to deduce the symmetry allowed $\psi_{flavor}\otimes\psi_{color}\otimes\psi_{spin}$ pentaquark wave functions.
In Table. \ref{flavor} of Appendix. \ref{sec10}, all the symmetrically allowed Young-Yamanouchi basis vector for the different flavor wave functions are listed.
By constructing all the possible $\psi_{flavor}\otimes\psi_{color}\otimes\psi_{spin}$ bases satisfied for $\{1234\}$, $\{123\}$, and $\{12\}\{34\}$ symmetry,
we can calculate the CMI matrices for the studied pentaquark states.
Here, we only present the expressions of CMI Hamiltonians for the $cccc\bar{n}$, $cccb\bar{n}$, and $ccbb\bar{n}$ pentaquark subsystems in Table \ref{nnnsQ} of Appendix. \ref{sec10}.
According to their similar symmetry properties, the expressions of CMI matrices for other pentaquark subsystems can be deduced for those of the the $cccc\bar{c}$, $cccb\bar{c}$, and $ccbb\bar{c}$ pentaquark subsystems.
\section{Mass Spectra and Decay Behaviors}\label{sec4}
As shown in Table \ref{flavor1}, according to symmetry properties, we can divide the $QQQQ\bar{q}$ pentaquark system into three groups: \\
\begin{enumerate}
\item[A.] The $cccc\bar{q}$ and $bbbb\bar{q}$ pentaquark subsystems;
\item[B.] The $cccb\bar{q}$ and $bbbc\bar{q}$ pentaquark subsystems;
\item[C.] The $ccbb\bar{q}$ pentaquark subsystem.
\end{enumerate}
For the $cccc\bar{q}$ and $bbbb\bar{q}$ subsystems, the wave functions should be to antisymmetric for the exchange between the 12, 13, 14, 23, 24, or 34 particles.
The $cccb\bar{q}$ and $bbbc\bar{q}$ wave functions are antisymmetric when exchanging the 12, 13, or 23 particles.
However, for the $ccbb\bar{q}$ subsystem, the antisymmetry is considered only for exchanging the 12 or 34 particles.
The fewer restrictions lead to more allowed wave functions.
Therefore, the number of basis states in Table \ref{flavor} increases from $cccc\bar{n}$ to $cccb\bar{n}$ and next to $ccbb\bar{n}$ due to the Pauli principle.
In the following, we will discuss the mass spectra and strong decay properties of $QQQQ\bar{q}$ pentaquark system group by group.
All of them are explicitly exotic states.
If such pentaquark states could be observed in experiment, its pentaquark state nature could be easily identified.
For simplicity, we use $\rm P_{content}$(Mass, $I$, $J^{P}$) to label a particular pentaquark state.
\begin{table*}[t]
\centering \caption{The estimated masses for the $QQQQ\bar{q}$ $(Q=c,b; q=n,s; n=u,d)$ system in units of MeV. The eigenvalues of $H_{\textrm{CMI}}$ matrix are listed in the second column.
The corresponding masses in the reference mass scheme are listed in third and/or fourth columns. The masses with the modified CMI model scheme are presented in the last column.}
\label{mass-QQQQq}
\renewcommand\arraystretch{1.25}
\begin{tabular}{cccc|c|cccc|c}
\bottomrule[1.5pt]
\bottomrule[0.5pt]
&\multicolumn{3}{c|}{reference mass}&{modified CMI}&&\multicolumn{3}{c|}{reference mass}&{modified CMI}\\
&\multicolumn{3}{c|}{scheme}&{model scheme}&&\multicolumn{3}{c|}{scheme}&{model scheme}
\\
\bottomrule[1pt]
\multicolumn{5}{l|}{$cccc\bar{n}$}&\multicolumn{5}{l}{$cccc\bar{s}$}\\
$J^P$&Eigenvalue&($D\Omega_{ccc}$)&&Mass&$J^P$&Eigenvalue&$(D_{s}\Omega_{ccc})$&&Mass\\
\bottomrule[0.7pt]
$\frac{3}{2}^{-}$ &
$26.4$&
$6761$&
&
$6761$&
$\frac{3}{2}^{-}$ &
$25.9$&
$6861$&
&
$6864$\\
$\frac{1}{2}^{-}$ &
$132.0$&
$6867$&
&
$6867$&
$\frac{1}{2}^{-}$ &
133.1&
6968&
&
6972\\
\bottomrule[0.5pt]
\multicolumn{5}{l|}{$bbbb\bar{n}$}&\multicolumn{5}{l}{$bbbb\bar{s}$}\\
$J^P$&Eigenvalue&($B\Omega_{bbb}$)&&Mass&$J^P$&Eigenvalue&$(B_{s}\Omega_{bbb})$&&Mass\\
\bottomrule[0.7pt]
$\frac{3}{2}^{-}$ &
$22.4$&
$19631$&
&
$19647$&
$\frac{3}{2}^{-}$ &
$21.3$&
$19720$&
&
$19736$\\
$\frac{1}{2}^{-}$ &
$56.0$&
$19664$&
&
$19681$&
$\frac{1}{2}^{-}$ &
$58.1$&
$19757$&
&
$19773$\\
\bottomrule[1.0pt]
\multicolumn{5}{l|}{$cccb\bar{n}$}&\multicolumn{5}{l}{$cccb\bar{s}$}\\
$J^P$&Eigenvalue&($B\Omega_{ccc}$)&($D\Omega_{ccb}$)&Mass&$J^P$&Eigenvalue&$(B_{s}\Omega_{ccc})$&$(D_{s}\Omega_{ccb})$&Mass\\
\bottomrule[0.7pt]
$\frac{5}{2}^{-}$ &
$37.6$&
$10110$&
$9816$&
$10110$&
$\frac{5}{2}^{-}$ &
$38.7$&
$10202$&
$9917$&
$10201$\\
$\frac{3}{2}^{-}$ &
$\begin{pmatrix}65.3\\22.0\\-40.6\end{pmatrix}$&
$\begin{pmatrix}10138\\10094\\10032\end{pmatrix}$&
$\begin{pmatrix}9843\\9800\\9738\end{pmatrix}$&
$\begin{pmatrix}10118\\10078\\9961\end{pmatrix}$&
$\frac{3}{2}^{-}$ &
$\begin{pmatrix}65.9\\21.3\\-43.1\end{pmatrix}$&
$\begin{pmatrix}10229\\10184\\10120\end{pmatrix}$&
$\begin{pmatrix}9944\\9900\\9835\end{pmatrix}$&
$\begin{pmatrix}10210\\10168\\10062\end{pmatrix}$\\
$\frac{1}{2}^{-}$ &
$\begin{pmatrix}110.4\\68.0\\-41.3\end{pmatrix}$&
$\begin{pmatrix}10183\\10140\\10031\end{pmatrix}$&
$\begin{pmatrix}9889\\9846\\9737\end{pmatrix}$&
$\begin{pmatrix}10134\\10062\\9946\end{pmatrix}$&
$\frac{1}{2}^{-}$ &
$\begin{pmatrix}111.6\\69.1\\-41.7\end{pmatrix}$&
$\begin{pmatrix}10275\\10232\\10121\end{pmatrix}$&
$\begin{pmatrix}9990\\9948\\9837\end{pmatrix}$&
$\begin{pmatrix}10228\\10166\\10047\end{pmatrix}$\\
\bottomrule[0.5pt]
\multicolumn{5}{l|}{$bbbc\bar{n}$}&\multicolumn{5}{l}{$bbbc\bar{s}$}\\
$J^P$&Eigenvalue&($D\Omega_{bbb}$)&($B\Omega_{bbc}$)&Mass&$J^P$&Eigenvalue&$(D_{s}\Omega_{bbb})$&$(B_{s}\Omega_{bbc})$&Mass\\
\bottomrule[0.7pt]
$\frac{5}{2}^{-}$ &
$49.6$&
$16320$&
$16544$&
$16318$&
$\frac{5}{2}^{-}$ &
$50.1$&
$16421$&
$16635$&
$16422$\\
$\frac{3}{2}^{-}$ &
$\begin{pmatrix}49.6\\16.6\\-94.5\end{pmatrix}$&
$\begin{pmatrix}16320\\16287\\16176\end{pmatrix}$&
$\begin{pmatrix}16544\\16511\\16400\end{pmatrix}$&
$\begin{pmatrix}16538\\16318\\16176\end{pmatrix}$&
$\frac{3}{2}^{-}$ &
$\begin{pmatrix}50.2\\16.1\\-94.4\end{pmatrix}$&
$\begin{pmatrix}16421\\16387\\16274\end{pmatrix}$&
$\begin{pmatrix}16636\\16601\\16489\end{pmatrix}$&
$\begin{pmatrix}16626\\16422\\16277\end{pmatrix}$\\
$\frac{1}{2}^{-}$ &
$\begin{pmatrix}72.2\\30.3\\-7.5\end{pmatrix}$&
$\begin{pmatrix}16343\\16301\\16263\end{pmatrix}$&
$\begin{pmatrix}16567\\16525\\16487\end{pmatrix}$&
$\begin{pmatrix}16574\\16523\\16315\end{pmatrix}$&
$\frac{1}{2}^{-}$ &
$\begin{pmatrix}74.0\\31.8\\-8.7\end{pmatrix}$&
$\begin{pmatrix}16445\\16403\\16362\end{pmatrix}$&
$\begin{pmatrix}16659\\16617\\16577\end{pmatrix}$&
$\begin{pmatrix}16663\\16611\\16418\end{pmatrix}$\\
\bottomrule[1.0pt]
\multicolumn{5}{l|}{$ccbb\bar{n}$}&\multicolumn{5}{l}{$ccbb\bar{s}$}\\
$J^P$&Eigenvalue&($B\Omega_{ccb}$)&($D\Omega_{bbc}$)&Mass&$J^P$&Eigenvalue&$(B_{s}\Omega_{ccb})$&$(D_{s}\Omega_{bbc})$&Mass\\
\bottomrule[0.7pt]
$\frac{5}{2}^{-}$ &
$42.1$&
$13358$&
$13199$&
$13244$&
$\frac{5}{2}^{-}$ &
$42.9$&
$13450$&
$13300$&
$13341$\\
$\frac{3}{2}^{-}$ &
$\begin{pmatrix}61.2\\24.3\\11.3\\-70.4\end{pmatrix}$&
$\begin{pmatrix}13377\\13340\\13327\\13246\end{pmatrix}$&
$\begin{pmatrix}13218\\13181\\13168\\13086\end{pmatrix}$&
$\begin{pmatrix}13383\\13231\\13221\\13100\end{pmatrix}$&
$\frac{3}{2}^{-}$ &
$\begin{pmatrix}61.7\\24.1\\12.0\\-72.6\end{pmatrix}$&
$\begin{pmatrix}13468\\13431\\13419\\13334\end{pmatrix}$&
$\begin{pmatrix}13319\\13281\\13269\\13185\end{pmatrix}$&
$\begin{pmatrix}13470\\13329\\13319\\13200\end{pmatrix}$\\
$\frac{1}{2}^{-}$ &
$\begin{pmatrix}90.3\\44.8\\-5.4\\-84.4\end{pmatrix}$&
$\begin{pmatrix}13406\\13361\\13311\\13232\end{pmatrix}$&
$\begin{pmatrix}13247\\13202\\13151\\13072\end{pmatrix}$&
$\begin{pmatrix}13414\\13242\\13212\\13086\end{pmatrix}$&
$\frac{1}{2}^{-}$ &
$\begin{pmatrix}91.7\\46.3\\-7.3\\-85.6\end{pmatrix}$&
$\begin{pmatrix}13498\\13453\\13399\\13321\end{pmatrix}$&
$\begin{pmatrix}13349\\13303\\13250\\13172\end{pmatrix}$&
$\begin{pmatrix}13502\\13343\\13307\\13187\end{pmatrix}$\\
\bottomrule[0.5pt]
\bottomrule[1.5pt]
\end{tabular}
\end{table*}
\begin{figure*}[t]
\begin{tabular}{cc}
\includegraphics[width=265pt]{ccccN.pdf}&
\includegraphics[width=265pt]{ccccS.pdf}\\
(a) \begin{tabular}{c} $cccc\bar{n}$ states\end{tabular} &(b) $cccc\bar{s}$ states\\
\includegraphics[width=265pt]{bbbbN.pdf}&
\includegraphics[width=265pt]{bbbbS.pdf}\\
(c) \begin{tabular}{c} $bbbb\bar{n}$ states\end{tabular} &(d) $bbbb\bar{s}$ states\\
\includegraphics[width=265pt]{cccbN.pdf}&
\includegraphics[width=265pt]{cccbS.pdf}\\
(e) \begin{tabular}{c} $cccb\bar{n}$ states\end{tabular} &(f) $cccb\bar{s}$ states\\
\includegraphics[width=265pt]{bbbcN.pdf}&
\includegraphics[width=265pt]{bbbcS.pdf}\\
(g) \begin{tabular}{c} $bbbc\bar{n}$ states\end{tabular} &(h) $bbbc\bar{s}$ states\\
\includegraphics[width=265pt]{ccbbN.pdf}&
\includegraphics[width=265pt]{ccbbS.pdf}\\
(i) \begin{tabular}{c} $ccbb\bar{n}$ states\end{tabular} &(j) $ccbb\bar{s}$ states\\
\end{tabular}
\caption{
Relative positions (units: MeV) for the $cccc\bar{n}$, $cccc\bar{s}$, $bbbb\bar{n}$, $bbbb\bar{s}$, $cccb\bar{n}$, $cccb\bar{s}$, $bbbc\bar{n}$, $bbbc\bar{s}$, $ccbb\bar{n}$, and $ccbb\bar{s}$ pentaquark states labeled with solid lines.
The dotted lines denote various S-wave baryon-meson thresholds, and the superscripts of the labels, e.g. $(\Omega_{ccc}D^{*})^{5/2,3/2,1/2}$, represent the possible total angular momenta of the channels.
We mark the relatively stable pentaquarks, unable to decay into the S-wave baryon-meson states, with ``$\diamond$" after their masses. We mark the pentaquark whose wave function overlaps with that of one special baryon-meson state more than 90\% with
``$\star$" after their masses.
}\label{fig-QQQQq}
\end{figure*}
\begin{table*}[t]
\centering \caption{The eigenvectors of the $cccc\bar{n}$, $cccc\bar{s}$, $bbbb\bar{n}$, $bbbb\bar{s}$, $cccb\bar{n}$, $cccb\bar{s}$, $bbbc\bar{n}$, $bbbc\bar{s}$, $ccbb\bar{n}$, and $ccbb\bar{s}$ pentaquark subsystems. The masses are all in units of MeV. See the caption of Fig. \ref{fig-QQQQq} for meanings of ``$\diamond$" and ``$\star$".
}\label{eigenvector-QQQQq}
\renewcommand\arraystretch{1.15}
\begin{tabular}{cl|cc|cccc|l|cc|cccc}
\bottomrule[1.5pt]
\bottomrule[0.5pt]
&\multicolumn{1}{c}{}&\multicolumn{2}{c|}{$ccc\bigotimes c\bar{n}$}&&\multicolumn{2}{c|}{$ccc \bigotimes c\bar{s}$}&&\multicolumn{1}{c}{}&\multicolumn{2}{c|}{$bbb\bigotimes b\bar{n}$}&&\multicolumn{2}{c|}{$bbb \bigotimes b\bar{s}$}\\
$J^P$&Mass&$\Omega_{ccc}D^{*}$&\multicolumn{1}{c|}{$\Omega_{ccc}D$}&\multicolumn{1}{c|}{Mass}&$\Omega_{ccc} D_{s}^{*}$&\multicolumn{1}{c|}{$\Omega_{ccc} D_s$}
&&Mass&$\Omega_{bbb}B^{*}$&$\Omega_{bbb}B$&\multicolumn{1}{c|}{Mass}&$\Omega_{bbb} B_{s}^{*}$&\multicolumn{1}{c|}{$\Omega_{bbb} B_s$}\\
\Xcline{1-7}{0.7pt}\Xcline{9-14}{0.7pt}
$\frac{3}{2}^{-}$&6761&0.456&-0.354&\multicolumn{1}{c|}{6864}&0.456&\multicolumn{1}{c|}{-0.354}&&19647&0.456&-0.354&\multicolumn{1}{c|}{19736}&0.456&\multicolumn{1}{c|}{-0.354}\\
$\frac{1}{2}^{-}$&6867&-0.577&&\multicolumn{1}{c|}{6972}&0.577&\multicolumn{1}{c|}{}&&19681&-0.577&&\multicolumn{1}{c|}{19773}&0.577&\multicolumn{1}{c|}{}\\
\bottomrule[0.5pt]
\multicolumn{2}{l|}{}&\multicolumn{2}{c}{$ccc\bigotimes b\bar{n}$}&\multicolumn{4}{c|}{$ccb\bigotimes c\bar{n}$}&\multicolumn{1}{l}{}&\multicolumn{2}{c}{$ccc\bigotimes b\bar{s}$}&\multicolumn{4}{c}{$ccb\bigotimes c\bar{s}$}\\
$J^P$&Mass&$\Omega_{ccc} B^{*}$&$\Omega_{ccc} B$
&$\Omega_{ccb}^{*}D^{*}$&$\Omega_{ccb}^{*}D$&$\Omega_{ccb}D^{*}$&$\Omega_{ccb}D$&
Mass&$\Omega_{ccc} B_{s}^{*}$&$\Omega_{ccc} B_{s}$
&$\Omega_{ccb}^{*}D_{s}^{*}$&$\Omega_{ccb}^{*}D_{s}$&$\Omega_{ccb}D_{s}^{*}$&$\Omega_{ccb}D_{s}$\\
\bottomrule[0.7pt]
$\frac{5}{2}^{-}$&$10110\star$&1.000&&0.333&&&&$10201\star$&1.000&&0.333&&&\\
$\frac{3}{2}^{-}$&$10118\star$&0.950&0.188&0.217&0.140&-0.255&&$10210\star$&0.939&0.197&0.234&0.127&-0.258\\
&$10078\star$&-0.261&0.917&0.333&-0.019&0.174&&10168&-0.289&0.895&0.345&-0.043&0.169\\
&$9961$&0.172&0.352&0.372&-0.450&-0.230&&10062&0.184&0.399&0.350&-0.452&-0.231\\
$\frac{1}{2}^{-}$
&$10134\star$&0.914&&-0.333&&-0.201&-0.121&10228&0.891&&-0.365&&-0.188&-0.106\\
&10062&-0.246&&0.498&&-0.415&-0.064&10166&-0.290&&0.477&&-0.430&-0.065\\
&9946&0.323&&0.220&&0.397&-0.451&10048&0.349&&0.217&&0.387&-0.455\\
\bottomrule[0.5pt]
\multicolumn{2}{l|}{}&\multicolumn{2}{c}{$bbb\bigotimes c\bar{n}$}&\multicolumn{4}{c|}{$bbc\bigotimes b\bar{n}$}&\multicolumn{1}{l}{}&\multicolumn{2}{c}{$bbb\bigotimes c\bar{s}$}&\multicolumn{4}{c}{$bbc\bigotimes b\bar{s}$}\\
$J^P$&Mass&$\Omega_{bbb} D^{*}$&$\Omega_{bbb} D$
&$\Omega_{bbc}^{*}B^{*}$&$\Omega_{bbc}^{*}B$&$\Omega_{bbc}B^{*}$&$\Omega_{bbc}B$&
Mass&$\Omega_{bbb} D_{s}^{*}$&$\Omega_{bbb} D_{s}$
&$\Omega_{bbc}^{*}B_{s}^{*}$&$\Omega_{bbc}^{*}B_{s}$&$\Omega_{bbc}B_{s}^{*}$&$\Omega_{bbc}B_{s}$\\
\bottomrule[0.7pt]
$\frac{5}{2}^{-}$
&$16318\star\diamond$&1.000&&0.333&&&&$16422\star\diamond$&1.000&&0.333&&&\\
$\frac{3}{2}^{-}$
&16538&-0.004&-0.052&0.508&-0.375&-0.213&&16626&-0.009&-0.056&0.509&-0.373&-0.213\\
&$16318\star$&0.999&-0.001&0.053&0.217&-0.139&&$16422\star\diamond$&0.999&-0.001&0.051&0.218&-0.140\\
&$16176\star\diamond$&0.001&0.999&0.189&0.187&0.204&&$16277\star\diamond$&0.001&0.998&0.187&0.188&0.204\\
$\frac{1}{2}^{-}$
&16574&-0.085&&-0.634&&0.135&0.146&16663&-0.092&&-0.634&&0.139&0.141\\
&16523&-0.074&&0.045&&0.580&-0.323&16611&-0.086&&0.052&&0.580&-0.321\\
&$16315\star\diamond$&0.994&&0.054&&0.127&-0.130&$16418\star\diamond$&0.992&&0.049&&0.121&0.315\\
\bottomrule[0.5pt]
\multicolumn{2}{l|}{}&\multicolumn{4}{c}{$bbc\bigotimes c\bar{n}$}&\multicolumn{4}{c|}{$ccb\bigotimes b\bar{n}$}\\
$J^P$&Mass&
$\Omega^{*}_{bbc}D^{*}$&\multicolumn{1}{c}{$\Omega^{*}_{bbc}D$}&$\Omega_{bbc}D^{*}$&\multicolumn{1}{c|}{$\Omega_{bbc}D$}&
$\Omega^{*}_{ccb}B^{*}$&\multicolumn{1}{c}{$\Omega^{*}_{ccb}B$}&\multicolumn{1}{c}{$\Omega_{ccb}B^{*}$}&\multicolumn{1}{c|}{$\Omega_{ccb}B$}\\
\Xcline{1-10}{0.7pt}
$\frac52^{-}$&13244&-0.577&\multicolumn{1}{c}{}&&\multicolumn{1}{c|}{}&-0.577&\multicolumn{1}{c}{}&\multicolumn{1}{c}{}&\multicolumn{1}{c|}{}\\
$\frac32^{-}$
&13383&-0.018&\multicolumn{1}{c}{-0.052}&-0.013&\multicolumn{1}{c|}{}&0.566&\multicolumn{1}{c}{-0.372}&\multicolumn{1}{c}{0.452}&\multicolumn{1}{c|}{}\\
&13231&0.621&\multicolumn{1}{c}{0.033}&-0.072&\multicolumn{1}{c|}{}&-0.077&\multicolumn{1}{c}{-0.454}&\multicolumn{1}{c}{-0.251}&\multicolumn{1}{c|}{}\\
&13221&-0.155&\multicolumn{1}{c}{-0.167}&-0.586&\multicolumn{1}{c|}{}&-0.319&\multicolumn{1}{c}{-0.040}&\multicolumn{1}{c}{0.410}&\multicolumn{1}{c|}{}\\
&13100&0.250&\multicolumn{1}{c}{-0.620}&0.310&\multicolumn{1}{c|}{}&0.209&\multicolumn{1}{c}{0.265}&\multicolumn{1}{c}{-0.096}&\multicolumn{1}{c|}{}\\
$\frac12^{-}$
&13414&-0.099&\multicolumn{1}{c}{}&0.053&\multicolumn{1}{c|}{0.024}&-0.710&\multicolumn{1}{c}{}&\multicolumn{1}{c}{-0.222}&\multicolumn{1}{c|}{-0.317}\\
&13242&0.641&\multicolumn{1}{c}{}&0.220&\multicolumn{1}{c|}{-0.107}&-0.124&\multicolumn{1}{c}{}&\multicolumn{1}{c}{0.385}&\multicolumn{1}{c|}{0.180}\\
&13212&0.197&\multicolumn{1}{c}{}&-0.535&\multicolumn{1}{c|}{0.052}&-0.078&\multicolumn{1}{c}{}&\multicolumn{1}{c}{-0.271}&\multicolumn{1}{c|}{0.510}\\
&13086&0.310&\multicolumn{1}{c}{}&0.155&\multicolumn{1}{c|}{0.634}&0.173&\multicolumn{1}{c}{}&\multicolumn{1}{c}{-0.301}&\multicolumn{1}{c|}{-0.155}\\
\Xcline{1-10}{0.5pt}
\multicolumn{2}{l|}{}&\multicolumn{4}{c}{$bbc\bigotimes c\bar{s}$}&\multicolumn{4}{c|}{$ccb\bigotimes b\bar{s}$}\\
$J^P$&Mass&
$\Omega^{*}_{bbc}D_{s}^{*}$&\multicolumn{1}{c}{$\Omega^{*}_{bbc}D_{s}$}&$\Omega_{bbc}D_{s}^{*}$&\multicolumn{1}{c|}{$\Omega_{bbc}D_{s}$}&
$\Omega^{*}_{ccb}B_{s}^{*}$&\multicolumn{1}{c}{$\Omega^{*}_{ccb}B_{s}$}&\multicolumn{1}{c}{$\Omega_{ccb}B_{s}^{*}$}&\multicolumn{1}{c|}{$\Omega_{ccb}B_{s}$}\\
\Xcline{1-10}{0.7pt}
$\frac52^{-}$&13341&-0.577&\multicolumn{1}{c}{}&&\multicolumn{1}{c|}{}&-0.577&\multicolumn{1}{c}{}&\multicolumn{1}{c}{}&\multicolumn{1}{c|}{}\\
$\frac32^{-}$
&13470&-0.025&\multicolumn{1}{c}{-0.056}&-0.013&\multicolumn{1}{c|}{}&0.569&\multicolumn{1}{c}{-0.366}&\multicolumn{1}{c}{0.453}&\multicolumn{1}{c|}{}\\
&13329&0.647&\multicolumn{1}{c}{0.041}&0.051&\multicolumn{1}{c|}{}&-0.122&\multicolumn{1}{c}{-0.448}&\multicolumn{1}{c}{-0.167}&\multicolumn{1}{c|}{}\\
&13319&-0.041&\multicolumn{1}{c}{-0.142}&-0.596&\multicolumn{1}{c|}{}&-0.292&\multicolumn{1}{c}{0.052}&\multicolumn{1}{c}{0.448}&\multicolumn{1}{c|}{}\\
&13200&-0.226&\multicolumn{1}{c}{0.626}&-0.293&\multicolumn{1}{c|}{}&0.218&\multicolumn{1}{c}{0.282}&\multicolumn{1}{c}{-0.104}&\multicolumn{1}{c|}{}\\
$\frac12^{-}$
&13502&-0.109&\multicolumn{1}{c}{}&0.057&\multicolumn{1}{c|}{0.025}&-0.713&\multicolumn{1}{c}{}&\multicolumn{1}{c}{-0.220}&\multicolumn{1}{c|}{-0.308}\\
&13343&0.648&\multicolumn{1}{c}{}&0.232&\multicolumn{1}{c|}{-0.085}&-0.108&\multicolumn{1}{c}{}&\multicolumn{1}{c}{0.380}&\multicolumn{1}{c|}{0.172}\\
&13307&0.201&\multicolumn{1}{c}{}&-0.531&\multicolumn{1}{c|}{0.055}&-0.069&\multicolumn{1}{c}{}&\multicolumn{1}{c}{-0.265}&\multicolumn{1}{c|}{0.515}\\
&13187&0.286&\multicolumn{1}{c}{}&0.150&\multicolumn{1}{c|}{0.637}&-0.177&\multicolumn{1}{c}{}&\multicolumn{1}{c}{0.313}&\multicolumn{1}{c|}{0.164}\\
\bottomrule[0.5pt]
\bottomrule[1.5pt]
\end{tabular}
\end{table*}
\begin{table*}[t]
\centering \caption{
The values of $k\cdot |c_{i}|^{2}$ for the the $cccc\bar{n}$, $cccc\bar{s}$, $bbbb\bar{n}$, $bbbb\bar{s}$, $cccb\bar{n}$, $cccb\bar{s}$, $bbbc\bar{n}$, $bbbc\bar{s}$, $ccbb\bar{n}$, and $ccbb\bar{s}$ pentaquark states. The masses are all in units of MeV.
The decay channel is marked with ``$\times$'' if kinetically forbidden.
See the caption of Fig. \ref{fig-QQQQq} for meanings of ``$\diamond$" and ``$\star$".
One can roughly estimate the relative decay widths between different decay processes of different initial pentaquark states with this table if neglecting the $\gamma_i$ differences.
}\label{value-QQQQq}
\renewcommand\arraystretch{1.15}
\begin{tabular}{cl|cc|cccc|l|cc|cccc}
\bottomrule[1.5pt]
\bottomrule[0.5pt]
&\multicolumn{1}{c}{}&\multicolumn{2}{c|}{$ccc\bigotimes c\bar{n}$}&&\multicolumn{2}{c|}{$ccc \bigotimes c\bar{s}$}&&\multicolumn{1}{c}{}&\multicolumn{2}{c|}{$bbb\bigotimes b\bar{n}$}&&\multicolumn{2}{c|}{$bbb \bigotimes b\bar{s}$}\\
$J^P$&Mass&$\Omega_{ccc}D^{*}$&\multicolumn{1}{c|}{$\Omega_{ccc}D$}&\multicolumn{1}{c|}{Mass}&$\Omega_{ccc} D_{s}^{*}$&\multicolumn{1}{c|}{$\Omega_{ccc} D_s$}
&&Mass&$\Omega_{bbb}B^{*}$&$\Omega_{bbb}B$&\multicolumn{1}{c|}{Mass}&$\Omega_{bbb} B_{s}^{*}$&\multicolumn{1}{c|}{$\Omega_{bbb} B_s$}\\
\Xcline{1-7}{0.7pt}\Xcline{9-14}{0.7pt}
$\frac{3}{2}^{-}$&6761&$\times$&67&\multicolumn{1}{c|}{6864}&$\times$&\multicolumn{1}{c|}{70}&&19647&66&84&\multicolumn{1}{c|}{19736}&62&\multicolumn{1}{c|}{86}\\
$\frac{1}{2}^{-}$&6867&151&&\multicolumn{1}{c|}{6972}&156&\multicolumn{1}{c|}{}&&19681&201&&\multicolumn{1}{c|}{19773}&204&\multicolumn{1}{c|}{}\\
\bottomrule[0.5pt]
\multicolumn{2}{l|}{}&\multicolumn{2}{c}{$ccc\bigotimes b\bar{n}$}&\multicolumn{4}{c|}{$ccb\bigotimes c\bar{n}$}&\multicolumn{1}{l}{}&\multicolumn{2}{c}{$ccc\bigotimes b\bar{s}$}&\multicolumn{4}{c}{$ccb\bigotimes c\bar{s}$}\\
$J^P$&Mass&$\Omega_{ccc} B^{*}$&$\Omega_{ccc} B$
&$\Omega_{ccb}^{*}D^{*}$&$\Omega_{ccb}^{*}D$&$\Omega_{ccb}D^{*}$&$\Omega_{ccb}D$&
Mass&$\Omega_{ccc} B_{s}^{*}$&$\Omega_{ccc} B_{s}$
&$\Omega_{ccb}^{*}D_{s}^{*}$&$\Omega_{ccb}^{*}D_{s}$&$\Omega_{ccb}D_{s}^{*}$&$\Omega_{ccb}D_{s}$\\
\bottomrule[0.7pt]
$\frac{5}{2}^{-}$
&$10110\star$&$\times$&&56&&&&$10201\star$&13&&53&&&\\
$\frac{3}{2}^{-}$
&$10118\star$&173&18&25&17&40&&$10210\star$&184&21&28&14&40\\
&$10078\star$&$\times$&217&43&0.3&15&&10168&$\times$&223&40&1&13\\
&$9961$&$\times$&$\times$&$\times$&94&$\times$&&10062&$\times$&$\times$&$\times$&98&$\times$\\
$\frac{1}{2}^{-}$
&$10134\star$&291&&64&&27&14&10228&296&&75&&23&11\\
&10062&$\times$&&76&&77&3&10166&$\times$&&75&&86&3\\
&9946&$\times$&&$\times$&&$\times$&104&10048&$\times$&&$\times$&&$\times$&111\\
\bottomrule[0.5pt]
\multicolumn{2}{l|}{}&\multicolumn{2}{c}{$bbb\bigotimes c\bar{n}$}&\multicolumn{4}{c|}{$bbc\bigotimes b\bar{n}$}&\multicolumn{1}{l}{}&\multicolumn{2}{c}{$bbb\bigotimes c\bar{s}$}&\multicolumn{4}{c}{$bbc\bigotimes b\bar{s}$}\\
$J^P$&Mass&$\Omega_{bbb} D^{*}$&$\Omega_{bbb} D$
&$\Omega_{bbc}^{*}B^{*}$&$\Omega_{bbc}^{*}B$&$\Omega_{bbc}B^{*}$&$\Omega_{bbc}B$&
Mass&$\Omega_{bbb} D_{s}^{*}$&$\Omega_{bbb} D_{s}$
&$\Omega_{bbc}^{*}B_{s}^{*}$&$\Omega_{bbc}^{*}B_{s}$&$\Omega_{bbc}B_{s}^{*}$&$\Omega_{bbc}B_{s}$\\
\bottomrule[0.7pt]
$\frac{5}{2}^{-}$
&$16318\star\diamond$&$\times$&&$\times$&&&&$16422\star\diamond$&$\times$&&$\times$&&&\\
$\frac{3}{2}^{-}$
&16538&0.02&3&89&94&27&&16626&0.1&4&82&94&26\\
&$16318\star$&$\times$&0.001&$\times$&$\times$&$\times$&&$16422\star\diamond$&$\times$&$\times$&$\times$&$\times$&$\times$\\
&$16176\star\diamond$&$\times$&$\times$&$\times$&$\times$&$\times$&&$16277\star\diamond$&$\times$&$\times$&$\times$&$\times$&$\times$\\
$\frac{1}{2}^{-}$
&16574&7&&248&&14&21&16663&8&&247&&15&20\\
&16523&5&&0.2&&164&79&16611&6&&$\times$&&160&78\\
&$16315\star\diamond$&$\times$&&$\times$&&$\times$&$\times$&$16418\star\diamond$&$\times$&&$\times$&&$\times$&$\times$\\
\bottomrule[0.5pt]
\multicolumn{2}{l|}{}&\multicolumn{4}{c}{$bbc\bigotimes c\bar{n}$}&\multicolumn{4}{c|}{$ccb\bigotimes b\bar{n}$}\\
$J^P$&Mass&
$\Omega^{*}_{bbc}D^{*}$&\multicolumn{1}{c}{$\Omega^{*}_{bbc}D$}&$\Omega_{bbc}D^{*}$&\multicolumn{1}{c|}{$\Omega_{bbc}D$}&
$\Omega^{*}_{ccb}B^{*}$&\multicolumn{1}{c}{$\Omega^{*}_{ccb}B$}&\multicolumn{1}{c}{$\Omega_{ccb}B^{*}$}&\multicolumn{1}{c|}{$\Omega_{ccb}B$}\\
\Xcline{1-10}{0.7pt}
$\frac52^{-}$&13244&119&\multicolumn{1}{c}{}&&\multicolumn{1}{c|}{}&$\times$&\multicolumn{1}{c}{}&\multicolumn{1}{c}{}&\multicolumn{1}{c|}{}\\
$\frac32^{-}$
&13383&0.3&\multicolumn{1}{c}{3}&0.2&\multicolumn{1}{c|}{}&154&\multicolumn{1}{c}{100}&\multicolumn{1}{c}{135}&\multicolumn{1}{c|}{}\\
&13231&111&\multicolumn{1}{c}{0.8}&2&\multicolumn{1}{c|}{}&$\times$&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c|}{}\\
&13221&5&\multicolumn{1}{c}{20}&135&\multicolumn{1}{c|}{}&$\times$&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c|}{}\\
&13100&$\times$&\multicolumn{1}{c}{127}&$\times$&\multicolumn{1}{c|}{}&$\times$&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c|}{}\\
$\frac12^{-}$
&13414&8&\multicolumn{1}{c}{}&3&\multicolumn{1}{c|}{0.7}&332&\multicolumn{1}{c}{}&\multicolumn{1}{c}{39}&\multicolumn{1}{c|}{96}\\
&13242&142&\multicolumn{1}{c}{}&23&\multicolumn{1}{c|}{9}&$\times$&\multicolumn{1}{c}{}&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c|}{$\times$}\\
&13212&5&\multicolumn{1}{c}{}&101&\multicolumn{1}{c|}{2}&$\times$&\multicolumn{1}{c}{}&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c|}{$\times$}\\
&13086&$\times$&\multicolumn{1}{c}{}&$\times$&\multicolumn{1}{c|}{164}&$\times$&\multicolumn{1}{c}{}&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c|}{$\times$}\\
\Xcline{1-10}{0.5pt}
\multicolumn{2}{l|}{}&\multicolumn{4}{c}{$bbc\bigotimes c\bar{s}$}&\multicolumn{4}{c|}{$ccb\bigotimes b\bar{s}$}\\
$J^P$&Mass&
$\Omega^{*}_{bbc}D_{s}^{*}$&\multicolumn{1}{c}{$\Omega^{*}_{bbc}D_{s}$}&$\Omega_{bbc}D_{s}^{*}$&\multicolumn{1}{c|}{$\Omega_{bbc}D_{s}$}&
$\Omega^{*}_{ccb}B_{s}^{*}$&\multicolumn{1}{c}{$\Omega^{*}_{ccb}B_{s}$}&\multicolumn{1}{c}{$\Omega_{ccb}B_{s}^{*}$}&\multicolumn{1}{c|}{$\Omega_{ccb}B_{s}$}\\
\Xcline{1-10}{0.7pt}
$\frac52^{-}$&13341&113&\multicolumn{1}{c}{}&&\multicolumn{1}{c|}{}&$\times$&\multicolumn{1}{c}{}&\multicolumn{1}{c}{}&\multicolumn{1}{c|}{}\\
$\frac32^{-}$
&13470&8&\multicolumn{1}{c}{3}&0.7&\multicolumn{1}{c|}{}&332&\multicolumn{1}{c}{39}&\multicolumn{1}{c}{96}&\multicolumn{1}{c|}{}\\
&13329&142&\multicolumn{1}{c}{23}&9&\multicolumn{1}{c|}{}&$\times$&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c|}{}\\
&13319&5&\multicolumn{1}{c}{101}&2&\multicolumn{1}{c|}{}&$\times$&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c|}{}\\
&13200&$\times$&\multicolumn{1}{c}{$\times$}&164&\multicolumn{1}{c|}{}&$\times$&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c|}{}\\
$\frac12^{-}$
&13502&10&\multicolumn{1}{c}{}&3&\multicolumn{1}{c|}{0.7}&328&\multicolumn{1}{c}{}&\multicolumn{1}{c}{38}&\multicolumn{1}{c|}{92}\\
&13343&148&\multicolumn{1}{c}{}&26&\multicolumn{1}{c|}{6}&$\times$&\multicolumn{1}{c}{}&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c|}{$\times$}\\
&13307&$\times$&\multicolumn{1}{c}{}&92&\multicolumn{1}{c|}{2}&$\times$&\multicolumn{1}{c}{}&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c|}{$\times$}\\
&13187&$\times$&\multicolumn{1}{c}{}&$\times$&\multicolumn{1}{c|}{173}&$\times$&\multicolumn{1}{c}{}&\multicolumn{1}{c}{$\times$}&\multicolumn{1}{c|}{$\times$}\\
\bottomrule[0.5pt]
\bottomrule[1.5pt]
\end{tabular}
\end{table*}
\subsubsection{The $cccc\bar{q}$ and $bbbb\bar{q}$ pentaquark states}
Here, we first discuss the $cccc\bar{q}$ and $bbbb\bar{q}$ pentaquark subsystems.
Because of the symmetrical constraint from Pauli principle, i.e., fully antisymmetric among the first four charm quarks, the ground $J^P=5/2^{-}$ pentaquark state with $cccc\bar{q}$ and $bbbb\bar{q}$ can not exist.
We only find two ground states: a $J^{P}=3/2^{-}$ state and a $J^{P}=1/2^{-}$ state for these subsystems.
Diagonalizing the Hamiltonians in Table \ref{nnnsQ} with the corresponding parameters in Table \ref{parameter2}, we can obtain the corresponding mass spectra for $cccc\bar{q}$ and $bbbb\bar{q}$ pentaquark subsystems and present them in Table \ref{mass-QQQQq}.
For the reference mass scheme, the only combination of meson-baryon reference systems are ($\Omega_{ccc}$)+($D$), ($\Omega_{ccc}$)+($D_{s}$), ($\Omega_{bbb}$)+($\bar{B}$), and ($\Omega_{bbb}$)+($\bar{B}_{s}$) for the $cccc\bar{n}$, $cccc\bar{s}$, $bbbb\bar{n}$, and $bbbb\bar{s}$ subsystems, respectively.
The modified CMI model scheme takes the chromoelectric interaction explicitly compared to the reference mass scheme, and therefore we use the results in this scheme for the following analysis.
Based on the results calculated from the modified CMI model scheme, we plot the mass spectra of the $cccc\bar{n}$, $cccc\bar{s}$, $bbbb\bar{n}$, and $bbbb\bar{s}$ subsystems in Fig. \ref{fig-QQQQq} (a)-(d), respectively.
Moreover, we also plot all the baryon-meson thresholds which they can decay to through quark rearrangement in Fig. \ref{fig-QQQQq} (a)-(d).
Meanwhile, we label the spin of the baryon-meson states with superscript.
When the spin of an initial pentaquark state is equal to the number in the superscript of a baryon-meson state,
it may decay into that baryon-meson channel through S wave.
Here, we define the relatively ``stable" pentaquark states as those which cannot decay into the S wave baryon-meson states.
Meanwhile, we label these stable pentaquark states with ``$\diamond$'' in the relevant figure and tables.
Based on the obtained $cccc\bar{q}$ and $bbbb\bar{q}$ pentaquark spectra,
we can discuss the possible decay patterns by considering different rearrangement of quarks in the corresponding pentaquark states. The discussion of possible decay patterns for these pentaquark states would be helpful for the observation in experiments.
From Fig. \ref{fig-QQQQq} (a)-(d), we can easily find that the $J^{P}=3/2^{-}$ state generally have smaller masses than that of the $J^{P}=1/2^{-}$ state in the $cccc\bar{q}$ and $bbbb\bar{q}$ pentaquark subsystems.
We also find that all the $cccc\bar{q}$ and $bbbb\bar{q}$ pentaquark states have strong decay channels, which indicates that in the modified CMI model, no stable $cccc\bar{q}$ and $bbbb\bar{q}$ pentaquark exists.
In addition to the mass spectra, the eigenvectors of pentaquark states will also provide important information about the two-body strong decay of multiquark states \cite{Jaffe:1976ig,Strottman:1979qu,Weng:2019ynva,Weng:2020jao,Zhao:2014qva,Wang:2015epa}.
The overlap for the pentaquark with a specific baryon $\otimes$ meson state can be calculated by transforming the eigenvectors of the pentaquark states into the baryon $\otimes$ meson configuration.
In the $QQQ\otimes Q\bar{q}$ configuration, the color wave function of the pentaquark falls into two categories:
the color-singlet $|(QQQ)^{1_{c}}(Q\bar{q})^{1_{c}}\rangle$ and the color-octet $|(QQQ)^{8_{c}}(Q\bar{q})^{8_{c}}\rangle$.
The former one can easily dissociate into an S-wave baryon and meson (the so-called Okubo-Zweig-Iizuka (OZI)-superallowed decays), while the latter one cannot fall apart without the gluon exchange force.
In this work, we only focus on the OZI-superallowed pentaquark decay process.
It means that only the $C_3$ color part in Eq.(\ref{eq-color4}) is considered in the color space.
For the two body decay via $L$-wave process, the expression describing partial decay width can be parameterized as \cite{Weng:2019ynva,Weng:2020jao}
\begin{eqnarray}\label{Eq20}
\Gamma_{i}=\gamma_{i}\alpha\frac{k^{2L+1}}{m^{2L}}\cdot|c_{i}|^{2},
\end{eqnarray}
where $\alpha$ is an effective coupling constant,
$m$ is the mass of the initial state,
$k$ is the momentum of the final states in the rest frame.
$c_i$ is the overlap between the pentaquark wave function and the meson + baryon wave function. Take $P_{c^{4}\bar{n}}(6761, 1/2, 3/2^{-})\to \Omega_{ccc} D^{*}$ as an example,
\begin{eqnarray}
c_i&\equiv& \langle\Omega_{ccc}\otimes D^{*} |P_{c^{4}\bar{n}}(6761, \frac12, \frac32^{-})\rangle\\
&=&\langle(ccc)^{1_c}_{S=1/2}\otimes (c\bar n)^{1_c}_{S=1} |P_{c^{4}\bar{n}}(6761, \frac12, \frac32^{-})\rangle=0.456. \nonumber\\
\end{eqnarray}
We show all possible overlaps between a pentaquark state and its possible $|(QQQ)^{1_{c}}(Q\bar{q})^{1_{c}}\rangle$ and $|(QQQ)^{1_{c}}(Q\bar{q})^{1_{c}}\rangle$ components in Table \ref{eigenvector-QQQQq}.
For the decay processes that we are interested in, $(k/m)^{2}$ is of $\mathcal{O}(10^{-2})$ or even smaller. Thus we only consider the $S$-wave decays.
As for the $\gamma_{i}$, it depends on the spatial wave functions of initial and final states, and may not be the same for different decay processes.
In the quark model, the spatial wave functions of the ground state scalar and vector meson are the same \cite{Weng:2019ynva}.
As a rough estimation, we introduce the following approximations to calculate the relative partial decay widths of the $cccc\bar{q}$ and $bbbb\bar{q}$ pentaquark states:
\begin{eqnarray}\label{eq:gamma2}
\gamma_{\Omega_{ccc}D^{*}}&=\gamma_{\Omega_{ccc}D},\quad\gamma_{\Omega_{ccc}D_{s}^{*}}=&\gamma_{\Omega_{ccc}D_{s}},\nonumber\\
\gamma_{\Omega_{bbb}B^{*}}&=\gamma_{\Omega_{bbb}B},\quad\gamma_{\Omega_{bbb}B^{*}_{s}}=&\gamma_{\Omega_{bbb}B_{s}}.
\end{eqnarray}
We present $k\cdot|c_{i}|^{2}$ for each decay process in Table \ref{value-QQQQq}. From Table \ref{value-QQQQq}, one can roughly estimate the relative decay widths between different decay processes of different initial pentaquark states if neglecting the $\gamma_i$ differences.
Such approximation have already been applied in Refs. \cite{Cheng:2019obk,Cheng:2020nho}.
We emphasised that we are only interested in its relative partial decay widths between different decay modes for a particular pentaquark state.
Based on Eqs. (\ref{Eq20})-(\ref{eq:gamma2}), we can calculate relative partial decay widths for the $cccc\bar{q}$ and $bbbb\bar{q}$ pentaquark subsystems.
For the $J^{P}=1/2^{-}$ $cccc\bar{n}$ pentaquark state, it cannot decay into S-wave $\Omega_{ccc} D$ because of the constraint of angular conservation law.
Same situation also happen in the $cccc\bar{s}$, $bbbb\bar{n}$, and $bbbb\bar{s}$ subsystems.
Due to small phase spaces, the $\rm P_{c^4\bar{n}}(6761,1/2,3/2^-)$ and $\rm P_{c^4\bar{s}}(6864,0,3/2^-)$ can only decay into $\Omega_{ccc}D$ and $\Omega_{ccc}D_s$ final states, respectively.
While for the two $J^{P}=3/2^{-}$ $bbbb\bar{q}$ pentaquark states, the $\rm P_{b^{4}\bar{n}}(19647,1/2,3/2^{-})$ and $\rm P_{b^{4}\bar{s}}(19736,0,3/2^{-})$, we find
\begin{equation}
\frac{\Gamma[P_{(b^{4}\bar{n})}(19647, 1/2, 3/2^{-})\rightarrow \Omega_{bbb}B]}{\Gamma[P_{(b^{4}\bar{n})}(19647, 1/2, 3/2^{-})\rightarrow \Omega_{bbb}B^{*}]}=1.3,
\end{equation}
and
\begin{equation}
\frac{\Gamma[P_{(b^{4}\bar{s})}(19736, 0, 3/2^{-})\rightarrow \Omega_{bbb}B_{s}]}{\Gamma[P_{(b^{4}\bar{s})}(19736, 0, 3/2^{-})\rightarrow \Omega_{bbb}B^{*}_{s}]}=1.4.
\end{equation}
respectively.
Thus the widths of the two modes do no differ very much.
\subsubsection{The $cccb\bar{q}$ and $bbbc\bar{q}$ pentaquark states}
Next we consider the $cccb\bar{q}$ and $bbbc\bar{q}$ pentaquark subsystems. The $cccb\bar{q}$ and $bbbc\bar{q}$ pentaquark subsystems include three identical heavy quarks.
The masses of $cccb\bar{q}$ and $bbbc\bar{q}$ pentaquark states can be determined in two schemes and shown in Table \ref{mass-QQQQq}.
In the reference mass scheme, we can exhaust two types of baryon-meson reference systems.
Specifically, we can use the $\Omega_{ccc}$+$B$ ($B_s$) and $\Omega_{ccb}$+$D$ ($D_s$) as reference systems to estimate the masses of the $cccb\bar{n}$ $(cccb\bar{s})$ pentaquark subsystem.
Similarly, the meson-baryon reference systems $\Omega_{bbb}$+$D$ ($D_s$) and $\Omega_{bbc}$+$B$ ($B_s$) are used to calculate the masses of the $bbbc\bar{n}$ $(bbbc\bar{s})$ pentaquark subsystem.
The obtained eigenvalues and masses of $cccb\bar{n}$ $(\bar{s})$ and $bbbc\bar{n}$ $(\bar{s})$ pentaquark states calculated from two types of reference systems are presented in the third and fourth columns of Table \ref{mass-QQQQq}, respectively.
We easily find that the mass spectra came from two different reference systems differ by more than 100 MeV for some studied subsystems.
The reason is that the dynamics and contributions from other terms in conventional meson and baryon potential are not elaborately considered in this model \cite{Zhou:2018pcv}.
However, the mass gaps under different reference systems are still same.
Thus, if one pentaquark state were observed, its partner states may be searched for with the relative positions presented in Table \ref{mass-QQQQq}.
Based on the results listed in the last column of Table \ref{mass-QQQQq}, we plot the mass spectra and relevant quark rearrangement decay patterns for the $cccb\bar{n}$, $cccb\bar{s}$, $bbbc\bar{n}$, and $bbbc\bar{s}$ subsystems in Fig. \ref{fig-QQQQq} (e)-(h), respectively.
Moreover, according to the modified CMI model, we can obtain the overlaps for $cccb\bar{n}$ $(cccb\bar{s})$ and $bbbc\bar{n}$ $(bbbc\bar{s})$ pentaquark states with different baryon $\otimes$ meson bases, and the results are shown in Table \ref{eigenvector-QQQQq}.
From Table \ref{eigenvector-QQQQq},
the $\rm P_{c^{3}b\bar{n}}(10110,1/2,5/2^{-})$ state couples completely to the $\Omega_{ccc}\bar{B}^{*}$ system , which can
be written as a direct product of a baryon $\Omega_{ccc}$ and a meson $\bar{B}^{*}$.
Moreover, for the $\rm P_{c^{3}b\bar{n}}(10118, 1/2, 3/2^{-})$, $\rm P_{c^{3}b\bar{n}}(10078, 1/2, 3/2^{-})$, and $\rm P_{c^{3}b\bar{n}}(10134, 1/2, 1/2^{-})$ states, they strongly couple to the $\Omega_{ccc}\bar{B}^{*}$, $\Omega_{ccc}\bar{B}$, and $\Omega_{ccc}\bar{B}^{*}$ bases, respectively.
This kind of pentaquark behaves similar to the ordinary scattering state made of a baryon and meson if the inner interaction is not strong, but could also be a resonance or bound state dynamically generated by the baryon and meson with strong interaction.
These kinds of pentaquarks deserve a more careful study with some hadron-hadron interaction models in future.
Thus, we label them with $``\star"$ in Tables \ref{eigenvector-QQQQq}, \ref{value-QQQQq}, and Fig. \ref{fig-QQQQq}.
Moreover, we find that the $J^{P}=5/2^{-}$ $QQQQ^{\prime}\bar{q}$ pentaquark states all have only one component $\Omega_{QQQ}B^{*}_{(s)}$ $(D^{*}_{(s)})$.
Therefore, the $J^{P}=5/2^{-}$ ground states are regarded as the states made of two hadrons.
For $cccb\bar{q}$ pentaquark states, they have two types of decay modes: $ccc-b\bar{q}$ and $ccb-c\bar{q}$.
Similarly, the $bbbc\bar{q}$ pentaquark states also have two types of decay modes: $bbb-c\bar{q}$ and $bbc-b\bar{q}$.
In the heavy quark limit, $\Omega^{*}_{ccb}$ ($\Omega^{*}_{bbc}$) and $\Omega_{ccb}$ ($\Omega_{bbc}$) have the same spatial wave function.
Thus, for a $cccb\bar{q}$ or $bbbc\bar{q}$ pentaquark state,
we use the following approximations
\begin{eqnarray}\label{eq:gamma3}
\gamma_{\Omega^{*}_{ccb}D^{*}}&=\gamma_{\Omega^{*}_{ccb}D}=\gamma_{\Omega_{ccb}D^{*}}=&\gamma_{\Omega_{ccb}D},\gamma_{\Omega_{ccc}B^{*}}=\gamma_{\Omega_{ccc}B},
\nonumber\\
\gamma_{\Omega^{*}_{ccb}D^{*}_{s}}&=\gamma_{\Omega^{*}_{ccb}D_{s}}=\gamma_{\Omega_{ccb}D^{*}_{s}}=&\gamma_{\Omega_{ccb}D_{s}},\gamma_{\Omega_{ccc}B^{*}_{s}}=\gamma_{\Omega_{ccc}B_{s}},
\nonumber\\
\gamma_{\Omega^{*}_{bbc}B^{*}}&=\gamma_{\Omega^{*}_{bbc}B}=\gamma_{\Omega_{bbc}B^{*}}=&\gamma_{\Omega_{bbc}B},
\gamma_{\Omega_{bbb}D^{*}}=\gamma_{\Omega_{bbb}D},\nonumber\\
\gamma_{\Omega^{*}_{bbc}B^{*}_{s}}&=\gamma_{\Omega^{*}_{bbc}B_{s}}=\gamma_{\Omega_{bbc}B^{*}_{s}}=&\gamma_{\Omega_{bbc}B_{s}}, \gamma_{\Omega_{bbb}D^{*}_{s}}=\gamma_{\Omega_{bbb}D_{s}}.\nonumber\\
\end{eqnarray}
Based on Table \ref{eigenvector-QQQQq}, we obtain $k\cdot|c_{i}|^{2}$ for each $cccb\bar{q}$ and $bbbc\bar{q}$ pentaquark state and present them in Table \ref{value-QQQQq}.
As an example, we only concentrate on the $cccb\bar{n}$ pentaquark subsystem. According to Table \ref{value-QQQQq}, we can see that the $\rm P_{c^{3}b\bar{n}}(9961, 1/2, 3/2^{-})$ and $\rm P_{c^{3}b\bar{n}}(9946, 1/2, 1/2^{-})$ states only decay into $\Omega^{*}_{ccb}D$ and $\Omega_{ccb}D$ final states, respectively.
For the $\rm P_{c^{3}b\bar{n}}(10062, 1/2, 1/2^{-})$ state, we have its relative partial decay width ratios as
\begin{equation}
\Gamma_{\Omega^{*}_{ccb}D^{*}}:\Gamma_{\Omega_{ccb}D^{*}}:\Gamma_{\Omega_{ccb}D}=24:24:1,
\end{equation}
which suggests that the partial decay width of the $\Omega^{*}_{ccb}D^{*}$ channel is nearly equal to that of the $\Omega_{ccb}D^{*}$ channel.
Note that if a state would be observed in the decay pattern $\Omega^{*}_{ccb}D^{*}$, $\Omega^{*}_{ccb}D$, $\Omega^{*}_{ccb}D^{*}$, and $\Omega_{ccb}D$ , it is a good candidate of a $cccb\bar{n}$ pentaquark state.
\subsubsection{The $ccbb\bar{q}$ pentaquark states}
The last group of the $QQQQ\bar{q}$ system is the pentaquark states with the $ccbb\bar{q}$ configuration.
The $ccbb\bar{q}$ pentaquark states have two pairs of identical heavy quarks, the $cc$ pair and $bb$ pair.
When we construct the wave functions of $ccbb\bar{q}$ pentaquark states, the Pauli principle should be satisfied simultaneously for these two pairs of heavy quarks.
In the reference mass scheme, there are also two types of meson-baryon reference systems for the $ccbb\bar{n}$ ($ccbb\bar{s}$) pentaquark subsystem, i.e., the $\Omega_{ccb}$+$B$ and $\Omega_{cbb}$+$D$ ($\Omega_{ccb}$+$B_s$ and $\Omega_{cbb}$+$D_s$).
Based on the results obtained from the modified CMI model in Table \ref{mass-QQQQq}, we plot the mass spectra and possible decay patterns via rearrangement of constituent quarks in $ccbb\bar{n}$ and $ccbb\bar{s}$ pentaquark states in Fig. \ref{fig-QQQQq} (i)-(j).
According to Fig. \ref{fig-QQQQq} (i)-(j),
we find that all $ccbb\bar{n}$ and $ccbb\bar{s}$ pentaquark states have strong decay channels.
i.e., from the modified CMI model analysis, there is no stable pentaquark state in $ccbb\bar{n}$ and $ccbb\bar{s}$ pentaquark subsystems.
To calculate the strong decay widths of the $ccbb\bar{n}$ and $ccbb\bar{s}$ pentaquark subsystems, we can use the following approximations
\begin{eqnarray}\label{eq:gamma4}
\gamma_{\Omega^{*}_{cbb}D^{*}}&=\gamma_{\Omega^{*}_{cbb}D}=\gamma_{\Omega_{cbb}D^{*}}=&\gamma_{\Omega_{cbb}D},\nonumber\\
\gamma_{\Omega^{*}_{ccb}B^{*}}&=\gamma_{\Omega^{*}_{ccb}B}=\gamma_{\Omega_{ccb}B^{*}}=&\gamma_{\Omega_{ccb}B},\nonumber\\
\gamma_{\Omega^{*}_{cbb}D_{s}^{*}}&=\gamma_{\Omega^{*}_{cbb}D_{s}}=\gamma_{\Omega_{cbb}D_{s}^{*}}=&\gamma_{\Omega_{cbb}D_{s}},\nonumber\\
\gamma_{\Omega^{*}_{ccb}B_{s}^{*}}&=\gamma_{\Omega^{*}_{ccb}B_{s}}=\gamma_{\Omega_{ccb}B_{s}^{*}}=&\gamma_{\Omega_{ccb}B_{s}}.
\end{eqnarray}
By introducing the above relations, $k\cdot|c_{i}|^{2}$ for $ccbb\bar{n}$ ($ccbb\bar{s}$) pentaquark states can be obtained and we present them in Table \ref{value-QQQQq}.
To discuss the strong decay behaviors of the $ccbb\bar{n}$ $(\bar{s})$ pentaquark states, we mainly focus on the relative partial decay widths of the $ccbb\bar{n}$ subsystem, and the $ccbb\bar{s}$ subsystem can be analyzed in a similar way.
From Table \ref{value-QQQQq}, the $J^{P}=5/2^{-}$, the lowest $J^{P}=3/2^{-}$, and the lowest $J^{P}=1/2^{-}$ states can only decay into $\Omega^{*}_{cbb}D^{*}$, $\Omega^{*}_{cbb}D$, and $\Omega_{cbb}D$, respectively.
The most important decay channel for the $\rm P_{c^{2}b^{2}\bar{n}}(13383, 1/2, 3/2^{-})$ is $\Omega^{*}_{bbc}D$ channel in $bbc-c\bar{n}$ decay mode.
The highest $J^P=3/2^-$ state $\rm P_{c^{2}b^{2}\bar{n}}(13414, 1/2, 1/2^{-})$ has many different decay channels and this state is expected to be broad.
For the $\rm P_{c^{2}b^{2}\bar{n}}(13383, 1/2, 3/2^{-})$, we have
\begin{equation}
\gamma_{\Omega^{*}_{cbb}D^{*}}:\gamma_{\Omega^{*}_{cbb}D}:\gamma_{\Omega_{cbb}D^{*}}=1.5:15:1,
\end{equation}
and
\begin{equation}
\gamma_{\Omega^{*}_{ccb}B^{*}}:\gamma_{\Omega^{*}_{ccb}B}:\gamma_{\Omega_{ccb}B^{*}}=1.5:1:1.4.
\end{equation}
For the $\rm P_{c^{2}b^{2}\bar{n}}(13414, 1/2, 1/2^{-})$, we have
\begin{equation}
\gamma_{\Omega^{*}_{cbb}D^{*}}:\gamma_{\Omega_{cbb}D^{*}}:\gamma_{\Omega_{cbb}D}=12:3.7:1,
\end{equation}
and
\begin{equation}
\gamma_{\Omega^{*}_{ccb}B^{*}}:\gamma_{\Omega_{ccb}B^{*}}:\gamma_{\Omega_{ccb}B}=3.5:0.4:1.
\end{equation}
Obviously, its dominant decay modes in $cbb-c\bar{n}$ and $ccb-b\bar{n}$ sectors are $\Omega^{*}_{cbb}D^{*}$ and $\Omega^{*}_{ccb}B^{*}$ channels, respectively.
Other three $J^{P}=3/2^{-}$ and three $J^{P}=1/2^{-}$ states only have $cbb-c\bar{n}$ decay mode.
The $ccb-b\bar{n}$ decay mode are strongly suppressed by the corresponding phase space.
\subsubsection{Comparison of other pentaquark systems}
In 2020, the LHCb collaboration studied the invariant mass spectrum of $J/\psi$ pairs, and they reported a narrow structure around 6.9 GeV \cite{Aaij:2020fnh}. Take this as an opportunity, the heavy flavored pentaquarks with four heavy quarks ($QQQQ\bar{q}$) and fully heavy pentaquarks ($QQQQ\bar{Q}$) are systematically discussed in this work and Ref. \cite{An:2020jix} within the modify CMI model. Here we can compare the differences between these two pentaquark systems.
Firstly, we discuss the mass differences among the ground states of $QQQQ\bar{n}$, $QQQQ\bar{s}$, $QQQQ\bar{c}$, and $QQQQ\bar{b}$ with the same $J^P$. From Table \ref{eigenvector-QQQQq} and Table IV of
Ref. \cite{An:2020jix}, the masses of $cccc\bar{n}$, $cccc\bar{s}$, $cccc\bar{c}$, and $cccc\bar{b}$ ground states with $J^{P}=3/2^{-}$ are 6761, 6864, 7864, and 11130 MeV, respectively.
Moreover, relative to the the $J^{P}=1/2^{-}$ states, their corresponding mass gaps are 106 MeV, 108 MeV, 85 MeV, and 47 MeV, respectively.
Other subsystems also have similar situations.
Thus, compared with the $QQQQ\bar{Q}$ system, the $QQQQ\bar{q}$ system has lighter masses and bigger mass gaps when a heavy antiquark is replaced by a light antiquark because $v_{ij} \propto 1/{m_{i}m_{j}}$.
Secondly, we study the relations between the pentaquark and their corresponding baryon-meson channels. The fully heavy pentaquarks $QQQQ\bar{Q}$ are more likely below all possible strong-decays channels and thus more stable compared to the $QQQQ\bar{q}$ systems. We have found two relatively stable states $\rm P_{c^{2}b^{2}\bar{b}}(17416, 0, 3/2^{-})$ and $\rm P_{c^{2}b^{2}\bar{b}}(17477, 0, 5/2^{-})$, which are below all allowed rearrangement decay channels in $QQQQ\bar{Q}$ system. However, we do not find any stable state for the $QQQQ\bar{q}$ multiquark systems. When both $QQQQ\bar{Q}$ and $QQQQ\bar{q}$ ground states are above their corresponding baryon-meson channel, the $QQQQ\bar{Q}$ mass would be closer to the threshold. For example,
the $cccc\bar{n}$, $cccc\bar{s}$, $cccc\bar{c}$, and $cccc\bar{b}$ states are above the corresponding minimum threshold ($\Omega_{ccc}$+ pseudoscalar meson) 106 MeV, 110 MeV, 94.5 MeV, and 69.5 MeV, respectively.
Unlike fully pentaquark $QQQQ\bar{Q}$, all $QQQQ\bar{q}$ pentaquark states can never mix with a triquark baryon and thus are explicit exotic states. Accurate measurement in future experiment and the comparison may help us understand the $Q\bar{Q}$ annihilation effects in the hadron spectrum.
\section{Summary}\label{sec5}
The observation of the $P_c(4312)$, $P_c(4440)$, and $P_c(4457)$ states achieved by the LHCb Collaboration and the study of the possible stable $QQ\bar{q}\bar{q}$ tetraquark states give us strong confidence to study the mass spectra of the $QQQQ\bar{q}$ pentaquark system within the framework of CMI model.
Similar to the fully-heavy $QQ\bar{Q}\bar{Q}$ tetraquark system \cite{Aaij:2020fnh}, the $QQQQ\bar{q}$ system consist of four heavy quarks are dominantly bounded by the gluon exchange interaction, and can hardly be considered as molecular states.
In this work, by including the flavor SU(3) breaking effect, we firstly construct the $\psi_{flavor}\otimes\psi_{color}\otimes\psi_{spin}$ wave functions based on the SU(2) and SU(3) symmetry and Pauli Principle.
Then we extract the effective coupling constants from the conventional hadrons. After that, we systematically calculate the chromomagnetic interaction Hamiltonian for the discussed pentaquark states and obtain the corresponding mass spectra.
As a modification to the CMI model, the effect of chromoelectric interaction is added in the modified CMI model.
So, we mainly discussed the results of mass spectra for the $QQQQ\bar{q}$ pentaquark system obtained from the modified CMI model. The results from the reference mass scheme are presented for comparison.
In addition to the eigenvalues, we also provide the eigenvectors to extract useful information about the decay properties for the studied pentaquark systems.
Finally, we analyze the stability, possible quark rearrangement decay channels and relative partial decay widths for all the $QQQQ\bar{q}$ pentaquark states.
Due to the constraint from Pauli principle, there is no ground $J^{P}=5/2^{-}$ $cccc\bar{q}$ and $bbbb\bar{q}$ pentaquark states.
From the obtained Tables and Figs for the $QQQQ\bar{q}$ pentaquark system, we find no stable candidate
in the modified CMI model.
However, due to the uncertainty of the modified CMI model, some of them may not truly be unstable states, and further dynamical calculations may help us to clarify their natures.
Especially, for some unstable states which are a little higher than the meson-baryon thresholds of lowest strong decay channels, they can be considered as narrow pentaquark states, and have opportunities to be found in future experiment.
Meanwhile, the whole mass spectra has a slight shift or down due to the mass deviation of constituent quarks.
While the mass gaps between different pentaquark states are relatively stable, if one pentaquark states are observed in experiment, we can use these mass gaps to predict their corresponding multiplets.
Among the studied $QQQQ\bar{q}$ pentaquark states, all of them are explicit exotic states. If such pentaquark states
are observed, their exotic nature can be easily identified.
However, up to now, none of $QQQQ\bar{q}$ pentaquark states is found.
More detailed dynamical investigations on these pentaquark systems are still needed.
Producing a $QQQQ\bar{q}$ pentaquark state seems to be a difficult task in experiment.
Our systematical study may provide theorists and experimentalists some preliminary understanding toward this pentaquark system.
We hope that the present study may inspire experimentalists and theorists to pay attention to this kind of pentaquark system.
\section*{ACKNOWLEDGMENTS}
This work is supported by the China National Funds for Distinguished Young Scientists under Grant No. 11825503, National Key Research and Development Program of China under Contract No. 2020YFA0406400, the 111 Project under Grant No. B20063, and the National Natural Science Foundation of China under Grant No. 12047501. This project is also supported by the National Natural Science Foundation of China under Grants No. 11705072, 11705069, and 11965016, CAS Interdisciplinary Innovation Team.
\begin{widetext}
|
\section{Introduction}
Triangle counting is a fundamental problem in network analysis. There has been a rich line of work on counting triangles in graphs~\cite{itai1978finding,alon1997finding,latapy2008main,seshadhri2013triadic,al2018triangle}. The triangle counts appear in form of different parameters such as \emph{clustering coefficient}~\cite{watts1998collective} and \emph{transitivity ratios}~\cite{wasserman1994social}. Triangle counting has many applications such as social networks analysis~\cite{pfeiffer2012fast}, indexing graph databases~\cite{khan2011neighborhood}, community discovery~\cite{palla2005uncovering}, and spam detection~\cite{becchetti2008efficient}.
Much of the rich history of triangle counting has focused on {static} graphs. But many real-world networks, such as communication networks, message networks, and social interaction networks are fundamentally \emph{temporal}. Every edge has an associated timestamp~\cite{kovanen2011temporal, gaumont2016finding, farajtabar2018coevolve}. We can model these attributed networks as \emph{temporal networks} where edges have timestamp. For example, in cryptocurrency transaction networks and email networks, each link is between a sender and a receiver and has a timestamp that could be represented as a directed edge with a timestamp in a temporal network.
\emph{Temporal triangle counts} provide a far richer set of counts than standard counts. These counts take into account the temporal ordering of edges in a triangle, and potentially impose constraints on the timestamp difference among edges. Temporal triangle and motif counting has applications in graph representation learning~\cite{tu2019gl2vec}, expressivity of graph neural networks (GNNs)~\cite{bouritsas2020improving}, network classification~\cite{tu2018network}, temporal text network analysis~\cite{vega2018foundations}, computer networks~\cite{valverde2005network}, and brain networks~\cite{dimitriadis2010tracking}.
Recently, there has been significant interest in temporal triangle and motif counting algorithms~\cite{mackey2018chronological,liu2018sampling,tu2018network,petrovic2019counting,sun2019new,tu2019gl2vec,boekhout2019efficiently,bouritsas2020improving,wang2020efficient}.
Counting temporal triangles in (directed) temporal networks introduces new challenges to that of triangle counting in static graphs. The first challenge is actually defining types of temporal triangles (or motifs). In essence, all definitions specify constraints on the time difference between edges of a triangle.
For example, Kovanen~{et al.}\xspace~\cite{kovanen2011temporal} restrict temporal triangles to cases where the gap between two consecutive edges in the temporal ordering is at most $\Delta$ time unit, and the two edges incident to each node are consecutive event of that node.
Paranjape-Benson-Leskovec (henceforth PBL) introduce \emph{$\delta$-temporal triangles}, where all edges of the triangle/motif have to occur within $\delta$ timesteps~\cite{paranjape2017motifs}. These varying definitions necessitate different algorithms. Our first motivating question is whether one can design algorithms for a more general notion of temporal triangles.
Secondly, there is a significant gap between the best static triangle counting algorithms and temporal triangle counters. Specifically, the classic and immensely practical triangle counting algorithm of Chiba-Nishizeki runs in time $O(m\kappa)$, where $m$ is the number of edges and $\kappa$ is the \emph{graph degeneracy} (or max core number). The current state-of-the-art temporal triangle counting algorithm of PBL runs in $O(m\sqrt{\tau})$ time, where $\tau$ is the total triangle count (of the underlying static graph). There is a large gap between $\kappa$ (which is typically in the hundreds and thought of as a constant) and $\tau$ (which is superlinear in $m$).
These twin issues motivate our study. \emph{Can we define a more general notion of temporal triangles, and give an algorithm whose asymptotic running time is closer to that of static triangle counting?}
\begin{figure*}[th]
\centering
\begin{subfigure}[b]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{Figures/speedup.pdf}
\caption{Speedup}
\label{fig:speedup}
\end{subfigure}
\begin{subfigure}[b]{0.33\textwidth}
\centering
\includegraphics[width=\columnwidth]{Figures/delta_vs_deltas.pdf}
\caption{Expressivity}
\label{fig:delta_vs_deltas}
\end{subfigure}
\begin{subfigure}[b]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{Figures/third_edge.pdf}
\caption{Triadic closure over time}
\label{fig:third_edge}
\end{subfigure}
\caption{(a): The Speedup of \textsc{DOTTT}{} for counting $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangles over the PBL\space algorithm for counting $\delta_{1,3}$-temporal triangles. (b): We fix $\delta_{1,3}$ to 1 hr. Blue bars show the ratio of (1 hr, 30 mins, 30 mins)-temporal triangles to (1 hr, 1 hr, 1 hr)-temporal triangle. The red bars illustrate the ratio for the case of (1 hr, 10 mins, 50 mins)-temporal triangles and is more restrictive. (c): We fix $\delta_{1,3}=2$ hrs and $\delta_{1,2}=1$ hr. At $t$ we plot the ratio of (2 hrs, 1 hr, t)-temporal triangles to (2 hrs, 1 hr, 1hr)-temporal triangles.}
\label{fig:count_dist_comp}
\end{figure*}
\subsection{Problem Description}
The input is a directed temporal graph $T = (V,E)$. Each edge is a tuple of the form $(u,v,t)$ where $u$ and $v$ are vertices in the temporal graph, and $t$ is a timestamp.
For notational convenience, we assume all timestamps in a temporal network are unique integers.
We introduce our notion of $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangles.
\begin{definition} \label{def:temp-tri}
Let $e_1 = (u_1,v_1,t_1), e_2= (u_2,v_2,t_2)$, and $e_3 = (u_3, v_3, t_3)$, be three directed temporal edges where the induced static graph on them is a triangle, and $t_1 < t_2 < t_3$.
$(e_1,e_2,e_3)$ is a \emph{$(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangle} if $t_2 - t_1 \leq \delta_{1,2}$, $t_3 - t_2 \leq \delta_{2,3}$, and $t_3 - t_1 \leq \delta_{1,3}$.
\end{definition}
Thus, we specify timestamp differences between \emph{every} pair of edges. When one also considers the direction of edges, there exist eight different types of temporal triangles as shown in~\Fig{temp_tri}. These types are distinguished by temporal ordering of edges and their direction. Thus, for any choice of $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$, there are eight different types of temporal triangles (one corresponding to each figure in~\Fig{temp_tri}).
We observe that the notion in \Def{temp-tri} subsumes most existing temporal triangle definitions. Specifically, a $(\delta_{1,3},\delta_{1,3},\delta_{1,3})$-temporal triangle becomes a $\delta_{1,3}$-temporal triangles as defined in PBL~\cite{paranjape2017motifs}.
Temporal triangles with respect to the temporal motif definition by by Kovanen {et al.}\xspace in~\cite{kovanen2011temporal} consider timestamp differences between consecutive edges in temporal ordering. By our definition, $(2\Delta,\Delta,\Delta)$-temporal triangles capture these types of temporal triangles. Although, the definition in~\cite{kovanen2011temporal} is more restrictive and requires that all edges incident to a node are consecutive events of that node.
Most existing temporal triangle counting literature uses these definitions~\cite{mackey2018chronological,liu2018sampling,sun2019new,wang2020efficient}.
We describe a simple example to see how \Def{temp-tri} offers richer temporal information. Let us measure time in hours, so $(2,1,1)$-temporal triangle is one where the first and second edge (of the triangle) are at most 1 hour apart, and similarly for the second and third edge. Now consider $(1.5, 1, 1)$-temporal triangles. The time gap between the first and second edge (as well as the second and third) is again 1 hour, but the entire triangle must occur within 1.5 hours. There is a significant difference between these cases, but previous definitions of temporal triangles would not distinguish these.
We note that more general temporal motifs, beyond triangles, have been defined. Yet, to the best of our knowledge, most fast algorithms that scale to millions of edges have been designed for triangles. Paranjape~{et al.}\xspace specialized algorithm for 3-edge triangle motifs (temporal triangles) is up to 56x faster than their general motif counting algorithm~\cite{paranjape2017motifs}. Our focus was on scalable algorithms, and hence, on triangle counting. We believe that generalizing \Def{temp-tri} (and our \textsc{DOTTT}{} algorithm) for general motifs would be compelling future work.
\begin{figure}[th]
\centering
\resizebox{\columnwidth}{!}{
\begin{tikzpicture}[nd/.style={scale=1,circle,draw,inner sep=2pt},minimum size = 6pt]
\matrix[column sep=0.8cm, row sep=0.4cm,ampersand replacement=\&]
{
\TriTypeOne \&
\TriTypeTwo \&
\TriTypeThree \&
\TriTypeFour \\
\TriTypeFive \&
\TriTypeSix \&
\TriTypeSeven \&
\TriTypeEight \\
};
\end{tikzpicture}
}
\caption{\label{fig:temp_tri} \small All possible temporal triangle types. The start point of the first edge (in temporal ordering of edges) is shown in red and the end point in green.}
\end{figure}
\subsection{Main Contributions} \label{sec:cont}
Our main result is the \emph{Degeneracy Oriented Temporal Triangle Totaler} algorithm, $\textsc{DOTTT}{}$ that counts $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangles as defined in \Def{temp-tri}. The running time is only a logarithmic overhead over static triangle counting. We detail our contributions below.
{\bf Theoretically bridging gap between temporal and static triangle counting:} Our main theorem is the following.
\begin{theorem}\label{thm:all_runtime}
Given $\delta_{1,3},\delta_{1,2}$ and $\delta_{2,3}$, the \textsc{DOTTT}{} algorithm exactly counts each of the eight types of $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangles (\Fig{temp_tri}) in a temporal graph in $O(m \kappa \log m)$ time. (Here, $m$ is the total number of temporal edges, and $\kappa$ is the degeneracy of the underlying static graph.)
\end{theorem}
Observe that, up to a logarithmic factor, our theoretical running time for temporal triangle counting matches the $O(m\kappa)$ bound for static triangle counting. As mentioned earlier, the previous best bound was $O(m\tau^{1/2})$. We stress that there is no dependence on the time intervals $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$.
The idea of degeneracy orientations is tailored to static graphs, and one of our contributions is to show it can help for temporal triangle counting. A key insight in \textsc{DOTTT}{} is to process (underlying) static edges in the exact order of the Chiba-Nishizeki algorithm, but carefully consider neighboring edges to capture all temporal triangles. By a non-trivial combinatorial analysis, we can prove that number of times that a temporal edge is processed is upper bounded by $\kappa$. We need additional data structure tricks to get the counts efficiently, leading to an extra logarithmic factor.
{\bf Excellent practical behavior of \textsc{DOTTT}:}
\textsc{DOTTT}{} consistently determines temporal triangle counts in less than ten minutes for datasets with tens of millions of edges. We only use a single commodity machine with 64GB memory, without any parallelization. We directly compare \textsc{DOTTT}{} with the state-of-the-art PBL algorithm. Our algorithm is consistently faster, and as illustrated in~\Fig{speedup} we typically get a factor 1.5 speedup for larger graphs. (We note that \textsc{DOTTT}{} can count a more general class of temporal triangles.)
We note that for the largest dataset in our experiments, Bitcoin (515.5M edges), \textsc{DOTTT}{} only uses 64GB memory and runs in less than an hour, while existing methods ran out of memory (details in~\Sec{experimental}).
{\bf Richer triadic information from $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangles:}
We demonstrate how $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangles can give a richer network
analysis method. Consider~\Fig{delta_vs_deltas}. For a collection of temporal datasets, we generate the counts of (1 hr, 30 min, 30 min)-temporal triangle counts, as well as those for (1 hr, 10 min, 50 min)-temporal triangles. We plot these numbers as a ratio of (1 hr, 1 hr, 1 hr)-temporal triangles. Across the datasets, the ratios are at most $75\%$. The red bars are typically at most $25\%$, showing the extra power of \Def{temp-tri} in distinguishing temporal triangles.
We note here that for each dataset, \textsc{DOTTT}{} has the same running time for obtaining the counts for (1 hr, 30 min, 30 min)-temporal triangle and (1 hr, 10 min, 50 min)-temporal triangles, as it has no dependency on the time intervals $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$.
An interesting study is presented in~\Fig{third_edge}. The transitivity and clustering coefficients are fundamental quantities of study in network science. In temporal graphs, in addition to these measure, the time it takes for a wedge (2-path) to close could also be of importance. (Zingnani~{et al.}\xspace proposed the triadic closure delay metric that capture the time delay between when a triadic closure is first possible, and when they occur~\cite{zignani2014link}.)
In \Fig{third_edge}, we fix $\delta_{1,3} = $ 2 hrs and $\delta_{1,2}$ = 1 hr. We then vary $\delta_{2,3}$ from zero to 60 minutes, and plot the ratio of $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangles to (2hrs, 1hr, 1hr)-temporal triangles. We can see the trends in triadic closure with respect to the time for the third edge.
We observe that, by and large, half the triangles are formed within 20 minutes of the first two edges appearing. And by 30 minutes, almost $75\%$ of these triangles are formed. These are examples of triadic analyses enabled by \textsc{DOTTT}.
\subsection{Main challenges} \label{sec:challenge}
In a temporal graph, the number of temporal edges is typically two to three times the number of underlying static edges. Since most triangle counting algorithms are based on some form of wedge enumerations, this leads to a significant increase in the number of edges.
One method used for temporal triangle counting is to simply prune the temporal edges based on the time period~\cite{mackey2018chronological,sun2019tm}. But such algorithms have a dependency on the time period and are inefficient for large
time periods.
Another significant challenge is the multiplicity of an individual edge can be extremely large. The same edge often occurs many hundreds to thousands of times in a temporal network (in the BitCoin network, there is an edge appearing 447K times~\Tab{runtime}). These edges create significant bottlenecks for enumeration methods. It is not clear how efficient methods on the underlying static graphs (which ignores multiplicities) can help with this problem. Triangle counting often works by finding a wedge (2-path) and checking for the third edge. With multiple temporal edges between the same pair of vertices, this method requires many edge lookups. Paranjape~{et al.}\xspace used a clever idea to process edges on a pair of vertices $O(\tau^{1/2})$ times. The challenge is to bound it by the degeneracy of the graph.
The time constraints expressed by $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangles create additional challenges. A clever wedge enumeration exploiting the degeneracy may produce wedges containing the first and second edges of the triangle, the first and third, or the second and third. This makes the lookup (or counting) of possible "matches" for the remaining edge challenging, since it appears we need to look at all multiple edges. On the other hand, if we enumerated wedges that only involved the first and second edge, we cannot benefit for the efficiencies of degeneracy-based methods. Some of these problems can be circumvented for $(\delta,\delta,\delta)$-temporal triangles, but the general case is challenging.
Overall, we can state the main challenge as follows. Fast triangle counting methods (such as degeneracy based methods) necessarily ignore time constraints while generating wedges, making it hard to look for the "closing" edge. On the other hand, a method that exploits the timestamps by (say) pruning cannot get the efficiency gains of degeneracy based methods. One of the insights of \textsc{DOTTT}{} is a resolution of this tension.
\section{Related Work} \label{sec:related}
There is rich history of work on triangle counting in static graphs. Various algorithm for triangle and motif counting in attributed graphs have also been proposed~\cite{WernickeRasche06,pfeiffer2014attributed,ribeiro2014discovering,mongiovi2018glabtrie,gu2018homogeneous,rossi2019heterogeneous}. Here we only focus on temporal networks and refer the reader to~\cite{al2018triangle} and the tutorial~\cite{seshadhri2019scalable} for a more detailed list of related work.
Graph orientation, in particular degeneracy ordering, is a classic idea in counting triangles and motifs in static graphs, pioneered by Chiba-Nizhizeki~\cite{chiba1985arboricity}. Recently, there has been a number of triangle counting and motif counting algorithms inspired by these techniques~\cite{jha2015path, eden2017approximately, PiSeVi17, jain2017fast, ortmann2017efficient, pashanasangi2020efficiently}. The main benefit of degeneracy ordering is that the out-degree of each vertex becomes small when we orient the static graph based on this ordering.
Kovanen~{et al.}\xspace called two temporal edges $\Delta T$-adjacent if they share a vertex and the difference of their timestamps are at most $\Delta T$~\cite{kovanen2011temporal}. In their definition of temporal motifs, temporal edges must represent consecutive events for a node. Redmond~{et al.}\xspace gave an algorithm for counting $\delta$-temporal motifs but their algorithm does not take the temporal ordering of edges into account~\cite{redmond2013temporal}, and only counts motifs where incoming edges occur before outgoing edges. Gurukar~{et al.}\xspace present a heuristic for counting temporal motifs~\cite{gurukar2015commit}.
More related to our work, Paranjape-Benson-Leskovec defined the $\delta$-temporal motifs where all edges occur inside a time period $\delta$ and also the temporal ordering of edges are taken into account~\cite{paranjape2017motifs}. They gave a general algorithm for counting $k$-node $\ell$-edge motifs in temporal networks. The main idea behind their algorithm is a moving time window of size $\delta$ over the sequence of all temporal edges for each static motif matching the underlying static motif of the temporal motif of interest. For temporal triangles, their algorithm runs in $O(\tau m)$ time where $\tau$ is the number of triangles in the underlying static graph of the input temporal graph $T$, as it might enumerate temporal edges on static edges with high multiplicity $O(\tau)$ time. They also presented a specialized, more efficient algorithm for counting 3-edge temporal triangles that runs in time $O(\tau^{1/2} m)$. We call their algorithm $PBL\space$ and use it as our baseline.
Mackey~{et al.}\xspace presented a backtracking algorithm for counting $\delta$-temporal motifs that maps edges of the motif to the edges of the host graph one by one in temporal (chronological) ordering. For each edge, it only searches through edges that occur in the correct temporal ordering and respect the time gap restriction.~\cite{mackey2018chronological}. Unlike the PBL\space algortihm and ours, this algorithm could be inefficient for large values of $\delta$ as its runtime depends on the value of $\delta$.
Liu~{et al.}\xspace~\cite{liu2021temporal} introduced a comparative survey of temporal motif models. Boekhout~{et al.}\xspace gave an algorithm for counting $\delta$-temporal multi-layer temporal motifs~\cite{boekhout2019efficiently}. Li~{et al.}\xspace, developed an algorithm for counting temporal motifs in heterogeneous information networks~\cite{li2018temporal}. Petrovic~{et al.}\xspace gave an algorithm for counting causal paths in time series data on networks~\cite{petrovic2019counting}.
There has also been recent progress on approximating the counts of temporal motifs and triangles~\cite{sun2019new, wang2020efficient}. Particularly , Liu~{et al.}\xspace presented a sampling framework for approximating the counts of $\delta_{1,3}$-temporal motifs~\cite{liu2018sampling}.
\section{Preliminaries}
The input graph is a directed temporal graph that we denote by $T(V,E)$. Let $|V|=n$ and $|E|=m$. Temporal graph $T$ is presented as a collection of $m$ directed temporal edges $e = (u,v,t)$ where $u,v \in V$, and $t$ is the timestamp for edge $e$ where $t\in \mathbb{R}$. We use $t(e)$ to denote the timestamp of a temporal edge $e$. Note that there could be multiple temporal edges on the same pair of nodes. We assume that all the timestamps in $T$ are unique. This assumption leads to the clean definition of different types of temporal triangles (~\Fig{temp_tri}), but is not a necessity of our algorithm. To be more specific, our algorithm also works for temporal graph including temporal edges with equal timestamp.
We denote the underlying undirected static graph of $T$ as $G=(V,E_s)$ and put $|E_s| = m_s$. Two vertices in the static graph $G$ are connected if there is at least one temporal edge between them. Formally, $E_s = \{\{u,v\} \mid \exists t: (u,v,t) \in E \vee (v,u,t) \in E\}$. For $v_1,v_2 \in V$, let $\sigma((v_1,v_2))$ denote the temporal multiplicity, that is the number of temporal edges on $\{v_1,v_2\}$ directed from $v_1$ to $v_2$.
As shown in~\Fig{temp_tri}, there are eight different types of temporal triangles. The time restrictions $\delta_{1,3}$, $\delta_{1,2}$, and $\delta_{2,3}$ is not involved in definition of these types and could be applied to each of them. Note that these different types account for all possible ordering of temporal edges in the triangle in addition to their directions.
In a directed graph $G$, we use $N^+(v):\{(v,u) \in E(G)\}$ to denote the out-neighborhood and $N^-(v):\{(u,v) \in E(G)\})$ to denote the in-neighborhood of a vertex $v$. For a vertex $v\in V(G)$, we define out-degree as $d^+(v) = |N^+(v)|$ and in-degree as $d^-(v) = |N^-(v)|$.
Vertex ordering is a central idea in triangle counting and motif analysis in general~\cite{chiba1985arboricity,PiSeVi17,jain2017fast,ortmann2017efficient,turk2019revisiting,pashanasangi2020efficiently,bera2020linear}. Let $G$ be an undirected static simple graph. Given any ordering $\prec$ of $V(G)$, we can construct a DAG $G_\prec$ by orienting each edge $\{u,v\} \in E(G)$ from $u$ to $v$ iff $u \prec v$.
Next we define \emph{degeneracy} and \emph{degeneracy ordering} formally.
\begin{definition}
The degeneracy of a graph $G$, denoted by $\kappa$, is the smallest integer $k$ such that there exists an ordering $\prec$ of $V(G)$ where $d^+_v(G_\prec) \leq k$ for each $v \in V(G)$.
\end{definition}
\begin{definition}
Degeneracy ordering of an undirected simple static graph $G$ is obtained by removing a vertex with minimum degree repeatedly. The order of the removal of vertices is the degeneracy ordering of $G$.
\end{definition}
There is an algorithm for finding the degeneracy ordering of a graph $G$ in $O(|E(G)|)$ time~\cite{matula1983smallest}. Let $\prec$ denote the degeneracy ordering.
\section{Main Ideas}
Our algorithm first enumerates static triangles in $G$, the underlying static graph of the input temporal graph $T$. Let $\{u,v,w\}$ be a static triangle. We consider all possible temporal orderings as shown in~\Fig{temp_order}, and all possible orientations as shown in~\Fig{temp_dir}, for a temporal triangle corresponding to $\{u,v,w\}$. A temporal ordering and a temporal orientation together determine the type of the temporal triangle. For example $\pi_1$ and $\rho_8$ correspond to $\mcT_1$.~\Tab{tri_type_conversion} lists all possible pairs of temporal ordering and orientations and their corresponding type of temporal triangle as a function $\psi$.
We store the input temporal edges of the input temporal graph $T$ in a data structure in the CSR format. Thus, we can assume that we have constant time access to temporal edges on each pair of vertices for each direction in the order of increasing timestamps. Let $\pi$ denote the temporal ordering, and $\rho$ denote the orientation for which we want to count the temporal triangles. In this section, from here we only consider temporal edges that follow the orientation $\rho$.
Assume that the timestamps of two of the edges of a temporal triangle corresponding to the static triangle $\{u,v,w\}$ is given. WLOG, assume that these temporal edges correspond to $\{u,v\}$ and $\{u,w\}$. We can use a binary search to find the number of temporal edges on the pair $\{v,w\}$ that are compatible with the two given temporal edges. Note that compatibility of timestamps is determined by the timestamp of edges and the temporal ordering $\pi$. Thus, all we need is to enumerate all possible pairs of temporal edges on $\{u,v\}$ and $\{u,w\}$. This could be an expensive enumeration if both these static edges have high multiplicity of temporal edges.
We show that we can obtain the counts of temporal triangles on $\{u,v,w\}$ without enumeration of all possible pairs of temporal edges on $\{u,v\}$ and $\{u,w\}$. Let $e_1,\ldots,e_{\sigma(\rho(\{u,v\}))}$ denote the sequence of temporal edges in the order of increasing timestamp on $\{u,v\}$, and $e_1^\prime,\ldots,e_{\sigma(\rho(\{u,w\}))}^\prime$ denote that of $\{u,w\}$. We enumerate each of these sequences of temporal edges separately, and for each edge we store cumulative counts of compatible temporal edge on $\{v,w\}$. In other words, for each edge $e$ in these two sequence, we store the counts of edges $e_3$ on $\{v,w\}$ that are compatible with $e$ or any other temporal edge in the same sequence with a smaller timestamp.
Then we enumerate the temporal edges on $\{u,v\}$, and for each edge $e_i$ we use binary search to find the sequence of temporal edges $e_j^\prime,\ldots,e_k^\prime$, in increasing order of timestamp, on $\{v,w\}$ that are compatible with $e_i$. We use the cumulative counts of compatible edges on $\{v,w\}$ that we stored for $e_i$, $e_k^\prime$, and $e_j^\prime$ to compute the counts of all temporal triangles on $\{u,v,w\}$ that include $e_i$.
Although we avoid the enumeration of pairs of temporal edges on $\{u,v\}$ and $\{u,w\}$, our algorithm could still be inefficient. The reason is that static edges $\{u,v\}$ could have high multiplicity of temporal edges and also participate in a large number of static triangles. Here is where we use the power of vertex ordering and graph orientation techniques.
\textsc{DOTTT}{} enumerates static triangles in $G_\prec$, and when processing a static triangle $\{u,v,w\}$ where $u$ comes first in the degeneracy ordering $\prec$ of $G$, it only enumerates temporal edges on $\{u,v\}$ and $\{u,w\}$. Thus, each temporal edge on a pair $\{x,y\}$ where $x \prec y$, is processed only for static triangles where the third vertex is in the out-neighborhood of $x$. But we know that the out-degree of each vertex is bounded by $\kappa$ in $G_\prec$. Therefore, each such temporal edge on $\{x,y\}$ is processed $O(\kappa)$ times.
\begin{figure}[t]
\centering
\resizebox{0.9\columnwidth}{!}{
\begin{tikzpicture}[nd/.style={scale=1,circle,draw,inner sep=2pt},minimum size = 6pt]
\matrix[column sep=1.0cm, row sep=0.4cm,ampersand replacement=\&]
{
\OrderOne \&
\OrderTwo \&
\OrderThree \\
\OrderFour \&
\OrderFive \&
\OrderSix \\
};
\end{tikzpicture}
}
\caption{\label{fig:temp_order} All possible ordering of temporal edges of a temporal triangle corresponding to a static triangle $\{u,v,w\}$.}
\end{figure}
\begin{figure}[t]
\centering
\resizebox{\columnwidth}{!}{
\begin{tikzpicture}[nd/.style={scale=1,circle,draw,inner sep=2pt},minimum size = 6pt]
\matrix[column sep=0.4cm, row sep=0.4cm,ampersand replacement=\&]
{
\DirOne \&
\DirTwo \&
\DirThree \&
\DirFour \\
\DirFive \&
\DirSix \&
\DirSeven \&
\DirEight \\
};
\end{tikzpicture}
}
\caption{\label{fig:temp_dir} All possible orientations of temporal edges of a temporal triangle corresponding to a static triangle $\{u,v,w\}$.}
\end{figure}
\begin{table}[t]
\caption{Conversion from temporal ordering and orientation to temporal triangle type.}
\begin{center}
\begin{tabular}{c|cccccccc}
$\psi$ & $\rho_1$ & $\rho_2$ & $\rho_3$ & $\rho_4$ & $\rho_5$ & $\rho_6$ & $\rho_7$ & $\rho_8$\\\hline
$\pi_1$ & $\mcT_7$ & $\mcT_6$ & $\mcT_5$ & $\mcT_8$ & $\mcT_3$ & $\mcT_2$ & $\mcT_4$ & $\mcT_1$ \\
$\pi_2$ & $\mcT_5$ & $\mcT_3$ & $\mcT_7$ & $\mcT_4$ & $\mcT_6$ & $\mcT_1$ & $\mcT_8$ & $\mcT_2$ \\
$\pi_3$ & $\mcT_3$ & $\mcT_2$ & $\mcT_1$ & $\mcT_4$ & $\mcT_7$ & $\mcT_6$ & $\mcT_8$ & $\mcT_5$ \\
$\pi_4$ & $\mcT_1$ & $\mcT_7$ & $\mcT_3$ & $\mcT_8$ & $\mcT_2$ & $\mcT_5$ & $\mcT_4$ & $\mcT_6$ \\
$\pi_5$ & $\mcT_6$ & $\mcT_1$ & $\mcT_2$ & $\mcT_8$ & $\mcT_5$ & $\mcT_3$ & $\mcT_4$ & $\mcT_7$ \\
$\pi_6$ & $\mcT_2$ & $\mcT_5$ & $\mcT_6$ & $\mcT_4$ & $\mcT_1$ & $\mcT_7$ & $\mcT_8$ & $\mcT_3$ \\
\end{tabular}
\end{center}
\label{tab:tri_type_conversion}
\end{table}
\section{Our Main Algorithm}\label{sec:mainalgorithm}
In this section we describe our algorithm for getting $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangles counts. Let $T=(V,E)$ be the input directed temporal graph given as a list of temporal edges sorted by timestamps. Although not necessary for our algorithm, assuming that edges are given in increasing order of timestamp is common in temporal networks as the edges are recorded in their order of occurrence~\cite{paranjape2017motifs}.
We first extract the static graph $G(V,E_s)$ from $T$. Then, we obtain the degeneracy ordering of $G$, denoted by $\prec$ using the algorithm by Matula and Beck~\cite{matula1983smallest}, and orient the edges of $G$ with respect to $\prec$ to get the DAG $G_\prec$.
We start by enumerating static triangles in $G_\prec$. This can be done in $O(m_s\kappa)$ where $\kappa$ is the degeneracy of $G$~\cite{chiba1985arboricity,PiSeVi17}.
Note that all triangles in $G_\prec$ are acyclic as $G_\prec$ is a DAG, so each triangle in $G$ correspond to an acyclic triangle in $G_\prec$. In order to enumerate all triangles in $G$, we enumerate all directed edges in $G_\prec$, and for each directed edge $(u,v)$ we enumerate $N^+(u)$. For each vertex $w \in N^+(u)$, we check whether $\{u,v,w\}$ is a triangle by checking the existence of an edge between $v$ and $w$.
We call vertex $u$ in a static triangle $\{u,v,w\}$ the \emph{source vertex} if $u \prec v$ and $u \prec w$. Let $\{u,v,w\}$ be the triangle being processed while enumerating triangles in $G_\prec$. WLOG, assume $u$ is the source vertex in $\{u,v,w\}$. Thus, the number of times we visit $\{u,v\}$ or $\{u,w\}$ in a static triangle are limited by $d^+_{G_\prec}(u)$ that is bounded by $\kappa$. But the number of times we visit the static edge $\{v,w\}$ is not bounded by $\kappa$, so we want to avoid enumerating temporal edges on $\{v,w\}$. Next, we show how to count the number of $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangles corresponding to the static triangle $\{u,v,w\}$.
We define the temporal ordering of a temporal triangle corresponding to a static triangle $\{u,v,w\}$ as a mapping $\pi: \{1,2,3\} \rightarrow \{\{u,v\},\{u,w\},\{v,w\}\}$. There are six different possible temporal orderings as shown in~\Fig{temp_order}.
We define the orientation of a temporal triangle corresponding to a static triangle $\{u,v,w\}$ as a mapping $\rho$ from each pair of vertices of $\{u,v,w\}$ to one of the two possible ordered pairs of the same pair of vertices. For example, $\rho_1(\{u,v\})=(u,v)$ for $\rho_1$ in~\Fig{temp_dir}. The orientation of a temporal triangle simply determines the direction of its temporal edges. Each such temporal edge can take two possible directions, so there are eight types of orientation such a temporal triangle can take as shown in~\Fig{temp_dir}. Note that orientation of a temporal triangle is independent of its temporal ordering.
It is easy to see that the temporal ordering and orientation determine the type of the temporal triangle. But different combinations of temporal orderings and orientation could result in the same type. The temporal triangle type for all possible pairs of temporal ordering and orientation are shown in~\Tab{tri_type_conversion}.
For a temporal ordering $\pi$ and for $i \in \{1,2,3\}$, we use $S_i(\pi, \rho)$ to denote the sequence of temporal edges between the pair of vertices $\pi(i)$ that have the direction $\rho(\pi(i))$, in sorted order of timestamp. When $\pi$ and $\rho$ are clear from the context, we use $S_i$ instead of $S_i(\pi,\rho)$. We assume that we have access to $S_1$, $S_2$, and $S_3$ in constant time. Let $\sigma_i$ denote the length of $S_i$. We use $S_i[\ell]$ to denote the $\ell$-th edge in the sequence $S_i$, and $S_i[\ell:\ell^\prime]$ to denote the consecutive subsequence of $S_i$ ranging from $S_i[\ell]$ to $S_i[\ell^\prime]$.
For a sequence $S$ of temporal edges in increasing order of timestamp and timestamps $t$ and $t^\prime$ where $t \leq t^\prime$, let $\edgeCount([t,t^\prime], S)$ denote the number of edges in $S$ with a timestamp in the time window $[t,t^\prime]$. For given $\delta_{1,3},\delta_{1,2}$, and $\delta_{2,3}$, let $\consecutivetriangleCount(\{u,v,w\}, \pi, \rho)$ denote the number of $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangles corresponding to the static triangle $\{u,v,w\}$, temporal ordering $\pi$, and orientation $\rho$.
\begin{lemma} \label{lem:delta_temporal_count}
For a static triangle $\{u,v,w\}$, a temporal ordering $\pi$, and an orientation $\rho$,
\begin{align*}
& \consecutivetriangleCount(\{u,v,w\}, \pi, \rho) = \sum_{e_2 \in S_2} \sum\limits_{\substack{e_1 \in S_1 \\ t(e_1) \in [t(e_2)-\delta_{1,2},t(e_2)]}} & \\ &\edgeCount([t(e_2),\min(t(e_2)+\delta_{2,3},t(e_1)+\delta_{1,3})],S_3)&
\end{align*}
\end{lemma}
\begin{proof}
If temporal edge $e_1 \in S_1$ is in a $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangle with edge $e_2 \in S_2$, then $t(e_1) \in [t(e_2)-\delta_{1,2},t(e_2)]$. Fix a pair of temporal edges $(e_1,e_2)$ in $S_1 \times S_2 = \{(e_1,e_2) \mid e_1 \in S_1 \wedge e_2 \in S_2\}$ where $t(e_2) \in [t(e_1), t(e_1) + \delta_{1,2}]$. A temporal edge $e_3 \in S_3$ composes a $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangle with $e_1$ and $e_2$ iff $t(e_3) \in [t(e_2), \min(t(e_2) + \delta_{2,3}, t(e_1) + \delta_{1,3})]$.
\end{proof}
For a triangle $\{u,v,w\}$ where $u$ is the source vertex, we divide all six possible temporal orderings into three categories based on the place of $\{v,w\}$ in them. Recall that we want to avoid enumerating temporal edges on $\{v,w\}$. In $\pi_1$ and $\pi_2$, $\{v,w\}$ is assigned to the third place. $\{v,w\}$ is assigned to the second place in $\pi_3$ and $\pi_4$, and finally to the first place in temporal ordering $\pi_5$ and $\pi_6$.
{\bf Temporal orderings $\pi_1$ and $\pi_2$:}
Using~\Lem{delta_temporal_count}, one can compute $\consecutivetriangleCount(\{u,v,w\}, \pi, \rho)$ by enumerating pairs of temporal edges in $S_1 \times S_2 = \{(e_1,e_2) \mid e_1 \in S_1 \wedge e_2 \in S_2\}$. For each pair we compute $\edgeCount([t(e_2), \min(t(e_2) + \delta_{2,3}, t(e_1) + \delta_{1,3}),S_3)$ using binary search. To get the final counts we sum $\edgeCount([t(e_2), \min(t(e_2) + \delta_{2,3}, t(e_1) + \delta_{1,3}),S_3)$ over all pairs $(e_1,e_2) \in S_1 \times S_2$. But enumerating $S_1 \times S_2$ could be expensive and this process overall runs in time $O(\sigma_1\sigma_2 \log(\sigma_3))$. Next, we show that we can compute the same count by enumerating edges in $S_1$ and $S_2$ separately and storing cumulative counts of compatible edges on $S_3$ for each edge.
For $i,j \in \{1,2,3\}$ where $i \neq j$, and $\ell, \ell^\prime \in \{1,\ldots,\sigma_i\}$ where $\ell \leq \ell^\prime$ we use $\cumulativeEdgeCount_{+\delta_{1,3}}(S_i[\ell:\ell^\prime],S_j)$ to denote the cumulative count of edges in $S_j$ with a timestamp in $[t(e),t(e)+\delta_{1,3}]$ for edges $e$ in the sequence $S_i[\ell:\ell^\prime]$. Formally
\begin{align*}
\cumulativeEdgeCount_{+\delta_{1,3}}(S_i[\ell:\ell^\prime],S_j) = \sum\limits_{\ell \leq r \leq \ell^\prime} \edgeCount([t(S_i[r]),t(S_i[r])+\delta_{1,3}], S_j).
\end{align*}
Cumulative counts $\cumulativeEdgeCount_{\infty}$, $\cumulativeEdgeCount_{-\delta_{1,3}}$, and $\cumulativeEdgeCount_{-\infty}$, are defined the same way with time intervals $[t(e),\infty)$, $[t(e)-\delta_{1,3},t(e)]$, and $(-\infty,t(e)]$, respectively. $\cumulativeEdgeCount_{-\delta_{1,2}}$, $\cumulativeEdgeCount_{+\delta_{1,2}}$, $\cumulativeEdgeCount_{-\delta_{2,3}}$, and $\cumulativeEdgeCount_{+\delta_{2,3}}$ are defined similarly.
Note that we can compute $\cumulativeEdgeCount(S_i[1:\ell],S_j)$ for each $\ell \in \{1,\ldots,\sigma_i\}$ with one pass over $S_i$, and once we have these counts, we can get the cumulative counts $\cumulativeEdgeCount(S_i[\ell^\prime:\ell^{\prime\prime}],S_j)$, for each consecutive subsequence $S_i[\ell^\prime:\ell^{\prime\prime}]$ of $S_i$ as follows.
\begin{align*}
\cumulativeEdgeCount_{+\delta_{1,3}}(S_i[\ell^\prime:\ell^{\prime\prime}],S_j) & = \cumulativeEdgeCount_{+\delta_{1,3}}(S_i[1:\ell^{\prime\prime}],S_j) \\ & - \cumulativeEdgeCount_{+\delta_{1,3}}(S_i[1:\ell^\prime-1],S_j).
\end{align*}
where $\cumulativeEdgeCount_{+\delta_{1,3}}(S_i[1:0],S_j) = 0$.
We first enumerate edges in $S_1$.
For each edge $e_1 \in S_1$ we compute $\cumulativeEdgeCount_{+\delta_{1,3}}(S_1[1:\ell],S_3)$ and $\cumulativeEdgeCount_{\infty}(S_1[1:\ell],S_3)$ for each $\ell \in \{1,\ldots,\sigma_1\}$ and store them for $e_1$.
Next, we enumerate edges in $S_2$ and compute $\cumulativeEdgeCount_{\infty}(S_2[1:\ell],S_3)$ for each $\ell \in \{1,\ldots,\sigma_2\}$.
Fix an edge $e_2 \in S_2$. Let $\ell_{f}$ and $\ell_{\ell}$ be the indices of the first and last edges in $S_1$ with a timestamp in $[t(e_2)-\delta_{1,2},t(e_2)]$. Also let $\ell_{\delta_{2,3}}$ be the index of the last edge in $S_1$ with a timestamp at most $t(e_2)-\delta_{1,3}+\delta_{2,3}$. We can find $\ell_{f},\ell_{\delta_{2,3}}$, and $\ell_{\ell}$ using a binary search on $S_1$. Note that $\delta_{1,3} \leq \delta_{1,2} + \delta_{2,3}$, thus $\ell_{f} \leq \ell_{\delta_{2,3}} \leq \ell_{\ell}$.
First consider the temporal edges $S_1[i]$ where $\ell_{f} \leq i \leq \ell_{\delta_{2,3}}$. For any such edge $t(S_1[i])+\delta_{1,3} \leq t(e_2)+\delta_{2,3}$, so the timestamp of compatible edges in $S_3$ lie in the interval $[t(e_2),t(S_1[i])+\delta_{1,3}]$. Having stored the cumulative counts described above, we can compute the number of pairs of temporal edges $(e_1,e_3) \in S_1[\ell_{f}:\ell_{\delta_{2,3}}] \times S_3$ that compose a $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangle on $\{u,v,w\}$ with $e_2$, complying with $\pi_1$ and $\rho$, as follows.
\begin{align*}
& \sum\limits_{\ell \leq i \leq \ell_{\delta_{2,3}}} \edgeCount([t(e_2),t(S_1(i))+\delta_{1,3}], S_3) = & \\
& \cumulativeEdgeCount_{+\delta_{1,3}}(S_1[\ell_{f}:\ell_{\delta_{2,3}}], S_3) - \cumulativeEdgeCount_{\infty}(S_1[\ell_{f}:\ell_{\delta_{2,3}}],S_3) & \\
+ & (\ell_{\delta_{2,3}}-\ell_{f}+1) \cdot \edgeCount([t(e_2),\infty),S_3) &
\end{align*}
Now, we count the number of temporal edges in $S_3$ that compose a triangle with $e_2$ and $S_1[i]$, where $\ell_{\delta_{2,3}} < i \leq \ell_{\ell}$. For a temporal edge $S_1[i]$ where $\ell_{\delta_{2,3}} < i \leq \ell_{\ell}$, we have $t(S_1[i]) + \delta_{1,3} > t(e_2)+\delta_{2,3}$. Thus, there are $\edgeCount([t(e_2), t(e_2)+\delta_{2,3}], S_3)$ edges on $S_3$ that compose a triangle with $e_2$ and $S_1[i]$. So the final count of pairs $(e_1,e_3) \in S_1 \times S_3$ that are in a temporal triangle with $e_2$ corresponding to the static triangle $\{u,v,w\}$ can be computed as follows.
\begin{align*}
& \sum\limits_{\substack{e_1 \in S_1, t(e_1) \in \\ [t(e_2)-\delta_{1,2},t(e_2)]}} \edgeCount([t(e_2),\min(t(e_2)+\delta_{2,3},t(e_1)+\delta_{1,3})],S_3) & \\
= & \sum\limits_{\ell_{f} \leq i \leq \ell_{\delta_{2,3}}} \edgeCount([t(e_2),t(S_1(i))+\delta_{1,3}],S_3) & \\
+ & (\ell_{\ell} - \ell_{\delta_{2,3}}) \cdot \edgeCount([t(e_2),t(e_2)+\delta_{2,3}],S_3) & \\
\end{align*}
By~\Lem{delta_temporal_count}, to get $\consecutivetriangleCount(\{u,v,w\}, \pi, \rho)$, we only need to sum these counts over edges in $S_2$. Let $\langle u,v,w \rangle$ denote a static triangle where $u \prec v \prec w$.~\Alg{tempOrder_1_count} formalizes the procedure described above for computing $\consecutivetriangleCount(\langle u,v,w \rangle, \pi, \rho)$ where $\pi$ is either $\pi_1$ or $\pi_2$.
\begin{algorithm}[th]
\setstretch{1.2}
\caption{Counting $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangles corresponding to a static triangle and temporal orientation $\pi_1$ or $\pi_2$}\label{alg:tempOrder_1_count}
\begin{algorithmic}[1]
\Procedure{TTC-vw3}{$\delta_{1,3}$, $\delta_{1,2}$, $\delta_{2,3}$,$\langle u,v,w \rangle$, $\pi$, $\rho$}
\LeftComment $\pi(\{v,w\})=3$
\State Enumerate $S_1$ and compute $\cumulativeEdgeCount_{+\delta_{1,3}}$ and $\cumulativeEdgeCount_{\infty}$ on $S_3$.
\State count = 0
\For{$i= 1,\ldots,\sigma_2$}
\State Let $\ell_{f} = \textsc{lowerBound}(t(S_2[i])-\delta_{1,2}, S_1)$
\State Let $\ell_{\delta_{2,3}} = \textsc{upperBound}(t(S_2[i])-\delta_{1,3}+\delta_{2,3}, S_1)$
\State Let $\ell_{\ell}= \textsc{upperBound}(t(S_2[i]), S_1)$
\LeftComment Edges in $S_1[\ell_{f}:\ell_{\delta_{2,3}}]$
\State count $+= \cumulativeEdgeCount_{+\delta_{1,3}}(S_1[\ell_{f}:\ell_{\delta_{2,3}}],S_3)$
\State count $-= \cumulativeEdgeCount_{\infty}(S_1[\ell_{f}:\ell_{\delta_{2,3}}],S_3)$
\State count $+= (\ell_{\delta_{2,3}}-\ell_{f}+1) \cdot \edgeCount([t(S_2[i]),\infty), S_3)$
\LeftComment Edges in $S_1[\ell_{\delta_{2,3}}+1:\ell_{\ell}]$
\State count $+= (\ell_{\ell} - \ell_{\delta_{2,3}}) \cdot \edgeCount([t(S_2[i]), t(S_2[i])+\delta_{2,3}], S_3)$
\EndFor
\State \Return count
\EndProcedure
\end{algorithmic}
\end{algorithm}
The algorithms for the category of $\pi_3$ and $\pi_4$ and the category of $\pi_5$ and $\pi_6$ are similar to~\Alg{tempOrder_1_count}, so we defer these to~\Apx{missing_algo}. Here, we suffice to say that similar to~\Alg{tempOrder_1_count}, in algorithms for the remaining temporal orderings, for each static triangle $\{u,v,w\}$ where $u$ is the source vertex, we only enumerate temporal edges on $\{u,v\}$ and $\{u,w\}$. And for each temporal edge we perform a constant number of binary searches on temporal edges on the other two static edges of the static triangle $\{u,v,w\}$.
\subsection{Getting the Counts for All Temporal Triangle Types}
Now that we can count $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangles for each combination of temporal ordering and orientation, it only remains to get the counts for each temporal triangle type (~\Fig{temp_tri}). Let $\psi(\pi,\rho)$ denote the triangle type for $\pi$ and $\rho$.~\Alg{count_all_delta} gets the counts for all eight types. Now, we can finally prove~\Thm{all_runtime}.
\begin{algorithm}[th]
\caption{Counting $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangles for each temporal triangle types}\label{alg:count_all_delta}
\begin{algorithmic}[1]
\Procedure{Count-temporal-triangles}{$T$, $\delta_{1,3}$, $\delta_{1,2}$, $\delta_{2,3}$}
\State Extract the static graph $G$ of $T$.
\State Find the degeneracy ordering $\prec$ of $G$.
\State Derive $G_\prec$ by orienting $G$ with respect to $\prec$.
\State Initialize Counts to 0 for $\mcT_1,\ldots,\mcT_8$.
\ForAll{Static triangles $\{u,v,w\}$}
\Statex\Comment{WLOG let $u\prec v \prec w$}
\ForAll{Temporal ordering $\pi$ and orientation $\rho$}
\State Counts($\psi(\pi,\rho)$) += $\consecutivetriangleCount(\langle u,v,w \rangle, \pi, \rho)$
\Comment{~\Tab{tri_type_conversion}}
\Statex\Comment{Using~\Alg{tempOrder_1_count},~\Alg{tempOrder_3_count}, and~\Alg{tempOrder_5_count}}
\EndFor
\EndFor
\EndProcedure
\end{algorithmic}
\end{algorithm}
\begin{proof}[Proof of Theorem~\ref{thm:all_runtime}]
Extracting the static graph $G$ from $T$ can be done in $O(m)$ time. We simply enumerate all temporal edges of $T$ and for each temporal edge $e=(v_1,v_2,t)$, we add an static edge between $\{v_1,v_2\}$ in $G$ if they are not connected already. The degeneracy ordering of $G$ could be obtained in $O(m_s)$ time~\cite{matula1983smallest}, and $G_\prec$ could also be derived in time $O(m_s)$. For enumerating static triangles we first enumerate each edge in $G_\prec$. For each edge $(u,v)$, we enumerate $N^+(u)$ which takes $O(\kappa)$ as $d^+_{G_\prec}(u) \leq \kappa$. We can lookup if there is an edge between $v$ and $w$ in constant time. Thus, enumerating triangles take $O(m_s \cdot \kappa)$ time overall.
Note that for each static triangle we only enumerate temporal edges on static edges incident to the source vertex. So for the static triangle $\langle u,v,w \rangle$, we only enumerate temporal edges on the pairs $\{u,v\}$ and $\{u,w\}$. While processing a temporal edge during enumeration of temporal edges on $\{u,v\}$ or $\{u,w\}$, we either perform a constant time operation, or spend $O(\log(\sigma_{\max}))$ time for a constant number of binary searches over the temporal edges of the other two static edges in the static triangle $\langle u,v,w \rangle$. Thus,
\begin{align*}
T(\mathcal A) = O(m_s \cdot \kappa + \sum\limits_{\langle u,v, w \rangle} (\sigma(u,v) + \sigma(u,w)) \log(\sigma_{\max}))
\end{align*}
where $\mathcal A$ denotes~\Alg{count_all_delta}, and $T(\mathcal A)$ denotes the worst case time complexity of $\mathcal A$.
For each vertex $u \in V$, $d^+_{G_\prec}(u) \leq \kappa$, so each edge $(u,v)$ in $G_\prec$, is a part of at most $\kappa$ static triangles where $u$ is the source vertex. Therefore, the temporal edges on each edge $\{u,v\}$ in $G$ are enumerated at most $O(\kappa)$ times. Thus,
\begin{align*}
T(\mathcal A) = O(m_s \cdot \kappa + \sum\limits_{\{u,v\} \in E_s} (\sigma(u,v) + \sigma(v,u)) \cdot \kappa \log(\sigma_{\max})).
\end{align*}
Hence,
\begin{align*}
T(\mathcal A) = O(m \kappa \log(\sigma_{\max})).
\end{align*}
\end{proof}
\section{Experimental Evaluations}\label{sec:experimental}
\begin{table*}[!ht]
\scriptsize
\caption{Descriptions of the datasets and runtime of \textsc{DOTTT}{} and PBL\space.}
\begin{center}
\begin{tabular}{l|rrrrrrr|rr}
\hline
\textbf{dataset} & $\#$\textbf{vertices} & $\#$\textbf{edges} & $\#$\textbf{static edges} & $\#$\textbf{static triangles} & \textbf{degeneracy} & \textbf{max multiplicity} & \textbf{time span (years)} & \textbf{\textsc{DOTTT}{} runtime} & \textbf{PBL\space runtime}\\\hline
CollegeMsg & 1.9K & 59.8K & 13.8K & 14.3K & 20 & 98 & 0.51 & 0.09 & 0.07\\
email-Eu-core & 986 & 332K & 16.1K & 105K & 34 & 2.8K & 2.2 & 2.31 & 3.37\\
MathOverflow & 24.7K & 390K & 188K & 1.4M & 78 & 225 & 6.46 & 3.17 & 3.6\\
SMS-A & 44.1K & 545K & 52.22K & 10K & 9 & 5.3K & 0.92 & 0.45 & 0.81\\
AskUbuntu & 157K & 727K & 456K & 680K & 48 & 154 & 7.09 & 2.23 & 5.08\\
SuperUser & 192K & 1.11M & 715K & 1.54M & 61 & 78 & 7.59 & 4.41 & 8.84\\
WikiTalk & 1.09M & 6.11M & 2.79M & 8.12M & 124 & 1.1K & 6.21 & 34 & 56\\
StackOverflow & 2.58M & 47.9M & 28.18M & 114.2M & 198 & 549 & 7.60 & 347 & 678\\
Wikipedia-DE & 2.17M & 86.21M & 39.71M & 169.9M & 265 & 347 & 10.18 & 576 & 987\\
Bitcoin & 59.61M & 515.5M & 366.4M & 706.2M & 604 & 447K & 5.98 & 2923 & 4374\\
\hline
\end{tabular}
\end{center}
\label{tab:runtime}
\end{table*}
\begin{figure*}[!ht]
\centering
\begin{subfigure}[b]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{Figures/dist.pdf}
\caption{Temporal triangle count distribution}
\label{fig:count_dist}
\end{subfigure}
\begin{subfigure}[b]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{Figures/closure_delta1.pdf}
\caption{Effect of $\delta_{1,2}$ and $\delta_{2,3}$ on triadic closure}
\label{fig:closure_delta1}
\end{subfigure}
\begin{subfigure}[b]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{Figures/cyclic_vs_acyclic.pdf}
\caption{Triadic closure in cyclic and acyclic cases}
\label{fig:closure_cyclic}
\end{subfigure}
\caption{(a):The distribution of (1 hr, 1 hr, 1 hr)-temporal triangle counts over all eight temporal triangle types as shown in~\Fig{temp_tri}. (b):We fix $\delta_{1,3}$ to 2 hrs. We vary $\delta_{1,2}$ from 0 to 60 minutes and plot the ratio of (2 hrs, $\delta_{1,2},\delta_{2,3}$)-temporal triangles to (2hrs, $\delta_{1,2}$, 1hr)-temporal triangles for $\delta_{2,3}$ ranging from 0 to 60 minutes. (c) We plot the ratio of (2 hrs, 1 hr, $\delta_{2,3}$)-temporal triangles to (2hrs, 1hr, 1hr)-temporal triangles for $\delta_{2,3}$ ranging from 0 to 60 minutes, for cyclic and acyclic triangles.}
\label{fig:count_exp}
\end{figure*}
We implemented our algorithm in C++ and used a commodity machine from AWS EC2: R5d.2xlarge to run our experiments. This EC2 instance has Intel(R) Xeon(R) Platinum 8175M CPU @ 2.50GHz and 64GB memory. On this AWS machine, PBL\space runs out of memory for the Bitcoin graph, so we used one with more than 256GB memory for this case. The implementation of \textsc{DOTTT}{} is available at~\cite{ettc}.
We performed our experiments on a collection of temporal graphs from SNAP~\cite{snapnets}, KONECT\cite{konect}, and the Bitcoin transaction dataset from~\cite{kondor2014inferring}, consisting of all transactions up to Feb 9, 2018. The timestamp of each transaction is the creation time of the block on the blockchain that contains it\cite{reid2013analysis}.
{\bf Running time:}
All the running times are shown in~\Tab{runtime}. We ran all experiments on a single thread. In most instances, \textsc{DOTTT}{} takes a few seconds to run. For graphs with tens of millions of temporal edges, \textsc{DOTTT}{} runs in less than ten minutes. Even for the Bitcoin graph with 515M edges, \textsc{DOTTT}{} takes less than an hour.
{\bf Running time independent of time periods:}
The running time of both \textsc{DOTTT}{} and PBL\space algorithms are independent of the time periods. \textsc{DOTTT}{} has the same running time for time restrictions ranging from 0 to the time span of the input dataset. For comparison with $PBL\space$, we set $\delta_{1,3}=\delta_{1,2}=\delta_{2,3}=$ 1 hr.
{\bf Comparison with PBL\space:}
We compare our algorithm with the PBL\space algorithm that counts $\delta_{1,3}$-temporal triangles, as it is the closest to our work. We typically get a 1.5x-2x speedup over PBL\space for large graphs (more than 0.5M edges) as shown in~\Fig{speedup}. Note that \textsc{DOTTT}{} computes $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangle counts while PBL\space only gets the counts of $\delta_{1,3}$-temporal triangles.
{\bf Distribution of counts over types of triangles:}
The distribution of (1 hr, 1 hr, 1 hr)-temporal triangle counts for our datasets are shown in~\Fig{count_dist}. As we expected~\cite{milo2004superfamilies,vazquez2004topological,yaverouglu2014revealing,paranjape2017motifs}, networks from similar domains have similar distributions. It is easy to see in~\Fig{count_dist}, that all the stack exchange networks have similar distributions. The same holds for the message networks CollegeMsg and SMS-A.
We observe that cyclic temporal triangles, $\mcT_4$ and $\mcT_8$, have a larger share in temporal triangle counts in messaging networks than in stack exchange networks.
{\bf Triadic closures in temporal networks:}
In static triangles, the transitivity measures the ratio of number of static triangles to the number of all wedges. In temporal graphs, in addition to transitivity, the time it takes for a wedge to appear and close is of importance~\cite{zignani2014link}. In~\Fig{closure_delta1}, we study the effect of the time it takes for a wedge to appear from an edge, on the time it takes to close for CollegeMsg graph. We fix $\delta_{1,3} = $ 2 hrs. For $\delta_{1,2}$ ranging from zero to 60 minutes (10 minute steps), we vary $\delta_{2,3}$ from zero to 60 minutes and plot the ratio of $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangles over (2hrs, $\delta_{1,2}$, 1hr)-temporal triangles. We observe that the set of ratios for all values of $\delta_{2,3}$ are almost identical for different values of $\delta_{1,2}$. For instance, for all values of $\delta_{1,2}$, roughly half the triangles are formed in 10-20 minutes. This implies that once a wedge is formed, the time it took to appear does not affect the time it takes to close.
As another demonstration of \textsc{DOTTT}{}, for $\delta_{1,3} = $ 2 hrs and $\delta_{1,2}=$ 1 hrs, we plot the ratio of $(\delta_{1,3},\delta_{1,2},\delta_{2,3})$-temporal triangles to (2hrs, 1hr, 1hr)-temporal triangles, this time separately for cyclic and acyclic temporal triangles in~\Fig{closure_cyclic}. We observe that for stack exchange networks, acyclic temporal triangles tend to take a shorter time to close from the moment their second edge appears than cyclic temporal triangles. As we see in~\Fig{closure_cyclic}, this is not the case for message networks.
\begin{acks}
The authors are supported by NSF DMS-2023495, NSF CCF-1740850, NSF CCF-1813165, CCF-1909790, and ARO Award W911NF1910294. Noujan Pashanasangi is supported by Jack Baskin and Peggy Downes-Baskin Fellowship.
\end{acks}
\bibliographystyle{acm}
|
\section{Introduction}
Multi-agent systems (MASs) have received considerable attention in recent years, and they have been applied in a wide variety of areas, such as automotive control system \cite{Pham2019b}, unmanned aerial vehicles \cite{Pham2020c}, sensor network \cite{Pham2020d}, etc. Besides, MASs are vulnerable to faults and attacks from the network. The faults or the attack occurring at an agent might spread to other agents. It means that the neighbor's agent should be affected by these faults or attacks through a communication topology of MASs \cite{Pham2020b}, \cite{Nguyen2020a}. The imperative purpose of fault detection and isolation (FDI) of MASs is thus becoming a very crucial, more challenging.
In fault detection and isolation problems for MASs, there are three methodologies: centralized, decentralized and distributed filter design. Among three aforementioned approaches, the centralized architecture is the fewest attraction because the structure of the MASs is distributed and not all measurements are available to each agent. Moreover, MASs could be burdened on a complex computation cost when the number of agents increases. Concerning the decentralize approach, a sufficient condition based on the $\mathcal H_{\infty}$ performance for large-scale systems is obtained by applying Lyapunov stability theory \cite{Li2009a}. Next, a local/decentralized detection and isolation for multi-robot systems was proposed \cite{Arrichiello2015}, where each robot is able to detect and isolate faults occurring on other robots. However, the decentralized approach is constructed by an observer, which contains the model of entire system. Thus, compared to the centralized and decentralized FDI methods, the distributed FDI for MASs has been the most attractive topic, because of using fewer network resources and having lower computation complexity. In the distributed FDI methodology, there are two attractive trends in considering, the first trend introduce a consensus protocol, which establishes a distributed dynamic model of the MASs for specific agent \cite{Shames2011, Liu2016, Quan2016, Gao2017}. Each agent in the network can update its information according to the consensus protocol and can send and receive information with its neighbor agents. Then, by constructing a bank of observers, each agent can detect not only its faults but also faults of its own neighbor agents. However, as the number of agents increases in the network, the FDI will place a heavy computational burden on the entire system. The second trend uses the sensor measurements both locally and from the agent's neighbors \cite{Davoodi2016, Nguyen2017d}.
On the basis of the above review, there are some limitations on the distributed FDI for MASs. Firstly, most of the literature considers the FDI for network of \textit{homogeneous dynamics} rather than \textit{heterogeneous dynamics}, which are different dynamics, as well as sensor faults are not considered in \cite{Liu2016}, \cite{Quan2016}, \cite{Shames2011}. Secondly, the FDI will suffer from a heavy computational burden on the whole system. To overcome this problem, we use the relative output measurements such as \cite{Nguyen2020a} to construct the virtual model of each agent. Thirdly, in order to detect faults, unknown input observer is a powerful technique, where the perfect unknown input decoupling condition is needed to guarantee. If this condition is not satisfied, which is common in practice, the methodologies proposed in \cite{Gao2017}, \cite{Quan2016} could not be applied. Therefore, instead of employing unknown input observer with the perfect unknown input decoupling condition, the distributed observer based on Luenberger is proposed in this paper, which has a simple structure. Moreover, the residual generation problem can be formulated as an $\mathcal{H}_2$ optimal filtering problem (the Kalman filtering) \cite{Ding}. Finally, although most of the proposed methodologies in previous works can achieve fault detection and isolation, there are rare methodologies, which achieves FDI and state estimation objectives at the same time.
\textbf{Contributions}. Motivated by the above works, the problem of distributed FDI for a network of heterogeneous MASs is addressed to detect and isolate the faults. A relative model is constructed by utilizing a combination of local sensor information of the nearest neighboring agents (called a vector of distributed relative output measurements). Therefore, the main contributions of this paper are summarized as follows: First, we develop a distributed observer for a team of time-invariant MASs, which utilizes a vector of distributed relative output measurements. To guarantee the existence of at least an observer matrix gain, the detectable property of closed loop system should be guaranteed. Furthermore, by using the distributed Luenberger observer, each agent is able to estimate its own states and the states of its nearest neighbors in the presence of the disturbances, faults, and the control inputs. Secondly, a multi-objective optimization, $\mathcal{H}_2$ and $\mathcal{H}_{-}$ performance indexes, is introduced to robustness against the disturbance signal on the residual signals and the sensitivity for the fault signals. Similar to the work in \cite{Nguyen2017d} the formulation proposed in this paper, in which each agent can detect not only its own faults but also the faults of its neighbors. Then, the distributes fault isolation strategy is proposed corresponding to the residual signals of the network.
This paper is organized as follows. In Section II, we describe the system description and problem formulation. The LMI-based solution to the distributed FDI problem is developed in Section III. To demonstrate the validity of the proposed approach, a numerical example is given in Section IV which is followed by conclusions in Section V.
\textbf{Notations}. The notation used in this paper is fairly standard. For a given matrix $A, A^T$ and $Trace(A)$ denote its transpose and trace, respectively. Using the notation $G:=(A,B,C,D)$
\begin{align*}
G:=\left[
\begin{array}{c|c}
A & B \\ \hline
C & D \\
\end{array}
\right] , [G_1\;\;G_2] :=\left[
\begin{array}{c|c}
A & B_1 \;\;\;\;B_2 \\ \hline
C & D_1 \;\;\;\;D_2 \\
\end{array}
\right]
\end{align*}
where $G_1:=(A,B_1,C,D_1)$ and $G_2:=(A,B_2,C,D_2)$. We drop the argument "s" in transfer matrices. The transfer matrix $G$ is proper if $G(\infty)=D$ and $G$ is strictly proper if $G(\infty)=0$. $\mathbb{RH}_{\infty}$ and $\mathbb{RH}_{2}$ denotes the set of stable and strictly proper transfer matrices. The $\mathcal H_2$ norm of $G$ is calculated as $\|G\|_2^2=Trace(B^TYB)=Trace(CQC^T)$ where $Y$ and $Q$ represent respectively the observability and controllability Gramaian ($A^TY+YA+C^TC=0$ and $AQ+QA^T+BB^T=0$). The notation $\|.\|_p$ denotes the $\mathcal{H}_2$ or $\mathcal{H}_{-}$ norm. Finally, we use $*$ to denote the symmetry entries of symmetry matrices.
\input{prob}
\input{mainresult}
\section{Simulation Results}
In this section, a heterogeneous MASs with two first-order agent 1 and agent 2 as well as one second-order agent 4 and one fourth-order agent 3 are considered. Their topology graph is shown in Fig. \ref{fig:1}. The following data associated with continuous-time model (\ref{1}) is considered.
\begin{figure}
\centering
\includegraphics[height=3.6cm]{1.jpg}
\caption{The topology of an MASs}\label{fig:1}
\end{figure}
\begin{align*}
A_{1}&=\begin{bmatrix}
{{-5}}&{{0.05}}\\
{{0}}&{{-13}}
\end{bmatrix}, B_{1}=\begin{bmatrix}
{{1}}\\
{{0}}
\end{bmatrix}, C_{1}=\begin{bmatrix}
{{3}}&{{0}}\\
{{0.01}}&{{0.1}}
\end{bmatrix},\\
B_{f1}&=\begin{bmatrix}
{{0.3}}\\
{{0.1}}
\end{bmatrix}, B_{d1}=\begin{bmatrix}
{{0.2}}\\
{{0.1}}
\end{bmatrix}, \\
D_{f1}&=\begin{bmatrix}
{{0.45}}\\
{{0.2}}
\end{bmatrix}, D_{d1}=\begin{bmatrix}
{{0.27}}\\
{{0.2}}
\end{bmatrix}\\
A_{2}&=\begin{bmatrix}
{{-2}}&{{0}}\\
{{0}}&{{-10}}
\end{bmatrix}, B_{2}=\begin{bmatrix}
{{1}}\\
{{0.1}}
\end{bmatrix}, C_{2}=\begin{bmatrix}
{{1}}&{{0}}\\
{{0}}&{{0.1}}
\end{bmatrix},\\
B_{f2}&=\begin{bmatrix}
{{1}}\\
{{0.1}}
\end{bmatrix}, B_{d2}=\begin{bmatrix}
{{0.3}}\\
{{0.1}}
\end{bmatrix},\\
D_{f2}&=\begin{bmatrix}
{{0.15}}\\
{{0.3}}
\end{bmatrix}, D_{d2}=\begin{bmatrix}
{{0.47}}\\
{{0.1}}
\end{bmatrix},\\
A_{3}&=\begin{bmatrix}
{{-1}}&{{0}}&{{0.05}}&{{0}}\\
{{0}}&{{-3.6}}&{{0}}&{{0.1}}\\
{{0}}&{{0}}&{{-2}}&{{0.1}}\\
{{-2}}&{{-10}}&{{-13}}&{{-9}}
\end{bmatrix}, B_{3}=\begin{bmatrix}
{{1}}\\
{{0}}\\
{{1}}\\
{{0}}
\end{bmatrix}, \\
B_{f3}&=\begin{bmatrix}
{{0.1}}\\
{{0.2}}\\
{{0}}\\
{{0}}
\end{bmatrix}, B_{d3}=\begin{bmatrix}
{{0.2}}\\
{{0.1}}\\
{{0}}\\
{{0}}
\end{bmatrix},D_{d3}=\begin{bmatrix}
{{0.35}}\\
{{0.2}}
\end{bmatrix},\\
C_{3}&=\begin{bmatrix}
{{0.7}}&{{0.5}}&{{0}}&{{0}}\\
{{0.5}}&{{0.1}}&{{0}}&{{0.1}}
\end{bmatrix}, D_{f3}=\begin{bmatrix}
{{0.17}}\\
{{1}}
\end{bmatrix}\\
A_{4}&=\begin{bmatrix}
{{-1}}&{{0}}&{{0.05}}\\
{{0}}&{{-3.6}}&{{0}}\\
{{-2}}&{{-10}}&{{-13}}
\end{bmatrix}, B_{4}=\begin{bmatrix}
{{1}}\\
{{0}}\\
{{1}}
\end{bmatrix}, \\
B_{f4}&=\begin{bmatrix}
{{0.11}}\\
{{0.2}}\\
{{0}}
\end{bmatrix}, B_{d4}=\begin{bmatrix}
{{0.1}}\\
{{0.2}}\\
{{0}}
\end{bmatrix},D_{d4}=\begin{bmatrix}
{{0.15}}\\
{{0.3}}
\end{bmatrix},\\
C_{4}&=\begin{bmatrix}
{{0.7}}&{{0.5}}&{{0.1}}\\
{{0.5}}&{{0.1}}&{{0.1}}
\end{bmatrix}, D_{f4}=\begin{bmatrix}
{{0.27}}\\
{{1}}
\end{bmatrix}
\end{align*}
the disturbances $d_1(t), d_2(t), d_3(t)$ and $d_4(t)$ are band-limited white noise with powers 0.001. The input signals $u_1(t), u_2(t), u_3(t)$ and $u_4(t)$ are taken as step inputs with amplitudes of 0.1, 0.5, 1 and -1, respectively. The initial states of agents 1,2,3 and 4 are $[0.4; \;0.4]$, $[0.2; \;0.2]$, $[0.4; -0.3; \;0.4; \;0.2]$ and $[-0.3; \;0.3;\;0.3]$, respectively. The initial states of observers of all agents are set be zero.
\begin{figure}
\centering
\includegraphics[height=3.5cm]{12.jpg}
\caption{The error of estimation of agent 4 and agent 1}
\label{fig:12}
\end{figure}
To demonstrate the performance of our proposed methodologies through considering different types of faults. $d_1(t), d_2(t), d_3(t)$ and $d_4(t)$ are assumed to occur in the agent 1, 2, 3 and 4, respectively. The fault signals in agent 1 ($f_1(t)$), agent 4 ($f_4(t)$) and agent 3 ($f_3(t)$) are simulated as a rectangular pulsed signal with an amplitude of -0.25, 0.5 and 0.15 that are connected during the time interval $[10 - 20]s$, $[30 - 40]s$ and $[60 - 70]s$, respectively. There is no faults in the agent 2 ($f_2(t)=0$). The estimation state error $e_{ijk}(t)=x_{Ni}(t) - \hat x_{Ni}(t)$ of agent 1 and agent 4 are represented in Fig. \ref{fig:12}. It shown that each agent estimates its own states and the states of its nearest neighbors in the presence of the disturbance $d_{Ni}(t)$, faults $f_{Ni}(t)$ and the control inputs $u_{Ni}(t)$.
\begin{figure*}
\centering
\begin{minipage}{0.8\textwidth}
\includegraphics[height=9.1cm]{5.jpg}
\caption{Actuator Faults: Residual signals of agent 1,2,3 and agent 4}
\label{fig:8}
\end{minipage}\\
\vspace{1cm}
\begin{minipage}{0.8\textwidth}
\includegraphics[height=9.1cm]{6.jpg}
\caption{Sensor Faults: Residual signals of agent 1,2,3 and agent 4}
\label{fig:8a}
\end{minipage}
\end{figure*}
To illustrate the effectiveness of our proposed method solving distributed FDI problem for system with parameter matrices given above. The relative mode system should be constructed for agent 1, 2, 3 and 4, respectively. Faults will take into account in two scenarios: actuator faults $\mathcal{B}_{fi} =\mathcal B_{ui}, \bar{\mathcal D}_{fi} =\bar{\mathcal D}_{di}=0, \; i=1,2,3,4$ and sensor faults $\mathcal{B}_{fi}=0, \bar{\mathcal D}_{fi} =I, \; i=1,2,3,4$, respectively.
\subsection{Scenario 1: Actuator Faults}
In the following, we illustrate the performance of our proposed methodology by considering actuator fault. The thresholds are obtained corresponding to the threshold functions of the agents in our method. In \textit{Theorem \ref{thm1}}: $J_{th1}=0.03, J_{th2}=0.05, J_{th3}=0.02,$ and $J_{th4}=0.025,$. Moreover, the residual evaluation functions $J_{r_1}(t), J_{r_2}(t), J_{r_3}(t),$ and $J_{r_4}(t)$ are shown in Fig. \ref{fig:8}. It is clearly realized that each fault can be detected from the other faults and the disturbances. The next step, fault can be isolated by using the Algorithm 1 for each method of our proposed methodologies.
For the fault $f_4(t)$, the fault pattern $\Upsilon_i$ are obtained as $\Upsilon_1= \left\{ \begin{bmatrix}
{{0}}\\
{{0}}
\end{bmatrix}, \begin{bmatrix}
{{0}}\\
{{0}}
\end{bmatrix}, \begin{bmatrix}
{{1}}\\
{{1}}
\end{bmatrix} \right\}, \Upsilon_2= \left\{ \begin{bmatrix}
{{0}}\\
{{0}}
\end{bmatrix}, \begin{bmatrix}
{{0}}\\
{{0}}
\end{bmatrix}\right\},
\Upsilon_3= \left\{ \begin{bmatrix}
{{0}}\\
{{0}}
\end{bmatrix}, \begin{bmatrix}
{{0}}\\
{{0}}
\end{bmatrix} \right\},
\Upsilon_4= \begin{bmatrix}
{{1}}\\
{{1}}
\end{bmatrix}$. All of elements of $\Upsilon_4=1$ that means the fourth agent is thus faulty, and the agent 1 is the nearest neighbor to the agent 4 based on observing the fault pattern $\Upsilon_1$. Moreover, the fault patterns for the fault signal $f_1(t)$ are obtained as $\Upsilon_1= \left\{ \begin{bmatrix}
{{1}}\\
{{1}}
\end{bmatrix}, \begin{bmatrix}
{{1}}\\
{{1}}
\end{bmatrix},\begin{bmatrix}
{{1}}\\
{{1}}
\end{bmatrix} \right\}
, \Upsilon_2=\begin{bmatrix}
{{1}}\\
{{1}}
\end{bmatrix}, \begin{bmatrix}
{{0}}\\
{{0}}
\end{bmatrix},
\Upsilon_3= \left \{ \begin{bmatrix}
{{1}}\\
{{1}}
\end{bmatrix}, \begin{bmatrix}
{{0}}\\
{{0}}
\end{bmatrix} \right \},
\Upsilon_4= \left \{\begin{bmatrix}
{{1}}\\
{{1}}
\end{bmatrix} \right \}$. It recognize that there are two fault patterns $\Upsilon_1, \Upsilon_4$ which have all of elements equal to one. In addition, the first element of fault patterns $\Upsilon_2= \Upsilon_3=1$ corresponding to flags $[r_{21}], [r_{31}]$ this mean that the fault will occur on the neighbor of agent 2 and agent 3, respectively. Consequently, the agent 1 is faulty and agent 2, 3, 4 are the nearest neighbors of agent 1, respectively. Besides, the fault patterns for the fault signal $f_3(t)$ are determined such that $\Upsilon_1= \left\{ \begin{bmatrix}
{{0}}\\
{{0}}
\end{bmatrix}, \begin{bmatrix}
{{1}}\\
{{1}}
\end{bmatrix}, \begin{bmatrix}
{{0}}\\
{{0}}
\end{bmatrix}\right\}
, \Upsilon_2= \begin{bmatrix}
{{0}}\\
{{0}}
\end{bmatrix}, \begin{bmatrix}
{{1}}\\
{{1}}
\end{bmatrix},
\Upsilon_3= \left\{ \begin{bmatrix}
{{1}}\\
{{1}}
\end{bmatrix}, \begin{bmatrix}
{{1}}\\
{{1}}
\end{bmatrix} \right \},
\Upsilon_4= \begin{bmatrix}
{{0}}\\
{{0}}
\end{bmatrix}$ and it conclude that the agent 3 is thus faulty because of all of elements of $\Upsilon_3=1$. Besides, the agent 1 and agent 2 are the nearest neighbors to the agent 4, respectively.
\subsection{Scenario 2: Sensor Faults}
In the subsection, the sensor fault will be investigated based on our proposed approach (which is can not implemented using the unknown input observer as \cite{Liu2016}, \cite{Quan2016}, \cite{Shames2011}). Similarly as mention above, the thresholds are obtained corresponding to the threshold functions of the agents in our method $J_{th1}=0.005, J_{th2}=0.0025, J_{th3}=0.0025,$ and $J_{th4}=0.005$. Moreover, the residual evaluation functions $J_{r_1}(t), J_{r_2}(t), J_{r_3}(t)$ and $J_{r_4}(t)$ are shown in Fig. \ref{fig:8a}. The analyses are the same in Actuator Faults.
\section{Conclusions}
In this work, robust distributed observer-based on Fault Detection and Isolation for a network of heterogeneous MASs is investigated. By using the proposed methodology not only each agent's faults (Actuator faults, and Sensor faults) but also faults of the agent's nearest neighbors can also be detected and isolated. Furthermore, the relative model system is utilized by a vector of relative output measurements. Each agent is thus able to estimate not only its own states but also states of its nearest neighbors and the dimension of observer at each node is reduced. Based on two performance indexes $\mathcal H_2$ and $\mathcal H_{-}$, such as a set of linear matrix inequality conditions, sufficient conditions for solvability of the optimization problem were obtained.
\bibliographystyle{IEEEtran}
\section{Main Results}
There are three performance indices from I to III that must be satisfied simultaneously for solving the distributed FDI problem. We assume that the pair ($\mathcal{A}_i, \bar{\mathcal C}_i$) is detectable. The underlying idea adopted here for solving such optimization problems is that the multi-objective optimization problem can be reduced to a single optimization problem with constraints. To this end, the well-established robust control theory and LMI-techniques have been used. In the following theorem, a feasible solution to the distributed FDI problem is obtained by simultaneously considering these indices.
\begin{theorem}\label{thm1}
Consider the relative model system (\ref{2}) and distributed observer (\ref{3}). The closed-loop system will be stable and simultaneously satisfied problems I -- III, if there exits positive given scalar $\beta_1>0, \beta_2>0$ and positive matrices $N_i,P_i$ for the following optimization $\mathcal H_2/ \mathcal H_{-}$ problem
\begin{align}\label{5f}
\mathop{maximize}\limits_{N_i,P_i} \beta_2 \gamma_2-\beta_1 \gamma_1
\end{align}
subject to
\begin{align}
\begin{bmatrix}\label{10}
{{Q_{i}}}&{{N_i^T}}\\
{{*}}&{{P_{i}}}
\end{bmatrix} \geq &0\\
trace(Q_i) - \gamma_1^2 +trace (\mathcal{C}_i Y_i\bar{\mathcal C}_i^T)\leq& \\
\mathbf{A}_i P_i - N_i\bar{\mathcal C}_i +P_i\mathbf{A}_i^T -\bar{\mathcal C}_i^TN_i^T+ \bar{\mathcal C}^T \bar{\mathcal C}_i <& 0 \\
\begin{bmatrix}\label{5b}
\bar{\mathcal D}_{fi}^T \bar{\mathcal D}_{fi}-\gamma_2^2 I&{{{P}_i\mathbf{B}_{fi}-N_i\bar{\mathcal D}_{fi}+\bar{\mathcal C}_i^T \bar{\mathcal D}_{fi}}}\\
{{*}}&{{F_i}}
\end{bmatrix} >& 0
\end{align}
where $F_i=\mathbf{A}_iP_i - N_i\bar{\mathcal C}_i +P_i\mathbf{A}_i^T -\bar{\mathcal C}_i^TN_i^T+ \bar{\mathcal C}_i^T \bar{\mathcal C}_i, \mathbf{A}_i=\mathcal{A}_i - \textbf L_i \bar{\mathcal {C}}_i, \mathbf{B}_{fi}=\mathcal{B}_{fi} - \textbf L_i \bar{\mathcal {C}}_i$. The observer gain matrix is calculated as
\begin{align*}
L_i =\textbf L_i+P^{-1}_iN_i
\end{align*}
where $\textbf L_i=Y_i\bar{\mathcal C}_i^T+\mathcal{B}_{di} \bar{\mathcal D}_{di}^T$ and $\Delta L_i = P^{-1}_i N_i$.
\end{theorem}
\input{proof1}
The \eqref{10}--\eqref{5b} are non-convex optimization problem. Instead of finding $\gamma_1, \gamma_2$, the optimization problem can be solved with $\gamma_1^2$ and $\gamma_2^2$. Let set $\alpha_1=\gamma_1^2$ and $\alpha_2=\gamma_2^2$, the non-convex optimization problem \eqref{10}--\eqref{5b} become convex optimization form.
\input{residual}
\section{Problem Formulation}
\subsection{System description}
Consider a network of $N$ heterogeneous agents, where each agent is expressed by a linear dynamic model such as
\begin{align}\label{1}
\begin{aligned}
\dot x_i(t)&= {A}_{i}x_i(t)+{B}_{i}u_i(t)+{B}_{fi}f_i(t)+{B}_{di}d_i(t)\\
y_i(t)&= {C}_{i}x_i(t)+{D}_{fi}f_i(t)+{D}_{di}d_i(t)
\end{aligned}
\end{align}
where $x_i(t) \in \mathbb{R}^{n_i}$ denotes the sate vector, $f_i(t)\in \mathbb{R}^{n_{fi}}$ denotes the faults signal (sensor faults, actuator faults and process faults), $d_i(t) \in \mathbb{R}^{n_{di}}$ denotes the disturbance and noise signals or uncertain components, $u_i(t) \in \mathbb{R}^{n_{u_i}}$ denotes the control inputs, and $y_i(t) \in \mathbb{R}^{n_{y_i}}$ denotes the measured outputs for the agent $i^{th}$ with $n_{y_i} \ge n_{f_i}$. In addition, system matrices ${A}_{i}, {B}_{i}, {B}_{fi}, {B}_{di}, C_{i}, {D}_{fi}$ and ${D}_{di}$ are constant matrices. The fault matrices ${B}_{fi}$ and ${D}_{fi}$ are specified according to faults that are to be detected in the components, sensors, and actuators. In the rest of paper, we omit the term $"t"$ in $x_i,u_i,f_i,d_i, y_i$.
With the assumption that all agents have the same number of outputs, we assume that each agent not only measures the trivial absolute output signal $y_i$ but also is equipped with the sensors for the relative output measurements, that is $z_{ij} = y_i - y_j, j\in N_i = {i_1,i_2,...,i_{|Ni|}} \subseteq [1,N]$ denotes the set of agents that agent $i^{th}$ can sense ($i^{th}$ agent's neighbors). The communication topology among the $N$ agents is represented by an undirected graph $\mathcal G = (V,\mathcal E)$ \cite{Pham2019c}, consisting of the node set $V = \{1,2,...,N\}$ and the edge set $\mathcal E \in V \times V$. We are interest in, at each time instant, the information available to agent $i^{th}$, which is the relative measurement of other agents with respect to itself.
The vector relative output of agent $i^{th}$ is thus expressed \begin{align*}
z_i&=(z_{i_{i_1}},z_{i_{i_2}},\cdots,z_{i_{|Ni|}})^T\\
&=(y_i - y_{i_1},y_i - y_{i_2},\cdots, y_i - y_{i_{|Ni|}})^T
\end{align*}
The relative model for $i$ agent could be expressed
\begin{align}\label{2}
\dot{x}_{Ni} &= \mathcal{A}_i x_{Ni}+\mathcal{B}_{ui} u_{Ni}+\mathcal{B}_{fi} f_{Ni}+\mathcal{B}_{di} d_{Ni}\nonumber\\
z_i&= \bar{\mathcal C}_i x_{Ni}+\bar{\mathcal D}_{fi}f_{Ni}+\bar{\mathcal D}_{di}d_{Ni}
\end{align}
where $x_{Ni}=(x_i^T,\cdots,x_{i_{|Ni|}}^T)^T, u_{Ni}=(u_i^T,\cdots,u_{i_{|Ni|}}^T)^T$ denote a state and control input vectors. $d_{Ni}=(d_i^T,\cdots,d_{i_{|Ni|}}^T)^T, f_{Ni}=(f_i^T,\cdots,f_{i_{|Ni|}}^T)^T$ stands for a disturbance and fault vectors. $\mathcal{A}_i, \mathcal{B}_{ui}, \mathcal{B}_{fi}, \mathcal{B}_{di}$ and $ \bar{\mathcal C}_i, \bar{\mathcal D}_{fi}, \bar{\mathcal D}_{di}$ are defined with the following notation.
For given matrices $H_i^{n_i \times n_i}, H_{i1}^{n_{i1} \times n_{i1}},...,H_{i_{|Ni|}}^{n_{i_{|Ni|}} \times n_{i_{|Ni|}}}$, where $\mathcal H_i$ denote the $\mu_i \times \xi_i$ matrices
\begin{align*}
\mathcal H_i &= \left[ \begin{matrix}
{{{H}_{i}}}&0&0&...&0\\
0& {{{H}_{i_1}}}&0&...&0\\
\vdots &\vdots&\vdots&\ddots&0\\
0&0&0&...&{{{H}_{i_{|Ni|}}}}
\end{matrix} \right]\\
\bar{\mathcal H}_i &= \left[ \begin{matrix}
{{{H}_{i}}}&{{-{H}_{i_1}}}&0&...&0\\
\vdots &\vdots&\vdots&\ddots&0\\
{{{H}_{i}}}&0&0&...&{{-{H}_{i_{|Ni|}}}}
\end{matrix} \right]
\end{align*}
where $\mu_i = n_i +\sum_{j=1}^{|Ni|}n_{ij}$ and $\xi_i =\sum_{j=1}^{|Ni|}m_{ij}$. $\mathcal{A}_i \in \mathbb{R}^{\mu_i \times \mu_i}, \mathcal{B}_{ui} \in \mathbb{R}^{\mu_i \times \mu_{ui}}, \mathcal{B}_{di} \in \mathbb{R}^{\mu_i \times \mu_{di}}$ and $\mathcal{B}_{fi} \in \mathbb{R}^{\mu_i \times \mu_{fi}}$ where $\mu_i = n_i +\sum_{j=1}^{|Ni|}n_{ij}, \mu_{ui} = n_{ui} +\sum_{j=1}^{|Ni|}n_{u_{ij}}, \mu_{fi} = n_{fi} +\sum_{j=1}^{|Ni|}n_{f_{ij}}$ and $\mu_{di} = n_{di} +\sum_{j=1}^{|Ni|}n_{d_{ij}}$. The matrices $\bar{\mathcal C_i} \in \mathbb{R}^{\xi_{yi} \times \mu_i}, \bar{\mathcal D}_{di} \in \mathbb{R}^{\xi_{yi} \times \xi_{di}}$ and $\bar{\mathcal D}_{fi} \in \mathbb{R}^{\mu_{yi} \times \xi_{fi}}$ where $\xi_i =\sum_{j=1}^{|Ni|}m_{ij}, \xi_{yi} =\sum_{j=1}^{|Ni|}m_{y_{ij}}, \xi_{di} =\sum_{j=1}^{|Ni|} m_{d_{ij}}$ and $\xi_{fi} =\sum_{j=1}^{|Ni|} m_{f_{ij}}$.
In this paper, we use the following version of the bounded real lemma.
\begin{lemma}\cite{Ding}\label{l1}
Given system $ C(sI- A+{LC})^{-1}( B_d-{LD}_d)$ and suppose that $( C, A)$ is detectable, $ D_d$ has full row rank with $ D_d D_d^T=I$ and $rank \begin{bmatrix}
{{{ A- j \omega I}}}&{{ B_d}}\\
{{ C}}&{{ D_d}}
\end{bmatrix}$ has full rank for all $\omega \in [0, \infty ]$, then the minimum
\begin{align*}
\mathop {min}\limits_{ L}\left \| { C} (sI-{{ A}}+{ L} { C})^{-1}({{ B}}_{d}-{ L} { D}_{d}) \right \|_2 =trace ({ C} Y { C}^T)^{\frac {1} {2}}
\end{align*}
is achieved by
$ L = Y C^T +{ B}_{d} { D}_{d}^T$
and matrix $ Y \ge 0$ solves the Riccati equation
\begin{align*}
&({ A}-{ B}_{d} { D}_{d}^T { C}) Y + Y({ A}-{ B}_{d} { D}_{d}^T { C})^T - Y { C}^T { C} Y \nonumber\\
&+{ B}_{d} { B}_{d}^T -{ B}_{d} D^T_{d}{ D}_{d} { B}_{d}^T =0
\end{align*}
\end{lemma}
\begin{lemma}\cite{Guang-RenDuan2013} \label{l2}
We take into account $G(s):=( A, B, C, D)$, $ A$ is stable.
$G(s) \in \mathcal{RH}_{\infty}^{m \times k}$ being injective $\forall \omega$
\begin{align*}
rank \begin{bmatrix}
{{{ A- j \omega I}}}&{{ B}}\\
{{ C}}&{{ D}}\\
\end{bmatrix}=n+k; {DD}^T-\gamma^2 I >0
\end{align*}
Then $\left \| G(s) \right \|_{-} > \gamma$ if only if $ X= X^T$ such that
\begin{align*}
&XA+ A^TX+ C^T C\\
&+(XB+ C^T D)(\gamma^2 I- D^T D)^{-1}( B^T X+ D^T C)>0
\end{align*}
\end{lemma}
\subsection{Distributed Observer}
A distributed observer for the $i$ agent, based on the relative model, is shown as
\begin{align}\label{3}
\begin{aligned}
\dot{\hat{x}}_{Ni} &= \mathcal{A}_i \hat x_{Ni}+\mathcal{B}_{ui}u_{Ni}+{L}_i(z_i-\hat z_i)\\
\hat z_i&= \bar{\mathcal C}_i\hat x_{Ni}
\end{aligned}
\end{align}
and the residual generator in distributed FDI system should described the inconsistency between the actual system variables and the mathematical model. It could be realized as a composition of state observer, and responses to faults, disturbances and modeling error
\begin{align}\label{3a}
\begin{aligned}
r_i&=z_i-\hat z_i\\
&=\bar{\mathcal C}_i e_{Ni}+\bar{\mathcal D}_{fi}f_{Ni}+\bar{\mathcal D}_{di}d_{Ni}
\end{aligned}
\end{align}
where $\hat x_{Ni} \in \mathbb{R}^{\mu_i}$ denote the state estimation derived by the observer. $L_i \in \mathbb{R}^{\mu_i \times \xi_{i} }$ is the observer gain matrices, and is to be determined.
We define a state estimation error between observer (\ref{3}) and the relative model (\ref{2}) as $e_{Ni}=x_{Ni}-\hat x_{Ni}$. It follows that the dynamic of $e_{Ni}$ can be re-expressed as
\begin{align}\label{4}
\dot e_{Ni} &=(\mathcal{A}_i-{L}_i \bar{\mathcal C}_i) e_{Ni} +(\mathcal{B}_{di}-{L}_i \bar{\mathcal D}_{di}) d_{Ni}\notag\\
&+(\mathcal{B}_{fi}-{L}_i \bar{\mathcal D}_{fi}) f_{Ni}\\
r_i&=\bar{\mathcal C}_i e_{Ni}+\bar{\mathcal D}_{fi}f_{Ni}+\bar{\mathcal D}_{di}d_{Ni})\label{4b}
\end{align}
By taking the Laplace transforms, it is easy to show that
\begin{align}\label{4a}
r_i=T_{r_{i}d_{Ni}}d_{Ni}+T_{r_{i}f_{Ni}}f_{Ni}
\end{align}
where
\begin{align*}
&[T_{r_{i}d_{Ni}}T_{r_{i}f_{Ni}}]\\
&=
\left[\begin{array}{c|c}
\mathcal{A}_i-{L}_i \bar{\mathcal C}_i & \mathcal{B}_{di}-{L}_i \bar{\mathcal D}_{di} \mathcal{B}_{fi}-{L}_i \bar{\mathcal D}_{fi} \\ \hline
\bar{\mathcal C}_i & \bar{\mathcal D}_{di}\;\;\;\;\;\;\;\;\;\;\;\;\bar{\mathcal D}_{fi} \\
\end{array}\right]
\end{align*}
are the transfer matrices from disturbances and faults to residual, respectively.
The proposed distributed FDI problem is now to answer the question \textit{"How each agent can detect and isolate not only both its own faults and simultaneously estimate states, but also faults and states of its neighbors using the relative outputs $z_i - z_{i_{i1}},...,z_i - z_{i_{Ni}}$"}. Moreover, it is easy from (\ref{4a}) that if we ignore the faults $f_{Ni}$, the residual $r_i$ only depends on the disturbances $d_{Ni}$ and the optimization problem such as the $\mathcal H_2$ filtering problem (the Kalman filtering). Similarly, if we disregard $d_{Ni}$ then the optimization problem is the $\mathcal H_{-}$ filtering problem.
We would like to solve the following multi-criterion optimization problem
\begin{align*}
I.& \;\left \| T_{r_i d_{Ni}} \right \|_{2} < \gamma_1\\
II.& \;\left \| T_{r_i f_{Ni}} \right \|_{-} > \gamma_2\\
III.& \;(\mathcal{A}_i-{L}_i \bar{\mathcal C}_i) \;\text{is \;stable}
\end{align*}
The first and second constraints ensure a trade-off between the robustness against the disturbances $d_{Ni}$ and the sensitivity for the faults $f_{Ni}$. The third condition ensures that the closed loop system (the relative model (\ref{2}) and distributed observer (\ref{3})) is stable and ensures dynamic errors of state estimation converges. To do this, we need to find $L_i$ such that satisfies multi-criterion optimization problem from I to III. We propose $L_i = \textbf L_i+\Delta L_i$, where $\textbf L_i$ is solution of Riccati equation in \textit{Lemma \ref{l1}}, and instead of finding $L_i$, we now need to find $\Delta L_i$ such as
\begin{align}\label{a}
&\;\;\;\;\;\;\;\;\;\;\;\;\;\mathop {maximize} \beta_2 \gamma_2-\beta_1\gamma_1\\
&subject \;\;to\notag\\
&\hspace*{1cm}I. \;\left \| T_{r_i d_{Ni}} \right \|_{2} < \gamma_1\notag\\
&\hspace*{1cm}II. \;\left \| T_{r_i f_{Ni}} \right \|_{-} > \gamma_2\notag\\
&\hspace*{1cm}III. \;[\mathcal{A}_i-(\textbf L_i+\Delta L_i) \bar{\mathcal C_i}] \;\text{is \;stable}\notag
\end{align}
We can see that the fault detection and robust against to disturbance can be improved by maximizing $\gamma_2$ and minimizing $\gamma_1$. Moreover, $\beta_1, \beta_2$ depends on designer, who want to emphasis the role of fault detection or of robust against the disturbance.
\subsection{Residual evaluation and threshold setting}
Following the generation of the residuals $r_i(t) \in \mathbb{R}^{\xi_{fi}}, \forall i \in V$, the next step in the distributed FDI methodology is to figure out the threshold function $J_{{th}_i}$ and the evaluation function $J_{{r}_i}(t)$. In this work, the threshold function $J_{{r}_i}$ and the residual evaluation $J_{{th}_i}$ are selected as
\begin{align} \label{17}
J_{th_i}&=\sup_{f_{Ni}=0, u_{Ni}, d_{Ni}}\| J_{r_i}(t)\|_{\infty}\\
J_{r_{ij_k}}(t) &= \|r_{ij_k}(t)\|_{2,L}\nonumber \\
&=\left[1/L \sum_{s=t-L}^{t} r^T_{ij_k}(s)r_{ij_k}(s)\right]^{1/2}\label{17a}
\end{align}
where $r_{ij_k}(t)$ and $J_{r_{ij_k}}(t)$ are the $k^{th}, k= 1,...,\mu_{yi}$ elements of $j^{th}$ elements (neighbor agents) of the residual signal $r_i(t)$ and evaluation function $J_{r_i}(t)$, and $L$ is the evaluation finite time window. The occurrence of a fault can then be detected by using the following decision logic: If $J_{r_i}(t) > J_{th_i}$, then $f_i(t) \ne 0$ or $f_j(t) \ne 0, j \in N_i$.
\subsection{Distributed fault isolation}
The final step, to determine the faulty agent in the team since a fault is detected. Based on the fault isolation techniques \cite{Davoodi2016}, the distributed strategy is proposed derived from flags, which are generated corresponding to the residual signal of the team agents. The so-called flags $[\epsilon_{ij}]_k, j \in V, j \in \{i,N_i\}$ in which $[\epsilon_{ij}]_k \in \mathbb R^{\mu_{yi}}$ is row vector and the $ij_k^{th}$ flag is 1 if $J_{r_{ij_k}}(t) > J_{th_i}$ and 0 otherwise. Consequently, it is assume that each agent constructs the defined fault pattern
\begin{align}\label{29}
\Upsilon_i=\left\{[\epsilon_{ij}]_k: j \in \{i,N_i\}, i \in V\right\}
\end{align}
Finally, the faulty agent can be isolated according to the following algorithm 1.
\begin{algorithm}
\SetAlgoLined
\caption{A distributed fault isolation procedure}
1. Calculate $J_{r_{ij_k}}(t)$, $J_{th_i}$ in \eqref{17} and \eqref{17a}.
2. Determine the flags and fault pattern $\Upsilon_i$ in \eqref{29}.
\If{all of elements of $\Upsilon_i=1$ (i.e., $[\epsilon_{ij}]_k =1$)}{
The $i^{th}$ agent is faulty
}
\If{only one element of $\Upsilon_i=1$}{
$j^{th}$ neighbor of agent $i^{th}$ is faulty
}
\If{$\Upsilon_{q_1}=\Upsilon_{q_n}=1, q_n \le i$ and the first element of $\Upsilon_{p_1},..,\Upsilon_{p_m}= 1 < q_n \ne p_m \le i$ corresponding to flags $[\epsilon_{ij}]_k$} {the fault will occur on the neighbor of agent $j$: It is agent $i$
}
\label{Al1}
\end{algorithm} |
\section{Exponential weight mixing}\label{sec: exp weights}
\begin{proof}[Proof of Theorem~\ref{thm:exp weights}]
From \eqref{eq:exp weights}, we have
\begin{eqnarray*}
\pi_r&\propto&\exp\left(-\sum_{(i,j)\in\Omega} \left(A_{ij}-\hat{P}^{(r)}_{ij}\right)^2\right)\\
&\propto&\exp\left(-\sum_{(i,j)\in\Omega} \left(A_{ij}-\hat{P}^{(r)}_{ij}\right)^2+\sum_{(i,j)\in\Omega} (A_{ij}-P_{ij})^2\right) =: \exp\left(-S^{(r)}\right),
\end{eqnarray*}
where $\propto$ denotes ``proportional to''.
Rewrite $S^{(r)}$ as follows:
\begin{eqnarray*}
S^{(r)} = \sum_{(i,j)\in\Omega}\left(P_{ij}-\hat{P}^{(r)}_{ij}\right)\left(2A_{ij}-P_{ij}-\hat{P}^{(r)}_{ij}\right)=:\sum_{(i,j)\in\Omega} X_{ij}^{(r)}.
\end{eqnarray*}
Denote $\rho = \max_{i,j}P_{ij}$.
Conditioning on $A(\Omega^c)$ (thus on $\hat{P}^{(r)}$), $S^{(r)}$ is the sum of independent random variables $X_{ij}^{(r)}$ with mean and variance given by
\begin{eqnarray*}
\mathbb{E} X_{ij}^{(r)} = \left(P_{ij}-\hat{P}^{(r)}_{ij}\right)^2, \quad
\text{var}\left(X_{ij}^{(r)}\right) = 4P_{ij}(1-P_{ij})\left(P_{ij}-\hat{P}^{(r)}_{ij}\right)^2 \le 4\rho\left(P_{ij}-\hat{P}^{(r)}_{ij}\right)^2.
\end{eqnarray*}
Since $\big|X_{ij}^{(r)}\big|\le 8$, by Bernstein's inequality,
\begin{eqnarray*}
\mathbb{P}\left(\big|S^{(r)}-\mathbb{E} S^{(r)}\big|>t\right)\le 2\exp\left(\frac{-t^2/2}{4\rho\displaystyle \sum_{(i,j)\in\Omega}\left(P_{ij}-\hat{P}^{(r)}_{ij}\right)^2+8t/3}\right).
\end{eqnarray*}
Since $m = o(n^2)$, we have
\begin{eqnarray}\label{eq:concentration S}
\Big|S^{(r)}-\sum_{(i,j)\in\Omega}\left(P_{ij}-\hat{P}_{ij}^{(r)}\right)^2\Big| \le C\left(\rho\log n \sum_{(i,j)\in\Omega} (P_{ij}-\hat{P}^{(r)}_{ij})^2\right)^{1/2}+C\log n
\end{eqnarray}
for all $1\le r \le m$ with high probability for a sufficiently large constant $C$. In other words, $S^{(r)}$ concentrates well around the mean $\sum_{(i,j)\in\Omega}\big(P_{ij}-P_{ij}^{(r)}\big)^2$, and the mixture of $P^{(r)}$ puts exponentially large weight on the estimator $P^{(r)}$ with the smallest error $\|P^{(r)}(\Omega)-P(\Omega)\|_F^2$.
\medskip
Let $r^* = \mathrm{argmin}_{1\le r \le m}\|P^{(r)}(\Omega)-P(\Omega)\|_F^2$ and denote
$$x_r=\|P^{(r)}(\Omega)-P(\Omega)\|_F, \quad x=\|P^{(r^*)}(\Omega)-P(\Omega)\|_F.
$$
Consider the index set
$$
I=\left\{r\in [m]:x_r\le x+2C\sqrt{\rho\log n}+ \frac{2C\log(nm)}{x}\right\}.
$$
By the triangle inequality,
\begin{eqnarray*}
\Big\|\sum_{r=1}^m \pi_r P^{(r)}(\Omega)-P(\Omega)\Big\|_F \le \sum_{r=1}^m \pi_r \|P^{(r)}(\Omega)-P(\Omega)\|_F
= \sum_{r\in I} \pi_r x_r + \sum_{r\not\in I} \pi_r x_r.
\end{eqnarray*}
From the definition of $I$ we get
\begin{eqnarray*}
\sum_{r\in I} \pi_r x_r \le x+2C\sqrt{\rho\log n}+\frac{2C\log(nm)}{x}.
\end{eqnarray*}
For the second sum, consider $r\not\in I$. Recall that $\pi_r\propto \exp(-S^{(r)})$ and by \eqref{eq:concentration S}, $S^{(r)}$ and $S^{(r^*)}$ concentrate around $x_r^2$ and $x^2$, respectively. Therefore with high probability we have
\begin{eqnarray*}
\pi_r = \frac{\exp(-S^{(r)})}{\sum_r \exp(-S^{(r)})} &\le& \exp\left(S^{(r*)}-S^{(r)}\right) \le \exp\left(x^2-x_r^2+2C\sqrt{\rho\log n}(x+x_r)+2C\log n\right).
\end{eqnarray*}
Since $r\not\in I$,
\begin{eqnarray*}
x_r^2-x^2-2C\sqrt{\rho\log n}(x_r+x) = (x_r+x)(x_r-x-2C\sqrt{\rho\log n}) \ge 4C\log(nm),
\end{eqnarray*}
and consequently,
\begin{eqnarray*}
\pi_r \le \exp\left(-2C\log(nm)\right)
= \left(nm\right)^{-2C}.
\end{eqnarray*}
Since $x_r\le n$, by choosing $C>1$, we get
$$
\sum_{r\not\in I} \pi_r \|P^{(r)}(\Omega)-P(\Omega)\|_F \le \sum_{r\not\in I} n(nm)^{-2C} = o(1).
$$
In summary, we have proved that with high probability,
\begin{eqnarray*}
\Big\|\sum_{r=1}^m \pi_r P^{(r)}(\Omega)-P(\Omega)\Big\|_F \le \min_{1\le r\le m}\big\|P^{(r)}(\Omega)-P(\Omega)\big\|_F+ \delta,
\end{eqnarray*}
where
\begin{eqnarray*}
\delta = 3C\Big(\log n\cdot\max_{i,j}P_{i,j}\Big)^{1/2}+\frac{3C\log(nm)}{\min_{1\le r\le m}\|P^{(r)}(\Omega)-P(\Omega)\|_F}.
\end{eqnarray*}
In case $\min_{1\le r\le m}\|P^{(r)}(\Omega)-P(\Omega)\|_F\le 1$, we can replace $\log(nm)/x$ in the definition of $I$ by $\log^{1/2}(nm)$ and repeat the above argument. Again, for constant $C>1$, the same derivations can still go through and the resulting error would be
$$\delta = 3C\Big(\log n\cdot\max_{i,j}P_{i,j}\Big)^{1/2}+3C\log^{1/2}(nm).
$$
Consequently, the error can be improved as follows:
\begin{eqnarray*}
\delta = 3C\Big(\log n\cdot\max_{i,j}P_{ij}\Big)^{1/2}+3C\min\left\{\frac{\log(nm)}{\min_{1\le r\le m}\|P^{(r)}(\Omega)-P(\Omega)\|_F}, \ \log^{1/2}(nm)\right\}.
\end{eqnarray*}
Rewriting $3C$ as a new constant $C$ completes the proof.
\end{proof}
\section{Non-negative linear aggregation}
\begin{proof}[Proof of Theorem~\ref{thm:nnls}]
Denote $E = A(\Omega)-P(\Omega)$. By the triangle inequality,
\begin{eqnarray}\label{eq:nnl ineq}
\|\Pi_\mathcal{C} A(\Omega)-P(\Omega)\|_F \le \|\Pi_\mathcal{C}E\|_F + \|\Pi_\mathcal{C}P(\Omega)-P(\Omega)\|_F.
\end{eqnarray}
We now show that $\|\Pi_\mathcal{C}E\|_F$ is small due to condition \eqref{eq:correlated assumption}.
Since $\mathcal{C}$ is a cone, whether $\|\Pi_\mathcal{C}E\|_F=0$ or $\|\Pi_\mathcal{C}E\|_F>0$, depending on the relative location of $E$ to $\mathcal{C}$. If $\|\Pi_\mathcal{C}E\|_F>0$ then
there exists $\nu\in\mathcal{C}$ with $\|\nu\|_F=1$ such that $\|\Pi_\mathcal{C}E\|_F=\langle\nu,E\rangle$.
Let $\nu^{(r)}$ be the normalized predictor
$$\nu^{(r)} = \|\hat{P}^{(r)}(\Omega)\|_F^{-1}\hat{P}^{(r)}(\Omega)$$
and $\nu = \sum_{r=1}^m \lambda_r\nu^{(r)}$ for some $\lambda_1,...,\lambda_m\ge 0$. Then by \eqref{eq:correlated assumption},
\begin{eqnarray*}
1 = \|\nu\|_F^2 = \sum_{1\le r,s\le m}\lambda_r\lambda_s\langle\nu^{(r)},\nu^{(s)}\rangle \ge \delta\Big(\sum_{r=1}^m\lambda_r\Big)^2,
\end{eqnarray*}
which implies $\sum_{r=1}^m \lambda_r \le \delta^{-1/2}$. Therefore
\begin{eqnarray*}
\|\Pi_\mathcal{C} E\|_F = \langle \nu,E\rangle = \sum_{r=1}^m \lambda_r \langle\nu^{(r)},E\rangle \le \delta^{-1/2}\max_{1\le r\le m} \langle\nu^{(r)},E\rangle.
\end{eqnarray*}
For each $\nu^{(r)}=\big(\nu^{(r)}_{ij}\big)$ with $\|\nu^{(r)}\|:=\max_{(i,j)\in\Omega}|\nu^{(r)}_{ij}|$, by Bernstein's inequality,
\begin{eqnarray*}
\mathbb{P}\big(|\langle\nu^{(r)},E\rangle|>t\big)&\le& \exp\left(\frac{-t^2/2}{\sum_{(i,j)\in\Omega} \big(\nu_{ij}^{(r)}\big)^2 \text{var}(E_{ij})+\|\nu^{(r)}\|_\infty t/3}\right) \\
&\le& \exp\left(\frac{-t^2/2}{\|P(\Omega)\|_\infty+\|\nu^{(r)}\|_\infty t/3}\right).
\end{eqnarray*}
Choosing $t \approx \max_{1\le r \le m}\|\nu^{(r)}\|_\infty\log(n+m) +(\|P(\Omega)\|_\infty\log(n+m))^{1/2}$ and applying the union bound, we obtain that with high probability,
\begin{eqnarray*}
\|\Pi_\mathcal{C}E\|_F \le C\delta^{-1/2}\left(\max_{1\le r \le m}\|\nu^{(r)}\|_\infty\log(n+m) +(\|P(\Omega)\|_\infty\log(n+m))^{1/2}\right).
\end{eqnarray*}
The proof is complete.
\end{proof}
\begin{proof}[Proof of Corollary~\ref{cor:comparison}]
Denote $E = A(\Omega)-P(\Omega)$. We can view $E$ and $\Pi_\mathcal{L}$ as a vector and a matrix in $\mathbb{R}^{|\Omega|}$ and $\mathbb{R}^{|\Omega|\times|\Omega|}$, respectively.
Then $\|\Pi_\mathcal{L}A(\Omega)-\Pi_\mathcal{L}P(\Omega)\|_F^2 = \|\Pi_\mathcal{L}E\|_F^2
$. Denoting by $\circ$ the entry-wise product, we get
\begin{eqnarray}
\label{eq:expected norm}\mathbb{E} \|\Pi_\mathcal{L}E\|_F^2 &=& \mathbb{E} \langle\Pi_\mathcal{L}E,\Pi_\mathcal{L}E\rangle = \text{trace}(\Pi_\mathcal{L}\text{diag}(P(\Omega)\circ (I-P(\Omega)))\Pi_\mathcal{L}^T)
\end{eqnarray}
So we have
\begin{equation}\label{eq:relation1}
m\sqrt{\rho} = m\cdot\min_{(i,j)\in\Omega} P_{ij}\cdot\Big(1-\max_{(i,j)\in\Omega} P_{ij}\Big)\le \mathbb{E} \|\Pi_\mathcal{L}E\|_F^2 \le m\cdot\max_{(i,j)\in\Omega}P_{ij}.
\end{equation}
Furthermore, by the Cauchy–Schwarz inequality,
\begin{equation}\label{eq:relation2}
\left(\mathbb{E} \|\Pi_\mathcal{L}E\|_F\right)^2 \le \mathbb{E} \|\Pi_\mathcal{L}E\|_F^2\le m\cdot\|P(\Omega)\|_\infty.
\end{equation}
Since $\|\Pi_\mathcal{L}\|=1$, by Hanson-Wright inequality \cite[Theorem 1.1]{Klochkov&Zhivotovskiy.hanson.wright.2020} and \eqref{eq:relation1}, for any $t\ge \max\{\mathbb{E} \|\Pi_\mathcal{L}E\|_F,1\}$ we have
\begin{eqnarray*}
\mathbb{P}\left(\big|\|\Pi_\mathcal{L}E\|_F^2-\mathbb{E} \|\Pi_\mathcal{L}E\|_F^2\big|>t\right) &\le& \exp\left(-c\min\left\{\frac{t^2}{\left(\mathbb{E} \|\Pi_\mathcal{L}E\|_F\right)^2},t\right\}\right) \\
&\le& \exp\left(-c\min\left\{\frac{t^2}{ \mathbb{E} \|\Pi_\mathcal{L}E\|_F^2},t\right\}\right).
\end{eqnarray*}
Choosing $t=\left(m\log n\cdot\|P(\Omega)\|_\infty\right)^{1/2}$ and using \eqref{eq:relation2}, we see that
$$t > \max\{\mathbb{E} \|\Pi_\mathcal{L}E\|_F,1\}. $$
Then \eqref{eq:relation1} leads to
\begin{equation}\label{eq:ols-upper}
\|\Pi_\mathcal{L}E\|_F^2 \le m\cdot\max_{(i,j)\in\Omega}P_{ij} + \left(m\log n\cdot\|P(\Omega)\|_\infty\right)^{1/2}
\end{equation}
with high probability. Combining the assumption $m\ge 4\rho^{-1}\norm{P(\Omega)}_{\infty}\log n$ and \eqref{eq:relation1}, we can see that
$$\mathbb{E} \|\Pi_\mathcal{L}E\|_F^2 \ge m\sqrt{\rho} > 2\sqrt{m\norm{P(\Omega)}_{\infty}\log{n}} =2t > t$$
and therefore with high probability
\begin{equation}\label{eq:ols-lower}
\|\Pi_\mathcal{L}E\|_F^2 \ge m\cdot\min_{(i,j)\in\Omega} P_{ij}\cdot\big(1-\|P(\Omega)\|_\infty\big) - \left(m\log n\cdot\|P(\Omega)\|_\infty\right)^{1/2} > \frac{m\sqrt{\rho}}{2}.
\end{equation}
Notice that \eqref{eq:ols-upper} directly indicates \eqref{eq:ols-oracle}. Now we have
\begin{eqnarray*}
\Delta &=& \norm{\hat{P}^{(\text{ols})}(\Omega) - P(\Omega)}_F^2 - \norm{\hat{P}^{(\text{nnl})}(\Omega) - P(\Omega)}_F^2\\
&=& \norm{\Pi_{\mathcal{L}}A(\Omega) - \Pi_{\mathcal{L}}P(\Omega)}_F^2 + \norm{\Pi_{\mathcal{L}}P(\Omega)-P(\Omega)}_F^2 \\
&& -\norm{\Pi_{\mathcal{C}}A(\Omega) - \Pi_{\mathcal{L}}P(\Omega)}_F^2- \norm{\Pi_{\mathcal{L}}P(\Omega) - P(\Omega)}_F^2\\
&\ge& \norm{\Pi_{\mathcal{L}}E}_F^2 - \norm{\Pi_{\mathcal{C}}E }_F^2 - \norm{\ \Pi_{\mathcal{C}}\Pi_{\mathcal{L}}P(\Omega)-\Pi_{\mathcal{L}}P(\Omega)}_F^2.
\end{eqnarray*}
On the other hand, by Theorem~\ref{thm:nnls} and \eqref{eq:ols-lower},
\begin{align*}
\Delta \ge \frac{m\rho^{1/2}}{2} - \frac{C^2}{\delta}\epsilon^2- \|\Pi_\mathcal{C}\Pi_\mathcal{L} P(\Omega)-\Pi_\mathcal{L} P(\Omega)\|_F^2
\end{align*}
with high probability, and the proof is completed by using the basic inequality $(a+b)^2 \le 2(a^2+b^2)$ for $\epsilon^2$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:self-regularizing}]
We will mainly follow the proof strategy of \cite{Slawski.NNLS.2011}. However, there is one key step in their proof (right after B.5) that may not go through. So our proof can be seen as a corrected version of that.
Recall the definition of $\hat{\beta}$ in \eqref{eq:nnls optimization}. Since this is a constrained linear regression problem, for the notation simplicity, let us denote $X_r = \hat{P}^{(r)}(\Omega)$, $Y=A(\Omega)$, $\mu = P(\Omega)$, and $E =A(\Omega)-P(\Omega)= Y-\mu$. We view them as column vectors and further denote $X=(X_1,...,X_m)$. The optimization problem \eqref{eq:nnls optimization} is equivalent to
\begin{eqnarray}\label{eq:nnls linear form}
\hat{\pi} = \mathrm{argmin}_{\pi \succeq 0} \|Y-X\pi\|^2.
\end{eqnarray}
Consider the oracle parameter
$$\pi^*=\mathrm{argmin}_{\pi\succeq 0}\|\mu-X\pi\|^2$$
and define
$\quad \delta=\pi^*-\pi$ for any $\pi \succeq 0$. In particular, we write $\hat{\delta}=\pi^*-\hat{\pi}.$
Our goal is to compare the NNL mixing estimate $X\hat{\pi}$ with the best linear approximation $X\pi^*$ in the noiseless setting when $\mu$ is known. We first rewrite the objective function in \eqref{eq:nnls linear form} as follows:
\begin{eqnarray*}
\nonumber\|Y-X\pi\|^2 &=& \|\mu - X\pi^* + X(\pi^*-\pi) +E\|^2\\
&=& \|\mu - X\pi^*\|^2+\|X\delta\|^2 + \|E\|^2 +2E^T(\mu-X\pi^*) +2E^TX\delta.
\end{eqnarray*}
Note that the constraint $\pi\succeq 0$ is equivalent to $\delta\preceq\pi^*$, and only the second and the last terms of the expression above depend on $\delta$, it follows that
\begin{eqnarray}\label{eq:nnls delta}
\hat{\delta} = \mathrm{argmin}_{\delta\preceq \pi^*}\left\{\|X\delta\|^2+2E^TX\delta\right\}.
\end{eqnarray}
Since the objective function on the right-hand side is zero when $\delta=0$, its value at the minimizer $\delta=\hat{\delta}$ is at most zero. Equivalently, we have
\begin{equation*}
\|X\hat{\delta}\|^2 \le 2E^TX\hat{\delta}.
\end{equation*}
For a vector $x\in\mathbb{R}^m$, denote
$$S_+(x)=\{i\in[m]:x_i\ge 0\}, \quad S_{-}(x)=\{i\in[m]:x_i< 0\}.$$
Let $x_P$ be the vector obtained from $x$ by setting all entries within $S_{-}(x)$ to zero and $x_N=x-x_P$. The inequality above implies
\begin{equation}\label{eq:Xdelta bound}
\|X\hat{\delta}\|^2 \le2\|E^TX\|_\infty\|\hat{\delta}\|_1 \le 2\|E^TX\|_\infty\big(\|\hat{\delta}_P\|_1+\|\hat{\delta}_N\|_1\big) \le 2\|E^TX\|_\infty\big(\|\pi^*\|_1+\|\hat{\delta}_N\|_1\big),
\end{equation}
where $\|\hat{\delta}_P\|_1\le \|\pi^*\|_1$ because $\hat{\delta}=\pi^*-\hat{\pi}\le \pi^*$ for $\hat{\pi}\succeq 0$. It remains to bound $\|\hat{\delta}_N\|_1$.
Denote $\Sigma=X^TX$. For each $\delta\le \pi^*$, let $\Sigma_{PP}$, $\Sigma_{NN}$ and $\Sigma_{NP}$ be the matrices obtained from $\Sigma$ by setting all entries with indices $(i,j)$ outside $S_+(\delta)\times S_+(\delta)$, $S_{-}(\delta)\times S_{-}(\delta)$ and $S_{-}(\delta)\times S_{+}(\delta)$ to zero, respectively.
Then \eqref{eq:nnls delta} is equivalent to
\begin{eqnarray*}
\hat{\delta}=\mathrm{argmin}_{\delta\preceq \pi^*}\left\{\delta_P^T\Sigma_{PP}\delta_P+\delta_N^T\Sigma_{NN}\delta_N+2\delta_N^T\Sigma_{NP}\delta_P+2E^TX(\delta_N+\delta_P)\right\}.
\end{eqnarray*}
Define $\mathcal{T} = \{x \in \mathbb{R}^m: x\le\pi^*, x_i \le 0, i \in S_-(\hat{\delta}), x_i=0, i\in S_+(\hat{\delta})\}$. Replacing $\delta_P$ in the objective function above on the right-hand side with $\hat{\delta}_P$, we see that
\begin{eqnarray*}
\hat{\delta}_N = \mathrm{argmin}_{\delta_N\in\mathcal{T}}\left\{\delta_N^T\Sigma_{NN}\delta_N+2\delta_N^T\Sigma_{NP}\hat{\delta}_P+2E^TX\delta_N\right\}.
\end{eqnarray*}
Since the objective function on the right-hand side is zero when $\delta_N=0\in\Omega$,
\begin{eqnarray*}
\hat{\delta}_N^T\Sigma_{NN}\hat{\delta}_N+2\hat{\delta}_N^T\Sigma_{NP}\hat{\delta}_P+2E^TX\hat{\delta}_N \le 0.
\end{eqnarray*}
Denote $\|\Sigma\|_\infty=\max_{1\le r,s\le m}|\Sigma_{rs}|$. By the self-regularizing property \eqref{eq:self-reg}, we have
\begin{eqnarray*}
0 &\ge& \kappa\|\hat{\delta}_N\|_1^2\|\Sigma\|_\infty -2\|\hat{\delta}_N\|_1\|\hat{\delta}_P\|_1\|\Sigma\|_\infty+2E^TX\hat{\delta}_N \\
&\ge& \kappa\|\hat{\delta}_N\|_1^2\|\Sigma\|_\infty -2\|\hat{\delta}_N\|_1\|\hat{\delta}_P\|_1\|\Sigma\|_\infty-2\|E^TX\|_\infty\|\hat{\delta}_N\|_1.
\end{eqnarray*}
It follows that
\begin{eqnarray}\label{eq:delta-N-bound}
\|\hat{\delta}_N\|_1 \le \frac{2}{\kappa}\left(\|\hat{\delta}_P\|_1+\frac{\|E^TX\|_\infty}{\|\Sigma\|_\infty}\right)\le \frac{2}{\kappa}\left(\|\pi^*\|_1+\frac{\|E^TX\|_\infty}{\|\Sigma\|_\infty}\right).
\end{eqnarray}
From \eqref{eq:Xdelta bound} and \eqref{eq:delta-N-bound}, we get
\begin{eqnarray*}
\|X\hat{\pi}-X\pi^*\|^2 \le \frac{4\|E^TX\|_{\infty}^2}{\kappa \|\Sigma\|_\infty}+\frac{(2\kappa+4)\|E^TX\|_\infty\|\pi^*\|_1}{\kappa}.
\end{eqnarray*}
By Bernstein's inequality on the Bernoulli error $E$, for each $1\le r\le m$, we have
\begin{eqnarray*}
\mathbb{P}\left(|E^TX_r|>t\right) \le 2\exp\left(\frac{-t^2/2}{\|P(\Omega)\|_\infty\|\hat{P}^{(r)}(\Omega)\|_F^2+t\|\hat{P}^{(r)}(\Omega)\|_\infty/3}\right).
\end{eqnarray*}
Choosing $t\approx \|\hat{P}^{(r)}(\Omega)\|_\infty\log(n+m) + \|\hat{P}^{(r)}(\Omega)\|_F(\|P(\Omega)\|_\infty\log(n+m))^{1/2}$
for each $r$ and using the union bound across $r$, we obtain the final bound for $\|X\hat{\pi}-X\pi^*\|^2 = \norm{\hat{P}(\Omega) - \Pi_{\mathcal{C}}P(\Omega)}_F^2$ with
$$\Phi = \max_r\|\hat{P}^{(r)}(\Omega)\|_\infty + \|P(\Omega)\|_\infty^{1/2}\max_r\|\hat{P}^{(r)}(\Omega)\|_F = \max_r\|\hat{P}^{(r)}(\Omega)\|_\infty + \|P(\Omega)\|_\infty^{1/2}\norm{\Sigma}_{\infty}.$$
The proof is complete.
\end{proof}
\subsection{Network mixing with exponential weights and estimation adaptivity}\label{sec:model selection}
We view the validation error $\|A(\Omega)-\hat{P}^{(r)}(\Omega)\|_F^2$ as a goodness-of-fit metric for the $r$th model. To match the performance of the optimal model in $\mathcal{M}$, we focus on the models with small validation errors. In particular, we propose the following simple rule to combine the models based on exponential weights:
\begin{eqnarray}\label{eq:exp weights}
\hat{P}^{\text{(exp)}} = \sum_{r=1}^m \pi_r \hat{P}^{(r)}, \qquad \pi_r= \frac{ \exp\left(-\|A(\Omega)-\hat{P}^{(r)}(\Omega)\|_F^2\right)}{ \sum_{r=1}^m\exp\left(-\|A(\Omega)-\hat{P}^{(r)}(\Omega)\|_F^2\right)}.
\end{eqnarray}
Compared with \eqref{eq:ECV-estimator}, this estimator is a soft-selection version of the cross-validation procedure. Despite its simplicity, we now show that it achieves model adaptivity, matching the performance of the unknown optimal estimator produced by $\mathcal{M}$. Since $\Omega$ is independent of the data, all theoretical analysis is conditioned on $\Omega$. The data-splitting proportion $p$ is a fixed constant, independent of the network size $n$.
\begin{thm}[Mixing estimator with exponential weights]\label{thm:exp weights}
Let $A$ be the adjacency matrix of a random network drawn from the inhomogeneous Erd\H{o}s-R\'{e}nyi model with probability matrix $P=\mathbb{E} A$.
Denote by $\hat{P}^{(r)}$, $1\le r \le m$, the estimators of $P$ obtained by applying the methods from $\mathcal{M}$ to $A(\Omega)$, and let $\hat{P}$ be the convex combination of these estimates defined by \eqref{eq:exp weights}. Assume $m = o(n^2)$. Then with high probability,
\begin{equation}\label{eq:exp-error}
\big\|\hat{P}^{\text{(exp)}}(\Omega)-P(\Omega)\big\|_F \le \min_{1\le r\le m}\big\|\hat{P}^{(r)}(\Omega)-P(\Omega)\big\|_F+ C\varepsilon,
\end{equation}
where $C>0$ is a constant and
\begin{eqnarray*}
\varepsilon = \Big(\log n\cdot\max_{i,j}P_{ij}\Big)^{1/2}+\min\left\{\frac{\log(nm)}{\min_{1\le r\le m}\|P^{(r)}(\Omega)-P(\Omega)\|_F}, \ \log^{1/2}(nm)\right\}.
\end{eqnarray*}
\end{thm}
\begin{rem}
Although Theorem~\ref{thm:exp weights} describes the error restricted to $\Omega$, extending the error bound to the entire matrix $P$ is trivial. Replacing $\Omega$ by $\Omega^c = [n]^2\setminus\Omega$ in the estimating procedure and applying Theorem~\ref{thm:exp weights} would generate another estimator $\breve{P}$ that admits the same type of error bound. Then combining $\hat{P}(\Omega)$ and $\breve{P}(\Omega^c)$ as a full estimator would produce the same error bound for the entire matrix $P$. For simplicity, we will state our results only in terms of $P(\Omega)$ in all theoretical discussions to follow.
\end{rem}
Theorem~\ref{thm:exp weights} states that the estimator given in \eqref{eq:exp weights} is nearly as good as the best estimate produced by a single method from $\mathcal{M}$. To better understand the additional error $\varepsilon$, assume that the network is generated from a stochastic block model, under which nodes are partitioned into $k$ groups according to a label vector $c\in[n]^k$ and edges are formed independently between nodes with probabilities $P_{ij}=B_{c_ic_j}$ for some fixed matrix $B\in\mathbb{R}^{k\times k}$. According to \cite{gao2015rate},
\begin{eqnarray*}
\min_{1\le r\le m}\big\|\hat{P}^{(r)}(\Omega)-P(\Omega)\big\|_F^2 \gtrsim k^2 + n\log k,
\end{eqnarray*}
while the square of the additional error is
\begin{eqnarray*}
\varepsilon^2 \lesssim \log n + \min\left\{\frac{\log^2(nm)}{k^2+n\log k},\log(nm)\right\} \approx \log n + \frac{\log^2(nm)}{k^2+n\log k}.
\end{eqnarray*}
It is easy to see that $\varepsilon^2$ is smaller than the rate-optimal error $k^2 + n\log k$. Therefore, our estimator can always match the optimal estimator from $\mathcal{M}$.
As a connection to similar properties of statistical estimation problems, \cite{catoni1997mixture} and \cite{yang2000mixing,yang2001adaptive} introduce mixing methods for regression and density estimation that can achieve estimation adaptivity. Unfortunately, their strategies are computationally expensive even in regression settings, let alone in network problems, the sample size of which scales in $O(n^2)$. In contrast, our estimator \eqref{eq:exp weights} achieves the same type of adaptivity with negligible computational cost in addition to the $m$ model-fitting procedures.
\subsection{Beyond adaptivity: Non-negative linear network mixing}
The exponential mixing estimator \eqref{eq:exp weights} performs well when at least one of the $m$ individual estimators is accurate. In practice, however, all the available estimators may be inaccurate, for example, when the network is very sparse. Given the $m$ estimators at hand, it is natural to be more ambitious: to seek a better candidate than any single model estimate.
Consider the class of linear combinations of the $m$ estimators
\begin{equation}\label{eq:generic-estimator}
\hat{P} = \sum_{r=1}^m \pi_r \hat{P}^{(r)}.
\end{equation}
The most natural option is the \emph{linear mixing estimator}, which uses the weights provided by solving the ordinary least squares (OLS) problem:
\begin{eqnarray}\label{eq:ols optimization}
\hat{\pi}^{\text{(ols)}} = \mathrm{argmin}_{\pi \in \mathbb{R}^m}\Big\|\sum_{r=1}^m \pi_r \hat{P}^{(r)}(\Omega) - A(\Omega)\Big\|_F^2.
\end{eqnarray}
Although linear mixing performs very well in many settings, it is a generic method and does not leverage many features of network estimation. For example, all the entries of $P$ are non-negative, and therefore it is expected that $\langle \hat{P}^{(r)}(\Omega),P(\Omega)\rangle>0$ for any reasonable estimate $\hat{P}^{(r)}(\Omega)$. Similarly, $\langle \hat{P}^{(r)}(\Omega),\hat{P}^{(s)}(\Omega) \rangle > 0$ for all $1\le r,s\le m$ (unless their supports are disjoint), and the angles between reasonably good estimators tend to be small. Moreover, in sparse networks the noises (entries of $A-P$) are heavy-tailed random variables, so the available estimators of $P$ can be very noisy. These observations motivate us to further improve the linear mixing method by introducing a natural non-negative constraint to the OLS weights.
In particular, we consider the \emph{non-negative linear} (NNL) mixing estimator based on the following weights:
\begin{eqnarray}\label{eq:nnls optimization}
\hat{\pi}^{\text{(nnl)}} = \mathrm{argmin}_{\pi \succeq 0}\Big\|\sum_{r=1}^m \pi_r \hat{P}^{(r)}(\Omega) - A(\Omega)\Big\|_F^2,
\end{eqnarray}
where $\pi \succeq 0$ denotes the constraint that all the entries of the weight vector $\pi$ are non-negative. This is a simple convex optimization problem and can be solved efficiently by either a projected quasi-Newton algorithm or a sequential coordinate descent algorithm \citep{kim2006new, chen2010nonnegativity}.
The NNL mixing estimator is defined to be
\begin{eqnarray}\label{eq:nnls estimate}
\hat{P}^{\text{(nnl)}} = \sum_{r=1}^m \hat{\pi}_r^{\text{(nnl)}} \hat{P}^{(r)}.
\end{eqnarray}
We can see that $\hat{P}^{\text{(nnl)}}(\Omega)$ is the projection of $A(\Omega)$ on the convex cone formed by the conical combination of $m$ individual estimators $\hat{P}^{(r)}(\Omega), 1\le r \le m$. The non-negative sign constraint directly imposes a regularization effect that significantly helps the method handle the potentially large number of individual estimators. This is crucial for a tuning-free procedure and matches our aim for an ``off-the-shelf" method.
The non-negative constraint has proved effective in high-dimensional linear regression problems \citep{Slawski.NNLS.2011, slawski2013non, meinshausen2013sign}, with properties similar to the LASSO estimator. However, in the current context, NNL estimation is not motivated by the curse of dimensionality. In network mixing problems, the sample size for \eqref{eq:ols optimization} is of order $n^2$, and $m$ is usually much smaller than $n^2$. Therefore, though $m$ can be large if one wants to ensure the expressiveness of the mixing estimator, \eqref{eq:ols optimization} is usually not an ultra-high-dimensional problem. Instead, the strong correlation between the $ \hat{P}^{(r)}$s and the low concentration of the adjacency matrix (due to the network sparsity) complicate the matter and result in the deterioration of the OLS estimator, even in a relatively low-dimensional setting.
To see why the non-negative constraint can help, let us assume that the true signal $P(\Omega)$ is within the convex cone of the estimators. Figure~\ref{fig:demo-normal} illustrates an ideal situation. When the two individual estimators form a large angle, and the perturbation range of $A$ around $P$ is not too large, the observed adjacency matrix is likely to stay within or close to the convex cone, so the OLS and NNL estimators perform similarly. When the perturbation range is large, as in Figure~\ref{fig:demo-correlation} (which happens in sparse networks \citep{le2017concentration}), $A$ may be far from the cone (and $P$), and the projection to the cone can result in a better estimator.
On the other hand, when the two estimators align well, the convex cone is smaller (Figure~\ref{fig:demo-sparse}), and $A$ is unlikely to belong to the cone. In this case, the NNL estimator can also outperform the OLS estimator. Overall, these observations show that sparse networks and strong alignment between estimators tend to make the OLS estimator vulnerable to increasing $m$, even when $m$ is much smaller than $n^2$.
\begin{figure}[H]
\centering
\begin{subfigure}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{./Figures/NormalFigure.pdf}
\caption{Good concentration and weak alignment}
\label{fig:demo-normal}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{./Figures/LargeDeviationFigure.pdf}
\caption{ Bad concentration and weak alignment}
\label{fig:demo-correlation}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{./Figures/SmallAngleFigure.pdf}
\caption{Good concentration and strong alignment}
\label{fig:demo-sparse}
\end{subfigure}
\caption{Illustration of the relation between the convex cone and the concentration of the adjacency matrix.}
\label{fig:demo}
\end{figure}
We now present the theoretical support for the NNL estimator and its claimed advantage over linear mixing. Let
\begin{eqnarray*}
\mathcal{L}=\Big\{\sum_{r=1}^m \pi_r\hat{P}^{(r)}(\Omega), \pi_r\in\mathbb{R},1\le r \le m\Big\}, \quad \mathcal{C}=\Big\{\sum_{r=1}^m\pi_r\hat{P}^{(r)}(\Omega), \pi_r\ge 0, 1\le r \le m\Big\}
\end{eqnarray*}
be the linear subspace and the cone generated by $\{\hat{P}^{(r)}(\Omega),1\le r\le m\}$. Denote by $\Pi_\mathcal{L}$ and $\Pi_\mathcal{C}$ the projections onto $\mathcal{L}$ and $\mathcal{C}$, respectively.
\begin{thm}[Non-negative mixing estimator, positive inner products]\label{thm:nnls}
Let $A$ be the adjacency matrix of a random network drawn from the inhomogeneous Erd\H{o}s-R\'{e}nyi model with probability matrix $P=\mathbb{E} A$. Denote by $\hat{P}^{(r)}$, $1\le r \le m$, the estimates of $P$ obtained by applying the methods from $\mathcal{M}$ to $A(\Omega^c)$, and let $\hat{P}^{\text{(nnl)}}$ be the NNL estimator defined by \eqref{eq:nnls estimate}. Assume that
\begin{equation}\label{eq:correlated assumption}
\delta = \min_{1\le r,s\le m}\frac{\langle \hat{P}^{(r)}(\Omega),\hat{P}^{(s)}(\Omega) \rangle}{\|\hat{P}^{(r)}(\Omega)\|_F\cdot\|\hat{P}^{(s)}(\Omega)\|_F}>0.
\end{equation}
Then there exists a constant $C>0$ such that with high probability,
\begin{eqnarray}\label{eq:nnls bound}
\|\hat{P}^{\text{(nnl)}}(\Omega)-P(\Omega)\|_F &\le& \|\Pi_\mathcal{C}P(\Omega)- P(\Omega)\|_F + C\delta^{-1/2}\varepsilon,
\end{eqnarray}
where $C>0$ is a constant and
\begin{equation}\label{eq:nnls-tail}
\varepsilon = \max_{1\le r \le m}\frac{\|\hat{P}^{(r)}(\Omega)\|_\infty}{\|\hat{P}^{(r)}(\Omega)\|_F}\cdot\log(n+m) +\sqrt{\|P(\Omega)\|_\infty\log(n+m)}.
\end{equation}
\end{thm}
Theorem~\ref{thm:nnls} states that the proposed estimator $\hat{P}^{\text{(nnl)}}(\Omega)$ is nearly as accurate as the non-negative linear oracle $\Pi_\mathcal{C}P(\Omega)$. Given a fixed $\delta$, the extra error term $\varepsilon$ can be at most of the order $\log(n+m)$, although it will grow much more slowly if the entries of $P$ and $\hat{P}^{(r)}$ are of similar orders and the network is relatively sparse.
As a consequence of Theorem~\ref{thm:nnls}, the next corollary shows that the NNL estimator can be strictly more accurate than the OLS estimator $\Pi_\mathcal{L} A(\Omega)$, especially when $m$ is large (a typical scenario in model aggregation).
\begin{coro}[Comparison of NNL and OLS estimators]\label{cor:comparison}
Let $A$ be the adjacency matrix of a random network drawn from the inhomogeneous Erd\H{o}s-R\'{e}nyi model with probability matrix $P=\mathbb{E} A$. Denote by $\hat{P}^{(r)}$, $1\le r \le m$, the estimators of $P$ obtained by applying the methods from $\mathcal{M}$ to $A(\Omega^c)$, and let $\hat{P}^{\text{(ols)}}$ be the OLS estimator defined by \eqref{eq:ols optimization}. With high probability,
\begin{equation}\label{eq:ols-oracle}
\norm{\hat{P}^{(\text{ols})}(\Omega) - P(\Omega)}_F^2 \le \norm{\Pi_{\mathcal{L}}P(\Omega) - P(\Omega)}_F^2 + m\|P(\Omega)\|_\infty + \left(m\log n\|P(\Omega)\|_\infty\right)^{1/2}.
\end{equation}
Furthermore, consider the setting in Theorem~\ref{thm:nnls} and denote
$$\rho = \big(1-\|P(\Omega)\|_\infty\big)^2\cdot\min_{(i,j)\in\Omega} P_{ij}^2, \quad \Delta = \|\hat{P}^{\text{(ols)}}(\Omega)-P(\Omega)\|_F^2 - \|\hat{P}^{\text{(nnl)}}(\Omega)-P(\Omega)\|_F^2.$$
Assume $m \ge C\|P(\Omega)\|_\infty\rho^{-1}\log n$ for some sufficiently large constant $C>0$. Then with high probability,
\begin{equation}\label{eq:Delta bound}
\Delta \ge \frac{m\rho^{1/2}}{2} - \Psi,
\end{equation}
where
\begin{eqnarray*}
\Psi = \frac{C'\log^2(n+m)}{\delta}\Big(\max_{1\le r \le m}\frac{\|\hat{P}^{(r)}(\Omega)\|_\infty^2}{\|\hat{P}^{(r)}(\Omega)\|_F^2} +\|P(\Omega)\|_\infty\Big)+ \|\Pi_\mathcal{C}\Pi_\mathcal{L} P(\Omega)-\Pi_\mathcal{L} P(\Omega)\|_F^2
\end{eqnarray*}
for some constant $C'>0$.
\end{coro}
Note that the term $\|\Pi_\mathcal{C}\Pi_\mathcal{L} P(\Omega)-\Pi_\mathcal{L} P(\Omega)\|_F$ is the distance from the linear oracle $\Pi_\mathcal{L}P(\Omega)$ to the cone $\mathcal{C}$, measuring the degree of violation of the non-negative cone assumption. Corollary~\ref{cor:comparison} demonstrates the effects of the main factors on performance. It can be seen that the advantage of NNL mixing over OLS mixing increases with $m$ and also with the correlation $\delta$, as we have discussed above. Next, we illustrate the effect of network density.
\begin{ex}\label{rem:degree}
Assume a sparse network and that all the entries of $P$ are of the same order $d/n = o(1)$, where $d$ is the average degree. Also, assume that the $m$ individual estimators are reasonable, so all the $\hat{P}^{(r)}_{ij}$s are also of the same order as the $P_{ij}$s. When the conical assumption holds and $m \gg n\log{n}/d$, taking the relative error to cancel the scaling effect of $P$, we have
$$\frac{\Delta}{\norm{P(\Omega)}_F^2} \gtrsim \frac{m}{nd} - \frac{\log^2(n+m)}{\delta d^2}.$$
For sparser networks with small $d$, the advantage of the NNL estimator is more dramatic. This theoretical prediction is further supported by our empirical evidence in Section~\ref{sec:simulation}.
\end{ex}
Theorem~\ref{thm:nnls} requires $\hat{P}^{(r)}(\Omega)$, $1\le r\le m$, to have positive inner products. This assumption is reasonable because they are all estimators of $P(\Omega)$ and must have non-negative entries. More importantly, we can directly calculate $\delta$ from data.
Next, we consider an even weaker assumption, which is true in all reasonable settings we can think of. We replace the assumption \eqref{eq:correlated assumption} in Theorem~\ref{thm:nnls} with the following assumption on $\hat{P}^{(r)}(\Omega)$, $1\le r\le m$, known as the {\em self-regularizing property} \citep{Slawski.NNLS.2011}. Denote by $\Sigma\in\mathbb{R}^{m\times m}$ the matrix with entries $\Sigma_{rs}=\langle \hat{P}^{(r)}(\Omega), \hat{P}^{(s)}(\Omega)\rangle$ and $\|\Sigma\|_\infty=\max_{1\le r,s\le m}|\Sigma_{rs}|$. We say that $\Sigma$ satisfies the self-regularizing property with constant $\kappa$, $0<\kappa\le 1$, if
\begin{eqnarray}\label{eq:self-reg}
\beta^T\Sigma\beta \ge \kappa\|\Sigma\|_\infty, \text{ for all $\beta\succeq 0$ and $\sum_{r=1}^m\beta_r = 1$}.
\end{eqnarray}
Notice that, like \eqref{eq:correlated assumption}, this condition can be numerically verified from the data by solving a convex problem. More importantly, in our current setting of network mixing, this condition will almost always hold due to the following property, taken directly from the discussion of \cite{Slawski.NNLS.2011}.
\begin{prop}\label{prop:self-reg}
If there exists a partition $\{ Q_t\}_{t=1}^T$ of $[m]$ such that
$$\min_{r, s \in Q_t} \langle \hat{P}^{(r)}(\Omega), \hat{P}^{(s)}(\Omega)\rangle \ge \kappa \max_{r\in Q_t}\norm{\hat{P}^{(r)}(\Omega)}_F^2 >0, \text{ for all } 1\le t \le T,$$
then $\Sigma$ is self-regularizing with constant $\kappa/T$.
\end{prop}
To see why \eqref{eq:self-reg} is weaker than \eqref{eq:correlated assumption}, notice that without loss of generality we can assume that all the $\hat{P}^{r}(\Omega)$s have the same norm, because rescaling does not change our linear fitting. So when \eqref{eq:correlated assumption} holds, $\Sigma$ also satisfies \eqref{eq:self-reg}, with $\kappa = \rho$. Moreover, even if we are in the extreme situation where
$$\langle \hat{P}^{(r)}(\Omega), \hat{P}^{(s)}(\Omega)\rangle = 0$$
for some $r, s$, Proposition~\ref{prop:self-reg} indicates that \eqref{eq:self-reg} can still hold as long as we separate such pairs in different groups. Even in the worst case, the $m$-way partition guarantees the self-regularizing property, with $\kappa = 1/m$.
\begin{thm}\label{thm:self-regularizing}
Consider the setting of Theorem~\ref{thm:nnls} with condition \eqref{eq:correlated assumption} replaced by the self-regularizing property \eqref{eq:self-reg}. Then there exists a constant $C>0$ such that with high probability,
\begin{eqnarray*}
\|\hat{P}^{\text{(nnl)}}(\Omega)-P(\Omega)\|_F &\le& \inf_{\eta \in \mathcal{C}}\|\eta-\Pi_\mathcal{L} P(\Omega)\|_F+ C\kappa^{-1/2}\epsilon',
\end{eqnarray*}
where
\begin{eqnarray*}
\epsilon' &=& \left(\frac{\Phi\log(n+m)}{\sqrt{\|\Sigma\|_\infty}} +\sqrt{\Phi\|\pi^*\|_1\log(n+m)}\right),\\
\Phi &=& \max_{1\le r \le m}\|\hat{P}^{(r)}(\Omega)\|_\infty + \|P(\Omega)\|_\infty^{1/2}\norm{\Sigma}_{\infty},
\end{eqnarray*}
and $\pi^*$ is the non-negative linear coefficient of the oracle estimator such that
$$\Pi_{\mathcal{C}}P(\Omega) = \sum_{r}\pi^*_r\hat{P}^{(r)}(\Omega).$$
\end{thm}
Compared with Theorem~\ref{thm:nnls}, the price we pay for the weaker assumption in Theorem~\ref{thm:self-regularizing} is greater additive error. As a simple demonstration, consider the setting of Example~\ref{rem:degree} and assume that all the $\hat{P}^{r}(\Omega)$s have the same norm. In this case, $\kappa = \delta$. At best, the $\epsilon'$ of Theorem~\ref{thm:self-regularizing} has
$$\epsilon' \approx \frac{d^{3/2}}{n^{1/2}}\log(n+m) + \frac{d^{5/4}}{n^{1/4}}\log^{1/2}(n+m). $$
In contrast, the $\epsilon$ of Theorem~\ref{thm:nnls} gives
$$\epsilon \approx \frac{1}{n}\log(n+m) + \frac{d^{1/2}}{n^{1/2}}\log^{1/2}(n+m), $$
which is clearly of a lower order.
\subsection{Candidate set and other practical considerations}\label{secsec:practical}
The mixing strategy involves determining a set of candidate estimators and the hold-out proportion, $p$. For the OLS and NNL mixing methods, the scales of the individual estimators do not matter, and they do not need to be accurate individually. This property is essential for a mixing method to be applicable for general link prediction problems (see Section~\ref{sec:data}). But exponential mixing does not make sense if the $\hat{P}^{(r)}$s are in the wrong scale or if they are all inaccurate.
In determining the candidate set $\mathcal{M}$, there is a necessary tradeoff between computational efficiency and expressiveness. We recommend using spectral clustering together with SBM fitting for $1 \le k \le K_{\max}$, spherical spectral clustering \citep{rohe2011spectral,lei2014consistency} and DCBM fitting for $1 \le k \le K_{\max}$, and the universal singular value thresholding (USVT) of \cite{chatterjee2015matrix}, with the improvement mentioned in \cite{zhang2015estimating}, by taking the first $n^{1/3}$ singular components. This set of candidate estimators can be calculated efficiently for large networks, with the main computational burden being the one-time calculation of SVD. The block models, despite their simplicity, have been shown to have great approximation power if they are used properly \citep{airoldi2013stochastic}, and USVT also comes with good expressiveness. When networks are sparse, the block components likely receive higher mixing weights, helping stabilize the estimator.
In contrast, when networks are sufficiently dense, USVT can be more effective. Empirically, we observe that this setup gives very effective estimation performance. In the above recommendation, $K_{\max}$ is a reasonably large positive integer, and in Sections~\ref{sec:simulation} and \ref{sec:data}, we show that our method is very robust to $K_{\max}$ and $\rho$ within reasonable ranges, which is again due to the regularizing property of the non-negative constraint.
The discussion so far has focused on one random split, $\Omega$. In practice, one can also randomly repeat the procedure multiple times and take the average as the output. This may slightly improve accuracy in our limited evaluation but will also increase the computational cost by a multiplicative factor. For this reason, we only consider single splits in this study.
The mixing method, and any aggregation methods, do have limitations. The primary performance metric of the mixing approach is estimation accuracy. Combining multiple estimators may destroy the structural interpretations of each individual estimator, such as block structures or smoothness. However, when special structures are desirable, one can always apply the same structural extraction strategy, such as community detection, to the estimated $\hat{P}$. A more accurate estimate of $P$ may lead to more accurate structure extraction.
\subsection{Comparison of mixing aggregation strategies}
We first want to compare different aggregation strategies for the mixing method, including linear mixing (OLS-m), NNL mixing (NNL-m), exponential mixing (EXP-m), and edge cross-validation model selection (ECV-m). Since $K_{\max}$ needs to be specified, we investigate its impact on the mixing method's performance. To demonstrate that the method is almost tuning-free, we want to show that it remains stable for a reasonable range of $K_{\max}$ values. Figure~\ref{fig:K-value} shows the estimation performance with varying $K_{\max}$ for networks with expected average node degree $20$.
\begin{figure}[H]
\centering
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{./Figures/Varying-K-AggregationComparison-BarsGraphon-d=20}
\caption{Model 1}
\label{fig:bars-graphon-varying-K}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{./Figures/Varying-K-AggregationComparison-RainbowGraphon-d=20}
\caption{Model 2}
\label{fig:rainbow-graphon-varying-K}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{./Figures/Varying-K-AggregationComparison-DiagGraphon-d=20}
\caption{Model 3}
\label{fig:diag-graphon-varying-K}
\end{subfigure}
\caption{Relative estimation error of $P$ with various values of $K_{\max}$ and $p=0.1$. The networks under evaluation have $n=1000$ nodes and average node degree $20$.}
\label{fig:K-value}
\end{figure}
As can be seen, small values of $K_{\max}$ result in underfitting and a large error. As $K_{\max}$ increases, the error drops until overfitting kicks in. However, overall, exponential mixing, NNL mixing, and cross-validation do not suffer from overparameterization of block approximations. Linear combination, in contrast, degrades when $K_{\max}$ is large. This observation matches our theoretical understanding of the linear estimator as we are fitting the model with an increasing number of variables. Exponential mixing significantly outperforms cross-validation, while NNL mixing is much better than both.
Next, we evaluate the robustness of the different strategies with respect to the hold-out proportion, $p$. We evaluate their performance when $p$ varies from 0.1 to 0.5 and $K_{\max}$ is fixed at 15 (Figure~\ref{fig:rho-value}). Overall, all the aggregation methods tend to improve as $p$ decreases within this range. This indicates that the aggregation weight determination step is easier and requires a smaller sample size than the individual model estimation step. Of all the aggregation methods, non-negative linear combination is the most robust. Linear combination also delivers competitive estimation accuracy. Based on this observation, we recommend using $K_{\max} = 15, p=0.1$ as the default configuration for network mixing estimation.
\begin{figure}[H]
\centering
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{./Figures/Varying-Rho-AggregationComparison-BarsGraphon-d=20}
\caption{Model 1}
\label{fig:bars-graphon-varying-rho}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{./Figures/Varying-Rho-AggregationComparison-RainbowGraphon-d=20}
\caption{Model 2}
\label{fig:rainbow-graphon-varying-rho}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{./Figures/Varying-Rho-AggregationComparison-DiagGraphon-d=20}
\caption{Model 3}
\label{fig:diag-graphon-varying-rho}
\end{subfigure}
\caption{Relative estimation error of $P$ with varying $p$ and $K_{\max} = 15$. The networks under evaluation have $m=1000$ and average node degree $20$.}
\label{fig:rho-value}
\end{figure}
Now we fix $K_{\max} = 15$ and $p=0.1$ and evaluate the estimation performance for a range of network sparsity levels, varying the expected average degree from 5 to 45 (Figure~\ref{fig:d-value}). In the sparse regime, NNL mixing outperforms the others, while linear mixing catches up as the network becomes denser. This is predicted in Corollary~\ref{cor:comparison} as well. Exponential mixing eventually coincides with the cross-validation method, which also matches our theory.
\begin{figure}[H]
\centering
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{./Figures/Varying-Density-AggregationComparison-BarsGraphon}
\caption{Model 1}
\label{fig:bars-graphon-varying-d}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{./Figures/Varying-Density-AggregationComparison-RainbowGraphon}
\caption{Model 2}
\label{fig:rainbow-graphon-varying-d}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{./Figures/Varying-Density-AggregationComparison-DiagGraphon}
\caption{Model 3}
\label{fig:diag-graphon-varying-d}
\end{subfigure}
\caption{Relative estimation error of $P$ with various expected average degrees. The networks under evaluation contain $n=1000$ nodes, and $K_{\max}=15, p = 0.1$ are used.}
\label{fig:d-value}
\end{figure}
Overall, NNL mixing is preferable compared to the other methods. Linear mixing is less robust to the choice of $K_{\max}$ and is inferior to non-negative combination in the sparse setting. However, when the network is denser, it outperforms the others. In all of the experiments to follow, we use $K_{\max}=15$ and $p=0.1$, and we believe this configuration can be used in almost all applicable tasks.
\subsection{Comparison with benchmark network estimation methods }
Now we compare the mixing method with a few benchmark network estimation methods. We split this into two parts. First, we consider the three graphon models in Figure~\ref{fig:graphons}, and we compare the mixing method to the graphon methods, which have theoretical guarantees and reasonable computational cost. These include the USVT estimator of \cite{chatterjee2015matrix}, the neighborhood smoothing (NS) method of \cite{zhang2015estimating}, and the sort-and-smooth (SAS) method of block approximation from \cite{airoldi2013stochastic} and \cite{chan2014consistent}. Second, we generate networks from two special parametric models: the SBM and the latent space model \citep{hoff2002latent}. Under these models, oracle estimations can be achieved by parametric model fitting, and they are included for comparison. The mixing, USVT, NS, and latent space model fitting \citep{ma2020universal} are all based on the R package {\em randnet} \citep{randnet}, while SAS is based on the R package {\em graphon} \citep{graphon}.
Figure~\ref{fig:density-benchmark} shows the model estimation performance of the proposed mixing strategies and the three graphon estimation benchmarks. All four variant mixing strategies uniformly outperform the benchmark methods with all three models. The advantage of the mixing method is clear when the networks are sparse. Of the three graphon estimation methods, SAS is more accurate for the first two models, while USVT and NS are better for the third one. The difference between various mixing strategies is negligible.
\begin{figure}[H]
\centering
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{./Figures/Varying-Density-GraphonBenchmarkComparison-BarsGraphon}
\caption{Model 1}
\label{fig:bars-graphon-varying-density-benchmark}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{./Figures/Varying-Density-GraphonBenchmarkComparison-RainbowGraphon}
\caption{Model 2}
\label{fig:rainbow-graphon-varying-density-benchmark}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{./Figures/Varying-Density-GraphonBenchmarkComparison-DiagGraphon}
\caption{Model 3}
\label{fig:diag-graphon-varying-density-benchmark}
\end{subfigure}
\caption{Estimation performance of the proposed mixing methods and three benchmark graphon estimation methods on synthetic networks generated from three graphon models. }
\label{fig:density-benchmark}
\end{figure}
Next, we generate networks from two parametric models. The first one is the SBM, with six communities of equal size, following the configuration of \cite{zhang2015estimating}. Under the SBM, we include two oracle estimators. Oracle1 requires the true community labels and estimates the probability matrix $P$ by averaging entries within corresponding blocks of the adjacency matrix $A$. Oracle2 knows the true SBM with six clusters but not the true community labels; it uses spectral clustering to find the node labels and averages entries within estimated blocks of $A$. In particular, notice that Oracle2 itself is automatically included as one of the individual models in the mixing.
The second model is the latent space model (LSM), with
$$\text{logit}(P_{ij}) = \alpha_i + \alpha_j + \langle Z_i, Z_j \rangle,$$
where $Z_i, Z_j$ are latent vectors in $\mathbb{R}^4$, generated by $N(0, I_4)$ and then centralized according to \cite{ma2020universal}. Oracle1 in this case is the oracle version of the model that uses the true model structure and also the true latent dimension, estimated by the gradient descent method of \cite{ma2020universal} with the recommended initialization. Oracle2 still assumes the correct model structure, but uses the wrong dimension (3 instead of 4), representing the possibility of dimensionality mis-specification. For reference, we also include an oracle version of the NNL mixing method, which has Oracle1 as an individual component.
Figure~\ref{fig:parametric-benchmark} shows the performance of all the methods under the two parametric models. In the SBM setting, the mixing methods are inferior only to the unbeatable Oracle1 and are even better than Oracle2. Since the mixing procedure includes Oracle2 as an individual component, this result demonstrates the effects of ensembling multiple models. In the difficult regime, including multiple models may help stabilize the estimation and further improve the estimate of the true model when it is fitted separately. In the LSM setting, the mixing method is again inferior only to the perfect oracle parametric estimation (Oracle1) in the dense setting and, again, it is even better than it in the sparse setting. USVT is also very effective in this setting, as indicated by the theory of \cite{chatterjee2015matrix}, and the mixing method adapts to it. NNL-m and USVT are even better than Oracle2 estimate. The oracle NNL-m adapts to Oracle1 in the dense setting and outperforms it in the sparse setting. The comparison between the mixing method and the oracle methods under these two parametric models highlights two advantages of the mixing approach:
\begin{itemize}
\item In the sparse regime, the estimation accuracy may be poor even if the true model is known. The mixing strategy provides a mechanism for combining simpler models to obtain a more stable estimate.
\item In the dense regime, the mixing approach may remain effective even if the mixing ensemble does not include the true model and can match the oracle otherwise.
\end{itemize}
\begin{figure}[H]
\centering
\begin{subfigure}[t]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{./Figures/Varying-Density-GraphonBenchmarkComparison-BlockGraphon}
\caption{Stochastic block model}
\label{fig:block-graphon-varying-density-benchmark}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{./Figures/Varying-Density-GraphonBenchmarkComparison-LSM}
\caption{Latent space model}
\label{fig:LSM-graphon-varying-density-benchmark}
\end{subfigure}
\caption{Comparison with benchmark graphon estimation methods on synthetic networks generated from two parametric models. }
\label{fig:parametric-benchmark}
\end{figure}
\section*{Acknowledgements}%
\addtocontents{toc}{\protect\vspace{6pt}}%
\addcontentsline{toc}{section}{Acknowledgements}%
}
\setlength{\textwidth}{15.3 truecm} \setlength{\textheight}{23.9
truecm}
\newcommand{\nonumber \\}{\nonumber \\}
\def\pr{\textsf{P}}
\def\ep{\textsf{E}}
\def\Cov{\textsf{Cov}}
\def\Var{\textsf{Var}}
\def\Cal#1{{\mathcal #1}}
\def\bk#1{{\mathbf #1}}
\def\bkg#1{\mbox{\boldmath{$#1$}}}
\def\smallbkg#1{\mbox{\scriptsize \boldmath{$#1$}}}
\def\text#1{\mbox{\rm #1}}
\def\overset#1#2{\stackrel{#1}{#2} }
\def\mathbf{\mathbf}
\def\mathrm{\mathrm}
\def\displaystyle\sum{\displaystyle\sum}
\def\displaystyle\int{\displaystyle\int}
\def\displaystyle\frac{\displaystyle\frac}
\renewcommand{\baselinestretch}{1}
\newcommand{\pkg}[1]{\textsf{#1}}
\newcommand{\printfnsymbol}[1]{%
\textsuperscript{\@fnsymbol{#1}}%
}
\usepackage{hyperref}
\hypersetup{
colorlinks=true,
linkcolor=blue,
filecolor=magenta,
urlcolor=cyan,
citecolor=blue
}
\urlstyle{same}
\title{\textbf{Network Estimation by Mixing: Adaptivity and More}
\thanks{Both authors contributed equally to this work.}}
\author{Tianxi Li\\ Department of Statistics, University of Virginia
\and Can M. Le\\ Department of Statistics, University of California, Davis
}
\date{\today}
\begin{document}
\maketitle
\begin{abstract}
\noindent
Networks analysis has been commonly used to study the interactions between units of complex systems. One problem of particular interest is learning the network's underlying connection pattern given a single and noisy instantiation. While many methods have been proposed to address this problem in recent years, they usually assume that the true model belongs to a known class, which is not verifiable in most real-world applications. Consequently, network modeling based on these methods either suffers from model misspecification or relies on additional model selection procedures that are not well understood in theory and can potentially be unstable in practice. To address this difficulty, we propose a mixing strategy that leverages available arbitrary models to improve their individual performances. The proposed method is computationally efficient and almost tuning-free; thus, it can be used as an off-the-shelf method for network modeling. We show that the proposed method performs equally well as the oracle estimate when the true model is included as individual candidates. More importantly, the method remains robust and outperforms all current estimates even when the models are misspecified. Extensive simulation examples are used to verify the advantage of the proposed mixing method. Evaluation of link prediction performance on 385 real-world networks from six domains also demonstrates the universal competitiveness of the mixing method across multiple domains.
\end{abstract}
\section{Introduction}
\input{Intro}
\section{The network mixing method and its properties}\label{sec:method}
\input{Setup}
\section{Simulation examples}\label{sec:simulation}
\input{Simulation}
\section{Link prediction in real-world networks}\label{sec:data}
\input{data}
\section{Discussion}\label{sec:disc}
\input{discussion}
\section*{Acknowledgement}
C. M. Le is supported in part by the NSF grant DMS-2015134. T. Li is supported in part by the NSF grant DMS-2015298 and the Quantitative Collaborative Award from the College of Arts and Sciences at the University of Virginia.
|
\section{Introduction}
\label{sec:intro}
\input{introduction}
\section{DP Preliminaries}
\label{sec:prelims}
\input{prelims}
\section{Privacy Loss Random Variables (PRVs)}
\label{sec:PRV}
\input{PRV}
\section{Numerical composition of privacy curves}
\label{sec:numerical_composition}
\input{numerical_composition.tex}
\section{Experiments}
\label{sec:experiments}
\input{experiments.tex}
\subsection*{Acknowledgements}
We would like to thank Janardhan Kulkarni and Sergey Yekhanin for several useful discussions and encouraging us to work on this problem.
L.W. would like to thank Daniel Jones and Victor R\"uhle for fruitful discussions and helpful guidance.
\FloatBarrier
\bibliographystyle{alpha}
\section{Missing Proofs in Error Analysis}
\label{sec:error_analysis_extra}
\subsection{Facts about Coupling Approximation}
Here we collect some useful properties of coupling approximations.
The following lemma shows that the coupling approximations satisfy a triangle inequality.
\begin{lemma}[Triangle inequality for couplings]
\label{lem:coupling_triangle_ineq}
Suppose $X,Y,Z$ are random variables such that $|X-Y|\le_{\eta_1} h_1$ and $|Y-Z|\le_{\eta_2} h_2$. Then $|X-Z|\le_{\eta_1 +\eta_2} h_1+h_2.$
\end{lemma}
\begin{proof}
There exists couplings $(X,Y)$ and $(Y,Z)$ such that
$$\Pr[|X-Y|\ge h_1] \le \eta_1 \text{ and } \Pr[|Y-Z|\ge h_2] \le \eta_2.$$ From these two couplings, we can construct a coupling between $(X,Z)$: sample $X$, sample $Y$ from $Y|X$ (given by coupling $(X,Y)$) and finally sample $Z$ from $Z|Y$ (given by coupling $(Y,Z)$). Therefore for this coupling, we have:
\begin{align*}
\Pr[|X-Z|\ge h_1+h_2] &\le \Pr[|(X-Y) + (Y-Z)|\ge h_1+h_2]\\
&\le \Pr[|X-Y| + |Y-Z|\ge h_1+h_2]\\
&\le \Pr[|X-Y|\ge h_1]+\Pr[|Y-Z|\ge h_2]\\
&\le \eta_1+\eta_2.
\end{align*}
\end{proof}
The following lemma shows that small total variation distance implies good coupling approximation.
\begin{lemma}
\label{lem:TV_coupling}
If the total variation distance $d_{TV}(X,Y)\le \eta$, then $|X-Y| \le_\eta 0$.
\end{lemma}
\begin{proof}
It is well known that for any two random variables $X,Y$, there exists a coupling such that $d_{TV}(X,Y)=\Pr[X\ne Y]$. This immediately implies what we want.
\end{proof}
\subsection{Bounding the error using tail bounds of PRVs}
The goal of this section is to bound the error of \textsf{ComposePRV} in terms of the tail bounds of the underlying PRVs.
\begin{theorem}
\label{thm:approximation}
Let $Y_1,Y_2,\dots,Y_k$ be PRVs and let $\widetilde{Y}$ be the approximation produced by the \textsf{ComposePRV} algorithm (Algorithm~\ref{alg:PrivacyComposition}) for $Y=\sum_{i=1}^k Y_i$ with truncation parameter $L$ and mesh size $$h=\frac{\epsilon_\mathrm{error}}{\sqrt{\frac{k}{2}\log \frac{2}{\eta_0}}}.$$ Then $$\delta_{\widetilde{Y}}(\epsilon+\epsilon_{\mathrm{error}}) -\delta_{\mathrm{error}}\le \delta_Y(\epsilon) \le \delta_{\widetilde{Y}}(\epsilon-\epsilon_{\mathrm{error}})+\delta_\mathrm{error}$$ where
\begin{align*}
\delta_\mathrm{error}&= \Pr\left[\left|\sum_{i=1}^k \widetilde{Y}_i\right|\ge L\right] + \sum_{i=1}^k \Pr[|Y_i|\ge L]+ \eta_0\\
&\le \Pr\left[\left|\sum_{i=1}^k Y_i\right|\ge L-\epsilon_\mathrm{error}\right] + 2\sum_{i=1}^k \Pr[|Y_i|\ge L]+ 2\eta_0.
\end{align*}
\end{theorem}
\begin{remark}
We can directly bound $\Pr\left[\left|\sum_{i=1}^k \widetilde{Y}_i\right|\ge L\right]$ using moment generating functions as $$\Pr\left[\left|\sum_{i=1}^k \widetilde{Y}_i\right|\ge L\right] \le \inf_{\lambda >0} \frac{\prod_{i=1}^k\E[\exp(\lambda \widetilde{Y}_i)]}{e^{\lambda L}} + \inf_{\lambda >0} \frac{\prod_{i=1}^k\E[\exp(-\lambda \widetilde{Y}_i)]}{e^{\lambda L}}.$$ Sometimes, if we already have good upper bound for $\Pr\left[ \left|\sum_i Y_i \right|\ge L\right]$, then the second bound on $\delta_\mathrm{error}$ is useful.
\end{remark}
The following key lemma shows that the \textsf{DiscretizePRV} algorithm (Algorithm~\ref{alg:discretizePRV}) produces a good coupling approximation to the PRV and preserves the mean.
\begin{lemma}
\label{lem:discretizePRV_coupling}
Given a PRV $Y$, let $Y^L = Y\big|_{|Y|\le L}$ be its truncation. The approximation $\widetilde{Y}$ returned by \textsf{DiscretizePRV} satisfies $\E[\widetilde{Y}]=\E[Y^L]$ and $|Y^L-(\widetilde{Y}-\mu)|\le_{0} \frac{h}{2}$ for some $\mu$ where $h$ is the mesh size. We also have $|Y^L - Y|\le_\eta 0$ where $\eta=\Pr[|Y|\ge L].$
\end{lemma}
\begin{proof}
Since $d_{TV}(Y,Y^L) \le \Pr[|Y|\ge L]=\eta$, by Lemma~\ref{lem:TV_coupling}, $|Y-Y^L|\le_\eta 0$. It is clear that by construction $\E[\widetilde{Y}]=\E[Y^L],$
$$\E[\widetilde{Y}]=\mu + \sum_{i=-n}^{n} ih \cdot q_i = \left( \E[Y^L] - \sum_{i=-n}^{n} ih \cdot q_i\right) + \sum_{i=-n}^{n} ih \cdot q_i =\E[Y^L].$$ We will now construct the coupling between $(Y^L,\widetilde{Y})$ such that $|Y^L-(\widetilde{Y}-\mu)|\le \frac{h}{2}$. The coupling is as follows: First sample $y\sim Y^L$. Suppose $y\in (ih-\frac{h}{2},ih+\frac{h}{2}]$ for some integer $i$ such that $-n\le i \le n$, then return $\widetilde{y} = \mu + ih$. Clearly, the distribution of $\widetilde{y}$ matches with $\widetilde{Y}$ and $|y-(\widetilde{y}-\mu)| = |y-ih|\le \frac{h}{2}$.
\end{proof}
Since our error bound on $\widetilde{Y}$ is slightly different from the assumption in Lemma \ref{lem:coupling_sum_independent}, we need the following generalization using the same proof.
\begin{lemma}
\label{lem:coupling_sum_independent2}
Suppose $Y_1,Y_2,\dots,Y_k$ and $\widetilde{Y}_1,\widetilde{Y}_2,\dots,\widetilde{Y}_k$ are two collections of independent random variables such that $|Y_i-(\widetilde{Y}_i - \mu_i)|\le_0 h$ for some $\mu_i$ and $\E[Y_i]=\E[\widetilde{Y}_i]$ for all $i$, then $$\left|\sum_{i=1}^k Y_i - \sum_{i=1}^k \widetilde{Y}_i\right|\le_{\eta}h\sqrt{2k\log{\frac{2}{\eta}}}.$$
\end{lemma}
In the algorithm, we only calculate the distribution of $Y_1 \oplus Y_2 \oplus \dots \oplus Y_k$ instead of $Y_1+Y_2+\dots+Y_k$. The following simple lemma shows that this is still a good approximation as long as $\sum_i Y_i$ stays within $[-L,L]$ with high probability.
\begin{lemma}
\label{lem:convolution_wrapping_around_error}
Let $Y_1,Y_2,\dots,Y_k$ be random variables supported on $(-L,L]$.
Then $$\left|\sum_{i=1}^k Y_i - \left( Y_1 \oplus_L Y_2 \oplus_L \cdots \oplus_L Y_k\right)\right| \le_\eta 0$$ where $$\eta = \Pr\left[\left|\sum_{i=1}^k Y_i\right|\ge L\right].$$
\end{lemma}
\begin{proof}
$$\Pr\left[\sum_{i=1}^k Y_i \ne \left( Y_1 \oplus_L Y_2 \oplus_L \cdots \oplus_L Y_k\right) \right] \le \Pr\left[\left|\sum_{i=1}^k Y_i\right|\ge L\right].$$ This clearly implies what we want.
\end{proof}
Combining all the above lemmas, we get the following corollary.
\begin{corollary}
\label{cor:composePRV_coupling}
Let $Y_1,Y_2,\dots,Y_k$ be random variables supported on and let $\widetilde{Y}_i$ be the discretization of $Y_i$ produced by \textsf{DiscretizePRV} algorithm with mesh size $h=\frac{h_0}{\sqrt{\frac{k}{2}\log \frac{2}{\eta_0}}}$ and truncation parameter $L$. Then
$$\left|(Y_1+Y_2+\dots+Y_k) - (\widetilde{Y}_1 \oplus \widetilde{Y}_2 \oplus \dots \oplus \widetilde{Y}_k)\right|\le_\eta h_0$$ where
$$\eta = \Pr\left[\left|\sum_{i=1}^k \widetilde{Y}_i\right|\ge L\right] + \sum_{i=1}^k \Pr[|Y_i|\ge L]+ \eta_0.$$ Furthermore, we can bound $$\Pr\left[\left|\sum_{i=1}^k \widetilde{Y}_i\right|\ge L\right] \le \Pr\left[\left|\sum_{i=1}^k Y_i\right|\ge L-h_0\right]+\sum_{i=1}^k \Pr[|Y_i|\ge L]+ \eta_0.$$
\end{corollary}
\begin{proof}
Let $Y^L\equiv Y_i\big|_{|Y_i|\le L}$ be the truncation of $Y_i.$
By Lemma~\ref{lem:discretizePRV_coupling}, $|Y_i^L - (\widetilde{Y}_i -\mu_i)| \le_{0} \frac{h}{2}$ for some $\mu_i$ and $|Y_i^L - Y_i|_{\xi_i} \le 0$ where $\xi_i = \Pr[|Y_i|\ge L].$ Now applying the triangle inequality for coupling approximations (Lemma~\ref{lem:coupling_triangle_ineq}), we have
$$\left|\sum_i Y_i - \sum_i Y_i^L\right| \le_{\eta_1} 0$$ where $\eta_1=\sum_i \xi_i=\sum_i \Pr[|Y_i|\ge L].$ By Lemma~\ref{lem:coupling_sum_independent2}, we have
$$\left|\sum_i Y_i^L - \sum_i \widetilde{Y}_i\right| \le_{\eta_0} \frac{h}{2}\cdot \sqrt{2k\log \frac{2}{\eta_0}} = h \sqrt{\frac{k}{2}\log \frac{2}{\eta_0}}=h_0.$$
By Lemma~\ref{lem:convolution_wrapping_around_error},
$$\left|\sum_{i=1}^k \widetilde{Y}_i - \left( \widetilde{Y}_1 \oplus_L \widetilde{Y}_2 \oplus_L \cdots \oplus_L \widetilde{Y}_k\right)\right| \le_{\eta_2} 0$$ where $\eta_2 = \Pr\left[\left|\sum_{i=1}^k \widetilde{Y}_i\right|\ge L\right].$ Finally applying triangle inequality (Lemma~\ref{lem:coupling_triangle_ineq}) once again, we get:
$$\left|(Y_1+Y_2+\dots+Y_k) - (\widetilde{Y}_1 \oplus_L \widetilde{Y}_2 \oplus_L \dots \oplus_L \widetilde{Y}_k)\right|\le_\eta h_0$$ where $\eta=\eta_0+\eta_1+\eta_2$. We can bound $\Pr\left[\left|\sum_{i=1}^k \widetilde{Y}_i\right|\ge L\right]$ as:
\begin{align*}
\Pr\left[\left|\sum_i \widetilde{Y}_i\right|\ge L\right] &= \Pr\left[\left|\sum_i (\widetilde{Y}_i-Y_i^L) + \sum_i (Y_i^L - Y_i) +\sum_i Y_i \right|\ge L\right]\\
&\le \Pr\left[\left|\sum_i (\widetilde{Y}_i-Y_i^L) \right|+\left|\sum_i (Y_i^L - Y_i) \right| +\left|\sum_i Y_i \right|\ge h_0 + 0 + L-h_0\right]\\
&\le \Pr\left[\left|\sum_i (\widetilde{Y}_i-Y_i^L) \right| > h_0\right] +\Pr\left[\left|\sum_i (Y_i^L - Y_i) \right| > 0\right] +\Pr\left[\left|\sum_i Y_i \right|\ge L-h_0\right]\\
&\le \eta_0 + \eta_1 + \Pr\left[\left|\sum_{i=1}^k Y_i\right|\ge L-h_0\right].
\end{align*}
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:approximation}]
Combining Corollary~\ref{cor:composePRV_coupling} (with $h_0=\epsilon_\mathrm{error}$) and Lemma~\ref{lem:coupling_to_privacy_curves}, we have Theorem~\ref{thm:approximation}.
\end{proof}
\subsection{Tail Bound for PRVs}
\label{sec:appendix_numerical_composition}
To finish the proof of our main theorem (Theorem~\ref{thm:approximation_eps_upper_bound}, we need a tail bound on PRVs in terms of their privacy curves. First, we need a lemma relating the PRVs of a privacy curve $\delta(P||Q)$ with the PRVs of $\delta(Q||P)$.
\begin{proposition}
\label{prop:finv_prv}
Let $(X,Y)$ be the PRVs for a privacy curve $\delta(P||Q)$. Then the PRVs for the privacy curve $\delta(Q||P)$ are $(-Y,-X).$
\end{proposition}
\begin{proof}
Let $(\widetilde{X},\widetilde{Y})$ be the PRVs for $\delta(Q||P)$. We know that $\delta(P||Q)=\delta(X||Y)$. So $\delta(Q||P)=\delta(Y||X).$ Then by Theorem~\ref{thm:PRV_LLRV},
\begin{align*}
\widetilde{X} &= \log\left( \frac{X(t)}{Y(t)}\right) \text{ where } t \sim Y\\
&= \log\left( e^{-t}\right) \text{ where } t \sim Y\\
& = -Y.
\end{align*}
\begin{align*}
\widetilde{Y} &= \log\left( \frac{X(t)}{Y(t)}\right) \text{ where } t \sim X\\
&= \log\left( e^{-t}\right) \text{ where } t \sim X\\
& = -X.
\end{align*}
\end{proof}
Now, we show our tail bound, which shows the PRVs $(X,Y)$ for a $(\epsilon,\delta)$-DP algorithm satisfies roughly that $\Pr(|Y| \geq \epsilon + 2) \leq 2 \delta$.
\PRVtail*
\begin{proof}
We have $\delta(X||Y)\le f_{\epsilon,\delta}$ and $\delta(Y||X)\le f_{\epsilon,\delta}$ where $f_{\epsilon,\delta}$ is the privacy curve of a $(\epsilon,\delta)$-DP algorithm.
By Theorem~\ref{thm:PRV_LLRV}, we have
\begin{align*}
\delta &\ge \int_0^\infty \Pr[Y\ge \epsilon+s] e^{-s} ds\\
&\ge \int_0^t \Pr[Y\ge \epsilon+s] e^{-s} ds\\
&\ge \Pr[Y\ge \epsilon+t]\int_0^t e^{-s} ds\\
&\ge \Pr[Y\ge \epsilon+t] (1-e^{-t}).
\end{align*}
By Proposition~\ref{prop:finv_prv}, the PRVs for $\delta(Y||X)$ are $(-Y,-X)$. Therefore by a similar argument, we have
$$\Pr[X \le -\epsilon-t]=\Pr[-X\ge \epsilon+t] \le \frac{\delta}{1-e^{-t}}.$$
Finally, note that $Y(s)=e^s X(s)$ for all $s\in \R$ and $Y(-\infty)=0$ by the definition of PRVs. Therefore $$\Pr[Y\le -\epsilon-t] \le e^{-\epsilon-t}\Pr[X\le -\epsilon-t].$$
Therefore we have:
\begin{align*}
\Pr[|Y|\ge \epsilon+t] &= \Pr[Y \ge \epsilon+t] + \Pr[Y \le -\epsilon-t]\\
&\le \Pr[Y \ge \epsilon+t] + e^{-\epsilon-t}\Pr[X \le -\epsilon-t]\\
&\le \left( 1 + e^{-\epsilon-t}\right)\frac{\delta}{1-e^{-t}}.
\end{align*}
\end{proof}
\subsection{Proof of Theorem~\ref{thm:approximation_eps_upper_bound}}
Now, we can prove our main theorem.
\mainthm*
\begin{proof}
By Lemma~\ref{lem:PRVs_tailbounds},
\begin{align*}
\Pr[|Y_i| \ge L] &= \Pr[|Y_i| \ge L-2 + 2]\\
&\le \delta_{Y_i}(L-2) \cdot \frac{1+e^{-L}}{1-e^{-2}}\\
&\le \delta_{Y_i}(L-2) \cdot \frac{1+e^{-2}}{1-e^{-2}}\\
&\le \delta_{Y_i}(L-2) \cdot \frac{4}{3}.
\end{align*}
Therefore we have $$\sum_{i=1}^k \Pr[|Y_i|\ge L]\le \frac{4}{3} \sum_{i=1}^k \delta_{Y_i}(L-2) \le \frac{4}{3} \cdot \frac{\delta_\mathrm{error}}{8} = \frac{\delta_\mathrm{error}}{6}.$$
Similarly $$\Pr[|Y| \ge L-\epsilon_\mathrm{error}] \le \frac{4}{3}\delta_Y(L-2-\epsilon_\mathrm{error}) \le \frac{\delta_\mathrm{error}}{3}.$$ Therefore by Theorem~\ref{thm:approximation}, setting $\eta_0=\frac{\delta_\mathrm{error}}{6}$, we get the desired result.
For the runtime, we note that the bottleneck of our algorithm is to compute the convolution, which can be done using FFT. In total, we need to compute $b+1$ many FFT for $b$ distinct algorithms, one for each for computing the Fourier transform and one of computing the inverse Fourier transform. Since the length of the array for the FFT is bounded by $O(L/h)$, this costs $O(b L/h \log(L/h))$ in total.
The step $\delta_{\widetilde{Y}}(\epsilon) = \E_{\widetilde{Y}}\left[ \left( 1-e^{\epsilon-\widetilde{Y}}\right)_+\right]$ can be computed in linear time by first computing the CDF of $\widetilde{Y}$ and the prefix sum $\E_{\widetilde{Y} \leq \alpha}\left[ e^{-\widetilde{Y}}\right]$ for all $\alpha$.
\end{proof}
\section{Privacy Loss Random Variables}
\label{sec:appendix_PRV}
In this section, we continue the discussion on privacy random variables in Section \ref{sec:PRV}. First, we give the proof of the formula for PRVs of $\delta(P||Q)$ and the formula for a privacy curve given its PRVs (Theorem \ref{thm:PRV_LLRV}).
{\renewcommand\footnote[1]{}\PRVLLRV*}
\begin{proof}
We will first verify that $Y(t)=e^t X(t).$ This is equivalent to proving that $\E_Y[\phi(Y)]=\E_X[\phi(X)e^X]$ for any test function $\phi:\overline{\R} \to [0,1]$. This is true since
\begin{align*}
\E_Y[\phi(Y)]&= \E_{\omega\sim Q}\left[ \phi\left(\log\left( \frac{Q(\omega)}{P(\omega)}\right)\rp\right] \\
&= \E_{\omega\sim P}\left[ \phi\left(\log\left( \frac{Q(\omega)}{P(\omega)}\right)\rp \frac{Q(\omega)}{P(\omega)}\right]\\
&= \E_X[\phi(X)e^X].
\end{align*}
We will now prove that $\delta(X||Y)=\delta(P||Q).$
We have
\begin{align*}
\delta(P||Q)(\epsilon) &= \sup_{S \subset \Omega} \Pr[Q\in S] - e^\epsilon\Pr[P\in S]\\
&= \Pr[Q\in S_\epsilon] - e^\epsilon\Pr[P\in S_\epsilon]
\end{align*}
where $$S_\epsilon = \left\{\omega\in \Omega: \frac{Q(\omega)}{P(\omega)} > e^\epsilon \right\}=\left\{\omega\in \Omega: \log\left(\frac{Q(\omega)}{P(\omega)}\right) > \epsilon \right\}.$$
Therefore $\Pr[Q\in S_\epsilon]=\Pr[Y > \epsilon]$ and $\Pr[P\in S_\epsilon]=\Pr[X > \epsilon]$. To complete the proof, note that
\begin{align*}
\delta(X||Y)(\epsilon) &= \sup_{T \subset \overline{\R}} \Pr[Y\in T] - e^\epsilon\Pr[X\in T]\\
&= \Pr[Y\in T_\epsilon] - e^\epsilon\Pr[X\in T_\epsilon]
\end{align*}
where $$T_\epsilon = \left\{t\in \overline{\R}: \frac{Y(t)}{X(t)} > e^\epsilon \right\}= \left\{t\in \overline{\R}: e^t > e^\epsilon \right\}=(\epsilon,\infty].$$
Putting it all together, we have:
\begin{align*}
\delta(P||Q)(\epsilon)&=\Pr[Y>\epsilon] - e^\epsilon \Pr[X>\epsilon]=\delta(X||Y).
\end{align*}
\end{proof}
{\renewcommand\footnote[1]{}\PRVprivacycurve*}
\begin{proof}
Since the PDFs of PRVs $(X,Y)$ satisfy the relation $Y(t)=e^tX(t)$, we can rewrite the equation~\ref{eqn:delta_PRV_simple} in terms of just $Y$ or just $X$.
\begin{align*}
\delta(\epsilon)&=\Pr[Y\ge \epsilon]-e^\epsilon \Pr[X\ge \epsilon]\\
&=\int_\epsilon^\infty Y(t) dt - \int_\epsilon^\infty e^\epsilon X(t) dt\\
&=\int_\epsilon^\infty Y(t) dt - \int_\epsilon^\infty e^{\epsilon-t} Y(t) dt \tag{Since $Y(t)=e^tX(t)$}\\
&=\int_\epsilon^\infty Y(t) (1-e^{\epsilon-t}) dt\\
&=\int_{-\infty}^\infty Y(t) (1-e^{\epsilon-t})_+ dt\\
&=\E_Y[(1-e^{\epsilon-Y})_+]
\end{align*}
To get the other form for $\delta(\epsilon)$, we use the integration by parts formula.
\begin{align*}
\delta(\epsilon)&=\int_{\epsilon}^\infty Y(t) (1-e^{\epsilon-t}) dt\\
&=\int_{\epsilon}^\infty Y(t) dt + \int_{\epsilon}^\infty \left( -Y(t) \right) e^{\epsilon-t} dt\\
&=\Pr[Y\ge \epsilon] + \left( \Pr[Y\ge t]e^{\epsilon-t}\Big|_\epsilon^\infty - \int_{\epsilon}^\infty \Pr[Y\ge t] \left( -e^{\epsilon-t}\right) dt \right)\\
&=\Pr[Y\ge \epsilon] - \Pr[Y\ge \epsilon] + \int_{\epsilon}^\infty \Pr[Y\ge t] e^{\epsilon-t} dt\\
&=\int_\epsilon^\infty e^{\epsilon-t} \Pr[Y\ge t] dt \\
&=\int_0^\infty e^{-z} \Pr[Y\ge \epsilon+z] dz \tag{Substituting $z=t-\epsilon$}\\
&=\Pr[Y\ge \epsilon+Z]. \tag{where $Z$ is an exponential random variable}
\end{align*}
We now prove the converse relation by differentiating the expression for $\delta(\epsilon)$ twice. We have:
\begin{align*}
&\delta(\epsilon)=\int_\epsilon^\infty Y(t) dt - e^{\epsilon}\int_\epsilon^\infty e^{-t} Y(t) dt\\
\implies & \delta'(\epsilon)=-Y(\epsilon)+e^{\epsilon}\cdot e^{-\epsilon}Y(\epsilon)-e^{\epsilon}\cdot \int_\epsilon^\infty e^{-t} Y(t) dt = -e^{\epsilon}\cdot \int_\epsilon^\infty e^{-t} Y(t) dt \\
\implies & e^{-\epsilon}\delta'(\epsilon)=-\int_\epsilon^\infty e^{-t} Y(t) dt \\
\implies & e^{-\epsilon}\delta''(\epsilon)-e^{-\epsilon}\delta'(\epsilon)=e^{-\epsilon} Y(\epsilon) \\
\implies & Y(\epsilon)=\delta''(\epsilon)-\delta'(\epsilon).
\end{align*}
\end{proof}
\subsection{Examples of privacy loss random variables}
In this section, we state the PRVs for a few standard mechanisms.
\begin{proposition}[Gaussian Mechanism]
The PRVs for $\delta(\mathcal N(\mu,1)||\mathcal N(0,1))$ are:
$$X=\mathcal N(-\mu^2/2,\mu^2) \text{ and } Y=\mathcal N(\mu^2/2,\mu^2).$$
\end{proposition}
\begin{proof}
Let $P=\mathcal N(\mu,1)$ and $Q=\mathcal N(0,1)$.
By Theorem~\ref{thm:PRV_LLRV},
\begin{align*}
Y &\sim \log\left( \frac{Q(t)}{P(t)}\right) \text{ where } t\sim Q\\
&\sim \log\left( \frac{\exp(-t^2/2)}{\exp(-(t-\mu)^2/2)}\right) \text{ where } t\sim \mathcal N(0,1)\\
&\sim \frac{(t-\mu)^2}{2} - \frac{t^2}{2} \text{ where } t\sim \mathcal N(0,1)\\
&\sim \frac{\mu^2}{2}-\mu t \text{ where } t\sim \mathcal N(0,1)\\
&= \mathcal N\left( \frac{\mu^2}{2},\mu^2 \right).
\end{align*}
A similar calculation shows that $X=\mathcal N\left( -\frac{\mu^2}{2},\mu^2\right)$
\end{proof}
\begin{proposition}[Laplace Mechanism]
The PRVs for the privacy curve $\delta\left(\mathsf{Lap}\left( \mu,1\right)||\mathsf{Lap}\left( 0,1\right)\rp$ are:
$$X=|Z|-|Z-\mu| \text{ and } Y= |Z-\mu|-|Z|$$ where $Z\sim \mathsf{Lap}(0,1).$
\end{proposition}
\begin{proof}
Let $P=\mathsf{Lap}(\mu,1)$ and $Q=\mathsf{Lap}(0,1)$.
By Theorem~\ref{thm:PRV_LLRV},
\begin{align*}
Y &\sim \log\left( \frac{Q(t)}{P(t)}\right) \text{ where } t\sim Q\\
&\sim \log\left( \frac{\exp(-|t|)}{\exp(-|t-\mu|}\right) \text{ where } t\sim \mathsf{Lap}(0,1)\\
&\sim |t-\mu|-|t| \text{ where } t\sim \mathsf{Lap}(0,1)\\
&= |Z-\mu|-|Z| \text{ where } Z\sim \mathsf{Lap}(0,1).
\end{align*}
A similar calculation shows that $X=|Z|-|Z-\mu|$ where $Z\sim\mathsf{Lap}(0,1).$
\end{proof}
\begin{proposition}[$(\epsilon,\delta)$-DP]
The PRVs for the privacy curve of a $(\epsilon,\delta)$-DP algorithm are
$$X=\begin{cases}
-\infty & \text{w.p. } \delta\\
-\epsilon & \text{w.p. } \frac{(1-\delta)e^\epsilon}{e^\epsilon+1}\\
\epsilon & \text{w.p. } \frac{1-\delta}{e^\epsilon+1},
\end{cases}$$
$$Y=\begin{cases}
-\epsilon & \text{w.p. } \frac{1-\delta}{e^\epsilon+1}\\
\epsilon & \text{w.p. } \frac{(1-\delta)e^\epsilon}{e^\epsilon+1}\\
\infty & \text{w.p. } \delta.
\end{cases}$$
\end{proposition}
\begin{proof}
It is easy to verify that $Y(t)=e^t X(t)$ for all $t\in \R$. We can also verify that $$\delta(\epsilon)=\Pr[Y> \epsilon] - e^\epsilon \Pr[X > \epsilon]=\delta.$$ Morever $X=-Y$, therefore the privacy curve $\delta(X||Y)$ is symmetric by Proposition~\ref{prop:finv_prv}, i.e., $\delta(X||Y)=\delta(Y||X)$. These conditions together imply that $X,Y$ are PRVs for the $(\epsilon,\delta)$-DP curve.
\end{proof}
Note that in the all the above examples, we have $X=-Y$ as the privacy curves are symmetric.
\subsection{Subsampling}
In this section, we calculate the PRVs for a subsampled mechanism given the PRVs for the original mechanism. Given two random variables $P,Q$ and a sampling probability $p\in [0,1]$, $p\cdot P+(1-p)\cdot Q$ denotes the mixture where we sample $P$ w.p. $p$ and $Q$ w.p. $1-p.$
\begin{proposition}
\label{prop:mixture_f_p_privacy_rv}
Let $(X,Y)$ be the PRVs for a privacy curve $\delta(P||Q)$. Let $(X_p,Y_p)$ be the PRVs for $\delta_p=\delta(P||\ p\cdot P+(1-p)\cdot Q)$. Then
\begin{align*}
X_p&=\log(1+p(e^X-1)),\\
Y_p&=\begin{cases}
\log(1+p(e^Y-1)) \text{ w.p. } p\\
\log(1+p(e^X-1)) \text{ w.p. } 1-p.
\end{cases}
\end{align*}
The CDFs of $X_p$ and $Y_p$ are given by:
\begin{align*}
\mathrm{CDF}_{X_p}(t) &= \begin{cases}
\mathrm{CDF}_X\left(\log\left( \frac{e^t-(1-p)}{p}\right) \right) &\text{ if } t\ge \log(1-p)\\
0 &\text{ if } t< \log(1-p)
\end{cases}\\
\mathrm{CDF}_{Y_p}(t) &= \begin{cases}
p\cdot \mathrm{CDF}_Y\left( \log\left( \frac{e^t-(1-p)}{p}\right) \right) + (1-p)\cdot \mathrm{CDF}_X\left( \log\left( \frac{e^t-(1-p)}{p}\right) \right) &\text{ if } t\ge \log(1-p)\\
0 & \text{ if } t <\log(1-p).
\end{cases}
\end{align*}
\end{proposition}
\begin{proof}
By Theorem~\ref{thm:PRV_LLRV},
\begin{align*}
X_p &= \log\left( \frac{pY(t)+(1-p)X(t)}{X(t)}\right) \text{ where } t\sim X\\
& = \log \left( p e^t + 1-p\right) \text{ where } t \sim X\\
& = \log \left( p e^X + 1-p\right).
\end{align*}
Similarly,
\begin{align*}
Y_p &= \log\left( \frac{pY(t)+(1-p)X(t)}{X(t)}\right) \text{ where } t\sim pY+(1-p)X\\
& = \log \left( p e^t + 1-p\right) \text{ where } t \sim pY+(1-p)X\\
& = \begin{cases}
\log(1+p(e^Y-1)) \text{ w.p. } p\\
\log(1+p(e^X-1)) \text{ w.p. } 1-p.
\end{cases}
\end{align*}
The CDF of $X_p$ is given by:
\begin{align*}
\Pr[X_p \le t] &= \Pr\left[ \log \left( p e^X + 1-p\right) \le t\right]\\
&= \Pr\left[ X \le \log\left( \frac{e^t-(1-p)}{p}\right) \right]
\end{align*}
The CDF of $Y_p$ is given by:
\begin{align*}
\Pr[Y_p \le t] &= p \Pr\left[ \log \left( p e^Y + 1-p\right) \le t\right] + (1-p) \Pr\left[ \log \left( p e^X + 1-p\right) \le t\right]\\
&= p\Pr\left[ Y \le \log\left( \frac{e^t-(1-p)}{p}\right) \right] + (1-p)\Pr\left[ X \le \log\left( \frac{e^t-(1-p)}{p}\right) \right].
\end{align*}
\end{proof}
\section{Effect of floating point arithmetic}
\label{sec:numerical_precision}
In this section, we demonstrate the effect of floating point inaccuracies on the computed privacy parameters.
Figure \ref{fig:num_error} compares lower and upper bounds of the privacy curve with the analytical solution for small values of $\delta$.
As mentioned in section \ref{sec:experiments}, we use a floating point representation with a resolution of at least $10^{-15}$.
The number of discretization points in this examples are on the order of $10^4$.
Consequently, we expect floating point inaccuracies to become dominant for values on the order of $10^{-11}$.
This can be also seen in the illustration, where the lower and upper bound fail to produce meaningful results for $\delta < 2 \times 10^{-11}$.
\begin{figure}[h]
\centering
\resizebox{!}{0.4\textwidth}{\large \input{Figures/convergence_exact-eps-delta-num_error}}
\caption{Setting $p=1$ and comparing to the analytical solution~(\ref{eqn:analytical_solution_p=1}) for values of $\delta$ beyond expected floating point accuracy.}
\label{fig:num_error}
\end{figure}
\subsection{Comparison with \cite{KoskelaJPH21}}
\label{sec:runtime_experiments}
\begin{figure}[h]
\centering
\resizebox{!}{0.4\textwidth}{\large \input{Figures/convergence_num_points}}
\caption{Comparison of error bounds of $\delta$ with varying number of discretisation points for $p=4\times10^{-3}, \sigma=0.8, \varepsilon=1.5, k=1000$.}
\label{fig:comparen}
\end{figure}
\begin{figure}[h]
\begin{subfigure}[b]{0.33\textwidth}
\hspace{-1.5cm}\resizebox{1.7\linewidth}{!}{\input{Figures/comparison-10-eps-delta}}
\caption{$k=10$}
\end{subfigure}
\begin{subfigure}[b]{0.33\textwidth}
\hspace{-1.5cm}\resizebox{1.7\linewidth}{!}{\input{Figures/comparison-100-eps-delta}}
\caption{$k=100$}
\end{subfigure}
\begin{subfigure}[b]{0.33\textwidth}
\hspace{-1.5cm}\resizebox{1.7\linewidth}{!}{\input{Figures/comparison-1000-eps-delta}}
\caption{$k=1000$}
\end{subfigure}
\caption{
Comparing different error bounds using the same mesh size $8\times10^{-4}$ under different number of steps $k = 10, 100, 1000$. (With $p=10^{-2}$, $\sigma=0.8$.)
\label{fig:comparek}}
\end{figure}
\begin{figure}[h]
\begin{subfigure}[b]{0.49\textwidth}
\resizebox{\linewidth}{!}{\large \input{Figures/runtime_comparison_runtime_combined}}
\end{subfigure}
\qquad
\begin{subfigure}[b]{0.49\textwidth}
\resizebox{\linewidth}{!}{\large \input{Figures/runtime_comparison_alignment}}
\end{subfigure}
\caption{(a) Comparing runtimes for our algorithm with that of~\cite{KoskelaJPH21} when aligned on accuracy for $\sigma = 0.8$, $p=4\times10^{-3}$.
We can see a significant reduction in runtime in particular for large number of DPSGD steps. We were not able to run the algorithm of~\cite{KoskelaJPH21} beyond 2,000 steps, since it becomes unstable beyond that point.\protect\footnotemark.
We also plot the speed up directly on the secondary $y$-axis. (b) Verification of the alignment of the error bounds of both algorithms at $\varepsilon=1.5$.}
\label{fig:runtimes}
\end{figure}
\footnotetext{We are using the implementation of~\cite{KoskelaJPH21} from~\cite{PLDAccountant}.}
In this section, we provide more results demonstrating the practical use of our algorithm.
We compare runtimes of our algorithm with~\cite{KoskelaJPH21}, which is the state-of-the-art, for typical values of privacy parameters ($\sigma = 0.8$, $p=4\times10^{-3}$, $\varepsilon=1.5$).
See Figure~\ref{fig:comparen} for the effect of the number of discretisation points $n$ on the accuracy of $\delta$.
Our algorithm requires about a few orders of magnitude smaller number of discretization points to converge compared to the algorithm of \cite{KoskelaJPH21}.
A similar picture can be seen in Figure~\ref{fig:comparek}.
While for a small number of compositions, the algorithm of ~\cite{KoskelaJPH21} gives reasonable estimates, for a large number of compositions, their error bounds worsen quickly.
We note that runtimes are directly proportional to the memory required by the algorithms and so a separate memory analysis is not required; the runtime and memory are dominated by the number of points in the discretization of PRV.
All experiments are performed on a Intel Xeon W-2155 CPU with 3.30GHz with 128GB of memory.
In order to compare runtimes, we align the accuracy of both FFT algorithms.
We find sets of numerical parameters (number of discretization bins and domain length) such that both algorithms give similarly accurate bounds and verify it visually (see Figure \ref{fig:runtimes} (b)).
Figure \ref{fig:runtimes} illustrates the runtimes for varying numbers of DPSGD steps.
We observe a significant reduction in the runtime using our algorithms.
\subsection{Our Contributions}
The main contribution of this work is a new algorithm with an improved analysis for computing the privacy curve of the composition of a large number of DP algorithms.
\begin{theorem}[Informal version of Theorem~\ref{thm:approximation_eps_upper_bound}]
\label{thm:informal_main}
Suppose $M_1,M_2,\dots,M_k$ are DP algorithms. Then the privacy curve $\delta_M(\epsilon)$ of adaptive composition $M=M_1 \circ M_2 \circ \dots \circ M_k$ can be approximated in time
\begin{equation}
\label{eqn:time_ours}
O\left(\frac{\epsilon_\mathrm{upper} \ k^{1.5}\log{k}\sqrt{\log \frac{1}{\delta_\mathrm{error}}}}{\epsilon_\mathrm{error}}\right),
\end{equation}
where $\epsilon_\mathrm{error}$ is the additive error in $\epsilon$, $\delta_\mathrm{error}$ is the additive error in $\delta$ and $\epsilon_\mathrm{upper}$ is an upper bound on $\max\left\{\epsilon_{M}(\delta_\mathrm{error}),\max_i \epsilon_{M_i}\left(\frac{\delta_\mathrm{error}}{k}\right)\right\}.$\footnote{$\epsilon_{M}(\delta)$ is the inverse of $\delta_{M}(\epsilon)$.}
\end{theorem}
If each $M_i$ satisfies $\left(\frac{1}{\sqrt{k}},\frac{o(1)}{k}\right)$-DP, then by advanced composition (Proposition~\ref{prop:advanced_composition}), we can set $\epsilon_\mathrm{upper}=O(1)$. Therefore the running time of our algorithm in this case is $\tilde{O}\left( \frac{k^{1.5}\sqrt{\log \frac{1}{\delta_\mathrm{error}}}}{\epsilon_\mathrm{error}}\right).$
We can save a factor of $k$, when we compose the same algorithm with itself $k$ times.
\begin{theorem}
Suppose $M$ is a DP algorithm. Then the privacy curve $\delta_{M^{\circ k}}(\epsilon)$ of $M$ (adaptively) composed with itself $k$ times can be approximated in time
\begin{equation}
\label{eqn:time_ours_single_alg}
O\left(\frac{\epsilon_\mathrm{upper} \ k^{\frac{1}{2}}\log{k}\sqrt{\log \frac{1}{\delta_\mathrm{error}}}}{\epsilon_\mathrm{error}}\right),
\end{equation}
where $\epsilon_\mathrm{error}$ is the additive error in $\epsilon$, $\delta_\mathrm{error}$ is the additive error in $\delta$ and $\epsilon_\mathrm{upper}$ is an upper bound on $\max\left\{\epsilon_{M^{\circ k}}(\delta_\mathrm{error}), \epsilon_M\left(\frac{\delta_\mathrm{error}}{k}\right)\right\}.$
\end{theorem}
Thus we improve the state-of-the-art by at least a factor of $k$ in running time. We also note that our algorithm improves the memory required by a factor of $k.$ See Figure~\ref{fig:ours_vs_Koskela} for a comparison of our algorithm with that of~\cite{KoskelaJPH21}. Also note that RDP Accountant (equivalent to the Moments Accountant) significantly overestimates the true $\epsilon$, while the GDP Accountant significantly underestimates the true $\epsilon.$ In contrast, the upper and lower bounds provided by our algorithm lie very close to each other.
\begin{figure}[!h]
\begin{subfigure}[b]{0.45\textwidth}
\resizebox{\linewidth}{!}{\large \input{Figures/dpsgd-vs_fourier}}
\caption{Our algorithm gives much closer upper and lower bounds on the true privacy curve compared to~\cite{KoskelaJPH21}, under the same mesh size of $4\times 10^{-5}$. Our upper and lower bounds are nearly coinciding.}
\end{subfigure}
\qquad
\begin{subfigure}[b]{0.45\textwidth}
\resizebox{\linewidth}{!}{\large \input{Figures/dpsgd_example}}
\caption{Our algorithm can improve significantly over the RDP Accountant. We also see that GDP Accountant can significantly underreport the true $\epsilon$. We have set $\epsilon_\mathrm{error}=0.1$, $\delta_\mathrm{error}=\delta/1000$ here.}
\end{subfigure}
\caption{Case study on DP-SGD. Sampling probability $p=10^{-3}$, noise scale $\sigma=0.8$, $\delta=10^{-7}$.}
\label{fig:ours_vs_Koskela}
\end{figure}
\paragraph{Our Techniques} Our algorithm (also the prior work of~\cite{koskela2020computing}) proceeds by approximating the privacy loss random variables (PRVs) by truncating and discretizing them. We then use Fast Fourier Transform (FFT) to convolve the distributions efficiently. The main difference is in the approximation procedure and the error analysis. In the approximation procedure, we correct the approximation so that the expected value of the discretization matches with the expected value of the PRV.
To analyze the approximation error, we introduce the concept of \emph{coupling approximation} (Definition~\ref{def:couplingapx}), which is a variant of Wasserstein (optimal transport) distance specifically tailored to this application.
We first show that the approximation output by our algorithm to each privacy random variable is a good coupling approximation. We then show that when independent coupling approximations are added, cancellation happens between the errors due to Hoeffding bound, producing a much better coupling approximation than one naively expects from the triangle inequality. This allows us to choose the mesh size in our discretization to be $\approx \frac{1}{\sqrt{k}}$, whereas~\cite{koskela2021computing} choose a mesh size of $\approx \frac{1}{k}$. The other improvement is the truncation procedure. We give a tight tail bound of the PRVs (Lemma~\ref{lem:PRVs_tailbounds}). This allows us to choose the domain size for in truncation to be $\approx \tilde{O}(1)$, whereas~\cite{koskela2021computing} choose $\approx \tilde{O}(\sqrt{k})$. Both ideas together saves a factor of $k$ in the run time and memory.
For the analysis, the previous paper analyzes the discretization error by studying the stability of convolution. This leads to complicated calculations with the runtime linear in $1/{\delta_\mathrm{error}}$ (see (\ref{eqn:time_Koskela})). Since $\delta_\mathrm{error} \ll \delta \ll 1/N$ is required to give meaningful privacy guarantee ($N$ is the number of users), this term $1/{\delta_\mathrm{error}}$ is huge. In this paper, we show various facts about how coupling approximation accumulates and use them to give a runtime depending only on $\sqrt{\log(1/{\delta_\mathrm{error}})}$.
\section{Error analysis}
\label{sec:error_analysis}
To analyze the discretization error, we introduce the notion of \emph{coupling approximation}, a variant of Wasserstein distance. Intuitively, a good coupling approximation is a coupling where the two random variables are close to each other with high probability.
\begin{definition}[coupling approximation]\label{def:couplingapx}
Given two random variables $Y_1,Y_2$, we write $|Y_1-Y_2|\le_{\eta} h$ if there exists a coupling between $Y_1,Y_2$ such that $\Pr[|Y_1-Y_2|>h]\le \eta.$
\end{definition}
The following lemma shows that if we have a good coupling approximation $\widetilde{Y}$ to a PRV $Y$, then the privacy curves $\delta_Y(\epsilon)$ and $\delta_{\widetilde{Y}}(\epsilon)$ should be close.
\begin{lemma}
\label{lem:coupling_to_privacy_curves}
If $Y$ and $\widetilde{Y}$ are two random variables such that $|Y-\widetilde{Y}|\le_\eta h$, then for every $\epsilon\in \R$,
$$\delta_{\widetilde{Y}}(\epsilon+h)-\eta \le \delta_{Y}(\epsilon) \le \delta_{\widetilde{Y}}(\epsilon-h)+\eta.$$
\end{lemma}
\begin{proof} By Theorem \ref{thm:PRV_LLRV}, $\delta_Y(\epsilon)=\Pr[Y\ge \epsilon+Z]$ and hence
\begin{align*}
\delta_Y(\epsilon)
&=\Pr[Y-\widetilde{Y} + \widetilde{Y} \ge \epsilon +Z]\\
&\le\Pr[Y-\widetilde{Y} \ge h] + \Pr[\widetilde{Y} \ge \epsilon -h +Z]\\
&\le\eta+ \delta_{\widetilde{Y}}(\epsilon-h).
\end{align*}
Similarly, we have $\delta_{\widetilde{Y}}(\epsilon) \le \eta + \delta_{Y}(\epsilon-h)$ for all $\epsilon\in \R.$
\end{proof}
Therefore the goal of our analysis is to show that the \textsf{ComposePRV} algorithm finds a good coupling approximation $\widetilde{Y}$ to $Y=\sum_{i=1}^k Y_i.$ We first show that the \textsf{DiscretizePRV} algorithm computes a good coupling approximation to the PRVs and crucially, it preserves the expected value after truncation. Lemma~\ref{lem:discretizePRV_coupling} shows that $|\widetilde{Y}-Y^L|\le_0 h$ where $\widetilde{Y}$ is the approximation of a PRV $Y$ output by Algorithm~\ref{alg:discretizePRV} and $Y^L$ is the truncation of $Y$ to $[-L,L].$
We then use the following key lemma which shows that when we add independent coupling approximations (where expected values match), we get a much better coupling approximation than what the triangle inequality predicts.
\begin{lemma}
\label{lem:coupling_sum_independent}
Suppose $Y_1,Y_2,\dots,Y_k$ and $\widetilde{Y}_1,\widetilde{Y}_2,\dots,\widetilde{Y}_k$ are two collections of independent random variables such that $|Y_i-\widetilde{Y}_i|\le_0 h$ and $\E[Y_i]=\E[\widetilde{Y}_i]$ for all $i$, then $$\left|\sum_{i=1}^k Y_i - \sum_{i=1}^k \widetilde{Y}_i\right|\le_{\eta}h\sqrt{2k\log{\frac{2}{\eta}}}.$$
\end{lemma}
\begin{proof}
Let $X_i=Y_i-\widetilde{Y}_i$ where $(Y_i,\widetilde{Y}_i)$ are coupled such that $|Y_i-\widetilde{Y}_i|\le h$ w.p. $1$. Then $X_i \in [-h,h]$ w.p. $1$. Note that $X_1,X_2,\dots,X_k$ are independent of each other. By Hoeffding's inequality,
$$\Pr\left[\left|\sum_i X_i\right| \ge t\right] \le 2\exp\left(- \frac{2t^2}{k(2h)^2}\right)=\eta$$ if we set $t=h\sqrt{2k\log{\frac{2}{\eta}}}$.
\end{proof}
This lemma shows that the error of $k$ times composition is around $\sqrt{k}\cdot h$ and hence setting $h \approx 1/\sqrt{k}$ gives small enough error. Next, we bound the domain size $L$. Naively, the domain size $L$ should be of the order of $\sqrt{k}$ because $Y$ is the sum of $k$ independent random variables with each bounded by a constant. In the appendix, we give a tighter tail bound of $Y$.
\begin{restatable}{lemma}{PRVtail}
\label{lem:PRVs_tailbounds}
Let $(X,Y)$ be the privacy random variables for a $(\epsilon, \delta)$-DP algorithm, then for any $t\ge 0$, we have
$$\Pr[|Y|\ge \epsilon + t ] \le \frac{\delta \left( 1+e^{-\epsilon-t}\right)}{1-e^{-t}}.$$
\end{restatable}
This shows that $\Pr[|Y| \ge \epsilon + 2] \le \frac{4}{3} \delta$ and hence truncating the domain with $L = 2 + \epsilon$ only introduces an additive $\delta$ error in the privacy curve. Therefore, if the composition satisfies a good privacy guarantee (namely $\epsilon = O(1)$ for small enough $\delta$), we can truncate the domain at $L = \Theta(1)$. Together with the fact that mesh size is $1/\sqrt{k}$, this gives a $O(\sqrt{k})$-time algorithm for computing the privacy curve when we compose the same mechanism with itself $k$ times. The following theorem gives a formal statement of the error bounds of our algorithm, it is proved in Appendix~\ref{sec:error_analysis}.
\begin{restatable}{theorem}{mainthm}
\label{thm:approximation_eps_upper_bound}
Let $\epsilon_\mathrm{error},\delta_\mathrm{error}>0$ be some fixed error terms.
Let $\mathcal M_1,\mathcal M_2,\dots,\mathcal M_k$ be DP algorithms with privacy curves $\delta_{\mathcal M_i}(\epsilon)$. Let $Y_i$ be the PRV corresponding to $\mathcal M_i$ such that $\delta_{\mathcal M_i}(\epsilon)=\delta_{Y_i}(\epsilon)$ for $\epsilon\ge 0$. Let $\mathcal M$ be the (adaptive) composition of $\mathcal M_1,\mathcal M_2,\dots,\mathcal M_k$ and let $\delta_{\mathcal M}(\epsilon)$ be its privacy curve.
Set $L\ge 2+\epsilon_\mathrm{error}$ sufficiently large such that
\begin{equation}
\label{eqn:L_delta}
\sum_{i=1}^k \delta_{\mathcal M_i}(L-2) \le \frac{\delta_\mathrm{error}}{8} \text{ and } \delta_{\mathcal M}(L-2-\epsilon_\mathrm{error}) \le \frac{\delta_\mathrm{error}}{4}.
\end{equation}
Let $\widetilde{Y}$ be the approximation of $Y=\sum_{i=1}^k Y_i$ produced by \textsf{ComposePRV} algorithm with mesh size $$h=\frac{\epsilon_\mathrm{error}}{\sqrt{\frac{k}{2}\log \frac{12}{\delta_\mathrm{error}}}}.$$ Then
\begin{equation}
\label{eqn:upperlower_delta}
\delta_{\widetilde{Y}}(\epsilon+\epsilon_{\mathrm{error}}) -\delta_{\mathrm{error}}\le \delta_Y(\epsilon) = \delta_{\mathcal M}(\epsilon) \le \delta_{\widetilde{Y}}(\epsilon-\epsilon_{\mathrm{error}})+\delta_\mathrm{error}.
\end{equation}
Furthermore, our algorithm takes $O\left( b \frac{L}{h} \log\left(\frac{L}{h}\right)\rp$ time where $b$ is the number of distinct algorithms among $\mathcal M_1,\mathcal M_2,\dots,\mathcal M_k$.
\end{restatable}
\begin{remark}
A simple way to set $L$ such that the condition (\ref{eqn:L_delta}) holds is by choosing an $L$ such that:
\begin{equation}
\label{eqn:L_eps}
L\ge 2+\max\left\{\epsilon_\mathrm{error}+\epsilon_{\mathcal M}\left( \frac{\delta_\mathrm{error}}{4}\right), \max_{i\in [k]}\ \epsilon_{\mathcal M_i}\left( \frac{\delta_\mathrm{error}}{8k}\right)\right\}
\end{equation}
where $\epsilon_{\mathcal A}(\delta)$ is the inverse of $\delta_\mathcal A(\epsilon)$. To set the value of $L$, we do not need the exact value of $\epsilon_{\mathcal M}$ (or $\epsilon_{\mathcal M_i}$). We only need an upper bound on $\epsilon_{\mathcal M}$, which can often by obtained by using the RDP Accountant or any other method to derive upper bounds on privacy.
\end{remark}
|
\section{Experiment Details}
\subsection{VQ-VAE Architecture}
For our image compression experiments, we used a convolutional VQ-VAE architecture similar to the one used in \citep{ramesh2021zero} with residual connections.
Both the input and the side information to this network have the shape $128 \times 256$ (i.e. vertical halves of a full $256 \times 256$ image).
The encoder scales down input image by a factor of $4\times$ or $8\times$ both vertically and horizontally, producing latent variables of shape $32 \times 64$ and $16 \times 32$, respectively. Each dimension of the latent variable is allotted different numbers of bits (i.e.~codebook bits), which range from $1$ to $8$ in our experiments. The decoder conversely takes a discrete latent variable and upscales it by a factor of $4\times$ or $8\times$ to produce a reconstruction of the original shape.
A detailed specification of the architecture is provided in \Cref{fig:vqvae_arch_conv}. The notation (\texttt{Tconv}) ``\texttt{Conv} $A \times B$ ($C \to D$)'' represents a 2D (transposed) convolution with kernel size $A$ and stride $B$ with input and output channels $C$ and $D$, respectively. The boxes ``\texttt{Residual Block} ($A\to B$)'' represent a two-layer residual network. We used GELU (Gaussian Error Linear Unit) activation for all layers except for the very last convolution of the decoder, for which we used sigmoid. The exact number of channels may differ for different rate-distortion points, and we refer the reader to the supplementary code submission for full details.
\input{figures/fig_vqvae_arch_conv}
We used a small network (denoted ``\texttt{SI Net}'') shared by the encoder and decoder to preprocess the side information. Whenever the encoder or decoder does not receive side information (for distributed and separate VQ-VAEs), we simply replace the output of \texttt{SI Net} with a zero tensor of the same shape.
We note that all three VQ-VAE variants have the same architecture and number of parameters for the autoencoder portion (the portion that the horizontal line goes through in \Cref{fig:vqvae_arch_conv}). Thus there is no architectural advantage among the VQ-VAE variants.
\subsection{Specifying Target Rate}
The shape and the number of codebook bits determine the total rate.
This is different from neural compression models that use a regular VAE, which control rate-distortion by reweighting the training objective (ELBO). As mentioned in \cref{sec:our_method}, the fact that we have a hard constraint on the rate is useful when working with a hard communication limit.
There are several ways to achieve a desired target rate. For example, an $8\times$ downscaling encoder with 4 codebook bits produces compressed message of size $16 \times 32 \times 4 = 2048$ bits. The same rate can be achieved using a $4\times$ downscaling encoder with 1 codebook bit: $32 \times 64 \times 1 = 2048$ bits. This flexibility is what allows us to have a fairly granular control over the target rate, at the cost of hyperparameter choices.
\textbf{Hyperparameters for KITTI Stereo Experiment. }
Due to the above choices, we use differently sized networks for each rate-distortion curve for our stereo image compression experiments. While the model definition is included in the code, we include the relevant hyperparameter information in \cref{table:kitti_hparams} below.
\input{figures/table_kitti_hparams}
\textbf{Hyperparameters for CelebA-HQ Experiment. }
For this experiment, we used the single $8\times$ downscaling encoder with codebook bits $\{1, 2, 3, 4\}$, resulting in the total rates of $\{512, 1024, 1536, 2048\}$.
\textbf{Distributed optimization. }
For gradient compression VQ-VAEs, we followed the same architecture as the image compression experiments, but replaced all convolutional layers with fully-connected layers. Full specification of the network is included in the code submission.
\subsection{Stereo Image Dataset}
For our stereo image compression experiment, we follow the setup of \citep{NDIC}. First, we construct training/test datasets using the files specified in the official repository for \citep{ayzikA2020dsin} (\url{https://github.com/ayziksha/DSIN/tree/master/src/data_paths}).
Then we apply the same preprocessing steps of \citep{NDIC}, where we first take the center crop of size $370 \times 740$, then resizing it to $128 \times 256$ using PyTorch transformations. This results in the training split containing 1576 pairs of stereo images and test split containing 790 pairs.
\subsection{Training Details}
\label{sec:dis_training}
For stereo image compression experiments, we trained our models for up to 1000 epochs on a DGX machine. Some training runs were early stopped because the validation performance started to plateau.
For CelebA-HQ experiments, we trained the VQ-VAEs for a total of 20 epochs distributed over two Nvidia GTX 2080 GPUs.
For gradient compression experiments, we trained the fully-connected VQ-VAEs for 500 epochs on a single GPU.
In all cases, we evaluated validation loss after each epoch and observed no overfitting. This leads us to believe that it may be possible to further improve the performance of our method by training a larger network, which we leave for future work.
\subsection{NDIC Training Details}
For the CelebA-HQ compression experiment, we trained NDIC \cite{NDIC} using the official code released by the authors. We used the ``Balle18'' \cite{balle2018variational} backbone and trained the model for 10 epochs ($\approx$ 300K examples). For evaluation, we used the model checkpoint with the best validation set performance. While \citet{NDIC} report training with a batch size of 1 for 500K steps, we chose to use a batch size of 20. This was done for several reasons. First, using a batch size of 1 is very inefficient as we do not benefit from parallelism of GPU. It also leads to high variance in the gradient, often leading to slower convergence. In our training, the loss plateaued well before reaching 300K total examples. We also used MSE instead of MS-SSIM as the objective, since our models were also trained using MSE loss. This ensures a fair comparison in both training as well as evaluation, as we report PSNR (which is equivalent to MSE). The rest of the model hyperparameters were not modified.
\section{Introduction}
Data compression plays an essential role in modern computer systems. From multimedia codecs running on consumer devices to cloud backups in large data centers, compression is a necessary component in any system that deals with high-volume or high-velocity sources.
Applications such as multi-camera surveillance systems, IoT sensing, 3D scene capture and stereo imaging create distributed data streams with very large volume that are highly correlated.
We are interested in distributed compression -- specifically the \textit{distributed source coding} problem.
In this setting, there are two correlated sources (input $\orig$ and side information $\sinfo$) that
are physically separated. Both must be compressed and sent to a common decoder.
We can assume that the side information $\sinfo$ is compressed in isolation and communicated to the decoder and now one can expect that the original source $\orig$ can be compressed at a higher-rate since $\sinfo$ is known to the decoder.
The core challenge is how to compress the original source $\orig$ when the correlated side information $\sinfo$ is available only at the decoder as shown in \cref{fig:dist-vs-joint} (left).
If side information is available to both the encoder and decoder as in \cref{fig:dist-vs-joint} (right), it is well known that the side information can be utilized to improve the compression rate of $\orig$.
Surprisingly, in 1973, Slepian and Wolf~\citep{Slepian-Wolf1973} showed that an encoder that has no access to the correlated side information can asymptotically achieve the same compression rate as when side information is available at both the encoder and the decoder.
In other words, \textit{distributed} compression is asymptotically as efficient as \textit{joint} compression. Later, Wyner and Ziv~\citep{Wyner-Ziv1976} extended this result to lossy compression. This is nothing short of a remarkable classical information theory result that defies intuition.
\input{figures/fig_joint_vs_dist}
\subsection{Motivating Example}
\label{sec:motivating_example}
To provide some intuition behind distributed source coding, we describe a simple example that illustrates how side information known \emph{only} to the decoder can be as useful as side information known to \emph{both} the encoder and decoder~\cite{DISCUS1999}.
Let $\orig$ and $\sinfo$ be the two correlated sources, here uniformly random 3-bit sources, which differ by at most one bit (i.e., Hamming distance between $\orig$ and $\sinfo$ is at most one). Clearly, to losslessly compress $\orig$ without any side information, the encoder has to send 3 bits to the decoder. If $\sinfo$ is known to both the encoder and decoder, then $\orig$ can be transmitted using only 2 bits instead, since the encoder can send the difference between $\orig$ and $\sinfo$, which is uniform in $\{000,100,010,001\}$. Therefore \textit{joint} compression using only $2$ bits is possible.
Now, if the side information $\sinfo$ is available only at the decoder, Slepian-Wolf theorem suggests that the encoder can still transmit $\orig$ using only $2$ bits.
How could this be possible? The key idea is to group $8$ possible values of $\orig$ into $4$ bins, each containing two bit-strings with maximal Hamming distance: $\mathcal{B}_0 = \{000,111\}$, $\mathcal{B}_1 = \{001,110\}$, $\mathcal{B}_2 = \{010,101\}$, $\mathcal{B}_3 = \{011,100\}$. Then the encoder simply transmits the bin index $\msg \in \set{0,1,2,3}$ for the bin containing $\orig$. The decoder can produce the reconstruction $\reconst$ based on the bin index $\msg$ and $\sinfo$; precisely,
$\hat{\orig} = \arg\max_{\orig \in \mathcal{B}_{\msg}} P(\orig|\sinfo)$. In this case, since the Hamming distance between $\orig$ and $\sinfo$ is at most one and the Hamming distance between the bit strings in each bin is $3$, the decoder can recover $\orig$ without error. The side information allows the decoder to correctly choose between the two candidates in the bin specified by the encoder.
\subsection{Practical Difficulties}
There are major challenges in designing practical distributed compression schemes for arbitrarily correlated sources.
First, the joint distribution $p(\orig,\sinfo)$ is required to design the encoder and decoder, but modeling high-dimensional joint distributions (e.g. for correlated images) is very challenging, especially without modern deep generative models.
Second, designing a distributed compression scheme with reasonable run time beyond simple structures remains an open question.
The significant gap between the theory of Slepian-Wolf and Wyner-Ziv and what is efficiently achieved in practice has been well-acknowledged in the information theory community. Two decades after the Slepian-Wolf theorem,
~\citet{Verdu1998} writes: ``despite the existence of potential applications, the conceptual importance of Slepian–Wolf coding has not been mirrored in practical data compression.'' Constructive compression schemes have been designed only for very special cases, such as correlated Gaussian sources~\cite{DISCUS1999} and stereo images (as elaborated in Section~\ref{sec:related}).
In this paper, we bridge this gap by leveraging the recent advances in deep generative modeling and show that it is possible to train an encoder and decoder for distributed compression of \textit{arbitrarily} correlated high-dimensional sources.
Specifically, our approach (denoted \textit{Neural DSC}) parametrizes the encoder-decoder pair as a Vector-Quantized VAE (VQ-VAE) \citep{van2017neural}, so that:
(a) we do not need a hand-crafted analytical model for the joint distribution; and (b) we can \textit{learn} the complicated encoder and decoder mappings via neural networks and execute them efficiently at inference time; and (c) the discrete latent representation of VQ-VAE provides a \textit{data-driven} quantization scheme that avoids the need for manual quantization and entropy coding to obtain the compressed message.
Our main contributions are as follows:
\begin{itemize}\setlength\itemsep{0em}
\item We introduce Neural DSC, a compression scheme for lossy distributed source coding based on VQ-VAE. Our encoder without side information creates the compressed message, which is then transmitted to the decoder that reconstructs the original signal by leveraging the side information.
\item We show that Neural DSC performs favorably to existing techniques on stereo camera image compression. Specifically, it outperforms existing methods for low rates and remains competitive at high rates.
\item We further validate that our approach can adapt to complex correlations. For this, we compress the top half of an image given the bottom part as side information. On this task, our method significantly outperforms the current state-of-the-art method and nearly matches the joint compression rate.
\item We provide a promising evidence that Neural DSC is effective in other modalities by applying it to gradient compression for distributed model training.
\item We compare our model to the theoretical optimum on synthetic data and provide evidence that our model indeed learns to perform distributed source coding. We also show that learned quantization scheme is critical to the model's performance.
\end{itemize}
\section{Neural Distributed Source Coding} \label{sec:ndsc}
\subsection{Vector-Quantized VAE}
\label{sec:vqvae}
Vector-Quantized VAE (VQ-VAE) \citep{van2017neural} is a specific type of VAE \citep{kingma2014auto,rezende2014stochastic} that has a \textit{discrete} latent variable, even though the input is continuous. Because the latent code is discrete and has fixed size, VQ-VAEs are a natural fit for lossy compression. Indeed, many existing works have explored its use in various compression tasks, ranging from music generation to high-resolution image synthesis \citep{garbacea2019low,razavi2019generating,roberts2018hierarchical}. Most recently, a variant of VQ-VAE was used to generate discrete embedding for natural images for zero-shot text-to-image generation \citep{ramesh2021zero}.
A VQ-VAE consists of three components: an encoder, a decoder, and a codebook. The main difference between VQ-VAE and a regular VAE is that the output of the encoder is quantized to the nearest vector in the codebook. During training, all three components are jointly optimized. The fact that VQ-VAE learns the quatization scheme from data is important, as we further explain in \cref{sec:discussion}. During inference, the input is represented by the index of the codebook vector that the encoder's output is quantized to.
\subsection{Our Method}
\label{sec:our_method}
\textbf{Notation and Setup. } We let $\orig$, $\sinfo$, and $\msg$ to denote the input signal, correlated side information, and compressed message, respectively. The encoder $\fencode$ generates the compressed message: $\msg = \fencode(\orig)$; the decoder $\fdecode$ creates the reconstruction $\reconst$ with the aid of the side information: $\reconst = \fdecode(\msg; \sinfo)$.
Note that the encoder only receives $\orig$, while the decoder receives both the compressed message $\msg$ and the side information $\sinfo$.
We refer to the number of bits required to transmit $\msg$ as the \textit{rate}, and the reconstruction performance (e.g. measured by $\ell_2$ error) as the \textit{distortion}. In general, higher rate leads to lower distortion, and vice versa.
We propose to learn the distributed encoder $\fencode$ and decoder $\fdecode$ for $(\orig, \sinfo)$ by parametrizing $\fencode$ and $\fdecode$ as a conditional VQ-VAE, where $\fdecode$ is conditioned on the side information $\sinfo$.
The network is trained on i.i.d. pairs $\set{(\orig^{(i)},\sinfo^{(i)})} \sim p(\orig, \sinfo)$ to minimize the $\ell_2$ distortion between the input $\orig$ and the reconstruction $\reconst$. This scheme allows us to directly train the encoder-decoder pair in an end-to-end manner without needing to model or manually specify the correlation structure between $\orig$ and $\sinfo$.
\textbf{Why VQ-VAE? }
Although much of the existing literature has focused on VAEs as the backbone for neural compression \citep{balle2017endtoend,theis2017lossy,balle2018variational,minnen2018joint,balle2021nonlinear}, we intentionally chose VQ-VAE instead for a few reasons.
First, VQ-VAE offers an explicit control over the rate through the latent dimension and codebook size -- unlike VAEs for which we can only estimate the rate after training the model. VQ-VAE also guarantees an upper bound on the resulting rate regardless of the input, which is often desired in practice (e.g. communication channel with strict bandwidth limit). The downside of using a VQ-VAE is that training is less stable than a VAE, but this did not pose much issue during our experiments.
\subsection{Related Work}\label{sec:related}
Unlike the vast literature that exists on distributed source coding (DSC), the use of deep learning on DSC has received relatively little attention. To the best of our knowledge, we are aware of two relevant studies, which we describe below.
\textbf{Distributed Stereo Image Compression}~\cite{ayzikA2020dsin} proposes a method to perform DSC for stereo camera image pairs with a high spatial correlation. Due to the large spatial similarity between the images, one of them can serve as the side information for the other.
The key component of their method is the ``Side Information Finder'', a module that finds similar image patches between the side information and the reconstructed signal produced by a pre-trained autoencoder. Since the two images have many approximately overlapping patches, the reconstruction is further improved by copying over matching image patches from the side information to the reconstruction.
While this leads to a considerable improvement in compression performance, this method is only applicable when the input and side information have large spatial overlap. On the other hand, our method is applicable to any correlated sources. We empirically validate this using data sources with substantially more complex correlation (see \Cref{fig:sample_stereo_vs_face}).
\textbf{Neural Distributed Image Compression}~\cite{NDIC} is the most recent and current state-of-the-art method that leverages a (regular) VAE for DSC of image data. The proposed architecture is designed to explicitly model the common information between the input and side information. Intuitively, the goal is to guide the encoder to compress only the portion that is not recoverable from the side information. Once trained, the encoder simply discards the common information and only transmits the residual, with the hope of decoder being able to reconstruct the common information from the side information.
As \citep{NDIC} is closely related to our work and is the current state-of-the-art method that improves upon \citep{ayzikA2020dsin}, we compare our method with both of these methods in our experiments.
\section{Experiments}
\subsection{Stereo Image Compression}
\textbf{Setup. }
We first evaluate our method on stereo image compression. Following \citep{ayzikA2020dsin}, we construct a dataset consisting of pairs of images obtained from the KITTI Stereo 2012 and 2015 \citep{geiger2012we,menze2015joint}. Each pair of images are taken by two cameras at a slightly different angle and share spatial similarity (see \cref{fig:sample_stereo_vs_face}, left). The goal is to compress one of the images in each pair, treating the other image as side information that only the decoder has access to. The performance of a compressor is evaluated by its rate-distortion points, where rate and distortion are measured in terms of bits/pixel (bpp) and Peak Signal-to-Noise Ratio (PSNR).
To the best of our knowledge, this is the only benchmark that currently exists for distributed image compression.
For each target rate, we train a conditional VQ-VAE as described in \cref{sec:our_method}. We then evaluate its distortion averaged over the test set, and plot the rate-distortion point. The exact details of how the dataset is constructed as well as the model hyperparameters are included in the Appendix.
\paragraph{Results.}
We compare our methods to existing distributed image compression methods DSIN \citep{ayzikA2020dsin} and NDIC \citep{NDIC}.
As shown in \cref{fig:kitti_bpp_psnr}, our method achieves new state-of-the-art distortion for rates below 0.25 bits/pixel. For higher rate, our method is slightly worse compared to NDIC.
\input{figures/fig_kitty_plot_and_samples}
In addition, our model is significantly more parameter-efficient compared to \citep{NDIC}. \Cref{table:param_efficiency} shows the parameter count for the models used to generate the points in \cref{fig:kitti_bpp_psnr}. As shown, our models are smaller by a factor of up to $\sim 6 \times$.
\input{figures/table_param_efficiency}
\subsection{Handling Complex Correlation}
\label{sec:complex_corr}
\textbf{Setup. }
To further investigate how well our method can handle complex correlation between the input and side information, we evaluate our method on a challenging distributed compression setup. First we create a dataset of correlated images from $256 \times 256$ CelebA-HQ dataset \citep{liu2015deep} containing images of celebrity faces. Each image is vertically split into top and bottom halves, where the top half is used as the input and the bottom half is used as side information. Following \citep{kingma2018glow}, we use $27000/3000$ split between training and test data.
While there is clearly some correlation between the top and bottom halves of an image of a human face, modeling this correlation (e.g. the conditional distribution over the top half of a face given the bottom half) is highly nontrivial. This experiment is thus designed to show our model's ability to leverage this correlation to improve compression performance.
\paragraph{Baseline VQ-VAEs.}
To analyze the gains from distributed compression, we train three different variants of VQ-VAEs: distributed (our method), joint, and separate. In the \textit{joint} model, both the encoder and decoder have access to the side information. This serves as a proxy to the intractable theoretical rate-distortion bound established by \citet{Wyner-Ziv1976} for lossy distributed compression. We expect this to be the upper limit on the performance of our method. The \textit{separate} model is positioned at the other extreme, where neither the encoder nor the decoder uses the side information. This serves as the lower limit on the performance of our method.
To ensure a fair comparison among these variants, they have identical architecture and number of parameters for the autoencoder portion, and only vary in the way they handle the side information. Architectural details and network hyperparameters are provided in the Appendix.
\paragraph{Results.}
In \Cref{fig:celeba_rate_distortion}, we present rate-distortion curves for our model as well as NDIC evaluated over the test set. We adopt the commonly used distortion metric Peak Signal-to-Noise Ratio (PSNR). %
\Cref{fig:celeba_ours_vs_ndic} shows that our method achieves a significantly better rate-distortion performance compared to NDIC \cite{NDIC} in the presence of complex correlation between the image pair. We posit that this is due to NDIC attempting to explicitly model the correlation even though it is not used during inference. In \Cref{fig:celeba_rate_psnr}, we see that the distributed VQ-VAE achieves nearly identical performance as the joint VQ-VAE, further proving the viability of our method as a distributed coding scheme.
\input{figures/fig_celeba_rate_distortion}
\paragraph{Effect of Side Information.}
It is possible that the distributed VQ-VAE learns to ignore the side information, effectively collapsing to a separate VQ-VAE. We investigate whether the distributed decoder actually makes use of the side information by intentionally providing incorrect input.
We can see in \Cref{fig:effect_sideinfo} that the side information plays a significant role in the quality of reconstruction. For example, providing a side information with a different face leads to the reconstruction having a matching skin tone that is different from the original. As expected, side information has no effect for the separate encoder.
\input{figures/fig_effect_sideinfo}
\subsection{Communication Constrained Distributed Optimization} \label{sec:gradients}
\paragraph{Background.}
Here we consider an interesting application of our method to distributed training of a neural network $\net$ with parameters $\params \in {\mathbb{R}}^d$.
Mini-batch SGD and other first order methods such as Adam~\citep{kingma2014adam} have become the de-facto standard for optimizing such functions~\citep{robbins1951stochastic, bottou2010large, tsitsiklis1986distributed}.
With the increasing size of deep learning models, it iis necessary to train a model across multiple workers\footnote{\footnotesize We use the terms node, client and worker interchangeably.}.
In particular, each worker $j$ is assigned a subset ${\mathcal{X}}_j \in {\mathcal{X}}$ of the training data. The workers then compute gradients
locally using SGD-like iterations and communicate the gradients back to a central parameter server. The server then aggregates the received gradients to perform a global update and broadcasts the updated parameters back to the workers.
An important bottleneck in this setting is the repeated communication of gradients between workers and the server,
often resulting in significant communication overhead, especially for large models.
To alleviate this issue, many gradient compression techniques have been developed.
Here we present a promising proof of concept that demonstrates that our distributed encoder is able to better compress gradients compared to several established techniques.
\paragraph{Experimental Setup.}
We focus on a simple setting with two workers interacting with a central server. The key observation is that, at each iteration $t$, the correlation between the client gradients ${\mathbf{g}}_t^1$ and ${\mathbf{g}}_t^2$ can be exploited to further improve compression performance by viewing this as a distributed compression task. Concretely, we treat ${\mathbf{g}}_t^1$ as side information and train a distributed encoder for ${\mathbf{g}}_t^2$, and show that the cost of communicating ${\mathbf{g}}_t^2$ can be substantially reduced.
The neural network $\net$ being trained using compressed gradients is a small convolutional network for MNIST \citep{lecun2010mnist} digit classification.
We partition the MNIST training data ${\mathcal{X}}$ into two equal subsets: ${\mathcal{X}}_{pre}$ and ${\mathcal{X}}_{train}$. ${\mathcal{X}}_{pre}$ is used to train the VQ-VAE encoder and ${\mathcal{X}}_{train}$ is used to train $\net$ using the trained encoder as gradient compressor.
We measure the performance of our method using two metrics: (a) \textbf{rate-distortion}
of the distributed compressor, and (b) \textbf{classification accuracy} of the model $\net$ trained using compressed gradients.
\textbf{Generating training data for encoder. }
We train $\net$ for $T$ steps using the Adam optimizer
across two nodes over ${\mathcal{X}}_{pre}$. This generates a sequence of $T$ gradient pairs $\set{{\mathbf{g}}_t^1, {\mathbf{g}}_t^2}_{t=1}^T$, which can be used to train the VQ-VAE encoders.
However, naively applying our method to this setting leads to a suboptimal VQ-VAE as these gradient pairs are highly correlated. Noticing that ${\mathbf{g}}_t^1$ and ${\mathbf{g}}_t^2$ are conditionally independent given the initial model weights and the time step $t$, we train $\net$ for multiple runs with different initialization for $\params$, while randomly sampling a subset of gradients from each run. Thus we generate a dataset of tuples $(t, {\mathbf{g}}_t^1, {\mathbf{g}}_t^2)$ sampled from multiple independent runs. We also update the architecture of the VQ-VAE so that both the encoder and decoder are conditioned on $t$. This way, the gradients become i.i.d. samples over random runs and time steps.
We consider the following baseline methods for our comparison:
\setlist{leftmargin=3mm}
\begin{itemize}
\item \textbf{Random-$k$}~\cite{wang2018atomo, koloskova2019decentralizeda}: Instead of communicating the full gradient vector $\mathbf{g} \in \mathbb{R}^d$, we only transmit randomly chosen $k$ coordinates, thus reducing the communication cost by a factor of $\frac{k}{d}$.
\item \textbf{Top-$k$}~\cite{shi2019understanding,stich2018sparsified}: Similar to Random-$k$, this approach also communicates $k$ of the $d$ coordinates of a gradient vector and improves the communication cost by a factor of $\frac{k}{d}$. However, this time we take the top $k$ coordinates with largest magnitude and discard the remaining $(d-k)$ coordinates.
\item \textbf{QSGD}~\cite{alistarh2017qsgd, basu2019qsparse}: Gradients normally stored using 32-bit floating point numbers are quantized using fewer bits before being transmitted.
\item \textbf{Coordinated Sampling}~\cite{shi2019distributed}:
The nodes \textit{cyclically} select a batch of $k$ coordinates at every iteration and communicate only those $k$ coordinates.
\end{itemize}
\paragraph{Pre-training VQ-VAE encoders.}
We first compare the performance of three different VQ-VAE variants over the course of a single training run of $\net$. As shown in \Cref{fig:grad_vqvae_mse}, we observe a substantial improvement in $\ell_2$ distortion for the distributed VQ-VAE, compared to the separate model. As training progresses, the gap between the joint and distributed encoders narrows, suggesting that a sufficiently large distributed VQ-VAE encoder can nearly match the performance of its joint counterpart.
\input{figures/fig_all_grad_plots}
\paragraph{Distributed Training.}
Given the pre-trained encoders, we train $\net$ using gradients compressed by the pre-trained encoder.
We compare it against the baseline gradient compression schemes listed above: Top-$k$, Rand-$k$, \textsc{QSGD} and Coordinated Sampling.%
All runs were repeated 20 times with different random seeds, and we report average performance along with standard error.
\Cref{fig:grad_performance} shows that $\net$ trained with gradients compressed using the distributed VQ-VAE leads to substantial accuracy improvement over the separate VQ-VAE without the extra communication cost. Moreover, we see that VQ-VAE based compression leads to about $3\times$ faster convergence than the baselines (see \cref{fig:grad_comparison_test_acc}).%
This shows that a learned distributed encoder can indeed benefit from the complex correlation between gradients distributed optimization.
\section{Discussion and Analysis}
\label{sec:discussion}
\paragraph{Role of Side Information.}
In \Cref{sec:motivating_example} we saw that grouping symbols with maximal distance from each other in the same group allowes the decoder to determine the correct symbol using the side information. Here we investigate to what extent our models perform such a grouping (also known as \textit{binning}). If approximate binning was occurring, the same codeword (i.e. bin index) would be decoded into different images for different side information. In other words, a distributed compressor should have high reconstruction diversity for a single codeword as we vary the side information. We refer to this as \textit{bin diversity}.
On the other hand, a model that does not perform binning should decode a single codeword to similar symbols regardless of what side information is given to the decoder. To test this hypothesis, we train two different models:
\textbf{Ours} -- trained with the correct side information as was done for other experiments;
\textbf{Uncorrelated SI} -- trained by replacing the side information with random ones from other, irrelevant samples in the dataset, thus making the input and side information completely independent. We expect this model to ignore side information and achieve lower bin diversity.
To quantitatively measure diversity, we used average pairwise distance with respect to $\ell_2$ norm and LPIPS distance \cite{zhang2018unreasonable}, which has been used in the literature \citep{srivastava2017veegan,metz2016unrolled} as a measure of sample diversity. Both metrics were computed and averaged over all images in the CelebA-HQ test dataset.
As shown in \cref{table:diversity}, the VQ-VAE trained with correlated side information exhibits much higher bin diversity in both metrics. In other words, the images within each bin are much farther from each other compared to the other model, suggesting that some form of approximate binning is happening. The uncorrelated VQ-VAE has particularly low bin diversity with respect to the LPIPS distance, meaning there is very little perceptual difference regardless of what side information is used.
\input{figures/table_diversity}
\paragraph{Optimality of the Learned Compressor.}
Another natural question to ask is how close our \textit{learned} distributed compressor is to the optimal distributed compressor. We investigate this using a synthetic data source studied in \cite{DISCUS1999} with two correlated Gaussian sources: $Y = X + N$, $X \sim \mathcal{N}(0, 1)$, and $N \sim \mathcal{N}(0, 0.1^2)$.
The asymptotically optimal (i.e.~in the limit of compressing infinitely many symbols together) rate-distortion curve of these sources is known analytically, so we can check how our model compares to the theoretical limit.
\Cref{fig:gaussian_rate_dist} shows that for low rates, both learned methods significantly outperform the asymptotically optimal encoder without side information (SI). This is a concrete evidence that our learned encoder is actually performing DSC (as opposed to simply achieving a very good compression rate for single source coding) Moreover, the distributed compression performance remains very close to that of joint compression, showing the efficacy of our practical distributed coding scheme.
At higher rates, however, the value of side information quickly diminishes, and the learned methods perform worse than even the encoder without SI. This is expected as the optimal curves presented in this figure are asymptotic in the limit of jointly compressing infinitely many symbols, which is clearly not the case for the learned methods that compress each symbol one at a time.
\input{figures/fig_optimality_and_quantization}
\paragraph{Importance of Learned Quantization.}
As mentioned in \cref{sec:vqvae}, we test the importance of \textit{learned} quantization of VQ-VAE, which likely contributes to the performance gains compared to NDIC given that NDIC uses uniform quantization. We do so by plotting the rate-distortion curve of an autoencoder with identical network architecture as our VQ-VAE models, but with uniform quantization.
Once trained, the continuous latent variables are uniformly quantized to different levels, thus producing multiple rate-distortion points as shown in \Cref{fig:learned_vs_static_quantization}. We see that VQ-VAE enjoys a significantly improved performance due to the data-driven quantization scheme. Even when we increase the latent dimension ($D=1$ vs $D=4$) for the autoencoder with uniform quantization, the resulting distortion is still worse than VQ-VAE, even at a significant cost of compression rate.
That said, a more sophisticated quantization scheme tailored to the downstream task may lead to substantial improvements. Studying better quantization schemes or even quantization-less compression techniques is an active area of research \cite{havasi2018minimal,ho2019flow++}, which we leave for future work.
\section{Conclusions}
We presented Neural DSC, a framework for distributed compression of correlated sources. Our method is built on the power of modern deep generative models, namely VQ-VAEs, which are excellent data-driven models to represent high-dimensional distributions using compressed discrete latent codes. For stereo image compression, our method performs better or competitively compared to existing techniques. We also show that our method is able to leverage complex correlations far beyond spatial similarity for better compression performance, approaching the joint compression rate. Finally, we show that our method is not limited to images and show a promising result for compressing gradients for distributed model training. We believe that our work provides encouraging initial results for practical data-driven distributed compression schemes and hope to see further developments.
\section*{Acknowledgments}
This research has been supported by NSF Grants CCF 1934932, AF 1901292,
2008710, 2019844 the NSF IFML 2019844 award as well as research gifts by Western Digital, WNCG and MLL, computing resources from TACC and the Archie Straiton Fellowship.
|
\section{Introduction}
\begin{figure}[t]
\centering
\subfigure{\includegraphics[width=1.0\linewidth]{figures/figure_1.pdf}}
\vspace{-2mm}
\caption{
Illustration of our proposed semi-supervised domain adaptive semantic segmentation (SSDAS): With a few labeled target samples as anchors, SSDAS achieves semi-supervised domain adaptation via 4 key processes including \textbf{A}: re-weight source features according to their similarity to the feature of the few labeled target samples, \ie, cross-domain similarity (CDS),
\textbf{B}: update feature centers with newly weighted source features, \textbf{C}: align target features to the updated feature centers, and \textbf{D}: replace dissimilar source features by high-confidence target features. The four processes run iteratively which keep assigning higher weights to target-alike source features and replacing target-unlike source features with confident target features adaptively and progressively (Best viewed in color).
}
\label{fig:intro}
\end{figure}
Domain adaptation has been investigated extensively for the task of semantic segmentation, largely for minimizing the gap between a labeled source domain (typically consisting of synthetic images with automatically generated labels) and an unlabeled target domain and accordingly alleviating the pain point of pixel-level annotation of large amounts of training images. Most existing domain adaptive semantic segmentation works are unsupervised \cite{hoffman2016fcns,kang2019contrastive,kang2018deep,sankaranarayanan2018learning,tsai2018learning,tzeng2017adversarial,luo2019taking,vu2019advent,Chen_2018_CVPR,chen2018domain} which assume a completely unannotated target domain. However, annotating a few target samples is often very manageable and worthwhile especially if it can substantially improve the segmentation performance in the target domain. By managing the amount of labeling in the target domain flexibly, such semi-supervised domain adaptation (SSDA) is also more scalable while facing very diverse domain gaps where unsupervised domain adaptation (UDA) often fails to handle well (e.g. when the domain gap is very large).
SSDA based semantic segmentation has been largely neglected despite its great values in various practical tasks. One intuitive approach to this new problem is to adapt existing UDA methods by including extra supervision from the few labeled target samples in the training process. However, existing UDA methods were designed without considering few-shot labelling in the semi-supervised setup, and recent studies \cite{saito2019semi,kim2020attract} show that this intuitive approach does not perform clearly better (sometimes even worse) than direct training over all labeled source and few-shot target samples in a supervised manner. Considering the similarity between semantic segmentation and image classification, another approach is to adapt the recent SSDA based image classification methods \cite{saito2019semi,kim2020attract} for the semantic segmentation problem. However, the SSDA based image classification methods do not perform well for the dense prediction of image pixels which often experiences much larger and more diverse variations across domains.
We propose SSDAS, a Semi-Supervised Domain Adaptive Segmentation network that positions a few labeled target samples as anchors for effective feature alignment between labeled source samples and unlabeled target samples. We introduce two novel designs for optimal feature alignment across domains. The first design is \textit{adaptive cross-domain alignment} that takes the few labeled target samples as references to gauge the cross-domain similarity (CDS) of each source feature and learn more from target-alike source features. Specifically, we raise (or lower) the weight of source features that are similar (or dissimilar) to the target features adaptively during the domain adaptation process as illustrated in Fig.~\ref{fig:intro}. The second design is \textit{progressive intra-domain alignment} where target-unlike source features are replaced by high-confidence target features progressively and the alignment becomes in-between confident and unconfident target features.
The proposed SSDAS adopts Jigsaw Puzzle classifiers to extract contextual features for domain adaptive semantic segmentation and it can also work with conventional features, more details to be discussed in experiments.
The contributions of this work can be summarized in three major aspects. \textit{First}, we design SSDAS – an innovative semi-supervised domain adaptive segmentation network that exploits a few labeled target samples for better domain adaptive semantic segmentation. To the best of our knowledge, this is the first work that explores semi-supervised domain adaptation (with few-shot labels) for the classical semantic segmentation task. \textit{Second}, we design two novel feature alignment strategies for semi-supervised domain adaptation problems. The strategies exploit a few labeled target samples and achieve adaptive cross-domain alignment and progressive intra-domain alignment effectively. \textit{Third}, extensive experiments over multiple domain adaptive segmentation tasks show that our proposed SSDAS achieves superior semantic segmentation consistently.
\section{Related Works}
\textbf{Unsupervised Domain Adaptation (UDA)} has been studied extensively for the task of semantic segmentation. Existing UDA-based segmentation methods can be broadly classified into three categories. The first category is \textit{adversarial training} based which utilizes a domain classifier to align source and target distributions in the feature, output or latent space \cite{hoffman2016fcns,long2016unsupervised,tzeng2017adversarial,luo2019taking,tsai2018learning,chen2018road,zhang2017curriculum,saito2017adversarial,saito2018maximum,vu2019advent,tsai2019domain,lee2019sliced,guan2021uncertainty,zhang2021detr,huang2021mlan}. The second category is \textit{image translation} based which translates images from source to target domains to mitigate domain gaps \cite{hoffman2018cycada, sankaranarayanan2018learning,chen2019crdoco, li2019bidirectional,Zhang2019lipreading,hong2018conditional,yang2020fda,huang2021fsdr}. The third category is \textit{self-training} based which utilizes ``pseudo labels" to guide iterative learning over unlabeled target data \cite{zou2018unsupervised,saleh2018effective,zhong2019invariance,zou2019confidence,guan2021scale,huang2021cross}.
\textbf{Semi-supervised Domain Adaptation (SSDA)} assumes the availability of a few labeled target samples beyond labeled source samples and a large amount of unlabeled target samples as in UDA. Several SSDA methods \cite{ao2017fast,donahue2013semi,yao2015semi,saito2019semi,kim2020attract,jiangbidirectional} have been proposed which addresses domain discrepancy by auxiliary constrain optimization~\cite{donahue2013semi}, subspace learning~\cite{yao2015semi}, label smoothing~\cite{ao2017fast}, entropy mini-max~\cite{saito2019semi}, intra-domain discrepancy minimization~\cite{kim2020attract} and bidirectional adversarial training~\cite{jiangbidirectional}. However, most existing SSDA works focus on image classification and the relevant semantic segmentation task involving dense pixel-level predictions is largely neglected. We focus on SSDA-based semantic segmentation (with few-shot target samples) and it is the first effort for this challenging task to the best of our knowledge.
\textbf{Jigsaw Puzzles} is a basic pattern recognition problem that aims to reconstruct an original image from its shuffled patches. In the field of computer science and artificial intelligence, solving jigsaw puzzles \cite{freeman1964apictorial,kosiba1994automatic} has been widely studied for a variety of tasks in image editing \cite{sholomon2014generalized,cho2009patch}, relic re-composition \cite{paumard2018image,brown2008system}, unsupervised visual representation learning \cite{santa2017deeppermnet,doersch2015unsupervised,noroozi2016unsupervised} and generalized network learning \cite{carlucci2019domain}. In this work, we employ Jigsaw Puzzle classifiers to learn contextual visual features (from labeled data) which are very suitable in semantic image segmentation. The learnt features are then exploited to align unlabeled target data for domain adaptive semantic segmentation.
\section{Method}
This section presents our proposed Semi-Supervised Domain Adaptive Segmentation (SSDAS) method. It consists of four subsections that focus on \textit{Task Definition}, \textit{Adaptive Cross-Domain Alignment (ACDA)} that performs adaptive inter-domain alignment according to the similarity of the learned
features, \textit{Progressive Intra-Domain Alignment (PIDA)} that replaces target-unlike source features by high-confidence target features for alignment between confident and unconfident target features, and \textit{Network Training}.
\subsection{Preliminaries}
\textbf{Task Definition:} We focus on the problem of semi-supervised domain adaptation (SSDA) in semantic segmentation. Given the labeled source images \{$X_{s} \subset \mathbb{R}^{H \times W \times 3}$, $Y_{s} \subset (1, C)^{H \times W}$\}, a few labeled target images \{$X_{t} \subset \mathbb{R}^{H \times W \times 3}$, $Y_{t} \subset (1, C)^{H \times W}$\} and a large number of unlabeled target images $X_{t^{u}} \subset \mathbb{R}^{H \times W \times 3}$ ($H$, $W$ and $C$ stands for image height, image width, and the number of semantic classes, respectively), the goal of SSDA-based semantic segmentation is to learn a model $G$ that performs well on unlabeled target-domain data $X_{t^{u}}$.
Under such data setup, a `S+T' model without any domain adaptation can be derived by training over the labeled source and target images in a fully supervised manner:
\begin{equation}
\begin{split}
\mathcal{L}_{s\text{+}t}(X_{s}, Y_{s}, X_{t}, Y_{t}; G) = l(G(X_{s}), Y_{s}) \\
+ l(G(X_{t}), Y_{t}),
\end{split}
\label{eq_S+T}
\end{equation}
where $l$ denotes the standard cross entropy loss.
\begin{figure*}[!ht]
\centering
\subfigure{\includegraphics[width=0.99\linewidth]{figures/figure_3.pdf}}
\vspace{-2pt}
\caption{
{Overview of the proposed semi-supervised domain adaptive segmentation (SSDAS) network: SSDAS is a bidirectional training framework that consists of two alternative learning processes, namely, labeled data flow (top part) and unlabeled data flow (bottom part). In labeled data flow, we first feed the labeled source and target images ($i.e.$, $X_{s}$ and $X_{t}$) into the segmentation model $G$ to acquire the predicted segmentation maps $P_{s}$ and $P_{t}$, and then employ two Jigsaw Puzzle classifiers ($i.e.$, $\mathcal{J}_{s}$ and $\mathcal{J}_{t}$) to learning context features by solving jigsaw puzzles on $P_{s}$ and $P_{t}$, respectively. In this process, Steps \textcolor{red}{A}, \textcolor{red}{B} and \textcolor{red}{D} are triggered for re-weighting each source feature, re-estimating feature centers with different treatments and removing the dissimilar source features, respectively. In unlabeled data flow ($i.e.$, for $X_{t^{u}}$), we fix the learnt $\mathcal{J}_{s}$ and update $G$ to enforce the the segmentation maps ($i.e.$, $P_{t^{u}}$) to have source-like context features by solving jigsaw puzzles. In this process, Steps \textcolor{red}{C} and \textcolor{red}{D} are triggered for
aligning unconfident target context features to estimated feature centers and adding high-confidence target features to update $G$ and $\mathcal{J}_{s}$, respectively. }
}
\label{fig:stru}
\end{figure*}
\textbf{Context Features Learning.}
Inspired by \cite{noroozi2016unsupervised,carlucci2019domain}, we employ Jigsaw Puzzle classifier $\mathcal{J}$ to learn context features by solving jigsaw puzzles on the predicted segmentation map $P=G(X)$:
\begin{equation}
\mathcal{L}_{jig}(P;G, \mathcal{J}) = l(\mathcal{J}(\text{S}(P, I)), I),
\label{eq_base_jigsaw}
\end{equation}
where `S' stands for a function that decomposes the input image into $n \times n$ patches, and shuffles and re-assigns each patch to one of the $n^{2}$ grid positions (Index $I$ records the original location index of every patch). $\mathcal{J}$ solves jigsaw puzzle ($i.e.$, restoring the shuffled patches) by predicting the original location index ($i.e.$, $I$) of patches.
Note context alignment is a simple yet efficient approach in UDA-based semantic segmentation \cite{tsai2018learning,vu2019advent,luo2019taking,tsai2019domain,huang2020contextual}. For example, \cite{tsai2018learning,vu2019advent,luo2019taking} employ adversarial learning \cite{goodfellow2014generative} to align context features at an image level. \cite{tsai2019domain,huang2020contextual} first sample and cluster patches on labeled data to discover regional context features and then conduct adversarial learning to align the regional context features at a region level. Different from the aforementioned methods, our Jigsaw Puzzle method works at both region and image levels and it learns different types of context features by adjusting the jigsaw puzzle sizes and locations. Please refer to \textcolor{red}{A.5.} in supplementary material for more details.
\subsection{Adaptive Cross-Domain Alignment}
This subsection describes adaptive cross-domain alignment (ACDA) that employs a few labeled target samples as references to re-weight the \textbf{features} of source samples. The features of unlabeled target samples will be aligned to the re-weighted source features for cross-domain alignment.
\textbf{Labeled Data Flow:} For labeled source images \{$X_{s}$, $Y_{s}$\} and labeled target images \{$X_{t}$, $Y_{t}$\}, we first feed them into a segmentation model $G$ to acquire segmentation maps $P_{s} = G(X_{s}) \subset \mathbb{R}^{H \times W \times C}$ and $P_{t} = G(X_{t}) \subset \mathbb{R}^{H \times W \times C}$. We then employ two Jigsaw Puzzle classifiers ($i.e.$, $\mathcal{J}_{s}$ and $\mathcal{J}_{t}$) to learn context features by solving jigsaw puzzles on the predicted segmentation maps:
\begin{equation}
\min_{G, \mathcal{J}_{s}, \mathcal{J}_{t}} \lambda_{j}(\mathcal{L}_{jig}(P_{s};G,\mathcal{J}_{s})\mathcal{M}_{cds} + \mathcal{L}_{jig}(P_{t};G,\mathcal{J}_{t})),
\label{eq_ACDA_label}
\end{equation}
where $\mathcal{M}_{cds} = 1- \mathcal{N}(|\mathcal{J}_{s}(\text{S}(P_{s}, I)) - \mathcal{J}_{t}(\text{S}(P_{s}, I))|)$ is the cross-domain similarity (CDS) map that is computed based on the prediction discrepancy between the source and target Jigsaw Puzzle classifiers. It will be used to re-weight the learning loss of the source context features. $S$ is defined in Eq.~\ref{eq_base_jigsaw} and $\mathcal{N}(x)$ is an unity-based normalization function that brings all values into the range $[0,1]$. As illustrated in Fig.~\ref{fig:intro}, the multiplication with $\mathcal{M}_{cds}$ corresponds to Step \textcolor{red}{A} for source feature re-weighting, and the optimization of this equation corresponds to Step \textcolor{red}{B} for feature center updating.
\textbf{Unlabeled Data Flow:} For unlabeled target images $X_{t^{u}}$, we fix the learnt Jigsaw Puzzle classifier $\mathcal{J}_{s}$ while updating the segmentation model $G$ to solve jigsaw puzzles on the predicted segmentation maps $P_{t^{u}} = G(X_{t^{u}})$. This will enforce the segmentation maps $P_{t^{u}}$ to have source-alike context features:
\begin{equation}
\min_{G} \lambda_{j}(\mathcal{L}_{jig}(P_{t^{u}};G,\mathcal{J}_{s}))
\label{eq_ACDA_unlabel}
\end{equation}
where the optimization of this equation corresponds to Step \textcolor{red}{C} that aligns target features to the updated feature centers as illustrated in Fig.~\ref{fig:intro}.
\subsection{Progressive Intra-Domain Alignment}
This subsection describes our progressive intra-domain alignment (PIDA) that replaces target-unlike source features by high-confidence target features iteratively in training. It leads to intra-domain alignment between confident and unconfident target context features.
\textbf{Labeled Data Flow:} Similar to Eq.~\ref{eq_ACDA_label}, a binary removing mask $\mathcal{M}_{rm}$ is employed to discard target-unlike source features continuously during the iterative training process, where the labeled target data flow remains unchanged:
\begin{equation}
\min_{G, \mathcal{J}_{s}, \mathcal{J}_{t}} \lambda_{j}(\mathcal{L}_{jig}(P_{s};G,\mathcal{J}_{s})\mathcal{M}_{rm} + \mathcal{L}_{jig}(P_{t};G,\mathcal{J}_{t})),
\label{eq_PIDA_label}
\end{equation}
where $\mathcal{M}_{rm}$ is obtained as described in Algorithm~\ref{algorithm_mask_rm}. As illustrated in Fig.~\ref{fig:intro}, the multiplication with the binary mask $\mathcal{M}_{rm}$ corresponds to the Step \textcolor{red}{D} (for removing target-unlike source features),
and the optimization of this equation correspond to
the Step \textcolor{red}{B}
.
\textbf{Unlabeled Data Flow:} Based on Eq.~\ref{eq_ACDA_unlabel}, a binary add mask $\mathcal{M}_{add}$ is employed to inject high-confidence target features progressively for updating $G$ and $\mathcal{J}_{s}$, where the unconfident target features will be aligned to confident target features by fixing $\mathcal{J}_{s}$ while updating $G$ during training:
\begin{equation}
\begin{split}
&\min_{G, \mathcal{J}_{s}} \lambda_{j}\mathcal{L}_{jig}(P_{t^{u}};G,\mathcal{J}_{s})\mathcal{M}_{add},\\
&\min_{G} \lambda_{j}\mathcal{L}_{jig}(P_{t^{u}};G,\mathcal{J}_{s})(1-\mathcal{M}_{add}),
\end{split}
\label{eq_PIDA_unlabel}
\end{equation}
where $\mathcal{M}_{add}$ is obtained as described in Algorithm~\ref{algorithm_mask_add}. As illustrated in Fig.~\ref{fig:intro}, the multiplication with the binary mask $\mathcal{M}_{add}$ corresponds to the Step \textcolor{red}{D} (injecting high-confidence target features), and the optimization of the first and second equations refer to Step \textcolor{red}{B} and Step \textcolor{red}{C}, respectively.
\begin{algorithm}[t]
\caption{Determination of $\mathcal{M}_{rm}$ in PIDA.
}\label{algorithm_mask_rm}
\begin{algorithmic}[1]
\REQUIRE Source images $X_{s}$; $epoch$ denotes the epoch times; segmentation model $G$; jigsaw puzzle classifiers $\mathcal{J}_{s}$ and $\mathcal{J}_{t}$
\ENSURE $\mathcal{M}_{rm}$
\STATE Calculate $\mathcal{M}_{cds}$ as in Eq.~\ref{eq_ACDA_label}
\STATE $\mathcal{M}_{cds}^{sort} = \text{sort}(\mathcal{M}_{cds}, \ \text{order=ascending})$
\STATE $N = \text{length}(\mathcal{M}_{cds}^{sort}) \times \frac{epoch}{max\_epoch}$
\STATE $\text{Thres} = \mathcal{M}_{cds}^{sort}[N]$
\STATE $\mathcal{M}_{rm}[\mathcal{M}_{cds}>=\text{Thres}] = 1$
\STATE $\mathcal{M}_{rm}[\mathcal{M}_{cds}<\text{Thres}] = 0$
\RETURN $\mathcal{M}_{rm}$
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[t]
\caption{Determination of $\mathcal{M}_{add}$ in PIDA.
}\label{algorithm_mask_add}
\begin{algorithmic}[1]
\REQUIRE Unlabeled target images $X_{t^{u}}$; $epoch$ denotes the epoch times; segmentation model $G$; jigsaw puzzle classifier $\mathcal{J}_{t}$
\ENSURE $\mathcal{M}_{add}$
\STATE Calculate jigsaw puzzle classification (JPC) probability $P_{t^{u}}^{jig} = \mathcal{J}_{t}(\text{S}(G(X_{t^{u}}), I)$ as in Eq.~\ref{eq_base_jigsaw}
\STATE Calculate JPC entropy $Ent_{t^{u}}^{jig} = P_{t^{u}}^{jig} \log P_{t^{u}}^{jig}$
\STATE $Ent_{t^{u}}^{jig\_sort} = \text{sort}(Ent_{t^{u}}^{jig}, \ \text{order=ascending})$
\STATE $N = \text{length}(Ent_{t^{u}}^{jig\_sort}) \times \frac{epoch}{max\_epoch}$
\STATE $\text{Thres} = Ent_{t^{u}}^{jig\_sort}[N]$
\STATE $\mathcal{M}_{add}[Ent_{t^{u}}^{jig}<=\text{Thres}] = 1$
\STATE $\mathcal{M}_{add}[Ent_{t^{u}}^{jig}>\text{Thres}] = 0$
\RETURN $\mathcal{M}_{add}$
\end{algorithmic}
\end{algorithm}
\subsection{Network Training}
With ACDA and PIDA as described in Sections 3.3 and 3.4, this subsection presents how ACDA and PIDA work together to achieve our proposed SSDAS. The \textit{labeled data flow} is optimized as follows:
\begin{equation}
\min_{G, \mathcal{J}_{s}, \mathcal{J}_{t}} \lambda_{j}(\mathcal{L}_{jig}(P_{s};G,\mathcal{J}_{s})\mathcal{M}_{cds}\mathcal{M}_{rm} + \mathcal{L}_{jig}(P_{t};G,\mathcal{J}_{t}))
\label{eq_SSDAS_label}
\end{equation}
The optimization function of the \textit{unlabeled data flow} remains unchanged as defined in Eq.~\ref{eq_PIDA_unlabel}
Note all optimization functions in Eqs.~\ref{eq_ACDA_label} - \ref{eq_SSDAS_label} are generic and can be applied for both image-level and region-level alignments. The only difference lies with the input. Specifically, the image-level alignment takes a whole segmentation map as input. But for \textit{region-level alignment}, the segmentation map $P$ ($i.e.$, $P_{s}$, $P_{t}$ and $P_{t^{u}}$) is cropped into $r \times r$ regions which are fed into the extra region-level Jigsaw Puzzle Classifiers ($\mathcal{J}_{s}^{region}$ and $\mathcal{J}_{t}^{region}$) for regional context features learning and alignment. The entire training pipeline is summarized in Algorithm.~\ref{algorithm_SSDAS_entire}.
\begin{algorithm}[t]
\caption{The proposed Semi-Supervised Domain Adaptive Segmentation (SSDAS).
}\label{algorithm_SSDAS_entire}
\begin{algorithmic}[1]
\REQUIRE Labeled source images $\{X_{s}, Y_{s}\}$ and target images $\{X_{t}, Y_{t}\}$; unlabeled target images $X_{t^{u}}$; $epoch$ denotes the epoch times; segmentation model $G$; jigsaw puzzle classifiers $\mathcal{J}_{s}^{image}$, $\mathcal{J}_{t}^{image}$, $\mathcal{J}_{s}^{region}$ and $\mathcal{J}_{t}^{region}$; region crop size $r$
\ENSURE Learnt parameters $\theta$ of segmentation model $G$
\FOR{$epoch = 1$ \textbf{to} $max\_epoch$}
\STATE \textbf{\emph{Supervised segmentation learning:}}
\STATE Update $G$ using Eq.~\ref{eq_S+T} with $\{X_{s}, Y_{s}\}$ and $\{X_{t}, Y_{t}\}$
\STATE \textbf{\emph{Image-level alignment:}} ACDA\&PIDA
\STATE Calculate segmentation probability map $P_{s} = G(X_{s})$, $P_{t} = G(X_{t})$ and $P_{t^{u}} = G(X_{t^{u}})$
\STATE Update $G$, $\mathcal{J}_{s}^{image}$ and $\mathcal{J}_{t}^{image}$ using Eq.~\ref{eq_SSDAS_label} and Eq.~\ref{eq_PIDA_unlabel} with $P_{s}$, $P_{t}$ and $P_{t^{u}}$
\STATE \textbf{\emph{Region-level alignment:}} ACDA\&PIDA
\STATE Crop maps into $r \times r$ regions $P_{s}^{region} = \{P_{s}^{0},P_{s}^{1},...,P_{s}^{r^{2}}\}$, $P_{t}^{region} = \{P_{t}^{0},P_{t}^{1},...,P_{t}^{r^{2}}\}$ and $P_{t^{u}}^{region} = \{P_{t^{u}}^{0},P_{t^{u}}^{1},...,P_{t^{u}}^{r^{2}}\}$
\STATE Update $G$, $\mathcal{J}_{s}^{region}$ and $\mathcal{J}_{t}^{region}$ using Eq.~\ref{eq_SSDAS_label} and Eq.~\ref{eq_PIDA_unlabel} with $P_{s}^{region}$, $P_{t}^{region}$ and $P_{t^{u}}^{region}$
\ENDFOR
\RETURN $G$
\end{algorithmic}
\end{algorithm}
\renewcommand\arraystretch{1.1}
\begin{table}[t]
\centering
\resizebox{\columnwidth}{!}{
\begin{tabular}{cccc|c}
\hline
\multicolumn{2}{c|}{\multirow{1}{*}{Image-level Alignment}} &\multicolumn{2}{c|}{\multirow{1}{*}{Region-level Alignment}} & \multicolumn{1}{c}{\multirow{2}{*}{mIoU}}
\\\cline{1-4}
ACDA &\multicolumn{1}{c|}{PIDA} &ACDA &\multicolumn{1}{c|}{PIDA}
\\\hline
& & & &37.9\\\hline
\multicolumn{1}{c}{\checkmark} & & & &43.6\\
&\multicolumn{1}{c}{\checkmark} & & &41.7\\
\multicolumn{1}{c}{\checkmark} &\multicolumn{1}{c}{\checkmark} & & &45.1\\\hline
& &\multicolumn{1}{c}{\checkmark} & &44.8\\
& & &\multicolumn{1}{c|}{\checkmark} &41.9\\
& &\multicolumn{1}{c}{\checkmark} &\multicolumn{1}{c|}{\checkmark} &46.2\\\hline
\multicolumn{1}{c}{\checkmark}& &\multicolumn{1}{c}{\checkmark} & &46.3\\
&\multicolumn{1}{c}{\checkmark} & &\multicolumn{1}{c|}{\checkmark} &44.5\\
\multicolumn{1}{c}{\checkmark} &\multicolumn{1}{c}{\checkmark} &\multicolumn{1}{c}{\checkmark} &\multicolumn{1}{c|}{\checkmark} &\textbf{48.5}\\
\hline
\end{tabular}}
\caption{
Ablation study of SSDAS over semi-supervised domain adaptive segmentation task GTA $\rightarrow$ Cityscapes. The \textit{1st} row shows ``S+T" model that is trained with supervised segmentation loss with labeled source and target samples only (without any alignment) as defined in Eq.~\ref{eq_S+T}.
The backbone is ResNet-101 and the setting is one-shot as evaluated in mIoU.
}
\label{tab:abla}
\end{table}
\begin{table*}[t]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{c|c|ccccccccccccccccccc|c}
\hline
Setting &Method & Road & SW & Build & Wall & Fence & Pole & TL & TS & Veg. & Terrain & Sky & PR & Rider & Car & Truck & Bus & Train & Motor & Bike & mIoU\\
\hline
\multirow{10}{0.1\linewidth}{\centering{1-shot}} &S+T &{82.1} &22.8 &72.8 &20.3 &22.0 &26.7 &31.1 &11.8 &75.7 &24.0 &76.8 &56.2 &23.2 &71.4 &21.6 &20.7 &0.1 &28.5 &32.5 &37.9\\
&AdaptSeg~\cite{tsai2018learning} &{87.0} &32.8 &80.5 &22.1 &22.6 &27.7 &34.3 &25.4 &83.4 &31.5 &81.7 &59.5 &18.6 &76.0 &35.4 &42.9 &1.6 &28.0 &27.1 &43.1\\
&ADVENT~\cite{vu2019advent} &{92.3} &52.8 &81.9 &30.0 &25.6 &28.8 &35.4 &24.0 &84.9 &39.1 &79.4 &56.1 &21.4 &86.3 &31.1 &33.5 &1.2 &30.8 &11.1 &44.5\\
&CRST~\cite{zou2019confidence} &{91.5} &51.7 &81.6 &28.3 &\textbf{28.5} &\textbf{43.9} &45.0 &26.7 &85.2 &35.4 &65.0 &\textbf{68.8} &29.2 &85.5 &32.8 &28.4 &1.3 &31.0 &\textbf{42.0} &47.5\\
&FDA~\cite{yang2020fda} &90.5 &40.9 &80.2 &26.0 &25.0 &31.1 &32.6 &34.5 &80.3 &33.7 &79.5 &54.0 &30.2 &83.8 &35.0 &41.2 &11.8 &21.1 &30.7 &45.4\\
&IDA~\cite{pan2020unsupervised} &{92.7} &53.2 &82.9 &31.9 &19.7 &29.9 &35.6 &22.5 &86.2 &45.8 &\textbf{82.4} &57.8 &26.6 &87.9 &33.2 &42.7 &1.1 &\textbf{32.9} &24.3 &46.8\\\cline{2-22}
&\textbf{SSDAS} &92.9 &53.9 &82.1 &31.5 &24.0 &36.4 &40.6 &33.8 &84.5 &44.9 &69.6 &60.5 &25.4 &85.1 &\textbf{46.9} &52.4 &2.7 &19.2 &34.3 &48.5\\
&\textbf{+ADVENT} &{93.2} &54.1 &82.4 &32.8 &25.4 &34.0 &39.5 &36.7 &85.0 &45.7 &80.4 &60.7 &26.8 &84.3 &42.5 &\textbf{52.5} &3.1 &28.5 &29.4 &49.3\\
&\textbf{+CRST} &\textbf{{94.1}} &60.8 &\textbf{83.9} &\textbf{35.3} &17.5 &40.1 &\textbf{50.6} &\textbf{47.1} &85.2 &45.6 &81.2 &66.3 &26.9 &\textbf{88.6} &33.7 &49.1 &10.7 &22.7 &35.8 &\textbf{51.3}\\
&\textbf{+FDA} &93.9 &52.0 &83.3 &28.2 &26.8 &38.9 &37.6 &38.6 &82.7 &40.7 &81.4 &54.4 &\textbf{32.9} &86.6 &42.1 &49.3 &\textbf{14.8} &22.8 &34.9 &49.6\\
&\textbf{+IDA} &93.7 &\textbf{56.7} &83.8 &33.9 &25.7 &35.8 &40.3 &38.2 &\textbf{86.6} &\textbf{46.1} &81.2 &60.0 &25.1 &87.8 &34.1 &47.1 &3.3 &26.1 &34.8 &49.5\\
\hline
\multirow{10}{0.1\linewidth}{\centering{3-shot}} &S+T &{73.2} &29.9 &75.4 &17.6 &20.4 &30.5 &34.7 &24.5 &80.8 &26.0 &76.0 &58.1 &28.1 &45.1 &34.8 &34.1 &0.6 &26.9 &37.0 &39.7\\
&AdaptSeg~\cite{tsai2018learning} &86.6 &43.3 &80.7 &22.1 &21.9 &26.1 &33.4 &27.0 &82.8 &28.8 &80.6 &58.1 &26.1 &77.9 &37.2 &42.0 &1.1 &24.8 &29.6 &43.7\\
&ADVENT~\cite{vu2019advent} &92.3 &51.1 &81.8 &29.7 &23.7 &31.9 &32.7 &18.2 &84.4 &35.5 &75.6 &57.8 &21.5 &86.7 &35.3 &48.0 &0.7 &28.4 &19.7 &45.0\\
&CRST~\cite{zou2019confidence} &92.2 &52.1 &81.6 &24.7 &27.5 &41.0 &\textbf{45.8} &27.6 &83.6 &34.2 &76.4 &63.6 &22.4 &86.3 &33.1 &48.0 &5.9 &28.3 &38.2 &48.0\\
&FDA~\cite{yang2020fda} &91.6 &45.5 &82.4 &25.8 &25.5 &30.7 &34.2 &30.2 &82.7 &29.2 &79.9 &60.3 &28.1 &87.0 &33.2 &37.4 &8.8 &22.6 &34.8 &45.8\\
&IDA~\cite{pan2020unsupervised} &92.6 &52.2 &83.3 &30.3 &26.7 &33.0 &35.7 &25.2 &85.2 &42.9 &79.5 &59.2 &27.0 &87.3 &37.0 &48.9 &4.5 &30.2 &12.5 &47.0\\\cline{2-22}
&\textbf{SSDAS} &92.4 &52.6 &83.7 &27.0 &20.9 &37.5 &41.5 &36.3 &85.3 &42.2 &78.7 &63.1 &32.0 &86.8 &45.7 &49.6 &5.9 &20.2 &42.3 &49.7\\
&\textbf{+ADVENT} &\textbf{93.9} &52.0 &83.3 &29.2 &26.8 &38.9 &40.6 &38.6 &\textbf{85.7} &41.7 &81.4 &59.4 &32.9 &86.6 &42.1 &49.3 &\textbf{14.8} &22.8 &34.9 &50.3\\
&\textbf{+CRST} &93.6 &\textbf{58.1} &83.5 &\textbf{32.1} &\textbf{28.1} &\textbf{42.7} &43.8 &\textbf{41.4} &85.5 &42.1 &81.8 &\textbf{64.9} &\textbf{33.7} &\textbf{87.9} &\textbf{46.8} &49.5 &13.5 &\textbf{31.9} &43.3 &\textbf{52.9}\\
&\textbf{+FDA} &93.3 &49.8 &83.1 &31.3 &25.7 &37.2 &39.1 &38.0 &85.6 &\textbf{43.2} &81.4 &61.9 &30.1 &86.4 &43.9 &\textbf{50.3} &9.4 &25.6 &\textbf{43.9} &50.5\\
&\textbf{+IDA} &93.6 &56.3 &\textbf{84.3} &25.3 &21.6 &38.2 &41.6 &38.7 &85.6 &40.0 &\textbf{82.5} &61.6 &29.4 &86.0 &42.2 &48.6 &7.6 &29.9 &42.6 &50.3\\
\hline
\end{tabular}
}
\caption{
Comparing SSDAS with state-of-the-art semantic segmentation methods: For semi-supervised domain adaptive semantic segmentation task GTA $\rightarrow$ \textbf{City}scapes, the proposed SSDAS consistently outperforms all state-of-the-art UDA methods that are adapted for the SSDA task. In addition, SSDAS is clearly complementary to all adapted UDA methods with clear performance gains. The backbone is ResNet-101 for all compared methods, and the setting includes 1-shot and 3-shot.
}
\label{table:gta2city}
\end{table*}
\begin{table*}[t]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{c|c|cccccccccccccccc|c|c}
\hline
Setting &Method & Road & SW & Build & Wall\textsuperscript{*} & Fence\textsuperscript{*} & Pole\textsuperscript{*} & TL & TS & Veg. & Sky & PR & Rider & Car & Bus & Motor & Bike & mIoU & mIoU\textsuperscript{*}\\
\hline
\multirow{10}{0.1\linewidth}{\centering{1-shot}} &S+T &56.3 &17.6 &76.3 &9.9 &2.1 &28.3 &14.3 &13.8 &80.0 &80.9 &51.2 &14.2 &43.1 &22.1 &19.0 &24.6 &34.6 &39.5\\
&AdaptSeg~\cite{tsai2018learning} &84.2 &41.4 &78.0 &10.1 &0.5 &27.8 &6.1 &11.7 &81.8 &79.5 &54.3 &21.9 &70.8 &35.3 &12.3 &32.8 &40.5 &46.9\\
&ADVENT~\cite{vu2019advent} &87.2 &45.3 &78.7 &9.2 &2.8 &23.6 &6.7 &14.7 &80.9 &83.3 &58.8 &22.0 &71.3 &31.4 &10.8 &35.1 &41.4 &48.2\\
&CRST~\cite{zou2019confidence} &69.9 &31.8 &74.9 &14.8 &3.6 &37.0 &22.5 &29.5 &81.6 &79.1 &58.2 &28.8 &83.6 &27.2 &22.9 &46.6 &44.5 &50.5\\
&FDA~\cite{yang2020fda} &78.2 &32.5 &73.0 &11.2 &2.4 &27.8 &16.1 &17.5 &80.0 &82.0 &52.5 &24.7 &74.3 &34.3 &20.0 &39.8 &41.6 &48.1\\
&IDA~\cite{pan2020unsupervised} &84.7 &38.3 &78.8 &10.5 &2.9 &27.0 &13.6 &12.3 &80.5 &\textbf{83.8} &57.7 &22.9 &72.0 &38.0 &20.7 &35.8 &42.5 &49.2 \\\cline{2-20}
&\textbf{SSDAS} &87.6 &44.4 &79.1 &13.2 &3.2 &29.8 &14.1 &18.4 &81.0 &80.8 &58.0 &26.3 &77.3 &39.7 &19.4 &38.6 &44.4 &51.1\\
&\textbf{+ADVENT} &\textbf{88.6} &\textbf{46.2} &78.8 &14.1 &2.2 &28.9 &16.8 &21.5 &80.9 &82.5 &58.9 &25.5 &78.7 &37.5 &20.3 &41.9 &45.2 &52.2\\
&\textbf{+CRST} &86.8 &45.7 &79.4 &\textbf{15.2} &3.1 &\textbf{39.7} &\textbf{24.0} &\textbf{31.9} &\textbf{82.5} &79.6 &57.7 &\textbf{29.1} &\textbf{84.4} &\textbf{41.5} &\textbf{25.2} &\textbf{48.7} &\textbf{48.4} &\textbf{55.1}\\
&\textbf{+FDA} &86.3 &43.6 &78.1 &14.5 &\textbf{4.1} &32.3 &18.6 &20.0 &80.1 &82.8 &57.4 &27.9 &79.8 &40.1 &22.8 &43.9 &45.8 &52.4\\
&\textbf{+IDA} &87.3 &44.9 &\textbf{80.2} &12.2 &2.1 &30.8 &12.3 &21.3 &81.8 &83.7 &\textbf{59.4} &27.1 &75.4 &40.9 &21.8 &41.9 &45.2 &52.2\\
\hline
\multirow{10}{0.1\linewidth}{\centering{3-shot}} &S+T &58.4 &24.3 &77.3 &9.8 &2.4 &27.4 &12.6 &16.1 &77.7 &78.7 &51.6 &18.7 &40.0 &28.7 &17.0 &30.3 &35.7 &40.9\\
&AdaptSeg~\cite{tsai2018learning} &84.7 &38.9 &78.1 &12.6 &2.1 &28.2 &8.8 &12.3 &80.9 &80.1 &55.4 &19.5 &72.7 &35.6 &14.6 &32.2 &41.0 &47.2\\
&ADVENT~\cite{vu2019advent} &86.8 &44.3 &79.2 &14.1 &4.1 &28.6 &13.2 &18.8 &81.4 &82.7 &56.7 &21.0 &77.2 &33.1 &12.6 &27.1 &42.6 &48.8\\
&CRST~\cite{zou2019confidence} &72.6 &36.3 &76.9 &15.2 &4.1 &37.5 &21.1 &28.6 &81.9 &82.5 &57.8 &27.4 &82.4 &30.5 &20.1 &42.2 &44.8 &50.8\\
&FDA~\cite{yang2020fda} &81.6 &36.6 &74.2 &15.6 &3.1 &26.5 &18.1 &19.9 &82.3 &83.0 &55.2 &18.6 &80.7 &27.5 &19.9 &34.5 &42.3 &48.6\\
&IDA~\cite{pan2020unsupervised} &85.5 &40.0 &78.6 &14.2 &3.1 &25.3 &16.3 &18.3 &80.5 &82.2 &54.9 &19.4 &75.9 &39.6 &19.8 &32.5 &42.9 &49.5
\\\cline{2-20}
&\textbf{SSDAS} &88.5 &45.1 &78.2 &15.0 &3.0 &29.2 &19.9 &21.3 &80.8 &82.6 &58.5 &26.0 &76.0 &37.9 &21.4 &40.0 &45.2 &52.0\\
&\textbf{+ADVENT} &88.7 &46.6 &79.9 &16.9 &4.1 &32.6 &20.2 &20.3 &81.9 &83.5 &57.5 &26.4 &80.7 &38.1 &22.3 &38.7 &46.2 &52.7\\
&\textbf{+CRST} &\textbf{89.1} &\textbf{46.7} &\textbf{80.2} &16.6 &4.2 &\textbf{40.5} &\textbf{23.8} &\textbf{32.9} &82.8 &83.3 &\textbf{59.7} &\textbf{29.3} &\textbf{84.6} &\textbf{41.0} &\textbf{24.1} &\textbf{47.3} &\textbf{49.1} &\textbf{55.8}\\
&\textbf{+FDA} &88.0 &43.7 &78.0 &\textbf{17.6} &\textbf{5.2} &31.8 &22.9 &20.4 &\textbf{83.4} &\textbf{83.8} &56.2 &27.7 &81.0 &38.1 &22.6 &42.9 &46.5 &53.0\\
&\textbf{+IDA} &87.9 &46.3 &79.4 &16.4 &4.0 &30.7 &22.0 &20.1 &82.6 &83.1 &57.3 &27.9 &80.8 &40.2 &20.4 &38.9 &46.1 &52.8\\
\hline
\end{tabular}
}
\caption{Comparing SSDAS with state-of-the-art semantic segmentation methods: For semi-supervised domain adaptive semantic segmentation task SYNTHIA $\rightarrow$ Cityscapes, the proposed SSDAS consistently outperforms all state-of-the-art UDA methods that are adapted for the SSDA task. In addition, SSDAS is clearly complementary to all adapted UDA methods with clear performance gains. The backbone is ResNet-101 for all compared methods, and the setting includes 1-shot and 3-shot.}
\label{table:synthia2city}
\end{table*}
\begin{table}[h]
\centering
\vspace{1mm}
\footnotesize
\begin{tabular}{l|c|cc} \hline
\multirow{2}{*}{Network} & \multirow{2}{*}{Method} &\multicolumn{2}{c}{GTA $\rightarrow$ City}\\
& & 1-shot & 3-shot\\\hline
\multirow{4}{*}{ResNet-101} &S+T &37.9& 39.7\\
&ENT~\cite{grandvalet2005semi} &42.8 &43.5\\
&MME~\cite{saito2019semi} &43.2 &43.8\\
&\textbf{SSDAS} &\textbf{48.5} &\textbf{49.7}\\
\hline
\end{tabular}
\caption{
Comparing our SSDAS with state-of-the-art SSDA classification methods (in mIoU): For domain adaptive semantic segmentation task GTA $\rightarrow$ \textbf{City}scapes, SSDAS outperforms the state-of-the-art by large margins consistently for both 1-shot and 3-shot.}
\label{tab:SSDA_classification_comp1}
\end{table}
\begin{table*}[t]
\begin{center}
\scalebox{0.82}{
\begin{tabular}{c|l|cccccccccccc|c}
\hline
Network& Method &R to C& R to P & R to A & P to R & P to C & P to A & A to P & A to C & A to R & C to R & C to A & C to P & MEAN \\\hline
\multirow{5}{*}{ResNet-34} & S+T & 55.7 & 80.8 & 67.8 & 73.1 & 53.8 & 63.5 & 73.1 & 54.0 & 74.2 & 68.3 & 57.6 & 72.3 & 66.2\\
&ENT~\cite{grandvalet2005semi} & 62.6 & 85.7 & 70.2 & 79.9 & 60.5 & 63.9 & 79.5 & 61.3 & 79.1 & 76.4 & 64.7 & 79.1 & 71.9\\
& MME~\cite{saito2019semi} & 64.6 & 85.5 & 71.3 & 80.1 & 64.6 & 65.5 & 79.0 & 63.6 & 79.7 & 76.6 & \textbf{67.2} & 79.3 & 73.1 \\
& APE~\cite{kim2020attract} & 66.4 & 86.2 & 73.4 & 82.0 & 65.2 & 66.1 & 81.1 & \textbf{63.9} & 80.2 & 76.8 & 66.6 & 79.9 & 74.0\\
& \textbf{SSDAS} &\textbf{69.1} &\textbf{86.9} &\textbf{76.2} &\textbf{83.4} &\textbf{66.8} &\textbf{67.5} &\textbf{83.5} &63.8 &\textbf{82.3} &\textbf{77.9} &67.0 &\textbf{81.1} &\textbf{75.5}\\
\hline
\end{tabular}}
\end{center}
\caption{
Comparing SSDAS with state-of-the-art SSDA classification methods: For domain adaptive image classification task (3-shot), the proposed SSDAS outperforms the state-of-the-art clearly in most of 12 adaptation scenarios in Office-home dataset.}
\label{tab:SSDA_classification_comp2}
\end{table*}
\begin{figure*}[t]
\begin{tabular}{p{3cm}<{\centering} p{3cm}<{\centering} p{3cm}<{\centering} p{3cm}<{\centering} p{3cm}<{\centering}}
\raisebox{-0.5\height}{Target Image}
& \raisebox{-0.5\height}{S+T}
& \raisebox{-0.5\height}{IDA~\cite{pan2020unsupervised}}
& \raisebox{-0.5\height}{\textbf{SSDAS (Ours)}}
& \raisebox{-0.5\height}{Ground Truth}
\\
\raisebox{-0.5\height}{\includegraphics[width=1.1\linewidth,height=0.55\linewidth]{figures/segmentation/frankfurt_000001_062016_leftImg8bit-min.png}}
& \raisebox{-0.5\height}{\includegraphics[width=1.1\linewidth,height=0.55\linewidth]{figures/segmentation/frankfurt_000001_062016_leftImg8bit_color_ST.png}}
& \raisebox{-0.5\height}{\includegraphics[width=1.1\linewidth,height=0.55\linewidth]{figures/segmentation/frankfurt_000001_062016_leftImg8bit_color_ida.png}}
& \raisebox{-0.5\height}{\includegraphics[width=1.1\linewidth,height=0.55\linewidth]{figures/segmentation/frankfurt_000001_062016_leftImg8bit_color_ours.png}}
& \raisebox{-0.5\height}{\includegraphics[width=1.1\linewidth,height=0.55\linewidth]{figures/segmentation/frankfurt_000001_062016_gtFine_color.png}}
\vspace{-2.5 pt}
\\
\end{tabular}
\begin{tabular}{p{5.18cm}<{\centering} p{5.18cm}<{\centering} p{5.18cm}<{\centering}}
\raisebox{-0.5\height}{\footnotesize{$\sigma_{w}^{2}=530.11$, $\sigma_{b}^{2}=42.60$}}
& \raisebox{-0.5\height}{\footnotesize{$\sigma_{w}^{2}=306.28$, $\sigma_{b}^{2}=46.25$}}
& \raisebox{-0.5\height}{\footnotesize{$\sigma_{w}^{2}=196.75$, $\sigma_{b}^{2}=57.71$}}
\vspace{-10 pt}
\\
\raisebox{-0.5\height}{\includegraphics[width=1\linewidth,height=0.8\linewidth]{figures/feature/baseline_530_11__42_60.png}}
& \raisebox{-0.5\height}{\includegraphics[width=1\linewidth,height=0.8\linewidth]{figures/feature/ida_306_28__46_25.png}}
& \raisebox{-0.5\height}{\includegraphics[width=1\linewidth,height=0.8\linewidth]{figures/feature/ssdas_196_75__57_71.png}}
\vspace{-2.5 pt}
\\
\raisebox{-0.5\height}{S+T}
& \raisebox{-0.5\height}{IDA~\cite{pan2020unsupervised}}
& \raisebox{-0.5\height}{\textbf{SSDAS (Ours)}}
\\
\end{tabular}
\vspace{2.5 pt}
\caption{
\textbf{First row:} Qualitative illustration of semi-supervised domain adaptive semantic segmentation for GTA5 $\rightarrow$ Cityscapes adaptation. SSDAS employs a few labeled target sample as anchors for adaptive and progressive feature alignment between labeled source samples and unlabeled target samples, which produces nice semantic segmentation especially for the challenging low-frequency categories such as pole, bus and traffic-light etc. \textbf{Second row:} t-SNE \cite{maaten2008visualizing} visualization of feature distribution for target images in task GTA $\rightarrow$ Cityscapes: Each colour represents one semantic class of image pixels with a digit showing the class centre. $\sigma_{w}^{2}$ and $\sigma_{b}^{2}$ on the top of each graph are intra-class variance and inter-class distance of the corresponding feature distribution. The proposed SSDAS greatly outperforms ``S+T" and ``IDA" baselines in domain adaptive semantic segmentation qualitatively and quantitatively.
}
\label{fig:results}
\end{figure*}
\section{Experiments}
This section presents the evaluation of our SSDAS including datasets and implementation details, comparisons with the state-of-the-art, ablation studies, and discussion, more details to be described in the ensuing subsections.
\subsection{Experiment Setups}
In our experiments, we followed the setting of ~\cite{saito2019semi} that focuses on SSDA-based image classification. The training data consist of three parts including labeled source samples, unlabeled target samples, and 1 or 3 labeled target samples that are randomly selected for 1-shot or 3-shot SSDA-based semantic segmentation, respectively.
\textbf{Datasets:} We evaluated SSDAS over two challenging domain adaptive segmentation tasks GTA5$\rightarrow$Cityscapes and SYNTHIA$\rightarrow$Cityscapes, which involve two synthetic source datasets and one real target dataset. Specifically, Cityscapes consists of $2975$ training images and $500$ validation images. GTA5 and SYNTHIA consist of $24,966$ and $9,400$ high-resolution synthetic images which share 19 and 16 classes with Cityscapes. We also evaluated SSDAS on domain adaptive classification over the dataset Office-Home~\cite{venkateswara2017deep} that has 65 image classes of 4 domains.
\textbf{Implementation Details.} All our experiments were implemented in Pytorch. The segmentation model $G$ uses ResNet101 \cite{he2016deep} (pre-trained with ImageNet \cite{deng2009imagenet}) with DeepLab-V2~\cite{chen2017deeplab}.
The optimizer is SGD \cite{bottou2010large} with a momentum of $0.9$ and a weight decay of $1e-4$. The learning rate is $2.5e-4$ initially and decreased by a polynomial policy with a power of $0.9$. Except parameter studies in Table~\ref{tab:abla_weight_jigsaw}, we set the trade-off parameter $\lambda_{j}$ at 0.1 and the number of Jigsaw Puzzle classes $N$ at 100 in all other experiments. The Jigsaw Puzzle classifiers (\ie, $\mathcal{J}_{s}$ and $\mathcal{J}_{t}$) are pre-trained with ``S+T" model to avoid noisy predictions at the initial training stage. We freeze Jigsaw Puzzle classifier $\mathcal{J}_{t}$ when its training loss is smaller than $\mathcal{J}_{s}$'s to avoid over-fitting with just a few labeled target samples. For SSDA-based classification, we follow the setting in~\cite{saito2019semi}.
\subsection{Ablation Studies}
We first examine different SSDAS components to study their contributions to SSDA-based semantic segmentation. Table \ref{tab:abla} shows experimental results over the validation set of Cityscapes, where the first row shows the result of ``S+T" model that is trained with supervised loss with labeled source and target samples only (with no alignment) as defined in Eq.~\ref{eq_S+T}. It can be seen that ``S+T" model does not perform well due to cross-domain gaps.
However, ACDA improves ``S+T" model clearly at both region and image levels, largely because ACDA employs few-shot target features to re-weight source features adaptively by increasing (or decreasing) the weight of target-alike (or target-unlike) source features. Further including PIDA improves mIoU by another $+1.4\%$ at both image and region levels, demonstrating its effectiveness in intra-domain alignment. It also shows that ACDA and PIDA are complementary by focusing on cross-domain alignment and intra-domain alignment, respectively.
The last three rows show that including image-level alignment and region-level alignment simultaneously outperforms adopting either one alone for both ACDA and PIDA. This shows that our proposed image-level and region-level alignment are complementary, where the image-level alignment focuses more on classes with big sizes (\eg, road, building, sky, etc.) while the region-level alignment focuses more on classes with small sizes (\eg, person, car, bike, etc.). Finally, including both alignment strategies at both image and region levels (\ie, the complete SSDAS model) performs clearly the best.
\subsection{Comparisons with the State-of-Art}
As there is few prior work on SSDA-based semantic segmentation, we conducted two sets of experiments to benchmark our SSDAS with the state-of-the-art. In the first set of experiments, we adapted state-of-the-art UDA methods for the SSDA task. Specifically, we included the labeled target samples and the corresponding supervised loss into the UDA methods to approximate SSDA-based semantic segmentation. Tables \ref{table:gta2city} and \ref{table:synthia2city} show representative UDA methods (`AdaptSeg', `ADVENT', `CRST', `FDA' and `CrCDA') and their results. It can be seen that our SSDAS outperforms all adapted UDA methods consistently for both tasks GTA5$\rightarrow$Cityscapes and SYNTHIA$\rightarrow$Cityscapes. The superior segmentation is largely attributed to the adaptive and progressive feature alignment in SSDAS that exploits the few-shot labeled target samples to guide the cross-domain and intra-domain alignment effectively.
In the second set of experiments, we benchmarked SSDAS with state-of-the-art SSDA-based image classification methods for both semantic segmentation and image classification tasks. To adapt SSDAS for image classification, we simply take the feature maps instead of segmentation maps as the Jigsaw Puzzle input with little fine-tuning. Tables \ref{tab:SSDA_classification_comp1} and \ref{tab:SSDA_classification_comp2} show experimental results. For the semantic segmentation in Table \ref{tab:SSDA_classification_comp1}, we can see that SSDAS outperforms SSDA classification methods by large margins (over 5.3\% in mIoU) for both 1-shot and 3-shot settings. For the image classification task in Table \ref{tab:SSDA_classification_comp2}, SSDAS outperforms state-of-the-art SSDA classification methods clearly as well. These experiments show that SSDAS is generic for different tasks.
\subsection{Discussion}
\textbf{Number of labeled target samples:} We studied how SSDAS behaves while including more labeled target samples in training. The experiments in Table~\ref{tab:abla_number_labeled} shows that domain adaptive segmentation can be improved consistently when more labeled target samples are included.
\renewcommand\arraystretch{1.1}
\begin{table}[t]
\centering
\scalebox{0.9}{
\begin{tabular}{cp{0.9cm}<{\centering}p{0.9cm}<{\centering}p{0.9cm}<{\centering}p{0.9cm}<{\centering}p{0.9cm}<{\centering}}
\hline
& \multicolumn{5}{c}{The number of the labeled target samples}
\\\hline
Method & \multicolumn{1}{c}{1} & \multicolumn{1}{c}{3} & \multicolumn{1}{c}{5} & \multicolumn{1}{c}{10} &\multicolumn{1}{c}{20}
\\\hline
S+T &37.9 &39.7 &40.4 &41.7 &44.6\\
\textbf{SSDAS} &48.5 &49.7 &50.1 &51.1 &52.6\\
\hline
\end{tabular}}
\caption{
The number of labeled target samples matters: Domain adaptive semantic segmentation keeps improving with the increase of labeled target samples (over the task GTA $\rightarrow$ Cityscapes).
}
\label{tab:abla_number_labeled}
\end{table}
\renewcommand\arraystretch{1.1}
\begin{table}[t]
\centering
\scalebox{0.9}{
\begin{tabular}{cccccc}
\hline
& \multicolumn{5}{c}{Parameter Analysis of $\lambda_{j}$ and $N$}
\\\toprule
$\lambda_{j}$ & \multicolumn{1}{c}{0.025} & \multicolumn{1}{c}{0.05} & \multicolumn{1}{c}{0.1} & \multicolumn{1}{c}{0.2} & \multicolumn{1}{c}{0.4}
\\\hline
SSDAS &47.9 &48.3 &48.5 &48.4 &48.1\\\bottomrule
$N$ & \multicolumn{1}{c}{30} & \multicolumn{1}{c}{50} & \multicolumn{1}{c}{100} & \multicolumn{1}{c}{300} & \multicolumn{1}{c}{500}
\\\hline
SSDAS &48.2 &48.3 &48.5 &48.4 &48.4
\\
\hline
\end{tabular}}
\caption{
The weight parameter $\lambda_{j}$ and the number of Jigsaw Puzzle classes $N$ mater: Domain adaptive semantic segmentation is slightly affected by $\lambda_{j}$ but tolerant to $N$ (for GTA $\rightarrow$ Cityscapes).
}
\label{tab:abla_weight_jigsaw}
\end{table}
\begin{table}
\centering
\vspace{1mm}
\scalebox{0.9}{
\begin{tabular}{l|cc|c} \hline
Method & Convention-feature &Context-feature &mIoU\\\hline
\multirow{3}{*}{SSDAS} &\multicolumn{1}{c}{\checkmark} & &47.1\\
& &\multicolumn{1}{c|}{\checkmark} &48.5 \\
&\multicolumn{1}{c}{\checkmark} &\multicolumn{1}{c|}{\checkmark} &49.7 \\
\hline
\end{tabular}}
\caption{
SSDAS is generic and can work for other features such as conventional features. For the task GTA $\rightarrow$ Cityscapes, we employ SSDAS to aligning the convention-feature and context-feature, and present the domain adaptive semantic segmentation performance in mIoU.}
\label{tab:SSDA_with_conven_F}
\end{table}
\textbf{The weight $\lambda_{j}$ and the number of Jigsaw Puzzle classes $N$:} Parameters $\lambda_{j}$ and $N$ control the weights of supervised and unsupervised losses and the difficulty of Jigsaw Puzzle, respectively. We studied the two parameters by changing $\lambda_{j}$ from $0$ to $1$ with a step of $1/6$ and setting $N$ at a few number as shown in Table~\ref{tab:abla_weight_jigsaw}. Experiments over the task GTA$\rightarrow$Cityscapes show that SSDAS is tolerant to both $\lambda_{j}$ and $N$. The major reason is that the adaptive and progressive learning in SSDAS can alleviate `negative alignment' with dissimilar features.
\textbf{Context vs conventional features:} We use context feature in this work but SSDAS can also work with conventional segmentation features, i.e. the features from the feature extractor $E$ without including the following Jigsaw Puzzle classifiers. To work with conventional features, the algorithm and optimization functions are the same (as with context features) except that the context features are replaced by conventional features. Please refer to Section \textcolor{red}{A.1.} of supplementary materials for details about the optimization functions and algorithms of aligning conventional features. Table \ref{tab:SSDA_with_conven_F} compares SSDAS while working with context and conventional features. We can observe that SSDAS can work with convention features well though the segmentation performance drops clearly. The performance drop is largely due to the fact that context features capture context information and are more effective in semantic segmentation. In addition, experiments show that context and conventional features are complementary while working together in domain adaptive semantic segmentation.
Moreover, we provide qualitative comparison and feature distribution visualization over the task GTA5$\rightarrow$Cityscapes in Fig.~\ref{fig:results}.
Due to the space limit, more discussion and qualitative comparisons over domain adaptive segmentation task are provided in the appendix.
\section{Conclusion}
In this work, we presented SSDAS, a Semi-Supervised Domain Adaptive image Segmentation network that employs a few labeled target samples as anchors for adaptive and progressive feature alignment between labeled source samples and unlabeled target samples. To the best of our knowledge, this is the first effort towards semi-supervised domain adaptive semantic segmentation (with few-shot target samples).
Extensive experiments demonstrate the superiority of our SSDAS over a number of baselines including UDA-based segmentation and SSDA-based classification methods. In addition, SSDAS is complementary and can be easily integrated with UDA-based methods with consistent improvements in segmentation. We will explore how to better make use of a few labeled target samples in SSDA-based semantic segmentation. In addition, we will also study how to extend the idea of our SSDAS to other computer vision tasks such as object detection and panoptic segmentation.
\section*{Acknowledgement}
This research was conducted in collaboration with Singapore Telecommunications Limited and supported/partially supported (delete as appropriate) by the Singapore Government through the Industry Alignment Fund - Industry Collaboration Projects Grant.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
{
Understanding the evolution of a system in response to stimuli is {\em the} central problem in many areas. The Machine Learning (ML) approach to this problem has been to learn a collection of latent variables and construct a dynamical model of the system directly from observational data.
While ML has achieved strong predictive performance in some applications, it has two central weaknesses.
The first is that it requires large datasets. The second is that the latent variables that the ML approach identifies often have no physical interpretation and do not correspond to any previously-identified quantities. }
{
One approach to dealing with these weaknesses has been to incorporate expert domain knowledge into ML models.
Most of the work using this approach has focused on incorporating \textit{high-level} knowledge about the underlying physical system, such as conservation of energy \cite{bertalan2019learning,greydanus2019hamiltonian,zhong2019symplectic}, independence of mechanism \cite{parascandolo2018learning}, monotonicity \cite{muralidhar2018incorporating}, or linearity \cite{guen2020disentangling}. In addition, there have been attempts to integrate domain-specific ``expert models'' into ML models to create ``hybrid'' models. Most of this work has employed expert models that directly issue predictions \cite{liu2019multi,wang2017physics,xu2015data,yao2018tensormol} or extract useful features from the raw measurements \cite{karpatne2017physics}. }
{
The approach taken in this paper begins with an expert model in the form of a system of Ordinary Differential Equations (ODEs) and integrates that expert model into a system of Neural ODEs \cite{chen2018neural}. The specific problem we address is that of predicting disease progression under medications; the specific expert model(s) come from Pharmacology \cite{katzung2012basic} -- but we think our approach may be much more widely applicable.} For a number of diseases, available pharmacological models, built on the basis of specialized knowledge and laboratory experiments, provide a description of the dynamics of carefully-chosen medically meaningful variables in terms of ODEs that govern the evolution of these states \cite{danhof2007mechanism,agoram2007role,gesztelyi2012hill}.
However, these models are typically not directly applicable in clinical environments, because they involve too few variables to fully describe a patient's health state \cite{spoorenberg2014pharmacokinetics,holford2013pharmacokinetic}, because the expert variables which the models employ may be observable in the laboratory setting but not in clinical environments \cite{frank2003clinical, aronson2017biomarkers}, and because the relationships between the expert variables and clinically observable quantities is not known \cite{falvey2015disease}. We will give a example later in Section \ref{sec:decision}.
This paper proposes a novel hybrid modeling framework, the Latent Hybridisation Model (LHM), that imbeds a given pharmacological model (a collection of expert variables and the ODEs that describe the evolution of these variables ) into a larger latent variable ML model (a system of Neural ODEs). In the larger model, we use observational data to learn {\em both} the evolution of the unobservable latent variables and the relationship between measurements and {\em all} the latent variables -- the expert variables from the pharmacological model {\em and} the latent variables in the larger model. The machine learning component provides links between the expert variables and the clinical measurements, the underlying pharmacological model improves sample efficiency, and the expert variables provide additional insights to the clinicians. A variety of experiments (using synthetic and real data) demonstrate the effectiveness of our hybrid approach.
\section{Problem setting}
We consider a set of hospitalized patients $[N] = \{1, \ldots, N\}$ over a time horizon $[0,T]$; $t=0$ represents the time of admission and $t=T$ represents the maximal length of stay.
The health status of each patient $i$ is characterized by a collection of \textit{observable} physiological variables ${\mathbf{x}}_i(t) \in \mathbb{R}^D$, $D\in\mathbb{N}^+$; because the physiological variables may include vital signs, bloodwork values, biomarkers, etc., ${\mathbf{x}}_i(t)$ is typically a high-dimensional vector. Although the physiological variables are observable, they are typically \textit{measured} only at discrete times, and with error.
To avoid confusion, we distinguish the measurements of these variables ${\mathbf{y}}(t)$ from the true values; i.e.
\begin{equation}
\label{eq:error}
{\mathbf{y}}_i(t) = {\mathbf{x}}_i(t) + \epsilon_{it}
\end{equation}
where the independent noise term $\epsilon_{it}$ accommodates the measurement error (modeling $\epsilon_t$ as an autocorrelated stochastic process is left as a future work).
For illustrative purposes, we also assume that $\epsilon_t$ follows a Normal distribution $N(0, \sigma^2_i)$, but any parametric distribution could be easily accommodated.
We denote the measurement times for each patient as $\mathcal{T}_i = \{t_{i1}, t_{i2}, \ldots\}$.
We write ${\mathbf{a}}_i(t) \in \mathbb{R}^A$, $A\in\mathbb{N}^+$ for the {\em treatments} the patient receives. Some treatments (e.g. intravenous medications) are continuous; others (e.g. surgical interventions) are discrete, so some components of ${\mathbf{a}}_i(t)$ may be continuous functions but others are (discontinuous) step functions.
It is convenient to write $\mathcal{A}_i[t_1:t_2] = \{{\mathbf{a}}_i(t) | t_1 \le t \le t_2\}$ and $\mathcal{Y}_i[t_1:t_2] = \{{\mathbf{y}}_i(t) | t_1 \le t \le t_2, t \in \mathcal{T}_i \}$ for the treatments and measurements (respectively) during the the time window $[t_1,t_2]$. Note that $\mathcal{Y}_i[0:t]$ and
$\mathcal{A}_i[0:t]$ represent {\em histories} at time $t$ while $\mathcal{A}_i[t:T]$ and $\mathcal{Y}_i[t:T]$ represent {\em treatment plans} and {\em predictions}, respectively.
Our objective is to
{\em predict }the future measurements under a given treatment plan ${\mathcal{A}}_i[t_0:T]$ given the history:
\begin{equation}
\label{eq:1}
\mathbb{P}(\mathcal{Y}_i[t_0:T]\ |\ \underbrace{\mathcal{Y}_i[0:t_0], \mathcal{A}_i[0:t_0]}_\text{Historical observations}, \underbrace{{\mathcal{A}}_i[t_0:T]}_\text{Treatment plan}).
\end{equation}
Understanding this distribution will allow us to compute both point estimates and credible intervals (reflecting uncertainty). Note that it is important for the clinician to understand uncertainty in order to balance risk and reward. When the context is clear, we will omit the subscript $i$ and the time index $t$.
\section{Method}
\begin{figure}[t!]
\centering
\includegraphics{figure/graph.pdf}
\caption{\footnotesize Dependency structure of the three models designed for the laboratory or clinical settings. Dashed nodes represent unobservable variables. The expert variables ${\mathbf{z}}^e$ are observable in the laboratory setting but not in the clinical setting. The pharmacological model does not contain the links to the clinical variables ${\mathbf{x}}$.}
\label{fig:1}
\end{figure}
\subsection{The pharmacological model}
We begin with a pharmacological model which describes the dynamics of a collection of ``expert'' variables ${\mathbf{z}}^e(t) \in \mathbb{R}^E$. Each expert variable captures a distinct and medically-meaningful aspect of the human body, e.g. the activation of immune system.
The pharmacological model describes the dynamics as a system of Ordinary Differential Equations (ODEs):
\begin{equation}
\label{eq:expert}
\dot{{\mathbf{z}}}^e(t) = f^e({\mathbf{z}}^e(t), {\mathbf{a}}(t); \theta^e),
\end{equation}
where we have written $\dot{{\mathbf{z}}}^e(t) $ for the time derivative of ${\mathbf{z}}^e$. The functional form of
$f^e: \mathbb{R}^E \times \mathbb{R}^A \rightarrow \mathbb{R}^E$ is specified but the unknown parameters $\theta^e$ (e.g., coefficients) need to be estimated from data.\footnote{The system in Equation (\ref{eq:expert}) is quite general; appropriate choices of expert variables allow it to capture both high-order ODEs and time-dependent ODEs \cite{perko2013differential}.}${}^,$\footnote{Some care must be taken because systems such as (\ref{eq:expert}) do not always admit unique global solutions. In practice, the pharmacological models are sufficiently well-behaved that global solutions exist and are unique. Although closed-form solutions may not be available, there are various efficient numerical methods for solution.}
It is important to note that the system of ODEs (\ref{eq:expert}) describes dynamics that are \textit{self-contained}, in the sense that the time derivatives $\dot{{\mathbf{z}}}^e(t)$ depend only on the {\em current} values of the expert variables ${\mathbf{z}}^e(t)$ and the {\em current} treatments
${\mathbf{a}}(t)$, and not on histories or on other variables. To ensure that this obtains, it may be necessary to limit the scope of the model and limit attention to a single system of the body (or perhaps to several closely related systems) \cite{danhof2016systems}. As a consequence of these limitations, the expert variables will usually not give a full picture of the health status of the patient and will usually not account for the full array of {\em observable} physiological variables ${\mathbf{x}}(t)$ \cite{frank2003clinical}.
\subsection{The latent hybridisation model: linking expert variables with measurements}
As we have already noted, pharmacological models are typically developed and calibrated in the laboratory, where the expert variables can be directly measured -- in patients, in laboratory animals, or even in vitro (Figure \ref{fig:1} A). In clinical environments, the expert variables are frequently not observed (Figure \ref{fig:1} B and C). To use the pharmacological models in clinical environments, we must establish links between the expert variables ${\mathbf{z}}^e(t)$ and the clinical measurements ${\mathbf{y}}(t)$. To do this we introduce additional latent variables
${\mathbf{z}}^m(t)\in \mathbb{R}^M$ and posit the following relationship between the latent variables ${\mathbf{z}}^e, {\mathbf{z}}^m$ and the observable physiological variables ${\mathbf{x}}(t)$:
\begin{equation}
\label{eq:physiological}
{\mathbf{x}}(t) = g({\mathbf{z}}^e(t), {\mathbf{z}}^m(t), {\mathbf{a}}(t);\ \gamma)
\end{equation}
The function $g: \mathbb{R}^{E \times M \times A} \rightarrow \mathbb{R}^D$ is a neural network with (unknown) weights $\gamma$, and maps the latent space to the ``physiological space''. We also posit that the dynamics of the latent variables
${\mathbf{z}}^m(t)$ follow a system of ODEs governed by its current values ${{\mathbf{z}}}^m(t)$, the treatments ${\mathbf{a}}(t)$ {\em and} the current values of the expert variables ${\mathbf{z}}^e(t)$.
\begin{equation}
\label{eq:latent}
\dot{{\mathbf{z}}}^m(t) = f^m({\mathbf{z}}^m(t), {\mathbf{z}}^e(t), {\mathbf{a}}(t);\ \theta^m),
\end{equation}
The function $f^m: \mathbb{R}^{M \times E \times A} \rightarrow \mathbb{R}^M$ is a neural network with (unknown) weights $\theta^m$.
Equations (\ref{eq:expert})-(\ref{eq:latent}) specify the dynamics of LHM. It is convenient to write ${\mathbf{z}}(t) = [{\mathbf{z}}^e(t)\ {\mathbf{z}}^m(t)]$ for the vector of all latent variables and $\Theta = (\theta^e, \theta^m, \gamma, \sigma)$ for the set of all (unknown) coefficients.
The coefficients $\Theta$ will be learned from data. However, even after these coefficients are learned, the initial state of the patient ${\mathbf{z}}_i(0)$ is still {\em unknown}.
In fact, the variation in initial states reflects the heterogeneity of the patient population.
If the coefficients and the initial state were known, the entire trajectory of ${\mathbf{z}}_i$ given the treatments could be computed (numerically). Because we have assumed that the noise/errors $\epsilon_t$ are independent, we have
\begin{equation}
\label{eq:independence}
\mathcal{Y}_i[t_0:T] \perp \!\!\! \perp \mathcal{Y}_i[0:t_0]\ |\ {\mathbf{z}}_i(0), \mathcal{A}_i[0:T], \Theta, \quad \forall t_0 < T
\end{equation}
However, because the initial state is unknown, it must be {\em learned} from the measurements ${\mathbf{y}}_i(t)$.
LHM would reduce to a pure latent neural ODE model \cite{chen2018neural,rubanova2019latent} if we omitted the expert variables (Figure \ref{fig:1} B). However, that would amount to {\em discarding prior (expert) information} and so is evidently undesirable. Indeed, as we have noted in the introduction, our approach is driven by the idea of incorporating this prior (expert) information into our hybrid model.
In the current work, we assume that the pharmacological model in Equation \ref{eq:expert} is correct. In practice, the model might be wrong in two ways. The obvious way is that the functional form of $f^e$ might be misspecified (e.g. a linear model might be specified when the truth is actually nonlinear). Many existing techniques can address such misspecification and could be integrated into LHM \cite{hamilton2017hybrid,parish2016paradigm,zhang2018real}; see the discussion in Appendix \ref{app:extensions}. Alternatively, it might be that the system of expert variables is {\em not} self-contained, and that their evolution actually depends on additional \textit{latent} variables, we leave this more challenging problem for future work.
Practical extensions to LHM such as including static covariates and modeling informative sampling are discussed in Appendix \ref{app:extensions}.
\subsection{Independent and informative priors}
It may be challenging to pinpoint the exact value of the latent variables ${\mathbf{z}}^e$ based on observations (e.g. due to measurement noise or sampling irregularity).
For this reason, we quantify the uncertainty around ${\mathbf{z}}^e$ using Bayesian inference.
In what follows, we assume the initial states ${\mathbf{z}}_i(0)$ of patients are independently sampled from a prior distribution ${\mathbf{z}}_i(0) \sim \mathbb{P}_0$. Two points are worth noting.
\textbf{Independent Priors.} We use {\em independent} prior distributions on the expert variables ${\mathbf{z}}^e$ and the latent variables ${\mathbf{z}}^m$, i.e.
$\mathbb{P}({\mathbf{z}}(0)) = \mathbb{P}({\mathbf{z}}^e(0)) \times \mathbb{P}({\mathbf{z}}^m(0))$.
This guarantees that information in ${\mathbf{z}}^{m}(0)$ does not duplicate (any of the) information in ${\mathbf{z}}^{e}(0)$, which captures our belief that the latent variables are incremental to the expert variables.
In addition, independent priors are also commonly used in Bayesian latent variable models such as variational autoencoders (VAEs) \cite{kingma2013auto, higgins2016beta}.
\textbf{Informative Priors } The prior distribution on the expert variables $\mathbb{P}({\mathbf{z}}^e(0))$ should reflect domain knowledge. Such knowledge is usually available from previous studies in Pharmacology \cite{holford2013pharmacokinetic}. Using an informative prior tends to improve the estimation of latent variables, especially in small-sample settings \cite{lee2018determining}. Moreover, the expert variables usually take values in specific ranges (e.g. $[0, 10]$ \cite{katzung2012basic}) and going beyond the valid range may lead to divergence. The informative prior can encode such prior knowledge to stabilize training.
\begin{figure}[t!]
\centering
\includegraphics[width=0.8\textwidth]{figure/Inference.pdf}
\caption{Illustration of the training and prediction procedure.}
\label{fig:inference}
\end{figure}
\subsection{Model training and prediction via amortized variational inference}
\label{sec:inference}
Given the training dataset $\mathcal{D} = \{(\mathcal{Y}_i[0:T], \mathcal{A}_i[0:T])\}_{i \in [N]}$, we use amortized variational inference (AVI) to estimate the global parameters $\Theta$ and the unknown initial condition ${\mathbf{z}}_i(0)$ \cite{zhang2018advances}.
Figure \ref{fig:inference} presents a diagram of the training procedure. We start by learning a \textit{variational distribution} to approximate the posterior $\mathbb{P}({\mathbf{z}}_i(0)|\mathcal{Y}_i[0:T], \mathcal{A}_i[0:T])$. As is standard in AVI \cite{kingma2013auto, zhang2018advances},
we use a Normal distribution with diagonal covariance matrix as approximation:
\begin{equation}
\label{eq:variational}
\mathbb{Q}({\mathbf{z}}_i(0)|\mathcal{Y}_i[0:T], \mathcal{A}_i[0:T]) = N(\mu_i, \Sigma_i); \quad \mu_i, \Sigma_i = e(\mathcal{Y}_i[0:T], \mathcal{A}_i[0:T];\ \phi)
\end{equation}
Here the parameters $\mu_i, \Sigma_i$ are produced by an inference network (also known as an encoder) $e(\cdot)$ with trainable weights $\phi$.
When the context is clear, we will denote the variational distribution defined by Equations (\ref{eq:variational}) as $\mathbb{Q}_\phi$.
The evidence lower bound (ELBO) for the global parameter $\Theta$ and the inference network parameters $\phi$ is defined as
\begin{equation}
\text{ELBO}(\Theta, \phi) = \mathbb{E}_{{\mathbf{z}}(0) \sim \mathbb{Q}_\phi}\big[\text{log}\mathbb{P}(\mathcal{Y}_i[0:T] | \mathcal{A}_i[0:T], {\mathbf{z}}(0), \Theta)\big] - \text{KL}[\mathbb{Q}_\phi | \mathbb{P}_0]
\end{equation}
To compute the ELBO for a given $\Theta$ and $\phi$, we sample ${\mathbf{z}}(0) \sim \mathbb{Q}_\phi$ and numerically solve the ODEs to obtain ${\mathbf{z}}(t)$, $\forall t \in [0, T]$ (Figure \ref{fig:inference}; Steps 1, 2). Then, we compute the inner log-likelihood function using the mapping $g$ and the noise distribution (Equations (\ref{eq:error}) and (\ref{eq:physiological})).
Finally, we use Monte Carlo sampling to evaluate the KL divergence term: $\mathbb{E}_{{\mathbf{z}}(0) \sim \mathbb{Q}_\phi}[\text{log}\mathbb{Q}_\phi({\mathbf{z}}(0)) - \text{log}\mathbb{P}_0({\mathbf{z}}(0))]$. This is because the informative prior $\mathbb{P}_0$ may not have an analytical KL divergence (unlike the standard Normal prior used in previous works \cite{kingma2013auto,rubanova2019latent}).
We optimize ELBO by stochastic gradient ascent and update all parameters jointly in an end-to-end manner (detailed in Appendix \ref{app:optimization}).
The prediction procedure follows the same steps as illustrated in Figure \ref{fig:inference}.
For a new patient with history $\mathcal{Y}_i[0:t_0], \mathcal{A}_i[0:t_0]$, we first estimate the variational posterior $\mathbb{Q}_\phi$ using the trained encoder. From Equation (\ref{eq:independence}), we can estimate the target distribution in Equation (\ref{eq:1}) as:
\begin{equation}
\mathbb{E}_{{\mathbf{z}}(0)\sim \mathbb{Q}_\phi}\big[\mathbb{P}\big({\mathbf{y}}(t) | {\mathbf{z}}(0), \mathcal{A}_i[0:t_0], {\mathcal{A}}_i[t_0:T]\big)\big], \ \forall t > t_0.
\end{equation}
where ${\mathcal{A}}_i[t_0:T]$ is a future treatment plan.
The outer expectation can be approximated by Monte Carlo sampling from $\mathbb{Q}_\phi$ and the inner probability is given by the likelihood function.
\textbf{Choice of variational distribution and encoder}. The training procedure above is agnostic to the exact choice of variational distribution and encoder architecture. We choose the Normal distribution to make fair comparisons with the previous works \cite{chen2018neural,rubanova2019latent} .
For the same reason, we use the reversed time-aware LSTM encoder proposed in \cite{chen2018neural}.
In the Appendix \ref{app:simulation}, we show additional experiments with more complex variational distributions, i.e. Normalizing Flows \cite{rezende2015variational}.
\subsection{Using LHM to provide clinical decision support}
\label{sec:decision}
In order for clinicians to properly treat patients, they need to predict the progression of disease given the treatments. Although machine learning models may demonstrate feature importance \cite{choi2016retain, alaaattentive}, they do not uncover the relationships between those features and the underlying pathophysiology. LHM can provide the missing link between clinical observations and disease mechanisms. In combination with clinical reasoning, this can provide treating clinicians with decision support in several complementary ways.
First, LHM can inform the clinicians about the values of the expert variables ${\mathbf{z}}^e(t)$ that cannot be observed in the clinical environment but are important for prognosis, choice of treatment, and anticipation of complications. For example, understanding and predicting immune response is pivotal when deciding on immunosuppresive therapy in the treatment of COVID-19: an extreme immune response may lead to a
potentially fatal cytokine storm \cite{fajgenbaum2020cytokine}, but a suppressed immune response may be equally dangerous in case of (secondary) infection \cite{van2020corticosteroid,koehler2020defining}. However, because immune response is not directly observable in the clinical environment, clinicians must rely on proxies such as C-reactive protein (CRP) for inflammation \cite{mortensen2001c}; by their very nature, such proxies are noisy and highly imperfect measures of the desired values.
Secondly, LHM can provide the clinician with predictions of the disease progression given the treatments, enabling the clinicians to design the best treatment plan for the patient at hand.
Finally, LHM can bridge the gap between the laboratory and clinical environments, helping to align model output with clinical reasoning, and thus to bring models to the patient bedside and also to foster translational research \cite{shillan2019use,fleuren2020machine}.
\section{Related works}
\label{sec:related_work}
\textbf{Hybrid models}.
Hybrid models combine a given expert model with ML \cite{willard2020integrating}. Depending on the type and functionality of the expert model, various approaches have been proposed.
\textit{Residual Models} and \textit{Ensembles} use expert models that can issue predictions directly \cite{liu2019multi,wang2017physics,wu2018physics, xu2015data,yao2018tensormol}.
A Residual Model fits a ML model to the residuals of the expert model while an Ensemble averages the ML and expert predictions.
\textit{Feature Extraction} makes use of an expert model that extracts useful features from the measurements \cite{karpatne2017physics}; an ML model then uses these features to make predictions.
These methods are not suitable for our setting because our expert model is an ODE that governs the latent variables (\ref{eq:latent}); it does not issue predictions of measurements nor does it extract features that the ML model can use. Appendix \ref{app:related_work} Table \ref{tab:1} summarizes these approaches.
ML inspired by physics uses
physical laws to guide the design of architectures \cite{schutt2017schnet,zepeda2019deep,thomas2018tensor}, loss functions \cite{yazdani2020systems,fioretto2020predicting}, and weight initialization \cite{read2019process}. Examples include Hamiltonian neural networks \cite{greydanus2019hamiltonian,bertalan2019learning,zhong2019symplectic}, which reflect the conservation of energy. These models utilize general physical laws rather than a specific expert model, and are rather different than the hybrid models discussed above.
\textbf{Neural ODEs}.
Neural ODEs approximate unknown ODEs by a neural network \cite{chen2018neural}, frequently using standard feed-forward networks. ODE\textsuperscript{2}VAE uses an architecture with identity blocks to approximate second-order ODEs \cite{yildiz2019ode2vae} and GRU-ODE uses an architecture inspired by the Gated Recurrent Unit (GRU) \cite{de2019gru,cho2014learning}.
Neural ODEs and extensions have achieved state-of-the-art performance in a variety of problems involving irregularly-sampled time series data \cite{rubanova2019latent,de2019gru, kidger2020neural}. We discuss other approaches to learning unknown ODEs from data in Appendix \ref{app:related_work}.
\textbf{Mechanistic models}.
Mechanistic models are widely applied in sciences such as Physics \cite{stronge2018impact}, Epidemiology \cite{yang2020modified,gilchrist2002modeling}, and Pharmacology \cite{gesztelyi2012hill,katzung2012basic}.
These models use ODEs to describe a system's continuous-time evolution under interventions. The intervention affects the system \textit{deterministically} through the governing ODEs; e.g., Equation (\ref{eq:expert}).
This deterministic notion of intervention effect is different from the probabilistic one adopted by the statistical causal inference literature \cite{rubin2005causal} (for a detailed discussion, see \cite{scholkopf2021toward} and Appendix \ref{app:related_work}).
LHM uses two ODEs to describe the disease progression under treatment: the known expert ODE $f^e$ and the data-driven Neural ODE $f^m$.
\textbf{Latent variable models}. Latent variable models are widely used in disease progression modeling \cite{wang2014unsupervised, alaaattentive}. These models attempt to infer a set of latent variables to predict complex high-dimensional measurements.
The latent variables sometimes have high-level interpretations (e.g. cluster membership), but do not usually correspond to any well-defined and clinically meaningful physiological variable. Moreover, without informative priors, the latent variables can usually be identified only up to certain transformations (e.g. permutation of cluster labels \cite{pakman2020neural}). By contrast, LHM involves medically meaningful expert variables driven by known governing equations and following informative priors.
\vspace{-.1in}
\section{Experiment and evaluation}
Here we present the results of two experimental studies, one with simulated data and one with real data. In both experiments, we study the effect of dexamethasone treatment for COVID-19 patients. Both studies are modeled on the real-life treatment of COVID-19 patients in the ICU.
\vspace{-.1in}
\subsection{Simulation study} \label{sec:simulation} In this simulation, we use LHM to predict the results of a single dexamethasone treatment. Each patient $i$ will receive a one-time treatment; with dosage $d_i \sim \text{uniform}[0,10]$ mg and time $s_i \sim \text{uniform}[0,14]$. Our objective is to predict future measurements.
\textbf{Datasets}. We generated a variety of datasets to evaluate the model performance under different scenarios.
To evaluate how the number of clinical measurements affects performance, we generated datasets with $D = 20, 40\ \mbox{or}\ 80$ observable physiological variables ${\mathbf{x}}$.
For each dataset, we set the number of un-modeled states ${\mathbf{z}}^m$ according to the number of variables in ${\mathbf{x}}$ to be $M = D/10 = 2, 4\ \mbox{or}\ 8$ (respectively). (We made this choice to reflect the fact that a larger number of physiological variables often necessitates a larger number of un-modeled states.)
We consider a time horizon of $T = 14$ days; this is the median length of stay in hospital for Covid-19 patients \cite{rees2020covid}.
After setting $M$ and $D$, we generate the data points within a dataset independently.
We use a pharmacological model adapted from \cite{dai2021prototype} that describes five expert variables ($E=5$) under dexamethasone treatment for COVID-19 patients. We use the same model in the real-data experiment. We specify the model and the expert variables in Appendix \ref{app:pharma}.
The un-modeled states ${\mathbf{z}}^m$ are governed by the nonlinear ODE $f^m$ shown in equation (\ref{eq:latent}); the true physiological variables ${\mathbf{x}}$ are generated by the function $g$ in Equation (\ref{eq:physiological}). The specifications of $f^m$ and $g$ are provided in Appendix \ref{app:simulation}.
For each patient $i$, each of the components of its initial condition ${\mathbf{z}}_i(0)$ were independently drawn from an exponential distribution with rate $\lambda = 100$.
Measurement noises are drawn independently from $\epsilon_{it} \sim N(0, \sigma^2)$, for
$\sigma=0.2, 0.4\ \mbox{or}\ 0.8$; Equation (\ref{eq:error}).
We first simulate all daily measurements at $t = 1, 2, \ldots, T$, and then randomly remove measurements with probability $0.5$ to proxy the fact that measurements are made irregularly.
\begin{figure}[t!]
\centering
\includegraphics[width=1\textwidth]{figure/sim_plot.png}
\caption{\footnotesize \textbf{Simulation Results}. Prediction accuracy on future measurements
$\mathcal{Y}[2:T]$ given the observed history $\mathcal{Y}[0:2]$ as measured by RMSE and CRPS. The left 2 panels show the results under different training sample sizes $N_0$. The right 2 panels show the results for different numbers of un-modeled variables $M$. The shaded areas represent 95\% confidence intervals.}
\label{fig:exp}
\end{figure}
\textbf{Prediction task} For a given patient $i$, we use the measurements $\mathcal{Y}[0:t_0]$ up to some time $t_0$ and the treatment plan for that particular patient to {\em predict} the future measurements
$\mathcal{Y}[t_0:T]$. (Note that treatment may have occurred prior to time $t_0$ or may be planned following time $t_0$.) To evaluate the performance under different lengths of observed history, we use $t_0 = 2, 5\ \mbox{or} \ 10$ days. We evaluate the prediction accuracy according to Root Mean Squared Error (RMSE) and the uncertainty calibration using the Continuous Ranked Probability Score (CRPS).
\textbf{Training and Evaluation.}
We partition each dataset into a training set, a validation set, and a testing set. We consider training sets consisting of $N_0 = 10, 100, 500\ \mbox{or}\ 1000$ data points; each validation set has 100 data points and each testing set has 1000 data points.
\textbf{Benchmarks}.
We compare the performance of our method (LHM) with the performance of four other methods: latent Neural ODE (NODE), the original Pharmacology model (Expert), the residual model (Residual), and the ensemble model (Ensemble) of Expert and NODE, described below.
The details of the optimization and hyper-parameter settings are reported in Appendix \ref{app:simulation}.
{\bf NODE} involves $Z$ latent variables ${\mathbf{z}}(t) \in \mathbb{R}^Z$ whose evolutions are described by $\dot{{\mathbf{z}}}(t) = f^m({\mathbf{z}}(t), {\mathbf{a}}(t);\ \theta^m)$, where $f^m$ is a neural network with trainable weights $\theta^m$.
The number of latent variables $Z$ is a \textit{hyper-parameter} that is set to be greater than $M+E$, which is the number of \textit{true} latent variables.\footnote{We found that the performance of NODE is not sensitive to the exact choice of $Z$ so long as it is sufficiently larger than $M+E$. (This is consistent with findings reported in the literature \cite{dupont2019augmented}.) } NODE predict the physiological variables as $\hat{{\mathbf{y}}}_N(t) = g_N({\mathbf{z}}(t), {\mathbf{a}}(t);\ \gamma_N)$, where $g_N$ is a a neural network with trainable weights $\gamma_N$.
{\bf Expert} is given the true governing equation (\ref{eq:expert}), which describes the expert variables ${\mathbf{z}}^e(t)$, but no un-modeled latent variables. We use a neural network $g_E$ with trainable weights $\gamma_E$ to predict the physiological variables: $\hat{{\mathbf{y}}}_E(t) = g_E({\mathbf{z}}^e(t), {\mathbf{a}}(t);\ \gamma_E)$.
{\bf Residual} Given a trained Expert model, we calculate its residuals ${\mathbf{r}}(t) = {\mathbf{y}}(t) - \hat{{\mathbf{y}}}_E(t)$. Then a NODE is trained to predict the residuals. The final prediction is $\hat{{\mathbf{y}}}_E(t) + \hat{{\mathbf{r}}}(t)$.
{\bf Ensemble }makes prediction as $w_{1t} \hat{{\mathbf{y}}}_N(t) + w_{2t} \hat{{\mathbf{y}}}_E(t)$, where $\hat{{\mathbf{y}}}_N(t)$ and $\hat{{\mathbf{y}}}_E(t)$ are the predictions issued by NODE and Expert respectively. The ensemble weights $w_{1t}$ and $w_{2t}$ are learned on the validation set to minimize the prediction error.
\textbf{Results.} Figure \ref{fig:exp} shows the predictive performance with $t_0=2$ and $\sigma=0.2$ (results for other settings are reported in Appendix \ref{app:simulation}). LHM achieves the best overall performance. Expert does not perform well because it leaves out ${\mathbf{z}}^m(t)$. As expected, its performance significantly degrades as the number $M$ of latent variables increases. NODE is more robust to increases in $M$ because it treats the number of latent variables as a hyper-parameter. Nevertheless, NODE is less sample efficient and it achieves worse performance for sample sizes up to 1000 (left half). Both Residual and Ensemble achieve performance gains over NODE and Expert alone, but they under-perform LHM because they perform averaging directly in the output space rather than trying to infer ${\mathbf{z}}^m$ in the latent space.
\subsection{Real-data experiments}
\label{sec:real_data}
\begin{table}[t]
\caption{Prediction accuracy (RMSE) on COVID-19 intensive care data under different training sample sizes $N$. Prediction horizon $H=24$ hours. The standard deviations are shown in the brackets.}
\label{real-table}
\footnotesize
\centering
\begin{tabular}{@{}lcccc@{}}
\toprule
Method \textbackslash $N_0$ & 100 & 250 & 500 & 1000 \\ \midrule
Expert & 0.718 (0.71) & 0.704 (0.02) & 0.702 (0.02) & 0.713 (0.01) \\
Residual & 0.958 (0.63) & 1.003 (0.03) & 0.717 (0.05) & 0.635 (0.04) \\
Ensemble & 0.707 (0.60) & 0.657 (0.05) & 0.628 (0.05) & 0.599 (0.05) \\
NODE & 0.662 (0.65) & 0.659 (0.02) & 0.644 (0.05) & 0.650 (0.04) \\
ODE2VAE & 0.674 (0.62) & 0.666 (0.02) & 0.643 (0.02) & 0.619 (0.02) \\
GRU-ODE & 0.722 (0.60) & 0.673 (0.05) & 0.623 (0.05) & 0.601 (0.05) \\
Time LSTM & 0.706 (0.63) & 0.649 (0.03) & 0.600 (0.03) & 0.631 (0.02) \\
LHM & \textbf{0.633 (0.51)} & \textbf{0.605 (0.02)} & \textbf{0.529 (0.02)} & \textbf{0.511 (0.02)} \\ \bottomrule
\end{tabular}
\end{table}
In this experiment, we use real data to evaluate the predictive performance of LHM and to illustrate its utility for decision support in a realistic clinical setting that closely tracks the actual treatment of COVID-19 patients in ICU.
\textbf{Dataset}. We used data from the Dutch Data Warehouse (DDW), a multicenter and full-admission anonymized electronic health records database of critically ill COVID-19 patients \cite{fleuren2021large}.
Up until March 2021, DDW has collected the health trajectories for 3464 patients admitted to intensive care units (ICU) across the Netherlands.
However, even if we use the entire DDW for training, the sample size is still relatively small compared to what is typically used by ML (tens or hundreds of thousands of samples \cite{johnson2016mimic}).
Furthermore, patients are even scarcer at the early stage of pandemic, arguably when a decision support tool is most needed: only 607 patients were admitted at the first peak (by April 2020).
After applying the eligibility criterion detailed in Appendix \ref{app:real_data}, we obtained a dataset of \textbf{2097} patients whose disease progression is characterized by an irregularly-sampled time series of \textbf{27} physiological variables ({Appendix} \ref{app:real_data}).
These variables capture the vital signals, respiratory mechanics, and biomarkers that are crucial for clinical decisions. In addition, we also included \textbf{11} static variables that are known to affect the progression of COVID-19, e.g. BMI ({Appendix} \ref{app:real_data}).
\textbf{Prediction task}. Denote $t_0$ as the time when the patient received the first dose of dexamethasone (we set $t_0=24$ for untreated patients). We use the history up to 24 hours before $t_0$, $\mathcal{Y}[t_0-24:t_0]$, to predict the future $\mathcal{Y}[t_0:t_0+24H]$ over a time horizon $H = 1, 3\ \mbox{or}\ 7$ days. We use $N_0 = 100, 250, 500\ \mbox{or}\ 1000$ patients for training, {97} for validation, and {1000} for testing. The pharmacological model and the prior distribution of expert variables are detailed in Appendix \ref{app:pharma}.
\textbf{Benchmarks}. In addition to all the benchmarks introduced in Section \ref{sec:simulation}, we compared the results with two extensions of NODE, GRU-ODE and ODE\textsuperscript{2}VAE, which achieved strong performance in medical time series prediction \cite{yildiz2019ode2vae, de2019gru}. We also used the Time LSTM as a strong baseline \cite{baytas2017patient}.
\textbf{Results}. The main results are shown in Table \ref{real-table} (additional results are shown in Appendix \ref{app:real_data}). LHM consistently outperformed the benchmarks.
Its performance with $N_0=100$ samples is close to the pure ML approaches' performance with $N_0=500$ samples.
As the sample size increases from $100$ to $1000$, the predictive accuracy of LHM improves by $19\%$ while Time LSTM improves by $11\%$ and NODE by less than $5\%$.
A larger improvement rate suggests LHM adapts to the newly available data faster, which is important when the samples are scarce.
As expected, the standalone expert model achieved poor performance because it is unable to capture the full array of clinical measurements.
\begin{figure}[t!]
\centering
\includegraphics[width=0.99\textwidth,trim={0.48cm 0 0 0},clip]{figure/patients.pdf}
\caption{\footnotesize \textbf{The observed measurements and the inferred expert variables for three illustrative patients}. Left: The observed physiological variable $x_1(t)$: C-Reactive Protein. Middle: The inferred expert variable $z^e_1(t)$: the immune response to viral infection. Right: The inferred expert variable $z^e_2(t)$: dexamethasone concentration. Vertical dotted lines mark the times of dexamethasone injections.}
\label{fig:real_interp}
\end{figure}
\textbf{LHM in action}. Here we show how LHM can support clinical decisions beyond predicting future clinical measurements.
Managing the level of immune response is pivotal when deciding on immunosuppresive therapy for COVID-19 patients \cite{fajgenbaum2020cytokine,van2020corticosteroid,koehler2020defining}.
This is a challenging task because the ``right'' level of immune response varies across patients \cite{keller2020effect}: for most patients, we would like to reduce the immune response to avoid cytokine storm and consequent organ failure \cite{fajgenbaum2020cytokine}, but for patients with other infections (e.g., a secondary bacterial or fungal infection), we would like to keep their immune systems activated \cite{koehler2020defining,cox2020co}.
Because immune response is not directly observable in the clinical settings, clinicians resort to unspecific inflammatory markers such as C-Reactive Protein (CRP) \cite{mortensen2001c}. However, better markers of the immune response such as the cytokine Type I IFNs can be measured in the laboratory setting and have been included in the pharmacological model as an expert variable $z^e_1$.
Moreover, the immune response is affected by dexamethasone concentrations in the lung tissue (the expert variable $z^e_2$). These concentrations are not easily or routinely measured in a clinical setting, and are therefore not available to clinicians.
Figure \ref{fig:real_interp} shows the \textit{measurements} of CRP ${\mathbf{x}}_1$, and the \textit{inferred} immune response $z^e_1$ and dexamethasone concentration $z^e_2$. The two expert variables are inferred by a trained LHM using the first five days of observations. We selected three representative patients based on the treatment regimen they received.
Patient A is representative of the 59.8\% of the population who did not receive dexamethasone; Patient C is representative of the 12.2\% who received dexamethasone according to the guidelines \cite{guideline}, and Patient B is representative of the 28.0\% of the population who received dexamethasone but whose treatment was not according to the guidelines.
For patient A, the initial level of CRP was moderately high, but then it rose and peaked at about 100 hours after admission. In the absence of contraindications, a clinician might begin dexamethasone treatment at this point, but LHM predicts that immune activity $z^e_1$ would decrease afterwards even without treatment. (The right panel is blank because dexamethasone was never administered.)
Patient B was admitted to the ICU with a very high level of the inflammatory marker CRP. Two doses of dexamethasone were given in rapid succession, preceding a decline in both CRP $x_1$ and immune activity $z^e_1$. However, after the dexamethasone depleted in the patient's body, the expert model predicts that the immune response will pick up again. This is reflected in the re-occurrence of the high CRP level.
Patient C was admitted to the ICU with a moderately high level of CRP. LHM also inferred a high level of immune activity $z^e_1$ at the initial stage. Inflammation was greatly reduced after dexamethasone treatments, which has an immunosuppressive effect; so $x_1$ and $z^e_1$ display the same downward trend. Because dexamethasone concentrations falls rapidly within 24 hours after treatment, it was repeated at 24 hour intervals, as seen in the right panel. This pattern is clinically expected under the treatment regimen.
\section{Discussion and future work}
\label{sec:discussion}
This paper has focused on a single disease (COVID-19), a single treatment (dexamethasone), and a single expert model. The ultimate goal is to build a model that encompasses a variety of diseases, a variety of treatments and multiple expert models. This is a challenge for future work.
\begin{ack}
This work was supported by the Office of Naval Research (ONR) and the NSF (Grant number:1722516).
We would like to thank the following individuals for providing domain expertise in Pharmacology and designing the expert ODEs.
Bernhard Steiert, PhD, Pharmaceutical Sciences, Roche Pharma Research and Early Development (pRED), Roche Innovation Center Basel, Basel, Switzerland
Richard Peck, MA, MB, BChir, FFPM, FRCP, Pharmaceutical Sciences, Roche Pharma Research and Early Development (pRED), Roche Innovation Center Basel, Basel, Switzerland
We would like to thank the following individuals for curating and sharing the Dutch Data Warehouse (DDW) for critically ill COVID-19 patients.
\textit{From collaborating hospitals having shared data:}
Diederik Gommers, MD, PhD, Department of Intensive Care, Erasmus Medical Center, Rotterdam, The Netherlands,
Olaf L. Cremer, MD, PhD, Intensive Care, UMC Utrecht, Utrecht, The Netherlands,
Rob J. Bosman, MD, ICU, OLVG, Amsterdam, The Netherlands,
Sander Rigter, MD, Department of Anesthesiology and Intensive Care, St. Antonius Hospital, Nieuwegein, The Netherlands,
Evert-Jan Wils, MD, PhD, Department of Intensive Care, Franciscus Gasthuis \& Vlietland, Rotterdam, The Netherlands,
Tim Frenzel, MD, PhD, Department of Intensive Care Medicine, Radboud University Medical Center, Nijmegen, The Netherlands,
Dave A. Dongelmans, MD, PhD, Department of Intensive Care Medicine, Amsterdam UMC, Amsterdam, The Netherlands,
Remko de Jong, MD, Intensive Care, Bovenij Ziekenhuis, Amsterdam, The Netherlands,
Marco Peters, MD, Intensive Care, Canisius Wilhelmina Ziekenhuis, Nijmegen, The Netherlands,
Marlijn J.A Kamps, MD, Intensive Care, Catharina Ziekenhuis Eindhoven, Eindhoven, The Netherlands,
Dharmanand Ramnarain, MD, Department of Intensive Care, ETZ Tilburg, Tilburg, The Netherlands,
Ralph Nowitzky, MD, Intensive Care, HagaZiekenhuis, Den Haag, The Netherlands,
Fleur G.C.A. Nooteboom, MD, Intensive Care, Laurentius Ziekenhuis, Roermond, The Netherlands,
Wouter de Ruijter, MD, PhD, Department of Intensive Care Medicine, Northwest Clinics, Alkmaar, The Netherlands,
Louise C. Urlings-Strop, MD, PhD, Intensive Care, Reinier de Graaf Gasthuis, Delft, The Netherlands,
Ellen G.M. Smit, MD, Intensive Care, Spaarne Gasthuis, Haarlem en Hoofddorp, The Netherlands,
D. Jannet Mehagnoul-Schipper, MD, PhD, Intensive Care, VieCuri Medisch Centrum, Venlo, The Netherlands,
Julia Koeter, MD, Intensive Care, Canisius Wilhelmina Ziekenhuis, Nijmegen, The Netherlands,
Tom Dormans, MD, PhD, Intensive care, Zuyderland MC, Heerlen, The Netherlands,
Cornelis P.C. de Jager, MD, PhD, Department of Intensive Care, Jeroen Bosch Ziekenhuis, Den Bosch, The Netherlands,
Stefaan H.A. Hendriks, MD, Intensive Care, Albert Schweitzerziekenhuis, Dordrecht, The Netherlands,
Sefanja Achterberg, MD, PhD, ICU, Haaglanden Medisch Centrum, Den Haag, The Netherlands,
Evelien Oostdijk, MD, PhD, ICU, Maasstad Ziekenhuis Rotterdam, Rotterdam, The Netherlands,
Auke C. Reidinga, MD, ICU, SEH, BWC, Martiniziekenhuis, Groningen, The Netherlands,
Barbara Festen-Spanjer, MD, Intensive Care, Ziekenhuis Gelderse Vallei, Ede, The Netherlands,
Gert B. Brunnekreef, MD, Department of Intensive Care, Ziekenhuisgroep Twente, Almelo, The Netherlands,
Alexander D. Cornet, MD, PhD, FRCP, Department of Intensive Care, Medisch Spectrum Twente, Enschede, The Netherlands,
Walter van den Tempel, MD, Department of Intensive Care, Ikazia Ziekenhuis Rotterdam, Rotterdam, The Netherlands,
Age D. Boelens, MD, Anesthesiology, Antonius Ziekenhuis Sneek, Sneek, The Netherlands,
Peter Koetsier, MD, Intensive Care, Medisch Centrum Leeuwarden, Leeuwarden, The Netherlands,
Judith Lens, MD, ICU, IJsselland Ziekenhuis, Capelle aan den IJssel, The Netherlands,
Roger van Rietschote, Business Intelligence, Haaglanden MC, Den Haag,The Netherlands,
Harald J. Faber, MD, ICU, WZA, Assen, The Netherlands,
A. Karakus, MD, Department of Intensive Care, Diakonessenhuis Hospital, Utrecht, The Netherlands,
Robert Entjes, MD, Department of Intensive Care, Admiraal De Ruyter Ziekenhuis, Goes, The Netherlands,
Paul de Jong, MD, Department of Anesthesia and Intensive Care, Slingeland Ziekenhuis, Doetinchem, The Netherlands,
Thijs C.D. Rettig, MD, PhD, Department of Intensive Care, Amphia Ziekenhuis, Breda, The Netherlands,
M.C. Reuland, MD, Department of Intensive Care Medicine, Amsterdam UMC, Universiteit van Amsterdam, Amsterdam, The Netherlands,
Laura van Manen, MD, Department of Intensive Care, BovenIJ Ziekenhuis, Amsterdam, The Netherlands,
Leon Montenij, MD, PhD, Department of Anesthesiology, Pain Management and Intensive Care, Catharina Ziekenhuis Eindhoven, Eindhoven, The Netherlands,
Jasper van Bommel, MD, PhD, Department of Intensive Care, Erasmus Medical Center, Rotterdam, The Netherlands,
Roy van den Berg, Department of Intensive Care, ETZ Tilburg, Tilburg, The Netherlands,
Ellen van Geest, Department of ICMT, Haga Ziekenhuis, Den Haag, The Netherlands,
Anisa Hana, MD, PhD, Intensive Care, Laurentius Ziekenhuis, Roermond, The Netherlands,
B. van den Bogaard, MD, PhD, ICU, OLVG, Amsterdam, The Netherlands,
Prof. Peter Pickkers, Department of Intensive Care Medicine, Radboud University Medical Centre, Nijmegen, The Netherlands,
Pim van der Heiden, MD, PhD, Intensive Care, Reinier de Graaf Gasthuis, Delft, The Netherlands,
Claudia (C.W.) van Gemeren, MD, Intensive Care, Spaarne Gasthuis, Haarlem en Hoofddorp, The Netherlands,
Arend Jan Meinders, MD, Department of Internal Medicine and Intensive Care, St Antonius Hospital, Nieuwegein, The Netherlands,
Martha de Bruin, MD, Department of Intensive Care, Franciscus Gasthuis \& Vlietland, Rotterdam, The Netherlands,
Emma Rademaker, MD, MSc, Department of Intensive Care, UMC Utrecht, Utrecht, The Netherlands,
Frits H.M. van Osch, PhD, Department of Clinical Epidemiology, VieCuri Medisch Centrum, Venlo, The Netherlands,
Martijn de Kruif, MD, PhD, Department of Pulmonology, Zuyderland MC, Heerlen, The Netherlands,
Nicolas Schroten, MD, Intensive Care, Albert Schweitzerziekenhuis, Dordrecht, The Netherlands,
Klaas Sierk Arnold, MD, Anesthesiology, Antonius Ziekenhuis Sneek, Sneek, The Netherlands,
J.W. Fijen, MD, PhD, Department of Intensive Care, Diakonessenhuis Hospital, Utrecht, The Netherland,
Jacomar J.M. van Koesveld, MD, ICU, IJsselland Ziekenhuis, Capelle aan den IJssel, The Netherlands,
Koen S. Simons, MD, PhD, Department of Intensive Care, Jeroen Bosch Ziekenhuis, Den Bosch, The Netherlands,
Joost Labout, MD, PhD, ICU, Maasstad Ziekenhuis Rotterdam, The Netherlands,
Bart van de Gaauw, MD, Martiniziekenhuis, Groningen, The Netherlands,
Michael Kuiper, Intensive Care, Medisch Centrum Leeuwarden, Leeuwarden, The Netherlands,
Albertus Beishuizen, MD, PhD, Department of Intensive Care, Medisch Spectrum Twente, Enschede, The Netherlands,
Dennis Geutjes, Department of Information Technology, Slingeland Ziekenhuis, Doetinchem, The Netherlands,
Johan Lutisan, MD, ICU, WZA, Assen, The Netherlands,
Bart P. Grady, MD, PhD, Department of Intensive Care, Ziekenhuisgroep Twente, Almelo, The Netherlands,
Remko van den Akker, Intensive Care, Adrz, Goes, The Netherlands,
Sesmu Arbous, MD, PhD, Intensivist, LUMC, Leiden, The Netherlands,
Tom A. Rijpstra, MD, Department of Intensive Care, Amphia Ziekenhuis, Breda, The Netherlands,
Roos Renckens, MD, PhD, Department of Internal Medicine, Northwest Clinics, Alkmaar, the Netherlands,
\textit{From collaborating hospitals having signed the data sharing agreement:}
Daniël Pretorius, MD, Department of Intensive Care Medicine, Hospital St Jansdal, Harderwijk, The Netherlands,
Menno Beukema, MD, Department of Intensive Care, Streekziekenhuis Koningin Beatrix, Winterswijk, The Netherlands,
Bram Simons, MD, Intensive Care, Bravis Ziekenhuis, Bergen op Zoom en Roosendaal, The Netherlands,
A.A. Rijkeboer, MD, ICU, Flevoziekenhuis, Almere, The Netherlands,
Marcel Aries, MD, PhD, MUMC+, University Maastricht, Maastricht, The Netherlands,
Niels C. Gritters van den Oever, MD, Intensive Care, Treant Zorggroep, Emmen, The Netherlands,
Martijn van Tellingen, MD, EDIC, Department of Intensive Care Medicine, afdeling Intensive Care, ziekenhuis Tjongerschans, Heerenveen, The Netherlands,
Annemieke Dijkstra, MD, Department of Intensive Care Medicine, Het Van Weel-Bethesda Ziekenhuis, Dirksland, The Netherlands,
Rutger van Raalte, Department of Intensive Care, Tergooi hospital, Hilversum, The Netherlands,
\textit{From the Laboratory for Critical Care Computational Intelligence:
}
Tariq A. Dam, MD, Department of Intensive Care Medicine, Laboratory for Critical Care Computational Intelligence, Amsterdam Medical Data Science, Amsterdam UMC, Vrije Universiteit, Amsterdam, The Netherlands,
Martin E. Haan, MD, Department of Intensive Care Medicine, Laboratory for Critical Care Computational Intelligence, Amsterdam Medical Data Science, Amsterdam UMC, Vrije Universiteit, Amsterdam, The Netherlands,
Mark Hoogendoorn, PhD, Quantitative Data Analytics Group, Department of Computer Science, Faculty of Science, VU University, Amsterdam, The Netherlands,
Armand R.J. Girbes, MD, PhD, EDIC, Department of Intensive Care Medicine, Laboratory for Critical Care Computational Intelligence, Amsterdam Medical Data Science, Amsterdam UMC, Vrije Universiteit, Amsterdam, The Netherlands,
Paul W.G. Elbers, MD, PhD, EDIC, Department of Intensive Care Medicine, Laboratory for Critical Care Computational Intelligence, Amsterdam Medical Data Science, Amsterdam UMC, Vrije Universiteit, Amsterdam, The Netherlands,
Patrick J. Thoral, MD, EDIC, Department of Intensive Care Medicine, Laboratory for Critical Care Computational Intelligence, Amsterdam Medical Data Science, Amsterdam UMC, Vrije Universiteit, Amsterdam, The Netherlands,
Dagmar M. Ouweneel, PhD, Department of Intensive Care Medicine, Laboratory for Critical Care Computational Intelligence, Amsterdam Medical Data Science, Amsterdam UMC, Vrije Universiteit, Amsterdam, The Netherlands,
Ronald Driessen, Department of Intensive Care Medicine, Laboratory for Critical Care Computational Intelligence, Amsterdam Medical Data Science, Amsterdam UMC, Vrije Universiteit, Amsterdam, The Netherlands,
Jan Peppink, Department of Intensive Care Medicine, Laboratory for Critical Care Computational Intelligence, Amsterdam Medical Data Science, Amsterdam UMC, Vrije Universiteit, Amsterdam, The Netherlands,
H.J. de Grooth, MD, PhD, Department of Intensive Care Medicine, Laboratory for Critical Care Computational Intelligence, Amsterdam Medical Data Science, Amsterdam UMC, Vrije Universiteit, Amsterdam, The Netherlands,
\textit{From Pacmed:}
Robbert C.A. Lalisang, MD, Pacmed, Amsterdam, The Netherlands,
Michele Tonutti, MRes, Pacmed, Amsterdam, The Netherlands,
Daan P. de Bruin, MSc, Pacmed, Amsterdam, The Netherlands,
Sebastiaan J.J. Vonk, MSc, Pacmed, Amsterdam, The Netherlands,
Mattia Fornasa, PhD, Pacmed, Amsterdam, The Netherlands,
Tomas Machado, Pacmed, Amsterdam, The Netherlands,
Michael de Neree tot Babberich, Pacmed, Amsterdam, The Netherlands,
Olivier Thijssens, MSc, Pacmed, Amsterdam, The Netherlands,
Lot Wagemakers, Pacmed, Amsterdam, The Netherlands,
Hilde G.A. van der Pol, Pacmed, Amsterdam, The Netherlands,
Tom Hendriks, Pacmed, Amsterdam, The Netherlands,
Julie Berend, Pacmed, Amsterdam, The Netherlands,
Virginia Ceni Silva, Pacmed, Amsterdam, The Netherlands,
Robert F.J. Kullberg, MD, Pacmed, Amsterdam, The Netherlands,
Taco Houwert, MSc, Pacmed, Amsterdam, The Netherlands,
Hidde Hovenkamp, MSc, Pacmed, Amsterdam, The Netherlands,
Roberto Noorduijn Londono, MSc, Pacmed, Amsterdam, The Netherlands,
Davide Quintarelli, MSc, Pacmed, Amsterdam, The Netherlands,
Martijn G. Scholtemeijer, MD, Pacmed, Amsterdam, The Netherlands,
Aletta A. de Beer, MSc, Pacmed, Amsterdam, The Netherlands,
Giovanni Cina, PhD, Pacmed, Amsterdam, The Netherlands,
Willem E. Herter, BSc, Pacmed, Amsterdam, The Netherlands,
Adam Izdebski, Pacmed, Amsterdam, The Netherlands,
\textit{From RCCnet:}
Leo Heunks, MD, PhD, Department of Intensive Care Medicine, Amsterdam Medical Data Science, Amsterdam UMC, Vrije Universiteit, Amsterdam, The Netherlands,
Nicole Juffermans, MD, PhD, ICU, OLVG, Amsterdam, The Netherlands,
Arjen J.C. Slooter, MD, PhD, Department of Intensive Care Medicine, UMC Utrecht, Utrecht University, Utrecht, the Netherlands,
\textit{From other collaborating partners:}
Martijn Beudel, MD, PhD, Department of Neurology, Amsterdam UMC, Universiteit van Amsterdam, Amsterdam, The Netherlands,
\end{ack}
\clearpage
|
\section{Introduction}
In recent years, advances in machine learning and improvements in computer performance have made it possible to make highly accurate predictions for a variety of tasks. However, there is generally a trade-off between prediction accuracy and interpretability, and methods that have high prediction accuracy such as neural networks and gradient boosting decision trees (GBDTs) tend to have low interpretability. However, as stated in the EU’s “General Data Protection Regulation” \cite{EU} and Japan’s “Draft Guidelines for AI Development” \cite{japan}, the interpretability and accountability of AI are becoming more important. In particular, in areas where the results of machine learning inferences can have serious consequences, such as finance, medicine, and autonomous driving, both interpretability and prediction accuracy are even more important. Thus, prediction models are needed that achieve both high prediction accuracy and interpretability.
There are two main directions for interpretability. The first is local interpretation of inference results. In this approach, the inference results of a machine learning models are interpreted individually, such as LIME \cite{LIME} and SHAP \cite{SHAP}. The second is global interpretation of the model, in which the model itself is structured such that it can be interpreted globally. In this study, we build a model for global interpretability from the viewpoint of generality.
The outline of this paper is as follows. In Section II, we present an overview of related works. In Section III, we discuss the problems of the Generalized Additive 2 Model (GA\textsuperscript{2}M), and explain the proposed method, Constrained Generalized Additive 2 Model plus (CGA\textsuperscript{2}M+), which addresses the problems of GA\textsuperscript{2}M. In Section IV, we describe numerical experiments conducted to verify the validity of the improvements and compare the prediction accuracy of CGA\textsuperscript{2}M+ and existing models. In addition, we introduce some examples of applications that take advantage of the interpretability of CGA\textsuperscript{2}M+. Finally, our conclusions and a summary are given in Section V.
\section{Related Work}
In this section, we divide existing machine learning models into the following three categories:
\begin{quote}
\begin{itemize}
\item Models that focus on prediction accuracy
\item Models that focus on interpretability
\item Balanced models
\end{itemize}
\end{quote}
First, we introduce support vector machine (SVM) \cite{SVM}, random forests \cite{RF}, the GBDT \cite{GBDT}, and neural networks as models that focus on prediction accuracy. An SVM is a model based on margin maximization. In combination with a kernel method, an SVM can capture high-dimensional nonlinearities and thus generally has high prediction accuracy. A random forest is an ensemble method in which multiple decision trees are trained in parallel using bootstrap sampling, and it generally has higher generalization performance than a normal decision tree. A GBDT is a model for training multiple decision trees sequentially and finally construct an ensemble of them. In particular, LightGBM \cite{LGBM}, which has been designed to reduce time complexity and space complexity, is widely used in various tasks. Neural networks are at the centre of the recent rise of AI and machine learning and have been used successfully in various fields such as image recognition and natural language processing. However, these models are capable of only local interpretation, such as feature importance, and do not have global interpretability.
Second, we introduce models that focus on interpretability. Linear regression is a classical model with high interpretability. Generalized linear models \cite{GLM} express the objective variable as a nonlinear transformation of a linear combination of features, which enables more flexible modelling. Generalized additive models \cite{GAM} are models that relax restrictions of the linear combination of generalized linear models, and the influence of each feature can be expressed as a nonlinear function. These models are highly interpretable, as the influence of each feature can be understood visually and intuitively. However, the prediction accuracy of the models is generally not very high, and they tend to be inaccurate for data having inherent higher-order interactions.
Finally, we introduce GA\textsuperscript{2}M, a model that achieves both high prediction accuracy and interpretability \cite{GA2M}. GA\textsuperscript{2}M is a generalized additive model with an additional interaction term. It is modelled as
\begin{equation}
y = \sum_{i \in Z^1} f_i(x_i)+\sum_{(ij)\in Z^2} f_{ij}(x_i,x_j)
\end{equation}
where $f_i(x_i)$ and $f_{ij}(x_i,x_j)$ are shape functions. $Z^1$ is an index set of all features, and we define $Z^2$ as $Z^2 := Z^{1} \times Z^{1}$. In this study, we used LightGBM as shape functions. Note that $i>j$ in the subscripts of the pairwise interaction terms. The pairwise interaction terms make it possible to model the interaction between the two features, and GA\textsuperscript{2}M is highly interpretable because the influence of each feature on the objective variable and the influence of the interaction between the two features can be visualized in a graph. In addition, the prediction accuracy is also high because nonlinear functions are used for each shape function and pairwise interactions are considered, rather than the effects of the features alone. In this study, we developed a model based on the idea of GA\textsuperscript{2}M that combines higher levels of prediction accuracy and interpretability. Another approach to an interpretation is to interpret the prediction results. For example, we can use LIME \cite{LIME} or SHAP \cite{SHAP} to interpret prediction results of a model like GBDT. However, these approach only provides local interpretations of the prediction results and cannot provide a global interpretation of the whole model. It should be noted that the purpose of this study is to interpret the whole model, which is fundamentally different from interpreting the prediction results by using LIME or SHAP.
\section{Proposed Method}
\subsection{Problems with GA2M}
GA\textsuperscript{2}M suffers from two major problems. The first one is the interpretability of the shape function. GA\textsuperscript{2}M can use any function as the shape function, and analysts can observe each shape function to infer how the individual feature affects the objective variable. However, the shape function may show illogical results due to noise in the data or lack of data. For example, consider the relationship between the housing prices and the number of rooms of it. In general, as the number of rooms increases, the real estate price rises. However, if we simply apply GA\textsuperscript{2}M to the data, due to the aforementioned causes, the result may be that the housing prices decreases even though the number of rooms increases, as shown in Fig. \ref{fig1}. This is not logically correct but also does not provide useful suggestions to analysts. Therefore, we apply monotonic constraints to some shape functions of a model to prevent it from producing illogical results. Returning to the previous example, we force the model to learn the relationship that housing prices increase as the number of rooms increases. We explain the details of this method in Section \ref{sec:propose}. The second problem is that GA\textsuperscript{2}M can not capture higher-order interactions. The shape functions in GA\textsuperscript{2}M include univariate terms and pairwise interaction terms but cannot consider more higher-order interactions, resulting in lower accuracy than models that can account for higher-order interactions. For example, higher-order interactions such as $x_1x_2x_3x_4$ difficult to approximate with a GA\textsuperscript{2}M. Therefore, we improved GA\textsuperscript{2}M by considering higher-order interactions, which do not have interpretability. Further, we devised a learning method that does not interpretability of the whole model.
\begin{figure}[b]
\centerline{\includegraphics[width=7cm]{image/limp.pdf}}
\caption{Example of the need to introduce monotonicity}
\label{fig1}
\end{figure}
\subsection{Constrained Generalized Additive 2 Model plus}\label{sec:propose}
We propose CGA\textsuperscript{2}M+, which solves the two abovementioned problems of GA\textsuperscript{2}M. CGA\textsuperscript{2}M+ is modelled as follows:
\begin{equation}
\label{eq:cga2m}
\begin{split}
y =& \sum_{i \in Z_c}f_i(x_i)+ \sum_{i \in Z_u}f_i(x_i)+\sum_{(ij) \in Z_{cc}} f_{ij}(x_i,x_j) \\& +\sum_{(ij) \in Z_{cu}}f_{ij}(x_i,x_j)+\sum_{(ij) \in Z_{uu}}f_{ij}(x_i,x_j) \\
&+f_{\mathrm{high}}(x_1,x_2,\dots,x_K)
\end{split}
\end{equation}
The symbols in Eq (\ref{eq:cga2m}) are as follows.
\begin{itemize}
\item $Z_c$ : an index set of features which are enforced monotonicity constraints on.
\item $Z_u$ : an index set of features which are not enforced monotonicity constraints on.
\item $Z_{cc}$ : $Z_{cc} = Z_c \times Z_c$
\item $Z_{cu}$ : $Z_{cu} = Z_c \times Z_u$
\item $Z_{uu}$ : $Z_{uu} = Z_u \times Z_u$
\end{itemize}
During training CGA\textsuperscript{2}M+, feature pairs are selected from a set $Z^2(=Z_{cc}\cup Z_{cu}\cup Z_{uu})$.
CGA\textsuperscript{2}M+ solves the problem of interpretability of shape functions in GA\textsuperscript{2}M. We train the model by enforcing monotonicity in the corresponding shape functions based on human knowledge, such as the fact that housing prices increase as the number of rooms increases. We chose LightGBM as the shape function because of its fast training speed and high accuracy. we can impose monotonicity constraints on any shape functions if necessary. The monotonicity constraint is implemented by imposing a constraint on decision tree branching. Specifically, we penalize branches that violate the monotonicity constraint. In addition, the branching is performed in compliance with the monotonicity constraint by transferring the constraint of the parent node to the child node. For details, refer to [12].
CGA\textsuperscript{2}M+ also solves the problem of GA\textsuperscript{2}M’s inability to capture higher-order interactions. To capture higher-order interactions without losing interpretability, we added the function $f_{\mathrm{high}} (x_1,x_2,...x_K)$ (called the higher-order term) to GA\textsuperscript{2}M. This function is not interpretable but can capture higher-order interactions, and the proposed learning method does not reduce interpretability of the whole model.
\begin{figure}[t]
\centerline{\includegraphics[width=\linewidth]{image/algo.pdf}}
\end{figure}
The training of the proposed CGA\textsuperscript{2}M+ is performed as in Algorithm 1. There are two points to note in the algorithm. First, the pair of features in $Z^2$ are not necessarily used for training. $Z^2$ is just a set of feature pairs that can be selected, and we select the pairs from this set sequentially. Therefore, unimportant feature pairs are deleted from the model in line 14. The next point is that $f_{\mathrm{high}}$ does not affect the learning of $f_i$ and $f_{ij}$. As can be seen in line 20, $f_i$ and $f_{ij}$ have already been learned when learning $f_{\mathrm{high}}$, and therefore they are unaffected. In other words, the interpretability of $f_i$ and $f_{ij}$ is ensured by this learning mechanism.
Next, we define importance of shape functions and the higher-order term. In the training of CGA\textsuperscript{2}M+, lines 14 and 18 of Algorithm 1 determine whether the shape function is effective, and the method is explained below. Assume that $F$ is a trained CGA\textsuperscript{2}M+:
\begin{equation}
F=\sum_{i} f_{i}\left(x_{i}\right)+\sum_{i<j} f_{i j}\left(x_{i}, x_{j}\right)+f_{\mathrm{high}}\left(x_{1}, x_{2}, \ldots x_{K}\right)
\end{equation}
Let $x_{in}$ be the value of the $i$th feature $x_i$ in the $n$th $(n \in \{1,2,...,N\})$ training sample. Also, let the value of the objective variable in the $n$th training data be $y_n$. The importance of $f_i$, $f_{ij}$, and $f_{\mathrm{high}}$ is defined as
\begin{align}
&\mathrm{importance ~ of}~ f_i=\frac{\mathrm{effect}_i}{\mathrm{effect_{all}}}\\
&\mathrm{importance ~ of}~ f_{ij}=\frac{\mathrm{effect}_{ij}}{\mathrm{effect_{all}}}\\
&\mathrm{importance ~ of}~ f_{\mathrm{high}}=\frac{\mathrm{effect_{high}}}{\mathrm{effect_{all}}}
\end{align}
where the respective definitions of $\mathrm{effect}_i$,$\mathrm{effect}_{ij}$, $\mathrm{effect_{high}}$ and $\mathrm{effect_{all}}$ are
\begin{align}
& \mathrm {effect}_{i}=\frac{\sum_{n=1}^{N}\left|f_{i}\left(x_{in}\right)-\bar{f}_{i}\right|}{\sum_{n=1}^{N}\left|y_{n}-\bar{y}\right|}\\
& \mathrm {effect}_{ij}=\frac{\sum_{n=1}^{N}\left|f_{ij}\left(x_{in}, x_{jn}\right)-\bar{f}_{ij}\right|}{\sum_{n=1}^{N}\left|y_{n}-\bar{y}\right|,}\\
& \mathrm{effect}_{\mathrm{high}}=\frac{\sum_{n=1}^{N}\left|f_{\mathrm{high}}\left(x_{1n}, x_{2n}, \ldots x_{Kn}\right)-\bar{f}_\text{high}\right|}{\sum_{n=1}^{N}\left|y_{n}-\bar{y}\right|,}\\
&\mathrm{effect}_{\mathrm{all}}=\sum_{i} \mathrm{ effect}_{i}+\sum_{ij} \mathrm{effect}_{ij}+\mathrm{effect}_{\mathrm{high}}
\end{align}
In (7) to (10), we use the following symbols.
\begin{align}
&\bar{f}_i=\frac{1}{N}\sum_{n=1}^{N}f_i(x_{in})\\
&\bar{f}_{ij}=\frac{1}{N}\sum_{n=1}^{N}f_{ij}(x_{in},x_{jn})\\
&\bar{f}_{\mathrm{high}}=\frac{1}{N}\sum_{n=1}^{N}f_{\mathrm{high}}(x_{n1},x_{n2},\dots,x_{nK})\\
&\bar{y}=\frac{1}{N}\sum_{n=1}^{N}y_{n}
\end{align}
In line 18 of Algorithm 1, shape functions with importance below a given threshold are deleted. As described in Section \ref{sec:propose}, the training method in Algorithm 1 controls the importance of the higher-order term, which is not interpretable. Therefore, the accuracy of the model can be improved without reducing interpretability of the whole model. In this study, the threshold is set to 0.01.
\section{Numerical Experiment}
We next introduce the datasets used in the experiments, evaluate the proposed monotonicity and higher-order term, compare the proposed method with existing methods, and finally outline some application examples.
\subsection{Datasets}
We conducted experiments on the two datasets described in Table \ref{tab:dataset}.
\begin{table}[b]
\caption{Dataset}
\label{tab:dataset}
\centering
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Data}&\textbf{Sample Size}&\textbf{Features}&\textbf{Monotonicity}\\
\hline
California housing prices&20,640&9&3\\
\hline
Oricon cafe satisfaction&6,537&37&31\\
\hline
\end{tabular}
\end{table}
``California housing prices'' is a dataset obtained from Kaggle \cite{california}, and it contains data about housing prices in various regions of California. ``Oricon cafe satisfaction'' is a dataset from a customer satisfaction survey about cafes conducted by Oricon Inc. We used the Oricon customer satisfaction survey data provided in the 2020 Data Analysis Competition sponsored by the Joint Association Study Group of Management Science \cite{compe}. In the California housing prices dataset, the task is to predict the housing prices based on features such as annual income, age and so on. In the Oricon cafe satisfaction dataset, the task is to predict how satisfied a customer is with a particular cafe based on cafe features and customer features. In Table \ref{tab:dataset}, \textbf{Features} is the number of features and \textbf{Monotonicity} is the number of features for which monotonicity is introduced when CGA\textsuperscript{2}M+ is applied. The ratio of training, validation, and test data was set to 60\%, 20\%, and 20\%, for each dataset.
\begin{table}[b]
\centering
\caption{Evaluation of Monotonicity and Higher-Order Term}
\label{tab:eval}
\begin{tabular}{|l|c|c|c|c|}
\hline
{} & \multicolumn{2}{|c|}{\bf California} & \multicolumn{2}{|c|}{\bf Oricon}\\
\hline
\bf Model & \bf \textit{training} & \bf \textit{test} & \bf \textit{training} & \bf \textit{test}\\
\hline
\begin{tabular}{l}GA\textsuperscript{2}M \end{tabular}& 31,263.74 & 50,488.58 & 0.953 & 1.167\\ \hline
\begin{tabular}{l}GA\textsuperscript{2}M\\+higher \end{tabular}& 27,587.23 & 49,923.73 & 0.927 & 1.162 \\ \hline
\begin{tabular}{l}GA\textsuperscript{2}M\\+monotonicity\end{tabular} & 39,436.42 & 48,543.97 & 1.016 & 1.141 \\ \hline
\begin{tabular}{l}GA\textsuperscript{2}M\\+ monotonicity\\+ higher\end{tabular} & 33,756.81 & \bf 47628.92 & 1.006 &\bf 1.136 \\ \hline
\end{tabular}
\end{table}
\subsection{Evaluating Monotonicity and Higher-Order Term}
Table \ref{tab:eval} shows the evaluation of the effect of introducing the monotonicity and the higher-order term proposed in this paper. We compare our proposed method, CGA\textsuperscript{2}M+, with other existing methods using the root mean squared error.
In both datasets, CGA\textsuperscript{2}M+ which introduced both monotonicity and a higher-order term shows the best results. We can also see that the introduction of the monotonicity reduced the test error, and as shown in Section \ref{sec:application}, improved interpretability. This suggests that the use of domain knowledge improves generalization performance. In contrast, although the effect of introducing the higher-order term is not as apparent with the Oricon data, some improvement in accuracy can be confirmed in the California housing data. The effectiveness of the higher-order term may vary depending on whether higher-order interactions are inherent in the data. The validity of interpreting CGA\textsuperscript{2}M+ with a non-interpretable higher-order term is discussed in Section \ref{sec:valid}.
\begin{table}[b]
\centering
\caption{Comparison with Existing Method}
\label{tab:exist}
\begin{tabular}{|c|c|c|c|c|}
\hline
{} & \multicolumn{2}{|c|}{\bf California} & \multicolumn{2}{|c|}{\bf Oricon}\\
\hline
\bf Model & \bf \textit{training} & \bf \textit{test} & \bf \textit{training} & \bf \textit{test}\\
\hline
Linear regression& 69,706.48 & 69,037.95 & 1.207 & 1.210\\ \hline
Random Forest& 18,313.34 & 49,829.04 & 0.419 & 1.160 \\ \hline
LightGBM& 41,893.29 & 48,880.43 & 1.090 & 1.148 \\ \hline
Multi-Layer Perceptron& 68,616.29 & 68,454.99 & 1.171 & 1.290 \\ \hline
CGA\textsuperscript{2}M+& 33,756.81 & \bf 47628.92 & 1.006 & \bf 1.136 \\ \hline
\end{tabular}
\end{table}
\subsection{Comparison with Existing Methods}
The results of comparing the proposed CGA\textsuperscript{2}M+ with existing methods are shown in Table \ref{tab:exist}. The models are compared by the root mean square error.
To implement the existing methods, we used the scikit-learn library [14] for linear regression, random forest, and multi-layer perceptron, and the LightGBM library [12] for LightGBM. In addition, the default settings were used for the hyperparameters of each method. From Table \ref{tab:exist}, we can see that CGA\textsuperscript{2}M+ gave the best results for both datasets. Linear regression is not highly accurate due to its lack of high expressive power, but it does not overfit the model, and there is little difference between the performance of training and test data. Random forest shows a considerable difference in accuracy between the training and test data, indicating that the model is overfitted. In addition, multi-layer perceptrons often require tuning of hyperparameters, such as layer depth and number of units, but in this case, the default values were used, resulting in unlearning.
\begin{figure*}[tb]
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[width=\linewidth]{image/california_importance.pdf}
\subcaption{California}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[width=\linewidth]{image/oricon_importance.pdf}
\subcaption{Oricon}
\end{minipage}
\caption{Feature importance for each dataset}
\label{fig:importance}
\end{figure*}
\subsection{Validity of Model Interpretation} \label{sec:valid}
Fig. \ref{fig:importance} shows a bar graph showing the importance of the univariate terms, pairwise interaction terms, and the higher-order term introduced in Section \ref{sec:propose}. The vertical axis is the number assigned to each term. Some features that have low importance are omitted for clarity. The importance of the higher-order term (bottom of each plot) is not that great. This is because most of the data can be explained by univariate and pairwise interaction terms, and higher-order interactions are not inherent in the data. In such a case, the validity of interpreting the whole model is relatively high, and both prediction accuracy and interpretability can be achieved. Conversely, when the importance of the higher-order term is large, the parts that can be explained by the univariate and pairwise interaction terms are relatively small, and the interpretation validity of the whole model becomes low. In other words, the importance of the higher-order term can be used as a measure for the interpretation validity of the univariate and pairwise interaction terms.
\subsection{Applications}\label{sec:application}
In this section, we introduce some examples of applications that take advantage of the interpretability of CGA\textsuperscript{2}M+. Note that the figures in this section can be prepared for all the univariate and pairwise interaction terms used in the trained model, but only some of them are shown for saving the space. Here, although explanations are given only for the univariate terms, the pairwise interaction terms can be interpreting in the same way.
\begin{figure}[tb]
\centerline{\includegraphics[width=\linewidth]{image/california_func.pdf}}
\caption{Plots of $f_i$ for California housing prices}
\label{fig:california}
\end{figure}
\paragraph{California Housing Prices}
Fig. \ref{fig:california} shows the relationship between each feature and the housing prices. We focus on the median income plot at the bottom middle. This is the median annual income in a certain region, and intuitively, the housing prices in that region is expected to increase monotonically with this value. Therefore, in this experiment, we imposed a monotonic increase constraint on this feature. The effect of the monotonic constraint shows that the housing prices monotonically increases. It can also be seen that the relationship between median income and housing prices is close to linear in the range of median income $<$ 10, and that median income does not affect housing prices when median income $\geq$ 10. Also, we imposed monotonic constraints on other features, such as average age and average number of rooms. These features can be interpreted in the same way as those of median income. We did not impose monotonicity on the other features, but if analysts have corresponding domain knowledge, they can consider introducing monotonicity to other features.
\paragraph{Oricon Cafe Satisfactions}
In this paragraph, we discuss how the cleanliness score in a cafe affects the satisfaction score of the cafe.
On the top center of Fig. \ref{fig:oricon}, when the cleanliness score for the cafe is 3.5, the cafe can increase the satisfaction score by 0.05, by improving the cleanliness score by 1. Also, the significantly lower Internet environment score has a large negative impact on the satisfaction score. This is probably because the Internet environment is an infrastructure element, and to some extent, it is taken for granted that it is installed. As a result, considering how the changes in each feature affect the objective variable leads to the interpretation of the whole model.
\section{Conclusion}
In this study, we proposed CGA\textsuperscript{2}M+, which has the high prediction performance and interpretability, by modifying the existing GA\textsuperscript{2}M model. The proposed model has two major improvements over GA\textsuperscript{2}M. The first is the use of domain knowledge, which improves the interpretability of the model and the generalization performance by imposing monotonicity constraints to some functions. The second is introduction of a higher-order term, which captures higher-order interactions that can not be explained by univariate and pairwise interaction terms, thereby improving the prediction accuracy.
In numerical experiments, we conducted a regression task using two datasets and confirmed that introducing the monotonicity and the higher-order term in CGA\textsuperscript{2}M+ is effective and that CGA\textsuperscript{2}M+ is superior to other models in terms of prediction accuracy. We also presented the two case studies, which demonstrate that CGA\textsuperscript{2}M+ has high interpretability.\\
\begin{figure}[h]
\centerline{\includegraphics[width=\linewidth]{image/oricon_func.pdf}}
\caption{Plots of $f_i$ for Oricon cafe satisfactions}
\label{fig:oricon}
\end{figure}
\newpage
|
\section{Introduction}\label{sec:intro}
The \textit{multi-armed bandit (MAB)} problem is a classical sequential decision-making problem in which an agent tries to maximize a cumulative stochastic reward~\cite{thompson1933likelihood, robbins1952some} under uncertainty. This problem, which is applicable to various areas such as recommender systems, online advertising and clinical trials, embodies the well known exploration-exploitation trade-off between learning the environment and acting optimally based on our current knowledge about the environment.
More formally, in the MAB problem at each time $t=1,\ldots,T$ an agent chooses an arm $i$ from the set $[k]=\{1,\ldots,k\}$ of $k$ arms, and obtains an iid reward $r^t$ drawn from the unknown distribution $R_i$ over $\{0,1\}$ with expectation $\mu_i=\expect{}{R_i}$. Let $a^*=\argmax_{a}\mu_{a}$ be an arm with the largest expected reward, and denote this reward by $\mu^*=\mu_{a^*}$. Let the (suboptimality) gap of an arm $a$ to be the gap between its expected reward and that of $a^*$, i.e., $\Delta_a = \mu^* - \mu_a$. The agent's goal is to maximize the total expected reward, or rather to minimize the expected regret $R(T)=T\cdot \mu^*-\expect{}{\sum_{i=1}^{T}r^t}$ defined to be the expected gap between the algorithm and the optimal algorithm that knows the distributions $R_i$.
In this work we address the privacy in such a setting.
As a motivating example, consider an advertisement system in which
the server presents to each user an
advertisement $a\in [k]$.
The user then decides whether to click on the
advertisement or not.
This click decision depends on different private characteristic of the user. The user then reports to the server whether it
clicked on the advertisement (in which case its reward is $r=1$) or not ($r=0$).
From this example, it is clear that $r$ is private information of the user, and using traditional algorithms for the MAB problem incautiously might leak user-private data.
In order to mathematically alleviate privacy concerns, Dwork et al.~\cite{dwork2006calibrating} defined the notion of \textit{differential privacy (DP)}, which requires that the output of the computation has a limited dependency on any single user's data. Formally, a mechanism $(\varepsilon,\delta)$-DP if for any pair of neighboring inputs (differing by a single user's data), the probability that the mechanism outputs a value in any set $B$ is not different by more than a multiplicative factor of $e^\varepsilon$ and an additive factor of $\delta$. \textit{Differential privacy} has been extensively studied under many different sub-models of privacy. On one end of the spectrum lies the \textit{centralized model} of differential privacy, where the users trust the server with their data, and the liability to protect user privacy lies on the server, who must make sure that any data published externally (e.g., aggregated statistics) respects the privacy constraints. On the other end of the spectrum lies the (strictly stronger) \textit{local model} of differentialy privacy (LDP), where the user privatizes its own data prior to sending it to the server.\footnote{Formally, in the non-interactive setting, a mechanism is $(\varepsilon,\delta)$-LDP if for any two user inputs, the probability that the privatizer sends the server a value in any set $B$ is not different by more than a multiplicative factor of $e^\varepsilon$ and an additive factor of $\delta$.}
Differentially private versions of the MAB problem have been considered in various previous works, where the private information are users' rewards (two neighboring inputs differ by the reward value of a single user), and the algorithm's output is the subsequent arm(s) it selects.
Table~\ref{tab:results} summarizes the best known distribution-dependent regret bounds for the various privacy models, together with our new result in the \textit{shuffle model} which we soon define formally.
\begin{table}
\caption{Best-known MAB regret upper and lower bounds for various DP models.}
\label{tab:results}
\centering
\begin{tabular}{ll}
\toprule
\textbf{Privacy model} & \textbf{Best-known regret upper and lower bounds\tablefootnote{The corresponding distribution-independent regret bounds usually simply replace the $\sum_{a\in [k]:\Delta_a>0}\frac{\log T}{\Delta_a}$ term with $\sqrt{kT\log T}$.}}\\
\midrule
\textbf{Centralized $(\varepsilon,0)$-DP} & $\Theta\left(\left(\sum_{a\in [k]:\Delta_a>0} \frac{\log T}{\Delta_a}\right)+\frac{k\log T}{\varepsilon}\right)$~\cite{sajed2019optimal,shariff2018differentially}\\
\textbf{Centralized $(\varepsilon,\delta)$-DP} & $O\left(\left(\sum_{a\in [k]:\Delta_a>0} \frac{\log T}{\Delta_a}\right)+\frac{k}{\varepsilon}\right)$~\cite{tossou2016algorithms}\\
\textbf{Local $(\varepsilon,0)$-DP} & $\Theta\left(\frac{1}{\varepsilon^2}\sum_{a\in [k]:\Delta_a>0}\frac{\log T}{\Delta_a}\right)$~\cite{ren2020multi}\tablefootnote{This lower bound can be extended from $(\varepsilon,0)$-LDP to $(\varepsilon,\delta)$-LDP using arguments from Bun et al.~\cite{bun2019heavy} and Cheu et al.~\cite{cheu2019distributed} since we focus on single-round (non-interactive) mechanisms in which the user can only send information to another party once.
}\\
\textbf{Shuffle $(\varepsilon,\delta)$-DP (ours)} & $O\left(\left(\sum_{a\in [k]:\Delta_a>0}\frac{\log T}{\Delta_a}\right)+\frac{k\sqrt{\log\frac{1}{\delta}}\log T}{\varepsilon}\right)$\\
\bottomrule
\end{tabular}
\end{table}
With real-world algorithms gradually moving away from the \textit{centralized model} of privacy, the immediate question is \enquote{can we design a private algorithm for MAB with privacy guarantees which are similar to local DP, but with similar regret to centralized DP (without a multiplicative $1/\varepsilon^2$ factor)?}.
To address the inevitable gap between the local and centralized models, which is in fact common in the literature of differential privacy, the alternative \textit{shuffle model}~\cite{bittau2017prochlo,cheu2019distributed,erlingsson2019amplification} explores the space in between the local and centralized models by introducing a trusted shuffler that receives user messages and permutes them (i.e., disassociates a message from its sender) before they are delivered to the server.
For privacy analysis, we assume that the shuffle is perfectly secure, i.e., its output contains no information about which user generated each of the messages. This is traditionally achieved by the shuffler stripping implicit metadata from the messages (e.g., timestamps, routing information), and frequently forwarding this data to remove time and order information.
The shuffle model ensures that sufficiently many reports are collected in each round so that any one report can hide in a shuffled batch. In order to apply the shuffle model to the MAB problem in the context of advertisements, we divide the algorithm into batches, where before each batch we decide on the fly its size $m$, and then present the $m$ next users the same advertisement $a$, and finally apply a private shuffle model mechanism to their rewards to communicate reward aggregate information to the server.\footnote{We remark that a given user does not know in advance the size of the batch in which it participates, since this size depends on the algorithm's run.}
A constantly growing body of work presents new and improved mechanisms in the shuffle model for basic statistical tasks~\cite{ghazi2019scalable, erlingsson2019amplification, erlingsson2020encode, balle2019privacy}, such as \textit{private binary summation}, in which the server must privately approximate the sum of a collection of values $x_1,..., x_m \in \{0, 1\}$ held by the $m$ users in the shuffle. For \textit{private binary summation}, the optimal achievable errors in the central, local and shuffle model are $\tilde{\Theta}(1/\varepsilon)$~\cite{dwork2006calibrating}, $\tilde{\Theta}(\sqrt{m}/\varepsilon)$~\cite{beimel2008distributed,chan2012optimal} and $\tilde{\Theta}(1/\varepsilon)$~\cite{cheu2019distributed} respectively, where the $\tilde{\Theta}(\cdot)$ hides poly-logarithmic terms. Similar errors hold for the \textit{private summation} problem which approximates the sum of real values in $[0,1]$.
\subsection{Our contributions}
To the best of our knowledge, our work is the first to consider the MAB problem under the shuffle model of differential privacy. In order to support the online nature of the MAB problem, we consider a variant of the shuffle model. As opposed to the classical shuffle model, in which the shuffle size is unbounded and the mechanism runs only once, we continuously run shuffle mechanisms for many disjoint batches of users who can only afford a single round of communication.\footnote{Beimel et al.~\cite{beimel2020round} showed that every centralized-DP mechanism can be emulated in the shuffle model in two (communication-intensive) rounds of communication, however these results are not applicable to our setting
since we assume that each online user participates only in one shuffle and
then disappears.}
We consider the paradigm where the server controls the different users who are cooperative and communicate only with the server. We give a rigorous definition of \textit{shuffle differential privacy (SDP)} for the multi-armed bandit problem (assuming binary rewards), and give and prove the first two such algorithms. Our algorithms \textit{Shuffle Differentially Private Arm Elimination (\mbox{SDP-AE}\xspace)} and \textit{Variable Batch Shuffle Differentially Private Arm Elimination (\mbox{VB-SDP-AE}\xspace)} are both based on the well-known \textit{arm elimination (\AE)} algorithm, using consecutive batches of users and together with an SDP private binary summation mechanism.
We show that the simpler but weaker \mbox{SDP-AE}\xspace achieves a distribution-dependent regret of
$O\left(\left(\sum_{a\in [k]:\Delta_a>0}\frac{\log T}{\Delta_a}\right)+\frac{k\log\frac{1}{\delta}}{\varepsilon^2}\right)$, and a distribution-independent regret of $O\left(\sqrt{kT\log T}+\frac{k\log\frac{1}{\delta}}{\varepsilon^2}\right)$.\footnote{In this paper, we assume that $T$ is known apriori. Otherwise, we can apply standard doubling arguments to get roughly the same results.}
We then describe \mbox{VB-SDP-AE}\xspace, a generalization of \mbox{SDP-AE}\xspace to exponentially growing batch sizes, and prove it has a distribution-dependent regret of $O\left(\left(\sum_{a\in [k]:\Delta_a>0}\frac{\log T}{\Delta_a}\right)+\frac{k\sqrt{\log\frac{1}{\delta}}\log T}{\varepsilon}\right)$, and a distribution-independent regret of $O\left(\sqrt{kT\log T}+\frac{k\sqrt{\log\frac{1}{\delta}}\log T}{\varepsilon}\right)$.
Note that, compared to the local model (Ren et al~\cite{ren2020multi}), the regret of
both \mbox{SDP-AE}\xspace and \mbox{VB-SDP-AE}\xspace is improved, by having the dependency on $1/\varepsilon$ be additive rather than multiplicative.
In addition, \mbox{VB-SDP-AE}\xspace almost matches the regret of the best known algorithms for the centralized model, that is the distribution-dependent regret of $O\left(\left(\sum_{a\in [k]:\Delta_a>0} \frac{\log T}{\Delta_a}\right)+\frac{k}{\varepsilon}\right)$ of Tossou and Dimitrakakis~\cite{tossou2016algorithms}.
\subsection{Related work}
The differentially private MAB problem has been considered in many previous works~\cite{jain2012differentially,guha2013nearly,mishra2015nearly}. Shi and Shen~\cite{shi2021federated} and Dubey and Pentland~\cite{dubey2020differentially} studied MAB and linear bandits respectively in the federated setting. Zheng et al.~\cite{zheng2020locally} studied contextual bandits with LDP.
Batched MAB with a predetermined number of batches was studied in~\cite{esfandiari2019regret,gao2019batched}.
For the private summation problem, Cheu et al.~\cite{cheu2019distributed} gave unbiased $(\varepsilon,\delta)$-SDP mechanisms over binary inputs and real inputs in $[0,1]$, with error roughly $\sqrt{\log ({1}/{\delta}})/\varepsilon$ and $\log ({1}/{\delta})/\varepsilon$, respectively. In several works of Balle et al.~\cite{balle2019privacy,balle2020private,balle2019differentially}, they gave a biased $(\varepsilon,\delta)$-SDP mechanism for real inputs in $[0,1]$ with similar error and a constant number of messages, and a single-message $(\varepsilon,\delta)$-SDP mechanism with optimal error.
\section{Background and preliminaries}\label{sec:prelim}
\subsection{Shuffle-model privacy}
In the well-studied setting of shuffle model privacy, there are $m$ users, each with data $x_i \in X$. Each user applies some encoder $E : X \to Y^*$ to their data and sends the messages $(y_{i,1},...,y_{i,p}) = E(x_i)$ to a shuffler $S : Y^* \to Y^*$. The shuffler then shuffles all the messages $y_{i,j}$ from all the users, and outputs them in a uniformly random order to an analyzer $A : Y^* \to Z$ to estimate some function $f(x_1,..., x_m)$. Thus, the mechanism $M$ consists of the tuple $(E, S,A)$. We say that such a mechanism $M$ is \textit{$(\varepsilon, \delta)$-shuffle differentially private} (or \textit{$(\varepsilon, \delta)$-SDP} for short) if the output of $S$ is $(\varepsilon, \delta)$-differentially private, or more formally:
A mechanism $M=(E, S,A)$ is $(\varepsilon,\delta)$-SDP if for any pair of inputs $\{x_i\}_{i=1}^{m}$ and $\{x_i'\}_{i=1}^{m}$ which differ in at most one value, we have for all $B\subseteq Y^*$:
$$P(S(\cup_{i=1}^{m}E(x_i))\in B)\leq e^\varepsilon\cdot P(S(\cup_{i=1}^{m}E(x_i'))\in B)+\delta.$$
In the mechanism used in this paper, $E$ outputs the user's reward bit together with a set of random bits, and $A$ sums all these bits and debiases the result to get an unbiased estimate of the sum of the users' rewards.
\subsection{Shuffle-model MAB}
Algorithms which are private in the shuffle model
typically apply the mechanism $M$ once over a set of $m$ users. Here we study the MAB problem which is an online problem, often deployed in real-world applications and with users which are end-devices such as cellphones with a possibly limited or unreliable internet connection. Hence, to adapt the MAB problem to the shuffle model, we batch sequences of consecutive users, and assume that each user can afford a single round of communication, and is never selected more than once.
\paragraph{Model and objective}
The \textit{shuffle-model MAB} setting involves repeating the following process until the $T$'th player pulls its arm:
\begin{enumerate}
\item The server selects a batch size $m$,
a \textit{batch} of $m$ random fresh new users, and an $m$-user single-round SDP mechanism $M$.\footnote{Note that always selecting $m=1$ reduces this setting to the Local model MAB.}
It then picks an arm
$a\in [k]$ that all $m$ users of the batch pull. (For concreteness think that the server picks an ad $a\in [k]$ and sends it to the a random batch of $m$ users.)
\item
Each user $i$ determines binary reward from pulling the arm $a$.\footnote{For simplicity, we assume that the rewards are binary. However, our algorithms and proofs naturally extend to the real $[0,1]$ reward setting, by replacing our private binary summation mechanism (defined later, see Appendix~\ref{sec:privateSummation}) with a private summation mechanism for real numbers in $[0,1]$ with similar guarantees.}
(For concreteness think
that each user decides whether to click on the ad or not. This defines the reward related to user $i$, which is $r_i=1$ if it clicks the ad and
$r_i=0$ otherwise.)
Since our $m$ users are random, these rewards are a sample of $m$ independent rewards from
the distribution $R_a$ associated with arm $a$.
\item The server computes $M(\{r_i\}_{i\in batch})$ using the rewards $r_i$.
\end{enumerate}
The objective is to minimize the (pseudo) regret, which is the expected difference between the sum of the rewards accumulated (over all the users) by the algorithm and the sum of the rewards of the optimal algorithm that apriori knows an arm with the largest expected reward $a^*$. Let $\mu_a=\expect{}{R_a}$ be the \textit{expected reward} (or simply \textit{mean}) of the arm $a$, let $\mu^*$ denote the expected reward of $a^*$, let $\Delta_a = \mu^* - \mu_a$ be the (suboptimality) gap of the arm $a$ which quantifies the gap between its expected reward and that of $a^*$, and let $N_a$ be the random variable which counts the total number of times the arm $a$ was pulled during the run of the algorithm.
Formally, the (pseudo) regret of the algorithm for $T\in \mathbb{N}$ users is defined to be
$$R(T) = T\cdot \mu^* - \expect{}{\sum_{i=1}^{T} r_{i}} =\expect{}{\sum_{a\in [k]}N_a \Delta_{a}}.$$
\paragraph{Privacy}
Since the private data of each user is its reward, the appropriate adaptation of shuffle model privacy for the multi-armed bandit problem is as follows.
An algorithm for the multi-armed bandits problem is \textit{$(\varepsilon,\delta)$-shuffle differentially private (SDP)} if for any batch of users, the shuffle mechanism that we apply over them in step 3 is $(\varepsilon, \delta)$-SDP with respect to the rewards of the users (as we recall, the reward of a user -- whether it clicked on an ad or not -- depends on its private features). Formally, for every batch we run a shuffle mechanism where the $m$ users are the users of the current batch, and the data $x_i$ of each user is its reward $r_i$, and we require that each such mechanism is $(\varepsilon, \delta)$-SDP.
\subsection{Concentration bounds}
We use the following standard definitions of Sub-Gaussian random variables and Hoeffding's inequality.
\begin{definition}[Sub-Gaussian random variable]\label{def:subgaussian}
A random variable $X$ with mean $\mu$ is called \textit{sub-Gaussian with variance $\sigma^2$}, i.e., $X\sim SG(\sigma^2)$ if:
$$\forall \lambda\in \mathbb{R},~\expect{}{\exp(\lambda(X-\mu))}\leq e^{\lambda^2\sigma^2/2}.$$
\end{definition}
An equivalent definition shows that if $\forall t>0,\max\left({P(X-\mu\geq t),P(X-\mu\leq -t)}\right)\leq \exp\left(\frac{-t^2}{2\sigma^2}\right)$, then $X$ is sub-Gaussian with variance $\sigma^2$ (up to constant factor).
It is well known that if $X_i\sim SG(\sigma_i^2)$ are independent random variables for $i=1,\ldots,n$, then $\sum_{i=1}^{n}X_n \sim SG(\sum_{i=1}^{n}\sigma_i^2)$ and for any $a,b>0$, $a\cdot X_1+b\sim SG(a^2\cdot \sigma_1^2)$. A bounded random variable $X\in[a,b]$ is $SG((b-a)^2/4)$.
Sub-Gaussian random variables satisfy the following concentration bound,
\begin{lemma}[Hoeffding's inequality~\cite{hoeffding1994probability}]\label{lma:hoeffding}
Let $\{X_i\}_{i=1}^n$ be independent $SG(\sigma^2)$ random variables,
then
$P\left(\abs{\sum_{i=1}^n X_i-\expect{}{\sum_{i=1}^n X_i}}\geq t\right)\leq 2\exp\left(\frac{-t^2}{2n\sigma^2}\right).$
\end{lemma}
\section{Differentially private MAB in the shuffle model}
In this paper, we use the shuffle model to give a private solution to the multi-armed bandit problem, attaining similar privacy guarantees to that of the local privacy model (LDP), without sacrificing the regret. That is, our algorithm almost matches the best known regret in the centralized model of differential privacy.
Our algorithms rely on the fact that algorithms for the multi-armed bandit
problem take decisions based on
sums of rewards received from the users. Hence, we rely on a particularly efficient and accurate mechanism for private binary summation in the shuffle model as a building block in our algorithms. Specifically, for any $\varepsilon,\delta\in (0,1)$, let $M_{sum}$ be a private binary summation mechanism, which for any number of users (batch size) $m\in \mathbb{N}$, is $(\varepsilon,\delta)$-SDP, unbiased, and has an error distribution which is independent of the input, and is sub-Gaussian with variance $\sigma_{\varepsilon,\delta}^2=O\left(\frac{\log \frac{1}{\delta}}{\varepsilon^2}\right)$.\footnote{Note that our methods and algorithms should work similarly if they were built over other common algorithms for the MAB problem in which batching makes sense, such as the \mbox{UCB}\xspace algorithm.} Note that our notation of $M_{sum}$ does not include $\varepsilon$ and $\delta$ which will always be clear from context or inherited from the algorithm which runs $M_{sum}$.
The challenge is to combine $M_{sum}$ with the well studied \textit{arm elimination (\AE)} MAB algorithm to get an SDP algorithm for the MAB problem with almost optimal regret.\footnote{For completeness, in Appendix~\ref{sec:privateSummation} we give a complete description and a proof of such a mechanism $M_{sum}$.}
\subsection{\mbox{SDP-AE}\xspace: Shuffle Differentially Private Arm Elimination}
\label{sec:sdpMAB}
We base our algorithm on the (non-private) \textit{arm elimination (\AE)} algorithm for the MAB problem, which informally maintains a set of viable arms (initially set to be $[k]$), and each phase pulls the set of viable arms sequentially. Once a phase ends, we search for arms which are noticeably suboptimal in comparison to some other arm, and we eliminate them from the set of viable arms.
To adapt \AE to the shuffle model, in each phase $t$ each arm is pulled not once, but rather by a whole batch of users. Once all the users in the batch pull the arm and receive their reward, we apply the private binary summation mechanism $M_{sum}$ to the batch's rewards, which gives the server an unbiased but noisy estimate of the sum of rewards in the batch. This estimate has two sources of error, which we account for when we compute the upper confidence bound -- the empirical error due to sampling the reward function of the arm, and the error due to the private binary summation mechanism $M_{sum}$.
\subsubsection{Algorithm outline}
In the algorithm below which we call \textit{Shuffle Differentially Private Arm Elimination (\mbox{SDP-AE}\xspace)}, we update the estimate of the mean reward of each arm $a$ after every batch of users who sample $a$.
The algorithm works in phases, and maintains a set of viable arms initially set to be $[k]$. In each phase, for every viable arm we have a single batch of users sampling it.
At the end of phase $t$, we denote by $\hat{S}_a^t$ the noisy estimate of the cumulative sum of the rewards from all previous samples of $a$ in the phases $1,\ldots,t$, and denote by $N_a^t$ the total number of previous samples of arm $a$ in the phases $1,\ldots,t$. The natural estimate for the mean reward of the arm $a$ (denoted by $\hat{\mu}_a^t$) is therefore $\hat{\mu}_a^t=\hat{S}_a^t/N_a^t$. We then calculate the upper and lower confidence bounds $UCB_a^t$ and $LCB_a^t$ respectively of each viable arm $a$ after each phase $t$ using a specific bound which takes into account both sources of error. We finally eliminate any remaining arm with an upper confidence bound which is strictly smaller than the lower confidence bound of some other arm. Algorithm~\ref{alg:batchae} consolidates the algorithm presentation above.
\medskip
\begin{algorithm}[H]\label{alg:batchae}
\SetAlgoLined
\textbf{Input:} privacy parameters $\varepsilon$ and $\delta$, batch size $m$ and horizon $T$.\\
\textbf{Initialize:} $\hat{S}_a^0=0$, $N_a^0=0$ and $\hat{\mu}_{a}^0=0$ for every $a\in [k]$\;
Let $\sigma_{\varepsilon,\delta}^2=O\left(\frac{\log \frac{1}{\delta}}{\varepsilon^2}\right)$ be the sub-Gaussian variance of the error distribution of $M_{sum}$\;
Let $V=[k]$ denote the set of viable arms\;
\For{phase $t\gets1,2,\ldots$}{
\For{arm $a\in V$}{
\For{each new user $i\gets1$ \KwTo $m$}{
User $i$ pulls the arm $a$ and observes reward $r_{a,i}^t$\;
If total arm samples in current algorithm run is $T$, exit\;
}
\textbf{Communication:} Perform private binary summation
$Z_a^t \gets M_{sum}\left(\{r_{a,i}^t\}_{i=1}^{m}\right)$\;
\textbf{Server update:} Update $\hat{S}_{a}^{t} \gets \hat{S}_{a}^{t-1} + Z_a^t $,
$N_{a}^{t} \gets N_{a}^{t-1} + m$, and finally $\hat{\mu}_{a}^{t} \gets \hat{S}_{a}^{t}/N_{a}^{t}$\;
}
\textbf{Confidence bounds:} For each arm $a$, calculate $I_a^t\gets\left(\frac{2\sqrt{t}\sigma_{\varepsilon,\delta}}{N_a^t}+\frac{1}{\sqrt{N_a^t}}\right)\cdot\sqrt{2\log T}$, and the upper and lower confidence bounds $UCB_a^t \gets \hat{\mu}_{a}^{t}+I_a^t$ and $LCB_a^t \gets \hat{\mu}_{a}^{t}-I_a^t$\;
\textbf{Elimination:} remove all arms $a$ from $V$ such that $UCB_a^t < \max_{a'\in S}LCB_{a'}^t$\;
}
\caption{\mbox{SDP-AE}\xspace (Shuffle Differentially Private Arm Elimination)}
\end{algorithm}
\subsubsection{Analysis}
The privacy is trivial, since each batch we use the $(\varepsilon,\delta)$-SDP mechanism $M_{sum}$. We now focus on regret.
Theorem~\ref{thm:batchAERegret} gives a bound on the regret of \mbox{SDP-AE}\xspace as a function of the batch size $m$. We follow a somewhat standard regret bound analysis for arm elimination, comprising two parts. The first part uses Hoeffding's inequality to derive a high probability bound on $\abs{\hat{\mu}_a^t-\mu_a}$, the error between the empirical average reward and the true mean reward of the arm $a$ at a given phase $t$. The second part uses this bound to bound the expected number of times we sample each suboptimal arm $a$, and summing over all suboptimal arms we get a bound on the regret.
\begin{theorem}\label{thm:batchAERegret}
The algorithm \mbox{SDP-AE}\xspace is $(\varepsilon,\delta)$-SDP, and has a distribution-dependent regret of $O\left(\sum_{a\in [k]:\Delta_a>0}\left(\frac{\log T}{\Delta_a}+\frac{\sigma_{\varepsilon,\delta}^2\log T}{m\Delta_a}+m\Delta_a\right)\right)$, and a distribution-independent regret of $O\left(\sqrt{\left(1+\frac{\sigma_{\varepsilon,\delta}^2}{m}\right)kT\log T}+mk\right)$.
\end{theorem}
\begin{proof}
First, note that for any arm $a$, and phase $t$, we have
$\hat{\mu}_a^t=\frac{\sum_{s=1}^{t}M_{sum}\left(\{r_{a,i}^s\}_{i=1}^{m}\right)}{N_a^t}$, where
$N_a^t=m\cdot t$.
We define the \textit{clean event} $C:=\left\{\forall a\in[k],\forall t\in[T]~~\abs{\hat{\mu}^t_a-\mu_a}\leq I_a^t\right\},$ where $I_a^t:=\left(\frac{2\sqrt{t}\sigma_{\varepsilon,\delta}}{N_a^t}+\frac{1}{\sqrt{N_a^t}}\right)\cdot \sqrt{2\log T}$ is a confidence bound interval. We now show that the clean event $C$ occurs with high probability, that is $P\left(C\right)\geq1-4T^{-2}$, and after that assume the event $C$ to simplify our analysis.
Indeed, for each arm $a$, we imagine both a reward tape of length $1\times T$, with each cell independently sampled from the distribution $D_a$ of arm $a$, and a private-binary-summation-error tape of length $1\times T$, with each cell independently sampled from the distribution of (the additive) error of the private binary summation mechanism $M_{sum}$ for $m$ users.
We assume that in the $j$'th time a given arm $a$ is pulled by the algorithm, its reward is taken from the $j$'th cell in this arm's reward tape, and similarly the $j$'th time we compute a private binary sum over rewards of a batch of users who pulled $a$, the (additive) error is taken from the $j$'th cell in the arm's private-binary-summation-error tape.\footnote{Here we rely on the fact that the distribution of the error of $M_{sum}$ is independent of the input.}$^,$\footnote{Note that sizes of both tapes have been chosen conservatively to be of size $T$. We never pass the end of any these tapes, since there are at most $T$ users in total, and at most $T$ batches (actually roughly $T/m$ in this case), and we may not use them all.}
Let $t\in [T]$ and $a\in [k]$, and let $\hat{v}_a^t$ be the approximated reward of arm $a$ that the algorithm would have held at the end of phase $t$ using the concrete values in the tapes defined above,
that is $ \hat{v}^t_a =
\frac{\sum_{s=1}^{t}d_s+\sum_{i=1}^{N_a^t}e_{i}}{N_a^t}$, where $d_{s}$ is the $s$'th cell of the private-binary-summation-error tape of $a$, and the $\{e_{i}\}_{i=1}^{N_a^t}$ are the total $N_a^t=m\cdot t$ cells of the reward tape of $a$ that we have used until the end of the $t$'th phase.
Our aim is to bound the term $\abs{\hat{v}^t_a-\mu_a}=\frac{\sum_{s=1}^{t}d_s+\sum_{i=1}^{N_a^t}(e_{i}-\mu_a)}{N_a^t}$. We first bound the first sum in the nominator, then bound the second sum in the nominator, and finally combine the bounds to get a bound for $\abs{\hat{v}^t_a-\mu_a}$.
To bound the $d_{s}$'s sum, we apply Hoeffding's inequality (Lemma~\ref{lma:hoeffding}) for a sum of $n\gets t$ random variables which are sub-Gaussian with variance $\sigma_{\varepsilon,\delta}^2$ and have zero mean (since $M_{sum}$ is unbiased) to get,
\begin{align}
P\left(\abs{\sum_{s=1}^{t} d_{s}}\leq 2\sigma_{\varepsilon,\delta}\cdot \sqrt{2t\log T}\right)
&\geq 1-2\exp\left(-\left(4\sigma_{\varepsilon,\delta}^2\cdot 2t\log T\right)/\left(2t\sigma_{\varepsilon,\delta}^2\right) \right)\nonumber\\
&= 1-2T^{-4}.\label{eq:dsbound}
\end{align}
To bound the $f_i=(e_{i}-\mu_a)$'s sum, we apply Hoeffding's inequality (Lemma~\ref{lma:hoeffding}) for the sum of $n\gets N_a^t$ random variables $f_i$, each sub-Gaussian with variance 1/4 (since it is bounded in the interval $[-\mu_a,1-\mu_a]$ of size 1), and with zero mean (since by its definition $\expect{}{e_i-\mu_a}=\expect{}{e_i}-\mu_a=0$), to get that
\begin{align}
P\left(\abs{\sum_{i=1}^{N_a^t}(e_{i}-\mu_a)}\leq \sqrt{2N_a^{t}\log T}\right)
&\geq1- 2\exp\left(-(2N_a^t\log T)/(2N_a^t/4)\right)=1-2T^{-4}.\label{eq:eibound}
\end{align}
Applying a union bound and the triangle inequality on Equation~\eqref{eq:dsbound} and Equation~\eqref{eq:eibound} gives that
$$P\left(\abs{\sum_{s=1}^{t} d_{s}+\sum_{i=1}^{N_a^t}(e_{i}-\mu_a)}\leq \left(2\sqrt{t}\sigma_{\varepsilon,\delta}+\sqrt{N_a^t}\right)\cdot \sqrt{2\log T}\right)\geq1-4T^{-4},$$
which by the definition of $\hat{v}^t_a$ and $I_a^t$ means that
\begin{align}
P\left(\abs{\hat{v}^t_a-\mu_a}\leq
I_a^t
\right)\geq1-4T^{-4}.\label{eq:VaMuaBound}
\end{align}
Since in the analysis above $t$ and $a$ are arbitrary, Equation~\eqref{eq:VaMuaBound} holds for every $t\in [T]$ and $a\in [k]$. Thus, we take a union bound over all arms $a\in [k]$ (assuming $k\leq T$) and all $t\in [T]$, to conclude that
\begin{align}
P\left(\forall a\in [k], \forall t\in [T]~\abs{\hat{v}^t_a-\mu_a}\leq
I_a^t
\right)\geq1-4T^{-2}.\label{eq:allBound}
\end{align}
Since the event in the probability above in Equation~\eqref{eq:allBound} is precisely the event that $C$ holds for a run of the algorithm using the randomness in the tapes as defined above (by the definitions of $\hat{v}^t_a$ and $\hat{\mu}^t_a$), we get that $$P\left(C\right)\geq1-4T^{-2}.$$
For the regret analysis, we assume the clean event $C$. Consider a suboptimal arm $a$ such that $\Delta_a=\mu^*-\mu_a>0$, and consider the last phase $t_0$ following which we did not remove the arm $a$ yet (or the last phase if $a$ remains active to the end). Since we assumed the clean event, an optimal arm $a^*$ cannot be disqualified, and since $a$ is not yet disqualified, the confidence intervals of the arms $a$ and $a^*$ at the end of the $t_0$'s phase must overlap. Therefore,
\begin{align}
\Delta_a=\mu^*-\mu_a\leq 2(I_a^{t_0}+I_{a^*}^{t_0}) =4I_a^{t_0}=\left(\frac{8\sqrt{t}\sigma_{\varepsilon,\delta}}{N_a^{t_0}}+\frac{4}{\sqrt{N_a^t}}\right)\cdot \sqrt{2\log T},\label{eq:deltaaBound}
\end{align}
where the third step follows since $a$ and $a^*$ were sampled using identical batch sizes throughout the algorithm, so at the end of the ${t_0}$'th phase, $N_a^{t_0}=N_{a^*}^{t_0}$ and therefore $I_a^{t_0}=I_{a^*}^{t_0}$, and the last step follows by the definition of $I_a^{t_0}$.
Observe that if $N_a^{t_0}> \frac{128\log T}{\Delta_a^2}$ then $\frac{4\cdot \sqrt{2\log T}}{\sqrt{N_a^{t_0}}}<\frac{\Delta_a}{2}$, and if $N_a^{t_0}>\frac{(16\sigma_{\varepsilon,\delta}\cdot \sqrt{2\log T})^2}{m\Delta_a^2}$ then $\frac{8\sqrt{{t_0}}\sigma_{\varepsilon,\delta}\cdot \sqrt{2\log T}}{N_a^{t_0}}=\frac{8\sqrt{N_a^{t_0}/m}\sigma_{\varepsilon,\delta}\cdot \sqrt{2\log T}}{N_a^{t_0}}=\frac{8\sigma_{\varepsilon,\delta}\cdot \sqrt{2\log T}}{\sqrt{mN_a^{t_0}}}<8\sigma_{\varepsilon,\delta}\cdot \sqrt{2\log T}\cdot \frac{\Delta_a}{16\sigma_{\varepsilon,\delta}\cdot \sqrt{2\log T}}=\frac{\Delta_a}{2}$, so their sum is $<\Delta_a$ in contradiction to Equation~\eqref{eq:deltaaBound}. Hence, $N_a^{t_0}\leq \max\left( \frac{128\log T}{\Delta_a^2},\frac{512\sigma_{\varepsilon,\delta}^2\cdot \log T}{m\Delta_a^2}\right)$.
Therefore the total regret on arm $a$ is
\begin{align}
R_a&\leq \Delta_a\cdot (N_a^{t_0}+m)\leq \max\left( \frac{128\log T}{\Delta_a},\frac{512\sigma_{\varepsilon,\delta}^2\cdot \log T}{m\Delta_a}\right)+m\Delta_a\nonumber\\
&\leq \frac{128\log T}{\Delta_a}+\frac{512\sigma_{\varepsilon,\delta}^2\cdot \log T}{m\Delta_a}+m\Delta_a,\label{eq:RaBound}
\end{align}
where the first step follows since since the arm $a$ is eliminated following phase ${t_0}+1$ (or if $t_0$ is the last phase, then we finish and don't sample $a$ after it) of batch size $m$ and is subsequently never pulled, and the second step follows by previous bound on $N_a^{t_0}$ and since $N_a^{{t_0}+1}=N_a^{t_0}+m$.
We sum up the regret over all arms, to obtain a bound for the total regret denoted by $R$:
$$R=\sum_{a\in [k]:\Delta_a>0}R_a\leq \sum_{a\in [k]:\Delta_a>0}\left(\frac{128\log T}{\Delta_a}+\frac{512\sigma_{\varepsilon,\delta}^2\cdot \log T}{m\Delta_a}+m\Delta_a\right).$$
To complete the analysis, we argue that the bad event in which $C$ does not hold contributes a negligible amount to the expected regret $R(T)$. Indeed,
\begin{align}
R(T)&=\expect{}{R\mid C}\cdot P(C)+\expect{}{R \mid \bar{C}}\cdot P(\bar{C})\nonumber\\
&\leq \sum_{a\in [k]:\Delta_a>0}\left(\frac{128\log T}{\Delta_a}+\frac{512\sigma_{\varepsilon,\delta}^2\cdot \log T}{m\Delta_a}+m\Delta_a\right)+T\cdot4T^{-2}\nonumber\\
&= O\left(\sum_{a\in [k]:\Delta_a>0}\left(\frac{\log T}{\Delta_a}+\frac{\sigma_{\varepsilon,\delta}^2\log T}{m\Delta_a}+m\Delta_a\right)\right),\label{eq:expectedRegret}
\end{align}
where the first step follows by the law of total expectation, and the second step follows since the regret is at most $T$, and by the previous bound on $P(C)$.
Now for the distribution-independent bound, assume the clean even $C$, and let $\gamma>0$ be a threshold whose exact value we will set later. We group the arms $a$ based on if $\Delta_a<\gamma$ or not, to get
\begin{align*}
R&=\sum_{a\in [k]\mid \Delta_a<\gamma} R_a
+\sum_{a\in [k]\mid \Delta_a\geq\gamma}R_a\\
&\leq T\cdot \gamma
+\sum_{a\in [k] \mid \Delta_a\geq\gamma}\left(\frac{128\log T}{\Delta_a}+\frac{512\sigma_{\varepsilon,\delta}^2\cdot \log T}{m\Delta_a}+m\Delta_a\right)\\
&\leq T\cdot \gamma + \frac{(128+512\sigma_{\varepsilon,\delta}^2/m)k\log T}{\gamma}+mk,
\end{align*}
where the first step follows from splitting the regret from before to two sums, the second step follows since in the first sum $\Delta_a<\gamma$ and since there are only $T$ samples in total throughout all arms, and in the second sum we apply Equation~\eqref{eq:RaBound}, and the final step follows since there are $k$ arms in total and since the elements in the sum satisfy $\Delta_a\in [\gamma,1]$.
We balance the first two terms by defining $\gamma$ to be $\gamma = \sqrt{\frac{(128+512\sigma_{\varepsilon,\delta}^2/m)k\log T}{T}}$, so the total regret is:
$R\leq 2\sqrt{(128+512\sigma_{\varepsilon,\delta}^2/m)k\cdot T\log T}+mk $.
By a similar argument to Equation~\eqref{eq:expectedRegret} conditioning on whether or not $C$ occurred, we get that the expected regret $R(T)$ satisfies
$$R(T)\leq 2\sqrt{(128+512\sigma_{\varepsilon,\delta}^2/m)k\cdot T\log T}+mk+T\cdot 4T^{-2}=O\left(\sqrt{\left(1+\frac{\sigma_{\varepsilon,\delta}^2}{m}\right)kT\log T}+mk\right).$$
\end{proof}
Recall that the private binary summation mechanism $M_{sum}$'s error distribution is sub-Gaussian with variance $\sigma_{\varepsilon,\delta}^2=O\left(\frac{\log \frac{1}{\delta}}{\varepsilon^2}\right)$. We fix a concrete batch size $m$ in \mbox{SDP-AE}\xspace, and apply Theorem~\ref{thm:batchAERegret} to get the following corollary.
\begin{corollary}\label{cor:batchedAESDPRegret}
\mbox{SDP-AE}\xspace with a batch size of $m=\ceil{\sigma_{\varepsilon,\delta}}=\Theta\left(\frac{\log\frac{1}{\delta}}{\varepsilon^2}\right)$ is $(\varepsilon,\delta)$-SDP and has a distribution-dependent regret of $$O\left(\sum_{a\in [k]:\Delta_a>0}\left(\frac{\log T}{\Delta_a}+\frac{\Delta_a\log\frac{1}{\delta}}{\varepsilon^2}\right)\right)=O\left(\left(\sum_{a\in [k]:\Delta_a>0}\frac{\log T}{\Delta_a}\right)+\frac{k\log\frac{1}{\delta}}{\varepsilon^2}\right),$$ and a distribution-independent regret of $O\left(\sqrt{kT\log T}+\frac{k\log\frac{1}{\delta}}{\varepsilon^2}\right)$.
\end{corollary}
\subsection{\mbox{VB-SDP-AE}\xspace: Variable Batch Shuffle Differentially Private Arm Elimination}\label{sec:optimalAlg}
In this section we modify \mbox{SDP-AE}\xspace to give an $(\varepsilon,\delta)$-SDP algorithm for the MAB problem with improved additional regret.
Recall that \mbox{SDP-AE}\xspace has a distribution-dependent regret of $O\left(\left(\sum_{a\in [k]:\Delta_a>0}\frac{\log T}{\Delta_a}\right)+\frac{k\log\frac{1}{\delta}}{\varepsilon^2}\right)$, whereas the best known regret for $(\varepsilon,\delta)$ centralized-DP is due to Tossou and Dimitrakakis~\cite{tossou2016algorithms}, and is $O\left(\left(\sum_{a\in [k]:\Delta_a>0} \frac{\log T}{\Delta_a}\right)+\frac{k}{\varepsilon}\right)$. Since any SDP algorithm can be emulated by a centralized-DP mechanism, we can only wish to match the regret of Tossou and Dimitrakakis~\cite{tossou2016algorithms}. Hence, the natural question is: \enquote{Can we reduce the dependence in $\varepsilon$ of the additive (second) regret term from $\frac{1}{\varepsilon^2}$ to $\frac{1}{\varepsilon}$?}
Before presenting our solution, we first indicate two intuitive approaches that fail. The first is to continue using batches of constant size $m=\Omega(1/\varepsilon^2)$. This idea fails since intuitively in the worst case we can expect a first batch regret of $m=\Omega(1/\varepsilon^2)$ which already surpasses $O(1/\varepsilon)$.
The second is to try to directly adapt the algorithm of Tossou and Dimitrakakis~\cite{tossou2016algorithms}, which can be interpreted as \AE with batches of size $\ceil{1/\varepsilon}$. Unfortunately, this adaptation fails since it requires that the added noise to the empirical mean of each arm decreases with time, whereas in the shuffle model, we must ensure privacy with respect to each batch equally and independently, so the total noise cannot decrease with time.
Now for our improved algorithm, observe that using a large batch size increases the regret a lot for arms which are very suboptimal. This is since rather than pulling these arms only a few times until we detect that they are suboptimal, we commit ourselves to pulling them throughout a large batch. On the other hand, fixing the batch size to be small increases the overall estimation error, since every application of the SDP summation algorithm $M_{sum}$ introduces an error which is of the same magnitude (specifically, the error is sub-Gaussian with variance $\sigma_{\varepsilon,\delta}^2=O\left(\frac{\log \frac{1}{\delta}}{\varepsilon^2}\right)$)
independently of the batch size. Hence, more executions of $M_{sum}$ translates to more noise due to privacy. We therefore extend \mbox{SDP-AE}\xspace to support variable size batches, which start small and gradually increase. Intuitively, the smaller batches initially, allow us to quickly eliminate very suboptimal arms with only a small number of pulls. We gradually increase the batch size to reduce the per-user error introduced by the private binary summation mechanism.
Specifically, we double the batch size after each phase.
\begin{algorithm}[H]\label{alg:variablebatchae}
\SetAlgoLined
\textbf{Input:} privacy parameters $\varepsilon$ and $\delta$ and horizon $T$.\\
\textbf{Initialize:} $\hat{S}_a^0=0$, $N_a^0=0$ and $\hat{\mu}_{a}^0=0$ for every $a\in [k]$\;
Let $\sigma_{\varepsilon,\delta}^2=O\left(\frac{\log \frac{1}{\delta}}{\varepsilon^2}\right)$ be the sub-Gaussian variance of the error distribution of $M_{sum}$\;
Let $V=[k]$ denote the set of viable arms\;
\For{phase $t\gets1,2,\ldots$}{
Let $m^t\gets2^t$\;
\For{arm $a\in V$}{
\For{each new user $i\gets1$ \KwTo $m^t$}{
User $i$ pulls the arm $a$ and observes reward $r_{a,i}^t$\;
If total number of arm samples is $T$, exit\;
}
\textbf{Communication:} Perform private binary summation
$Z_a^t \gets M_{sum}\left(\{r_{a,i}^t\}_{i=1}^{m^t}\right)$\;
\textbf{Server update:} Update $\hat{S}_{a}^{t} \gets \hat{S}_{a}^{t-1} + Z_a^t $,
$N_{a}^{t} \gets N_{a}^{t-1} + m^t$, and finally $\hat{\mu}_{a}^{t} \gets \hat{S}_{a}^{t}/N_{a}^{t}$\;
}
\textbf{Confidence bounds:} For each arm $a$, calculate $I_a^t\gets\left(\frac{2\sqrt{t}\sigma_{\varepsilon,\delta}}{N_a^t}+\frac{1}{\sqrt{N_a^t}}\right)\cdot\sqrt{2\log T}$, and the upper and lower confidence bounds $UCB_a^t \gets \hat{\mu}_{a}^{t}+I_a^t$ and $LCB_a^t \gets \hat{\mu}_{a}^{t}-I_a^t$\;
\textbf{Elimination:} remove all arms $a$ from $V$ such that $UCB_a^t < \max_{a'\in S}LCB_{a'}^t$\;
}
\caption{\mbox{VB-SDP-AE}\xspace (Variable Batch Shuffle Differentially Private Arm Elimination)}
\end{algorithm}
\subsubsection{Algorithm outline}
Algorithm~\ref{alg:variablebatchae} above, which we call \textit{Variable Batch Shuffle Differentially Private Arm Elimination (\mbox{VB-SDP-AE}\xspace)}, consolidates the previous presentation using a different batch size $m^t=2^t$ for each phase $t$.
\subsubsection{Analysis}
The privacy is trivial, since in each batch we use the $(\varepsilon,\delta)$-SDP mechanism $M_{sum}$. We now focus on regret.
To give a regret bound on \mbox{VB-SDP-AE}\xspace, we follow a similar proof to that of Theorem~\ref{thm:batchAERegret}.
\begin{theorem}\label{thm:variablebatchAERegret}
The algorithm \mbox{VB-SDP-AE}\xspace is $(\varepsilon,\delta)$-SDP, and has a distribution-dependent regret of $O\left(\left(\sum_{a\in [k]:\Delta_a>0}\frac{\log T}{\Delta_a}\right)+k\sigma_{\varepsilon,\delta}\log T\right)$, and a distribution-independent regret of $O\left(\sqrt{kT\log T}+k\sigma_{\varepsilon,\delta}\log T\right)$.
\end{theorem}
\begin{proof}
We continue identically to the proof of Theorem~\ref{thm:batchAERegret}, except the fact that we need the $t$'th index of each arm's private-binary-summation-error tape to contain an iid sample of the error of the private binary summation mechanism $M_{sum}$ for $m^t=2^t$ users. By the definition of $M_{sum}$, which is sub-Gaussian with the same variance for any number of users (batch size), the application of Hoeffding inequality as in the proof of Theorem~\ref{thm:batchAERegret} still follows. We conclude that the \textit{clean event} $C:=\left\{\forall a\in[k], \forall t\in[T]~~\abs{\hat{\mu}^t_a-\mu_a}\leq I_a^t\right\},$ where $I_a^t:=\left(\frac{2\sqrt{t}\sigma_{\varepsilon,\delta}}{N_a^t}+\frac{1}{\sqrt{N_a^t}}\right)\cdot \sqrt{2\log T}$ occurs with high probability, that is $P\left(C\right)\geq1-4T^{-2}$.
For the regret analysis, we assume the clean event $C$. Let $a$ be a suboptimal arm, and let $t_0$ be the last phase following which we did not remove the arm $a$ yet (or the last phase if $a$ remains active to the end). As in the proof of Theorem~\ref{thm:batchAERegret}, we get that
\begin{align}
\Delta_a\leq \left(\frac{8\sqrt{t_0}\sigma_{\varepsilon,\delta}}{N_a^{t_0}}+\frac{4}{\sqrt{N_a^{t_0}}}\right)\cdot \sqrt{2\log T}.\label{eq:variabledeltaaBound}
\end{align}
We now diverge from the proof of Theorem~\ref{thm:batchAERegret}.
Observe that if both $N_a^{t_0}> \frac{128\log T}{\Delta_a^2}$ and $N_a^{t_0}>\frac{16\sqrt{{t_0}}\sigma_{\varepsilon,\delta}\cdot \sqrt{2\log T}}{\Delta_a}$, then we get a contradiction to Equation~\eqref{eq:variabledeltaaBound} since $\frac{4\cdot \sqrt{2\log T}}{\sqrt{N_a^{t_0}}}<\frac{\Delta_a}{2}$ and $\frac{8\sqrt{{t_0}}\sigma_{\varepsilon,\delta}\cdot \sqrt{2\log T}}{N_a^{t_0}}< \frac{\Delta_a}{2}$ respectively. Hence, $N_a^{t_0}\leq \max\left( \frac{128\log T}{\Delta_a^2},\frac{16\sqrt{{t_0}}\sigma_{\varepsilon,\delta}\cdot \sqrt{2\log T}}{\Delta_a}\right)$.
Therefore the total regret on arm $a$ is
\begin{align}
R_a&\leq 4\Delta_aN_a^{t_0}\leq \max\left( \frac{512\log T}{\Delta_a},64\sqrt{{t_0}}\sigma_{\varepsilon,\delta}\cdot \sqrt{2\log T}\right)\nonumber\\
&\leq \frac{512\log T}{\Delta_a}+64\sqrt{2}\sigma_{\varepsilon,\delta}\cdot \log T,\label{eq:variableRaBound}
\end{align}
where the first step follows since arm $a$ is eliminated following phase ${t_0}+1$ (or if $t_0$ is the last phase, then we finish and don't sample $a$ after it) with batch size $2^{t_0+1}=2\cdot 2^{t_0}\leq 3\cdot N_a^{t_0}$, and the third step follows since $m^{t_0}=2^{t_0}$ and ${t_0}\geq 1$, so $T\geq N_a^t=\sum_{s=1}^{{t_0}}m^s=2^{t_0+1}-2\geq 2^{t_0}$ and therefore ${t_0}\leq \log T$.
Similarly to the proof of Theorem~\ref{thm:batchAERegret} which uses Equation~\eqref{eq:RaBound} to get the distribution-dependent bound, here we use the analogous Equation~\eqref{eq:variableRaBound} to conclude that the distribution-dependent bound is
\begin{align*}
R(T)&=O\left(\sum_{a\in [k]:\Delta_a>0}\left(\frac{\log T}{\Delta_a}+\sigma_{\varepsilon,\delta}\log T\right)\right)=O\left(\left(\sum_{a\in [k]:\Delta_a>0}\frac{\log T}{\Delta_a}\right)+k\sigma_{\varepsilon,\delta}\log T\right).
\end{align*}
Now for the distribution-independent bound, similarly to the proof of Theorem~\ref{thm:batchAERegret}, assuming the clean event $C$, for any $\gamma>0$ it holds that the total regret
$$R\leq T\cdot \gamma + \frac{512k\log T}{\gamma}+64\sqrt{2}k\sigma_{\varepsilon,\delta}\cdot \log T,$$
and specifically for $\gamma = \sqrt{\frac{512k\log T}{T}}$, the total regret
$R\leq \sqrt{2048k\cdot T\log T}+64\sqrt{2}k\sigma_{\varepsilon,\delta}\cdot \log T $.
Similarly to the argument in Equation~\eqref{eq:expectedRegret}, conditioning on whether the clean event $C$ occurred or not, we conclude that the expected regret $R(T)$ satisfies
$$R(T)\leq \sqrt{2048k\cdot T\log T}+64\sqrt{2}k\sigma_{\varepsilon,\delta}\cdot \log T+T\cdot 4T^{-2}=O\left(\sqrt{kT\log T}+k\sigma_{\varepsilon,\delta}\log T\right).$$
\end{proof}
We recall that the private binary summation mechanism $M_{sum}$'s error distribution is sub-Gaussian with variance $\sigma_{\varepsilon,\delta}^2=O\left(\frac{\log \frac{1}{\delta}}{\varepsilon^2}\right)$, i.e., $\sigma_{\varepsilon,\delta}=O\left(\frac{\sqrt{\log \frac{1}{\delta}}}{\varepsilon}\right)$, and apply Theorem~\ref{thm:variablebatchAERegret} to get the following corollary
\begin{corollary}\label{cor:variablebatchedAESDPRegret}
\mbox{VB-SDP-AE}\xspace is $(\varepsilon,\delta)$-SDP and has a distribution-dependent regret of $O\left(\left(\sum_{a\in [k]:\Delta_a>0}\frac{\log T}{\Delta_a}\right)+\frac{k\sqrt{\log\frac{1}{\delta}}\log T}{\varepsilon}\right)$, and a distribution-independent regret of $O\left(\sqrt{kT\log T}+\frac{k\sqrt{\log\frac{1}{\delta}}\log T}{\varepsilon}\right)$.
\end{corollary}
\section{Conclusion and future work}
In this paper, we gave and analyzed differentially private algorithms for the MAB problem, closing the inevitable multiplicative $\Omega(1/\varepsilon^2)$ regret gap between the local model and the centralized model, by considering the (intermediate) shuffle model. Our algorithms are batched variants of \AE, which use a private binary summation mechanism for the shuffle model as a building block. Compared to the non-private \AE algorithm's regret, our first algorithm \mbox{SDP-AE}\xspace has an additive factor of $\frac{k\log \frac{1}{\delta}}{\varepsilon^2}$ using constant size batches, and our second algorithm \mbox{VB-SDP-AE}\xspace improves the additive factor to $\frac{k\sqrt{\log \frac{1}{\delta}}\log T}{\varepsilon}$ by using exponentially growing batches, which enable the early detection and elimination of very suboptimal arms.
A natural future work is to extend our results (i.e., the usage of a private binary summation mechanism for the shuffle model) to more general RL settings such as linear/contextual bandits or Markov decision processes. It would also be interesting to study whether our $\log T$ term in the additional additive regret factor can be shaved through a more sophisticated algorithm, or an alternative analysis.
\section*{Disclosure of Funding}
This work is partially supported by Israel Science Foundation (grants 993/17,1595/19,1871/19), German-Israeli Foundation (grant 1367/2017), Len Blavatnik and the Blavatnik Family Foundation, the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation program (grant agreement 882396), the Yandex Initiative for Machine Learning at Tel Aviv University.
\bibliographystyle{abbrv}
\newpage
|
\section{Introduction}
\label{sec:intro}
A huge experimental programme on accelerator-based neutrino oscillation experiments has been developed in recent years with the scope of improving our knowledge of neutrino properties and of potentially answering one of the fundamental open questions of modern physics, that is the origin of the matter/antimatter asymmetry in the universe, by measuring the weak CP-violating phase~\cite{Abe:2019vii}.
The success of this program partly relies on the control of systematic errors, largely due to uncertainties in the description of the neutrino interactions with the detector, typically made of medium-weight nuclei like carbon, oxygen or argon.
With this motivation, intense theoretical activity has been carried out in parallel in order to provide an accurate description of neutrino-nucleus reactions
in the GeV region, relevant for the experiments, for different processes ranging from quasielastic up to deep inelastic scattering, encompassing the excitation of nucleon resonances and the emission of two or more nucleons (see \cite{Alvarez-Ruso:2017oui} for a comprehensive review of recent progresses and open challenges in the field).
In this article we review our work on the SuSA (Super-Scaling Approach) model, originally developed in Ref.~\cite{Amaro:2004bs} and subsequently refined and improved to the updated version SuSAv2~\cite{Gonzalez-Jimenez:2014eqa,Megias:2016fjk} (see also Refs.~\cite{Amaro20,Barbaro:2021psv} for other recent reviews).
The original SuSA, as will be explained in more detail in Sect.~\ref{sec:scal}, is a phenomenological model which, while retaining the relativistic aspect of the Relativistic Fermi Gas (RFG), provides by construction - unlike the RFG - a good description of inclusive electron scattering data $(e,e')$ in the quasielastic region. In the model initial and final state interaction effects, absent in the RFG, are directly extracted from $(e,e')$ data.
The updated version, SuSAv2, implements inputs from the relativistic mean field model (RMF), which provides a microscopic interpretation of the basic features of SuSA and also includes a more detailed description of the different spin and isospin channels.
Moreover, the model has been extended to include two-body currents, able to excite two-particle-two-hole (2p2h) states; these, as will be shown in Sect.~\ref{sec:results}, play an important role in the analysis of neutrino-nucleus experimental data.
The scheme of this review is the following. In Sect.~\ref{sec:incl} we focus on the inclusive reaction, presenting the formalism, describing the model and testing it with electron scattering data.
In Sect.~\ref{sec:semi_electrons} we discuss the semi-inclusive reaction an its relation to the inclusive one. In Sect.~\ref{sec:results} we show some selected results and discuss the comparison with experimental data. Finally, in Sect.~\ref{sec:concl} we draw our conclusion and outline the future developments of our work.
\section{Inclusive scattering}
\label{sec:incl}
\subsection{General formalism}
\label{sec:form}
In this review we mainly focus on the charged current (CC) inclusive process
\begin{equation}
\nu_l (\overline\nu_l)+A \longrightarrow l^-(l^+) + X \,,
\end{equation}
where a neutrino (antineutrino) of a given flavour $l$ having four-momentum $K^\mu=(\varepsilon,{\bf k})$ hits a nucleus $A$ and a charged lepton $l^-$ ($l^+$) is detected in the final state with four-momentum $K^{\prime\mu}=(\varepsilon^\prime,{\bf k}^\prime)$ and scattering angle $\theta$, while the residual system $X$ is unobserved. The four-momentum transferred from the probe to the nucleus is $Q^\mu=(\omega,{\bf q})$, with $\omega=\varepsilon-\varepsilon^\prime$ and ${\bf q}={\bf k}-{\bf k^\prime}$. We shall assume the neutrino to be massless, while the outgoing lepton has finite mass $m_l$.
The corresponding cross section is obtained by contracting the leptonic and hadronic tensors
\begin{eqnarray}
l_{\mu\nu}({\bf q},\omega) = 2\left(K_\mu K^\prime_\nu+K^\prime_\mu K_\nu- K\cdot K^\prime g_{\mu\nu}+ i \chi \epsilon_{\mu\nu\rho\sigma}K^\rho K^{\prime\sigma}\right)
\label{eq:lmunu}
\\
W^{\mu\nu}({\bf q},\omega) = \sum_n <A|J^{\mu\dagger}({\bf q},\omega)|n><n|J^\nu({\bf q},\omega)|A> \,\delta(\omega+E_A-E_n) ,
\label{eq:Wmunu}
\end{eqnarray}
where $\chi=+1 (-1)$ in the neutrino (antineutrino) case, $|A>$ is the initial nuclear ground state and $|n>$ are all the intermediate hadronic states that can be reached through the weak current operator $J^\mu$.
The double differential cross section with respect to the momentum and scattering angle of the outgoing lepton can be expressed
in terms of five response functions~\cite{Amaro:2004bs} (for the semi-inclusive reaction $(\nu_l,l^-N)$, which will be briefly discussed in Sect.~\ref{sec:semi_electrons}, the response functions are ten)
\begin{equation}
\frac{d^2\sigma}{d\varepsilon^\prime d\cos\theta}
=
\frac{(G_F\cos\theta_c)^2}{4\pi}
v_0 \frac{k^\prime}{\varepsilon^\prime}
\left(
V_{CC} R_{CC}+
2{V}_{CL} R_{CL}
+{V}_{LL} R_{LL}+
{V}_{T} R_{T}
+
2\chi{V}_{T'} R_{T'}
\right),
\label{eq:cs}
\end{equation}
where $G_F$ the Fermi weak constant,
$\theta_c$ the Cabibbo angle and $v_0= (\varepsilon+\varepsilon^\prime)^2-{\bf q}^2$. The indices $C$, $L$, $T$ refer to the Coulomb, longitudinal and transverse components of the leptonic and hadronic currents
with respect to ${\bf q}$.
The leptonic coefficients $V_K$ are related to the components of the tensor \eqref{eq:lmunu} and are given by
\begin{eqnarray}
&&V_{CC } = 1-\delta^2 \frac{|Q^2|}{v_0},\
V_{CL } = \frac{\omega}{q}+\frac{\delta^2}{\rho^\prime} \frac{|Q^2|}{v_0},\
V_{LL } = \frac{\omega^2}{q^2}+ \delta^2 \frac{|Q^2|}{v_0}
\left(1+\frac{2\omega}{q\rho^\prime}
+\rho\delta^2\right),\nonumber\\
&&V_{T } = \frac{|Q^2|}{v_0}
+\frac{\rho}{2}-\delta^2\frac{|Q^2|}{v_0}
\left(\frac{\omega}{q\rho^\prime}+\frac{\rho\delta^2}{2}\right) ,\
V_{T' } = \frac{|Q^2|}{v_0}
\left(\frac{1}{\rho^\prime}-\frac{\omega\delta^2}{q}\right) ,
\end{eqnarray}
where the variables $\delta^2=\frac{m_l^2}{|Q^2|}$, $\rho=\frac{|Q^2|}{q^2}$ and $\rho^\prime=\frac{q}{\varepsilon+\varepsilon^\prime}$ have been introduced.
The response functions
\begin{equation}
R_{CC } = W^{00 },\
R_{CL } = -\frac{1}{2}\left(W^{03 }+W^{30}\right),\
R_{LL } = W^{33 },\
R_{T } = W^{11 } + W^{22 },\
R_{T' } = -\frac{i}{2}\left(W^{12 }-W^{21}\right)
\end{equation}
embody the nuclear dynamics and are specific components of the hadronic tensor $W^{\mu\nu}$, depending only upon $\omega$ and $\bf q$ .
The above decomposition into response functions is valid for all reaction channels - elastic, quasielastic, inelastic - each of which characterized by a
different current operator $J^\mu$ in the hadronic tensor \eqref{eq:Wmunu}. In the case of quasielastic scattering, corresponding to the interaction of the probe with a single nucleon, the weak current operator is
\begin{equation}
J^\mu({\bf q},\omega)=F_1(Q^2)\gamma^\mu+\frac{iF_2(Q^2)}{2m_N}\sigma^{\mu\nu}Q_\nu-G_A(Q^2)\gamma^\mu\gamma_5-\frac{G_P(Q^2)}{m_N}Q^\mu\gamma_5,
\end{equation}
where $F_1$, $F_2$, $G_A$ and $G_P$ are the Pauli, Dirac, axial and pseudoscalar weak form factors, respectively.
In neutrino oscillation experiments the incident beam is not monochromatic. As a consequence, before comparing with experimental data, the cross
section \eqref{eq:cs} must be folded with the normalized neutrino flux $\phi(\varepsilon)$ $(\int d\varepsilon \,\phi(\varepsilon)=1)$:
\begin{equation}
\langle \frac{d^2\sigma}{d\varepsilon^\prime d\cos\theta} \rangle =
\int d\varepsilon \,\phi(\varepsilon) \,\frac{d^2\sigma}{d\varepsilon^\prime d\cos\theta}.
\end{equation}
\subsection{Scaling and SuSAv2}
\label{sec:scal}
The phenomenon of scaling is sometimes found to occur in different fields, including solid state, atomic, molecular, nuclear and hadronic physics, when an interacting probe scatters from composite many-body systems. Under some circumstances it is found that the response of the complex system no longer depends on two independent variables, but only on a particular combination of those, called the scaling variable. This phenomenon is very well known in high energy physics where the inelastic nucleon response functions are shown to depend only on the Bjorken variable $x$~\cite{bjorkxy}. A similar phenomenon is observed in lepton-nucleus scattering, where the nuclear response functions are found to scale with a single scaling variable, denoted as $y$. This indicates that the probe (lepton) interacts with the nucleus' constituents, in this case with the nucleons in the nucleus.
The phenomenon of $y$-scaling emerges from the analysis of quasielastic (QE)
$(e,e')$ reactions and has been studied in detail in \cite{Day90,DS199,DS299}. The scaling function is defined as the QE $(e,e')$ differential cross section divided by a single-nucleon cross section averaged over the Fermi gas.
For high enough values of the momentum transfer, $q$, the scaling function does only depend
on a single variable, $y$, given as a particular combination of the energy ($\omega$) and momentum ($q$) transferred in the process.
The scaling variable $y$ is (up to a sign) the minimum value of the missing momentum allowed by
kinematics~\cite{Day90,DS199,DS299}.
In the QE domain the basic mechanism in $(e,e')$ reactions on nuclei corresponds to elastic scattering from
individual nucleons in the nuclear medium. This implies that the inclusive $(e,e')$ cross section is mainly constructed from the
exclusive $(e,e'N)$ process, including the contribution of all nucleons in the target and integrating over all (unobserved) ejected
nucleon variables (see also the following section). This approach constitutes the basis of the Impulse Approximation (IA). Thus, the double differential ($e,e'$) inclusive cross section is given as the sum of two response functions corresponding
to the longitudinal, $L$, and transverse, $T$, channels,
\begin{equation}\label{ecsec}
\displaystyle \frac{d^2\sigma}{d\Omega_e d\omega}=\sigma_{Mott}\left[v_LR_L(q,\omega)+v_TR_T(q,\omega)\right],
\end{equation}
where $\sigma_{Mott}$ is the Mott cross section and the $v_{L,T}$
are kinematical factors that involve leptonic variables (see~\cite{Amaro20,Kel96} for explicit expressions).
In terms of the scaling functions the nuclear responses are
\begin{equation}\label{eresp}
\displaystyle R_{L,T} (q,\omega)=\frac{1}{k_F}\Bigl[f_{L,T}(q,\omega)G_{L,T}(q,\omega)\Bigr] ,
\end{equation}
where $k_F$ is the Fermi momentum and $G_{L,T}$ are defined as the
responses of a moving nucleon and include relativistic corrections arising from the presence of the medium.
Their explicit expressions can be found in \cite{Gonzalez-Jimenez:2014eqa,Amaro20,Amaro07}. The terms $f_{L,T}$
are the scaling functions that show a mild dependence
upon the momentum transfer $q$ (first-kind scaling) and very weak dependence on the nuclear
system considered (second-kind scaling). The occurrence of both types of scaling is denoted as superscaling.
Hence to the extent that at some kinematics the above $f$ functions are the same for all nuclei and do only depend on a
single variable, denoted as the scaling variable $\psi(q,\omega)$, but not separately on $q$, one says that superscaling occurs.
The superscaling variable $\psi$ is given in terms of $q$ and $\omega$ as
\begin{equation}
\psi = \frac{1}{\sqrt{\xi_F} } \frac{\lambda-\tau}{\sqrt{(1+\lambda)\tau+\kappa\sqrt{\tau(1+\tau)}}} \, ,
\label{eq:psi}
\end{equation}
where the dimensionless Fermi kinetic energy $\xi_F=\frac{T_F}{m_N}$, energy transfer $\lambda=\frac{\omega}{2m_N}$, momentum transfer
$\kappa=\frac{q}{2m_N}$ and squared four-momentum transfer $\tau=\kappa^2-\lambda^2$ have been introduced.
Here the $y$-variable introduced above is given approximately by
\begin{equation}
y \cong k_F \psi
\label{ytopsi}
\end{equation}
as discussed in \cite{DS299}.
A phenomenological energy shift $E_{shift}$, fitted for each nucleus to the experimental position of the quasielastic peak~\cite{MDS02}, is included in the definition of the scaling variable $\psi'\equiv \psi(q,\omega-E_{shift})$.
Scaling and superscaling properties of electron-nucleus interactions have been analyzed in detail in a series of previous
works~\cite{Day90,DS199,DS299,MDS02,Barbaro:1998gu,PRL05,jac06} . The importance of this phenomenon to test the validity of any nuclear model aiming to describe electron scattering reactions has been clearly proven. The model, denoted as the Superscaling Approach (SuSA), is entirely based on the phenomenology, making use of a unique, universal, scaling function extracted from the analysis of the longitudinal electron scattering data (see Fig.~\ref{fig:scaling}). Notice that the behavior and properties of the experimental superscaling function constitute a
strong constraint for any theoretical model describing QE electron
scattering. Not only should the superscaling behavior be fulfilled, but also the specific shape of the longitudinal scaling function, $f^{exp}_L$, must be reproduced. The SuSA model assumes the longitudinal phenomenological scaling function to be valid also in describing the transverse
channel, {\it i.e.,} $f_L=f_T$.
\begin{figure}[htbp]
\centering
\includegraphics[width=.325\textwidth,angle=270]{Figures/susarmf.ps}\hspace*{0.25cm}
\includegraphics[width=.325\textwidth,angle=270]{Figures/fL_rmf_vs_data.ps}
\caption{{\it Left panel:} Phenomenological SuSA scaling function versus $\psi'$ in the QE region together with ($e,e'$) longitudinal scaling data from~\cite{Jourdan}. The RFG scaling function is also shown as reference.
{\it Right panel:} RMF longitudinal scaling functions for ($e,e'$) at different $q$ values compared with ($e,e'$) longitudinal scaling data from~\cite{Jourdan}. In the inner smaller plot a reduced-$\chi^2$ analyses shows a minimum at $q=650$ MeV/c. Figure adapted from~\cite{Gonzalez-Jimenez:2014eqa}.}\label{fig:scaling}
\end{figure}\vspace*{-0.15cm}
In recent years we have explored in detail the extension of SuSA to neutrino-nucleus scattering. Due to the complexity of the weak process, with an increased number of nuclear responses compared with the pure electromagnetic one, and the particular role played by the axial term in the weak current, we have developed an improved version of the superscaling model, called SuSAv2~\cite{Gonzalez-Jimenez:2014eqa}.
Contrary to the original SuSA~\cite{Amaro:2004bs,Amaro07,Amaro:2006tf} where a universal scaling function based on electron scattering data is used, the new SuSAv2 model incorporates relativistic mean field (RMF) effects~\cite{PRL05,jac06} in the longitudinal and
transverse nuclear responses, as well as in the isovector and isoscalar channels that is of great importance in order to describe charged-current (CC) neutrino reactions that are purely isovector~\cite{Caballero:2007tz}.
The origin of the SuSAv2 approach is based on the capability of the RMF to describe properly the scaling behavior of the electron scattering data. As shown in previous works~\cite{PRL05,jac06}, RMF is one of the few microscopic models capable of reproducing the asymmetric shape of the phenomenological scaling function with a long tail extended to high values of the transfer energy (large values of $\psi^\prime$). Moreover, RMF produces an enhancement in the transverse scaling function, a genuine dynamical relativistic effect linked to the lower components in the wave functions, that is supported by the analysis of data. The RMF framework to finite nuclei has proven to successfully reproduce the scaling behavior shown by data at low to intermediate $q$ values (see Fig.~\ref{fig:scaling}). However, the model clearly fails at higher momentum transfers where Final State Interactions (FSI) are expected to be weaker. This is due to the RMF strong energy-independent scalar and vector potentials used in the final state that lead to too much asymmetry in the scaling functions and shift the QE peak to very high transfer energies, in clear disagreement with data. To remedy this shortfall of the RMF model, the SuSAv2 incorporates both the RMF scaling functions at low-to-intermediate $q$ values and the Relativistic Plane Wave Impulse Approximation (RPWIA) ones at higher $q$ by using a $q$-dependent blending function that smoothly connects the two regimes (see~\cite{Megias:2016fjk,Megias:2016lke} for details). A similar solution to this drawback of the RMF model has been taken in the recent Energy-Dependent RMF (ED-RMF) approach~\cite{Gonzalez-Jimenez_edRMF,Gonzalez-Jimenez_edRMF2} where RMF potentials are multiplied by a blending function inspired by the SuSAv2 one that scales them down as the kinetic energy of the scattered nucleon increases, also preventing non-orthogonality issues. This model predicts both lepton and nucleon kinematics, showing a similar agreement on electron and neutrino data with SuSAv2.
In summary, the SuSAv2 model and, for extension, the ED-RMF one reproduce the experimental longitudinal scaling data, gives rise to an enhancement in the electromagnetic transverse channel, {\it i.e.}, $f_T^{(e,e')} > f_L^{(e,e')}$, takes into account the differences in the isoscalar/isovector scaling functions, of crucial interest for neutrino scattering processes, and finally avoids the problems of the RMF model in the region of high momentum transfer, where FSI effects are negligible. One of the basic merits of SuSAv2 is the translation of sophisticated and demanding microscopic calculations into relatively straightforward parametrizations and, hence, easing its implementation in the MonteCarlo simulations employed in the analysis of neutrino oscillation experiments.
To conclude, it is important to point out that SuSAv2 is not restricted to the QE kinematics domain. On the contrary, it can be extended to the inelastic region by generalizing the superscaling variable \eqref{eq:psi} to the excitation of any inelastic state having invariant mass $W$. One can thus define~\cite{Barbaro:2003ie}
\begin{equation}
\psi_W = \frac{1}{\sqrt{\xi_F} } \frac{\lambda-\tau\rho_W}{\sqrt{(1+\lambda\rho_W)\tau+\kappa\sqrt{\tau(1+\tau\rho_W^2)}}}
\,,\ \ \rho_W=1+\frac{1}{4\tau}\left(W^2/m_N^2-1\right)
\end{equation}
and evaluate the inelastic hadronic tensor as the integral
\begin{equation}
W^{\mu\nu}_{inel} = \frac{1}{m_N^2} \int_{W_{\rm min}}^{W_{\rm max}} dW W f(\psi_W)\, w^{\mu\nu}_{inel}
\end{equation}
between the limits $W_{\rm min}, W_{\rm max}$ imposed by the kinematics, where $f(\psi_W)$ is the SuSAv2 superscaling function previously described.
This procedure is based on the assumption that the nuclear effects, encoded in the function $f$, are the same in the QE and inelastic regions and requires the knowledge of the elementary tensor $w^{\mu\nu}_{inel}$ across the full inelastic spectrum.
\begin{figure}[htbp]
\centering
\includegraphics[width=.345\textwidth,angle=270]{Figures/680_60_final_inelastic2.ps}\hspace*{0.25cm}
\includegraphics[width=.345\textwidth,angle=270]{Figures/1108_37p5_inelastic.ps}
\caption{
Comparison of inclusive $^{12}$C($e,e'$) double differential cross sections and predictions for the inelastic regime of the Bosted-Christy parametrization~\cite{Bosted:2007xd,Christy:2007ve} (dot-dashed lines), Bodek-Ritchie parametrization~\cite{BodekRitchie1,BodekRitchie2} (solid lines) and GRV98 PDFs~\cite{GRV98_1,GRV98_2} (dashed lines) at different kinematics (incident electron beam and scattering angle) in terms of the energy transferred to the nucleus ($\omega$). Experimental data taken from~\cite{QESarchive,QESarxiv}. The y-axis represents $d^2\sigma/d\Omega/d\omega$ in nb/GeV/sr. The value of $q$ at the QE peak ($q_{QE}$) is shown as reference.}\label{fig:susav2inelastic}
\end{figure
The SuSAv2 inelastic model has been applied to the electron scattering case using phenomenological fits of the single-nucleon inelastic structure functions $w_1$ and $w_2$ extracted from e-p and e-d data~\cite{Bosted:2007xd,Christy:2007ve}. In Fig.~\ref{fig:susav2inelastic}, a comparison of the SuSAv2-inelastic model with $^{12}$C($e,e'$) data using different inelastic structure functions shows a preference for the Bosted-Christy parametrization~\cite{Bosted:2007xd,Christy:2007ve}.
In general, the SuSAv2 model (QE+inelastic) provides a very good description of data for very different kinematics once Meson Exchange Currents (MEC) are also incorporated, denoted as SuSAv2-MEC (see Sect.~\ref{sec:electron} and \cite{Megias:2016fjk,Megias:2017cuh,Megias:2018ujz}).
The inclusion of the full inelastic spectrum in the SuSAv2 model for weak interactions is still in progress.
An alternative approach to the study of the resonant pion production was taken in Refs.~\cite{Amaro:2004bs,Maieron:2009an,Ivanov16}, where a scaling function to be used in the $\Delta$ resonance region, different from the quasielastic one, was extracted from electron scattering data and multiplied by the appropriate weak $N\to\Delta$ transition form factors to get the neutrino-nucleus cross section in this region. This method provides a phenomenological description valid at transferred energies below the $\Delta$ peak, while at higher $\omega$ it fails due to the opening of other inelastic channels. Some results corresponding to this method, referred to as SuSA-$\Delta$ approach, will be shown in Sect.~\ref{sec:results}.
\subsection{Relativistic model for CC MEC and 2p2h responses}
\label{sec:2p2h}
Multinucleon knockout processes give a non-negligible contribution to
the inclusive neutrino cross section for the intermediate energies
involved in the experiments \cite{Megias:2016fjk,Mar09,Nieves:2011pp,Rocco:2018mwt,Sob20}.
In Ref. \cite{Sim16} we developed a model
of two-particle two-hole excitations of the RFG induced by weak
meson-exchange currents for inclusive
CC neutrino scattering. The model is fully relativistic and includes
the diagrams of Fig. \ref{fig_feynman}, involving one-pion exchange
and $\Delta$ excitation, taken from the pion production model of
\cite{Her07}.
The 2p2h matrix element of MEC
depends on the momenta, spin and isospin coordinates
$(1',2';1,2) \equiv ({\bf p}'_1s'_1t'_1,{\bf p}'_2s'_2t'_2;{\bf h}_1s_1t_1,{\bf h}_2s_2t_2$)
of the two
holes, ${\bf h}_1$, ${\bf h}_2$, and
the two particles, ${\bf p}'_1$, ${\bf p}'_2$. It is the
sum of four contributions
\begin{equation}
j^\mu(1',2';1,2)
\equiv j^\mu({\bf p}'_1s'_1t'_1,{\bf p}'_2s'_2t'_2;{\bf h}_1s_1t_1,{\bf h}_2s_2t_2) =
j_{\rm sea}^\mu +j_{\rm \pi}^\mu + j_{\rm pole}^\mu + j_{\rm \Delta}^\mu,
\end{equation}
corresponding in Fig. \ref{fig_feynman} to the seagull (diagrams
a,b), pion in flight (c), pion-pole (d,e) and $\Delta(1232)$
excitation (f,g,h,i). Their explicit expressions are given in Ref.
\cite{Sim16}.
\begin{figure}[htbp]
\centering
\includegraphics[width=8cm,bb=110 310 500 690]{Figures/diagmec.ps}
\caption{Feynman diagrams for the electroweak MEC. }\label{fig_feynman}
\end{figure}
The inclusive hadronic tensor in the 2p-2h channel is computed by integration over all
the 2p-2h excitations of the RFG.
Momentum conservation enforces $\bf p'_2= h_1+h_2+q-p'_1$. Hence
\begin{equation}
W^{\mu\nu}_{\rm 2p2h}
=
\frac{V}{(2\pi)^9}\int
d^3p'_1
d^3h_1
d^3h_2
\frac{m_N^4}{E_1E_2E'_1E'_2}
w^{\mu\nu}({\bf p}'_1,{\bf p}'_2;{\bf h}_1,{\bf h}_2)
\Theta(p'_1,h_1)\Theta(p'_2,h_2)
\delta(E'_1+E'_2-E_1-E_2-\omega) ,
\label{amaro-hadronic}
\end{equation}
where $\Theta(p',h) \equiv
\theta(p'-k_F)
\theta(k_F-h)$. The tensor inside the integral is
\begin{equation}
w^{\mu\nu}({\bf p}'_1,{\bf p}'_2;{\bf h}_1,{\bf h}_2) \equiv \frac{1}{4}
\sum_{s_1s_2s'_1s'_2}
\sum_{t_1t_2t'_1t'_2}
j^{\mu}(1',2';1,2)^*_A
j^{\nu}(1',2';1,2)_A \, ,
\label{amaro-elementary}
\end{equation}
where $j^\mu(1',2',1,2)_A$ is the antisymetrized MEC matrix element
\begin{equation} \label{amaro-anti}
j^{\mu}(1',2',1,2)_A
\equiv j^{\mu}(1',2',1,2)-
j^{\mu}(1',2',2,1) \,.
\end{equation}
The factor $1/4$ in Eq.~(\ref{amaro-elementary}) accounts for the
antisymmetry of the two-body wave function in isospin formalism,
to avoid double counting
in the number of final 2p-2h states.
Due to azimuthal symmetry around the $z$ axis
---in the ${\bf q}$ direction--- we fix the azimuthal angle of particle 1'
$\phi'_1=0$, and multiply by a factor $2\pi$. The
energy delta-function enables integrating over $p'_1$.
Then Eq.~(\ref{amaro-hadronic}) is reduced to a seven
dimensions integral that is computed numerically
\cite{Sim14a,Sim14b}.
The Dirac matrix elements of the currents are also computed numerically.
The 2p-2h inclusive cross section requires one to compute
the five weak response functions, $R^{CC,CL,LL,T,T'}$, for
$(\nu_\mu,\mu^-)$. All of these responses were computed and analyzed in
Ref.~\cite{Sim16}.
The five response functions have
been parametrized in the kinematic range $100< q< 2000$
\cite{Megias:2016fjk};
the parametrization is
convenient because in neutrino scattering there is an additional
integration over the incident neutrino flux.
This parametrization
has been implemented in the Monte Carlo event generator GENIE
\cite{Dol20,Megias:2018ujz}.
Using the parametrization of \cite{Megias:2016fjk} does not allow one to modify
the internal parameters of the MEC model. This is why
we have developed approximations to the 2p2h responses in order to speed up the calculation. Specifically, in the so-called frozen nucleon approximation we set the momenta of the two holes $h_1=h_2=0$, allowing us to integrate over the initial states analytically
\cite{Rui17}.
In the modified convolution approximation (MCA) \cite{Rui18} we
write the 2p2h responses as a convolution of two 1p-1h responses
multiplied by the elementary hadronic tensor
$w^{\mu\nu}({\bf p}'_1,{\bf p}'_2;{\bf h}_1,{\bf h}_2)$ evaluated for convenient
averaged values of ${\bf h}_1$ and ${\bf h}_2$. The resulting MCA responses are
a good approximation to the exact result with 4D integration only
over the momentum and energy comunicated to one of the nucleons.
\subsection{Validation vs electron scattering}
\label{sec:electron}
The model introduced in the previous sections allows one to describe inclusive lepton-nucleus scattering in the kinematic region including the quasielastic
reaction, the excitation of 2p2h states and the inelastic spectrum. Validation against the large amount of existing high-precision electron scattering data is a necessary test to be performed before using nuclear models in the analysis of neutrino oscillation experiments.
Such tests have been successfully carried out for the SuSAv2 model in a very wide range of kinematics from low/intermediate energies up to the highly inelastic regime. Only at very low energy and momentum transfers ($q<$300 MeV/c and $\omega<$50 MeV) does the model fail to reproduce the data, as do all models based on the impulse approximation, which is clearly not appropriate in the low-energy regime.
SuSAv2 predictions have been extensively compared with the available $(e,e')$ data on carbon and oxygen in~\cite{Megias:2016lke,Megias:2017cuh} and with the more recent JLab data on argon and titanium in~\cite{SuSAv2JLab}. Here we just show a few representative examples, choosing the kinematics
of particular interest for ongoing neutrino experiments.
\begin{figure}[htbp
\centering
\includegraphics[width=.29\textwidth,angle=270]{Figures/560_60_final.ps}
\includegraphics[width=.29\textwidth,angle=270]{Figures/12C_q570_RLRT.ps}\\\vspace*{-0.295cm}
\includegraphics[width=.29\textwidth,angle=270]{Figures/16O_737_37p1_kf230_new.ps}
\includegraphics[width=.29\textwidth,angle=270]{Figures/Ar_JLab_new.ps}\vspace*{-0.15cm}
\caption{{\it Top left panel:} The $^{12}$C$(e,e')$ cross sections from~\cite{QESarchive,QESarxiv} compared with SuSAv2-MEC predictions. The separate QE, 2p-2h and inelastic contributions are also shown.
{\it Top right panel:} $^{12}$C$(e,e')$ longitudinal (solid) and transverse (dashed) responses at $q=$ 570 MeV/c. QE, 2p2h and inelastic contributions are shown, respectively, as green, blue and orange lines. The total response is shown by the black lines. Data from~\cite{Jourdan}. {\it Bottom panels:} The $(e,e')$ cross section (left) for $^{16}$O from~~\cite{Anghinolfi:1996vm} and for $^{40}$Ar (right) from~\cite{Dai:2018gch} compared with the SuSAv2-MEC model. The separate QE, 2p-2h and inelastic contributions are also displayed.}\label{fig:Ceep}
\end{figure
In Fig.~\ref{fig:Ceep} (top left panel) the double differential $^{12}$C$(e,e')$ cross section
is shown as a function of the energy transfer $\omega$, for an incident electron beam of energy $E$
and scattering angle $\theta$,
corresponding to the momentum transfer $q_{_{\rm QE}}$
at the quasielastic peak.
The separate QE, 2p2h and inelastic contributions are shown. The agreement with the data is excellent across the full spectrum. Note that 2p2h excitations are essential in order to describe the "dip" region between the QE and $\Delta$ peaks.
An important feature of the SuSAv2 model, based on the RMF theory, is its capability to reproduce not only cross sections but also
the separated longitudinal and transverse responses, as shown in the top right panel of Fig.~\ref{fig:Ceep}, where $R_L$ (solid) and $R_T$ (dashed) are plotted versus $\omega$ for $q=$570 MeV/c, together with the separate QE, 2p2h and inelastic contributions. Note that while the longitudinal response is almost purely quasielastic, with a small 2p2h contributions arising from relativistic corrections, the transverse response receives contributions from all the three processes, which sizeably overlap at these kinematics.
In Fig.~\ref{fig:Ceep} (bottom panels) the same kind of comparison is shown for the oxygen (left) and argon (right) nuclei, showing that the superscaling approach allows for a consistent description of different nuclei.
These results give us confidence on the reliability of the model for the application to neutrino scattering. Before showing the comparison with neutrino data, in the next section we address the study of semi-inclusive electron scattering.
\section{Semi-inclusive scattering}
\label{sec:semi_electrons}
A pivotal difference between electron and neutrino scattering experiments is that the neutrino energy is known only as a broad distribution, while the energy in electron beams is typically defined with high precision. Hence, in electron scattering experiments by, for instance, detecting one nucleon in coincidence with the scattered electron, the hadronic final state is completely determined if one restricts one's attention to excitation energies of the residual system below the two-nucleon emission threshold. On the contrary, in the neutrino case, even in scenarios in which one or a few hadrons are detected in coincidence with the scattered lepton, there is no way of fully determining the hadronic final state because the beam energy and consequently the excitation energy of the residual system is unknown. This means that one needs to `integrate' for all possible final states compatible with the given kinematics. Formally, this integration is equivalent to the integration over the neutrino flux (or flux folding procedure) introduced in Sect.~\ref{sec:form}.
The semi-inclusive process for CCQE reactions was recently discussed in Refs.~\cite{VanOrden19,Gonzalez-Jimenez21}.
To obtain a somewhat deeper understanding of how the inclusive cross sections and scaling emerge, in this section we briefly discuss semi-inclusive (coincidence) electron scattering, focusing on the reaction $(e,e'N)$ in which the scattering electron and a nucleon are assumed to be detected in the final state. Then, in addition to the electron kinematical variables introduced above, we have an outgoing nucleon with 4-momentum $P^\mu_N = (E_N,{\bf p}_N)$ involving 3-momentum ${\bf p}_N$ and polar and azimuthal angles $\theta_N$ and $\phi_N$, respectively, together with energy $E_N = \sqrt{p_N^2 + m_N^2}$. No other particles are assumed to be detected, although, depending on the specific kinematics, they must be present (see below). The magnitude of the nucleon's 3-momentum is given by $p_N=|{\bf p}_N|$.
Apart from the detected nucleon, the final state contains an undetected hadronic system having missing 4-momentum $(E_B, {\bf p}_B)$, namely, a total energy of $E_B$ and a missing 3-momentum ${\bf p}_B \equiv {\bf p}_m$. In the following we shall assume that the detected nucleon is a proton. One then has
\begin{eqnarray}
{\bf p}_m = {\bf q} - {\bf p}_N\,.
\end{eqnarray}
The undetected hadronic system has invariant mass $M_B$ ($M_B^0$ at threshold with $M_B \ge M_B^0$) and total energy
\begin{eqnarray}
E_B =T_B+M_B=\sqrt{ (M_B)^2 + {p_m}^2 }\,,\label{Emiss1}
\end{eqnarray}
which defines the kinetic energy of the unobserved final-state system, $T_B$. From (\ref{Emiss1}) one has
\begin{eqnarray}
E_B = \varepsilon - \varepsilon^\prime - T_N + (M_A^0-m_N)\,,\label{Em1}
\end{eqnarray}
where $M_A^0$ is the target ground-state mass and $T_N=E_N-m_N$ is the kinetic energy of the detected nucleon. This leads to an expression for the so-called missing-energy,
\begin{equation}
E_m = (M_B-M_B^0)+E_s
= \varepsilon - \varepsilon^\prime - T_N - T_B\,,\label{eq:E_m}
\end{equation}
where $E_s=M_B^0+m_N-M_A^0$ is the separation energy and the (typically very small) recoil kinetic energy difference has been neglected. Defining the excitation energy of the residual system ${\cal E}=M_B-M_B^0$, we have
\begin{eqnarray}
E_m = {\cal E} +E_s\,.\label{eq:calE}
\end{eqnarray}
The magnitude of the missing-momentum $p_m$ is given by
\begin{equation}
p_m=\bigl[k^2+{k^\prime}^2+p_N^2-2kk^\prime\cos\theta_l-2kp_N\cos\theta_N
+2k^\prime p_N(\cos\theta_l \cos\theta_N+\sin\theta_l \sin\theta_N\cos\phi_N)\bigr]^\frac{1}{2}\,.\label{eq:p_missing}
\end{equation}
Depending on the value of the missing-energy, the residual system may be the daughter nucleus in its ground state (this defines the threshold for the reaction to become possible); or it may be in a discrete excited state (lying below the threshold where a second nucleon can be ejected), and, although they de-excite by $\gamma$-decay, that process is slow on the nuclear timescale and thus these states may be treated effectively as stationary states. Then, at a well-defined threshold a second nucleon must be emitted (this is not optional: there are no nuclear states involving one nucleon and a residual bound nucleus above this point). As $E_m$ continues to increase more and more particles enter in the final state in addition to the one special nucleon that is assumed to be detected. At even larger missing energy (roughly 140 MeV) pion production becomes possible (still with the lepton and one nucleon assumed to be detected).
\begin{figure}[htbp]
\centering
\includegraphics[width=.34\textwidth,angle=270]{Figures/trajectories_thetaN.ps}
\caption{The $E_m-p_m$ trajectories are shown for selected ``typical'' kinematics: $E_l=3800$ MeV, $\theta_l=7$ deg, $T_N=140$ MeV, $\phi_N=180$ deg for the reaction $^{16}$O$(e,e'p)$. Each line corresponds to a different value of the proton scattering angle $\theta_N$ (in degrees). Here, we plot the Rome spectral function as a background to allow one to easily identify the different regions of the spectral function that are crossed by the trajectories. }\label{fig:trajectories}
\end{figure}
From (\ref{eq:E_m} -- \ref{eq:p_missing}) it is clear that for fixed values of the observable parameters the values of $E_m$ and $p_m$ are determined for each value of $E$. In Fig.~\ref{fig:trajectories} trajectories are shown for selected kinematics, namely, $E_l=3800$ MeV, $\theta_l=7$ deg, $T_N=140$ MeV, $\phi_N=180$ deg, representing a ``typical'' situation. What is varied here is the polar angle for the detected proton, $\theta_N$. As stated above, as one goes along a given trajectory the electron energy $E$ that determines where on the trajectory one finds oneself must vary. The lower boundary defines the threshold for the semi-inclusive reaction to occur. In effect, each event where an electron and a proton are detected in coincidence corresponds to a specific trajectory and point on that trajectory.
One sees a striking pattern to the behavior one should expect when going along a given trajectory. The strength in the Rome spectral function~\cite{Benhar94,Benhar05}, which should provide a good starting point for the characteristics to be expected in semi-inclusive reactions, is extremely localized. One sees the largest concentration of strength where the $p$-shells are located (at around $E_m = 20$ MeV) with less where the broad $s$-shell is located (at around $E_m = 50$ MeV); at still larger values of $E_m$ (and $p_m$) the spectral function does have some strength, although it is spread over a wide region in the $E_m-p_m$ plane and is too small to be seen in this representation. Furthermore, we note that pion production cannot occur until one reaches $E_m \sim m_\pi$ and that it is not appreciable until $E_m \sim m_\Delta - m_N \sim 300$ MeV.
We notice that in the case of the CCQE interaction, studied in Ref.~\cite{Gonzalez-Jimenez21}, one gets a $E_m-p_m$ trajectory plot that is almost identical to the one in Fig.~\ref{fig:trajectories}. The reason is that the muon mass does not play a strong role in these trajectories, especially for the kinematics shown here for which $E_l$ is much larger than the muon mass.
In passing we note that, while other choices of two variables to replace $E_m$ and $p_m$ can of course be made, the generic behavior seen here strongly suggests that the present choice is a good one and that other choices may not reflect the highly localized nature of the nuclear response.
\begin{figure}[htbp]
\centering
(a)\includegraphics[width=.3\textwidth,angle=270]{Figures/d6sigma_vs_Em_thetaN.ps}\
(b)\includegraphics[width=.3\textwidth,angle=270]{Figures/d6sigma_vs_Em_thetaN_log.ps}
\caption{The six-fold $^{16}$O$(e,e'p)$ differential cross section as a function of the missing-energy $E_m$ (lower x-axis) and the neutrino energy (upper x-axis) on linear (a) and semi-log (b) scales. The electron and proton variables are fixed to: $E_l=3800$ MeV, $\theta_l=7$ deg, $T_N=140$ MeV, $\phi_N=180$ deg, as in Fig.~\ref{fig:trajectories}. }\label{fig:6fold}
\end{figure}
In Fig.~\ref{fig:6fold} we represent the six-fold differential cross section for fixed electron and proton kinematics as a function of the missing-energy. The model used was PWIA together with the Rome spectral function~\cite{Benhar94,Benhar05}, as described, {\it e.g.,} in Refs.~\cite{VanOrden19,Franco-Patino20}.
Equation~\eqref{eq:E_m} tells us that for fixed nucleon and final lepton energies, if one neglects the kinetic energy of the residual system $T_B$, which under typical QE conditions is always very small, then one finds a one-to-one linear relation between the initial lepton energy and the missing energy. Thus, in Fig.~\ref{fig:6fold} the beam energy is shown as a second x-axis.
By varying the angle $\theta_N$, the cross section changes its magnitude and its shape. One observes two prominent peaks corresponding to the $p$-shells, a wide bump for the $s$-shell and a background that extends up to high missing-energies. Clearly, as expected, the $p$-shell strength is largest, the $s$-shell strength is smaller and the high-$E_m$ strength is completely negligible, being down by several orders of magnitude. By examining these results in the light of the trajectories shown in Fig.~\ref{fig:trajectories} we see that the general behavior we expect to occur is borne out. For example, the trajectories for $\theta_N = 60$ and $80$ degrees both pass though the $p$-shell region near its peak. However, one trajectory intersects the $s$-shell region more than the other one does and this results in relatively different amounts from the $s$-shell compared with the $p$-shell.
Or, large values of $\theta_N$ correspond to very small cross sections, as they should, since neither the $p$- nor $s$-shell regions are crossed.
\begin{figure}[htbp]
\centering
\includegraphics[width=.75\textwidth,angle=0]{Figures/Donnelly_kin.eps}
\caption{Excitation energy ${\cal E} = E_m - E_s$ versus missing momentum $p=p_m$. The shaded area represents the kinematically allowed region for $y<0$ (left panel) and $y>0$ (right panel).}\label{fig:regions}
\end{figure}
We do have some knowledge about this generic behavior of the distribution of strength from inclusive electron scattering, $(e,e')$. Inclusive scattering corresponds to performing integrals over specific regions in the $E_m-p_m$ plane~\cite{Amaro20,Day90,Moreno14}. In Fig.~\ref{fig:regions} we show the typical situation for kinematics at values of $q$ and $\omega$ where one is below the QE peak ($y < 0$, left panel) or above the peak ($y > 0$, right panel). In each case the inclusive QE cross section is obtained by integrating the semi-inclusive $(e,e'N)$ cross sections over the shaded regions. Accordingly one can see that the two classes of cross section are intrinsically related. Clearly, if one had complete knowledge of the semi-inclusive response for a wide range of kinematics then the integrations could be performed to yield the inclusive response. Unfortunately this is not the case. Note that having a model for the inclusive cross section (the total hadronic cross section) does not mean that such a model will be valid for the semi-inclusive cross section (which constitutes the integrand of the former). An example of this is the RFG model which is not unreasonable for the total (inclusive) QE cross section, but is poor for the semi-inclusive response.
An example is that of the model used for the semi-inclusive response discussed above. If employed for the inclusive cross section one finds for $^{16}$O that somewhat over 50\% of the inclusive cross section stems from the $p$-shells, about 25\% comes from the $s$-shell region and the rest comes from a broad region at higher missing-energy. This strength at higher missing energy is partially responsible for the asymmetry found in the scaling functions. Note that inclusive scattering at high momentum transfers and hence the scaling functions involve broad integrals, whereas semi-inclusive scattering at fixed final-state electron and nucleon momenta involves a trajectory in the $E_m - p_m$ plane as the beam energy is varied. In detailed analyses one finds much more strength in that case coming from the valence knockout region, with much less arising from the high-$E_m$ region. This has consequences for CC$\nu$ reactions as discussed in Ref.~\cite{Gonzalez-Jimenez21}.
\section{Results }
\label{sec:results}
In Sect.~\ref{sec:electron}, the validation of the SuSAv2-MEC model based on the RMF theory with ($e,e'$) data has been proven as a solid benchmark to assess the validity of a nuclear model before its application to neutrino reactions. Here, we show the capability of the SuSAv2-MEC model to describe a wide range of kinematics of interest for neutrino oscillation experiments. In particular, we focus on the comparisons of the SuSAv2-MEC model with charged-current neutrino cross sections from different experiments, in particular T2K and MINERvA. Our analysis is mainly related to charged-current quasielastic-like events, also called CC0$\pi$, which are characterized by having no pions detected in the final state and are dominated by the QE and 2p2h channels. Finally, we extend the analysis to cross section measurements where pion production
is also included.
\begin{figure}[htbp
\begin{center}\vspace{0.80cm}
\hspace*{-0.84cm}\includegraphics[scale=0.167, angle=270]{Figures/T2K_12C_000060_new.ps}\hspace*{-0.64cm}%
\includegraphics[scale=0.167, angle=270]{Figures/T2K_12C_080085_new.ps}\hspace*{-0.74cm}
\includegraphics[scale=0.167, angle=270]{Figures/T2K_12C_094098_new.ps}\hspace*{-0.64cm}%
\includegraphics[scale=0.167, angle=270]{Figures/T2K_12C_098100_new.ps}\\
\hspace{-0.84cm}\includegraphics[scale=0.167, angle=270]{Figures/T2K_16O_00000600_kf230.ps}\hspace{-0.64cm}\includegraphics[scale=0.167, angle=270]{Figures/T2K_16O_08000850_kf230_2.ps}\hspace{-0.64cm}\includegraphics[scale=0.167, angle=270]{Figures/T2K_16O_09250975_kf230_2.ps}\hspace{-0.64cm}\includegraphics[scale=0.167, angle=270]{Figures/T2K_16O_09751000_kf230_2.ps}\hspace{-0.4cm}
\end{center
\caption{Comparison of the SuSAv2-MEC model with the T2K flux-integrated CCQE and 2p2h
double-differential cross section for neutrino for
scattering on $^{12}$C (top panels) and $^{16}$O (bottom panels) in units of 10$^{-39}$ cm$^2$/GeV per nucleon (for carbon) and per nucleon target (for oxygen).
The CC$0\pi$ T2K data are from Ref.~\cite{Abe:2016tmq} and~\cite{T2Kwater}.
\label{fig:T2K_d2snew}}
\end{figure}
In Fig.~\ref{fig:T2K_d2snew} we show the comparison of the SuSAv2-MEC model with T2K CC0$\pi$ double differential cross sections on $^{12}$C (top panels) and $^{16}$O (bottom panels). These data are compared with the QE and 2p2h MEC contributions showing an overall good agreement. Monte Carlo event generators also include other contributions that can mimic a CCQE-like event in the data analysis such as pion-absorption processes in the nucleus. Nevertheless, as it will be discussed later, these contributions are not particularly relevant at T2K kinematics.
It is also important to note that the QE and 2p2h contributions in the SuSAv2-MEC model can be easily extrapolated from one nucleus to another by means of scaling rules, which are different for each nuclear regime. This is based on the assumption of 2nd-kind scaling, already mentioned in Sect.~\ref{sec:scal}, and that has been proven in the SuSAv2-MEC approach for the QE and 2p2h channels~\cite{SuSAv2JLab,2p2hscaling}. It is worth commenting on the case of the most forward-angle region, associated with low kinematics, {\it i.e.,} low energy and momentum transferred to the nucleus.
\begin{figure}[htbp]
\begin{center}
\hspace*{-0.26cm}\includegraphics[scale=0.16, angle=270]{Figures/T2K_12C_minus100000_rmfrpwia.ps}\hspace*{-0.6cm}\includegraphics[scale=0.16, angle=270]{Figures/T2K_12C_08000850_rmfrpwia.ps}\hspace*{-0.6cm}\includegraphics[scale=0.16, angle=270]{Figures/T2K_12C_094098_rmfrpwia.ps} \hspace*{-0.7cm}\includegraphics[scale=0.16, angle=270]{Figures/T2K_12C_098100_rmfrpwia.ps}
\begin{center}
\end{center}
\end{center}\vspace*{-0.79cm}
\caption{Comparison of double differential cross sections on
$^{12}$C (solid lines) and $^{16}$O (dashed lines) at T2K
kinematics within the SuSAv2 (1p1h) and RMF models. Results are
displayed from forward to backward angles.}\label{rmfv2}
\end{figure}
Within this region, the SuSAv2-MEC model slightly overestimates the T2K data for C and O, being more noticeable in the latter case. This is mainly due to RMF scaling violations, related to low-energy nuclear effects and to different binding energies for each nucleus, which are not properly accounted for in the SuSAv2 approach. This limitation of the model has been solved in the recent ED-RMF approach~\cite{Gonzalez-Jimenez_edRMF,Gonzalez-Jimenez_edRMF2} where the goodness of the RMF strong vector and scalar potentials at low-intermediate kinematics are present while also retaining the benefits of the SuSAv2 model description at larger kinematics, as described in Sect.~\ref{sec:scal}. The differences introduced by this more accurate description of low-energy nuclear effects in the RMF (and ED-RMF) model can be observed in Fig.~\ref{rmfv2} where large discrepancies between oxygen and carbon predictions are observed between the RMF model and the SuSAv2-MEC one at very forward angles, being smaller as we move to more backward kinematics. At the same time, these differences are more prominent in the case of oxygen.
This is connected with the assumptions made in the SuSAv2 approach which is based on the RMF analysis on $^{12}$C but on the application of scaling rules to describe other nuclear targets instead of relying on particular RMF predictions for each nucleus. Therefore, at kinematics where scaling violations are present, these differences, while noticeable in the nucleus of reference ($^{12}$C), can be even more important for other nuclear targets. On the contrary, SuSAv2 assumes that scaling works well even at low kinematics, thus implying minor differences between $^{12}$C and $^{16}$O, mainly due to the scaling rule and differences in the energy shift introduced for each nucleus. This analysis is of relevance for T2K~\cite{T2KCO} where the accuracy of the C to O extrapolation plays an important role on the oscillation analysis and whose predictions have shown some discrepancies between C and O for very forward-going muons which may be explained by the RMF predictions.
\begin{figure}[ht]\vspace{-0.128cm}
\hspace*{-0.295cm}\includegraphics[scale=0.214, angle=270]{Figures/minerva2018_qelike_1752.ps}\hspace*{-0.584cm}%
\hspace*{-0.295cm}\includegraphics[scale=0.214, angle=270]{Figures/minerva2018_qelike_3252.ps}\hspace*{-0.584cm}%
\hspace*{-0.295cm}\includegraphics[scale=0.214, angle=270]{Figures/minerva2018_qelike_12502.ps}\hspace*{-0.584cm}\\
\hspace*{-0.295cm}\includegraphics[scale=0.214, angle=270]{Figures/minerva2018_qelike_175_neutrino.ps}\hspace*{-0.584cm}%
\hspace*{-0.295cm}\includegraphics[scale=0.214, angle=270]{Figures/minerva2018_qelike_325_neutrino.ps}\hspace*{-0.584cm}%
\hspace*{-0.295cm}\includegraphics[scale=0.214, angle=270]{Figures/minerva2018_qelike_1250_neutrino.ps}
\caption{(Color online) The MINERvA ``QE-like" and ``CCQE"
double differential cross sections for $\bar\nu_\mu$ (top
panels) and $\nu_\mu$ (bottom panels) scattering on
hydrocarbon versus the muon transverse momentum, in bins of
the muon longitudinal momentum (in GeV/c). The curves
represent the prediction of the SuSAv2+2p2h-MEC (blue) as
well as the separate quasielastic (red) and 2p2h-MEC
(orange) channels. MINERvA data and experimental
fluxes are from Refs.~\cite{Patrick:2018gvi}
and~\cite{PhysRevD.99.012004}.
}
\label{fig:fig1mnv}
\end{figure}
The SuSAv2-MEC model is also compared in Fig.~\ref{fig:fig1mnv} with the MINERvA CCQE-like double differential (anti)neutrino measurements on hydrocarbon (CH) in terms of the transverse momentum of the outgoing muon (with respect to the antineutrino beam), in bins of the muon longitudinal momentum. Due to its relativistic nature, the SuSAv2-MEC model is well suited to describe these data~\cite{Megias:2018ujz,Patrick:2018gvi} where the mean neutrino energy is around 3.5 GeV. An overall good agreement is reached without resorting to any tuning or additional parameters. According to Refs.~\cite{Megias:2018ujz,Patrick:2018gvi}, the MINERvA ``QE-like'' cross sections entail, besides pure quasielastic contributions, events that have post-FSI final states without mesons, prompt photons above nuclear de-excitation energies, heavy baryons, or protons above a kinetic energy threshold of 120 MeV, thus including zero-meson final states arising from resonant pion production followed
by pion absorption in the nucleus and from multi-nucleon interactions.
Apart from the ``QE-like'' points, the MINERvA ``CCQE'' signal is also shown, and corresponds to events initially generated in the GENIE event generator~\cite{Andreopoulos:2009rq} as quasielastic (that is, no resonant or deep inelastic scatters, but including scatters from nucleons in correlated pairs with zero-meson final states), regardless of the final-state particles produced, thus
including CCQE and 2p2h reactions.
The difference between the two data sets is mainly related to pion production plus re-absorption and goes from $\sim15\%$ to $\sim5\%$ depending on the kinematics. The present SuSAv2-MEC results do not include processes
corresponding to these pion re-absorption processes inside the
nucleus and the comparison should be done with the
``CCQE'' data rather than with the ``QE-like'' ones.
A more detailed analysis of these results together with a $\chi^2$ test can be found in~\cite{Megias:2018ujz}, where the SuSAv2 $\chi^2$ shows its compatibility with data and with the MINERvA/GENIE predictions.
It is also important to note that due to MINERvA's acceptance, the muon scattering angle is limited to $\theta_\mu<$ 20$^{\circ}$ as well as the muon kinematics (1.5 GeV $< p_{||} < $ 15 GeV, $p_T<$ 1.5 GeV) in both experimental and theoretical results. This implies important phase-space restrictions for large energy and momentum transfer to the nuclear target and makes the available phase space not so different from the T2K one, as shown in~\cite{megias2019axial}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.349\linewidth, angle=0]{Figures/T2K_12C_000_rmf_cutPf_alt.ps}\hspace*{-0.295cm}
\includegraphics[width=0.349\linewidth, angle=0]{Figures/T2K_12C_0450_rmf_cutPf_alt.ps}\hspace*{-0.295cm}
\includegraphics[width=0.349\linewidth, angle=0]{Figures/T2K_12C_0650_rmf_cutPf_alt.ps}
\includegraphics[width=0.349\linewidth, angle=0]{Figures/T2K_12C_07000800_rmf_cutPf_alt.ps}\hspace*{-0.295cm}
\includegraphics[width=0.349\linewidth, angle=0]{Figures/T2K_12C_08000850_rmf_cutPf_alt.ps}\hspace*{-0.295cm}
\includegraphics[width=0.349\linewidth, angle=0]{Figures/T2K_12C_0875_rmf_cutPf_alt.ps}
\includegraphics[width=0.349\linewidth, angle=0]{Figures/T2K_12C_092_rmf_cutPf_alt.ps}\hspace*{-0.295cm}
\includegraphics[width=0.349\linewidth, angle=0]{Figures/T2K_12C_096_rmf_cutPf_alt.ps}\hspace*{-0.295cm}
\includegraphics[width=0.349\linewidth, angle=0]{Figures/T2K_12C_099_rmf_cutPf_alt2.ps}
\end{center}
\caption{Comparisons of data and model predictions for differential CC0$\pi$ muon-neutrino cross sections on $^{12}$C in the T2K neutrino beam as a function of the muon kinematics when there are no protons with momenta above 500 MeV. Two 1p1h predictions are shown (one from RMF, the other from SuSAv2 implemented in GENIE), in addition to the SuSAv2 2p2h and pion absorption contributions from GENIE. The total contributions when using each of the two 1p1h models is also shown. Goodness of fit are calculated to be $\chi^2_{RMF}=171.87$ (59 bins) and $\chi^2_{SuSA}=168.92$ (60 bins), where the latter includes a single extra bin from -1.0 to -0.3 $\cos{\theta}$ (not shown). The data points are taken from~\cite{Abe:2018pwo}.}
\label{fig:ssincT2KComp}
\end{figure}
The previous T2K and MINERvA results are only related to the final lepton kinematics, the so-called inclusive measurements. At this point, it is worth mentioning that other models~\cite{Nieves:2011pp,Nieves:2011yp,Martini:2010ex,Martini12PRD,Gallmeister16,Meucci,Meucci15,Rocco16,Rocco:2018mwt,Golan12,Lovato:2015qka,Pandey:2016ee,Butkevich:2017mnc,Ivanov:2013saa,PhysRevD.99.093001,Martini:2011wp,Nieves:2013nubar,Martini:2013sha,Mosel:2014lja,Meucci:2014bva,MartiniCP,Mosel,Dolan:2018sbb} have been also developed to address these CC inclusive neutrino interactions and, although similar agreement with data can be obtained, they are based on different assumptions about the nuclear properties and dynamics. Combined analyses of these models with more exclusive neutrino measurements where hadron kinematics and other nuclear effects can be analyzed in more detail would help to improve model selection for data analysis (see also the discussions in Sect.~\ref{sec:semi_electrons}). The advantage of SuSAv2-MEC (and RMF) is that of being a fully relativistic model that has shown an overall good agreement with electron and neutrino scattering data and that can be extended without further approximations to the full-energy range of interest for present and future neutrino experiments.
For this reason, the SuSAv2-MEC model (1p1h and 2p2h) has been recently implemented in the GENIE neutrino event generator~\cite{Andreopoulos:2009rq} with the aim of improving the characterisation of the nuclear effects in neutrino cross section measurements and work is now in progress to implement this model together with the ED-RMF one in the NEUT event generator~\cite{Hayato:2009zz} for its application on the T2K oscillation analysis.
Based on these works, in Fig.~\ref{fig:ssincT2KComp} we study the T2K CC0$\pi$ measurement of interactions with protons less than 500 MeV~\cite{Abe:2018pwo} in comparison with the SuSAv2-1p1h, SuSAv2-2p2h and pion-absorption predictions from GENIE (see~\cite{Dol20} for details). At proton momentum below 500 MeV/c a clear dominance of the SuSAv2-1p1h channel is observed. Note that SuSAv2-GENIE's also shows an overestimation of data at very forward angles, as observed in Fig.~\ref{fig:T2K_d2snew} which is mainly due to low energy transfer scaling violations which are absent in the SuSAv2-model but present in the RMF theory. In order to analyze the low-energy nuclear effects and these scaling violations, we also show the comparison with the RMF model for the 1p1h channel together with the SuSAv2-2p2h and pion-absorption results from GENIE, noticing a remarkable improvement in the data comparison at very forward angles.
Here, one of the drawbacks of the SuSAv2 model is that it only predicts lepton kinematics so its implementation in MC event generators has to rely on a factorization approach and on the information available in these generators to determine hadron kinematics (see~\cite{Dol20} for details). This limitation will be addressed in the forthcoming implementation of the ED-RMF model in event generators which will provide full information about lepton and hadron kinematics in the final state together with a consistent description of the nuclear dynamics.
Due to aforementioned dominance of the 1p1h channel on CC0$\pi$ interactions with low momentum protons observed in Fig.~\ref{fig:ssincT2KComp}, it remains difficult to draw clear conclusions about the goodness of the 2p2h description.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.49\linewidth]{Figures/dptCompFull.ps}\hspace*{-0.15cm}
\includegraphics[width=0.509\linewidth]{Figures/dptSOMerge.ps}
\end{center}
\caption{The regularised T2K measurement of CC0$\pi$ muon-neutrino cross sections on $^{12}$C at T2K kinematics as a function of the Single Transverse Variable~\cite{Lu:2015tcr} $\delta p_T$ compared to predictions from the GENIE-implemented SuSAv2 and Valencia 1p1h+2p2h models, each of which is added to GENIE's pion absorption prediction. The total contributions when using SuSAv2 and Valencia models is also displayed. A shape only comparison is also shown (right panel). Goodness of fit are calculated as follows. For $\delta p_T$: $\chi^{2}_{SuSA}=20.5$, $\chi^{2}_{Valencia}=27.1$. The data points are taken from~\cite{Abe:2018pwo}.}
\label{fig:stvT2KComp}
\end{figure}
Nevertheless, this can be explored further using more semi-inclusive measurements with measurements of proton and muon kinematics as shown in Fig.~\ref{fig:stvT2KComp}, where the SuSAv2-MEC model in GENIE is compared with the so-called single transverse variables (STV), and in particular with the transverse momentum imbalance, $\delta p_T$, defined in terms of the momentum imbalances between the outgoing muon and highest momentum proton in the plane transverse to the incoming neutrino (see~\cite{Dol20,Lu:2015tcr} for details). These transverse kinematic imbalances allow one to study initial-state nuclear dynamics but also to better isolate the 2p2h channel. In $\delta p_T$ the 1p1h channel is not expected to contribute significantly beyond the initial state nucleon momentum ($\sim$230~MeV/c for carbon in a Fermi gas approach), thus implying that the high-$\delta p_T$ values will be dominated by 2p2h and other contributions, as observed in Fig.~\ref{fig:stvT2KComp}. The overestimation at high-$\delta p_T$ in the left panel may indicate that the 2p2h contribution is too strong. However, as discussed in~\cite{Dolan:2018zye}, this overall over-prediction could potentially be improved by stronger nucleon FSI, which may improve the data agreement in the tail. On the other hand, this may also be explained by the approximations taken to produce semi-inclusive predictions from inclusive models, as described in~\cite{Dol20}. It is expected that these drawbacks can be addressed in further works via the implementation of RMF models for the 1p1h channel and a full semi-inclusive 2p2h model. Apart from these limitations, it is worth noticing the almost perfect description of the shape of $\delta p_T$ in the right panel, improving the agreement reached by other descriptions also implemented in generators.
To conclude this section, we show the comparison of the SuSAv2-MEC model with measurements that also consider non-QE contributions, mainly pion production. This reaction channel is also of relevance for present and future neutrino oscillation experiments. Neutral current $\pi^0$ production constitutes an important background in the electron neutrino and antineutrino appearance analyses but also neutrino-induced pions emitted can mimic QE-like events if they are not properly detected or if they are absorbed due to FSI effects. Thus, the analysis and detection of these pion production events in coincidence with the final lepton and other hadrons is of paramount importance for the neutrino energy reconstruction.
\begin{figure}[htbp]
\centering
\includegraphics[width=.345\textwidth,angle=270]{Figures/t2k-qe.ps}
\hspace{0.25cm}
\includegraphics[width=.345\textwidth,angle=270]{Figures/sciboone.ps}
\caption{ (Left panel) We show the CC-inclusive T2K flux-folded
$\nu_e$-$^{12}$C $Q^2_{QE}$ differential cross section per
nucleon. (Right panel) The CC $\nu_\mu$ total cross section on C$_8$H$_8$ is
presented. Experimental data are from T2K~\cite{T2Kinclusive14}
and SciBooNE~\cite{SciBooNE11}. Theoretical predictions for QE, non-QE
(1$\pi$) and the 2p2h MEC are shown
separately.
Plots from Ref.~\cite{Amaro20}. }
\label{fig:T2KQE-SciB}
\end{figure}
In order to analyze the neutrino-induced pion production channel we make use of the phenomenological SuSA-$\Delta$ approach which extends the superscaling arguments observed in the QE regime to the $\Delta$ resonance region (see~\cite{Ivanov16} for details). In Fig.~\ref{fig:T2KQE-SciB} (left panel) we show the $\nu_e$-$^{12}$C inclusive differential cross section averaged with the T2K flux versus the reconstructed four-momentum transfer, $Q^2_{QE}$, and in Fig.~\ref{fig:T2KQE-SciB} (right panel) the $\nu_\mu$ total cross section on C$_8$H$_8$ target. Three contributions are shown in these plots, namely, the SuSAv2 QE and SuSAv2 2p2h-MEC channels together with the SuSA-$\Delta$ approach.
Although an overall good agreement with data is observed in both panels, some underestimations are present at large kinematics, {\it i.e.,} high $E_\nu$ and $Q^2_{QE}$, which reveal the need for including higher resonances and deep inelastic scattering in the description. To address this, the SuSAv2-inelastic, which has been successfully applied for the analysis of the full inelastic spectrum in electron scattering (see Sect.~\ref{sec:electron} for details), will be soon extended to the analysis of the neutrino sector.
\section{Conclusions}
\label{sec:concl}
In this paper we have summarized the basic ingredients that go into scaling analyses of inclusive electron scattering and charge-changing neutrino reactions with nuclei. Importantly, good agreement in the former case is viewed as a pre-requisite to being able to predict the latter --- clearly if one fails to account for inclusive electron scattering then it is unreasonable to expect that inclusive CC$\nu$ reactions will be adequately modeled.
In particular, we have emphasized the use of scaling of the first kind (independence of $q$ at high energies) and of scaling of the second kind (independence of nuclear species). When both kinds of scaling are invoked we refer to it as Super-Scaling and accordingly our present focus has been placed on the so-called Super-Scaling Approach (SuSA) in its original form and in extensions of that original form. Such extensions have been introduced to account for the modest level of scaling violations seen in direct comparisons with inclusive electron scattering data.
In addition we have introduced the way 2p-2h MEC effects have been incorporated, showing that, while they are usually corrections to the cross sections being represented via the scaling functions obtained either phenomenologically or through use of specific models, they typically are required to get a successful picture of the response.
We have briefly summarized these basic ideas and then proceeded to show examples of the excellent agreement found with inclusive electron scattering cross sections measured for several light to medium-weight nuclei.
To place the discussions in context we have also provided a section outlining how semi-inclusive scattering and inclusive scattering are related, and specifically how the former probes particular regions of missing-energy and -momentum (characterized by so-called trajectories), while the latter involves integrations over the $E_m - p_m$ plane of the semi-inclusive response. On the one hand, inclusive scattering, being a total hadronic cross section, is less dependent on the underlying details of nuclear structure. In fact, given that the kinematics of the single-nucleon knockout are handled relativistically and that there are sum rules that determine the integral of the inclusive response, then a single parameter can be used to fix the width of the response. For this reason many models yield the rough behaviour of the inclusive cross section, even the relativisitic Fermi gas model. On the other hand, the semi-inclusive cross section depends critically on how the strength is distributed as a function of $E_m$ and $p_m$. Accordingly, models that fare reasonably well for inclusive scattering cannot be relied upon to properly represent the underlying nuclear structure needed for semi-inclusive reactions.
Finally, having discussed the foundations of Super-Scaling for electron scattering we show some selected results for CC$\nu$ reactions. The agreements with existing data are found to be excellent, giving us confidence that at least for inclusive neutrino reactions the problem appears to be well in hand. Note, however, from the statements made in the previous paragraph, that this is not to be taken as proof that semi-inclusive CC$\nu$ modeling should be expected to be robust. Indeed, different models yield quite different results for cross sections in which both a charged lepton and a nucleon are detected in coincidence (a trend in modern experiments), even when the corresponding inclusive cross sections do not differ significantly. The modeling of semi-inclusive reactions is part of our present projects.
\section{Acknowledgments}
This work was supported in part by the Project BARM-RILO-20 of University of Turin and from INFN, National Project NUCSYS (M.B.B.); by the Madrid Government and Complutense University under, project PR65/19-22430 (R.G.-J.);
by the Spanish Ministerio de Ciencia, Innovaci\'on y Universidades and ERDF (European Regional Development Fund) under contracts FIS2017-88410-P, and by the Junta de Andalucia (grants No. FQM160 and SOMM17/6105/UGR) (J.A.C.); by the University of Tokyo ICRR's Inter-University Research Program FY2020\&FY2021 and by the European Union's Horizon 2020 research and innovation programme under the Marie Sk{\l}odowska-Curie grant agreement No. 839481 (G.D.M.); by the Spanish Ministry of Science through grant FIS2017-85053-C2-1-P, and by Junta de Andalucia (grant No. FQM-225) (J.E.A. and I.R.-S); by the Office of Nuclear Physics of the US Department of Energy under Grant Contract DE-FG02-94ER40818 (T.W.D.).
|
\section{Introduction and definitions}
In 2017 \cite{klaassen} a definition of spiral tilings was given, thereby
answering a question posed by Gr\"unbaum and Shephard in the late 1970s.
The author had the pleasure to discuss the topic via e-mail with Branko
Gr\"unbaum in his 87th year. During this correspondence the question
arose whether a spiral structure (given a certain definition of it)
could be recognized automatically or whether ``to some
extent, at least, the spiral effect is psychological\textquotedblright ,
as Gr\"unbaum and Shephard had conjectured in 1987 (see exercise section of chapter 9.5 in \cite{grunbaum}). In this
paper, an algorithm for automatic detection of such a tiling's spiral
structure and its first implementation results will be discussed. Finally, the definitions for several types of spiral tilings will be refined based on this investigation.
\begin{center}
\includegraphics[width=12.5cm]{Fig1}
\end{center}
\begin{description}
\small
\item [{Figure\ 1}] Spiral structure ``by coloring'' (left) vs. ``by construction'' (right)
\end{description}
If in Figure 1 all colors of tiles were erased and only the tile structure
remained, in the left tiling (of simple squares) nobody could
find a spiral character. On the other hand, the right tiling contains
the spiral structure by construction, although not so easily recognized
without coloring.
One key aspect of the definition of spiral tilings in \cite{klaassen} is that it gives a basis to distinguish between tilings in which the spirals were just introduced by coloring
and those which incorporate a spiral structure.
The two tilings of Figure 1 represent both types of spirals. For the latter type we will define a
further distinction at the end of this paper.
During this study, we assume that all tiles are closed topological disks. If not specified explicitly, we assume that no singular points exist, where the tiles are clustered. All investigated tilings without such singular points are assumed to be $k$-hedral, which means that there are only finitely many congruence classes.
We will refer to the definitions from \cite{klaassen} throughout this paper,
so, we decided to put them into the appendix to have them at hand.
First we need the term \textit{L-tiling} which can be summarized using
ordinary language:
``An \textit{L-tiling} allows a partitioning into several parts (called
\textit{arms}), in each of which we can draw a continuous, unlimited
curve (called \textit{thread}) running through the interior of each tile (of the part)
exactly once and winding infinitely often around a certain point''.
In the appendix the reader may have a look for this definition in
strict mathematical terms, but for the further understanding this
one-sentence-version should be sufficient. (The left hand part of
Figure 1 serves as an example of an L-tiling.)
Also the term \textit{S-tiling} from \cite{klaassen} can be summarized in
ordinary words: ``An\textit{ S-tiling} must have the properties of
an L-tiling plus an extra property that neighboring tiles within each
arm are positioned to each other in a way that cannot occur with
two neighbored tiles from different arms (except at the beginning
of an arm)''. E.g., a closer inspection of Figure 1 (right) shows
that within each arm (equal color) there are just two different constellations
of neighboring tiles, and both constellations do not occur with tiles
of different colors. So, this is an S-tiling. (See again the appendix
for a more rigid definition.)
Then for our algorithm we need another pair of definitions.
\begin{spacing}{0}
\noindent \begin{center}
\includegraphics[width=10cm]{edg2}
\par\end{center}
\end{spacing}
\begin{description}
\small
\item [{Figure\ 2}] 6-square subset \textit{M} with \textit{DG(M)} in
grey (left) and \textit{CG(M)} (right)
\end{description}
\begin{description}
\item [{Definition}] \textbf{\textit{direct contact graph (DG)}}\textbf{ and }\textbf{\textit{
contact graph (CG)}}\textbf{:}
\end{description}
Let $M$ be a connected set of some tiles of a tiling. Then the \textit{``contact
graph of}$M$'' or \textbf{\textit{CG(M)}} is the graph in which each
tile of $M$ is represented by a node and two of such nodes are connected
by an edge iff the corresponding tiles ``have contact'', i.e., have
non-empty intersection \cite{buchsbaum}. We can construct a subgraph of \textit{CG(M)}
called ``\textit{direct contact graph of} $M$'' or \textbf{\textit{DG(M)}}
by deleting all edges for each pair of tiles which share only a finite
number of points of their boundaries. (In graph theory this would be called \textit{dual graph} where the
tiling is interpreted as a planar graph.)
Figure 2 shows a simple example of \textit{DG} and \textit{CG} for
a small subset of the square tiling. Observe that \textit{CG} in many
cases will not be planar. Both\textit{ DG} and \textit{CG} can be
finite or infinite, depending on the choice of \textit{M}.
\section{The algorithm}
Looking at the above-mentioned definitions for S-tilings, we
observe that they start with a partitioning of the tiling, but do
not tell us how to find it (in our example in Figure 1 ``partition''
and ``coloring'' are equivalent). So, if any automatic recognition
is possible, it must deliver a partitioning into ``spiral
arms''. Given these partitions (or arms in terms of our definition)
it is clear how to proceed further: Check whether a continuous curve
(a so-called thread) can be found satisfying the necessary conditions.
For practical reasons, we decided to search for Hamilton paths \cite{harel}
within each candidate for a spiral arm. Although this is not equivalent
to definition L or S, for a huge subset of S-tilings (maybe for all of them)
the spiral arms can be regarded as Hamiltonian w.r.t \textit{DG} or\textit{ CG}. This can
be easily implemented using graph libraries. (A Hamilton path within a connected component of
\textit{DG} or\textit{ CG} means that we can walk through the component
along its edges meeting every node exactly once.)
Let us first describe the main ideas of the algorithm just by words:
\begin{itemize}
\item Build classes of neighboring tile pairs according to their relative
position to each other
\item For each possible subset of these classes cut the tiling at the intersection
of each tile pair belonging to one of the selected classes
\item After each cut check the resulting connected components whether they
allow a Hamilton path running through each component of the (direct) contact graph winding around a central point
\end{itemize}
To give an example, in Figure 2.1 we can find four classes of tile
pairs (one connected by a short edge and three others sharing a long
edge in different ways).
\begin{frame}{}
\hbox{\hspace{-0.5em} \includegraphics[width=8cm]{F952B_new} }
\end{frame}
\begin{description}
\small
\item [{Figure\ 2.1}] Example: spiral tiling (left) and result from the algorithm (right)
\end{description}
It is obvious that just the connections via
``short'' edges had to be deleted by the algorithm to find the spiral
structure. In this case it is a two-arm spiral where both arms meet
at the center. We see (at the right half of Figure 2.1) the direct
contact graph (\textit{DG}) of the tiling after the described cut.
Now the more formal description must follow: For an algorithmic approach,
we have to restrict our scope to finite portions of a given tiling.
Throughout this section let $M$ be the investigated portion of a
tiling that should be checked by our algorithm. In section 4 the appropriate
choice of\textit{ M} will be addressed. Those tiles in $M$ which
are (in the unlimited tiling) neighbors of tiles lying outside of
$M$ are called the \textit{border} $B_{M}$. Then \textit{CG(M)}
(or \textit{CG} for short) is the contact graph of \textit{M} and\textit{
DG(M)} (or \textit{DG}) the corresponding direct contact graph. We
start with classifying all edges in these graphs depending on how
their corresponding tile pairs in $M$ are positioned to each other:
For each edge $k$ of $DG$ we form the class {[}$k${]} consisting
of all the edges $k'$ of $DG$ for which the two tiles (determined
by the endpoints of $k'$) are congruent to the corresponding
tiles determined by $k$, through an orientation-preserving isometry
of the plane (that is, by translation or rotation). Let the set of
all such classes be denoted $K_{M}=\left\{ [k_{1}],[k_{2}],\text{\dots}\right\} $.\footnote{One could call such classes \emph{edge classes} or \emph{tile pair classes}, which is equivalent here.}
(During the algorithm we will also need additional edge classes from
$CG$, constructed in the same way.) For a class {[}$k${]} we consider
the set $E(k)$ of edges in {[}$k${]}. For
each subset of classes \textbf{\textit{K}} $\subset K_{M}$ we write
\textit{E(}\textbf{\textit{K}}\textit{)}
for the corresponding edge set as a union of $E(k)$.
An edge from \textit{E(}\textbf{\textit{K}}\textit{)} between the tiles $T_1$ and $T_2$ should be denoted as $(T_1,T_2)$.
For each of these subsets of edge classes
in $DG$ we can check what happens if all these edges were deleted.
How do the remaining connected components of $DG$ ``behave''? Several
steps were included in order to exit the loops as early as possible.
For shortness, we will use the term ``component''
for ``connected component''.
\begin{description}
\item [{Algorithm:}] First check whether there are at least three congruent
tiles differing in orientation or reflection within $M$.\\
If not, end the algorithm with empty result.\\
Else, form the set of classes $K_{M}$ as described above. \\
Next we define an operation to be performed on each nonempty \textbf{\textit{K}}$\subset$\,$K_{M}$.
\item [{Operation\ A:}] (using \textbf{\textit{K }}as input and returning
either \textbf{\textit{K }}plus a graph or the result ``discarded''
if \textbf{\textit{K}} cannot fulfill a condition)\end{description}
\begin{verse}
Check whether all components of $\bigcup\limits_ {(T_i,T_j)\in E(\boldsymbol{K})} \!\!\!\!\!\!\!\!\! T_i \cap T_j \,\,\,$
are connected to the border $B_{M}$; [Remark: Represents boundaries of spiral arms.]
if not, discard \textbf{\textit{K}} and finish Operation A with result
``discarded'';
if yes, delete the edges \textit{E(}\textbf{\textit{K}}\textit{)}
from $DG$, the result is called $G$;
if all components of $G$ are connected to $B_{M}$ and allow a Hamilton path without self-intersections, go directly
to ({*});
else, do the following steps with \textbf{\textit{K}} plus any combination
of edge classes from $CG$, called ``\textbf{\textit{K}}-extension'', each of which generates a new $G$:
If for such an extended \textbf{\textit{K}} a component of the new
$G$ is not connected to $B_{M}$ or does not allow a self-avoiding Hamilton path, ignore this \textbf{\textit{K}}-extension and try the next possible one;
if a tile in $M$ has more than two vertices where it meets other tiles at single points (being connected to these tiles by edges in the new $G$), ignore this \textbf{\textit{K}}-extension and try the next possible one; [Remark: Excluding cases like the checkers tiling in [1] Figure 3, where a spiral arm is not simply connected.]
({*})
If for each component of $G$ the number of tile equivalence classes w.r.t translation is less than 3 \textendash{} discard (extended) \textbf{\textit{K}} and continue with the next one (if
extensions were needed);
if extensions were needed, return all non-discarded variants of \textbf{\textit{K}}
plus $G$ or ``discarded'' as result when all extensions were checked;
else return (\textbf{\textit{K}}, $G$) if non-discarded or else
return ``discarded''.
(\textbf{End of A})
Perform Operation A with all nonempty subsets of $K_{M}$. All non-discarded
subsets are candidates for spiral partitions. Sort the non-discarded
subsets by the number of components of $G$ (= nbr. of arms) in increasing
order.\end{verse}
\begin{description}
\item [{Operation\ B:}] (using each non-discarded $G$ as input if there
is any)\end{description}
\begin{verse}
For each component of $G$: Find a continuous piece-wise linear
curve through the corresponding tiles following the possible Hamilton
paths and modify it to check whether the conditions for being a thread
can be fulfilled. {[}Remark: this section of the algorithm is not
difficult for the human eye but needs considerable programming efforts.
On the other hand, by methods of computer graphics and optimization
this task could be handled in principle. Since the above-mentioned
``psychological effect'' is not needed here, we decided not to code
this section of the algorithm.{]}
If one component does not allow a thread, $G$ has to be discarded. As a final result the components of $G$ each with a valid thread
represent the spiral arms.
\end{verse}
\textbf{End of Algorithm}
\
This algorithm (except for Operation B) was implemented in Python,
which is by far not the fastest language but offers a lot
of libraries for graph operations.
Let us return to the example in Figure 2.1. The separation into two
arms cannot be managed by the implemented algorithm, but the spiral structure
was recognized. Only the direct contact graph is needed in this case,
but there will be some examples with \textit{CG} in the following
section.
\section{Results\ }
As a set of test cases we took several tilings from \cite{wichmann} with spiral structure.
\begin{spacing}{0}
\noindent \begin{center}
\includegraphics[width=12.2cm]{Results}
\par\end{center}
\end{spacing}
\begin{description}
\small
\item [{Figure\ 3.1a}] Tilings and resulting graphs from the algorithm
\end{description}
In the first and third column of Figure
3.1a and 3.1b we show the tilings and right hand besides them in the
second and fourth column the resulting graphs from our algorithm.
For the majority of tilings the algorithm works with \textit{DG}.
The list of examples is continued with Figure 3.1b (still with
usage of \textit{DG} instead of \textit{CG}).
\begin{spacing}{0}
\noindent \begin{center}
\includegraphics[width=12.2cm]{results2}
\par\end{center}
\end{spacing}
\begin{description}
\small
\item [{Figure\ 3.1b}] Tilings and resulting graphs from the algorithm
\end{description}
\begin{spacing}{0}
\noindent \begin{flushleft}
\includegraphics[width=12.2cm]{results4}
\par\end{flushleft}
\end{spacing}
\begin{description}
\small
\item [{Figure\ 3.2}] Tilings and resulting graphs with usage of \textit{CG}
\end{description}
There are some rare cases where \textit{CG} is needed (see Figure 3.2). So, it is recommended to start with \textit{DG} and only
if nothing could be found, a second round with \textit{CG} should
be performed.
It is interesting to note that the algorithm also works for one-armed
spirals. Though the definition for this type differs slightly from
Definition S (see Appendix: Definition O for more details) the algorithm
(up to Operation B) can be applied without any changes. Operation B can be performed here in simplified version, since only the spiral boundary curve has to be checked if it is winding around its starting point. In the above-mentioned collection
of spiral tilings \cite{wichmann} there are two examples for this case (Figure
3.3).
\begin{spacing}{0}
\noindent \begin{flushleft}
\includegraphics[width=12.2cm]{results3}
\par\end{flushleft}
\end{spacing}
\begin{description}
\small
\item [{Figure\ 3.3}] Tilings and resulting graphs for the one-armed case
\end{description}
There are some special situations, where the results indicate more
than one spiral partitioning. In Figure 3.4 we show two different
spirals for the same tiling that were both found by the algorithm
\includegraphics[width=11cm]{F6A-6}
\begin{description}
\small
\item [{Figure\ 3.4}] Two different S-partitions for the same tiling
\end{description}
The spiral arms on the right side of Figure 3.4 do not look very ``natural''
compared to the spirals on the left half, but they fulfill all conditions
for an S-tiling. Only the heuristic argument could be applied that
the partition with lower number of arms should be preferred. This
is the reason for the final part of the algorithm, where a sorting
of the resulting graphs has been proposed to find the result with
lowest number of connected components.
\section{Complexity and other algorithmic aspects}
It is quite obvious that an algorithm containing a loop over all subsets
of a given finite set must have exponential complexity (w.r.t. the
number of edge classes $\vert K_{M}\vert$). Hence, there will be
cases where the algorithm's runtime outruns all practical limits.
It should be noted here that the whole investigation did not aim on
efficient implementation, but to answer the question whether such
an algorithm exists at all.
What can be done now in cases of extremely long runtime? Such examples
exist, but we are lucky that they are rare. For these few cases we
propose to apply an algorithmic test ``by hand'' in a way that the
following items should be checked to decide whether the algorithm
will (or won't) be successful. We use again the simple structure of
Figure 2.1 to illustrate the steps:
\begin{itemize}
\item classify all edges of the direct contact graph\textit{ DG} by assigning
integers for each class of direct neighbors to define the edge classes
$K_{M}$ (In Figure 2.1 there are four classes: Let us assign 1 to
the neighbors sharing a short edge and 2, 3 and 4 to the other classes
of neighbors sharing a long edge.)
\item if spiral arms can be observed by the human eye:
Consider the spiral arms as subsets of $M$ and
run along their boundaries
to find the specific subset \textbf{\textit{K}} of $K_{M}$.
If a spiral arm locally shrinks to a single point, as in
Figure 1 (right), go back to the previous item but use \textit{CG}
(In our example in Figure 2.1 just \textit{DG} is needed and the arms'
boundaries are easily characterized just by the short edges, so we
choose \textbf{\textit{K}} = \{1\}.)
\item check whether the chosen \textbf{\textit{K}} finishes Operation A
without being discarded (This is easily checked in our example since
the tiling contains more than three tiles with different rotation
angles and each component of the resulting $G$ - after deleting
the connections via short edges - can be naturally traversed by a
Hamilton path. All these paths are connected to the outside border
region, which is also true for the arms' boundary.)
\item perform Operation B for the components of the non-discarded results
of Operation A, i.e., find a thread - or maybe several of them - following
the Hamilton path(s) (In our example this is done straight forward
with two threads starting close to the tiling's center.)
\end{itemize}
If by these checks a single subset\textbf{\textit{ K}} is found not
being discarded by Operation A, it is shown that the algorithm must
find this result within finite time. All remaining tilings from the
literature (less than 10) were investigated with the result that in
all cases where definition S is satisfied, the algorithm will return
a spiral partition. Also the somehow unexpected spiral structure within
the Hirschhorn tiling can be detected by this analysis (discussed in the last section, see Figure 5.2).
In the same section we will see another simple example which demonstrates the
advantages of the algorithm's application ``by hand''.
There is one interesting case in Brian Wichman's collection \cite{wichmann}
showing kind of disrupted spiral arms (see Figure 4.1). The two arms
indicated by two different colors are following a spiral structure
from inside to outside, but it is not possible to draw a continuous
path following the spiral within the interior of each arm.
\
\begin{center}
\includegraphics[scale=0.4, angle=90]{F19}
\end{center}
\begin{description}
\small
\item [{Figure\ 4.1}] A tiling from \cite{wichmann} with disrupted spiral arms
\end{description}
Our algorithm (here not by hand but by software) returned a negative
result in this case, which is correct since neither definition S nor
even L can be satisfied.
By its nature, an algorithm working on a finite portion of the tiling
cannot in all cases distinguish between ``true'' spirals and partitions
which start like spirals but later stop the spiral behavior. (Figure
11 in \cite{klaassen} shows an example of such a pseudo-spiral partition.)
Therefore one could start the algorithm with a smaller part of the tiling as described
and then add further tiles outside of the so-called border region.
Then one could check whether the orientation of the tiles within a
spiral arm will further change or remains in one or two fixed angular
positions. In addition, a further difficulty could occur: It is not sure that
the spiral center is always placed in the middle of the finite portion
of the tiling. So, one might first look for this center by searching
for the part where the highest number of tiles with different orientation
are clustered.
The proposed refinements from this section are all possible in principle
but the described version of the algorithm worked well enough without
it.
\section{Discussion and further refinements}
The main result of this paper is the fact that an algorithm can be
designed to decide whether a given tiling has or doesn't have a spiral
structure. This is done by a method of partitioning into spiral arms.
As we have seen in the results section, the proposed algorithm can
be applied to a wide range of tilings. We can claim that all known
spiral tilings from the literature (in the meaning of definition S
or O resp.) can be detected by the algorithm. The vast majority was
covered by our Python implementation while the remaining part (less
than 10) could be analyzed ``by hand'' following the algorithmic
check list described in the previous section. So, the algorithm is
working as desired with the limitation of not being very efficient
for all cases due to its exponential complexity.
This application ``by hand'' can also be used to decide
whether a given tiling contains more than one spiral structure.
\begin{spacing}{0}
\noindent \begin{center}
\includegraphics[width=10cm]{Fig_one_armed4}
\par\end{center}
\end{spacing}
\begin{description}
\small
\item [{Figure\ 5.1}] The algorithm applied ``by hand'' to analyze a tiling
\end{description}
We can demonstrate this with a tiling presented in \cite{klaassen} to find
out whether more than one spiral arm exists in this case (see Figure
5.1). First, we cut the tiling at those edges shared by two tringles
(= thick line) to get the left hand version (one spiral arm). Alternatively
- on the right hand side - we cut the same tiling at those edges where
a triangle meets a rhombus (= edges shared between dark grey and light
grey tiles) to find the right hand partitioning (two spiral arms).
This means that the algorithm ``by hand'' can also be used as a
method to partition a given tiling for a better understanding of its
structure.
The concept of structure analysis developed in this paper can be used for other structures than spirals, as well. In any case the final result is a tile set partition, where in each part the tiles are positioned to each other in a different way than on the parts' boundaries. So, one could ask, what typical structures could be found in this way: For the large domain of periodic tilings, we will often find partitions in form of stripes or patches. For non-periodic tilings, especially with rotational symmetry - but not restricted to those - we will detect ring-like structures, where each ring is surrounded by a larger one. For such ring partitions, we distinguish two types which can be defined in a way analogous to definition L and S.
\begin{description}
\item [{Definition}] \textbf{\textit{weak ring partition}: }
A tile set partition of a plane tiling into infinitely many parts is called a \textit{weak ring partition} if each part (as union of its tiles)
contains a closed Jordan curve $\theta$ (called \textit{thread}) around a fixed central point,
$\theta(t)=r(t)\exp{(i\varphi(t))}$ with the plane identified with $\ensuremath{\mathbb{C}}$,
$r(t)>0$, $t\in [0,1]$ and $\varphi$ being monotonic with $\varphi([0,1])= [0,2\pi]$.
For each tile $T$ in the part the intersection of the interior
of $T$ with the image of $\theta$ is nonempty and connected.
The threads do not meet or cross each other.
\end{description}
Note that this definition could also be used for tilings with a singular point, where arbitrarily small tiles are clusterd. Apart from this, a huge number of tilings allow weak ring partitions, however, it is not a simple question how to characterize the family of tilings that share this property. We can pose this as an open problem so far.
For the further analysis we need a stronger version of this definition. The condition is analogous to S2 from definition S with `arm' replaced by `part':
\begin{description}
\item [{Definition}] \textbf{\textit{strong ring partition}: }
A tile set partition of a plane tiling with all properties of a weak ring partition is called a \textit{strong ring partition} if an additional condition holds:
If any two tiles $T_{1},T_{2}$ in a part are
direct neighbors and can be respectively
mapped by an operation $\tau$ (composed of translation, rotation or scaling) onto another tile pair $\tau(T_{1})$
and $\tau(T_{2})$, these must also be direct neighbors within
a part. ($T_{1},T_{2}$ from the same part are called
\textit{direct neighbors} if $T_{1}\text{\ensuremath{\cap}}T_{2}$
is cut\footnote{Here and in all other occurrences "cut by the thread" means that the thread (by passing from $T_{1}$ to $T_{2}$) intersects $T_{1}\text{\ensuremath{\cap}}T_{2}$, which might also be just a single vertex.} by the part's thread or contains more than a finite number
of points.)
\end{description}
Figure 5.2 (left) shows an example of a strong ring partition. The scaling operation was inserted here to make this definition applicable also in the context of tilings with singular points, see below in this section. It is obvious that by the techniques of the algorithm presented in section 2 one could automatically check whether a tiling allows or doesn't allow a strong ring partition.
Now we can separate tilings with a spiral structure from those with a ring structure, which sometimes can both occur simultaneously.
\begin{spacing}{0}
\noindent \begin{center}
\includegraphics[width=11cm]{Hirschhorn2}
\par\end{center}
\end{spacing}
\begin{description}
\small
\item [{Figure\ 5.2}] A tiling with strong ring partition (left) and S-partition (right)
\end{description}
\begin{description}
\item [{Definition:}]
A $k$-hedral tiling is called a \textit{strong spiral tiling} (respectively \textit{strong S-tiling} or \textit{strong O-tiling})
if it is an S- or O-tiling and additionally \textit{doesn't} allow a strong ring partition. (Hence, strong spiral tilings and strong ring partitions exclude each other.)
\end{description}
A closer inspection shows that most of the known S- or O-tilings are also strong S- respectively strong O-tilings. A famous example where this is \textit{not} the case is the Hirschhorn tiling. In Figure 5.2 we can observe on the left side the obvious ring structure and on the right side the spiral arms allowing an S-partition.
In the context of tilings with one singular point, we can do the same to separate the ring structure from the spirals.
\begin{description}
\item [{Definition:}]
A tiling with one singular point and finitely many similarity classes is called a \textit{strong spiral tiling} (or \textit{strong P-tiling})
if it allows a partition according to definition P but \textit{doesn't} allow a strong ring partition.
\end{description}
Tilings with strong ring partition and spiral structure - regardless of having a singular point or not - as shown in Figure 5.2 or in \cite{staana} often have the property that the spirals are in some sense hidden or visually dominated by the ring structure. They can be viewed as ``picture puzzles''. So, though we have demonstrated that spirals (and other structures) can be detected principally without human aid by algorithms, in the context of perception the quote from the beginning remains true that ``to some
extent, at least, the spiral effect is psychological\textquotedblright.
\section*{Acknowledgements}
Although it is no longer possible in person, I would like to express
my gratitude to the late Branko Gr\"unbaum for the fruitful discussion
during the development of this paper and thanks to Brian A. Wichmann
for his support with the tilings from his great collection.
|
\section{Introduction}
Motivated by the model uncertainty in finance, Peng \cite{Peng2005, P07a,
P08a} established the theory of $G$-expectation which is a consistent
sublinear expectation and does not require a probability space. The
representation of $G$-expectation as the supremum of expectations over a set
of nondominated probability measures was obtained in \cite{DHP11, HP09}. Due
to this set of nondominated probability measures, the backward stochastic
differential equation (BSDE for short) is completely different from the
classical one. Hu et al. \cite{HJPS1} obtained an existence and uniqueness
theorem for a new kind of BSDE driven by $G$-Brownian motion. In addition,
there are other advances in this direction. Denis and Martini
\cite{DenisMartini2006} developed quasi-sure stochastic analysis. Soner et al.
\cite{STZ11} obtained an existence and uniqueness theorem for a new type of
BSDE ($2$BSDE) under a family of nondominated probability measures.
Recently, Hu and Ji \cite{HJ1} studied the following stochastic recursive
optimal control problem under $G$-expectation
\begin{equation}
\left \{
\begin{array}
[c]{rl
dX_{s}^{t,x,u}= & b(s,X_{s}^{t,x,u},u_{s})ds+h_{ij}(s,X_{s}^{t,x,u
,u_{s})d\langle B^{i},B^{j}\rangle_{s}+\sigma(s,X_{s}^{t,x,u},u_{s})dB_{s},\\
X_{t}^{t,x,u}= & x,
\end{array}
\right. \label{ine1
\end{equation
\begin{equation
\begin{array}
[c]{rl
Y_{s}^{t,x,u}= & \Phi(X_{T}^{t,x,u})+\int_{s}^{T}f(r,X_{r}^{t,x,u
,Y_{r}^{t,x,u},Z_{r}^{t,x,u},u_{r})dr+\int_{s}^{T}g_{ij}(r,X_{r}^{t,x,u
,Y_{r}^{t,x,u},Z_{r}^{t,x,u},u_{r})d\langle B^{i},B^{j}\rangle_{r}\\
& -\int_{s}^{T}Z_{r}^{t,x,u}dB_{r}-(K_{T}^{t,x,u}-K_{s}^{t,xu}),\text{
s\in \lbrack t,T].
\end{array}
\label{ine2
\end{equation}
The value function is defined a
\begin{equation}
V(t,x):=\underset{u\in \mathcal{U}[t,T]}{ess\inf}Y_{t}^{t,x,u}. \label{ine3
\end{equation}
As pointed out in \cite{HJ1}, the value function defined in (\ref{ine3}) is a
$\inf \sup$ problem, which is known as the robust optimal control problem. For
recent development of robust control problem under a set of nondominated
probability measures, we refer the readers to \cite{DK, EJ-1, EJ-2, MPZ, PZ}
and the references therein. When $G$ is linear, the above optimal control
problem is classical stochastic recursive optimal control problem, which was
first studied by Peng in \cite{peng-dpp}. For the development of classical
stochastic recursive optimal control problem, we refer the readers to
\cite{BH, BL, EPQ, HJX, LW, MPY, MY, Tang, WY, J.Yong} and the references therein.
The nonlinear part with respect to $\partial_{xx}^{2}V$ in the HJB equation
related to the optimal control problem (\ref{ine1}) and (\ref{ine2}) is the
$\inf \sup$ of a family of linear part with respect to $\partial_{xx}^{2}V$. Up
to our knowledge, this $\inf \sup$ representation is the only result that has
been made so far in the optimal control problem. In order to obtain the fully
nonlinear representation, we want to study the stochastic recursive optimal
control problem under $\tilde{G}$-expectation. Here $\tilde{G}$ is any
function dominated by $G$ in the meaning of (\ref{newe1}). More precisely, we
consider the following BSDE under $\tilde{G}$-expectation
\begin{equation}
Y_{s}^{t,x,u}=\mathbb{\tilde{E}}_{s}\left[ \Phi(X_{T}^{t,x,u})+\int_{s
^{T}f(r,X_{r}^{t,x,u},Y_{r}^{t,x,u},u_{r})dr+\int_{s}^{T}g_{ij}(r,X_{r
^{t,x,u},Y_{r}^{t,x,u},u_{r})d\langle B^{i},B^{j}\rangle_{r}\right] .
\label{ine4
\end{equation}
The new optimal control problem is (\ref{ine1}) and (\ref{ine4}), and the
value function is still defined as (\ref{ine3}). It is worth pointing out that
the BSDE (\ref{ine4}) under $\tilde{G}$-expectation does not contain $Z$,
which is an important open problem.
In this paper, we study the dynamic programming principle (DPP) and HJB
equation for optimal control problem (\ref{ine1}) and (\ref{ine4}). Firstly,
we establish the comparison theorem for BSDE (\ref{ine4}), which is new in the
literature. Secondly, for each $\xi \in L_{G}^{2}(\Omega)$, we prove that there
exists a sequence of simple random variables $\xi_{k}\in L_{G}^{2}(\Omega)$
such that $\xi_{k}$ converges to $\xi$ in the sense of $L_{G}^{2}$. Based on
this approximation, we give a new method to prove the DPP, which still holds
for the optimal control problem (\ref{ine1}) and (\ref{ine2}) and is easier
than the implied partition method in \cite{HJ1}. At last, we prove that $V$ is
the unique viscosity solution of a type of fully nonlinear HJB equation, which
is not the $\inf \sup$ representation with respect to $\partial_{xx}^{2}V$.
This paper is organized as follows. We recall some basic results on
$G$-expectation and $\tilde{G}$-expectation in Section 2. In Section 3, we
formulate our stochastic recursive optimal control problem under $\tilde{G
$-expectation. In Section 4, we prove the properties of the value function and
obtain the DPP. We prove that the value function is the unique viscosity
solution of a type of fully nonlinear HJB equation in Section 5.
\section{Preliminaries}
Let $T>0$ be given and let $\Omega_{T}=C_{0}([0,T];\mathbb{R}^{d})$ be the
space of $\mathbb{R}^{d}$-valued continuous functions on $[0,T]$ with
$\omega_{0}=0$. The canonical process $B_{t}(\omega):=\omega_{t}$, for
$\omega \in \Omega_{T}$ and $t\in \lbrack0,T]$. Se
\[
Lip(\Omega_{T}):=\{ \varphi(B_{t_{1}},B_{t_{2}}-B_{t_{1}},\ldots,B_{t_{N
}-B_{t_{N-1}}):N\geq1,t_{1}<\cdots<t_{N}\leq T,\varphi \in C_{b.Lip
(\mathbb{R}^{d\times N})\},
\]
where $C_{b.Lip}(\mathbb{R}^{d\times N})$ denotes the space of bounded
Lipschitz functions on $\mathbb{R}^{d\times N}$.
Let $G:\mathbb{S}_{d}\rightarrow \mathbb{R}$ be a given monotonic and sublinear
function, where $\mathbb{S}_{d}$ denotes the set of $d\times d$ symmetric
matrices. Peng \cite{P07a, P08a} constructed a $G$-expectation space
$(\Omega_{T},Lip(\Omega_{T}),\mathbb{\hat{E}},(\mathbb{\hat{E}}_{t
)_{t\in \lbrack0,T]})$, which is a consistent sublinear expectation space. The
canonical process $(B_{t})_{t\in \lbrack0,T]}$ is called $G$-Brownian motion
under $\mathbb{\hat{E}}$. Throughout this paper, we suppose that $G$ is
non-degenerate, i.e., there exists a $\underline{\sigma}^{2}>0$ such that
$G(A)-G(B)\geq \frac{1}{2}\underline{\sigma}^{2}\mathrm{tr}[A-B]$ for any
$A\geq B$. Furthermore, let $\tilde{G}:\mathbb{S}_{d}\rightarrow \mathbb{R}$ be
any given monotonic function dominated by $G$, i.e., for $A_{1}$, $A_{2
\in \mathbb{S}_{d}$
\begin{equation}
\left \{
\begin{array}
[c]{l
\tilde{G}(0)=0,\\
\tilde{G}(A_{1})\geq \tilde{G}(A_{2})\text{ if }A_{1}\geq A_{2},\\
\tilde{G}(A_{1})-\tilde{G}(A_{2})\leq G(A_{1}-A_{2}).
\end{array}
\right. \label{newe1
\end{equation}
Peng also constructed a $\tilde{G}$-expectation space $(\Omega_{T
,Lip(\Omega_{T}),\mathbb{\tilde{E}},(\mathbb{\tilde{E}}_{t})_{t\in \lbrack
0,T]})$ in \cite{Peng2005, P2019}, which is a consistent nonlinear expectation
space satisfyin
\begin{equation}
\mathbb{\tilde{E}}_{t}[X]-\mathbb{\tilde{E}}_{t}[Y]\leq \mathbb{\hat{E}
_{t}[X-Y]\text{ for }X,Y\in Lip(\Omega_{T}),\text{ }t\in \lbrack0,T].
\label{e1
\end{equation}
Denote by $L_{G}^{p}(\Omega_{T})$ the completion of $Lip(\Omega_{T})$ under
the norm $||X||_{L_{G}^{p}}:=(\mathbb{\hat{E}}[|X|^{p}])^{1/p}$ for $p\geq1$.
For each $t\in \lbrack0,T]$, the conditional $G$-expectation and $\tilde{G
$-expectation can be continuously extended to $L_{G}^{1}(\Omega_{T})$ under
the norm $||\cdot||_{L_{G}^{1}}$, and still satisfy the relation (\ref{e1})
for $X,Y\in L_{G}^{1}(\Omega_{T})$.
\begin{definition}
Let $M_{G}^{0}(0,T)$ be the space of simple processes in the following form:
for each $N\in \mathbb{N}$ and $0=t_{0}<\cdots<t_{N}=T$
\[
\eta_{t}=\sum_{k=0}^{N-1}\xi_{k}I_{[t_{k},t_{k+1})}(t),
\]
where $\xi_{k}\in Lip(\Omega_{t_{k}})$ for $k=0,1,\ldots,N-1$.
\end{definition}
Denote by $M_{G}^{p}(0,T)$ the completion of $M_{G}^{0}(0,T)$ under the norm
$||\eta||_{M_{G}^{p}}:=(\mathbb{\hat{E}}[\int_{0}^{T}|\eta_{t}|^{p}dt])^{1/p}$
for $p\geq1$. For each $\eta^{k}\in M_{G}^{2}(0,T)$, $k=1,\ldots,d$, denote
$\eta=(\eta^{1},\ldots,\eta^{d})^{T}\in M_{G}^{2}(0,T;\mathbb{R}^{d})$, the
$G$-It\^{o} integral $\int_{0}^{T}\eta_{t}^{T}dB_{t}$ is well defined, see
Peng \cite{P07a, P08a, P2019}.
\begin{theorem}
(\cite{DHP11, HP09}) There exists a weakly compact set of probability measures
$\mathcal{P}$ on $(\Omega_{T},\mathcal{B}(\Omega_{T}))$ such tha
\[
\mathbb{\hat{E}}[X]=\sup_{P\in \mathcal{P}}E_{P}[X]\text{ for all }X\in
L_{G}^{1}(\Omega_{T}).
\]
$\mathcal{P}$ is called a set that represents $\mathbb{\hat{E}}$.
\end{theorem}
For this $\mathcal{P}$, we define capacit
\[
c(A):=\sup_{P\in \mathcal{P}}P(A)\text{ for }A\in \mathcal{B}(\Omega_{T}).
\]
A set $A\in \mathcal{B}(\Omega_{T})$ is polar if $c(A)=0$. A property holds
\textquotedblleft quasi-surely" (q.s. for short) if it holds outside a polar
set. In the following, we do not distinguish two random variables $X$ and $Y$
if $X=Y$ q.s.
\section{Stochastic optimal control problem}
Let $U$ be a given compact set of $\mathbb{R}^{m}$. For each $t\in \lbrack
0,T]$, we denote by
\[
\mathcal{U}[t,T]:=\{u:u\in M_{G}^{2}(t,T;\mathbb{R}^{m})\text{ with values in
}U\}
\]
the set of admissible controls on $[t,T]$.
In the following, we use Einstein summation convention. For each given
$t\in \lbrack0,T]$, $\xi \in L_{G}^{p}(\Omega_{t};\mathbb{R}^{n})$ with $p\geq2$
and $u\in \mathcal{U}[t,T]$, we consider the following forward and backward
SDEs
\begin{equation}
\left \{
\begin{array}
[c]{rl
dX_{s}^{t,\xi,u}= & b(s,X_{s}^{t,\xi,u},u_{s})ds+h_{ij}(s,X_{s}^{t,\xi
,u},u_{s})d\langle B^{i},B^{j}\rangle_{s}+\sigma(s,X_{s}^{t,\xi,u
,u_{s})dB_{s},\\
X_{t}^{t,\xi,u}= & \xi,
\end{array}
\right. \label{e2
\end{equation}
an
\begin{equation}
Y_{s}^{t,\xi,u}=\mathbb{\tilde{E}}_{s}\left[ \Phi(X_{T}^{t,\xi,u})+\int
_{s}^{T}f(r,X_{r}^{t,\xi,u},Y_{r}^{t,\xi,u},u_{r})dr+\int_{s}^{T
g_{ij}(r,X_{r}^{t,\xi,u},Y_{r}^{t,\xi,u},u_{r})d\langle B^{i},B^{j}\rangle
_{r}\right] , \label{e3
\end{equation}
where $s\in \lbrack t,T]$, $\langle B\rangle=(\langle B^{i},B^{j
\rangle)_{i,j=1}^{d}$ is the quadratic variation of $B$.
Suppose that $b$, $h_{ij}:[0,T]\times \mathbb{R}^{n}\times U\rightarrow
\mathbb{R}^{n}$, $\sigma:[0,T]\times \mathbb{R}^{n}\times U\rightarrow
\mathbb{R}^{n\times d}$, $\Phi:\mathbb{R}^{n}\rightarrow \mathbb{R}$, $f$,
$g_{ij}:[0,T]\times \mathbb{R}^{n}\times \mathbb{R}\times U\rightarrow
\mathbb{R}$ are deterministic functions and satisfy the following conditions:
\begin{description}
\item[(H1)] There exists a constant $L>0$ such that for any $(s,x,y,u)$,
$(s,x^{\prime},y^{\prime},v)\in \lbrack0,T]\times \mathbb{R}^{n}\times
\mathbb{R}\times U$
\
\begin{array}
[c]{l
|b(s,x,u)-b(s,x^{\prime},v)|+|h_{ij}(s,x,u)-h_{ij}(s,x^{\prime},v)|+|\sigma
(s,x,u)-\sigma(s,x^{\prime},v)|\\
\ \ \leq L(|x-x^{\prime}|+|u-v|),\\
|\Phi(x)-\Phi(x^{\prime})|\leq L|x-x^{\prime}|,\\
|f(s,x,y,u)-f(s,x^{\prime},y^{\prime},v)|+|g_{ij}(s,x,y,u)-g_{ij}(s,x^{\prime
},y^{\prime},v)|\\
\ \ \leq L(|x-x^{\prime}|+|y-y^{\prime}|+|u-v|);
\end{array}
\]
\item[(H2)] $h_{ij}=h_{ji}$ and $g_{ij}=g_{ji}$; $b,$ $h_{ij},$ $\sigma,$ $f,$
$g_{ij}$ are continuous in $s$.
\end{description}
We have the following theorems.
\begin{theorem}
(\cite{P2019}) Let Assumptions (H1) and (H2) hold. Then, for each $\xi \in
L_{G}^{2}(\Omega_{t};\mathbb{R}^{n})$ and $u\in \mathcal{U}[t,T]$, there exists
a unique solution $(X,Y)\in M_{G}^{2}(t,T;\mathbb{R}^{n+1})$ for the
forward-backward SDE (\ref{e2}) and (\ref{e3}).
\end{theorem}
\begin{theorem}
\label{th-ex}(\cite{HJ1, P2019}) Let Assumptions (H1) and (H2) hold, and let
$\xi,$ $\xi^{\prime}\in L_{G}^{p}(\Omega_{t};\mathbb{R}^{n})$ with $p\geq2$
and $u,$ $v\in \mathcal{U}[t,T]$. Then, for each $\delta \in \lbrack0,T-t]$, we
hav
\
\begin{array}
[c]{l
\mathbb{\hat{E}}_{t}[|X_{t+\delta}^{t,\xi,u}-X_{t+\delta}^{t,\xi^{\prime
,v}|^{2}]\leq C(|\xi-\xi^{\prime}|^{2}+\mathbb{\hat{E}}_{t}[\int_{t
^{t+\delta}|u_{s}-v_{s}|^{2}ds]),\\
\mathbb{\hat{E}}_{t}[|X_{t+\delta}^{t,\xi,u}|^{p}]\leq C(1+|\xi|^{p}),\\
\mathbb{\hat{E}}_{t}\left[ \underset{s\in \lbrack t,t+\delta]}{\sup
|X_{s}^{t,\xi,u}-\xi|^{p}\right] \leq C(1+|\xi|^{p})\delta^{p/2},
\end{array}
\]
where $C$ depends on $T$, $G$, $p$ and $L$.
\end{theorem}
Our stochastic optimal control problem is to find $u\in \mathcal{U}[t,T]$ which
minimizes the objective function $Y_{t}^{t,x,u}$ for each given $x\in
\mathbb{R}^{n}$. For this purpose, we need the following definition of
\ essential infimum of $\{Y_{t}^{t,x,u}:u\in \mathcal{U}[t,T]\}$.
\begin{definition}
(\cite{HJ1}) The essential infimum of $\{Y_{t}^{t,x,u}:u\in \mathcal{U
[t,T]\}$, denoted by $\underset{u\in \mathcal{U}[t,T]}{ess\inf}Y_{t}^{t,x,u}$,
is a random variable $\zeta \in L_{G}^{2}(\Omega_{t})$ satisfying:
\begin{description}
\item[(i)] for any $u\in \mathcal{U}[t,T]$, $\zeta \leq Y_{t}^{t,x,u}$ q.s.;
\item[(ii)] if $\eta$ is a random variable satisfying $\eta \leq Y_{t}^{t,x,u}$
q.s. for any $u\in \mathcal{U}[t,T]$, then $\zeta \geq \eta$ q.s.
\end{description}
\end{definition}
For each $(t,x)\in \lbrack0,T]\times \mathbb{R}^{n}$, we define the value
functio
\begin{equation}
V(t,x):=\underset{u\in \mathcal{U}[t,T]}{ess\inf}Y_{t}^{t,x,u}. \label{e4
\end{equation}
In the following we will prove that $V(\cdot,\cdot)$ is deterministic and
$V(t,\xi)=\underset{u\in \mathcal{U}[t,T]}{ess\inf}Y_{t}^{t,\xi,u}$ for each
$\xi \in L_{G}^{2}(\Omega_{t};\mathbb{R}^{n})$. Furthermore, we will obtain the
dynamic programming principle and the related fully nonlinear HJB equation.
\section{Dynamic programming principle}
In the following, the constant $C$ will change from line to line in our proof.
We use the following notations: for each given $0\leq t\leq s\leq T$
\
\begin{array}
[c]{l
Lip(\Omega_{s}^{t}):=\{ \varphi(B_{t_{1}}-B_{t},\ldots,B_{t_{N}}-B_{t
):N\geq1,t_{1},\ldots,t_{N}\in \lbrack t,s],\varphi \in C_{b.Lip}(\mathbb{R
^{d\times N})\};\\
L_{G}^{2}(\Omega_{s}^{t}):=\{ \text{the completion of }Lip(\Omega_{s
^{t})\text{ under the norm }||\cdot||_{L_{G}^{2}}\};\\
M_{G}^{0,t}(t,T):=\{ \eta_{s}=\sum_{k=0}^{N-1}\xi_{k}I_{[t_{k},t_{k+1
)}(s):t=t_{0}<\cdots<t_{N}=T,\xi_{k}\in Lip(\Omega_{t_{k}}^{t})\};\\
M_{G}^{2,t}(t,T):=\{ \text{the completion of }M_{G}^{0,t}(t,T)\text{ under the
norm }||\cdot||_{M_{G}^{2}}\};\\
\mathcal{U}^{t}[t,T]:=\{u:u\in M_{G}^{2,t}(t,T;\mathbb{R}^{m})\text{ with
values in }U\};\\
\mathbb{U}[t,T]:=\{u=\sum_{k=1}^{N}I_{A_{k}}u^{k}:N\geq1,u^{k}\in
\mathcal{U}^{t}[t,T],I_{A_{k}}\in L_{G}^{2}(\Omega_{t}),(A_{k})_{k=1
^{N}\text{ is a partition of }\Omega \}.
\end{array}
\]
In order to prove that $V(\cdot,\cdot)$ is deterministic, we need the
following two lemmas. The first lemma can be found in \cite{HJ1}.
\begin{lemma}
(\cite{HJ1}) Let $u\in \mathcal{U}[t,T]$ be given. Then there exists a sequence
$(u^{k})_{k\geq1}$ in $\mathbb{U}[t,T]$ such tha
\[
\lim_{k\rightarrow \infty}\mathbb{\hat{E}}\left[ \int_{t}^{T}|u_{s}-u_{s
^{k}|^{2}ds\right] =0.
\]
\end{lemma}
\begin{lemma}
Let Assumptions (H1) and (H2) hold, and let $\xi \in L_{G}^{2}(\Omega
_{t};\mathbb{R}^{n})$, $u\in \mathcal{U}[t,T]$ and $v=\sum_{k=1}^{N}I_{A_{k
}v^{k}\in \mathbb{U}[t,T]$. Then there exists a constant $C$ depending on $T$,
$G$ and $L$ such tha
\[
\mathbb{\hat{E}}\left[ \left \vert Y_{t}^{t,\xi,u}-\sum_{k=1}^{N}I_{A_{k
}Y_{t}^{t,\xi,v^{k}}\right \vert ^{2}\right] \leq C\mathbb{\hat{E}}\left[
\int_{t}^{T}|u_{s}-v_{s}|^{2}ds\right] .
\]
\end{lemma}
\begin{proof}
Similar to the proof of Lemma 15 in \cite{HJ1}, we can ge
\[
X_{s}^{t,\xi,v}=\sum_{k=1}^{N}I_{A_{k}}X_{s}^{t,\xi,v^{k}}\text{ and
Y_{s}^{t,\xi,v}=\sum_{k=1}^{N}I_{A_{k}}Y_{s}^{t,\xi,v^{k}}\text{ for
s\in \lbrack t,T].
\]
Since $\tilde{G}$-expectation $\mathbb{\tilde{E}}$ is dominated by
$G$-expectation $\mathbb{\hat{E}}$, by (\ref{e3}), we obtain
\[
|Y_{s}^{t,\xi,u}-Y_{s}^{t,\xi,v}|\leq C\mathbb{\hat{E}}_{s}\left[
|X_{T}^{t,\xi,u}-X_{T}^{t,\xi,v}|+\int_{s}^{T}(|Y_{r}^{t,\xi,u}-Y_{r
^{t,\xi,v}|+|X_{r}^{t,\xi,u}-X_{r}^{t,\xi,v}|+|u_{r}-v_{r}|)dr\right] ,
\]
where $s\in \lbrack t,T]$ and $C$ depends on $G$ and $L$. By the H\"{o}lder
inequality, we get
\[
|Y_{s}^{t,\xi,u}-Y_{s}^{t,\xi,v}|^{2}\leq C\mathbb{\hat{E}}_{s}\left[
|X_{T}^{t,\xi,u}-X_{T}^{t,\xi,v}|^{2}+\int_{s}^{T}(|Y_{r}^{t,\xi,u
-Y_{r}^{t,\xi,v}|^{2}+|X_{r}^{t,\xi,u}-X_{r}^{t,\xi,v}|^{2}+|u_{r}-v_{r
|^{2})dr\right] ,
\]
where $s\in \lbrack t,T]$ and $C$ depends on $T$, $G$ and $L$. By the Gronwall
inequality under $\mathbb{\hat{E}}$ (see Theorem 3.10 in \cite{HJPS}), we
deduc
\begin{equation}
|Y_{t}^{t,\xi,u}-Y_{t}^{t,\xi,v}|^{2}\leq C\mathbb{\hat{E}}_{t}\left[
|X_{T}^{t,\xi,u}-X_{T}^{t,\xi,v}|^{2}+\int_{t}^{T}(|X_{r}^{t,\xi,u
-X_{r}^{t,\xi,v}|^{2}+|u_{r}-v_{r}|^{2})dr\right] , \label{e5
\end{equation}
where $C$ depends on $T$, $G$ and $L$. By Theorem \ref{th-ex}, we have
\begin{equation}
\mathbb{\hat{E}}_{t}\left[ |X_{s}^{t,\xi,u}-X_{s}^{t,\xi,v}|^{2}\right] \leq
C\mathbb{\hat{E}}_{t}\left[ \int_{t}^{T}|u_{r}-v_{r}|^{2}dr\right] ,
\label{e6
\end{equation}
where $s\in \lbrack t,T]$ and $C$ depends on $T$, $G$ and $L$. Thus we obtain
the desired result by (\ref{e5}) and (\ref{e6}).
\end{proof}
\begin{theorem}
\label{th-vtx}Let Assumptions (H1) and (H2) hold. Then the value function
$V(t,x)$ exists an
\[
V(t,x)=\inf_{u\in \mathcal{U}^{t}[t,T]}Y_{t}^{t,x,u}.
\]
\end{theorem}
\begin{proof}
The proof is the same as Theorem 17 in \cite{HJ1}. We omit it.
\end{proof}
Now we study the properties of $V(\cdot,\cdot)$.
\begin{proposition}
\label{pro-v}Let Assumptions (H1) and (H2) hold. Then there exists a constant
$C$ depending on $T$, $G$ and $L$ such that, for any $t\in \lbrack0,T]$, $x,$
$y\in \mathbb{R}^{n},
\[
|V(t,x)-V(t,y)|\leq C|x-y|\text{ and }|V(t,x)|\leq C(1+|x|).
\]
\end{proposition}
\begin{proof}
Similar to the proof of inequality (\ref{e5}), we can obtain that, for any
$u\in \mathcal{U}^{t}[t,T]$
\begin{equation}
|Y_{t}^{t,x,u}-Y_{t}^{t,y,u}|^{2}\leq C\mathbb{\hat{E}}_{t}\left[
|X_{T}^{t,x,u}-X_{T}^{t,y,u}|^{2}+\int_{t}^{T}|X_{r}^{t,x,u}-X_{r
^{t,y,u}|^{2}dr\right] , \label{e7
\end{equation}
where $C$ depends on $T$, $G$ and $L$. By Theorem \ref{th-ex}, we hav
\begin{equation}
\mathbb{\hat{E}}_{t}\left[ |X_{s}^{t,x,u}-X_{s}^{t,y,u}|^{2}\right] \leq
C|x-y|^{2}, \label{e8
\end{equation}
where $s\in \lbrack t,T]$ and $C$ depends on $T$, $G$ and $L$. Thus we get
$|V(t,x)-V(t,y)|\leq C|x-y|$ by (\ref{e7}) and (\ref{e8}). Similarly, we can
obtain $|V(t,x)|\leq C(1+|x|)$.
\end{proof}
\begin{theorem}
\label{th-vxi}Let Assumptions (H1) and (H2) hold. Then, for any $\xi \in
L_{G}^{2}(\Omega_{t};\mathbb{R}^{n})$, we hav
\[
V(t,\xi)=\underset{u\in \mathcal{U}[t,T]}{ess\inf}Y_{t}^{t,\xi,u}.
\]
\end{theorem}
\begin{proof}
For each given $u\in \mathcal{U}[t,T]$, we first prove that $V(t,\xi)\leq
Y_{t}^{t,\xi,u}$ q.s.
For each $\varepsilon>0$, we can find a $\xi_{\varepsilon}=\sum_{k=1}^{\infty
}x_{k}I_{A_{k}}$ such that $|\xi-\xi_{\varepsilon}|\leq \varepsilon$, where
$x_{k}\in \mathbb{R}^{n}$ and $\{A_{k}\}_{k=1}^{\infty}$ is a $\mathcal{B
(\Omega_{t})$-partition of $\Omega$. By Proposition \ref{pro-v}, we ge
\begin{equation}
\left \vert V(t,\xi)-\sum_{k=1}^{\infty}V(t,x_{k})I_{A_{k}}\right \vert
=\left \vert V(t,\xi)-V(t,\xi_{\varepsilon})\right \vert \leq C\varepsilon.
\label{e9
\end{equation}
Similar to the proof of inequalities (\ref{e7}) and (\ref{e8}), we can ge
\[
|Y_{t}^{t,\xi,u}-Y_{t}^{t,x_{k},u}|\leq C|\xi-x_{k}|,\text{ }k\geq1,
\]
where $C$ depends on $T$, $G$ and $L$. Then, we obtai
\begin{equation}
\left \vert Y_{t}^{t,\xi,u}-\sum_{k=1}^{\infty}Y_{t}^{t,x_{k},u}I_{A_{k
}\right \vert =\sum_{k=1}^{\infty}|Y_{t}^{t,\xi,u}-Y_{t}^{t,x_{k},u}|I_{A_{k
}\leq C|\xi-\xi_{\varepsilon}|\leq C\varepsilon. \label{e10
\end{equation}
By (\ref{e4}), we hav
\begin{equation}
\sum_{k=1}^{\infty}V(t,x_{k})I_{A_{k}}\leq \sum_{k=1}^{\infty}Y_{t}^{t,x_{k
,u}I_{A_{k}},\text{ q.s.} \label{e11
\end{equation}
It follows from (\ref{e9}), (\ref{e10}) and (\ref{e11}) tha
\[
V(t,\xi)\leq Y_{t}^{t,\xi,u}+C\varepsilon,\text{ q.s.,
\]
where $C$ is independent of $\varepsilon$. Thus we obtain $V(t,\xi)\leq
Y_{t}^{t,\xi,u}$ q.s.
Second, if $\eta$ is a random variable satisfying $\eta \leq Y_{t}^{t,\xi,u}$
q.s. for any $u\in \mathcal{U}[t,T]$, then we prove that $V(t,\xi)\geq \eta$ q.s.
It is easy to verify that the constant $C$ in inequality (\ref{e10}) is
independent of $u$. Then we obtai
\[
\eta \leq \sum_{k=1}^{\infty}Y_{t}^{t,x_{k},u}I_{A_{k}}+C\varepsilon,\text{
q.s., for any }u\in \mathcal{U}[t,T],
\]
where $C$ depends on $T$, $G$ and $L$. By Theorem \ref{th-vtx} and the above
inequality, we can ge
\begin{equation}
\eta \leq \sum_{k=1}^{\infty}V(t,x_{k})I_{A_{k}}+C\varepsilon,\text{ q.s.}
\label{e12
\end{equation}
Thus we obtain $V(t,\xi)\geq \eta$ q.s. by (\ref{e9}) and (\ref{e12}), which
implies the desired result.
\end{proof}
Finally, we study the dynamic programming principle. The following lemma is
useful in deriving the dynamic programming principle.
\begin{lemma}
\label{le-apxi}Let $\xi \in L_{G}^{2}(\Omega_{s})$ with fixed $s\in \lbrack
0,T]$. Then there exists a sequence $\xi_{k}=\sum_{i=1}^{N_{k}}x_{i
^{k}I_{A_{i}^{k}}$, $k\geq1$, such tha
\[
\lim_{k\rightarrow \infty}\mathbb{\hat{E}}\left[ |\xi-\xi_{k}|^{2}\right]
=0,
\]
where $x_{i}^{k}\in \mathbb{R}$, $I_{A_{i}^{k}}\in L_{G}^{2}(\Omega_{s})$,
$i\leq N_{k}$, $k\geq1$ and $(A_{i}^{k})_{i=1}^{N_{k}}$ is a $\mathcal{B
(\Omega_{s})$-partition of $\Omega$.
\end{lemma}
\begin{proof}
Since $L_{G}^{2}(\Omega_{s})$ is the completion of $Lip(\Omega_{s})$ under the
norm $||\cdot||_{2}$, we only need to prove the case
\[
\xi=\varphi(B_{t_{1}},B_{t_{2}}-B_{t_{1}},\ldots,B_{t_{N}}-B_{t_{N-1}}),
\]
where $N\geq1$, $0<t_{1}<\cdots<t_{N}\leq s$, $\varphi \in C_{b.Lip
(\mathbb{R}^{d\times N})$.
By Theorem 3.20 in \cite{HWZ}, we know that
\[
I_{\{(B_{t_{1}},B_{t_{2}}-B_{t_{1}},\ldots,B_{t_{N}}-B_{t_{N-1}})\in \lbrack
c,c^{\prime})\}}\in L_{G}^{2}(\Omega_{s})
\]
for each $c$, $c^{\prime}\in \mathbb{R}^{d\times N}$ with $c\leq c^{\prime}$.
For each $k\geq1$, we can find
\[
A_{i}^{k}=\{(B_{t_{1}},B_{t_{2}}-B_{t_{1}},\ldots,B_{t_{N}}-B_{t_{N-1}
)\in \lbrack c_{i,k},c_{i,k}^{\prime})\},i=1,\ldots,N_{k}-1,
\]
such that $[-ke,ke)=\cup_{i\leq N_{k}-1}[c_{i,k},c_{i,k}^{\prime})$ with
$e=[1,\ldots,1]^{T}\in \mathbb{R}^{d\times N}$, $|c_{i,k}^{\prime}-c_{i,k}|\leq
k^{-1}$ and $A_{i}^{k}\cap A_{j}^{k}=\emptyset$ for $i\neq j$. Set $A_{N_{k
}^{k}=\Omega \backslash \cup_{i\leq N_{k}-1}A_{i}^{k}$ and
\[
\xi_{k}=\sum_{i=1}^{N_{k}-1}\varphi(c_{i,k})I_{A_{i}^{k}}+0I_{A_{N_{k}}^{k}}.
\]
Then we obtai
\[
|\xi-\xi_{k}|\leq \frac{L_{\varphi}}{k}+\frac{M_{\varphi}}{k}(|B_{t_{1
}|+|B_{t_{2}}-B_{t_{1}}|+\cdots+|B_{t_{N}}-B_{t_{N-1}}|),
\]
where $L_{\varphi}$ is the Lipschitz constant of $\varphi$ and $M_{\varphi}$
is the bound of $\varphi$. Thu
\[
\mathbb{\hat{E}}\left[ |\xi-\xi_{k}|^{2}\right] \leq \frac{C}{k^{2}},
\]
which yields the desired result.
\end{proof}
In order to give the dynamic programming principle, we define the following
backward semigroup $\mathbb{G}_{t,t+\delta}^{t,x,u}[\cdot]$ which was first
introduced by Peng in \cite{peng-dpp-1}.
For each given $(t,x)\in \lbrack0,T)\times \mathbb{R}^{n}$, $\delta \in
\lbrack0,T-t]$, $u\in \mathcal{U}[t,t+\delta]$ and $\eta \in L_{G}^{2
(\Omega_{t+\delta})$, defin
\[
\mathbb{G}_{s,t+\delta}^{t,x,u}[\eta]=\tilde{Y}_{s}^{t,x,u}\text{ for
s\in \lbrack t,t+\delta],
\]
where $(X_{s}^{t,x,u},\tilde{Y}_{s}^{t,x,u})_{s\in \lbrack t,t+\delta]}$ is the
solution of the following forward and backward SDEs:
\begin{equation}
\left \{
\begin{array}
[c]{rl
dX_{s}^{t,x,u}= & b(s,X_{s}^{t,x,u},u_{s})ds+h_{ij}(s,X_{s}^{t,x,u
,u_{s})d\langle B^{i},B^{j}\rangle_{s}+\sigma(s,X_{s}^{t,x,u},u_{s})dB_{s},\\
X_{t}^{t,x,u}= & x,
\end{array}
\right. \label{newe2
\end{equation}
an
\begin{equation}
\tilde{Y}_{s}^{t,x,u}=\mathbb{\tilde{E}}_{s}\left[ \eta+\int_{s}^{t+\delta
}f(r,X_{r}^{t,x,u},\tilde{Y}_{r}^{t,x,u},u_{r})dr+\int_{s}^{t+\delta
g_{ij}(r,X_{r}^{t,x,u},\tilde{Y}_{r}^{t,x,u},u_{r})d\langle B^{i},B^{j
\rangle_{r}\right] . \label{e13
\end{equation}
The following lemma is the comparison theorem of backward SDE under
$\mathbb{\tilde{E}}$.
\begin{lemma}
\label{le-com}Let Assumptions (H1) and (H2) hold, and let $(t,x)\in
\lbrack0,T)\times \mathbb{R}^{n}$, $\delta \in \lbrack0,T-t]$, $u\in
\mathcal{U}[t,t+\delta]$ and $\eta_{1}$, $\eta_{2}\in L_{G}^{2}(\Omega
_{t+\delta})$ be given. If $\eta_{1}\geq \eta_{2}$ q.s., then $\mathbb{G
_{t,t+\delta}^{t,x,u}[\eta_{1}]\geq \mathbb{G}_{t,t+\delta}^{t,x,u}[\eta_{2}]$ q.s.
\end{lemma}
\begin{proof}
Denote $Y_{s}^{1}=\mathbb{G}_{s,t+\delta}^{t,x,u}[\eta_{1}]$, $Y_{s
^{2}=\mathbb{G}_{s,t+\delta}^{t,x,u}[\eta_{2}]$, $\hat{Y}_{s}=Y_{s}^{1
-Y_{s}^{2}$ for $s\in \lbrack t,t+\delta]$, and $\hat{\eta}=\eta_{1}-\eta_{2}$.
For each given $\varepsilon>0$, just like the proof of Theorem 3.6 in
\cite{HJPS}, we can find $(a_{s}^{\varepsilon})_{s\in \lbrack t,t+\delta]}$,
$(m_{s}^{\varepsilon})_{s\in \lbrack t,t+\delta]}$, $(c_{s}^{ij,\varepsilon
})_{s\in \lbrack t,t+\delta]}$, $(n_{s}^{ij,\varepsilon})_{s\in \lbrack
t,t+\delta]}\in M_{G}^{2}(t,t+\delta)$ such that $|a_{s}^{\varepsilon}|\leq
L$, $|c_{s}^{ij,\varepsilon}|\leq L$, $|m_{s}^{\varepsilon}|\leq2L\varepsilon
$, $|n_{s}^{ij,\varepsilon}|\leq2L\varepsilon$
\[
f(r,X_{r}^{t,x,u},Y_{r}^{1},u_{r})-f(r,X_{r}^{t,x,u},Y_{r}^{2},u_{r
)=a_{r}^{\varepsilon}\hat{Y}_{r}+m_{r}^{\varepsilon
\]
an
\[
g_{ij}(r,X_{r}^{t,x,u},Y_{r}^{1},u_{r})-g_{ij}(r,X_{r}^{t,x,u},Y_{r}^{2
,u_{r})=c_{r}^{ij,\varepsilon}\hat{Y}_{r}+n_{r}^{ij,\varepsilon}.
\]
The
\begin{equation}
\hat{Y}_{s}=\mathbb{\tilde{E}}_{s}\left[ \hat{\eta}+\tilde{\eta}+\int
_{s}^{t+\delta}(a_{r}^{\varepsilon}\hat{Y}_{r}+m_{r}^{\varepsilon})dr+\int
_{s}^{t+\delta}(c_{r}^{ij,\varepsilon}\hat{Y}_{r}+n_{r}^{ij,\varepsilon
})d\langle B^{i},B^{j}\rangle_{r}\right] -\mathbb{\tilde{E}}_{s}[\tilde{\eta
}], \label{e15
\end{equation}
where $s\in \lbrack t,t+\delta]$ and $\tilde{\eta}=\eta_{2}+\int_{t}^{t+\delta
}f(r,X_{r}^{t,x,u},Y_{r}^{2},u_{r})dr+\int_{t}^{t+\delta}g_{ij}(r,X_{r
^{t,x,u},Y_{r}^{2},u_{r})d\langle B^{i},B^{j}\rangle_{r}$.
For each given $k\geq1$, set $t_{l}^{k}=t+l\delta k^{-1}$, $l=0$,$1$,$\ldots
$,$k$. By (\ref{e15}), one can check that, for $s\in \lbrack t_{l}^{k
,t_{l+1}^{k}]$, $l=k-1$,$\ldots$,$0$
\begin{equation}
\hat{Y}_{s}=\mathbb{\tilde{E}}_{s}\left[ \hat{Y}_{t_{l+1}^{k}}+\tilde{\eta
}+\int_{s}^{t_{l+1}^{k}}(a_{r}^{\varepsilon}\hat{Y}_{r}+m_{r}^{\varepsilon
})dr+\int_{s}^{t_{l+1}^{k}}(c_{r}^{ij,\varepsilon}\hat{Y}_{r}+n_{r
^{ij,\varepsilon})d\langle B^{i},B^{j}\rangle_{r}\right] -\mathbb{\tilde{E
}_{s}[\tilde{\eta}]. \label{e16
\end{equation}
Define $(\hat{Y}_{l}^{k})_{l=0}^{n}$ backwardly as follows: set $\hat{Y
_{k}^{k}=\hat{\eta}$, for $l=k-1$,$\ldots$,$0$
\begin{equation}
\hat{Y}_{l}^{k}=\mathbb{\tilde{E}}_{t_{l}^{k}}\left[ \hat{Y}_{l+1}^{k
+\tilde{\eta}+\int_{t_{l}^{k}}^{t_{l+1}^{k}}(a_{r}^{\varepsilon}\hat{Y
_{l+1}^{k}+m_{r}^{\varepsilon})dr+\int_{t_{l}^{k}}^{t_{l+1}^{k}
(c_{r}^{ij,\varepsilon}\hat{Y}_{l+1}^{k}+n_{r}^{ij,\varepsilon})d\langle
B^{i},B^{j}\rangle_{r}\right] -\mathbb{\tilde{E}}_{t_{l}^{k}}[\tilde{\eta}].
\label{e17
\end{equation}
Note that $|\int_{s_{1}}^{s_{2}}\zeta_{r}d\langle B^{i},B^{j}\rangle_{r
|\leq(\mathbb{\hat{E}}\left[ |B^{i}|^{2}\right] \mathbb{\hat{E}}\left[
|B^{j}|^{2}\right] )^{1/2}\int_{s_{1}}^{s_{2}}|\zeta_{r}|dr$ for each $s_{1
$, $s_{2}\in \lbrack t,t+\delta]$ and $\zeta \in M_{G}^{1}(t,t+\delta)$, then
one can verify tha
\[
\left \vert \int_{t_{l}^{k}}^{t_{l+1}^{k}}a_{r}^{\varepsilon}dr+\int_{t_{l
^{k}}^{t_{l+1}^{k}}c_{r}^{ij,\varepsilon}d\langle B^{i},B^{j}\rangle
_{r}\right \vert \leq C\int_{t_{l}^{k}}^{t_{l+1}^{k}}(|a_{r}^{\varepsilon
}|+|c_{r}^{ij,\varepsilon}|)dr\leq Ck^{-1
\]
an
\[
\left \vert \int_{t_{l}^{k}}^{t_{l+1}^{k}}m_{r}^{\varepsilon}dr+\int_{t_{l
^{k}}^{t_{l+1}^{k}}n_{r}^{ij,\varepsilon}d\langle B^{i},B^{j}\rangle
_{r}\right \vert \leq C\int_{t_{l}^{k}}^{t_{l+1}^{k}}(|m_{r}^{\varepsilon
}|+|n_{r}^{ij,\varepsilon}|)dr\leq C\varepsilon k^{-1},
\]
where $C$ is dependent of $L$ and $\delta$ and independent of $l$. For each
$k\geq k_{0}$ with $Ck_{0}^{-1}\leq2^{-1}$, we hav
\[
\hat{Y}_{k-1}^{k}\geq \mathbb{\tilde{E}}_{t_{k-1}^{k}}[\tilde{\eta
}-C\varepsilon k^{-1}]-\mathbb{\tilde{E}}_{t_{k-1}^{k}}[\tilde{\eta
}]=-C\varepsilon k^{-1
\]
an
\[
\hat{Y}_{k-2}^{k}\geq \mathbb{\tilde{E}}_{t_{k-2}^{k}}[-(1+Ck^{-1})C\varepsilon
k^{-1}+\tilde{\eta}-C\varepsilon k^{-1}]-\mathbb{\tilde{E}}_{t_{k-2}^{k
}[\tilde{\eta}]=-[(1+Ck^{-1})+1]C\varepsilon k^{-1}.
\]
Continuing this process, we obtain
\begin{equation}
\hat{Y}_{0}^{k}\geq-C\varepsilon k^{-1}\sum_{l=0}^{k-1}(1+Ck^{-1})^{l
\geq-(e^{C}-1)\varepsilon. \label{e18
\end{equation}
For each given $\eta \in Lip(\Omega_{t+\delta})$, define $\phi(s_{1
,s_{2})=\mathbb{\hat{E}}[|\mathbb{\tilde{E}}_{s_{1}}[\eta]-\mathbb{\tilde{E
}_{s_{2}}[\eta]|]$ for $s_{1}$, $s_{2}\in \lbrack t,t+\delta]$. By the
definition of $\mathbb{\tilde{E}}_{s}[\eta]$, one can verify that $\phi$ is a
continuous function. Then we ge
\begin{equation}
\sup_{|s_{1}-s_{2}|\leq \delta k^{-1}}\mathbb{\hat{E}}[|\mathbb{\tilde{E
}_{s_{1}}[\eta]-\mathbb{\tilde{E}}_{s_{2}}[\eta]|]\rightarrow0\text{ as
}k\rightarrow \infty \text{.} \label{e19
\end{equation}
Note tha
\[
Y_{s}^{2}=\mathbb{\tilde{E}}_{s}[\tilde{\eta}]-\int_{t}^{s}f(r,X_{r
^{t,x,u},Y_{r}^{2},u_{r})dr-\int_{t}^{s}g_{ij}(r,X_{r}^{t,x,u},Y_{r}^{2
,u_{r})d\langle B^{i},B^{j}\rangle_{r},
\]
then, by (\ref{e19}) and $\tilde{\eta}\in L_{G}^{2}(\Omega_{t+\delta})$, one
can check tha
\begin{equation}
\sup_{|s_{1}-s_{2}|\leq \delta k^{-1}}\mathbb{\hat{E}}[|Y_{s_{1}}^{2}-Y_{s_{2
}^{2}|]\rightarrow0\text{ as }k\rightarrow \infty \text{.} \label{e20
\end{equation}
Similarly, the relation (\ref{e20}) still holds for $Y^{1}$. Thus we obtai
\begin{equation}
\gamma_{k}:=\sup_{|s_{1}-s_{2}|\leq \delta k^{-1}}\mathbb{\hat{E}}[|\hat
{Y}_{s_{1}}-\hat{Y}_{s_{2}}|]\rightarrow0\text{ as }k\rightarrow \infty \text{.}
\label{e21
\end{equation}
Define $\Delta_{l}^{k}=\hat{Y}_{t_{l}^{k}}-\hat{Y}_{l}^{k}$ for $l=0
,$1$,$\ldots$,$k$. By (\ref{e16}), (\ref{e17}) and (\ref{e21}), we ge
\begin{equation}
\mathbb{\hat{E}}[|\Delta_{l}^{k}|]\leq(1+Ck^{-1})\mathbb{\hat{E}
[|\Delta_{l+1}^{k}|]+Ck^{-1}\gamma_{k}, \label{e22
\end{equation}
where $l=k-1$,$\ldots$,$0$, $\Delta_{k}^{k}=0$, $C$ depends on $L$ and
$\delta$. Similar to (\ref{e18}), we deduc
\begin{equation}
\mathbb{\hat{E}}[|\Delta_{0}^{k}|]=\mathbb{\hat{E}}[|\hat{Y}_{t}-\hat{Y
_{0}^{k}|]\leq(e^{C}-1)\gamma_{k}. \label{e23
\end{equation}
It follows from (\ref{e18}), (\ref{e21}) and (\ref{e23}) that $\hat{Y}_{t
\geq-(e^{C}-1)\varepsilon$ q.s. Since $\varepsilon$ is arbitrary, we obtain
the desired result.
\end{proof}
The following theorem is the dynamic programming principle.
\begin{theorem}
\label{DPP}Let Assumptions (H1) and (H2) hold. Then, for each $(t,x)\in
\lbrack0,T)\times \mathbb{R}^{n}$, $\delta \in \lbrack0,T-t]$, we hav
\begin{equation}
V(t,x)=\underset{u\in \mathcal{U}[t,t+\delta]}{ess\inf}\mathbb{G}_{t,t+\delta
}^{t,x,u}[V(t+\delta,X_{t+\delta}^{t,x,u})]=\inf_{u\in \mathcal{U
^{t}[t,t+\delta]}\mathbb{G}_{t,t+\delta}^{t,x,u}[V(t+\delta,X_{t+\delta
}^{t,x,u})]. \label{e14
\end{equation}
\end{theorem}
\begin{proof}
By Theorem \ref{th-vtx}, we hav
\[
\underset{u\in \mathcal{U}[t,t+\delta]}{ess\inf}\mathbb{G}_{t,t+\delta
^{t,x,u}[V(t+\delta,X_{t+\delta}^{t,x,u})]=\inf_{u\in \mathcal{U
^{t}[t,t+\delta]}\mathbb{G}_{t,t+\delta}^{t,x,u}[V(t+\delta,X_{t+\delta
}^{t,x,u})].
\]
For any $u\in \mathcal{U}^{t}[t,T]$, by Theorem \ref{th-vxi}, we ge
\[
Y_{t+\delta}^{t,x,u}=Y_{t+\delta}^{t+\delta,X_{t+\delta}^{t,x,u},u}\geq
V(t+\delta,X_{t+\delta}^{t,x,u})\text{ q.s.
\]
Then, by Lemma \ref{le-com}, we obtai
\[
Y_{t}^{t,x,u}=\mathbb{G}_{t,t+\delta}^{t,x,u}[Y_{t+\delta}^{t,x,u
]\geq \mathbb{G}_{t,t+\delta}^{t,x,u}[V(t+\delta,X_{t+\delta}^{t,x,u})],
\]
which implie
\[
V(t,x)\geq \inf_{u\in \mathcal{U}^{t}[t,t+\delta]}\mathbb{G}_{t,t+\delta
}^{t,x,u}[V(t+\delta,X_{t+\delta}^{t,x,u})].
\]
Now we prove the converse inequality. For each given $\varepsilon>0$, there
exists a $v\in \mathcal{U}^{t}[t,t+\delta]$ such tha
\begin{equation}
\mathbb{G}_{t,t+\delta}^{t,x,v}[V(t+\delta,X_{t+\delta}^{t,x,v})]\leq
\varepsilon+\inf_{u\in \mathcal{U}^{t}[t,t+\delta]}\mathbb{G}_{t,t+\delta
}^{t,x,u}[V(t+\delta,X_{t+\delta}^{t,x,u})]. \label{e25
\end{equation}
Since $X_{t+\delta}^{t,x,v}\in L_{G}^{2}(\Omega_{t+\delta}^{t})$, by Lemma
\ref{le-apxi}, we can find a sequence $\xi_{k}=\sum_{l=1}^{N_{k}}x_{l
^{k}I_{A_{l}^{k}}$, $k\geq1$, such tha
\begin{equation}
\mathbb{\hat{E}}\left[ |X_{t+\delta}^{t,x,v}-\xi_{k}|^{2}\right] \leq
k^{-1}, \label{e24
\end{equation}
where $x_{l}^{k}\in \mathbb{R}$, $I_{A_{l}^{k}}\in L_{G}^{2}(\Omega_{t+\delta
}^{t})$, $l\leq N_{k}$, $k\geq1$ and $(A_{l}^{k})_{l=1}^{N_{k}}$ is a
$\mathcal{B}(\Omega_{t+\delta}^{t})$-partition of $\Omega$. For each
$x_{l}^{k}$, we can find $v_{l}^{k}\in \mathcal{U}^{t+\delta}[t+\delta,T]$ such
tha
\begin{equation}
V(t+\delta,x_{l}^{k})\leq Y_{t+\delta}^{t+\delta,x_{l}^{k},v_{l}^{k}}\leq
V(t+\delta,x_{l}^{k})+\varepsilon. \label{e26
\end{equation}
Se
\[
v^{k}(s)=\sum_{l=1}^{N_{k}}v_{l}^{k}(s)I_{A_{l}^{k}}\text{ for }s\in \lbrack
t+\delta,T],
\]
an
\[
u^{k}(s)=v(s)I_{[t,t+\delta)}(s)+v^{k}(s)I_{[t+\delta,T]}(s)\text{ for
s\in \lbrack t,T],
\]
it is easy to verify that $v^{k}\in \mathcal{U}[t+\delta,T]$ and $u^{k
\in \mathcal{U}^{t}[t,T]$. Thus we ge
\begin{equation}
V(t,x)\leq Y_{t}^{t,x,u^{k}}=\mathbb{G}_{t,t+\delta}^{t,x,v}[Y_{t+\delta
}^{t,x,u^{k}}]=\mathbb{G}_{t,t+\delta}^{t,x,v}[Y_{t+\delta}^{t+\delta
,X_{t+\delta}^{t,x,v},v^{k}}]. \label{e27
\end{equation}
Similarly to the proof of inequality (\ref{e5}), we obtain tha
\begin{equation}
\left \vert \mathbb{G}_{t,t+\delta}^{t,x,v}[Y_{t+\delta}^{t+\delta,X_{t+\delta
}^{t,x,v},v^{k}}]-\mathbb{G}_{t,t+\delta}^{t,x,v}[V(t+\delta,X_{t+\delta
}^{t,x,v})]\right \vert ^{2}\leq C\mathbb{\hat{E}}\left[ \left \vert
Y_{t+\delta}^{t+\delta,X_{t+\delta}^{t,x,v},v^{k}}-V(t+\delta,X_{t+\delta
}^{t,x,v})\right \vert ^{2}\right] \label{e28
\end{equation}
an
\begin{equation}
\mathbb{\hat{E}}\left[ \left \vert Y_{t+\delta}^{t+\delta,X_{t+\delta
^{t,x,v},v^{k}}-Y_{t+\delta}^{t+\delta,\xi_{k},v^{k}}\right \vert ^{2}\right]
\leq C\sup_{s\in \lbrack t+\delta,T]}\mathbb{\hat{E}}\left[ \left \vert
X_{s}^{t+\delta,X_{t+\delta}^{t,x,v},v^{k}}-X_{s}^{t+\delta,\xi_{k},v^{k
}\right \vert ^{2}\right] , \label{e29
\end{equation}
where $C$ depends on $T$, $G$ and $L$. By Theorem \ref{th-ex}, (\ref{e24}) and
(\ref{e29}), we hav
\begin{equation}
\mathbb{\hat{E}}\left[ \left \vert Y_{t+\delta}^{t+\delta,X_{t+\delta
^{t,x,v},v^{k}}-Y_{t+\delta}^{t+\delta,\xi_{k},v^{k}}\right \vert ^{2}\right]
\leq C\mathbb{\hat{E}}\left[ \left \vert X_{t+\delta}^{t,x,v}-\xi
_{k}\right \vert ^{2}\right] \leq Ck^{-1}, \label{e30
\end{equation}
where $C$ depends on $T$, $G$ and $L$. It is easy to check tha
\begin{equation}
Y_{t+\delta}^{t+\delta,\xi_{k},v^{k}}=\sum_{l=1}^{N_{k}}Y_{t+\delta
^{t+\delta,x_{l}^{k},v_{l}^{k}}I_{A_{l}^{k}}. \label{e31
\end{equation}
It follows from (\ref{e26}) and (\ref{e31}) tha
\begin{equation}
V(t+\delta,\xi_{k})\leq Y_{t+\delta}^{t+\delta,\xi_{k},v^{k}}\leq
V(t+\delta,\xi_{k})+\varepsilon. \label{e32
\end{equation}
By Proposition \ref{pro-v} and (\ref{e32}), we obtai
\begin{equation}
\mathbb{\hat{E}}\left[ \left \vert Y_{t+\delta}^{t+\delta,\xi_{k},v^{k
}-V(t+\delta,X_{t+\delta}^{t,x,v})\right \vert ^{2}\right] \leq C\left(
\varepsilon^{2}+\mathbb{\hat{E}}\left[ \left \vert X_{t+\delta}^{t,x,v
-\xi_{k}\right \vert ^{2}\right] \right) \leq C(\varepsilon^{2}+k^{-1}),
\label{e33
\end{equation}
where $C$ depends on $T$, $G$ and $L$. By (\ref{e25}), (\ref{e27}),
(\ref{e28}), (\ref{e30}) and (\ref{e33}), we deduce tha
\[
V(t,x)\leq C(\varepsilon+\sqrt{k^{-1}})+\inf_{u\in \mathcal{U}^{t}[t,t+\delta
]}\mathbb{G}_{t,t+\delta}^{t,x,u}[V(t+\delta,X_{t+\delta}^{t,x,u})],
\]
which implies the desired result by letting $k\rightarrow \infty$ and then
$\varepsilon \downarrow0$.
\end{proof}
\begin{remark}
In the above proof, we use Lemma \ref{le-apxi} to find $v^{k}$, which can be
used to simplify the proof of the dynamic programming principle and is easier
than the implied partition method in \cite{HJ1}.
\end{remark}
Now we use the dynamic programming principle to prove the continuity of
$V(\cdot,\cdot)$ in $t$.
\begin{lemma}
\label{newle333}Let Assumptions (H1) and (H2) hold. Then the value function
$V(\cdot,\cdot)$ is $\frac{1}{2}$ H\"{o}lder continuous in $t$.
\end{lemma}
\begin{proof}
For each $(t,x)\in \lbrack0,T)\times \mathbb{R}^{n}$, $\delta \in \lbrack0,T-t]$,
by Theorem \ref{DPP}, we ge
\begin{equation}
|V(t,x)-V(t+\delta,x)|\leq \sup_{u\in \mathcal{U}^{t}[t,t+\delta]
|\mathbb{G}_{t,t+\delta}^{t,x,u}[V(t+\delta,X_{t+\delta}^{t,x,u
)]-V(t+\delta,x)|. \label{e34
\end{equation}
For each given $u\in \mathcal{U}^{t}[t,t+\delta]$, by the definition of the
backward semigroup, we know $\mathbb{G}_{t,t+\delta}^{t,x,u}[V(t+\delta
,X_{t+\delta}^{t,x,u})]=Y_{t}$, where $(Y_{s})_{s\in \lbrack t,t+\delta]}$ is
the solution of the following backward SDE
\[
Y_{s}=\mathbb{\tilde{E}}_{s}\left[ V(t+\delta,X_{t+\delta}^{t,x,u})+\int
_{s}^{t+\delta}f(r,X_{r}^{t,x,u},Y_{r},u_{r})dr+\int_{s}^{t+\delta
g_{ij}(r,X_{r}^{t,x,u},Y_{r},u_{r})d\langle B^{i},B^{j}\rangle_{r}\right] .
\]
By Assumptions (H1), (H2) and Proposition \ref{pro-v}, one can verify tha
\[
|Y_{s}-V(t+\delta,x)|\leq C\mathbb{\hat{E}}_{s}\left[ |X_{t+\delta
^{t,x,u}-x|+\int_{s}^{t+\delta}(1+|x|+|X_{r}^{t,x,u}|+|Y_{r}-V(t+\delta
,x)|)dr\right] ,
\]
where $C$ depends on $T$, $G$ and $L$. It follows from the Gronwall inequality
under $\mathbb{\hat{E}}$ tha
\[
|Y_{t}-V(t+\delta,x)|\leq C\mathbb{\hat{E}}_{t}\left[ |X_{t+\delta
^{t,x,u}-x|+\int_{t}^{t+\delta}(1+|x|+|X_{r}^{t,x,u}|)dr\right] ,
\]
where $C$ depends on $T$, $G$ and $L$. Since $\mathbb{\hat{E}}_{t
[|X_{t+\delta}^{t,x,u}-x|]\leq(\mathbb{\hat{E}}_{t}[|X_{t+\delta
^{t,x,u}-x|^{2}])^{1/2}$ and $\mathbb{\hat{E}}_{t}[|X_{r}^{t,x,u
|]\leq(\mathbb{\hat{E}}_{t}[|X_{r}^{t,x,u}|^{2}])^{1/2}$, we obtai
\[
|\mathbb{G}_{t,t+\delta}^{t,x,u}[V(t+\delta,X_{t+\delta}^{t,x,u
)]-V(t+\delta,x)|\leq C(1+|x|)\sqrt{\delta
\]
by Theorem \ref{th-ex}, where $C$ depends on $T$, $G$ and $L$. Thus we obtain
$|V(t,x)-V(t+\delta,x)|\leq C(1+|x|)\sqrt{\delta}$ by inequality (\ref{e34}).
\end{proof}
\section{The viscosity solution of HJB equation}
The following theorem establishes the relationship between the value function
$V(\cdot,\cdot)$ and the fully nonlinear second-order partial differential
equation. For the definition of the viscosity solution, the readers can refer
to \cite{CIP}.
\begin{theorem}
\label{newth222}Let Assumptions (H1) and (H2) hold. Then the value function
$V(\cdot,\cdot)$ defined in (\ref{e4}) is the unique viscosity solution of the
following second-order partial differential equation
\begin{equation}
\left \{
\begin{array}
[c]{l
\partial_{t}V(t,x)+\underset{u\in U}{\inf}H(t,x,V,\partial_{x}V,\partial
_{xx}^{2}V,u)=0,\\
V(T,x)=\Phi(x),\text{ }x\in \mathbb{R}^{n},
\end{array}
\right. \label{e35
\end{equation}
wher
\[
H(t,x,v,p,A,u)=\tilde{G}(F(t,x,v,p,A,u))+\langle p,b(t,x,u)\rangle
+f(t,x,v,u),
\
\[
F_{ij}(t,x,v,p,A,u)=(\sigma^{T}(t,x,u)A\sigma(t,x,u))_{ij}+2\langle
p,h_{ij}(t,x,u)\rangle+2g_{ij}(t,x,v,u),
\]
$(t,x,v,p,A,u)\in \lbrack0,T]\times \mathbb{R}^{n}\times \mathbb{R
\times \mathbb{R}^{n}\times \mathbb{S}_{n}\times U$, $\tilde{G}$ is defined in
(\ref{newe1}).
\end{theorem}
\begin{remark}
\label{new-re111}According to Theorem C.3.5 in \cite{P2019}, for the case
that
\[
\Phi \in C_{0}(\mathbb{R}^{n})=\{ \phi \in C(\mathbb{R}^{n}):\lim
_{|x|\rightarrow \infty}\phi(x)=0\},
\]
the viscosity solution of PDE (\ref{e35}) is unique; for the case that $\Phi$
$\in C(\mathbb{R}^{n})$ satisfying $|\Phi(x)|\leq C(1+|x|^{p})$ for some
positive constants $C$ and $p$, the meaning of uniqueness is that, for each
$\Phi_{k}\in C_{0}(\mathbb{R}^{n})$ such that $\Phi_{k}$ converges uniformly
to $\Phi$ on each compact set and $|\Phi_{k}|\leq C(1+|x|^{p})$, we have
$V^{\Phi_{k}}(t,x)\rightarrow V^{\Phi}(t,x)$ for $(t,x)\in \lbrack
0,T]\times \mathbb{R}^{n}$.
\end{remark}
In order to prove this theorem, we need the following lemmas. Let $\varphi \in
C_{b}^{2,3}([0,T]\times \mathbb{R}^{n})$ be given. Here $C_{b}^{2,3
([0,T]\times \mathbb{R}^{n})$ denotes the set of real-valued functions that are
continuously differentiable up to the second order (resp. third order) in
$t$-variable (resp. $x$-variable) and whose derivatives are bounded. For each
given $(t,x)\in \lbrack0,T)\times \mathbb{R}^{n}$, $\delta \in \lbrack0,T-t]$ and
$u\in \mathcal{U}[t,t+\delta]$, we consider the following BSDE
\begin{equation}
Y_{s}^{u}=\mathbb{\tilde{E}}_{s}\left[ \varphi(t+\delta,X_{t+\delta
^{t,x,u})+\int_{s}^{t+\delta}f(r,X_{r}^{t,x,u},Y_{r}^{u},u_{r})dr+\int
_{s}^{t+\delta}g_{ij}(r,X_{r}^{t,x,u},Y_{r}^{u},u_{r})d\langle B^{i
,B^{j}\rangle_{r}\right] , \label{e36
\end{equation
\begin{equation}
Y_{s}^{1,u}=\mathbb{\tilde{E}}_{s}\left[ \int_{s}^{t+\delta}F_{1
(r,X_{r}^{t,x,u},Y_{r}^{1,u},u_{r})dr+\int_{s}^{t+\delta}F_{2}^{ij
(r,X_{r}^{t,x,u},Y_{r}^{1,u},u_{r})d\langle B^{i},B^{j}\rangle_{r}\right]
\label{e37
\end{equation}
an
\begin{equation}
Y_{s}^{2,u}=\mathbb{\tilde{E}}_{s}\left[ \int_{s}^{t+\delta}F_{1
(r,x,0,u_{r})dr+\int_{s}^{t+\delta}F_{2}^{ij}(r,x,0,u_{r})d\langle B^{i
,B^{j}\rangle_{r}\right] , \label{e38
\end{equation}
where $s\in \lbrack t,t+\delta]$, $(X_{s}^{t,x,u})_{s\in \lbrack t,t+\delta]}$
is the solution of the SDE (\ref{newe2})
\[
F_{1}(s,x,y,u)=\partial_{t}\varphi(s,x)+\langle b(s,x,u),\partial_{x
\varphi(s,x)\rangle+f(s,x,y+\varphi(s,x),u),
\
\[
F_{2}^{ij}(s,x,y,u)=\frac{1}{2}F_{ij}(s,x,y+\varphi(s,x),\partial_{x
\varphi(s,x),\partial_{xx}^{2}\varphi(s,x),u).
\]
\begin{lemma}
\label{new-1}For each $u\in \mathcal{U}[t,t+\delta]$, we hav
\[
Y_{s}^{1,u}=Y_{s}^{u}-\varphi(s,X_{s}^{t,x,u})\text{ for }s\in \lbrack
t,t+\delta].
\]
\end{lemma}
\begin{proof}
Applying It\^{o}'s formula to $\varphi(r,X_{r}^{t,x,u})$ on $[s,t+\delta]$, we
obtain that $(Y_{s}^{u}-\varphi(s,X_{s}^{t,x,u}))_{s\in \lbrack t,t+\delta]}$
satisfies the backward SDE (\ref{e37}), which implies the desired result by
the uniqueness of the solution.
\end{proof}
\begin{lemma}
\label{new-2}For each $u\in \mathcal{U}^{t}[t,t+\delta]$, we hav
\[
|Y_{t}^{1,u}-Y_{t}^{2,u}|\leq C(1+|x|^{3})\delta^{3/2},
\]
where the constant $C$ is dependent on $T$, $G$, $L$ and independent of $u$.
\end{lemma}
\begin{proof}
Noting that $\varphi \in C_{b}^{2,3}([0,T]\times \mathbb{R}^{n})$ and $U$ is
compact, one can verify tha
\[
|F_{1}(r,x,0,u_{r})|\leq C(1+|x|)\text{ and }|F_{2}^{ij}(r,x,0,u_{r})|\leq
C(1+|x|^{2}),
\]
where $C$ is dependent on $L$ and independent of $u$. Thu
\begin{equation}
|Y_{s}^{2,u}|\leq C(1+|x|^{2})\delta \text{ for }s\in \lbrack t,t+\delta],
\label{e39
\end{equation}
where $C$ is dependent on $G$, $L$ and independent of $u$. Set $\hat{Y
_{s}=Y_{s}^{1,u}-Y_{s}^{2,u}$ for $s\in \lbrack t,t+\delta]$, by (\ref{e37})
and (\ref{e38}), we ge
\[
|\hat{Y}_{s}|\leq C\mathbb{\hat{E}}_{s}\left[ \int_{s}^{t+\delta}(\hat{F
_{r}+|\hat{Y}_{r}|)dr\right] ,
\]
where $C>0$ is dependent on $G$, $L$ and independent of $u$,
\[
\hat{F}_{r}=|F_{1}(r,X_{r}^{t,x,u},Y_{r}^{2,u},u_{r})-F_{1}(r,x,0,u_{r
)|+|F_{2}^{ij}(r,X_{r}^{t,x,u},Y_{r}^{2,u},u_{r})-F_{2}^{ij}(r,x,0,u_{r})|.
\]
Note that $Y_{t}^{1,u}\in \mathbb{R}$ and $Y_{t}^{2,u}\in \mathbb{R}$ for each
$u\in \mathcal{U}^{t}[t,t+\delta]$, then, by the Gronwall inequality under
$\mathbb{\hat{E}}$, we obtai
\begin{equation}
|Y_{t}^{1,u}-Y_{t}^{2,u}|\leq C\mathbb{\hat{E}}\left[ \int_{t}^{t+\delta
\hat{F}_{r}dr\right] , \label{e40
\end{equation}
where $C>0$ is dependent on $T$, $G$, $L$ and independent of $u$. One can
check tha
\begin{equation}
\hat{F}_{r}\leq C\left[ (1+|x|^{2})|X_{r}^{t,x,u}-x|+|X_{r}^{t,x,u
-x|^{2}+|Y_{r}^{2,u}|\right] , \label{e41
\end{equation}
where $C$ is dependent on $L$ and independent of $u$. It follows from
(\ref{e39}), (\ref{e40}), (\ref{e41}) and Theorem \ref{th-ex} tha
\begin{align*}
|Y_{t}^{1,u}-Y_{t}^{2,u}| & \leq C\left \{ (1+|x|^{2})\delta \left(
\mathbb{\hat{E}}\left[ \sup_{r\in \lbrack t,t+\delta]}|X_{r}^{t,x,u
-x|^{2}\right] \right) ^{1/2}+\delta \mathbb{\hat{E}}\left[ \sup
_{r\in \lbrack t,t+\delta]}|X_{r}^{t,x,u}-x|^{2}\right] +(1+|x|^{2})\delta
^{2}\right \} \\
& \leq C(1+|x|^{3})\delta^{3/2},
\end{align*}
where $C$ is dependent on $T$, $G$, $L$ and independent of $u$.
\end{proof}
\begin{lemma}
\label{newlem111}Let $\eta=(\eta^{ij})_{i,j=1}^{d}\in M_{G}^{1}(0,T;\mathbb{S
_{d})$. Then, for each $s\leq T$, we hav
\[
\mathbb{\tilde{E}}_{s}\left[ \int_{s}^{T}\eta_{r}^{ij}d\langle B^{i
,B^{j}\rangle_{r}-\int_{s}^{T}\tilde{G}(2\eta_{r})dr\right] =0.
\]
\end{lemma}
\begin{proof}
For each $\eta$, $\tilde{\eta}\in M_{G}^{1}(0,T;\mathbb{S}_{d})$, one can
verify that
\begin{align*}
& \mathbb{\hat{E}}\left[ \left \vert \mathbb{\tilde{E}}_{s}\left[ \int
_{s}^{T}\eta_{r}^{ij}d\langle B^{i},B^{j}\rangle_{r}-\int_{s}^{T}\tilde
{G}(2\eta_{r})dr\right] -\mathbb{\tilde{E}}_{s}\left[ \int_{s}^{T
\tilde{\eta}_{r}^{ij}d\langle B^{i},B^{j}\rangle_{r}-\int_{s}^{T}\tilde
{G}(2\tilde{\eta}_{r})dr\right] \right \vert \right] \\
& \leq C\mathbb{\hat{E}}\left[ \int_{s}^{T}|\eta_{r}-\tilde{\eta
_{r}|dr\right] ,
\end{align*}
where $C$ only depends on $G$. Thus we only need to prove the case $\eta \in
M_{G}^{0}(0,T;\mathbb{S}_{d})$, i.e.
\[
\eta_{r}=\sum_{k=0}^{N-1}\eta_{t_{k}}I_{[t_{k},t_{k+1})}(r),
\]
where $s=t_{0}<\cdots<t_{N}=T$, $\eta_{t_{k}}\in Lip(\Omega_{t_{k}
;\mathbb{S}_{d})$. Since $\mathbb{\tilde{E}}_{s}[\cdot]=\mathbb{\tilde{E}
_{s}[\mathbb{\tilde{E}}_{t_{k}}[\cdot]]$, we only need to prov
\begin{equation}
\mathbb{\tilde{E}}_{t_{k}}\left[ \eta_{t_{k}}^{ij}(\langle B^{i},B^{j
\rangle_{t_{k+1}}-\langle B^{i},B^{j}\rangle_{t_{k}})-\tilde{G}(2\eta_{t_{k
})(t_{k+1}-t_{k})\right] =0. \label{e42
\end{equation}
Applying It\^{o}'s formular to $\langle \eta_{t_{k}}(B_{r}-B_{t_{k}
),B_{r}-B_{t_{k}}\rangle$ on $[t_{k},t_{k+1}]$, we ge
\[
\mathbb{\tilde{E}}_{t_{k}}\left[ \eta_{t_{k}}^{ij}(\langle B^{i},B^{j
\rangle_{t_{k+1}}-\langle B^{i},B^{j}\rangle_{t_{k}})\right] =\mathbb{\tilde
{E}}_{t_{k}}\left[ \langle \eta_{t_{k}}(B_{t_{k+1}}-B_{t_{k}}),B_{t_{k+1
}-B_{t_{k}}\rangle \right] .
\]
For each given $A\in \mathbb{S}_{d}$, define
\[
u(t,x)=\mathbb{\tilde{E}}\left[ \langle A(x+B_{t}),x+B_{t}\rangle \right]
\text{ for }(t,x)\in \lbrack0,\infty)\times \mathbb{R}^{d}.
\]
By Theorem C.3.5 in \cite{P2019}, we know that $u$ is a viscosity solution of
the following PD
\begin{equation}
\partial_{t}u-\tilde{G}(\partial_{xx}^{2}u)=0,\text{ }u(0,x)=\langle
Ax,x\rangle. \label{e43
\end{equation}
On the other hand, by the proof of Theorem 3.8.2 in \cite{P2019}, we hav
\begin{equation}
u(t,x)=\langle Ax,x\rangle+\mathbb{\tilde{E}}\left[ \langle AB_{t
,B_{t}\rangle \right] =\langle Ax,x\rangle+\mathbb{\tilde{E}}\left[ \langle
AB_{1},B_{1}\rangle \right] t. \label{e44
\end{equation}
By (\ref{e43}) and (\ref{e44}), we obtain $\mathbb{\tilde{E}}\left[ \langle
AB_{1},B_{1}\rangle \right] =\tilde{G}(2A)$, which implies $\mathbb{\tilde{E
}\left[ \langle AB_{t},B_{t}\rangle \right] =\tilde{G}(2A)t$. Thus we hav
\[
\mathbb{\tilde{E}}_{t_{k}}\left[ \eta_{t_{k}}^{ij}(\langle B^{i},B^{j
\rangle_{t_{k+1}}-\langle B^{i},B^{j}\rangle_{t_{k}})\right] =\tilde{G
(2\eta_{t_{k}})(t_{k+1}-t_{k}),
\]
which implies (\ref{e42}).
\end{proof}
\begin{remark}
It is important to note that we can not derive $\mathbb{\tilde{E}}\left[
\langle AB_{1},B_{1}\rangle \right] =\tilde{G}(2A)$ by $u(t,x)=\langle
Ax,x\rangle+\tilde{G}(2A)t$ satisfying (\ref{e43}). Because, in this case of
$u(0,x)=\langle Ax,x\rangle \not \in C_{0}(\mathbb{R}^{n})$, the meaning of
uniqueness of viscosity solution is stated as in Remark \ref{new-re111}.
\end{remark}
\begin{lemma}
\label{new-3}We hav
\[
\inf_{u\in \mathcal{U}^{t}[t,t+\delta]}Y_{t}^{2,u}=\int_{t}^{t+\delta
F_{0}(r,x)dr,
\]
wher
\[
F_{0}(r,x)=\inf_{v\in U}\{F_{1}(r,x,0,v)+\tilde{G}(2(F_{2}^{ij
(r,x,0,v))_{ij=1}^{d})\}.
\]
\end{lemma}
\begin{proof}
For each $u\in \mathcal{U}^{t}[t,t+\delta]$, by Lemma \ref{newlem111}, we get
\begin{align*}
Y_{t}^{2,u} & =\mathbb{\tilde{E}}_{t}\left[ \int_{t}^{t+\delta
F_{1}(r,x,0,u_{r})dr+\int_{t}^{t+\delta}F_{2}^{ij}(r,x,0,u_{r})d\langle
B^{i},B^{j}\rangle_{r}\right] \\
& \geq \mathbb{\tilde{E}}_{t}\left[ \int_{t}^{t+\delta}F_{0}(r,x)dr+\int
_{t}^{t+\delta}F_{2}^{ij}(r,x,0,u_{r})d\langle B^{i},B^{j}\rangle_{r}-\int
_{t}^{t+\delta}\tilde{G}(2(F_{2}^{ij}(r,x,0,u_{r}))_{ij=1}^{d})dr\right] \\
& =\int_{t}^{t+\delta}F_{0}(r,x)dr.
\end{align*}
Hence, $\inf_{u\in \mathcal{U}^{t}[t,t+\delta]}Y_{t}^{2,u}\geq \int
_{t}^{t+\delta}F_{0}(r,x)dr$. On the other hand, we can choose a deterministic
control $u^{\ast}\in \mathcal{U}^{t}[t,t+\delta]$ such tha
\[
\int_{t}^{t+\delta}[F_{1}(r,x,0,u_{r}^{\ast})+\tilde{G}(2(F_{2}^{ij
(r,x,0,u_{r}^{\ast}))_{ij=1}^{d})]dr=\int_{t}^{t+\delta}F_{0}(r,x)dr.
\]
Then we obtain $Y_{t}^{2,u^{\ast}}=\int_{t}^{t+\delta}F_{0}(r,x)dr$ by Lemma
\ref{newlem111}, which implies $\inf_{u\in \mathcal{U}^{t}[t,t+\delta]
Y_{t}^{2,u}\leq \int_{t}^{t+\delta}F_{0}(r,x)dr$. Thus we obtain the desired result.
\end{proof}
\noindent \textbf{Proof of Theorem \ref{newth222}.} By Proposition \ref{pro-v}
and Lemma \ref{newle333}, we know that $V(\cdot,\cdot)$ is continuous on
$[0,T]\times \mathbb{R}^{n}$. Now, we first prove that $V(\cdot,\cdot)$ is the
viscosity subsolution of (\ref{e35}).
For each given $(t,x)\in \lbrack0,T)\times \mathbb{R}^{n}$, suppose $\varphi \in
C_{b}^{2,3}([0,T]\times \mathbb{R}^{n})$ such that $\varphi(t,x)=V(t,x)$ and
$\varphi \geq V$ on $[0,T]\times \mathbb{R}^{n}$. For each $\delta \in
\lbrack0,T-t]$, by Theorem \ref{DPP}, we ge
\[
V(t,x)=\inf_{u\in \mathcal{U}^{t}[t,t+\delta]}\mathbb{G}_{t,t+\delta
^{t,x,u}[V(t+\delta,X_{t+\delta}^{t,x,u})].
\]
Since $\varphi(t+\delta,X_{t+\delta}^{t,x,u})\geq V(t+\delta,X_{t+\delta
}^{t,x,u})$, by Lemma \ref{le-com}, we obtain $\mathbb{G}_{t,t+\delta
^{t,x,u}[V(t+\delta,X_{t+\delta}^{t,x,u})]\leq Y_{t}^{u}$. It follows from
$\varphi(t,x)=V(t,x)$, Lemmas \ref{new-1} and \ref{new-2} tha
\begin{align*}
\inf_{u\in \mathcal{U}^{t}[t,t+\delta]}Y_{t}^{2,u} & \geq \inf_{u\in
\mathcal{U}^{t}[t,t+\delta]}Y_{t}^{1,u}-C(1+|x|^{3})\delta^{3/2}\\
& =\inf_{u\in \mathcal{U}^{t}[t,t+\delta]}(Y_{t}^{u}-\varphi(t,x))-C(1+|x|^{3
)\delta^{3/2}\\
& \geq-C(1+|x|^{3})\delta^{3/2},
\end{align*}
where $C$ is dependent on $T$, $G$, $L$. By Lemma \ref{new-3}, we ge
\[
\delta^{-1}\int_{t}^{t+\delta}F_{0}(r,x)dr\geq-C(1+|x|^{3})\delta^{1/2}.
\]
One can verify that $F_{0}(\cdot,x)$ is continuous in $r$. Hence we obtain
$F_{0}(t,x)\geq0$ by letting $\delta \downarrow0$, which implies that
$V(\cdot,\cdot)$ is the viscosity subsolution of (\ref{e35}). By the same
method, we can prove that $V(\cdot,\cdot)$ is the viscosity supersolution of
(\ref{e35}). Thus $V(\cdot,\cdot)$ is the viscosity solution of (\ref{e35}).
For the uniqueness of viscosity solution, we only need to prove the case
$\Phi \in C_{0}(\mathbb{R}^{n})$ according to Remark \ref{new-re111}.
Hoverever, by the proof of Theorem C.2.9 with $l=0$ in \cite{P2019}, we see
that in order to get the uniqueness we just need to know that $\inf_{u\in
U}H(t,x,v,p,A,u)$ satisfies assumption (G$^{\prime}$). For each $t\in
\lbrack0,T)$, $x$, $y\in \mathbb{R}^{n}$, $v\in \mathbb{R}$, $\alpha>0$, $A$,
$B\in \mathbb{S}_{n}$ such tha
\[
\left(
\begin{array}
[c]{cc
A & 0\\
0 & B
\end{array}
\right) \leq3\alpha \left(
\begin{array}
[c]{cc
I_{n} & -I_{n}\\
-I_{n} & I_{n
\end{array}
\right) ,
\]
we hav
\begin{align*}
& \underset{u\in U}{\inf}H(t,x,v,\alpha(x-y),A,u)-\underset{u\in U}{\inf
}H(t,y,v,\alpha(x-y),-B,u)\\
& \leq \underset{u\in U}{\sup}\left[ H(t,x,v,\alpha(x-y),A,u)-H(t,y,v,\alpha
(x-y),-B,u)\right] \\
& \leq \underset{u\in U}{\sup}G(F(t,x,v,\alpha(x-y),A,u)-F(t,y,v,\alpha
(x-y),-B,u))+L(|x-y|+\alpha|x-y|^{2})\\
& \leq \underset{u\in U}{\sup}G(\sigma^{T}(t,x,u)A\sigma(t,x,u)+\sigma
^{T}(t,y,u)B\sigma(t,y,u))+C(|x-y|+\alpha|x-y|^{2})\\
& \leq \underset{u\in U}{\sup}G(3\alpha(\sigma(t,x,u)-\sigma(t,y,u))^{T
(\sigma(t,x,u)-\sigma(t,y,u)))+C(|x-y|+\alpha|x-y|^{2})\\
& \leq C(|x-y|+\alpha|x-y|^{2}),
\end{align*}
where $C$ depends on $L$ and $G$. Thus $\inf_{u\in U}H(t,x,v,p,A,u)$ satisfies
assumption (G$^{\prime}$), which implies that $V(\cdot,\cdot)$ is the unique
viscosity solution of (\ref{e35}). $\Box$
\bigskip
|
\section{Introduction}
Drug discovery is the process of using multi-disciplinary knowledge such as biology, chemistry and pharmacology to discover proficient medications amongst candidates according to safety and efficacy requirements. Modern drug discovery often features a cost-ineffective virtual screening process to select candidates from general chemical databases such as \textit{ChEMBL}~\cite{gaulton2012chembl} and \textit{OpenChem}~\cite{kim2016pubchem} with large volume of molecular data to build a significant smaller in-house database for further synthesis.
Traditional machine learning algorithms such as random forest~\cite{jayaraj2016gpurfscreen}, support vector machine~\cite{liew2009svm}, k nearest neighbors~\cite{arian2020protein}, and gradient boosting~\cite{wu2018moleculenet} have been investigated in drug discovery applications. Such algorithms use molecular representations as input to predict molecular properties. However, because of limited sophistication, deep and complex structural information within a molecule is generally overlooked by those models. Thus, they typically do not exhibit strong capability in learning the features and only achieve sub-par performance. On the other hand, inspired by the recent success from other applications such as computer vision, neural network models have been increasingly applied in drug discovery. GNN learns representations by aggregating nodes and neighbouring information for molecular property predictions under different drug discovery objectives. However, molecular graphs often requires pre-processing or featurization. Extended-connectivity fingerprints (ECFP) is one of the most common featurization method that converts molecular graphs into fixed length representations, or fingerprints~\cite{rogers2010extended}. Such featurization algorithms usually requires comprehensive efforts using chemical tool-chains such as RDKit~\cite{landrum2013rdkit}.
This paper takes a radical departure from common machine learning methods including neural networks by developing an ultra-low-cost brain-inspired hyperdimensional computing (HDC) model that requires less pre-processing efforts and is easier to implement. Inspired by the attributes of brain circuits including high-dimensionality and fully distributed holographic representation, this emerging computing paradigm postulates the generation, manipulation, and comparison of symbols represented by high-dimensional vectors. Compared with DNNs, the advantages of HDC include smaller model size, less computation cost, and one/few-shot learning, making it a promising alternative computing paradigm~\cite{karunaratne2020memory}. Recently, HDC has demonstrated success on various application domains such as robotics~\cite{mitrokhin2019learning}, natural language processing~\cite{thapa2021spamhd}, biomedical signal analysis~\cite{rahimi2016hyperdimensional}, and biological sequence matching~\cite{imani2018hdna}.
In this paper, we develop \textbf{MoleHD}, an HDC-based method to predict molecular properties in drug discovery. \textbf{MoleHD} ~first tokenizes SMILES strings into numerical list of tokens, and then develop HDC encoding mechanisms to project realistic features into their high-dimensional space representations: hypervectors. Next, \textbf{MoleHD} ~leverages hypervector properties to train an HDC model that can be used to perform molecule classification tasks.
The qualitative advantages of \textbf{MoleHD} ~compared to existing neural network-based classifiers for drug discovery are: (1) back-propagation free: \textbf{MoleHD} ~does not need backpropagation to train a set of parameters; instead, it uses one/few-shot learning to establish abstract patterns that can represent specific symbols. (2) efficient computing: unlike neural networks, \textbf{MoleHD} ~does not need complicated arithmetic operations such as convolutions which presents a major computing/energy burden to computing platforms; instead, it only uses simple arithmetic operations such as addition between two vectors. Thus, \textbf{MoleHD} ~only needs to run on commodity CPU and can finish both training and testing on the reported datasets within minutes, while GNN requires around 5 days for training using Nvidia GPU~\cite{wang2021molclr}. (3) smaller model size: \textbf{MoleHD} ~only needs to store a set of vectors for comparison during inference, while SOTA neural networks often need millions of parameters and requires memory in 100MB scale to store the parameters (e.g., weights and activation values)~\cite{ma2020multi}.
The main contributions of this paper are summarized as below:
\begin{enumerate}
\item We propose \textbf{MoleHD}, an ultra-low-cost novel learning model based on hyperdimensional computing. This promising results of \textbf{MoleHD} ~provide a viable option and alternative to existing learning methods in drug discovery domain.
\item We develop a complete molecular-specific pipeline for HDC-based drug discovery. First, \textbf{MoleHD} ~tokenizes SMILE strings into tokens representing the substructures and then project them into hypervectors during encoding. Then, \textbf{MoleHD} ~uses the encoded hypervectors to train and evaluate the classification model.
\item We perform an extensive evaluation of \textbf{MoleHD} ~on 29 classification tasks from 3 widely-used molecule datasets under three splits methods. By a comprehensive comparison with 8 baseline models including SOTA neural networks, \textbf{MoleHD} ~is able to achieve highest ROC-AUC score on random and scaffold splits on average across 3 datasets and achieve second-highest on stratified split. More importantly, \textbf{MoleHD} ~achieves such performance with significantly reduced computing cost than GNNs, e.g., ~10 mins training using CPU VS. 5 days on GPU, and 80 KB VS 100MB model size.
\item We conduct a design space exploration of \textbf{MoleHD} ~by developing two tokenization schemes (\textbf{MoleHD}-PE and \textbf{MoleHD}-char), two gram sizes (uni-gram and bi-gram), and evaluate their corresponding performance.
\end{enumerate}
\section{Related Works}
Hyperdimensional computing (HDC), also known as vector-symbolic architectures (VSA), was introduced as an alternative computational model mimicking the ``human brain'' at the functionality level~\cite{kanerva2009hyperdimensional}.
HDC has been used in modern robotics to perform active perception by integrating the sensory perceptions experienced by an agent with its motoric capabilities, which is vital to autonomous learning agents~\cite{mitrokhin2019learning}. HDC has also been used in biomedical signal processing and exhibits 97.8\% accuracy on hand gesture recognition based on EMG, which surpasses support vector machine by 8.1\%~\cite{rahimi2016hyperdimensional}. Recent works also show that HDC outperforms other machine learning methods in DNA sequencing~\cite{imani2018hdna, kim2020geniehd}.
Machine learning algorithms are used in drug discovery mostly in predicting molecular properties to determine if they satisfy the drug discovery objective.
Recently, emerging machine learning algorithms such as GNNs are increasingly applied to drug discovery for achieving higher performance. GNNs leverages fingerprints derived from the molecular graph to learn the representations. Direct message passing neural network (D-MPNN) is an evolution of message passing neural networks that centers on bonds between atoms which is able to maintain two representations~\cite{yang2019learned, swanson2019message}. Contrastive learning is also applied into GNNs to fuse drug discovery domain knowledge and molecular properties to augment learning of representations~\cite{fang2021knowledge, wang2021molclr}. In addition to GNNs, natural language processing (NLP) models such as recurrent neural networks (RNNs) are also introduced in drug discovery. Compared with GNNs, RNNs typically do not rely on complex fingerprint conversion process using toolchains such as \textbf{RDKit}~\cite{quan2018system, lin2020novel}. However, RNNs still require word embeddings tools such as \textbf{Smi2Vec}, to fully extract features from the molecule SMILES representation.
\section{Preliminaries on HDC}
\subsection{Hypervectors} Hypervectors (HV) are high-dimensional (usually higher than 10,000), holographic (not micro-coded) vectors with (pseudo-)random and i.i.d. elements~\cite{kanerva2009hyperdimensional}. An HV with $d$ dimensions can be denoted as $\overrightarrow{H} = \langle h_1, h_2, \dots, h_d\rangle$, where $h_i$ refers to the elements inside the HV. HVs are fundamental blocks in HDC that are able to accommodate and represent information in different scales and layers. When the dimensionality is sufficiently high (e.g., $D=10,000$), any two random HVs are nearly orthogonal~\cite{kanerva2009hyperdimensional}. HDC utilizes different operations HVs support as means of producing aggregations of information or creating representations of new information.
\subsection{Operations}
In HDC, addition ($+$), multiplication ($*$) and permutation ($\rho$) are the three basic operations HVs can support. Additions and multiplications take two input HVs as operands and perform \textbf{element-wise} add or multiply operations on the two HVs. Permutation takes one HV as the input operand and perform \textbf{cyclic rotation} by a specific amount. All the operations do not modify the dimensionality of the input HVs, i.e. the input and the output HVs are in the same dimension.
These three operations also have their corresponding physical meanings. Addition is used to aggregate same-type information, while multiplication is used to combine different types of information together to generate new information. Permutation is used to reflect spatial or temporal changes in the information, such as time series or spatial coordinates~\cite{kanerva2009hyperdimensional}.
\subsection{Similarity Measurement}
In HDC, the similarity metric $\delta$ between the information that two HVs represent is measured by similarity check. Different algorithms can be used to calculate the similarity, such as the Euclidean ($L2$) distance, the Hamming distance (for binary HVs), and cosine similarity (which we use in this paper). A higher similarity $\delta$ between two HVs shows that these two HVs have more information in common, or vice versa. Because of the high dimensionality of HVs, addition generally results in a new HV that is approximately 50\% similar to the two original HVs, while multiplication and permutation result in HVs that are orthogonal to the original HVs, i.e., not similar.
\begin{figure*}
\centering
\includegraphics[page=1, width=1.93\columnwidth]{./figures/moleculeHD.pdf}
\caption{Overview of \textbf{MoleHD}. \textbf{MoleHD} ~has 5 major steps: \textbf{Tokenization}, \textbf{Encoding}, \textbf{Training}, \textbf{Retraining} and \textbf{Inference}.}
\label{fig:moleculeHD}
\end{figure*}
\section{\textbf{MoleHD} ~Framework}
In this section, we will introduce the proposed framework \textbf{MoleHD} ~and how it utilizes HDC to perform learning tasks in drug discovery. An overview of \textbf{MoleHD} ~is presented in Fig.~\ref{fig:moleculeHD}.
\subsection{Tokenization}
In \textbf{MoleHD}, tokenization is the process of converting molecule features into their corresponding set of numerical tokens. It basically consists of three procedures: converting the SMILES string into a list of tokens and then assign number for the tokens to obtain a list of numerical tokens which are ready for HDC processing.
We develop two tokenization schemes for \textbf{MoleHD}: \textbf{MoleHD}-char and \textbf{MoleHD}-PE. \textbf{MoleHD}-char is the basic tokenization strategy that treats the input SMILES string as a textual string. \textbf{MoleHD}-char split the textual string into characters to obtain a list of tokens. Each unique character inside the string is then assigned with a unique random number to form the numerical tokens. \textbf{MoleHD}-PE uses the open-source SMILES Pair Encoding (SMILES-PE) model to extract the sub-structures in the input SMILES strings then assign a unique number based on their appearance frequency ranking to tokenize them. SMILES-PE is a data-driven algorithm to find substructures from a SMILES string~\cite{li2020smiles}. \textbf{MoleHD}-PE ~uses SMILES-PE as-is as an add-on and does not require additional time for the pre-train. Due to model size limitation or other user-specific constraints, only $m$ tokens will be stored. For missing tokens in \textbf{MoleHD}, a special token `0' is assigned.
\begin{figure*}
\centering
\subfigure[HDC Encoding]{
\includegraphics[page=1, width=1.94\columnwidth]{./figures/hdc.pdf}
}
\subfigure[HDC Training and Retraining]{
\includegraphics[page=2, width=1.99\columnwidth]{./figures/hdc.pdf}
}
\subfigure[HDC Inference]{
\includegraphics[page=3, width=1.94\columnwidth]{./figures/hdc.pdf}
}
\caption{HDC processing: \textbf{Encoding}, \textbf{Training}, \textbf{Retraining} and \textbf{Inference}.}
\label{fig:hdc}
\end{figure*}
\subsection{HDC Encoding}
Encoding is the process to project real-world features into their high-dimensional space representations: the HVs. In \textbf{MoleHD}, encoding process projects tokenized sample into its representing sample HV (sHV, or $\overrightarrow{S}$) via a combination of pre-defined HD operations as shown in Fig.~\ref{fig:hdc}(a).
\mypara{Item Memory}
Item memory is generated from the token dictionary in tokenization. The item memory contains base HVs (bHV, or $\overrightarrow{B}$) in the same number ($m+1$, considering the missing entry assigned as `0') as the entries in the token dictionary, i.e., each HV serves as the high-dimensional representation of a token. The item memory is fully random generated using a seed to ensure the i.i.d. properties. We note item memory as $\mathbb{B} = \{\overrightarrow{B_0}, \overrightarrow{B_1}, ... , \overrightarrow{B_m}\}$ where $\overrightarrow{B_i}$ is the base HV with index $i$.
\mypara{HD operations in Encoding}
In \textbf{MoleHD}, encoding schemes can be flexible and data-specific. Algorithm~\ref{alg:unigram} shows the process of uni-gram encoding as an example. Tokenized sample first uses its tokens iteratively in the item memory to index and fetch the corresponding base HVs. The base HVs permutate by their order in the tokenized sample and added up to establish the sample HV (Line 2 - 4). \textbf{MoleHD} ~also bipolarizes the elements inside the sample HV according to their relation with zero (Line 5 - 11). \textbf{MoleHD} ~also features bi-gram and tri-gram encoding which resembles the uni-gram encoding but instead permutes every 2 or 3 tokens aggregated together by HV multiplication.
\begin{algorithm}
\small
\caption{\textbf{MoleHD} ~Encoding (uni-gram)}
\algrenewcommand\algorithmicrequire{\textbf{Input}}
\algrenewcommand\algorithmicensure{\textbf{Output}}
\begin{algorithmic}[1]
\Require tokenized sample $T = \{t_0, t_1, t_k\}$, item memory $\mathbb{B}$.
\Ensure sample HV $\overrightarrow{S}$
\State $setZero(\overrightarrow{S})$
\For{$t_i$ in $T$} \textit{\textbackslash * Perform uni-gram encoding. */}
\State $\overrightarrow{S} = \overrightarrow{S} + \rho^i(\overrightarrow{B_{t_i}})$
\EndFor
\For{$s_i$ in $\overrightarrow{S}$} \textit{\textbackslash * Bipolarize the sample HV. */}
\If{$s_i > 0$} $s_i = 1$
\ElsIf{$s_i < 0$} $s_i = -1$
\EndIf
\EndFor
\end{algorithmic}
\label{alg:unigram}
\end{algorithm}
\subsection{HDC Training}
Training is the process of establishing the associative memory $\mathbb{C} = {\overrightarrow{C_1}, \overrightarrow{C_2}, ... , \overrightarrow{C_p}}$ using the training set. Associative memory (AM) contains $p$ class HVs (cHV, or $\overrightarrow{C}$), each representing a class in a learning task. Using a binary classification task as example shown in Fig.~\ref{fig:hdc}(b), AM contains the class HV representing positive ($\overrightarrow{C_P}$) and negative ($\overrightarrow{C_N}$). For each training sample, \textbf{MoleHD} ~adds its HV to the corresponding class HV according to the label, as shown in Eq.~\ref{eqn:retrain}. This process is to aggregate the information from sample HVs together into the AM. However, one-epoch training is usually not enough to train a reliable AM for learning tasks, it is necessary to perform additional epochs for fine-tuning or retraining.
\begin{equation}
\begin{aligned}
\overrightarrow{C_P} = \sum \overrightarrow{S_p}, & & \overrightarrow{C_N} = \sum \overrightarrow{S_n}
\end{aligned}
\label{eqn:train}
\end{equation}
\subsection{HDC Retraining}
Retraining is the process of fine-tuning the associative memory to enhance its accuracy using the training set, as shown in Fig.~\ref{fig:hdc}(b). For each training sample, \textbf{MoleHD} ~tries to use the AM to predict its label. If the prediction is correct, \textbf{MoleHD} ~proceeds to the next training sample. However, if the prediction is wrong, it indicates that the correct information of the sample HV has not been aggregated into the AM, or the information in the AM is not properly represented. Therefore, \textbf{MoleHD} ~performs an update to the AM to remove the erroneous and add the correct information, by subtracting the sample HV from the wrongly predicted class HV ($\overrightarrow{C_W}$) and adding it to the correct class HV ($\overrightarrow{C_R}$), as shown in Eq.~\ref{eqn:retrain}.
\begin{equation}
\begin{aligned}
\overrightarrow{C_W} = \overrightarrow{C_W} - \overrightarrow{S}, & & \overrightarrow{C_R} = \overrightarrow{C_R} + \overrightarrow{S}
\end{aligned}
\label{eqn:retrain}
\end{equation}
\subsection{HDC Inference}
Inference is the process of using unseen data from the inference set to evaluate the trained model's performance. As illustrated in Fig.~\ref{fig:hdc}(c), \textbf{MoleHD} ~calculates the cosine similarity ($\delta$) between the sample HV from the inference set with unknown label (referred to as query HV (qHV, $\overrightarrow{Q_?}$) and each cHV in the AM to obtain the similarity values. The cHV with the most similarity indicates having the most overlap as to the preserved information with the qHV. Therefore, class of the qHV, i.e., the class of the inference sample, is subsequently predicted as $x$ as described in ~\ref{eqn:sim}.
\begin{equation}
\begin{aligned}
x = argmax(\delta(\overrightarrow{Q_?}, \mathbb{C}))
\end{aligned}
\label{eqn:sim}
\end{equation}
\section{Experimental Results}
\label{sec:results}
\subsection{Experimental Setup}
\label{sec:setup}
\mypara{Datasets} We use 29 binary classification tasks in total from 3 datasets in the popular \textbf{MoleculeNet} benchmark suite for molecule machine learning~\cite{wu2018moleculenet}. For each dataset, we perform 0.8/0.2 random, stratified and scaffold split to build our training and inference set and repeat 5 experiments to get average performance with error bars. Details of the datasets are as follows:
\begin{itemize}
\item \textbf{BBBP}~\cite{martins2012bayesian} contains 2052 drug compounds and their binary label (positive or negative) of permeability to the blood-brain barrier.
\item \textbf{Clintox}~\cite{gayvert2016data} contains 1491 drug compounds and their binary label (positive or negative) of 1) clinical trial toxicity and 2) FDA approval status.
\item \textbf{SIDER}~\cite{kuhn2016sider} contains 1428 marketed drugs and their adverse drug reactions (ADR) in 27 individual tasks per \textbf{MedDRA} classifications~\cite{brown1999medical}. Each task aims to classifying the positive (active) or negative (inactive) relationship between the drug compound and the ADR disorders of system organs.
\end{itemize}
\mypara{Baseline Models} We compare \textbf{MoleHD} ~with various baseline methods which are roughly in three categories: traditional learning models, GNNs and RNNs.
\begin{itemize}
\item \textbf{Traditional learning Models} including logistic regression \textbf{(LR)}, random forest \textbf{(RF)}, and support vector machine \textbf{(SVM)} implemented and reported in the \textit{MoleculeNet} benchmark~\cite{wu2018moleculenet} and \textit{DeepChem}~\cite{ramsundar2019deep} framework.
\item \textbf{Weave}~\cite{kearnes2016molecular}, which is a graph convolution method that takes both local chemical environment and atom connectivity in featurization.
\item \textbf{MolCLR}~\cite{wang2021molclr}, which is a GNN with contrastive learning of representations with augmentations of atom masking, bond deletion, and subgraph removal.
\item \textbf{D-MPNN}~\cite{yang2019learned, swanson2019message}, which is the directed message passing neural network that operates on molecular graphs.
\item \textbf{LSTM}~\cite{quan2018system}, which applies a modified version of the Smi2Vec tool to convert SMILE strings into atom vectors and then apply long short term memory (LSTM) RNN for classification.
\item \textbf{BiGRU}~\cite{lin2020novel}, which also uses Smi2Vec. It leverages the bidirectional gated recurrent unit (BiGRU) RNN to train sample vectors embedded in the atomic matrix.
\end{itemize}
\subsection{Metrics}
Drug discovery datasets are mostly significantly imbalanced, thus accuracy is generally not considered as a valid metric to reflect performance of a model. Receiver operating characteristics (ROC) curves and ROC Area-under-curve (AUC) scores are mostly embraced as the metric for model prediction performance, as suggested by benchmark datasets along with majority of literature~\cite{wu2018moleculenet, ramsundar2019deep, mayr2018large}. Since HDC models are predicting using similarities, the ``probability'' used in calculating the ROC-AUC score requires specific definition. We propose to use ``confidence level'' (for being positive) $\eta$ in Eq.~\ref{eqn:score}s. Confidence level is derived from similarities between query HV and the class HVs. The larger the difference, the higher the confidence of the HDC model prediction. Because the range of similarity difference is [-2, 2], to perform linear transformation to map the range of confidence level to [0, 1], we accordingly set 1/2 as the average value and 1/4 for coefficient of similarity difference, conforming with the probabilities.
\begin{equation}
\eta = \frac{1}{2} + \frac{1}{4}(\delta(\overrightarrow{Q_?}, \overrightarrow{C_P}) - \delta(\overrightarrow{Q_?}, \overrightarrow{C_N}))
\label{eqn:score}
\end{equation}
\begin{table*}[h!]
\centering
\begin{tabular}{c|ccc|ccc|ccc}
\toprule
split & \multicolumn{3}{c|}{random} & \multicolumn{3}{c|}{stratified} & \multicolumn{3}{c}{scaffold} \\
dataset & Clintox & BBBP & SIDER & Clintox & BBBP & SIDER & Clintox & BBBP & SIDER \\
\midrule
\textbf{MoleHD} & $\mathbf{0.976}^{(1)}$ & $0.879^{(3)}$ & $0.599^{(4)}$ & $0.973^{(2)}$ & $0.916^{(3)}$ & $0.61^{(2)}$ & $\mathbf{0.987}^{(1)}$ & $0.844^{(2)}$ & $0.566^{(4)}$ \\
tokenization & char & char & PE & char & PE & PE & char & char & PE \\
gram size & trigram & trigram & unigram & trigram & unigram & trigram & bigram & bigram & bigram \\
\midrule
LR & 0.733 & 0.737 & 0.643 & - & 0.728 & - & - & 0.699 & - \\
RF & 0.551 & 0.811 & 0.567 & - & 0.736 & - & 0.712 & 0.770 & 0.549 \\
SVM & 0.669 & 0.67 & \textbf{0.656} & - & 0.587 & - & 0.669 & 0.729 & \textbf{0.682} \\
Weave & 0.948 & 0.832 & 0.581 & - & - & - & 0.823 & 0.837 & 0.543 \\
MolCLR & - & - & - & - & - & - & 0.932 & 0.736 & 0.68 \\
D-MPNN & 0.892 & \textbf{0.92} & 0.639 & 0.898 & 0.932 & \textbf{0.655} & 0.874 & \textbf{0.915} & 0.606 \\
LSTM & - & 0.832 & - & - & 0.876 & 0.530 & - & - & - \\
BiGRU & - & 0.889 & - & \textbf{0.978} & \textbf{0.946} & 0.607 & - & - & - \\
\bottomrule
\end{tabular}%
\caption{\textbf{MoleHD} ~vs. Baselines on 3 datasets by average ROC-AUC score. Bold: the highest score. ``-'': data unavailable. Superscript ``(k)'': \textbf{MoleHD} ~ranks $k$-th place amongst all the available models under current dataset and split method.}
\label{tab:beyond}%
\end{table*}%
The experimental results are presented in two parts: the comparison between \textbf{MoleHD} ~and other baseline models as well as the comparison within \textbf{MoleHD} ~configurations.
\subsection{\textbf{MoleHD} ~vs. Baselines}
Most of the baseline models report results not as exhaustive as \textbf{MoleHD} ~in terms of split strategy. Therefore, we are performing comparison by ``best effort'', i.e., we compare best performing \textbf{MoleHD} ~with the baseline with data available under each split method of all the tasks. We decide not to report error bars or variations for this comparison because 1). baseline models may also use different numbers of runs for average and/or cross-validations, and 2). some of the baselines just simply did not report the error bars or variation. However, for all the \textbf{MoleHD} ~versions we implemented, we report all the error bars in Table~\ref{tab:pe} and Table~\ref{tab:char}.
We can observe from the results at Table~\ref{tab:beyond} that, in general, \textbf{MoleHD} ~is achieving high ROC-AUC scores across datasets. For each split, \textbf{MoleHD} ~achieves a dataset-average ROC-AUC scores of \textbf{0.818}, \textbf{0.833} and \textbf{0.799} respectively, \textbf{ranking first on random and scaffold split and second on stratified split}, amongst the models with data available. Particularly, \textbf{MoleHD} ~performs greatly on the Clintox dataset particularly with scaffold split which are often regarded more challenging than the other splits where most of other baseline models are suffering from degradation, the score of \textbf{MoleHD} ~even increases instead.
We have an interesting observation that traditional models such as LR, RF and SVM exhibit poor performance over Clintox and BBBP datasets with significantly low score, however, they show competitive score for the SIDER dataset. On the contrary, while NNs usually performs greatly on Clintox and BBBP datasets, they only show sub-par score even lower than some traditional models, e.g., SVM achieves highest accuracy on SIDER dataset with random and scaffold split.
Robustness-wise, \textbf{MoleHD} ~also outperforms other models. For example, some GNNs are able to achieve top score on a specific dataset, however, they present much lower score on other datasets. For example, for \textbf{D-MPNN}, although it shows high scores at the BBBP dataset by ranking first at random and scaffold split, its score on Clintox seems mediocre. For Weave, it shows significantly degraded score on Clintox dataset from random split to scaffold split. Such variation on performance would arouse questions on those models' transferability, while for \textbf{MoleHD}, the performance is largely consistent across different datasets and split methods.
\begin{table*}[t]
\vspace{-.5em}
\centering
\begin{tabular}{c|ccc|ccc|ccc}
\toprule
split & \multicolumn{3}{c|}{random} & \multicolumn{3}{c|}{stratified} & \multicolumn{3}{c}{scaffold} \\
gram & {uni-gram} & {bi-gram} & {tri-gram} & {uni-gram} & {bi-gram} & {tri-gram} & {uni-gram} & {bi-gram} & {tri-gram} \\
\midrule
Clintox & {$0.941^{0.010}_{0.015}$} & {$0.897^{0.013}_{0.012}$} & {$0.881^{0.024}_{0.013}$} & {$0.960^{0.014}_{0.024}$} & {$0.970^{0.008}_{0.013}$} & {$0.932^{0.025}_{0.018}$} & {$0.952^{0.009}_{0.014}$} & {$0.966^{0.009}_{0.014}$} & {$0.930^{0.020}_{0.021}$} \\
BBBP & {$0.886^{0.014}_{0.024}$} & {$0.875^{0.012}_{0.021}$} & {$0.834^{0.021}_{0.021}$} & {$0.916^{0.014}_{0.014}$} & {$0.908^{0.016}_{0.025}$} & {$0.884^{0.016}_{0.026}$} & {$0.785^{0.024}_{0.023}$} & {$0.802^{0.021}_{0.021}$} & {$0.801^{0.020}_{0.018}$} \\
SIDER & 0.599 & 0.588 & 0.574 & 0.584 & 0.594 & 0.610 & 0.556 & 0.566 & 0.554 \\
\bottomrule
\end{tabular}%
\caption{\textbf{MoleHD}-PE performance on ROC-AUC score comparison on 3 datasets by average. Superscript and subscript refer to the ceiling and floor of errors. For SIDER dataset, the ROC-AUC score is task-average.}
\label{tab:pe}%
\end{table*}%
\begin{table*}[h!]
\vspace{-.5em}
\centering
\begin{tabular}{c|ccc|ccc|ccc}
\toprule
split & \multicolumn{3}{c|}{random} & \multicolumn{3}{c|}{stratified} & \multicolumn{3}{c}{scaffold} \\
gram & {uni-gram} & {bi-gram} & {tri-gram} & {uni-gram} & {bi-gram} & {tri-gram} & {uni-gram} & {bi-gram} & {tri-gram} \\
\midrule
Clintox & {$0.955^{0.023}_{0.023}$} & {$0.971^{0.015}_{0.016}$} & {$0.976^{0.016}_{0.025}$} & {$0.956^{0.011}_{0.028}$} & {$0.971^{0.019}_{0.020}$} & {$0.973^{0.010}_{0.014}$} & {$0.966^{0.008}_{0.010}$} & {$0.987^{0.001}_{0.001}$} & {$0.982^{0.002}_{0.002}$} \\
BBBP & {$0.850^{0.022}_{0.028}$} & {$0.879^{0.034}_{0.029}$} & {$0.879^{0.026}_{0.020}$} & {$0.860^{0.017}_{0.020}$} & {$0.877^{0.014}_{0.013}$} & {$0.865^{0.012}_{0.015}$} & {$0.805^{0.009}_{0.011}$} & {$0.844^{0.006}_{0.010}$} & {$0.828^{0.004}_{0.002}$} \\
SIDER & 0.580 & 0.544 & 0.525 & 0.578 & 0.544 & 0.514 & 0.553 & 0.541 & 0.565 \\
\bottomrule
\end{tabular}%
\caption{\textbf{MoleHD}-char performance on ROC-AUC score comparison on 3 datasets by average. Superscript and subscript refer to the ceiling and floor of errors. For SIDER dataset, the ROC-AUC score is task-average.}
\label{tab:char}%
\end{table*}%
\subsection{Results between different \textbf{MoleHD} ~versions}
In addition to comparing with baseline models, we also evaluate an intensive set of \textbf{MoleHD} ~and dataset configurations, including: two tokenization schemes (\textbf{MoleHD}-PE and \textbf{MoleHD}-char), two gram sizes (uni-gram and bi-gram), and three dataset split methods (random, stratified and scaffold split).
In general, the performance of \textbf{MoleHD} ~is overall consistent, thus, there is no single configuration that can dominate other configurations for most, if not all, the datasets and split methods. However, we do observe that for the scaffold split, the score variation is generally smaller than that of random and stratified split, as suggested by the error bars.
\subsection{Ultra-Low-Cost Computing of \textbf{MoleHD}}
We elaborate the computing cost of \textbf{MoleHD} ~and compare it with SOTA neural networks.
1) Unlike GNNs, \textbf{MoleHD} ~does not require specific additional effort on pre-training the model. 2) For all the reported datasets together, \textbf{MoleHD} ~is able to achieve the reported accuracy within 10 minutes using CPU only from the commodity desktop (AMD Ryzen\texttrademark ~5 3600 3.6 GHz). Note that this includes both training and inference for each dataset. 3) \textbf{MoleHD} ~also requires less space for model storage as for one binary classification task, model size of \textbf{MoleHD} ~is only around 80kB and during run-time, the memory footprint of \textbf{MoleHD} ~is also generally less than 10MB. For GNNs as a comparison, extensive pre-training can be necessary, e.g., MolCLR requires around 5 days of pre-training using Nvidia\textregistered ~Quadro RTX\texttrademark ~6000 as reported in the corresponding literature~\cite{wang2021molclr}. Neural network models such as GNNs and RNNs are also harder to implement considering the effort of establish multiple layers with considerable amount of nodes with the model size at 100MB level, especially considering the necessity of performing back-propagation during training with millions of parameters in total~\cite{ma2020multi}.
\section{Conclusion}
In this paper, we propose \textbf{MoleHD}, an ultra-low-cost learning model which leverages the novel brain-inspired hyperdimensional computing for molecule property prediction in drug discovery. \textbf{MoleHD} ~projects SMILES strings of drug compound into hypervectors in the hyperdimensional space to extract features. The hypervectors are then used during training, retraining and inference of the HDC model to perform learning tasks. We evaluate \textbf{MoleHD} ~on 29 classification tasks from 3 widely-used benchmark datasets and compare \textbf{MoleHD} ~performance with 8 baseline machine learning models including SOTA GNNs and RNNS. According to experimental results, \textbf{MoleHD} ~is able to achieve highest ROC-AUC score on random and scaffold splits on average across 3 datasets and achieve second-highest on stratified split. Compared with traditional models and NNs, \textbf{MoleHD} ~also requires less training efforts, smaller model size, as well as smaller computation costs. This work marks the potential of using hyperdimensional computing as an alternative to the existing models in the drug discovery domain.
|
\section{Introduction}
"Spooky action at the distance" between particles was firstly noted in the famous Einstein-Podolsky-Rosen (EPR) argument against the completeness of the Copenhagen interpretation of quantum mechanics~\cite{PhysRev.47.777}. In response to the EPR work, Schrödinger described a peculiar phenomenon, where one party can \textit{steer} the other, but not conversely~\cite{Schrodinger1935}. Nonlocal correlations were firstly considered as unnatural, even though they do not allow communication. For three decades the EPR paradox remained almost a philosophical question. However, in 1964 Bell proposed a test of realistic theories~\cite{PhysicsPhysiqueFizika.1.195}, which resulted in a great progress in theoretical and experimental research on quantum nonlocality~\cite{RevModPhys.86.419}. The systematic research on spatial steering initiated by Wieseman \textit{et al.}~\cite{PhysRevLett.98.140402} lead numerous results regarding various aspects of this phenomenon and its relation to other types of quantum correlations \cite{RevModPhys.92.015001}. Remarkably, we can distinguish at least three different types of quantum correlations, i.e., entanglement, Bell nonlocality, and EPR steering. What is more, there exists a hierarchical relation between them. That is, every Bell nonlocal state is steerable and every steerable state is entangled, but not vice versa~\cite{PhysRevLett.98.140402}. As it turns out this hierarchy can be tested experimentally~\cite{Jirakova2021}.
\par In 1985 Legget and Garg, assuming \textit{macroscopic realism} (MR) and \textit{noninvasive measurability}, proposed a Bell-like experiment to test the MR of a single system at two different times \cite{Leggett1985,Emary2013}. Thereafter, other tests of MR were proposed, e.g., Chen \textit{et al.}~\cite{PhysRevA.89.032112} introduced a temporal counterpart of EPR steering inequality, which was further studied in relation to quantum key distribution protocols in~\cite{bartkiewicz2016temporal} or non-Markovianity~\cite{chen2016quantifying}. Next to these theoretical investigations, temporal steering was also reported experimentally~\cite{bartkiewicz2016experimental}. Another type of temporal correlations is the quantum causality. Confidence of validity of definite causal order is fundamental in our understanding of physical reality. However, recent research on quantum phenomena indicates that causality cannot be treated as an axiom. For example, there can exist superpositions of causual structures~\cite{Brukner2014} or other couterintuitive phenomena~ \cite{MacLean2017,Oreshkov2012,Chiribella2013,fitzsimons2015quantum}. The degree of casuality can be quantified by a monotone of causality proposed in~ \cite{fitzsimons2015quantum}. Hierarchy of temporal correlations under the assumption of no-signaling in time (NSIT) for two-level systems was shown in~\cite{mal2016probing,PhysRevA.98.022104}. Similarly to the spacial case, temporal Bell nonlocality implies temporal steering, and temporal steering implies temporal counterpart of entanglement, i.e., the quantum causal effect.
\par The goal of this paper is to investigate relations between temporal correlations in systems of arbitrary dimensions. For systems larger than qubits, the spooky action over time requires more study. Theese can exhibit unusual properties as it happens to be the case with studying quantum contextuality~\cite{Budroni2021}, which can be only studied in systems larger than qubits. To analyze such systems, we extend the notion of a pseudo density operator (PDO), originally introduced by Fitzsimons \textit{et al.}~\cite{fitzsimons2015quantum} for multi-qubit system, to describe odd-dimensional systems, and compare it to a Wigner function formulation from Ref.~\cite{WOOTTERS19871}. In Sec.~\ref{sec:results}, it is proven that the hierarchy between temporal nonlocality, temporal steering, and temporal entanglement holds under the NSIT assumption. However, as we also demonstrate here, in systems which do not satisfy the NSIT condition the hierarchy can be broken.
\par Quantification and description of the quantum correlations can be performed in various ways, including, e.g., a geometric approach used for EPR steering in Ref.~\cite{Ku2018} or matrices of moments applied for detecting various nonclassical correlations (see, e.g., Ref.~\cite{Miranowicz2010}). However, to treat all types of temporal correlations consistently, we decided to quantify robustness of quantum correlations. In Sec.~\ref{sec:framework} we introduce new quantifiers of robustness of quantum correlations, that is \textit{temporal nonlocality robustness} (TNR) and \textit{temporal entanglement robustness} (TER) as direct counterparts of spatial nonlocality robustness~\cite{PhysRevA.93.052112} and spacial entanglemnt robustness~\cite{PhysRevA.59.141}, respectively. The comparison of TNR, TER, and \textit{temporal steering robustness} (TSR) included in this paper provides a systematic picture of the correlations. Our concussions are summarized in Sec.~\ref{sec:conclusions}.
\section{Theoretical framework}\label{sec:framework}
\subsection{States over time and temporal entanglement}
The notion of quantum mechanical states defined over time is far from being a trivial counterpart of a single-time multidimensional state~\cite{PhysRevA.79.052110,PhysRevA.88.052130,fitzsimons2015quantum,Horsman2017}. The definition of a pseudo-density operator (PDO), that we extend in this paper, was introduced in Ref.~\cite{fitzsimons2015quantum} in order to establish a formalism which treats space and time indiscriminately. This proposition of state over time was further studied in \cite{Zhao2018,Pisarczyk2019,Zhang2020,Zhang_2020-2}.Analogously to $n$ qubits in space, let us consider $n$ events in space-time, where at each event a measurement of a single Pauli operator can be performed. A density matrix describing system composed of $n$ qubits can be written as
\begin{equation}
\rho = \frac{1}{2^n}\sum_{i_1=0}^3\dots\sum_{i_n=0}^3\left\langle
\bigotimes_{j=1}^n\sigma_{i_j}\right\rangle\bigotimes_{j=1}^n\sigma_{i_j}, \label{eq:nqubit_dm}
\end{equation}
where $\sigma_i$ are Pauli matrices. This leads to Fitzsimons' \textit{et al.} proposal for a state over time~\cite{fitzsimons2015quantum}:
\begin{equation}
R^{\text{PDO}} = \frac{1}{2^n}\sum_{i_1=0}^3\dots\sum_{i_n=0}^3\left\langle\left\{\sigma_{i_j}\right\}_{j=1}^n\right\rangle\bigotimes_{j=1}^n\sigma_{i_j},\label{eq:R_PDO}
\end{equation}
where $\langle\{\sigma_{i_j}\}_{j=1}^n\rangle=\textrm{tr}[(\bigotimes_{j=1}^n\sigma_{i_j})R^{\text{PDO}}]$ is the expectation value of the product of the results of measurements. In spite of the fact that definition (\ref{eq:R_PDO}) for two instances refers to measurements performed at $t_A$ and $t_B$, $R^{\text{PDO}}$ describes an initial state $\rho_A$ evolving to $\rho_B$ under the quantum channel $\mathcal{E}:\rho_A\mapsto\rho_B$, represented via Choi--Jamiołkowski isomorphism \cite{jamiolkowski1972linear,choi1975completely} as operator
\begin{equation}
E_{B\vert A}=\sum_{i,j}\vert i\rangle\langle j\vert_A\otimes\mathcal{E}(\vert j\rangle\langle i\vert_A).\label{eq:E_BA}
\end{equation}
Density matrices $\rho_A$ and $\rho_B$ are spanned on $\mathcal{H}_A$ and $\mathcal{H}_B,$ respectively. Hence, the PDO is explicitly determined by the initial state and the channel.
\par However, a state over time can be defined in more than one way. For example, in Ref.~\cite{WOOTTERS19871} authors construct it using the quasi-probabilities from a discrete Wigner representation \cite{Gibbons2004,Gross2006}, which is defined by a set of phase-space point operators $\{K_i\}_{i=0}^{d^2-1}$. Operators $K_i$ form a basis for a $d^2$-dimensional Hilbert space $\mathcal{H}$ with the Hilbert-Schmidt product $\langle X\vert Y\rangle=\tr(XY)$. They satisfy $\sum_{i}K_i=d\openone,$ $\tr(K_i K_j)=d\delta_{ij}$ and $\tr(K_i)=1$. Since arbitrary density matrix $\rho_A$ and operator $E_{B\vert A}$ can be written as $\sum_{i}r(i)K_i^A$ and $\sum_{i,j}r(j\vert i)K_i^A\otimes K_j^B$, respectively, one may describe an evolution of a quantum system between events $A$ and $B$ by a different operator
\begin{equation}
R^{\text{W}} = \sum_{i,j}r(j\vert i)r(i)K_{i}^{A}\otimes K_{j}^B,\label{eq:R_W}
\end{equation}
where quasi-probability distributions $r(i)$ and $r(j\vert i)$ are given by $\tr(\rho_A K_i)$ and $\tr[E_{B\vert A}(K_i^A\otimes K_j^B)]$, respectively. In general, the properties of various operators $R$ can be studied via the star product $\star$ (see Ref.~\cite{Horsman2017}), a binary operation $\star:X\times X\mapsto X\in\mathcal{H}_A\otimes\mathcal{H}_B$.
In this framework a whole class of definitions of a sate over time
\begin{equation}
R = \rho_A\star E_{B\vert A}\label{eq:star}
\end{equation}
can be represented regardless of their explicit form.
\par A given state over time does not need to be positive semi-definite.
If $R\geq 0$, then it might be considered as a regular density matrix. However, negative eigenvalues of the state over time imply time-like correlations between events. A proper causality measure for these events $\Phi(R)$ has to distinguish between space-like and time-like correlations and be a causality monotone. A given function $\Phi(R)$ is a causality monotone if \cite{fitzsimons2015quantum}:
\begin{itemize}
\item $\Phi(R)\geq 0$, with $\Phi(R)=0$ if $R$ is completely positive,
and $f(R)$ is maximal for any $R$ obtained from two consecutive measurements on a closed system,
\item $\Phi(R)$ is invariant under unitary operations,
\item $\Phi(R)$ is non-increasing under local operations,
\item $\sum_ip_i\Phi(R_i)\geq \Phi(\sum_ip_iR_i)$.
\end{itemize}
The above criteria are similar to these for an entanglement monotone \cite{RevModPhys.81.865}. Therefore, we refer to the phenomena for which $\Phi(R)\neq 0$ as \textit{temporal entanglement}. In Ref.~\cite{fitzsimons2015quantum} authors proposed a simple monotone, i.e., $f(R)=\|R\|_{\tr}-1$. A separable state over time can be expressed as
\begin{equation}
R^{\mathrm{SEP}}=\sum_k p_k\rho_k^A\otimes\rho_k^B,\label{eq:separableR}
\end{equation}
where $p_k$ stands for a certain probability distribution, $\rho_k^A$, $\rho_k^B$ are density matrices spanned on Hilbert spaces $\mathcal{H}_A$ and $\mathcal{H}_B,$ respectively. This can be interpreted as a space-like PDO being generated from a maximally entangled space-like state, where subsystem $B$ is subjected to a dissipative channel. However, $R\geq 0$ does not imply separability, i.e., there can still exist space-like temporal correlations. A quantum state is \textit{spatially entangled,} if it cannot be written as a convex combination of product states (\ref{eq:separableR}). Otherwise the state is \textit{separable}.
\par Quantum entanglement of a given state $\sigma$ can be quantified, e.g., as the minimal amount of noise $\tau$ that makes the state separable~\cite{PhysRevA.59.141}. This entanglement robustness (ER) can be expressed as
\begin{eqnarray}
\text{ER}=&&\,\min\,\gamma,\nonumber\\
\text{s.t. }&& \frac{\sigma+\gamma\tau}{1+\gamma}=\rho,\nonumber\\
&&\rho=\sum_k p_k\rho_k^A\otimes\rho_k^B,\nonumber\\
&&p_k\geq 0,\quad\rho_k^A,\rho_k^B\geq 0\quad\forall k,\nonumber\\
&&\tr(\tau)=\tr(\rho_k^A)=\tr(\rho_k^B)=1\quad\forall k,\nonumber\\
&&\sum_kp_k=1,\quad\tau\geq 0,\quad\gamma\geq 0.\label{eq:ER}
\end{eqnarray}
By analogy, we introduce temporal entanglement robustness (TER) as minimal amount of noise $\mathfrak{R}$ needed to destroy temporal causality. The relevant optimization problem reads
\begin{eqnarray}
\text{TER}=&&\,\min\,\gamma\nonumber\\*
\text{s.t. }&&\frac{R+\gamma\mathfrak{R}}{1+\gamma}=\rho,\nonumber\\*
&&\mathfrak{R}:\;\text{a pseudo-density operator},\nonumber\\*
&&\rho\geq 0,\quad\gamma\geq 0.\label{eq:TER}
\end{eqnarray}
By substituting the product $\gamma\mathfrak{R}$ with unnormalized state over time $\tilde{\mathfrak{R}}$ we can easily rewrite the above problem in a linear SDP form,
\begin{eqnarray}
&&\text{TER}=\min\left(\tr\tilde{\mathfrak{R}}\right),\nonumber\\*
&&\text{s.t. }R+\tilde{\mathfrak{R}}\geq 0.\label{eq:TER2}
\end{eqnarray}
We shown in Appendix\ref{App:A} that TER (\ref{eq:TER}) satisfies the above-listed criteria for a causality monotone.
\subsection{Temporal steering}
Consider two parties $A$ (Alice) and $B$ (Bob). At time $t_A$ Alice receives an initial state $\rho_A$. Then, she performs a positive-operator valued measurement (POVM) $\{M_{a\vert x}\}$ and obtains result $a$ with probability $p(a\vert x)=\tr_A(\rho_A M_{a\vert x})$. Next, the state
\begin{equation}
\rho_{a\vert x}\left(0\right)=\sqrt{M_{a\vert x}}\rho_A\sqrt{M_{a\vert x}}\label{eq:init_conditional_state}
\end{equation}
is delivered to $B$ through a channel represented by a Choi-Jamiołkowski operator $E_{B\vert A}$. Bob receives the conditional state
\begin{eqnarray}
\rho_{a\vert x}\left(t\right)&&\equiv\rho_{a\vert x}=\mathcal{E}\left(\sqrt{M_{a\vert x}}\rho_A\sqrt{M_{a\vert x}}\right)\nonumber\\*
&&{}=\tr_A\left(E_{B\vert A}\sqrt{M_{a\vert x}}\rho_A\sqrt{M_{a\vert x}}\right).\label{eq:conditional_state}
\end{eqnarray}
If Alice and Bob repeat their protocol many times, they can reconstruct their shared set of unnormalized conditional states $\{\tilde{\rho}_{a\vert x}=p(a\vert x)\rho_{a\vert x}\},$ referred to as \textit{assemblage}.
\par If Bob wants to explain receiving $\tilde{\rho}_{a\vert x}$ with a \textit{hidden state model} (HSM), he can assume that Alice sends state $\tilde{\rho}_{a\vert x}$ through different channels according to probability distribution $p(\lambda)$. Hence, there exists a classical variable $\lambda$ which determines both the Alice's outcome and the conditional state $\rho_{a\vert x}=\rho_\lambda$. Thus, according to Bob, a conditional state can be written as
\begin{equation}
\rho_{a\vert x}=\sum_\lambda p(\lambda)D(a\vert x,\lambda)\rho_\lambda\quad\forall a,x.\label{eq:HSM}
\end{equation}
If the state $\rho_{a\vert x}$ can be expressed as in Eq.~(\ref{eq:HSM}), Bob concludes that the state at time $t_A$ does not influence the assemblage $\rho_{a\vert x}$. Thus, he excludes the possibility of Alice steering his state. In such case $\rho_{a\vert x}$ is \textit{temporally unsteerable}. Conversely, if a conditional states do not admit the decomposition (\ref{eq:HSM}), then Bob concludes that Alice steers his state. We refer to assemblages which do not admit the HSM as \textit{temporal steerable}.
\par In this paper we quantify the steerability of an assemblage with \textit{temporal steering robustness} (TSR), introduced in \cite{PhysRevA.94.062126,Chen2017SciRep} as a temporal couterpart of steering robustness \cite{PhysRevLett.114.060404}. TSR is defined as the minimal amount of noise needed to destroy the temporal steerability of a given assemblage:
\begin{eqnarray}
\text{TSR}=&&\,\min\,\alpha\nonumber\\*
\text{s.t. }&&\left\{\frac{\tilde{\rho}_{a\vert x}+\alpha\tau_{a\vert x}}{1+\alpha}\right\}_{a,x}\text{ temporal unsteereable,}\nonumber\\*
&&\left\{\tau_{a\vert x}\right\}:\,\text{a noisy assemblage.}\label{eq:TSR}
\end{eqnarray}
The optimization problem (\ref{eq:TSR}) can be rewritten as a semi-definite program (SDP)
\begin{eqnarray}
\text{TSR}=&&\min\left(\tr\sum_\lambda\tilde{\rho}_\lambda-1\right)\nonumber\\*
\text{s.t. }&&\sum_\lambda D(a\vert x,\lambda)\tilde{\rho}_\lambda-\rho_{a\vert x}\geq0\quad\forall a,x,\nonumber\\*
&&\tilde{\rho}_\lambda\geq0\quad\forall\lambda,\label{eq:TSR_SDP}
\end{eqnarray}
where $\tilde{\rho}\equiv(1+\alpha)\rho_\lambda$ and $D(a\vert x,\lambda)=\delta_{a,\lambda(x)}.$
\subsection{Temporal counterpart of Bell nonlocality}
Another type of correlations, different than steering, can be detected by a Bell-type experiment. Let us consider a scenario, where Alice and Bob make measurements of observables $A_x$ and $B_y,$ respectively. Quantum mechanics predicts that the probability of obtaining results $a$ and $b$ is given by the Born rule, i.e.,
\begin{eqnarray}
P(a,b\vert x,y)=\tr\left[M_{b\vert y}\mathcal{E}\left(\sqrt{M_{a\vert x}}\rho_A\sqrt{M_{a\vert x}}\right)\right]\nonumber\\
\forall a,b,x,y,\label{eq:p(ab|xy)}
\end{eqnarray}
where $M_{c\vert z}$ for $c=a,b$ and $z=x,y$ stands for a POVM describing the measurement of value $c$ associated with observable $C_z$ (i.e. $A_z$ or $B_z$). The joint probability distribution $\{P(a,b\vert x,y)\}_{a,b,x,y}$ is also referred to as \textit{behavior} of the system.
\par A \textit{local hidden variable} (LHV) model can be constructed, if the behavior $P(a,b\vert x,y)$ can be expressed as
\begin{equation}
P(a,b\vert x,y)=\sum_{\mu,\nu}p(\mu,\nu)D(a\vert x,\mu)D(b\vert y,\nu),\label{eq:LHV}
\end{equation}
where $\mu$ and $\nu$ are classical random variables given by a preexisting distribution $p(\mu,\nu),$ and $D(a\vert x,\mu)$ and $D(b\vert y,\nu)$ are local deterministic response functions. The existence of an LHV model implies that the behavior $P(a,b\vert x,y)$ is macro-realistic and that correlations between $A$ and $B$ can origin in a third party controlling measurement devices at $A$ and $B$.
\par We refer to the sets of probability distributions which cannot be expressed by Eq.~(\ref{eq:LHV}) as \textit{nonlocal}. We quantify the degree of Bell-nonlocality with a temporal version of nonlocality robustness from Ref.~\cite{PhysRevA.93.052112}. Temporal nonlocality robustness (TNR) is defined as the minimal amount of behavior $Q(a,b\vert x,y)$ needed to make $P(a,b\vert x,y)$ local,
\begin{eqnarray}
\text{TNR}&&{}=\min\,\beta,\nonumber\\*
\text{s.t. }&&\left\{\frac{P(a,b\vert x,y)+\beta Q(a,b\vert x,y)}{1+\beta}=R(a,b\vert x,y)\right\},\nonumber\\*
&&R(a,b\vert x,y)=\sum_{\mu,\nu}r(\mu,\nu)D(a\vert x,\mu)D(b\vert y,\nu),\nonumber\\*
&&\left\{Q(a,b\vert x,y)\right\}\text{: a behavior,}\quad\beta\geq 0.\label{eq:TNR}
\end{eqnarray}
Additionally, we can introduce local hidden variable temporal nonlocality robustness (LHV TNR) by requiting that the noise is an LHV behavior, i.e.,
\begin{eqnarray}
Q(a,b\vert x,y)=&&{}\sum_{\mu,\nu}q(\mu,\nu)D(a\vert x,\mu)D(b\vert y,\nu),\nonumber\\*
&&q(\mu,\nu)\geq 0\quad\forall\mu,\nu.
\end{eqnarray}
Following the techniques presented in \cite{PhysRevLett.114.060404}, one may rewrite this problem in the linear SDP form
\begin{eqnarray}
\text{TNR}&&{}=\min\left[\frac{\sum_{x,y,\mu,\nu}\tilde{r}(\mu,\nu)}{\sum_{x,y,a,b}P(a,b\vert x,y)}-1\right],\nonumber\\*
\text{s.t. }&&\sum_{\mu,\nu}\tilde{r}(\mu,\nu)D(a\vert x,\mu)D(b\vert y,\nu)\geq P(a,b\vert x,y),\nonumber\\*
&&\tilde{r}(\mu,\nu)\geq 0\quad\forall\mu,\nu,a,b,x,y,\label{eq:TNR_SDP}
\end{eqnarray}
where $\tilde{r}(\mu,\nu)$ stands for the unnormalized probability distribution $(1+\beta)r(\mu,\nu)$. For details of the derivation see Appendix~\ref{App:B}. In the case of LHV TNR we require that there exists such $\tilde{q}(\mu,\nu)$ that
\begin{eqnarray}
\sum_{\mu,\nu}\tilde{r}&&(\mu,\nu)D(a\vert x,\mu)D(b\vert y,\nu)\nonumber\\*
=&&{}\sum_{\mu,\nu}\tilde{r}(\mu,\nu)D(a\vert x,\mu)D(b\vert y,\nu)-P(a,b\vert x,y),\nonumber\\*
&&{}\tilde{q}(\mu,\nu)\geq 0,\quad\forall a,b,x,y.
\end{eqnarray}
\par Both temporal Bell inequalities and LHS model assume non-invasive measurements, for which a necessary condition is the no-signaling in time (NSIT) condition \cite{Kofler2013,Halliwell2017}. NSIT is obeyed if the Alice's measurement does not change the outcome statistics of the Bob's measurement. It implies that the elements of assemblage have to satisfy the following relation,
\begin{equation}
\sum_a\varrho_{a\vert x}=\tr_A\left(E_{B\vert A}\rho_A\right)\quad\forall x.\label{eq:NSIT}
\end{equation}
To demonstrate that a state does not satisfy the NSIT condition it is sufficient to find such a pair of $x_1,x_2$ for which
\begin{equation}
\sum_a\varrho_{a\vert x_1}\neq\sum_a\varrho_{a\vert x_2}.
\end{equation}
This is the case if we choose the initial state to be a pure state, i.e., an eigenstate of $x_1$ and the channel is the identity channel. Then the eigenvectors of operator $x_2$ form mutually unbiased set with respect to the eigenvectors of $x_1.$ In this extreme case, one side of the equality corresponds to a pure state, while the other to a maximally mixed state. Thus, we can expect pure states to violate the NSIT condition for some channels.
\section{Results}\label{sec:results}
\subsection{Bipartite pseudo-density operator of a \textit{d}-dimensional system}
Let us consider a quantum system AB composed of two qudits. Assume that operators $G_i^A$ and $G_j^B$ form orthogonal basis (i.e., $\tr\left(G_iG_j\right)=d\delta_{i,j},$ where $d>0$) in $\mathcal{H}_A$ and $\mathcal{H}_B$, respectively. A density matrix describing AB can be written in terms of $G_i^A$ and $G_j^B$ as
\begin{equation}
\rho_{AB}=\frac{1}{\mathcal{N}}\sum_{i=0}^{d_A^2-1}\sum_{j=0}^{d_B^2-1}C_{ij}G_i^A\otimes G_j^B,\label{eq:rho1}
\end{equation}
where $\mathcal{N}$ is a normalization factor such that $\tr(\rho_{AB})=1$ and $C_{ij}=\mathcal{N}\tr[\rho_{AB}(G_i^A\otimes G_j^B)]$ is a correlation tensor. To make the following expressions more legible, from this point we will denote all operators simply as $G_i$ and assume that $d_A = d_B=d$.
\par Equation (\ref{eq:rho1}) suggests the following generalization of the notion of the PDO,
\begin{equation}
R^{\text{PDO}}=\frac{1}{\mathcal{N}}\sum_{i,j=0}^{d^2-1}\left\langle G_i\otimes G_j\right\rangle G_i\otimes G_j,\label{eq:new_PDO}
\end{equation}
where $\langle G_i\otimes G_j\rangle$ stands for the expectation value of the product of the result of measurements of operators $G_i$ and $G_j$ at $t_A$ and $t_B$, respectively. Formally, this reads
\begin{eqnarray}
\left\langle G_i\otimes G_j\right\rangle&=&\tr\left[R^{\text{PDO}}\left(G_i\otimes G_j\right)\right]\nonumber\\*
&=&\sum_a a\tr\left[E_{B\vert A}\left(\rho_{i,a}^A\otimes G_j\right)\right].\label{eq:correlation_tensor}
\end{eqnarray}
Operator $\rho_{i,a}^A=\Pi_{i,a}\rho_A\Pi_{i,a}$ denotes the subnormalized state of $A$
after projection $\Pi_{i,a}$ onto the eigenspace corresponding to an eigenvalue $a$ of operator $G_i.$
\par Since in 2D case the PDO is defined by the use of Pauli matrices, it would be natural to establish $G_i$ as generators of special unitary group $\text{SU}(d)$. However, due to contextuality~\cite{Budroni2021} between measurements of the $\text{SU}(d)$ generators, such definition is dependent on a choice of basis of $\mathcal{H}_A$ and $\mathcal{H}_B$. One may explicitly verify that $R^{\text{PDO}}$ has different spectrum for different pure initial states subjected to the identity channel. Entanglement is not base-dependent so we cannot quantify it reliably utilizing base-dependent quantities. Therefore, we have to use non-contextual scenarios to define a proper state over time. For systems of odd, prime dimension the subset of quantum states known to be classically simulable, and thus forming non-contextual scenario, is prescribed by the Wigner polytope \cite{Gross2006,Dawkins2015PRL},
\begin{equation}
\text{Wigner polytope}=\{\rho\,\vert\,\tr(\rho K_i)\geq 0\;\forall i\}.
\end{equation}
Hence, by setting $G_i=K_i$ we make $R^{\text{PDO}}$ independent of a chosen basis for quantum systems being in the Wigner polytope. It is worth emphasizing that definitions (\ref{eq:R_W}) and (\ref{eq:new_PDO}) are not equivalent. Quasi-probability distribution $r(j\vert i)r(i)$ differs from expectation values $\langle G_i\otimes G_j\rangle$ used in equations (\ref{eq:R_PDO}) and (\ref{eq:new_PDO}).
\par In particular, for a qutrit (see Ref~\cite{Dawkins2015PRL}) the discrete Wigner-space point operators $K_i$ operators can be selected as
\begin{equation*}
K_1=\left[
\begin{array}{ccc}
1 & 0 & 0 \\
0 & 0 & 1 \\
0 & 1 & 0 \\
\end{array}\right],\,
K_2=\left[
\begin{array}{ccc}
0 & 1 & 0 \\
1 & 0 & 0 \\
0 & 0 & 1 \\
\end{array}\right],\,
K_3=\left[
\begin{array}{ccc}
0 & 0 & 1 \\
0 & 1 & 0 \\
1 & 0 & 0 \\
\end{array}\right],
\end{equation*}
\begin{eqnarray*}
K_4&=&\left[
\begin{array}{ccc}
1 & 0 & 0 \\
0 & 0 & \frac{-1-\sqrt{3}i}{2} \\
0 & \frac{-1+\sqrt{3}i}{2} & 0 \\
\end{array}\right],\\
K_5&=&\left[
\begin{array}{ccc}
0 & \frac{-1-\sqrt{3}i}{2} & 0 \\
\frac{-1+\sqrt{3}i}{2} & 0 & 0 \\
0 & 0 & 1 \\
\end{array}\right],\\
K_6&=&\left[
\begin{array}{ccc}
0 & 0 & \frac{-1-\sqrt{3}i}{2} \\
0 & 1 & 0 \\
\frac{-1+\sqrt{3}i}{2} & 0 & 0 \\
\end{array}\right],\\
\end{eqnarray*}
\begin{equation}
K_7 = K_4^T,\qquad
K_8 = K_5^T,\qquad
K_9 = K_6^T.
\end{equation}
These matrices differ from matrices usually used to express a density matrix of three-level system, i.e.,
Patera-Zassenhaus~\cite{Patera1988JMP} matrices, Gell-Mann~\cite{Gell-Mann1962PR} matrices, or matrices of spin-1 spherical symmetries~\cite{Hofmann2004PRA}. We confirmed numerically our prediction that applying matrices $K$ in the definition of a PDO for qutrits makes the spectrum of a PDO independent of a particular choice the pure initial state.
\subsection{Relations between temporal entanglement and separability for qubits and three-level systems}
\begin{figure*}
\includegraphics[width=1\linewidth]{fig1.pdf}
\caption{(color online) The dynamics of (a)-(c) TER and (d)-(f) ER for a qubit initially in state evolving under [(a),(d)] amplitude damping channel, [(b),(e)] phase damping channel, and [(c),(f)] depolarizing channel. Time $t$ is measured in units of decay rate $\gamma$. Initially the system is in: (solid curve) vacuum state $\omega_1:\vert 0\rangle,$, (dashed curve) balanced superposition $\omega_2:\frac{1}{\sqrt{2}}(\vert 0\rangle + \vert 1\rangle),$ and (dotted curve) maximally mixed state $\omega_3:\openone/2.$ We observe that for systems initially in pure states positive semi-definiteness of PDO does not imply separability. For depolarization we observe vanishing of TER and ER for the same finite time. Also ER calculated for $\omega_3$ (in contrast to pure states) is above $1/2,$ which is an idicator of a physical inequivalence between TER and ER.}
\label{fig:ERvsTER}
\end{figure*}
In Ref.~\cite{PhysRevA.98.022104} the authors proved that a PDO describing a maximally mixed qubit transmitted through amplitude damping, phase damping or depolarizing channel is separable, if $R^{\text{PDO}}$ is positive semi-definite. Let us firstly investigate the case of a pure initial state. In the cases of qubit-qubit and qubit-qutrit bipartite systems Peres-Horodecki positive partial transpose (PPT) criterion \cite{PhysRevLett.77.1413,HORODECKI19961} is both necessary and sufficient condition for separability of the system. This criterion can be applied to simplify the optimization problem (\ref{eq:ER}) to a linear SDP form, i.e.,
\begin{eqnarray}
\text{ER}=&&{}\min\left[\tr\left(\mathfrak{R}\right)-1\right],\nonumber\\*
\text{s.t. }&&R+\mathfrak{R}=\rho,\nonumber\\*
&&\rho\geq 0,\nonumber\\*
&&\rho^{\mathrm{PT}}\geq 0,\label{eq:ER-2level}
\end{eqnarray}
where $\mathrm{PT}$ stands for a partial transposition. By considering a simple example of a qubit in a pure state at time $t_A$ subjected to amplitude damping, phase damping or depolarizing channel we can demonstrate that TER (\ref{eq:TER}) and ER (\ref{eq:ER-2level}) are not equivalent (see Fig.~\ref{fig:ERvsTER}). It is clear that space-like separable state over time is also temporally separable.
\par In the case of a system of an arbitrary dimension the sufficient condition is that for all positive maps $\Lambda$ acting on $\mathcal{H}$, a separable density matrix of a composite system $\rho\in\mathcal{H}\otimes\mathcal{H}$ satisfies $(\Lambda\otimes\openone)\rho\geq 0$. Equivalently, one can find an entanglement witness $W\in\mathcal{H}\otimes\mathcal{H}$, Choi-isomorphic to $\Lambda$, such that $\tr(W\rho)\geq 0$ necessarily for a separable matrix $\rho$ and $\tr(W\rho)\leq 0,$ otherwise~\cite{GUHNE20091,RevModPhys.81.865}. However, testing $\rho$ under all of positive maps $\Lambda$ is impractical. Hence, we are forced to use necessary but not sufficient criteria \cite{GUHNE20091,RevModPhys.81.865} in order to detect spacial entanglement. These criteria are not universal but they can be easily applied and allow to distinguish many separable states. In Appendix~\ref{App:C} we show that linear separability criteria based on correlation matrix, which have been recently unified in Ref.~\cite{PhysRevA.101.012341}, are satisfied by a positive semi-definite PDO representing a single qutrit, initially in maximally mixed state, undergoing depolarizing, phase damping, and amplitude damping. Hence, it is reasonable to assume, as in Ref.~\cite{PhysRevA.98.022104} for the two-level case, that
\begin{eqnarray}
&&R\geq 0\,\Rightarrow\,\exists\,\rho_k^A,\,\rho_k^B,\nonumber\\*
&&\text{such that }R=\sum_kp_k\rho_k^A\otimes\rho_k^B,\,\rho_A=\tfrac{1}{3}\openone.\label{eq:proposition}
\end{eqnarray}
{However, proving this proposition is beyond the scope of this paper.}
\subsection{Hierarchy of two-time three-level correlations}
\begin{figure*}
\includegraphics[width=\linewidth]{fig2.pdf}
\caption{(color online) The dynamics of (a)-(c) TER, (d)-(f) TSR, and (g)-(i) TNR for a qutrit evolving under [(a),(d),(g)] amplitude damping channel, [(b),(e),(h)] phase damping channel, [(c),(f),(i)] depolarizing channel. Time $t$ is measured in units of decay rate $\gamma$. Initially the system is in (solid curve) vacuum state $\rho_1:\ket{0}$ state, balanced superposition (dashed curve) $\rho_2:(\ket{0}+\ket{1}+\ket{2})/\sqrt{3},$ or (dotted curve) maximally mixed state $\rho_3:\openone/3.$ We can see that in all cases $\text{TSR}\geq\text{TNR}$, which demonstrates the hierarchical relation between temporal steering and nonlocality. If the initial state satisfies NSIT condition the standard hierarchy of correlations is conserved, $\text{TER}\geq\text{TSR}$. However, panels (c) and (f) demonstrate that in general even a positive semi-definite state over time can describe a temporal steerable system. Note that in all cases involving depolarizing channel we can observe the sudden death of temporal entanglement.}
\label{fig:hierarchy_pdm}
\end{figure*}
\par Hierarchical relations between two-time correlations in systems of arbitrary dimension, obeying NSIT condition, can be proven in the same manner as in two-level case \cite{PhysRevA.98.022104}. Let us start with a state that satisfies Eq.~(\ref{eq:NSIT}), i.e., $\rho_A=\openone/d$. To show the hierarchy, let us recall the scenario demonstrating temporal steering. At time $t_A$ Alice performs a measurement represented by a set of POVMs $\{M_{a\vert x}\}_{x,a}$ on a state $\rho_A$ and sends to Bob a temporal assemblage $\{\rho_{a\vert x}\}$ given by Eq.~(\ref{eq:conditional_state}). It is easy to see that for maximally mixed initial states both formulations of a state over time (\ref{eq:R_W}, \ref{eq:new_PDO}) reduce to the Choi-Jamiolkowski operator (\ref{eq:E_BA}). Hence, the assemblage can be written in terms of a state over time $R_{AB}$ as
\begin{equation}
\rho_{a\vert x}(t)=\tr_A\left(R_{AB}\sqrt{M_{a\vert x}}\rho_A\sqrt{a\vert x}\right).
\end{equation}
If the state over time is separable, i.e. $R_{AB}=\sum_kp_k\rho_k^A\otimes\rho_k^B,$ we can express $\rho_{a\vert x}(t)$ as
\begin{equation}
\rho_{a\vert x}(t)=\sum_k p_k\tr\left(\sqrt{M_{a\vert x}}\rho_k^A\sqrt{M_{a\vert x}}\right)\rho_k^B.
\end{equation}
State sent to Bob can be described as HSM, if the quantum mechanical description can be replaced with local response functions $D(a\vert x,k)$, i.e.,
\begin{equation}
\tr\left(\sqrt{M_{a\vert x}}\rho_k^A\sqrt{M_{a\vert x}}\right)\to D(a\vert x,k).\label{eq:temporalHSM}
\end{equation}
Every separable state over time admits temporal HSM. Thus, inseparability of $R$ is a necessary condition for temporal steering under the NSIT assumption. Note that above considerations are true for all operators of the form (\ref{eq:star}) which satisfy $R_{AB}=\rho_A E_{B\vert A}$ for $\rho_A=\openone/d$.
\par However, numerical calculations show that for initial vacuum state $\vert 0\rangle$ undergoing phase damping a temporally steerable system may not be time-like entangled (see Fig.~\ref{fig:hierarchy_pdm}). Therefore, we conclude that temporal steering implies time-like entanglement, if NSIT condition is satisfied. If the NIST condition is violated, we ce can see in Fig.~\ref{fig:hierarchy_pdm} that the violation of hierarchy is possible.
\par A hierarchical relation between temporal steering and temporal Bell nonlocality holds for an arbitrary quantum system. Equation~(\ref{eq:temporalHSM}) yields the following probability distribution
\begin{equation}
p(a,b\vert x,y)=\sum_k p_k q(a\vert x,k)\tr\left(M_{b\vert y}\rho_k^B\right),\label{eq:temporalLHV}
\end{equation}
where $q(a\vert x,k) = \tr\left(\sqrt{M_{a\vert x}}\varrho_k^A\sqrt{M_{a\vert x}}\right)$ and $M_{b\vert y}$ denotes Bob's POVM at time $t_B$. The above distribution is an LHV model, if probabilities $q(a\vert x,k)$ and $\tr(M_{b\vert y}\varrho_k^B)$ are equivalent to $D(a\vert x,k)$ and $D(b\vert y,k),$ respectively. Therefore, every state over time admitting temporal LHV model is also temporally unsteerable. However, there can exist unsteerable states with no corresponding LHV model (i.e., nonlocal in time). Thus, the set of steerable states is a subset of nonlocal states.
\par To visualize relations between temporal correlations, we simulated the case of single qutrits initially in states $\openone/3$, $\vert 0\rangle$ and $\tfrac{1}{\sqrt{3}}(\vert 0\rangle+\vert 1\rangle+\vert 2\rangle)$, transmitted through standard quantum channels. Note that for initially pure states NSIT condition does not hold, which implies temporal steerability. The robustness-based measures of temporal entanglement, temporal steering, and temporal Bell nonlocality are shown in Fig.~\ref{fig:hierarchy_pdm}. Both TNR and TSR shown in Fig.~\ref{fig:hierarchy_pdm} were calculated for optimal PVMs (projection-valued measures) instead of POVMs, which were too computationally challenging to find in reasonable time. However, for a few points we found that the respective values of robustness for suboptimal POVMs are larger than for optimal PVMs. The depicted quantitative relations $\text{TER}\geq\text{TSR}\geq\text{TNR}$ and $\text{ER}\geq\text{TER}$ in segments (a), (d) and (g) illustrate to the hierarchy of correlations. Note that the same results as for POVMs hold by the same reasoning for PVMs (subsets of POVMs).
\section{Conclusions}\label{sec:conclusions}
In this paper we have provided a comparison of $d$-dimensional temporal correlations. In order to standardize a description of different types of temporal correlations, we have applied a consistent robustness-based measures and introduced \textit{temporal entanglement robustness} and \textit{temporal nonlocality robustness} as temporal counterparts of \textit{entanglement robustness} defined by Vidal and Tarrach in Ref.~\cite{PhysRevA.59.141} and \textit{nonlocality robustness} described in Ref.~\cite{PhysRevA.93.052112}. Additionally, we have proven that TER is a proper causality monotone.
\par We have also extended the notion of PDO, firstly defined in multi-qubit systems~\cite{fitzsimons2015quantum}, to odd-dimensional (\ref{eq:new_PDO}). It seems that thought measurements in the PDO definition have to be non-contextual in order to make the PDO independent of a chosen basis. This might be connected with the formulation of the PDO in terms of decoherence functionals \cite{Zhang2020}. Expectation values of the product of the result of the thought measurements can be expressed in terms of diagonal decoherence functionals, which eliminate contextuality.
\par We have numerically studied the robustness of the
above-mentioned temporal correlations for a 3-level system initially in pure and mixed states with protective measurements. Our results indicate that in contrast to TSR, some level of quantum coherence is required for the initial state to exhibit nonzero TNR. As demonstrated in Fig.~\ref{fig:hierarchy_pdm} the standard hierarchy of correlations is conserved only in these cases, where the initial state satisfies NSIT condition. In the other cases (initially pure states) we can even witness the lack of temporal-entanglement and at the same time equality between nonzero values of TSR and TNR. The latter has deep physical meaning as this is numerical evidence of equivalence of TSR and TNR under certain conditions, possibly this equivalence could be not limited to pure states. It is also relevant to note that our analysis of temporal nonlocality (for zero evolution time) can be conceptually related to quantifying quantum contextuality with contextual fraction~\cite{Abramsky2017}.
\par A universal and sufficient condition for separability of a general $d$-dimensional system is not known, thus, we could only use state-dependent criteria for separability in order to compare space-like and temporal entanglement. The complete characterization of the nontrivial relation between these two types of temporal correlations requires further research. However, we believe that our result will stimulate further research on defining temporal entanglement in a way that will not require the NSIT condition to be satisfied and will conserve the hierarchy of correlations known for spatially correlated systems.
\par In our study of the above-mentioned temporal correlations we have shown that for systems satisfying NSIT condition the hierarchy holds for every definition of a state over time which preserves the classical limit. We have also found an example of system violating both NSIT and the hierarchy between the correlations for two different definitions of a state over time. This answers the question formulated in the remarks of Ref.~\cite{PhysRevA.98.022104}. Relations between quantum temporal correlations do not depend on a particular definition of the star product.
|
\section{Introduction}
Machine learning~(ML) techniques are being adapted in an increasingly large variety of scientific and engineering disciplines.
The field of computational fluid dynamics may be particularly suitable, with various ML applications possible.
The fluid dynamical simulations play an important role in a number of scientific and engineering spheres, especially for instance relativistic codes for high energy physics \cite{Scheid:1974zz,Stoecker:1986ci,Clare:1986qj, Rischke:1995ir,Kolb:2003dz,Petersen:2008dd,Gale:2013da} and astrophysics \cite{Oechslin:2001km,Baiotti:2004wn,Janka:2006fh,Bauswein:2013yna,Baiotti:2016qnr,Hanauske:2017oxo,Most:2018eaw}, but also including molecular biology, aerospace engineering, meteorology, and many other disciplines.
It has been shown that ML can provide valuable inputs for the interpretation and simulation of relativistic fluid dynamical simulations \cite{Pang:2016vdc,Huang:2018fzn,du:2020identifying} and in the field of high energy and nuclear physics in general \cite{Bernhard:2016tnd,Utama:2016tcl,Haake:2017dpr,Zhou:2018ill,Urban:2018tqv,Mori:2017nwj,Shanahan:2018vcv,Tanaka:2017niz}.
Given this, it may be of great benefit to improve various aspects of fluid dynamical applications by means of ML.
Such improvements may comprise speeding up the time-consuming hydrodynamic simulations, as well as other tasks like enforcing constraints on conservation laws, or classification of the numerical hydro solutions.
Perhaps the most natural task is to improve the computational speed of fluid dynamical simulations.
Typically, this procedure demands a trade-off with accuracy and stability, thus, it is crucial to keep these two aspects at an acceptable level.
In many cases involving scientific and industrial applications maintaining the required accuracy level while achieving the necessary speed-up of the simulations is not possible even when the implementation is with parallelization on Graphics Processing Units (GPU’s)~\cite{chen2020gpu}.
The main issue preventing fast direct simulations comes from the Courant convergence criterion~\cite{guenther1996partial}, which limits the maximum size of the time step in numerical simulations.
Other issues include the necessity of a detailed treatment of boundary conditions in Smooth-Particle-Hydro (SPH) based methods~\cite{monaghan2005smoothed,liu2003smoothed}, and in some cases the necessity to use the dynamic viscosity models~\cite{balescu1976equilibrium}.
Numerous efforts have been made to overcome these difficulties, typically sacrificing either the resolution or the robustness of the approach.
Alongside the
purely numerical approaches to solving partial differential equations of fluid dynamics,
several combined methods were also developed,
where the numerical simulations are applied to a coarse-grained system, where the detailed resolution is achieved by data augmentation based on generative adversarial networks~\cite{um2018liquid,xie2018tempoGAN}.
Such
combined methods have shown prominent results in the field of computer gaming and cinema, creating realistic visualizations of fluid dynamics.~\cite{zaspel2011photorealistic,Harris:HarrisGPU}
In the present work we use deep neural network (DNN) as a fast solver of fluid dynamical equations.
This allows to forgo the computationally expensive numerical simulations that would require extensive computing resources.
The DNN approach emulates the solution of partial differential equations by learning the mathematical mapping encoded in the fluid dynamical equations.
As an exploratory study, we consider one-dimensional (1D) fluid dynamics.
The neural network maps the initial state characterized by the fluid dynamical profiles of density, velocity and pressure to their state at a later time moment.
We cover in details the various aspects of the approach such as the determination of the optimal network structure, the optimal choice of the training set, variable vs fixed time step, and the performance of the network to interpolate and extrapolate.
Additionally, we separately incorporate a denoising autoencoder, which is found to increase the network accuracy substantially. The analogy between the mapping performed by the DNN and the mapping performed by the system of fluid dynamic equations makes the current approach rather transparent in its application. It opens the possibility to interpret the performance of the DNN-mapping in fairly simple way and thus, provides greater confidence for the possibility of its implementation in applied science.
The paper is organized as follows. Section~\ref{sec-meth} describes the DNN implementation for 1D fluid dynamics, with a focus on the network structure and the training set choice.
Section~\ref{sec-outside} investigates the ability of the network to
extrapolate the solution beyond the original training set.
Section~\ref{sec-fix-steps} explores the usage of a modified network structure with a fixed time step.
Section~\ref{sec-extend-training} studies the dependence of the performance on the content of the training set.
Section~\ref{sec:noisegate} discusses the application of a noise gate filter to improve the accuracy of the DNN output.
In Sec.~\ref{sec:performance} the achievable speed-ups in the DNN performance over the conventional methods are presented.
Summary in Section~\ref{summary} closes the article.
\section{Methodology}
\label{sec-meth}
We use a DNN to emulate the solution of fluid dynamical equations.
In the present work we study a one dimensional inviscid fluid described by Eulers fluid dynamical equations:
\eq{
&\der{\rho}{t}+u\der{\rho}{x}+\rho\der{u}{x}=0~,\\
&\der{u}{t}+u\der{u}{x}+\frac{1}{\rho}\der{p}{x}=0~,\\
&\der{\varepsilon}{t}+u\der{\varepsilon}{x}+\frac{p}{\rho}\der{u}{x}=0~.
}
Here $\rho$, $u$, and $p$ are the density, hydrodynamic velocity, and pressure, respectively. $\varepsilon$, $v$, $x$, and $t$ are the internal specific energy, specific volume, spatial coordinate, and time.
While the Eulerian case is chosen here for simplicity, conceptually the method can also be applied in more elaborate cases.
The neural network maps the profiles of the hydrodynamic velocity, density and pressure fields from an initial time moment $t_0$ to a later time $t>t_0$:
\eq{
\rho(x,t_0),u(x,t_0),p(x,t_0)\rightarrow \rho(x,t),u(x,t),p(x,t)~.
}
The fluid dynamical equations are supplemented with a polytropic equation of state, $\ p {V^\gamma } = const$, where $p,V$ are the corresponding pressure and volume
values and $\gamma$ is the isentropic expansion factor.
The calculations are performed at a constant temperature.
Knowing $\rho(x,t)$, $u(x,t)$ and for a fixed equation of state, it is possible to exclude $p(x,t)$ from the system of dynamical equations and without additionally supplementing the equation of state (EoS) obtain all three profiles.
However,
supplying all three $\rho(x,t_0)$, $u(x,t_0)$, $p(x,t_0)$ quantities in the input layer
opens
the possibility
to crosscheck
the quality of the network performance by ensuring that the profiles of pressure, velocity and density produced by the network satisfy the EoS.
Therefore, the DNN performs a mapping of all three fields, $\rho(x,t)$, $u(x,t)$, $p(x,t)$.
While in the present study we work with the polytropic EoS, our method can be applied to other equations of state as well, for instance van der Waals~\cite{hansen1990theory}, quantum van der Waals~\cite{Vovchenko:2015vxa}, or Sugie-Lu~\cite{adachi1986development} models.
Along with the consistency with the EoS, other constraints may be introduced to improve the performance of the DNN.
The application of such constraints for the DNN
fluid dynamics simulations
will be studied
in a future paper.
Here we use periodic boundary conditions for all the training samples.
Other choices of boundary conditions are also possible, for instance hard walls.
A different choice of boundary conditions will require re-training of the network.
In choosing the network structure we follow the recommendations of Ref.~\cite{muller2001introduction}, which suggest that the best informational capacity of the model is achieved when the chosen network structure is a little over-descriptive.
Thus, we are using a network with 3 hidden layers and one 10\% dropout layer.
The input layer contains 151 neurons: 50 for each of the $\rho(x,t)$, $u(x,t)$, $p(x,t)$ dependencies and one for $\delta t$ time shift parameter. The output layer contains 150 neurons: 50 for each of the $\rho(x,t+\delta t)$, $u(x,t+\delta t)$, $p(x,t+\delta t)$ dependencies.
The detailed description of the network structure and training process is presented in Appendix~\ref{app-a}.
The structure and content of the training set appear to be amongst the most important factors influencing the quality of the performance of the present approach.
The main challenge of the training set preparation is the particular choice of the curves included in the training set.
The structure and content of the training set have a large influence on the performance of the DNN.
It is important that the training set covers a broad variety of different profiles of density, velocity and pressure.
To achieve the goal we consider three profile functions characterizing the density and velocity profiles,
\eq{\label{f1}
{f_1(x)}&={a_1}{e^{{b_1}{x^2}}}, \\\label{f2}
{f_2(x)}&={a_2}\sin \left( {{b_2}x} \right) +{c_2}, \\\label{f3}
{f_3(x)}&={a_3}{\left| {{x}} \right|}^{\frac{1}{{{b_3}}}}{\mathop{\rm sgn}} \left( x \right) + {c_3},
}
with randomly varied values of the coefficients ${a_i}$, ${b_i}$, ${c_i}$.
Such a set of functions allows to cover a broad variety of possible input scenarios.
It also resembles various general characteristics peculiar to the input curves of fluid dynamics.
Each of the functions $f_1,f_2,f_3$ is chosen randomly to characterize density or velocity profiles\footnote{The pressure profile is evaluated from $\rho(x,t)$ via the equation of state.}
The considered ranges for the randomly
selected values of coefficients are the following:
\eq{\nonumber
{a_1}&\in \left[ { - 5,5} \right],~{b_1} \in \left[ { - 500,0} \right],~{a_2}\in \left[ {0,10} \right],~{b_2} \in \left[ { - 15,15} \right],\\
{c_2} &\in \left[ { - \pi ,\pi } \right],
{a_3} \in \left[ { - 10,10} \right],~{b_3} \in \left[ {1,10} \right],~{c_3} \in \left[ {0,10} \right].\label{ranges}
}
Note that negative values of $f_1,f_2,f_3$ are allowed only for the velocity profiles, but not for density.
Other choices of initial profile functions are possible. These can, for instance, be based on systems of orthogonal functions like Hermite polynomials, Bessel functions, and the various Fourier series~\cite{rudin1976principles,andrews1999special}, depending on the geometric symmetries of the system.
We leave these choices for future studies.
The DNN was trained on a set of $2.5\times 10^5$ samples. Each sample was generated via the following procedure:
\begin{enumerate}
\item
The initial profiles of density and velocity at $t = t_0$ are generated by randomly choosing one of the curve families for each of the profiles independently in Eqs.~(\ref{f1})-(\ref{f3}). The coefficients defining the curve from the chosen family are being randomly generated within the defined boundaries [Eq.~(\ref{ranges})]. For the density profiles the values of the coefficients are constrained such that to ensure the positivity of the density profile.
\item For each pair of the initial density and velocity profiles the pressure profile is computed using the EoS.
\item The Euler equations are solved numerically to obtain the density, pressure and velocity profiles are at a future time moment $t = t_0 + \delta t$, where $\delta t$ at this particular case is chosen to be equal to 1.
\item The combination of initial states at $t_0$ and final states at $t_0+\delta t$ are organized into training pairs
\eq{
\{\delta t;\rho_0(x),u_0(x),p_0(x)|\rho_{\delta t}(x),u_{\delta t}(x),p_{\delta t}(x)\}
}
and incorporated into the training set.
\end{enumerate}
The DNN training procedure consists of $23$ training epochs, where the saturation of training and validation error is observed.
The resulting training set accuracy level, quantified by the $R^2$ metric, is about 89\%.
We explored also a simpler DNN structure consisting of only a single hidden layer, in that case the DNN performance was inferior, with the accuracy level not exceeding 83\%.
Thus, the structure with 3 hidden layers and a 10\% dropout layer was found to be reasonable for the task at hand.
\begin{figure*}
\includegraphics[width=1\textwidth]{Fig1_merged_mod.pdf}
\caption{\label{test}
Comparison of the DNN performance (red solid lines) with the analytical solutions of the numerical solver (blue dashed lines) for the following Riemann Problem Test cases: (top) Sod Shock Tube, (middle) collision, and (bottom) expansion at $t = 2$. Here $\rho(x)$, $u(x)$, $p(x)$ are the dimensionless density, velocity and pressure dependencies correspondingly.
}
\end{figure*}
In addition to the training set accuracy, we also subjected the network to the Riemann Problem~\cite{menikoff1989riemann}, which is a classical test problem for evaluating the accuracy and stability of a hydro solver.
For this purpose the training set has been extended to incorporate a family of profiles that are similar to the Riemann Problem's initial conditions.
The comparison of the DNN generated profiles with an accurate numerical solution corresponding to the classical cases of the Sod shock tube Riemann case~\cite{monaghan1983shock}, a collision as well as expansion scenario is depicted in top, middle, and bottom panels of Fig.~\ref{test}, respectively.
While a few percent level deviations from the numerical solution are visible, the network captures accurately all the qualitative features of the Riemann Problem solutions.
\begin{figure}[h!]
\includegraphics[width=0.49\textwidth]{Extrapolation_Regions.pdf}
\caption{\label{extrapolation_reg}
Regions in the parameter space of the initial hydro profiles and regions along the spatial and time axes in which the network's extrapolation ability was studied. The regions in which the network demonstrated positive performance are marked with the check mark. The regions, where the results are negative are marked with the red cross. The performance of the network of the present structure is still acceptable just outside the training set. But shortly after the end of the training set instabilities start to emerge. This behavior is demonstrated with the corresponding gradient.
}
\end{figure}
\begin{figure}[h!]
\includegraphics[width=0.4\textwidth]{4_1.pdf}
\includegraphics[width=0.4\textwidth]{Fig4_2.pdf}
\caption{\label{initial}
Initial density and pressure profiles, which are contained within the interval, that was removed from the complete test set in order to test the network's capacity to interpolate. The curves from the extirpated interval are used as the input for the network with the diminished training set.
}
\end{figure}
\begin{figure}[h!]
\includegraphics[width=0.49\textwidth]{5.pdf}
\caption{\label{performance}
Comparison of the performance of the two neural networks at $t = 2.5$ -- the one trained on the complete training set, containing the initial density and pressure profiles shown in Fig.~\ref{initial} (red solid lines) and another trained on the diminished training set, which does not contain the aforementioned subset of curves (green dashed lines). The initial density and pressure profiles at $t = 0$ used for this test case are shown in Fig.~\ref{initial}.
}
\end{figure}
\begin{figure*}
\includegraphics[width=0.86\textwidth]{Fig6_merged.pdf}
\caption{\label{dvp3}
Comparison of the DNN performance (red solid lines) with the analytical solutions of the numerical solver (blue dashed lines) at time $t = 5.1$ (top) and $t = 6$ (bottom) which is beyond the training region $t \in [0,5]$.
The performance of the DNN with a variable time-shift parameter deteriorates right beyond the training interval.
}
\end{figure*}
\section{DNN Performance outside the training set}
\label{sec-outside}
We have shown that the network is able to approximate the solution with high accuracy when the solution is within the boundaries of the training interval.
An important question now is: how good are the network capabilities to extrapolate outside the training set?
Here we study three different aspects of the network's extrapolation ability:
\begin{enumerate}
\item The performance on hydro profiles that are not included in the training set
\item The ability to extrapolate to time intervals not included in the training set
\end{enumerate}
These two cases are schematically depicted in the top and bottom panels of Fig.~\ref{extrapolation_reg}.
First, we explore the extrapolatory ability of the network to cover profiles not included in the training~(region "1" in Fig.~\ref{extrapolation_reg}).
This is achieved by excluding a certain subset of initial $\rho,u,p$ curves~[Eqs.~(\ref{f1})-(\ref{f3})] from the training set, which correspond to a particular interval of the parameters ${a_i}$, ${b_i}$, ${c_i}$.
Namely, we omit density profiles corresponding to parameter intervals $a\in[5,8]$, $b\in[-7,-8]$ for the $f_1$ family of initial curves and velocity profiles corresponding to $a\in[2,3]$, $b\in[-4,-5]$, $c\in[0.5,1.5]$ for the same $f_1$ family of initial curves.
We then analyze the performance of such a network on an input profile that was omitted from the training set.
This procedure is illustrated in Fig.~\ref{initial}.
To assess the extrapolation performance of the network, we compare its output to the one produced by a network based on the full training set.
The comparison, presented in Fig.~\ref{performance}, indicates that the network performance is virtually unaffected by excluding a broad set of profiles from the training data.\footnote{Note, that this interpolation test differs from the classical “test-set” trial of the network performance, when just some random samples are excluded from the training set and are moved to the test set. Given the vast initial training set of $2.4~10^5$ curves,
even with multiple of the curves being
excluded,
there will still remain plenty of quite close curves.
}.
Thus, the network demonstrates reasonable ability to extrapolate from the different initial profiles.
Next, we study the extrapolation ability along the time axis.
Namely, we analyze the network output for time moments exceeding the training interval of $t \in [0,5]$~ (region "2" in Fig.~\ref{extrapolation_reg}).
We take two values: (i) $t = 5.1$ which is slightly outside the training interval and (ii) $t = 6$ which is further away.
The results for these two cases are depicted in top and bottom panels of Fig.~\ref{dvp3}, respectively.
The network performance is still relatively decent in the first case just outside the training interval, although the developing instabilities are visible.
For $t = 6$ the network output is quite off from the reference solution, at best capturing only the general qualitative features of the solution.
This indicates that the extrapolation ability of the present network with a variable time step is weak.
For this reason we explore a different treatment of the time step.
It is also possible to explore other scenarios of interpolation. For instance, when a certain period of time-evolution of the system was omitted and must be reconstructed.
\section{Fixing the time step}
\label{sec-fix-steps}
\begin{figure*}
\includegraphics[width=0.86\textwidth]{Fig7_mod.pdf}
\caption{\label{dvp}
Comparison of the altered DNN performance trained on the original training set (red solid lines) and extended training set (blue solid lines) with the analytical solutions of the numerical solver (green dashed lines)
for the Sod shock tube problem at time parameter values $t = \delta t \equiv 1$ (top), $t = 5$ (middle), and $t = 10$ (bottom).
}
\end{figure*}
As shown in the previous section, the DNN performance along the time axis becomes unsatisfactory as one goes beyond the training range.
This is most likely a reflection of an unbalanced structure of neurons in the network used in Secs.~\ref{sec-meth} and \ref{sec-outside}.
Whereas $50 \times 3$ parameters were used for the spatial variable, only a single neuron in the input layer was dedicated to the time variable.
This one neuron alone defines the whole shape of the output curves, as each different time-shift $\delta t$ corresponds to different $\rho, u, p $ mappings.
The output $\rho, u, p $ profiles strongly depend on the time-shift $\delta t$ parameter.
This leads to an
asymmetry in the distribution of weights and biases, as this single, time shift dedicated, neuron in the input layer
is significantly
more important than any other.
The asymmetry in the “significance” leads to non-uniform distribution of weights and biases encountered by the network in the input layer, consequently affecting its ability to extrapolate.
To mitigate the situation we modify the network structure in this section.
We do not specify a variable time-shift and remove the corresponding neuron from the DNN.
Instead we specify a fixed
time shift between the initial and the final states:
$\rho \left( {x,t} \right),u\left( {x,t} \right),v\left( {x,t} \right) \to \rho \left( {x,t + \delta t} \right),u\left( {x,t + \delta t} \right),v\left( {x,t + \delta t} \right)$.
In this way we eliminate the non-uniformity in the distribution of values in the input layer.
This also reduces the required informational capacity, as the network only needs to learn the mapping for a single fixed time step rather than for a multitude of steps.
The training of the new network is performed similarly to the previous one.
The training data consists of the $\rho, u, p $ profiles mapped between the initial time moment $t_0$ and a subsequent time moment $t + \delta t$ with $\delta t = 1$.
These profiles are combined into training pairs
\eq{
\{\rho_0(x),u_0(x),p_0(x)|\rho_{0.1}(x),u_{0.1}(x),p_{0.1}(x)\}
}
and the training is performed.
The updated network allows to perform mapping $f\left( {x,{t_0}} \right) \to f\left( {x,{t_0} + \delta T} \right)$ across large time steps $\delta T$, provided that it can be presented as a multiple of $\delta t$, i.e. $\delta T = n\delta t$, $n \in N$.
This is achieved by successively applying the network $n$ times, by using the output from the previous step as an input to the next one.
To test the performance of the new network along the time axis we will subject it to the Sod shock tube test.
More specifically, we study the performance in a time interval $t \in [0,10]$ by consequently performing 10 steps in increments of $\delta t = 1$ as described above.
The performance of the altered network is presented in Fig.~\ref{dvp} (solid red lines) and compared to the numerical solution~(dashed blue lines).
The top panel corresponds to $t = 1$, this short time interval coincides with the training interval, thus, it illustrates the interpolating capability of the network.
The network reproduces the numerical solution with high accuracy, thus its interpolating ability is good.
The middle and bottom panels of Fig.~\ref{dvp} correspond to $t = 5$ and $t = 10$, respectively.
These intervals are outside the training range, thus, the extrapolation capability of the network is probed.
The updated network clearly
shows much better capacity to
extrapolate
in comparison with the previous method employed in Secs.~\ref{sec-meth} and \ref{sec-outside}.
However, already for $t = 5$ one can see deviations of the network generated profiles from the numerical solver, especially for the velocity profile.
Thus, further improvements are required.
\section{Extending the training set}
\label{sec-extend-training}
The DNN performance can be further improved by modifications to the training procedure.
Here we explore two modifications: (i) an extended set of training profiles and (ii) training the network across several time steps.
The first one addresses a possible gap in the set of profiles that the network is able to map accurately while the second modification is designed to improve the network extrapolation performance across large time periods.
\subsection{Extending the training set with harmonic functions}
\label{sec-extend-1}
The results of the previous section show considerable error accumulation as the number of time steps is increased, in particular in the velocity profile.
We surmise that the error accumulation energy due to an incomplete range of velocity profiles covered in the training set, namely that the $f_{1,2,3}$ curve families used in training are not close enough to accurately capture the features of the velocity profile in the Sod shock tube test.
Therefore, here we extend the training set with an additional family of curves, corresponds to a linear combination of harmonic functions:
\eq{\label{f4}
f_4 = \sum\limits_{i = 1}^n {{a_i}\sin \left( {{b_i}x + {c_i}} \right)}.
}
Here $a_i$, $b_i$, $c_i$ are random parameters chosen uniformly
for the following ranges:
\eq{\nonumber
{a_i}\in \left[ { 0,10} \right],~{b_i} \in \left[ { - 15,15 } \right],~{c_i}\in \left[ { - \pi ,\pi } \right],~{n}=1,...,10 .}
With this extension the training set allows a significantly more accurate coverage of discontinuous profile functions, as illustrated in Fig.~\ref{rangeextension} for the case of a step function and similar discontinuous profiles.
We generated $2.5\times 10^5$ samples for training and $8\times 10^4$ for validation of the modified network.
For this network structure without the time-shift parameter
in the input layer with the periodic boundary conditions
within $24$ training epochs it was possible to obtain an accuracy level of $92\%$.
The performance of the modified DNN for the Sod shock tube test is shown in Fig.~\ref{dvp} by the solid green lines.
One sees that the extension of the training set for the velocity significantly improves the performance of the network.
The improved performance is particularly
evident at
$t=10$, which corresponds to consecutive 10 network iterations.
\begin{figure}[h!]
\includegraphics[width=.49\textwidth]{Fig8.pdf}
\caption{\label{rangeextension}
An illustration of how the extended parameter range allows to better approximate discontinuous profiles.
Here the blue and green curves show the best approximation achievable by, respectively, the original and extended parameter sets for a step function $f(x) = 10 \, \operatorname{\theta}(x)$.
}
\end{figure}
\subsection{Training over multiple time steps}
Another extension of the training set is achieved by utilizing multiple time steps in the training procedure.
This is achieved in the following way.
We take the initial profiles from the extended data-set $A$ described in Sec.~\ref{sec-extend-1}.
Then, for each curve the numerical solver is run up to a time moment $t = 6$.
We then take the profiles at time moments $t = 0$, 1, 2, 3, 4, 5, 6, and include them into the training set.
In this way the new extended training set ($C$) contains not only the set of profiles $f_1,...,f_4$ and their mapping to a subsequent $t+\delta t$ time moment, but also that for the profiles obtained from each of the steps from one ($t+\delta t$) to six ($t+6*\delta t$).
The training utilizing this extended set is then done from scratch.
It is observed that the extended training set $C$ leads to a significantly improved performance of the network across long time intervals.
This is illustrated in Fig.~\ref{r2_metric}, which shows the behavior of the $R^2$ error metric for the different studied training sets across a long time interval extending to $t = 200$.
The error increases substantially after a certain number of network iterations performed.
This behavior is attributed to encountering new curve configurations, that are not anymore covered accurately by the standard training set ($A$).
The training set $B$ incorporates a single additional time step $t+\delta t$.
This allows to preserve a satisfactory network performance to larger times, however the accuracy drop substantially at $t ~ 30-50$.
The full extended training set "C", on the other hand, preserves a relatively constant level of the $R^2$ error metric across long time intervals.
Figure~\ref{further_set_extension} compares the velocity profile output of the network based on training set $C$ at $t = 30$ with the numerical solution.
\begin{figure}
\includegraphics[width=.49\textwidth]{U_comparisonNoiseGate.pdf}
\caption{\label{r2_metric}
The dependence of the $R^2$ metric between the solver and network performance on the time (number of network iterations) for the velocity profile for different training sets used. Set "C" includes 6 consequent generations of curves computed at the corresponding time values from ($t_0$) to ($t_0+6*\delta t$). Set "B" includes only one additional time step($t_0+\delta t$). Set "A" presents only the modified network structure with no time-shift parameter in the input layer. Additionally the effect of the noise gate filter application is shown. Note that the alterations in the structure of training sets alter the content of the set but not the total number of curves present in it.
}
\end{figure}
\begin{figure}
\includegraphics[width=.4\textwidth]{FIg9.pdf}
\caption{\label{further_set_extension}
The effect of further extension of the training set at t=30 (after 30 network iterations) on the velocity dependency.}
\end{figure}
A substantial performance increase due to the extension of the training set with profiles from multiple time steps is evident.
This is due to the fact, that the training set $A$,
which did not contain training profiles after multiple time steps,
remained incomplete,
i.e. the set of trained configurations was not sufficient for
the network to perform the mapping at long time intervals.
The result effectively underlines that the particular choice of the initial training set appears to be one of the most important factors influencing the capability of the method.
This choice will generally depend on multiple other factors, including the geometry of the system, equation of state and boundary conditions.
In this work we use the periodic boundary conditions, if these are changed to other boundary conditions such as open or hard-walls, the network should be re-trained.
It should also be emphasized that here we trained the network for a fixed value of the time shift, thus the network is only capable of performing mappings in the units of this fixed time shift.
If a different value of the time shift is desired, this will also require the re-training of the network.
\section{noise gate filter}
\label{sec:noisegate}
As discussed in the previous section, the performance of the network drops after a large number of time steps even in the case of the most complete training set "C".
This is reflected by the appearance of noise in the hydrodynamic fields~(Fig.~\ref{further_set_extension}).
Although the level of noise remains fairly stable, it does notably affect the performance of the method, the $R^2$ accuracy metric reaching about $0.83$ at large times, as seen in~Fig.~\ref{r2_metric}.
Here we explore a possibility to alleviate this issue by implementing a noise gate filter.
The benefit of knowing the numerical solution along with the prediction of the neural network allows one to train the separate convolutional autoencoder structure, which performs the role of a noise gate filter.
This noise gate filter is incorporated in the following way.
First, we generate multiple clean training curves belonging to the aforementioned 6 iterations of evolution from $t_0$ to $t_0+i*\delta t$, where ${i} \in \left[ { 0,6} \right]$.
Then, we prepare the training input-output pairs by adding (i) random Gaussian noise of various types, (ii) uniformly distributed noise and (iii) noiseless samples.
The latter ensures that the autoencoder learns not to filter the already clean samples.
$10^5$ training pairs are prepared in this way, then this autoencoder structure is applied to the output obtained from the neural network trained on the extended set "C" in the previous section.
The resulting effect of the autoencoder on the performance of the network trained on the extended set "C" in the previous section is depicted in Fig.~\ref{r2_metric} by the dashed green line.
The application of the noise gate filter considerably improves the network accuracy at large times $t > 100$.
The main benefit of the noise gate filter is to smoothen the hydro profiles.
This is illustrated in Fig.~\ref{denoising_process}, which depicts a typical case of a network-produced velocity profile with and without the application of the noise gate filter compared to the reference numerical solution.
While the output of the regular network is able to capture the overall numerical profile, it does exhibit characteristic noise pattern.
The noise gate filter is able to largely remove this noise and resemble the reference profile even more accurately.
The application of the noise gate thus allows to substantially increase the accuracy of the neural network based model.
\section{performance}
\label{sec:performance}
The main purpose of applying the DNN to fluid dynamics is to improve the speed of the associated computations over the conventional methods while preserving an acceptable level of accuracy.
This can be quantified by a speed-up factor $s$.
Its value will depend on the particular task under consideration.
Let us consider the following generic problem: the fluid dynamical simulation is run over a time period $\delta t_e$, with an intermediate output at $n_{\rm int}$ time steps.
For instance, such a setup can occur in a case of real-time simulation of the flow for visualization purposes, where at least 24 frames per second are required to render.
The maximum speed of the conventional method can be achieved by maximizing the value of a single time-step, which will minimize the total number of steps needed to be performed.
This maximum value, $\delta t_{\rm CFL}$, is constrained by the Courant–Friedrichs–Lewy~(CFL) condition.
The neural network, on the other hand, is not constrained by the CFL condition.
Instead, the time step per single DNN iteration is $\delta t_e / n_{\rm int}$, reflecting the fact that one needs to output $n_{\rm int}$ intermediate time steps.
The speed-up factor thus reads
\eq{
s = \frac{{\delta {t_{\rm{e}}} / n_{{\rm{int}}}}}{{\delta {t_{{\rm{CFL}}}}}} \, \frac{{{\tau _{\rm num}}}}{{{\tau _{\rm net}}}}~.
}
Here $\tau_{\rm num}$ is the time duration of a single computational step that computes the evolution from $t^*$ to $t^*+\delta t_{\rm CFL}$ and $\tau_{\rm net}$ is the time duration of a single computational step of neural network application.
The performance increase due to the neural network comes from two sources.
First, the neural network application is not constrained by the CFL criterion, thus $\frac{{\delta {t_{\rm{e}}} / n_{{\rm{int}}}}}{{\delta {t_{{\rm{CFL}}}}}} > 1$ is possible.
Second, the neural network can perform a single time step considerably faster than the conventional methods.
For the applications considered in the present work, we have ${\tau_{\rm net}}/{\tau_{\rm num}} \sim 10^2$.
The values of $\frac{{\delta {t_{\rm{e}}} / n_{{\rm{int}}}}}{{\delta {t_{{\rm{CFL}}}}}}$ were observed to vary in range $15-50$, depending on the exact network configuration employed.
The speedup factors listed here correspond to both the numerical method and the network running on CPU.
Additional speed-up up to a few orders magnitude can be obtained by employing hardware tailored for DNN-related computations, such as NVIDIA tensor core GPUs.
\begin{figure*}
\includegraphics[width=0.8\textwidth]{fig11.pdf}
\caption{\label{denoising_process}
The effect of the application of the noise gate layer on the velocity profile. The velocity profile containing noise is presented in the middle panel. The result of the noise - gate layer application is presented in the right panel and the conventional reference solution from the numerical scheme is presented at the left panel.
}
\end{figure*}
\section{summary}
\label{summary}
We investigated the possibility of a deep neural network based simulation of one-dimensional fluid dynamics.
The DNN performance vastly depends on the network structure and training set, an acceptable accuracy can be obtained with an appropriate choice.
For the network structure, a network with 3 hidden layers and one dropout layer was found to be a decent choice, yielding a 90\% accuracy on the training set that consists of a vast amount of fluid dynamical profiles.
Regarding the network's ability to extrapolate along the time axis, it shows a limited ability when the time step is varied and part of the network input.
On the other hand, a decent performance is achieved when the time step is fixed and the network is applied successively by using the output from the previous time step as input to the next step.
The performance is further optimized by extending the training set with profiles obtained numerically after multiple time steps.
We also find that the noise in the profiles output by the network at future times can be to a large extend eliminated by the application of a noise gate filter.
The time step of the DNN simulations is not restricted by the Courant criterion, which is a major source of the speed-up of DNN computations over the conventional numerical methods.
Also, the time spent per single time step in the DNN is observed to be about two orders of magnitude smaller than that for the conventional computations.
This performance can be further improved substantially by utilizing dedicated hardware tailored for DNN applications, such as modern GPU's with tensor cores.
While the presented studies indicate that a careful choice of the network structure and the training set is not trivial and crucial to achieve an acceptable performance, it is reasonable to expect even greater benefits of the DNN in higher dimensions, where the conventional methods become very time consuming.
\section*{Acknowledgments}
K.T. and R.P. express gratitude to the Stiftung Polytechnische Gesellschaft for supporting the research project.
K.Z. and J.S. thank the AI grant at FIAS by SAMSON AG and the BMBF funding through the ErUM-Data project.
R.P. acknowledges the support of the Target Program of Fundamental Research of the Department of Physics and Astronomy of the National Academy of Sciences of Ukraine
(N 0120U100857).
V.V. acknowledges the support through the
Feodor Lynen program of the Alexander von Humboldt
foundation and the U.S. Department of Energy,
Office of Science, Office of Nuclear Physics, under contract number
DE-AC02-05CH11231231. H.St. acknowledges the Walter Greiner Gesellschaft zur F\"orderung der physikalischen Grundlagenforschung e.V. through the Judah M. Eisenberg Laureatus Chair at Goethe Universit\"at Frankfurt am Main. K.Z. also thank NVIDIA Corporation with the generous donation of NVIDIA GPU cards.
|
\section{Introduction}
Let $M$ be a differentiable manifold. An {\it almost complex structure} on $M$ is an endomorphism $f$ of the tangent bundle $TM$ such that $f^2+I=0$, where $I$ denotes the identity endomorphism. It is easy to see that in order to support an almost complex structure, $M$ has to be even-dimensional. The odd-dimensional analogue is an {\it almost contact structure}, i.e.\ an endomorphism $f$ of $TM$ together with a vector $E$ and a 1-form $\eta$ such that $f^2+I=E\otimes \eta$. Yano \cite{Yano63} introduced the notion of {\it $f$-structure} i.e.\ an endomorphism of $TM$ such that $f^3+f=0$, of which almost complex structures and almost contact structures are both particular cases. In a different direction, {\it almost tangent structures} \cites{Eliopoulos65, Lehmann-Lejeune66} are endomorphisms $f$ of $TM$ such that $f^2=0$. Motivated by the quartic case, Goldberg and Yano \cite{GoldbergYano70} (see also \cite{GoldbergPetridis73}) looked at general {\it polynomial structures} i.e.\ endomorphisms $f$ of $TM$ satisfying $P(f)=0$ for some polynomial $P(x)$ with real constant coefficients and such that $I, f,\ldots,f^{\deg(P)-1}$ induce linearly independent endomorphisms on each fiber of $TM$.
Hitchin \cite{Hitchin03} has shown that certain geometric structures arising from the study of the supersymmetric sigma-model can be understood as {\it generalized almost complex structures}, i.e.\ as endomorphisms $\varphi$ of the generalized tangent bundle $\mathbb{T} M= TM\oplus T^*M$ that are skew-symmetric with respect to the tautological inner product and such that $\varphi^2+I=0$. Any classical almost complex structure $f$ gives rise to a generalized almost complex structure by setting $\varphi=f\oplus (-f^{t})$. Just as in the classical case, generalized almost complex structures can only exist on even-dimensional manifolds. {\it Generalized $F$-structures} i.e.\ skew-symmetric endomorphisms $\varphi$ of the generalized tangent bundle such that $\varphi^3+\varphi=0$ were introduced by Vaisman \cite{Vaisman08} as a common extension of the notions of generalized almost complex structure and of classical $f$-structure.
The goal of the present paper is to fill a gap in the literature and study the analogue of classical polynomial structures in the context of generalized geometry i.e.\ skew-symmetric endomorphisms $\varphi$ of the generalized tangent bundle such that $P(\varphi)=0$ for some polynomial $P(x)$ with real constant coefficients. We call these structures {\it generalized polynomial structures}. In addition to generalized almost complex structures and generalized $F$-structures, the quadratic case has recently received some attention with progress being made for {\it generalized almost tangent structures} \cite{BlagaCrasmareanu14}, and more generally for {\it generalized metallic structures} \cite{BlagaNannicini20}. With the exception of the somewhat degenerate case of generalized almost tangent structures, a common feature of the particular examples of generalized polynomial structures studied in the literature up to this point is that the polynomial $P(x)$ has simple roots and thus $\mathbb{T} M\otimes \mathbb C$ splits into the direct sum of the corresponding eigenbundles. In general, this is no longer the case and $\varphi$ may very well be non-diagonalizable.
Our first main observation, Theorem \ref{theorem:40} in Section \ref{sec:3}, is that given a generalized polynomial structure $\varphi$, the complexified generalized tangent bundle decomposes into the direct sum of the generalized eigenbundles $L_\lambda$ of $\varphi$. Moreover, each $L_\lambda$ is an isotropic eigenbundle and for each non-zero $\lambda$ the restriction of the tautological inner product to $E_\lambda=L_\lambda\oplus L_{-\lambda}$ is non-degenerate. In particular, every generalized polynomial structure admits a Jordan-Chevalley decomposition into its semisimple and nilpotent parts, both of which are also generalized polynomial structures. Another important consequence is that every generalized polynomial structure induces a canonical (up to overall shifts and involutions) multigrading on (complexified) differential forms. In the particular case of generalized almost complex structures, we recover the familiar decomposition of the complexified generalized tangent bundle into the $\pm\sqrt{-1}$-eigenbundles and the corresponding grading of complex differential forms \cite{Gualtieri11} (later extended to generalized F-structures in \cite{AldiGrandini17}). Finally, using results by Burgoyne and Cushman \cite{BurgoyneCushman77} in the case where $M$ is a point, we offer a detailed description of the possibly indecomposable blocks that can arise.
Generalized complex structures are characterized \cites{Gualtieri11, Cavalcanti06} among all generalized almost complex structures in three equivalent ways: 1) by the vanishing of their Courant-Nijenhuis torsion, 2) by the closure of their eigenbundles with respect to the Dorfman bracket, and 3) the decomposition $d=\partial + \overline \partial$, with $\partial$ (respectively $\overline \partial$) of degree $1$ (respectively $-1$) with respect to the induced grading on complex differential forms. In Theorem \ref{theorem:56} of Section \ref{sec:4}, our second main result, we extend this characterization to arbitrary generalized polynomial structures. Naively, based on the example of generalized complex structures and on the definition of integrability for classical polynomial structures \cites{GoldbergPetridis73,IshiharaYano65, Vanzura76}, one might expect that imposing the vanishing of the Courant-Nijenhuis torsion leads to an interesting class of generalized polynomial structures. Unfortunately, as we point out in Section \ref{sec:4} this condition is too strong as it forces the generalized polynomial structure to have at most two distinct eigenvalues. This is in sharp contrast with the classical case, where the vanishing of the Nijenhuis torsion imposes no limit on the number of distinct eigenvalues of a polynomial structure. As it turns out, this discrepancy is closely related to the fact that if the restriction of the tautological inner product to a proper subbundle $E\subseteq \mathbb{T} M\otimes \mathbb C$ is non-degenerate, then $E$ cannot be closed under the Dorfman bracket. Hence, the natural extension of condition 2) above is the requirement that if $\lambda\neq - \mu$ then the Dorfman bracket ${[}\!{[} L_\lambda, L_\mu {]}\!{]}$ is contained in $L_\lambda+L_\mu$. In particular no condition is imposed on the Dorfman bracket of two sections of $L_0$, the generalized eigenbundle corresponding to the eigenvalue $0$. Correspondingly, the natural extension of condition 1) to arbitrary generalized polynomial structures is the vanishing of what we call the {\it shifted Courant-Nijenhuis torsion} of the semisimple part. The appearance of the shifted Courant-Nijenhuis is not entirely unexpected because its vanishing precisely encodes the notion of {\it strong integrability} of generalized F-structures \cites{PoonWade11, AldiGrandini15}. The fact that only the semisimple part plays a role in Theorem \ref{theorem:56} is also not entirely unexpected since the generalized eigenbundles and the multigrading of differential forms that they induce only depend on the semisimple part.
Having obtained in Section \ref{sec:4} a satisfactory understanding of how the semisimple part of a generalized polynomial structure interacts with the de Rham operator, we would like to grasp the geometric meaning of the nilpotent part. We tackle this task in Section \ref{sec:5}. Our starting point is the observation \cites{Guttenberg07, Tomasiello08} that the integrability of a generalized almost complex structure $\varphi$ is equivalent to the vanishing of the derived bracket with respect to the operator $({\rm ad}_{\widetilde \varphi}^2+I)(d)$, where $\widetilde \varphi$ is a lift of $\varphi$ to the space of $\mathbb R$-linear operators acting on differential forms and ${\rm ad}_{\widetilde \varphi}$ denotes the corresponding adjoint action (with respect to the natural graded commutator of operators acting on forms). As pointed out in \cite{AldiGrandini17} a similar result holds for a generalized F-structure $\varphi$, whose strong integrability is equivalent to the vanishing of the derived bracket with respect to the operator $({\rm ad}_{\widetilde \varphi}^3+{\rm ad}_{\widetilde\varphi})(d)$. With this motivation in mind, in the present paper we introduce the {\it minimal torsion} $\mathcal M_\varphi$ of a generalized polynomial structure $\varphi$ with minimal polynomial $P(x)$ as the derived bracket for the operator $P({\rm ad}_{\widetilde{\varphi}})(d)$. The terminology is justified by the fact that if $P(x)$ is quadratic (including the case of generalized almost complex structures discussed above) then $\mathcal M_\varphi=2\mathcal T_\varphi$, where $\mathcal T_\varphi$ is the usual Courant-Nijenhuis torsion of $\varphi$. Similarly, if $P(x)$ is cubic then $\mathcal M_\varphi = -3 \mathcal S_\varphi$, where $S_\varphi$ is the shifted Courant-Nijenhuis torsion introduced in Section \ref{sec:4}. Our first result in this direction, Theorem \ref{theorem:57}, explicitly relates the minimal torsion of a generalized polynomial structure to what we call {\it (shifted) higher Courant-Nijenhuis torsions} (i.e.\ the natural extension of higher Nijenhuis and Haantjes torsions \cites{Kosmann-Schwarzbach19, TempestaTondo18a, TempestaTondo21} to sections of the generalized tangent bundle). Our second result on minimal torsion, Theorem \ref{theorem:83}, characterizes the vanishing of the minimal torsion in terms of a suitable decomposition of the de Rham operator.
We say that a generalized polynomial structure is {\it minimal} if its minimal torsion vanishes. Our perspective is that minimality provides a useful notion of compatibility with the structure of Courant algebroid on the generalized tangent bundle that involves both the semisimple and the nilpotent parts. A natural question at this point is how minimality relates to the compatibility conditions, which do not involve the nilpotent part, discussed in Section \ref{sec:4}. As it turns out, for a generic generalized polynomial structure, minimality implies that the semisimple part satisfies the equivalent conditions of Theorem \ref{theorem:56}. As a consequence, we have that, generically, the decomposition of the de Rham operator provided by Theorem \ref{theorem:83} coincides with the decomposition into multigraded components guaranteed by Theorem \ref{theorem:56}. The non-generic case is characterized by a sort of ``resonance'' condition in which an eigenvalue can be written as a non-trivial sum of three other eigenvalues. This phenomenon is rather unexpected and it would be interesting to understand it better. Here we limit ourselves to show, by means of an example, that this condition cannot be entirely removed.
Throughout the paper, we place particular emphasis on the geometrically relevant \cite{GoldbergYano70} quartic case. In particular, we show that minimal generalized polynomial structures with minimal polynomial $(x^2+1)^2$ are nothing but commuting pairs consisting of a generalized complex structure and a weak generalized tangent structure. Furthermore, minimal generalized polynomial structures with minimal polynomial $x^2(x^2+1)$ are nothing but commuting pairs consisting of a strongly integrable generalized F-structure and a weak generalized tangent structure with vanishing bivariate Courant-Nijenhuis torsion.
In Section \ref{sec:6} we further illustrate the main constructions of the paper with concrete examples of polynomial structures on low-dimensional Lie groups. We treat these cases by constructing a basis of the underlying Lie algebra that is adapted to the block decomposition of a given generalized polynomial structure. This allows us to explicitly compute the higher Courant-Nijenhuis torsions and their shifted counterparts. While our focus is on particular examples, our calculations are algorithmic in nature and could conceivably be carried out in a systematic fashion with the aid of suitable computer algebra systems. We hope to come back to this point in future work.
\vskip.3in
{\bf Acknowledgements:} We would like to thank the anonymous referees for the many useful suggestions.
\section{Preliminaries}
The goal of this section is to make the paper reasonably self-contained by providing the necessary background on the geometry of the generalized tangent bundle of a smooth manifold. We find it convenient to systematically employ the language, developed in \cite{AldiGrandini17} (see also \cite{Buttin74} for the classical case), of $\mathbb R$-linear operators acting on the sheaf of differential forms. The well-known concept of derived bracket (\cites{Vinogradov90, Kosmann-Schwarzbach04}) is of crucial importance in the reminder of the paper.
\subsection{Operators acting on differential forms}
Let $M$ be a smooth manifold of dimension $n$. We denote by $\Omega_M$ the sheaf of differential forms on $M$ and by $\mathcal E_M$ the sheaf of algebras of $\mathbb R$-linear endomorphisms of $\Omega_M$. The usual grading $\Omega_M=\bigoplus_{i=0}^n \Omega_M^i$ induces a compatible grading on $\mathcal E_M$. Sections of $\mathcal E_M$ are endowed with the natural composition product, giving rise to the standard (graded) commutator $[\psi,\varphi]=\psi\varphi-(-1)^{kl}\varphi\psi$. Each section $\zeta \in \mathcal E_M^l$ acts on $\mathcal E_M$ as a left-derivation of degree $l$ via the adjoint action ${\rm ad}_\zeta(\varphi)=[\zeta,\varphi]$.
\begin{example}
By letting differential forms act by left multiplication $\eta \omega=\eta \wedge \omega$ (the wedge notation will be omitted), $\Omega_M$ can be canonically identified with a subsheaf of $\mathcal E_M$ . With this identification in mind, we observe that $[\Omega_M,\Omega_M]=0$.
\end{example}
\begin{example}
The tangent sheaf $\mathcal X_M^1$ is canonically identified with the subsheaf of $\mathcal E_M$ consisting of those local sections $X$ such that $X(1)=0$ and $[X,\omega]=\iota_X \omega$, for any $\omega\in \Omega_M$. Taking linear combinations of compositions of vectors (now identified with sections of $\mathcal E_M$) yields the full sheaf $\mathcal X_M$ of all polyvector fields acting on $M$. It is easy to see that $\mathcal X_M$ and $\mathcal X_M\oplus \Omega_M$ are both closed under the commutator.
\end{example}
\begin{example}
The de Rham differential is a global section of $\mathcal E_M^1$. In our notation, $[d,d]=2d^2=0$, $[X,d]=\mathcal L_X$ for every $X\in TM$, and $[d,\omega]=d\omega$ for every $\omega\in \Omega_M$.
\end{example}
\begin{mydef}
Let $D$ be an $\mathbb R$-linear endomorphism of $\mathcal E_M$, of degree $k$. The {\it derived bracket} associated to $D$ is the bilinear operation
\begin{equation}
{[}\!{[} \psi, \varphi{]}\!{]}_D = (-1)^{kl+1}[D\psi, \varphi]\,,
\end{equation}
for all $\psi\in \mathcal E^l_M$ and $\varphi\in \mathcal E_M$. In particular, if $D={\rm ad}_\delta$, for some $\delta\in \mathcal E_M$ we use the shorthand notation ${[}\!{[} \psi,\varphi {]}\!{]}_\delta$ for ${[}\!{[} \psi,\varphi {]}\!{]}_{{\rm ad}_\delta}=[[\psi,\delta],\varphi]$.
\end{mydef}
\begin{example}
If $\delta=d$, the resulting bracket is known as the {\it Dorfman bracket}, for which we will reserve the special notation ${[}\!{[}\,\, ,\,\, {]}\!{]}$. More generally, if $H$ is a closed 3-form on $M$ then the derived bracket corresponding to $\delta=d+H$ is known as a {\it twisted Dorfman bracket}. For simplicity, in the rest of the paper we mostly focus on the untwisted case.
\end{example}
\subsection{The generalized tangent bundle}
The {\it generalized tangent bundle} of $M$ is the bundle $\mathbb T M$ associated to the sheaf $\mathcal X_M^1\oplus \Omega_M^1$. The {\it tautological inner product} is the symmetric, $\Omega_M^0$-bilinear map $\langle \,\, , \,\,\rangle$ defined by the formula
\begin{equation}
\langle \mathbf{x},\mathbf{y}\rangle = \frac{1}{2} [\mathbf{x},\mathbf{y}]
\end{equation}
for all $\mathbf{x},\mathbf{y} \in \Omega_M$.
\begin{rem}
The generalized tangent bundle is closed under any twisted Dorfman bracket.
\end{rem}
\begin{lem}\label{lemma:7}
Let $\varphi,\delta$ be sections of $\mathcal E_M$, with $\delta$ of odd degree. Then
\begin{enumerate}[1)]
\item $[\varphi,\mathbb T M]\subseteq \Omega_M^0$ if and only if $\varphi\in \Omega_M^0\oplus \mathbb T M$;
\item $[\varphi,\mathbb T M]=0$ if and only if $\varphi\in \Omega_M^0$;
\item ${[}\!{[} \mathbb{T} M,\mathbb{T} M{]}\!{]}_\delta =0$ if and only if $\delta\in \mathbb{T} M$.
\end{enumerate}
\end{lem}
\begin{proof}The backward implications are all straightforward from the definitions.
Assume without loss of generality that $\varphi\neq 0$ and that is of degree $k$. If $[\varphi,\mathbb{T} M]\subseteq \Omega_M^0$, then
\begin{equation}
[\varphi,[\alpha,X]]=[[\varphi,\alpha],X]+(-1)^k[\alpha,[\varphi,X]]=0
\end{equation}
for every $\alpha\in \Omega^1_M$, $X\in \mathcal X_M^1$. Since every function is locally expressible as the tautological inner product of a 1-form and a vector, we conclude that $\left[\varphi,\Omega_M^0\right]=0$. Since $\iota_X\varphi(1)\in \Omega_M^0$ for all $X\in \mathcal X_M^1$ we have $\varphi(1)\in \Omega^1_M\oplus \Omega^0_M$. Let $\varphi(1)\neq 0$: in this case $k\in\{0,1\}$ and by degree comparison, $\left[\varphi, \Omega_M^1\right]=0$, which in turn implies
that $\varphi(\omega)=\varphi(1)\omega$ for all $\omega\in \Omega_M$, i.e. $\varphi\in \Omega_M^k$ with $k\in\{0,1\}$. Now, let $\varphi(1)=0$:
for all $\alpha\in \Omega_M^1$ we have that the linear map $\alpha\mapsto\varphi(\alpha)\in\Omega_M^0$ is tensorial and hence of the form $\iota_X$ for some $X\in \mathcal X_M^1$. Also note that from $\varphi\neq 0$ and $\left[\varphi,\Omega_M^1\right]\subseteq \Omega_M^0$ follows $X\neq 0$, so that $k=-1$. Now, again by degree comparison $\left[\varphi,\mathcal X^1_M\right]=0$ whence $\varphi(\omega)=\iota_X\omega$ for all $\omega\in \Omega_M$. This concludes the proof of 1). The proofs of 2) and 3) follow immediately.
\end{proof}
\begin{lem}\label{lem:8}
Let $\delta\in \mathcal E_M$ be $\Omega_M^0$-linear and odd. Then the restriction of the derived bracket associated to $\delta$ to $\mathbb{T} M$ is $\Omega_M^0$-bilinear and skew-symmetric.
\end{lem}
\begin{proof} To see that the derived bracket is skew-symmetric we calculate
\begin{equation}
{[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]}_\delta-{[}\!{[} \mathbf{y},\mathbf{x}{]}\!{]}_\delta= [\delta,[\mathbf{x},\mathbf{y}]]=0
\end{equation}
for all $\mathbf{x},\mathbf{y}\in \mathbb{T} M$. Similarly, for all $\mathbf{x},\mathbf{y}\in \mathbb{T} M$ and for all $f\in \Omega_M^0$
\begin{equation}
{[}\!{[} f \mathbf{x},\mathbf{y}{]}\!{]}_\delta -f{[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]}_\delta = [[f,\delta]\mathbf{x},\mathbf{y}]=0 \,.
\end{equation}
\end{proof}
\begin{mydef}
A {\it skew-symmetric endomorphism} of $\mathbb{T} M$ is an $\Omega_M^0$-linear bundle endomorphism $\varphi$ of $\mathbb{T} M$ such that $\langle \varphi(\mathbf{x}),\mathbf{y}\rangle + \langle \mathbf{x}, \varphi(\mathbf{y})\rangle =0$ for all $\mathbf{x},\mathbf{y}\in \mathbb{T} M$.
\end{mydef}
\begin{prop}\label{prop:9}
Let $\varphi$ be a skew-symmetric endomorphism of $\mathbb{T} M$. Then there is a unique even element $\widetilde \varphi\in \mathcal E_M$ such that
\begin{enumerate}[1)]
\item $[\widetilde \varphi, \Omega_M^0]=0$;
\item $[\widetilde \varphi,\mathbf{x}]=\varphi(\mathbf{x})$ for all $\mathbf{x}\in \mathbb{T} M$;
\item $\widetilde \varphi(1)\in \Omega^2_M$.
\end{enumerate}
\end{prop}
\begin{proof}
Define $\widetilde \varphi(1)=\epsilon$, where $\epsilon\in \Omega_M^2$ is uniquely defined by $\iota_X\epsilon=-\varphi(X)(1)$ for all $X\in \mathcal{X}_M^1$ (note that $\epsilon$ is alternating precisely because $\varphi$ is skew-symmetric).
Condition 1) suggests to define $\widetilde \varphi(f)=f\epsilon$ for all $f\in \Omega_M^0$. Moreover, condition 2) applied to all $\mathbf{x}\in \Omega_M^1$ inspires to extend recursively $\widetilde \varphi$ to all of $\Omega_M$:
\begin{equation}\label{eq:1}
\widetilde \varphi(\eta\gamma)=\eta\widetilde \varphi(\gamma)+\varphi(\eta)(\gamma)
\end{equation}
for all $\eta\in \Omega_M^1$ and $\gamma\in \Omega_M^k$. In order to show that $\widetilde \varphi$ so constructed is the required element it remains only to prove that $\widetilde \varphi$ satisfies condition 2) for all $\mathbf{x}\in \mathcal{X}_M^1$ as well. Let $X\in \mathcal X_M^1$ and let $\omega\in \Omega_M^k$. We need to prove that
\begin{equation}\label{eq:prop10}\widetilde\varphi(\iota_X\omega)-\iota_X\widetilde\varphi(\omega)=\varphi(X)(\omega)\end{equation}
by induction on $k$: if $k=0$ the identity (\ref{eq:prop10}) follows from condition 1) and the definition of $\epsilon$. Suppose now $k>0$ and that (\ref{eq:prop10}) holds for $0\leq h<k$. Then, we can assume $\omega=\alpha\beta$, with $\alpha\in \Omega_M^1$ and $\beta\in\Omega_M^{k-1}$. Then
\begin{eqnarray} \widetilde\varphi(\iota_X\omega)-\iota_X\widetilde\varphi(\omega)&=& \widetilde\varphi(\iota_X(\alpha\beta))-\iota_X\widetilde\varphi(\alpha\beta)\nonumber\\
&=& \widetilde\varphi((\iota_X\alpha)\beta-\alpha(\iota_X\beta))-\iota_X\widetilde\varphi(\alpha\beta)\nonumber\\
&=& \widetilde\varphi((\iota_X\alpha)\beta)-\widetilde\varphi(\alpha(\iota_X\beta))-\iota_X\widetilde\varphi(\alpha\beta)\nonumber\\
&=& (\iota_X\alpha)\widetilde\varphi(\beta)-\alpha\widetilde\varphi(\iota_X\beta)-\varphi(\alpha)(\iota_X\beta)-\iota_X\left(\alpha\widetilde\varphi(\beta)+\varphi(\alpha)(\beta)\right)\nonumber\\
&=&\alpha\iota_X\widetilde\varphi(\beta)-\alpha\widetilde\varphi(\iota_X\beta)-\varphi(\alpha)(\iota_X\beta)-\iota_X\left(\varphi(\alpha)(\beta)\right)\nonumber\\
&=&- \alpha\varphi(X)(\beta)-[\varphi(\alpha),X](\beta)\nonumber\\
&=&- \alpha\varphi(X)(\beta)+[\alpha,\varphi(X)](\beta)\nonumber\\
&=&\varphi(X)(\omega),\nonumber
\end{eqnarray}
which concludes the inductive step and the proof of the existence of the required element. The uniqueness of $\widetilde \varphi$ is an immediate consequence of conditions 1), 2), 3), and Lemma \ref{lemma:7}. \end{proof}
\begin{rem}\label{rem:11}
Condition 3) in Proposition \ref{prop:9} is somewhat ad hoc, its sole purpose being to ensure uniqueness. Any two operators satisfying both 1) and 2), which we may think of as different lifts of $\varphi$ to $\mathcal E_M$, differ by addition of a section of $\Omega_M^0$.
\end{rem}
\subsection{Quasi split structures and gradings}
A {\it quasi split structure} is a subbundle $E$ of $\mathbb{T} M\otimes \mathbb{C}$ that satisfies the following requirements:
\begin{enumerate}[i)]
\item $E$ is {\it non-degenerate}, i.e.\ the restriction of the tautological inner product to $E$ is non-degenerate;
\item $E$ is {\it split}, i.e.\ $E$ splits fiberwise as direct sum of two isotropic subspaces.
\end{enumerate}
\begin{rem} Every non-degenerate subbundle $E\subseteq \mathbb{T} M\otimes \mathbb C$ is a quasi split structure if and only if the rank of $E$ is equal to twice the {\it Witt index} (i.e.\ the maximum dimension of an isotropic subspace) of $E_p$, for each $p\in M$.
\end{rem}
\begin{rem}
Let $E$ be a subbundle of $\mathbb{T} M$. If the restriction of the tautological inner product to $E$ has signature $(k,k)$ for some $k\in \{1,\ldots, \dim(M)\}$, then $E\otimes \mathbb{C}$ is a quasi split structure. Quasi split structures of this form are called {\it split structures} \cites{AldiGrandini16, AldiGrandini17}.
\end{rem}
\begin{rem}\label{rem:13}
A feature of the Dorfman bracket is that if $f\in \Omega_M^0$ and $\mathbf{x},\mathbf{y}\in \mathbb{T} M$, then \begin{equation}\label{eq:2}
{[}\!{[} f\mathbf{x},\mathbf{y}{]}\!{]} = f{[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]} - 2\langle d f,\mathbf{y}\rangle \mathbf{x} + 2\langle \mathbf{x},\mathbf{y} \rangle d f\,.
\end{equation}
The last term of \eqref{eq:2} is not present if the Courant bracket (i.e.\ the skew-symmetrization of the restriction of the Dorfman bracket to the generalized tangent bundle) is used instead. In particular, this shows that an isotropic subbundle $L\subseteq \mathbb{T} M$ is closed under the Courant bracket if and only if it is closed under the Dorfman bracket. However for non-isotropic subbundles, the last term of \eqref{eq:2} does not vanish in general.
\end{rem}
\begin{lem}\label{lem:14}
$\mathbb{T} M\otimes \mathbb C$ is the only non-zero quasi split structure that is closed under the Dorfman bracket.
\end{lem}
\begin{proof}
Let $E$ be a non-zero quasi split structure. By non-degeneracy, $E$ admits (in a fixed local neighborhood) at least two sections $\mathbf{x}, \mathbf{y}$ such that $\langle \mathbf{x},\mathbf{y}\rangle\neq 0 $ and the projection of $\mathbf{x}$ onto $TM$ is non-zero. Since $f$ in \eqref{eq:2} can be chosen arbitrarily, we have $T^*M\subseteq E$ and thus $E=\mathbb{T} M$.
\end{proof}
\begin{rem}
Let $E$ be a quasi split structure and let $E^\perp$ be its orthogonal complement in $\mathbb T M\otimes\mathbb{C}$ with respect to the tautological inner product. If $L$ is a maximal isotropic subbundle of $E$, then any maximal isotropic subspace of $\mathbb{T} M\otimes \mathbb{C}$ that contains $L$ is of the form $L\oplus L_{\perp}$, for some $L_{\perp}$ maximal isotropic in $E^{\bot}$. In particular, $E^{\bot}$ is also a quasi split structure: in fact, the Witt index of $E^{\perp}$ satisfies
\begin{equation}
{\rm Witt}(E^{\bot})=n-{\rm Witt}(E)=n-\frac{1}{2}{\rm rk}(E)=\frac{1}{2}{\rm rk}(E^{\bot})\,.
\end{equation}
\end{rem}
\begin{rem}
The formula $\mathbf{x} \mathbf{x} \omega = \langle \mathbf{x}, \mathbf{x}\rangle \omega$, valid for every $\mathbf{x} \in \mathbb{T} M \otimes \mathbb C$ and every $\omega\in \Omega_M\otimes \mathbb C$, shows the existence of a $\mathbb C$-linear map from the Clifford algebra $C\ell (\mathbb{T} M\otimes \mathbb C)$ to $\mathcal E_M\otimes \mathbb C$ known as the {\it standard spin representation} \cites{Gualtieri11, Meinrenken13}. The standard spin representation is irreducible, surjective, and faithful. It provides a fiberwise identification of $C\ell (\mathbb{T} M\otimes \mathbb C)$ with $\mathcal E_M\otimes \mathbb C$. Given a quasi split structure $E$, the standard spin representation restricts to two commuting actions of $C\ell(E)$ and $C\ell(E^\perp)$.
\end{rem}
\begin{mydef}
Let $E\subseteq \mathbb{T} M\otimes \mathbb C$ be a quasi split structure and let $L$ be maximal isotropic in $E$. The {\it canonical bundle} of $L$ is the subsheaf $K_L\subseteq \Omega_M\otimes \mathbb C$ of all sections $\rho$ such that $l\rho=0$ for all $l\in L$.
\end{mydef}
\begin{rem}\label{rem:14}
Fiberwise, the canonical bundle is an example of a pure subspace in the sense of \cite{Batista14}. As proved in \cite{Batista14}, $K_L$ has rank $2^{\dim(M)-{\textrm{rank}}(L)}$, showing in particular that $K_L$ is indeed a vector bundle. Also proved in \cite{Batista14} is the identification $L={\rm Ann}(K_L)$.
\end{rem}
\begin{prop}\label{prop:16}
Let $E$ be a quasi split structure of rank $2k$. Each decomposition $E=L\oplus L'$ into maximal isotropic subspaces gives rise to a canonical grading
\begin{equation}\label{eq:grading}
\Omega_M\otimes \mathbb C=\bigoplus_{r=0}^k (\mathord{\adjustbox{valign=B,totalheight=.6\baselineskip}{$\bigwedge$}}^r L')\otimes K_L\,.
\end{equation}
\end{prop}
\begin{proof}
Pick a decomposition $E^\perp=L_\perp\oplus L_\perp'$ into maximal isotropic subspaces. By construction, $C\ell(E^\perp)$ acts on $K_L$. The $C\ell(E^\perp)$-submodule generated by the pure spinor line $K_{L\oplus L_\perp}$ is (see e.g.\ Proposition 3.10 in \cite{Meinrenken13}) isomorphic to the spinor module $\mathord{\adjustbox{valign=B,totalheight=.6\baselineskip}{$\bigwedge$}}^{\bullet} L_\perp'$. Since the latter is of rank $2^{\dim(M)-k}$, we conclude from Remark \ref{rem:14} that $K_L=(\mathord{\adjustbox{valign=B,totalheight=.6\baselineskip}{$\bigwedge$}}^\bullet L_\perp')\otimes K_{L\oplus L_\perp}$. Hence the grading \eqref{eq:grading} is obtained from the canonical grading (see e.g.\ \cite{Gualtieri11}) of $\Omega_M\otimes \mathbb C$ induced by the maximal isotropic decomposition $\mathbb{T} M\otimes \mathbb C=(L\oplus L_\perp)\oplus (L'\oplus L_\perp')$ upon forgetting the information coming from the the choice of $L_\perp$ and $L_\perp'$.
\end{proof}
\begin{mydef}
We refer to the grading \eqref{eq:grading} as the {\it $(L,L')$-grading}.
\end{mydef}
\begin{rem}
As shown in \cite{Gualtieri11}, $K_{L'\oplus L_\perp'}=\det(L'\oplus L_\perp')\otimes K_{L\oplus L_\perp}$. Acting on both sides by $C\ell(E^\perp)$, we obtain the identification $K_{L'}=\det(L')\otimes K_L$ and thus of $(\mathord{\adjustbox{valign=B,totalheight=.6\baselineskip}{$\bigwedge$}}^{k-r} L)\otimes K_{L'}$ with $(\mathord{\adjustbox{valign=B,totalheight=.6\baselineskip}{$\bigwedge$}}^r L) \otimes K_{L'}$. Hence, the $(L,L')$-grading is related to the $(L',L)$-grading by an overall involution.
\end{rem}
\begin{cor}\label{cor:19}
Consider a decomposition $\mathbb{T} M\otimes \mathbb C=\bigoplus_{i=1}^s E_i$ into quasi split structures $E_1,\ldots, E_s$ of respective rank $2k_1,\ldots,2k_s$. The additional datum of isotropic decompositions $E_i=L_i\oplus L'_i$ for all $1\le i \le t\le s$ gives rise to a canonical multigrading
\begin{equation}\label{eq:multigrading}
\Omega_M\otimes \mathbb C = \bigoplus_{r_1=0}^{k_1}\cdots\bigoplus_{r_t=0}^{k_t}(\mathord{\adjustbox{valign=B,totalheight=.6\baselineskip}{$\bigwedge$}}^{r_1} L_1')\otimes \cdots \otimes (\mathord{\adjustbox{valign=B,totalheight=.6\baselineskip}{$\bigwedge$}}^{r_t} L_t')\otimes K_{L_1\oplus \cdots \oplus L_t} \,.
\end{equation}
\end{cor}
\begin{proof}
This follows immediately from Proposition \ref{prop:16} applied to the $(L_1\oplus\cdots \oplus L_t,L_1'\oplus\cdots \oplus L_t')$-grading.
\end{proof}
\begin{lem}\label{lem:23}
Let $\delta\in \mathcal E_M\otimes \mathbb C$ be an operator such that $[\delta,\mathbb{T} M\otimes \mathbb C]\subseteq \mathbb{T} M\otimes \mathbb C$ and let $L,L'\subseteq \mathbb{T} M\otimes \mathbb C$ be isotropic subbundles such that $L\oplus L'$ is a quasi split structure of rank $2k$. If ${[}\!{[} L,L{]}\!{]}_\delta\subseteq L$ and ${[}\!{[} L',L'{]}\!{]}_\delta\subseteq L'$, then $\delta$ decomposes into components of degree $-1,0,1$ with respect to the $(L,L')$-grading.
\end{lem}
\begin{proof}
Our proof is modeled after the proof of Theorem 2.9 in \cite{Gualtieri11}. Let $F_r\subseteq \Omega_M\otimes \mathbb C$ consist of all forms annihilated by $\mathord{\adjustbox{valign=B,totalheight=.6\baselineskip}{$\bigwedge$}}^{r+1}L$ so that, in particular, $F_0=K_L$. The assumption ${[}\!{[} L,L{]}\!{]}_\delta\subseteq L$ implies ${[}\!{[} L,L{]}\!{]}_\delta F_0=0$ and, unraveling the definition of the derived bracket, $\delta(F_0)\subseteq F_1$. In a similar way, proceeding by induction on $r$, we prove that $\delta(F_r)\subseteq F_{r+1}$ for all non-negative integers $r$. Similarly, $\delta(F'_r)\subseteq F'_{r+1}$, where $\{F'_r\}_r$ is the filtration of $\Omega_M\otimes \mathbb C$ induced by $L'$. Hence,
\begin{equation}
\delta((\mathord{\adjustbox{valign=B,totalheight=.6\baselineskip}{$\bigwedge$}}^r L')\otimes K_L)\subseteq \delta(F_r)\cap \delta(F'_{k-r})\subseteq F_{r+1}\cap F'_{k-r+1}
\end{equation}
which concludes the proof.
\end{proof}
\begin{lem}\label{lem:24}
Let $L,L'\subseteq \mathbb{T} M\otimes \mathbb C$ be isotropic subbundles such that $E=L\oplus L'$ is a quasi split structure. Assume $\delta \in \mathcal E_M$ is such that
\begin{enumerate}[1)]
\item $\delta$ is of degree 1 with respect to a grading that commutes with the $(L,L')$-grading;
\item ${[}\!{[} L,L'{]}\!{]}_\delta \subseteq E^\perp$;
\item ${[}\!{[} L,L\oplus E^\perp{]}\!{]}_\delta\subseteq L$ and ${[}\!{[} L',L'\oplus E^\perp{]}\!{]}_\delta\subseteq L'$.
\end{enumerate}
Then $\delta$ is of degree $0$ with respect to the $(L,L')$-grading.
\end{lem}
\begin{proof}
Let $\rho\in K_L$ be a (locally defined) pure spinor so that $L''={\rm Ann}(\rho)$ is maximal isotropic and contains $L$. Using 3) and expanding the derived bracket of $\delta$, we obtain
\begin{equation}
0={[}\!{[} L'',L{]}\!{]}_\delta \rho = L''L\delta \rho\,.
\end{equation}
Hence there exists a local function $g$ such that $L\delta \rho = g\rho$. On the other hand, condition 1) and degree comparison force $g=0$ and thus $\delta(K_L)\subseteq K_L$. Let $\{F_r\}$ be the filtration of $\Omega_M\otimes \mathbb C$ defined by $L$ introduced in the proof of Lemma \ref{lem:23}. In particular, $LF_r=F_{r-1}$ and $L'F_r=F_{r+1}$. Let us assume $\delta(F_r)\subseteq F_{r}$ so that, as a consequence of 2), we have ${[}\!{[} L',L{]}\!{]}_\delta F_r\subseteq F_r$. On the other hand, expanding the definition of the derived bracket and using induction on $r$ we have
\begin{equation}
{[}\!{[} L,L'{]}\!{]}_\delta F_r\subseteq L\delta F_{r+1} + F_r\,.
\end{equation}
Therefore, $\delta(F_{r+1})\subseteq F_{r+1}$ which, by induction on $r$, proves that $\delta(F_r)\subseteq F_r$ for all $r$. Using a similar argument for $L'$, we conclude that $\delta$ preserves the $(L,L')$-grading.
\end{proof}
\section{Generalized Polynomial Structures}\label{sec:3}
In this section we introduce the notion of generalized polynomial structure, modeled after the classical polynomial structures of Goldberg and Yano \cite{GoldbergYano70}. We show that the generalized tangent bundle decomposes into the direct sum of generalized eigenbundles.
\subsection{Definition and basic properties}
\begin{mydef}
Let $N$ be a positive integer. A skew-symmetric endomorphism $\varphi$ of $\mathbb{T} M$ is called a \emph{generalized polynomial structure of degree} $N$ if there is a monic polynomial $P(x)\in \mathbb{R}[x]$ of degree $N$ such that, for all $p\in M$, $P(x)$ is the minimal polynomial of the endomorphism $\varphi_p:\mathbb{T}_p M\rightarrow \mathbb{T}_p M$. We will refer to the polynomial $P(x)$ as the \emph{minimal polynomial} of $\varphi$.
\end{mydef}
\begin{example}
Let $f:TM\to TM$ be a polynomial structure with minimal polynomial $p(x)$. If $f^T$ denotes the transpose action on $T^*M$, then $\varphi=f\oplus (-f^T)$ is by construction a skew-symmetric endomorphism of $\mathbb{T} M$. Moreover, for every $X\in TM$ and $\alpha\in T^*M$, $0=\langle p(f)(X),\alpha\rangle = \langle X,p(f^T)\alpha\rangle$ implies that $p(\varphi)p(-\varphi)$ vanishes on $\mathbb{T} M$. Hence $\varphi$ is a generalized polynomial structure whose minimal polynomial $P(x)$ divides $p(x)p(-x)$. More precisely, if $q(x)$ denotes the largest (monic) even divisor of $p(x)$, then
\begin{equation*}
P(x)=(-1)^{\deg(p)-\deg(q)}\frac{p(x)p(-x)}{q(x)}\,.
\end{equation*}
\end{example}
\begin{lem}\label{lem:GPS}
Let $P(x)\in \mathbb{R}[x]$ be a monic polynomial of degree $N$, and let $\varphi$ be a skew-symmetric endomorphism of $\mathbb{T} M$. Then $\varphi$ is a generalized polynomial structure with minimal polynomial $P(x)$ if and only if the following conditions are satisfied:
\begin{enumerate}[1)]
\item $P(\varphi)(\mathbf{x})=0$ for all $\mathbf{x}\in \mathbb{T} M$;
\item for all $p\in M$, the endomorphisms $I_p,\varphi_p,\varphi^2_p,\dots\varphi^{N-1}_p$ of $\mathbb{T}_p M$ are linearly independent.
\end{enumerate}
\end{lem}
\begin{proof}
Suppose $\varphi$ is a generalized polynomial structure with minimal polynomial $P(x)$. Let us fix a point $p\in M$ and let $\mathbf{x}\in \mathbb{T}_p M$. Since $P(x)$ is the minimal polynomial of $\varphi_p$, then $P(\varphi_p)(\mathbf{x})=0$. Moreover, suppose there exist real numbers $a_0,a_1,\ldots,a_{N-1}$ such that
\begin{equation}
a_0I_p+a_1\varphi_p+a_2\varphi_p^2+\dots+a_{N-1}\varphi_p^{N-1}=0.
\end{equation}
Then the polynomial $Q(x)=a_0+a_1x+a_2x^2+\dots+a_{N-1}x^{N-1}$ satisfies $Q(\varphi_p)=0$. Since ${\rm deg}\ Q(x) < {\rm deg}\ P(x)$, this forces $Q(x)=0$ i.e.\ $a_i=0$ for all $i$. Vice versa, suppose that conditions 1) and 2) are satisfied. It follows from 1) that $P(x)$ is multiple of $Q_p(x)$, the (monic) minimal polynomial of $\varphi_p$. On the other hand, 2) implies ${\rm deg}\ Q_p(x)\geq {\rm deg}\ P(x)$ and thus $Q_p(x)=P(x)$.
\end{proof}
\begin{rem}
Lemma \ref{lem:GPS} shows that generalized polynomial structures naturally extend the notion of classical polynomial structure as introduced by Goldberg and Yano \cite{GoldbergYano70} (and further studied in \cite{GoldbergPetridis73}) to the generalized tangent bundle.
\end{rem}
\begin{example}
A generalized polynomial structure with minimal polynomial $P(x)=x^2+1$ is nothing but a {\it generalized almost complex structure} in the sense of Hitchin and Gualtieri \cites{Hitchin03,Gualtieri11}. Similarly, generalized polynomial structures with minimal polynomial $P(x)=x^3+x$ are precisely (non-zero) {\it generalized $F$-structures} in the sense of Vaisman \cite{Vaisman08} that are not generalized almost complex structures.
\end{example}
\begin{example}
Generalized polynomial structures with minimal polynomial of the form $P(x)=x^2-p x-q$ are also known as skew-symmetric {\it generalized metallic structures} \cite{BlagaNannicini20}. (Throughout the present paper we will assume generalized metallic structures to be additionally skew-symmetric with respect to the tautological inner product, in contrast to \cite{BlagaNannicini20}.) In particular, {\it generalized almost tangent structures} \cite{BlagaCrasmareanu14} are generalized metallic structures such that $p=q=0$.
\end{example}
\begin{rem}\label{rem:29}
Let $f$ be a (classical) metallic structure i.e.\ a polynomial structure with minimal polynomial $p(x)=x^2-ax-b$ and let $\varphi=f\oplus(-f^T)$. If $a=0$, then $\varphi$ is a generalized metallic structure with minimal polynomial $p(x)$. If $a\neq 0$, then $\varphi$ is a generalized polynomial structure of degree 4 whose with minimal polynomial is $p(x)p(-x)$.
\end{rem}
\begin{rem}
Given a skew-symmetric endomorphism $\varphi:\mathbb{T} M\rightarrow \mathbb{T} M$, consider the ideal
\begin{equation}
{\mathcal I}(\varphi)=\{P(x)\in\mathbb{R}[x]: P(\varphi)(\mathbf{x})=0\ \mbox{ for all }\mathbf{x}\in \mathbb{T} M\}\,.
\end{equation}
If $\varphi$ is a generalized polynomial structure, then its minimal polynomial $P(x)$ belongs to $\mathcal I(\varphi)$. In particular, ${\mathcal I}(\varphi)$ is non-trivial. Among all the elements of $\mathcal I(\varphi)$, $P(x)$ is the unique monic polynomial of smallest degree. This motivates the following.
\end{rem}
\begin{mydef} \label{def: WGPS}
A skew-symmetric endomorphism $\varphi$ of $\mathbb{T} M$ is called a \emph{weak generalized polynomial structure} if ${\mathcal I}(\varphi)\neq 0$. The monic polynomial $P(x)$ of smallest degree in ${\mathcal I}(\varphi)$ is called the {\it minimal polynomial of $\varphi$}, while ${\rm deg}\ P(x)$ is called the {\it degree} of $\varphi$.
\end{mydef}
\begin{lem}
Let $M$ be a smooth connected manifold. A skew-symmetric endomorphism $\varphi$ of $\mathbb{T} M$ is a weak generalized polynomial structure if and only if its spectrum (the set of the eigenvalues, regardless of the multiplicities) is constant.
\end{lem}
\begin{proof}
If $\varphi$ has a constant spectrum, then its characteristic polynomial $Q(x)={\rm det}(\varphi-xI)$ has constant coefficients and positive degree. Moreover, by the Cayley-Hamilton Theorem, $Q(x)\in {\mathcal I}(\varphi)$. Vice versa, if $Q(x)$ is any polynomial of positive degree in ${\mathcal I}(\varphi)$, then $Q(\lambda)=0$ for every eigenvalue $\lambda$ of $\varphi$. Since by definition all polynomials in $\mathcal I(\varphi)$ have constant coefficients, then the spectrum is also constant.
\end{proof}
\begin{lem}\label{lem:36}
A skew-symmetric endomorphism $\varphi$ is a weak generalized polynomial structure if and only if there is a monic polynomial $P(x)\in \mathbb{R}[x]$ of positive degree $N$ such that
\begin{enumerate}[1)]
\item $P(\varphi)(\mathbf{x})=0$ for all $\mathbf{x}\in\mathbb{T} M$;
\item the endomorphisms $I,\varphi,\varphi^2,\dots,\varphi^{N-1}$ are $\mathbb{R}$-linearly independent.
\end{enumerate}
\end{lem}
\begin{proof}
The proof is a straightforward adaptation of the proof of Lemma \ref{lem:GPS}.
\end{proof}
\begin{rem}
Lemma \ref{lem:36} shows that weak generalized polynomial structures of constant rank extend the notion of polynomial structure described by Goldberg and Yano \cite{GoldbergYano70} to the generalized tangent bundle.
\end{rem}
\begin{rem}
It follows from Remark \ref{rem:29} that every generalized polynomial structure is also a weak generalized polynomial structure (and the two definitions of the minimal polynomial coincide). The converse is not necessarily true as the monic polynomial of smallest degree in $\mathcal I(\varphi)$ is a priori only a multiple of the minimal polynomials of the fiberwise endomorphisms $\varphi_p$.
\end{rem}
\begin{example}
Given a generalized polynomial structure $\varphi$ with minimal polynomial $P(x)=x^2$ and a function $g\in \Omega_M^0$, consider $\varphi'=g\varphi$. Then, the minimal polynomial $Q_p(x)$ of $\varphi'_p$ satisfies
\begin{equation}
Q_p(x)=\left\{\begin{array}{lc} x^2&\mbox{ if } g(p)\neq 0,\\
x&\mbox{ if } g(p)=0.\end{array}\right.
\end{equation}
In particular, suppose $g$ is non-zero and non-invertible. Then $\varphi'$ is not a generalized polynomial structure, but still a weak generalized polynomial structure. Moreover, $P(x)=x^2$ is still the minimal polynomial of $\varphi'(x)$ in the sense of Definition \ref{def: WGPS}.
\end{example}
\begin{lem}
Let $\varphi$ be a weak generalized polynomial structure such that all eigenvalues of $\varphi$ have constant multiplicity $1$ (i.e.\ $\varphi$ is semisimple). Then $\varphi$ is a generalized polynomial structure.
\end{lem}
\begin{proof}
Let $P(x)$ be the polynomial in ${\mathcal I}(\varphi)$ of smallest degree, and let $p\in M$. Since $P(\varphi_p)=0$ and $P(x)$ has only simple roots, then $P(x)$ is also the minimal polynomial of $\varphi_p$ for each $p\in M$.
\end{proof}
\begin{rem}\label{rem:34}
Let $\varphi$ be a weak generalized polynomial structure. Since $\varphi$ is in particular a skew-symmetric endomorphism of $\mathbb{T} M$, its spectrum must be symmetric with respect to $0\in \mathbb C$. As a consequence, the minimal polynomial $P(x)$ of $\varphi$ is either even or odd, i.e. of the form $P(x)=x^k Q(x^2)$, for some non negative integer $k$ and some $Q(x)\in \mathbb{R}[x]$. For example, imposing that a generalized metallic structure with minimal polynomial $x^2-px-q$ is skew-symmetric implies the condition $p=0$.
\end{rem}
\begin{example}
Let $f:T M\to T M$ be a polynomial structure and let $\lambda$ be an eigenvalue of the associated generalized polynomial structure $\varphi=f\oplus (-f^t)$. Then $L_\lambda=(L_\lambda\cap TM)\oplus (L_\lambda \cap T^*M)$. The first summand is the generalized eigenbundle with eigenvalue $\lambda$ for the action of $f$ of $TM$. The second summand is the generalized eigenspace with eigenvalue $-\lambda$ for the transpose action of $f^t$ on $T^*M$.
\end{example}
\subsection{Generalized eigenbundles}
Let $\varphi$ be a weak generalized polynomial structure and let $\lambda$ be an eigenvalue of $\varphi$. We have a sequence of subsheaves of $\mathbb{T} M\otimes\mathbb{C}$:
\begin{equation}\label{filt1}
{\rm Ker}(\varphi-\lambda I)\subseteq {\rm Ker}((\varphi-\lambda I)^2)\subseteq {\rm Ker}((\varphi-\lambda I)^3)\subseteq \cdots\subset {\rm Ker}((\varphi-\lambda I)^n)\subseteq\cdots.
\end{equation}
The inclusions in (\ref{filt1}) are all proper, until the sequence stabilizes. The smallest integer $m(\lambda)$ at which (\ref{filt1}) stabilizes coincides with the {\it multiplicity} of $\lambda$ in the minimal polynomial of $\varphi$. We denote
\begin{equation}
L_{\lambda}:={\rm Ker}((\varphi-\lambda I)^{m(\lambda)})
\end{equation}
the corresponding subsheaf. We also denote by $\Sigma(\varphi)$ the spectrum of $\varphi$ i.e.\ the set of all eigenvalues of $\varphi$. We reserve the notation $\Sigma_+(\varphi)$ for any fixed subset $\Sigma_+(\varphi)\subset \Sigma(\varphi)$ such that $\Sigma_+(\varphi)$ and $-\Sigma_+(\varphi)$ partition $\Sigma(\varphi)\setminus\{0\}$.
\begin{rem}
If $\varphi$ is a generalized polynomial structure, then $L_{\lambda}$ is a vector subbundle of $\mathbb{T} M\otimes \mathbb{C}$, whose complex rank equals the multiplicity of $\lambda$ in the minimal polynomial. In this case we refer to $L_\lambda$ as the {\it generalized eigenbundle of $\varphi$ with eigenvalue $\lambda$}. In light of Remark \ref{rem:34}, $L_{\lambda}$ and $L_{-\lambda}$ have equal rank.
\end{rem}
\begin{rem}\label{rem:im}
Similarly to \eqref{filt1}, we have the filtration
\begin{equation}\label{filt2}
{\rm Im}(\varphi-\lambda I)\supseteq {\rm Im}((\varphi-\lambda I)^2)\supseteq {\rm Im}((\varphi-\lambda I)^3)\supseteq \cdots\supseteq {\rm Im}((\varphi-\lambda I)^n)\supseteq\cdots.
\end{equation}
which stabilizes at $n=m(\lambda)$. If $\varphi$ is a generalized polynomial structure, then the sheaf
\begin{equation}
R_{\lambda}:={\rm Im}((\varphi-\lambda I)^{m(\lambda)})=\bigoplus_{\mu\in \Sigma(\varphi)\setminus\{\lambda\}}L_{\mu}
\end{equation}
is a vector bundle. More generally, given any subset
$S$ of $\Sigma(\varphi)$, we have
\begin{equation}
{\rm Im}\left(\prod_{\lambda\in S}(\varphi-\lambda I)^{m(\lambda)}\right)=\bigoplus_{\mu\in \Sigma(\varphi)\setminus S}L_{\mu}
\end{equation}
and
\begin{equation}
\qquad{\rm Ker}\left(\prod_{\lambda\in S}(\varphi-\lambda I)^{m(\lambda)}\right)=\bigoplus_{\mu\in \Sigma(\varphi)\setminus S}R_{\mu}\,.
\end{equation}
\end{rem}
\begin{theorem} \label{theorem:40}
Let $\varphi$ be a generalized polynomial structure. The following properties are satisfied:
\begin{enumerate}[i)]
\item For all $\lambda\in\Sigma(\varphi)\setminus\{0\}$, the bundles $L_{\lambda}$ are isotropic;
\item for all $\lambda\in\Sigma(\varphi)\setminus\{0\}$, the bundles $$E_{\lambda}:=L_{\lambda}\oplus L_{-\lambda}$$ are quasi split structures;
\item the bundle $E_0:=L_0$ is a quasi split structure;
\item the bundle $\mathbb{T} M\otimes{\mathbb{C}}$ splits in the orthogonal sum
\begin{equation}\label{eq:splitting}\mathbb{T} M\otimes{\mathbb{C}}=E_0\oplus\bigoplus_{\lambda\in \Sigma_+(\varphi)}E_{\lambda}\end{equation}
\end{enumerate}
\end{theorem}
\begin{proof} It suffices to show that if $\lambda\in \Sigma(\varphi) $, then
\begin{equation}
(L_{\lambda})^{\bot}=\bigoplus_{\mu\in \Sigma(\varphi)\setminus\{-\lambda\}} L_{\mu}\,.
\end{equation}
To see this, let $P(x)$ be the minimal polynomial of $\varphi$, and let
$F_{\lambda}(x)=P(x)/(x-\lambda)^{m(\lambda)}.$ From Remark \ref{rem:im} it follows that $L_{\lambda}$ is the image of $F_{\lambda}(\varphi)$.
Also note that $(F_{\lambda}(\varphi))^*=(-1)^{N-m(\lambda)}F_{-\lambda}(\varphi)$, where $N={\rm deg}\ P(x)$.
This implies that $\mathbf{x}\in (L_{\lambda})^{\bot}$ if and only if $\mathbf{x}\in {\rm Ker}( F_{-\lambda}(\varphi))$. Again from Remark \ref{rem:im}, the last condition is equivalent to
$\mathbf{x}\in R_{-\lambda}$.
\end{proof}
\begin{cor}\label{cor:42}
Let $\varphi$ be a generalized polynomial structure that is neither invertible nor nilpotent, then $L_0$ is not closed under the Dorfman bracket.
\end{cor}
\begin{proof}
Since $\varphi$ is not invertible, Theorem \ref{theorem:40} implies that $L_0$ is a non-zero quasi split structure. On the other hand $L_0$ is not nilpotent and thus a proper subbundle of $\mathbb{T} M$. The result then follows from Lemma \ref{lem:14}.
\end{proof}
\begin{rem}
The splitting \eqref{eq:splitting} does not determine uniquely $\varphi$. However, there is a unique polynomial structure $\varphi_s$ such that $L_{\lambda}$ is the $\lambda$-eigenbundle of $\varphi_s$. Setting $\varphi_n:=\varphi-\varphi_s$, the decomposition $\varphi=\varphi_s+\varphi_n$ coincides fiberwise with the Jordan-Chevalley decomposition of $\varphi$, where $\varphi_s$ is the {\it semisimple part of $\varphi$} and $\varphi_n$ is the {\it nilpotent part of $\varphi$}.
\end{rem}
\begin{rem}
The decomposition \eqref{eq:splitting} can be computed explicitly as follows. Let $a_{\lambda,i}$ be the coefficients of the partial fraction decomposition
\begin{equation}
\frac{1}{P(x)}=\sum_{\lambda\in \Sigma(\varphi)}\sum_{i=1}^{m(\lambda)}\frac{a_{\lambda,i}}{(x-\lambda)^i}\,.
\end{equation}
For each $\lambda\in \Sigma(\varphi)$ and for each $i\in\{1,\ldots,m(\lambda)\}$, let $Q_{\lambda,i}$ be the polynomial defined by the relation $(x-\lambda)^i Q_{\lambda,i}(x)=P(x)$. Then the projector $\mathcal P_\lambda :\mathbb{T} M\otimes \mathbb C\to L_\lambda$ satisfies
\begin{equation}\label{eq:9}
\mathcal P_\lambda(\mathbf{x})= \sum_{i=1}^{m(\lambda)}a_{\lambda,i}Q_{\lambda,i}(\varphi)(\mathbf{x})
\end{equation}
for all $\mathbf{x}\in \mathbb{T} M\otimes \mathbb C$. As a consequence of \eqref{eq:9}, the semisimple part of $\varphi$ can be written as
\begin{equation}\label{eq:44}
\varphi_s = \sum_{\lambda \in \Sigma(\varphi)}\lambda \mathcal P_\lambda = \sum_{\lambda\in \Sigma(\varphi)} \sum_{i=1}^{m(\lambda)}\lambda a_{\lambda,i}Q_{\lambda,i}(\varphi)\,.
\end{equation}
In particular, we recover the known fact \cite{Humphreys78} that the semisimple and the nilpotent parts of $\varphi$ are both polynomials in $\varphi$.
\end{rem}
\begin{example}
Let $c$ be a non-zero real number and let $\varphi$ be a generalized polynomial structure with minimal polynomial $P(x)=(x^2+c)^2$. The projectors on $L_{\pm \sqrt{-c}}$ are
\begin{align*}
\mathcal P_{\pm\sqrt{-c}}&=\pm\frac{1}{4c\sqrt{-c}}(\varphi\mp \sqrt{-c})(\varphi\pm \sqrt{-c})^2-\frac{1}{4c}(\varphi\pm \sqrt{-c})^2\\
&= \mp\frac{1}{4c\sqrt{-c}}(\varphi^3+3c\varphi\mp 2c\sqrt{-c})
\end{align*}
from which we calculate the semisimple part of $\varphi$ to be $\varphi_s=\frac{1}{2c}\varphi^3+\frac{3}{2}\varphi$. Hence the nilpotent part of $\varphi$ is $\varphi_n=-\frac{1}{2c}\varphi^3-\frac{1}{2}\varphi$.
\end{example}
\subsection{Block decomposition}\label{decompos}
The content of this subsection is adapted from \cite{BurgoyneCushman77}. Let $\varphi$ be a polynomial structure, and let $p\in M$. A \emph{real block} of $\varphi$ at $p$ is a non-degenerate subspace $V\subseteq \mathbb{T}_p M$ such that $\varphi(V)\subseteq V$. Analogously, a \emph{complex block} of $\varphi$ at $p$ is a non-degenerate subspace $V\subseteq \mathbb{T}_p M\otimes\mathbb{C}$ such that $\varphi(V)\subseteq V$.
Moreover, a real block (respectively, a complex block) $V$ is called \emph{indecomposable} if there is no real block (respectively, complex block) $V'$ such that $V'\subseteq V$ and $V'\neq V, V'\neq \{0\}$. Note that since $\varphi$ is skew, $\mathbb{T}_p M$ splits into an orthogonal sum of real indecomposable blocks, and similarly $\mathbb{T}_p M\otimes \mathbb{C}$ splits into an orthogonal sum of complex indecomposable blocks. Let ${\rm deg}(V)$ denote the degree of nilpotency of $\varphi_n$ on a block $V$, that is the smallest positive integer $k$ such that restriction of $\varphi_n^k$ to $V$ vanishes.
The general structure of indecomposable blocks is described by the following
\begin{lem}\label{lem:block1}
For every indecomposable block $V$ of degree $k$, there exist a subspace $W\subset V$ satisfying the following properties:
\begin{enumerate}[(i)]
\item $V=W\oplus K$, where $K={\rm Ker}\left(\varphi_{n|V}^{k-1}\right)$;
\item $\varphi_s(W)\subset W$;
\item the bilinear form $\tau_{k-1}(\mathbf{w},\mathbf{w}'):=\langle \mathbf{w}, \varphi_n^{k-1}(\mathbf{w}')\rangle$ (symmetric for $k$ odd and skewsymmetric for $k$ even) is non-degenerate on $W$;
\item $V=W\oplus \varphi_n(W)\oplus \varphi^2_n(W)\oplus\dots\oplus \varphi^{k-1}_n(W)$;
\item if $V$ is a complex block and $\overline{V}=V$, then $\overline{W}=W$;
\item for any non negative integer $j<k-1$, the bilinear forms $\tau_{j}(\mathbf{w},\mathbf{w}'):=\langle \mathbf{w}, \varphi_n^{j}(\mathbf{w}')\rangle$ vanish identically on $W$;
\item there are no proper subspaces $W'\subset W$, $W'\neq 0$ that satisfy properties (i)-(iii).
\end{enumerate}
\end{lem}
\begin{proof}
The existence of a subspace $W$ satisfying (i) and (ii) follows from $\varphi_s(K)\subset K$ and the fact that $\varphi_s$ is semisimple.
Now, we prove that $W$ satisfies (iii): if $\mathbf{w}\in W$ is such that $\tau_{k-1}(\mathbf{w}',\mathbf{w})=0$ for all $\mathbf{w}'\in W$, then $\langle \mathbf{v}, \varphi_n^{k-1}(\mathbf{w})\rangle=0$ for all $\mathbf{v}\in V$, whence $\mathbf{w}\in W\cap K=0$. Moreover, note that $W$ generates a splitting
$$V':=W\oplus \varphi_n(W)\oplus\dots \oplus\varphi_n^{k-1}(W)\subseteq V.$$
Since $\tau_{k-1}$ is non-degenerate we deduce that $V'$ is a block, so that by indecomposability $V'=V$ and (iv) is proved. A similar reasoning shows (vii).
Next, we prove that we can modify $W$ so that (v) is satisfied as well: let $\overline{V}=V$, let $\sigma: V\rightarrow V$ be the complex conjugation map and let $\pi: V\rightarrow K$ the projection onto $K$ with respect to the decomposition $V=W\oplus K$. Now, we can replace $W$ with the graph of the map
$$f:W\rightarrow K, \qquad f=-\frac{1}{2}\sigma \pi\sigma_{|W}$$
which satisfies (i)-(v). It remains to prove that we can further modify $W$ so that (vi) is also satisfied. Let
$\widehat{\tau}_j: W\rightarrow W^*$ such that $\widehat{\tau}_j(\mathbf{w})(\mathbf{w}')=\tau_j(\mathbf{w},\mathbf{w}')$. Suppose $j<k-1$ is such that $\widehat{\tau}_{h}=0$ for $j<h<k-1$.
Then, the graph $W''$ of the map
$$f_j:W\rightarrow K \qquad f_j=\frac{(-1)^{k-j}}{2}N^{k-1-j}\widehat{\tau}^{-1}_{h-1}\widehat{\tau}_j$$
is still a complement of $K$, satisfies properties (i)-(v) and additionally $\tau_h(W'',W'')=0$ for $j\leq h<k-1$. Replacing $W$ with $W''$ and repeating this construction, after a finite number of iterations we obtain the requested property. \end{proof}
\begin{mydef}
We will refer to a subspace $W$ satisfying the properties of the previous lemma as a \emph{semisimple component} of the block $V$.
\end{mydef}
\begin{lem}\label{lem:block2}
For any indecomposable complex block $V$ with ${\rm deg}(V)=k$ and semisimple component $W$, one of the following conditions is satisfied:
\begin{enumerate}[i)]
\item $W\subset E_0$, $k$ is odd and ${\rm dim}(W)=1$. In particular, $W={\rm Span}_{\mathbb{C}}(\mathbf{w})$ where $\mathbf{w}\in E_0$ and $\langle \mathbf{w}, \varphi_n^{k-1}(\mathbf{w})\rangle=1$;
\item $W\subset E_{\lambda}$ for some $\lambda\in\Sigma(\varphi)$ and ${\rm dim}(W)=2$. In particular, $W={\rm Span}_{\mathbb{C}}(\mathbf{w}_+,\mathbf{w}_-)$ where $\mathbf{w}_{\pm}\in L_{\pm\lambda}$ and $\langle \mathbf{w}_+, \varphi_n^{k-1}(\mathbf{w}_-)\rangle=\frac{1}{2}$.
\end{enumerate}
\end{lem}
\begin{proof}
From the indecomposability $V\subset E_{\lambda}$ for some $\lambda$. If $\lambda\neq 0$, from Lemma \ref{lem:block1} follows that $W=(W\cap L_{\lambda})\oplus(W\cap L_{-\lambda})$ where the summands are isotropic with respect to the bilinear form $\tau_{k-1}$, and that ${\rm dim}(W)=2$. If $\lambda=0$ and $k$ is even, $\tau_{k-1}$ is a indecomposable, non-degenerate skewsymmetric bilinear form, so that ${\rm dim}(W)=2$. If $\lambda=0$ and $k$ is odd, $\tau_{k-1}$ is a indecomposable, non-degenerate symmetric bilinear form, so that ${\rm dim}(W)=1$.
\end{proof}
\begin{rem}
We denote by $\Delta_k(0)$ and $\Delta_k(\lambda,-\lambda)$ the families of indecomposable blocks (also called \emph{ indecomposable complex types}) satisfying (i) and (ii) respectively in Lemma \ref{lem:block2}. Note that for $k$ odd, the family $\Delta_k(0,0)$ does not contain any indecomposable blocks. However, for $k$ odd we can redefine $\Delta_k(0,0):=\Delta_k(0)\oplus\Delta_k(0)$, that is the family of blocks that are orthogonal sum of two blocks of type $\Delta_k(0)$.
\end{rem}
\begin{rem}\label{ex: blocks}
If ${\rm dim}\ M=3$, then at any $p\in M$ the complex block decomposition of $\varphi_p$ is of one of the following forms:
\begin{enumerate}
\item $\Delta_5(0)\oplus\Delta_1(0)$;
\item $\Delta_2(\lambda_1,-\lambda_1)\oplus\Delta_1(\lambda_2,-\lambda_2)$;
\item $\Delta_3(\lambda,-\lambda)$;
\item $\Delta_3(0)\oplus\Delta_1(\lambda,-\lambda)\oplus\Delta_1(0)$;
\item $\Delta_1(\lambda_1,-\lambda_1)\oplus \Delta_1(\lambda_2,-\lambda_2)\oplus \Delta_1(\lambda_3,-\lambda_3)$.
\end{enumerate}
Each decomposition corresponds to a minimal polynomial which may depend on the eigenvalues involved. Note that there are constraints on the eigenvalues: in the decompositions (2), (3) and (4) the eigenvalues must be either real or imaginary, while in (5) if $\lambda_i$ is neither real or imaginary, then there is a $\lambda_j$ with $j\neq i$ such that $\lambda_j=\pm\overline{\lambda_i}$ while the remaining eigenvalues $\pm \lambda_k$ are either real or imaginary.
\end{rem}
\noindent We would like to describe the indecomposable real blocks as well. Let $V$ be a indecomposable complex block of degree $k$ and semisimple component $W$. Moreover, let $\tau=\tau_{k-1}$ be the non-degenerate bilinear form as before.\\ \\ Let first $V=\overline{V}$, so that $W=\overline{W}$ and $V$ is of type either $\Delta_k(0)$ or $\Delta_k(\lambda,-\lambda)$ for $\lambda\in \mathbb{R}\cup\sqrt{-1}{\mathbb{R}}$. Then, $W=W_0\oplus\sqrt{-1}W_0$ and $V=V_0\oplus\sqrt{-1}V_0$, where $$V_0=W_0\oplus \varphi_{n}(W_0)\oplus \varphi^2_{n}(W_0)\oplus\dots\oplus \varphi^{k-1}_{n}(W_0)$$
and $V_0$ is a real indecomposable block, with semisimple part $W_0$. We denote by ${\rm sign}(V_0):=(n_+,n_-)$ the signature of the tautological inner product on $V_0$.
\begin{lem}\label{lem:block3} Let $V$ be an indecomposable complex block with semisimple part $W$ such that $\overline{V}=V$ and $\overline{W}=W$. One of the following conditions is satisfied:
\begin{enumerate}[(i)]
\item $V\in \Delta_{2h+1}(0)$: $W_0={\rm Span}_{\mathbb{R}}(\mathbf{w})$ where $\varphi_s(\mathbf{w})=0$ and $\langle \mathbf{w}, \varphi_n^{2h}(\mathbf{w})\rangle=\epsilon=\pm 1$.
Moreover, when $\epsilon=+1$ we have ${\rm sign}(V_0)=(h+1,h)$ if $h$ is even, while ${\rm sign}(V_0)=(h,h+1)$ if $h$ is odd.
When $\epsilon=-1$ we have ${\rm sign}(V_0)=(h,h+1)$ if $h$ is even, while ${\rm sign}(V_0)=(h+1,h)$ if $h$ is odd.
\item $V\in \Delta_{k}(\lambda_0,-\lambda_0)$, with $\lambda_0\in \mathbb{R}$: $W_0={\rm Span}_{\mathbb{R}}(\mathbf{w}_+,\mathbf{w}_-)$ where $\mathbf{w}_{\pm}\in W_0\cap L_{\pm\lambda_0}$ and $\langle \mathbf{w}_+, \varphi_n^{k-1}(\mathbf{w}_-)\rangle=\frac{1}{2}$.
Moreover, ${\rm sign}(V_0)=(k,k)$;
\item $V\in \Delta_{2h+1}(\sqrt{-1}\lambda_0,-\sqrt{-1}\lambda_0)$, with $\lambda_0\in \mathbb{R}$: $W_0={\rm Span}_{\mathbb{R}}(\mathbf{u},\mathbf{v})$ where $\varphi_s(\mathbf{u})=\lambda_0 \mathbf{v}$, $\varphi_s(\mathbf{v})=-\lambda_0 \mathbf{u}$ and $\langle u, \varphi_n^{2h}(\mathbf{v})\rangle=0$, $\langle \mathbf{u}, \varphi_n^{2h}(\mathbf{u})\rangle=\langle \mathbf{v}, \varphi_n^{2h}(\mathbf{v})\rangle=\epsilon=\pm 1$.
Moreover, when $\epsilon=+1$ we have ${\rm sign}(V_0)=(2h+2,2h)$ if $h$ is even, while ${\rm sign}(V_0)=(2h,2h+2)$ if $h$ is odd.
When $\epsilon=-1$ we have ${\rm sign}(V_0)=(2h,2h+2)$ if $h$ is even, while ${\rm sign}(V_0)=(2h+2,2h)$ if $h$ is odd.
\item $V\in \Delta_{2h}(\sqrt{-1}\lambda_0,-\sqrt{-1}\lambda_0)$, with $\lambda_0\in \mathbb{R}$: $W_0={\rm Span}_{\mathbb{R}}(\mathbf{u},\mathbf{v})$ where $\varphi_s(\mathbf{u})=\lambda_0 \mathbf{v}$, $\varphi_s(\mathbf{v})=-\lambda_0 \mathbf{u}$ and $\langle \mathbf{u}, \varphi_n^{2h-1}(\mathbf{v})\rangle=\pm\frac{1}{2}$.
Moreover, ${\rm sign}(V_0)=(2h,2h)$.
\end{enumerate}
\end{lem}
\begin{proof}In cases (i) and (ii), we can choose the basis vectors (as described in Lemma \ref{lem:block2}) in $W_0$. In case (iii) we can choose $\mathbf{w}_-=\overline{\mathbf{w}}_+$ and write $\mathbf{w}_+=\mathbf{u}-\sqrt{-1}\mathbf{v}$. In case (iii) we can choose $\mathbf{w}_-=\sqrt{-1}\overline{\mathbf{w}}_+$ and write $\mathbf{w}_+=\mathbf{u}-\sqrt{-1}\mathbf{v}$. Finally, the calculation of the signatures is a straightforward consequence of Lemma \ref{lem:block1}.\end{proof}
\noindent We will denote by $\Delta^{\pm}_{2h+1}(0)$, $\Delta^0_{k}(\lambda_0,-\lambda_0)$, and $\Delta_{k}^{\pm}(\sqrt{-1}\lambda_0,-\sqrt{-1}\lambda_0)$ the families of the real parts $V_0$ corresponding respectively to the conditions (i), (ii), and the condition (iii) for $k=2h+1$ or (iv) for $k=2h$. In particular, note that if $k$ is even, $\Delta_{k}^-(\sqrt{-1}\lambda_0,-\sqrt{-1}\lambda_0)=\Delta_{k}^{+}(-\sqrt{-1}\lambda_0,\sqrt{-1}\lambda_0)$, so that we can set the sign to be a $+$ after reordering the eigenvalues.\\ \\
Finally, consider the case of an indecomposable complex block $V$ such that $\overline{V}\neq V$: in this case we have $\overline{V}\cap V=0$ and $V\in \Delta_k(\lambda,-\lambda)$ with $\lambda\notin \mathbb{R}\cup\sqrt{-1}\mathbb{R}$. In this case, $VV:=V\oplus \overline{V}$ admits a real part $VV_0$ and the subspace $WW:=W\oplus \overline{W}$ admits a real part $WW_0$, such that
$$VV_0=WW_0\oplus \varphi_n(WW_0)\oplus\dots \oplus\varphi^{k-1}_n(WW_0)$$
and $VV_0$ is an indecomposable real block, with semisimple part $WW_0$.
\begin{lem}Let $\lambda=a+\sqrt{-1}b$, where $a\neq 0\neq b$. Then
$WW_0={\rm Span}_{\mathbb{R}}\{\mathbf{u}_+,\mathbf{u}_-,\mathbf{v}_+,\mathbf{v}_-\}$ such that
$$\varphi_s(\mathbf{u}_{\pm})=\pm(a\mathbf{u}_{\pm}+b\mathbf{v}_{\pm}), \qquad \varphi_s(\mathbf{v}_{\pm})=\pm(-b\mathbf{u}_{\pm}+a\mathbf{v}_{\pm})$$
and
$$\langle \mathbf{u}_{\pm},\varphi_n^{k-1}(\mathbf{u}_{\pm})\rangle=\langle \mathbf{v}_{\pm},\varphi_n^{k-1}(\mathbf{v}_{\pm})\rangle=\langle \mathbf{u}_{\pm},\varphi_n^{k-1}(\mathbf{v}_{\pm})\rangle=\langle \mathbf{u}_{\pm},\varphi_n^{k-1}(\mathbf{v}_{\mp})\rangle=0$$
$$\langle \mathbf{u}_{+},\varphi_n^{k-1}(\mathbf{u}_-)\rangle=\frac{1}{2}, \qquad \langle \mathbf{v}_{+},\varphi_n^{k-1}(\mathbf{v}_-)\rangle=-\frac{1}{2}.$$
Moreover, ${\rm sign}(VV_0)=(2k,2k)$.
\end{lem}
\begin{proof}
Calculation analogous to the proof of Lemma \ref{lem:block3}.
\end{proof}
\noindent The families of the real subspaces $VV_0$ described above are denoted $\Delta_k^0(\lambda,-\lambda,\overline{\lambda},-\overline{\lambda})$. We refer to the families
\begin{equation}
\Delta^{\pm}_{2h+1}(0)\, ,\quad \Delta^0_{k}(\lambda_0,-\lambda_0)\, , \quad \Delta_{k}^{\pm}(\sqrt{-1}\lambda_0,-\sqrt{-1}\lambda_0)\, , \textrm{ and }\quad \Delta_k^0(\lambda,-\lambda,\overline{\lambda},-\overline{\lambda})
\end{equation}
as \emph{indecomposable real types}. From the above discussion follows the
\begin{prop}Let $\varphi:\mathbb{T} M\rightarrow \mathbb{T} M$ be a skew-symmetric endomorphism. Then, at any point $p\in M$, the generalized tangent space $\mathbb{T}_p M$ splits into an orthogonal sum of real blocks, each of type either $\Delta^{\pm}_{2h+1}(0)$, $\Delta^0_{k}(\lambda_0,-\lambda_0)$, $\Delta_{k}^{\pm}(\sqrt{-1}\lambda_0,-\sqrt{-1}\lambda_0)$ or $\Delta_k^0(\lambda,-\lambda,\overline{\lambda},-\overline{\lambda})$.
\end{prop}
\begin{example} If $\varphi$ is a generalized almost complex structure, its real block decomposition is fiberwise of type
$m\cdot\Delta^+_{1}(\sqrt{-1},-\sqrt{-1})\oplus m\cdot\Delta^-_{1}(\sqrt{-1},-\sqrt{-1})$, where $2m=n$. More generally, if $\varphi$ is a generalized $f$-structure, the real block decomposition is of type
$m_1\cdot\Delta^+_{1}(\sqrt{-1},-\sqrt{-1})\oplus m_2\cdot\Delta^-_{1}(\sqrt{-1},-\sqrt{-1})\oplus n_1\cdot \Delta^+_{1}(0)\oplus n_2\cdot \Delta^-_{1}(0)$, where $2m_1+n_1=n=2m_2+n_2$.
\end{example}
\begin{example}\label{ex:54}
Suppose $M$ is 3-dimensional, fix $p\in M$, fix a basis $\{v_1,v_2,v_3\}$ of $T_pM$, and dual basis given by $\{\alpha_1,\alpha_2,\alpha_3\}$. Consider the endomorphism $\varphi:\mathbb{T}_pM\rightarrow \mathbb{T}_p M$ such that $\varphi(v_1)=v_2$, $\varphi(v_2)=v_3+\alpha_3$, $\varphi(v_3)=-\alpha_2$,
$\varphi(\alpha_1)=0$, $\varphi(\alpha_2)=-\alpha_1$, and $\varphi(\alpha_3)=-\alpha_2$. Then $\varphi$
has a decomposition of real type $\Delta_5^+(0)\oplus\Delta_1^-(0)$ and minimal polynomial $P(x)=x^5$.
\end{example}
\begin{example}\label{ex:55}
Continuing with the notation of Example \ref{ex:54}, suppose now that $\varphi:\mathbb{T}_pM\rightarrow \mathbb{T}_p M$ is defined by
$\varphi(v_1)=v_1-2\alpha_2$, $\varphi(v_2)=-v_2+2\alpha_1$, $\varphi(v_3)=\alpha_3$,
$\varphi(\alpha_1)=-\alpha_1$, $\varphi(\alpha_2)=\alpha_2$, and $\varphi(\alpha_3)=v_3$. Then $\varphi$
has decomposition of real type $\Delta_2^0(1,-1)\oplus\Delta_1^0(1,-1)$ and minimal polynomial $P(x)=(x^2-1)^2$.
\end{example}
\begin{example}\label{ex:56}
Again in the notation of Example \ref{ex:54}, suppose that $\varphi:\mathbb{T}_pM\rightarrow \mathbb{T}_p M$ is such that
$\varphi(v_1)= v_2-\alpha_2$, $\varphi(v_2)= \alpha_1-\alpha_3$, $\varphi(v_3)=v_2+\alpha_2$,
$\varphi(\alpha_1)=0$, $\varphi(\alpha_2)=-\alpha_1-\alpha_3$, and $\varphi(\alpha_3)=0$. Then $\varphi$
has decomposition of type $\Delta^+_3(0)\oplus \Delta^-_3(0)$ and minimal polynomial $P(x)=x^3$.
\end{example}
\begin{rem} If $M$ is 3-dimensional and orientable, then it is parallelizable \cite{BenedettiLisca18}. Interpreting $\{v_1,v_2,v_3\}$ as a global frame and $\{\alpha_1,\alpha_2,\alpha_3\}$ as the associated dual frame, Examples \ref{ex:54}-\ref{ex:56} apply to $M$.
\end{rem}
\section{Weak generalized Nijenhuis operators}\label{sec:4}
In this section we investigate the compatibility of generalized polynomial structures with the Dorfman bracket or, equivalently, with the de Rham operator. Our main finding is that the semisimple part of a generalized polynomial structure controls the Dorfman involutivity of the generalized eigenbundles and the decomposition of the de Rham operator induced by the associated multigrading of $\Omega_M$.
\begin{mydef}
Let $\varphi$ be an endomorphism of $\mathbb{T} M$. The {\it Courant-Nijenhuis torsion of $\varphi$} is defined as
\begin{equation}
\mathcal T_\varphi(\mathbf{x},\mathbf{y})=\varphi^2{[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]} +{[}\!{[} \varphi(\mathbf{x}),\varphi(\mathbf{y}){]}\!{]} - \varphi({[}\!{[} \varphi(\mathbf{x}),\mathbf{y}{]}\!{]} +{[}\!{[} \mathbf{x},\varphi(\mathbf{y}){]}\!{]})
\end{equation}
for all $\mathbf{x},\mathbf{y}\in \mathbb{T} M$. Similarly, we define the {\it shifted Courant-Nijenhuis torsion of $\varphi$} as
\begin{equation}\label{eq:12}
\mathcal S_\varphi(\mathbf{x},\mathbf{y})=\mathcal T_\varphi(\varphi(\mathbf{x}),\mathbf{y})+ \mathcal T_\varphi(\mathbf{x},\varphi(\mathbf{y}))
\end{equation}
for all $\mathbf{x},\mathbf{y}\in \mathbb{T} M$.
\end{mydef}
\begin{mydef}[\cite{BoualemBrouzet12}] A {\it generalized Nijenhuis operator} is an endomorphism $\varphi$ of $\mathbb{T} M$ with vanishing Courant-Nijenhuis torsion. A {\it weak generalized Nijenhuis operator} is an endomorphism of $\mathbb{T} M$ with vanishing shifted Courant-Nijenhuis torsion.
\end{mydef}
\begin{rem}
If follows from \eqref{eq:12} that every generalized Nijenhuis operator is also a weak generalized Nijenhuis operator.
\end{rem}
\begin{example}[\cite{Gualtieri11}]\label{ex:52}
Let $\varphi$ be a generalized almost complex structure. Then $\varphi$ is a generalized complex structure if and only if it is a generalized Nijenhuis operator.
\end{example}
\begin{example}[\cite{BlagaCrasmareanu14}]\label{ex:53}
Let $\varphi$ be a weak generalized almost tangent structure. Then by definition $\varphi$ is a weak generalized tangent structure if and only if it is a generalized Nijenhuis operator (and not merely a weak generalized Nijenhuis operator, as the terminology might suggest).
\end{example}
\begin{rem}\label{rem:54}
Let $\varphi$ be a non-zero semisimple generalized polynomial structure. Then
\begin{equation}
\mathcal T_\varphi(\mathbf{x},\mathbf{y})=(\varphi-\mu I)(\varphi-\nu I){[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]}
\end{equation}
for all $\mathbf{x}\in L_\mu$, $\mathbf{y}\in L_\nu$. Since ${\rm Ker}((\varphi-\mu I)(\varphi-\nu I))=L_\mu+L_\nu$ we conclude that $\varphi$ is a generalized Nijenhuis operator if and only if ${[}\!{[} L_\mu,L_\nu{]}\!{]} \subseteq L_\mu+L_\nu$. In particular, taking $\mu=\pm\lambda$ we see that the quasi split structure $E_\lambda$ is closed under the Dorfman bracket. Since $\varphi$ is non-zero and semisimple we have $\lambda\neq 0$. On the other hand, Remark \ref{rem:13} implies that $E_\lambda=\mathbb{T} M\otimes \mathbb C$ and thus $\varphi$ has precisely two eigenvalues: $\pm \lambda$.
\end{rem}
\begin{rem}
Classically, integrability of polynomial structures is defined by requiring the vanishing of their Nijenhuis torsion. Thus, naively, one might be tempted to define a generalized polynomial structure $\varphi$ to be integrable if and only if $\varphi$ is a generalized Nijenhuis operator. Remark \ref{rem:54} shows that such a definition would be too restrictive, even for semisimple generalized polynomial structures with more than two eigenvalues. As noted in \cite{Kosmann-Schwarzbach19}, the necessity of relaxing the vanishing of the Nijenhuis torsion when working with general eigenbundles was already pointed in the classical case by Haantjes \cite{Haantjes55}. The connection between the shifted Courant-Nijenhuis torsion and the generalized analogues of the higher torsions introduced in \cite{Haantjes55} plays an important role in the reminder of this paper.
\end{rem}
\begin{theorem}\label{theorem:56}
Let $\varphi$ be a generalized polynomial structure. The following are equivalent:
\begin{enumerate}[1)]
\item the semisimple part of $\varphi$ is a weak generalized Nijenhuis operator;
\item $(\mu+\nu){[}\!{[} L_\mu,L_\nu{]}\!{]}\subseteq L_\mu+L_\nu$ for all $\mu,\nu\in \Sigma(\varphi)$;
\item there exists a decomposition of the de Rham operator $d=\sum_{\lambda\in \Sigma(\varphi)} \delta_\lambda$ such that for every $\lambda\in \Sigma(\varphi)\setminus\{0\}$, $\delta_\lambda$ is of degree $1$ with respect to the $(L_\lambda,L_{-\lambda})$-grading and of degree $0$ with respect to the $(L_\mu,L_{-\mu})$-grading whenever $\mu\notin\{0,\pm\lambda\}$.
\end{enumerate}
\end{theorem}
\begin{proof} Since the three statements only depend on the semisimple part of $\varphi$, we may (without loss of generality) assume that $\varphi$ is semisimple. With this assumption, the equivalence of 1) and 2) follows from the formula
\begin{equation}
\mathcal S_\varphi(\mathbf{x},\mathbf{y})=(\mu+\nu)(\varphi-\mu I)(\varphi-\nu I){[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]}\,,
\end{equation}
which is valid for all $\mathbf{x},\in L_\mu$, $\mathbf{y}\in L_\nu$. Assume that a decomposition of the de Rham operator as in 3) is given. Viewed as operators acting on $\Omega_M\otimes \mathbb C$, sections of $L_\lambda$ are of degree $-1$ with respect to the $(L_\lambda,L_{-\lambda})$-grading. Since the generalized tangent bundle is closed under the Dorfman bracket, given $\mathbf{x}\in L_\mu$, $\mathbf{y}\in L_\nu$ such that $\mu+\nu\neq 0$ we have
\begin{equation}
{[}\!{[} \mathbf{x}, \mathbf{y} {]}\!{]} = \sum_{\lambda \in \Sigma(\varphi)}{[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]}_{\delta_\lambda} = {[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]}_{\delta_\mu} + {[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]}_{\delta_\nu}\in L_\mu+L_\nu\, ,
\end{equation}
which implies 2). We are left to prove that 2) implies 30). The statement is trivial if $\varphi$ is nilpotent so we may use induction on the number of non-zero eigenvalues. Suppose $\lambda\in \Sigma(\varphi)\setminus\{0\}$. In particular, this implies that $L_{\pm\lambda}$ is closed under the Dorfman bracket. By Lemma \ref{lem:23}, $d$ has component of degree $1$, $-1$, and $0$ with respect to the $(L_\lambda,L_{-\lambda})$-grading. Define $\delta_{\pm\lambda}$ to be the components of degree $\pm 1$ and let $\delta=d-\delta_\lambda-\delta_{-\lambda}$. If $\pm \lambda$ are the only non-zero eigenvalues then we may set $\delta_0=\delta$ and the Theorem is proved. If there is another eigenvalue $\mu\notin\{0,\pm \lambda\}$, then using the closure of $L_\mu$ under the Dorfman bracket and Lemma \ref{lem:23} again we may decompose $\delta$ into component of degree $1$, $-1$, $0$ with respect to $(L_\mu,L_{-\mu})$-grading and, if necessary, iterate. Since the order in which the non-zero eigenvalues are selected is irrelevant, we are left to show that $\delta_\lambda$ is of degree $0$ with respect to the $(L_\mu,L_{-\mu})$-grading whenever $\mu\notin\{0,\pm \lambda\}$. To see this, we note that $\delta_\lambda$ is of degree $1$ with respect to the $(L_\lambda,L_{-\lambda})$-grading (which commutes with the $(L_\mu,L_{-\mu})$-grading). Together with 2), this shows that the the assumptions of Lemma \ref{lem:24} are met proving that $\delta_\lambda$ does indeed preserve the $(L_\mu,L_{-\mu})$-grading.
\end{proof}
\begin{rem}\label{rem:61}
Let $\varphi$ be a generalized polynomial structure with exactly two eigenvalues. This condition is equivalent to requiring that the minimal polynomial of $\varphi$ is of the form $P(x)=(x^2+c)^N$ for some non-zero real number $c$. Then the semisimple part $\varphi_s$ of $\varphi$ is a generalized Nijenhuis operator if and only if $\varphi_s$ is a weak generalized Nijenhuis operator if and only if $L_{\sqrt{-c}}$ and $L_{-\sqrt{-c}}$ are both closed under the Dorfman bracket. In the special case $N=c=1$, the equivalent conditions of Theorem \ref{theorem:56} are also equivalent to the statement that the generalized almost complex structure $\varphi=\varphi_s$ is integrable. Furthermore, in this case $\delta_{\sqrt{-1}}=\partial$ and $\delta_{-\sqrt{-1}}=\overline \partial$.
\end{rem}
\begin{example}\label{ex:58}
Let $\varphi$ be a generalized polynomial structure with exactly three eigenvalues, so that its minimal polynomial is of the form $P(x)=x^{N_1}(x^2+c)^{N_2}$ for some non-zero real number $c$. Then $\varphi_s$ is a generalized Nijenhuis operator if and only if
\begin{equation}\label{eq:15}
{[}\!{[} L_{\pm\sqrt{-c}},L_{\pm\sqrt{-c}}{]}\!{]} \subseteq L_{\pm\sqrt{-c}}\quad \textrm { and } \quad {[}\!{[} L_0,L_{\pm\sqrt{-c}}{]}\!{]} \subseteq L_0\oplus L_{\pm\sqrt{-c}}\,.
\end{equation}
Let us further specialize to the case $N_1=N_2=c=1$ i.e.\ to the case in which $\varphi=\varphi_s$ is a generalized $F$-structure. If $\varphi$ is a weak generalized Nijenhuis operator, then the first condition in \eqref{eq:15} shows that $\varphi$ is a generalized CRF structure. Generalized $F$-structure satisfying both conditions in \eqref{eq:15} are sometimes referred to \cites{PoonWade11,AldiGrandini15,AldiGrandini17} as {\it strongly integrable}. Furthermore, in this case the decomposition of the de Rham operator is the one discussed in \cite{AldiGrandini17}.
\end{example}
\begin{rem} Let $\varphi$ be a generalized polynomial structure. Assume $\varphi$ is a weak generalized Nijenhuis operator and let $d=\sum_\lambda \delta_\lambda$ be the decomposition of the de Rham operator whose existence is guaranteed by Theorem \ref{theorem:56}. Setting equal to zero the components of definite multidegree in $d^2=0$, we obtain
$(\mu+\nu)[\delta_\mu,\delta_\nu]=0$ for every $\lambda,\mu\in \Sigma(\varphi)$. In particular, the components of the de Rham operator corresponding to non-zero eigenvalues are differentials. We also have the additional equation
\begin{equation}
\delta_0^2=-\frac{1}{2}\sum_{\lambda\in \Sigma_+(\varphi)} [\delta_\lambda,\delta_{-\lambda}]\,.
\end{equation}
\end{rem}
\begin{rem}
In the classical case, Vanzurova \cite{Vanzurova98} restated and studied the integrability of semisimple polynomial structures in terms of a decomposition of the de Rham operator.
\end{rem}
\section{Minimality}\label{sec:5}
The goal of this section is to introduce a novel notion of compatibility between generalized polynomial structures and the de Rham operator that depends on both their semisimple and nilpotent parts.
\subsection{The minimal torsion and the Courant tensor}
Let $\varphi$ be a generalized polynomial structure on $M$ with minimal polynomial $P$ and let $\widetilde \varphi$ be the lift of $\varphi$ to $\mathcal E_M$, whose existence is guaranteed by Proposition \ref{prop:9}. We define the {\it minimal operator of $\varphi$} as $\delta_\varphi=P({\rm ad}_{\widetilde \varphi})(d)$, where $d$ is the de Rham operator.
\begin{example}
Let $\varphi$ be a generalized almost complex structure. Then $P(x)=x^2+1$ and $\delta_\varphi=[\widetilde \varphi,[\widetilde\varphi,d]]+d=-{[}\!{[} \widetilde \varphi,\widetilde\varphi{]}\!{]}+d$.
\end{example}
\begin{lem}\label{lem:40}
The minimal operator of a generalized polynomial structure is $\Omega_M^0$-linear.
\end{lem}
\begin{proof}
Since $[{\rm ad}_{\widetilde \varphi},f]=0$ for all $f\in \Omega_M^0$, then $[\delta_\varphi,f]=P({\rm ad}_{\widetilde\varphi})(df)=P(\varphi)(df)=0$
\end{proof}
\begin{mydef}
The {\it minimal torsion of $\varphi$}, denoted by $\mathcal M_\varphi:\mathord{\adjustbox{valign=B,totalheight=.6\baselineskip}{$\bigwedge$}}^2 \mathbb{T} M\to \mathbb{T} M$, is by definition the derived bracket with respect to the minimal operator of $\varphi$ i.e.\ $\mathcal M_\varphi(\mathbf{x},\mathbf{y})={[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]}_{\delta_\varphi}$ for all $\mathbf{x},\mathbf{y}\in \mathbb{T} M$. The {\it Courant tensor of $\varphi$} is the map $\mathcal C_\varphi:\mathord{\adjustbox{valign=B,totalheight=.6\baselineskip}{$\bigwedge$}}^3 \mathbb{T} M\to \Omega_M^0$ defined by the formula $\mathcal C_\varphi(\mathbf{x},\mathbf{y},\mathbf{z})=\langle \mathcal M_\varphi(\mathbf{x},\mathbf{y}),\mathbf{z}\rangle$ for all $\mathbf{x},\mathbf{y},\mathbf{z}\in \mathbb{T} M$.
\end{mydef}
\begin{rem}
More generally, if $\delta\in \mathcal E_M$ is an operator such that ${[}\!{[} \mathbb{T} M, \mathbb{T} M{]}\!{]}_\delta \subseteq \mathbb{T} M$ one can introduce the $\Omega_M^0$-valued ternary operation
\begin{equation}
T_\delta(\mathbf{x},\mathbf{y},\mathbf{z})=\langle {[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]}_\delta ,\mathbf{z} \rangle
\end{equation}
for all $\mathbf{x},\mathbf{y},\mathbf{z}\in \mathbb{T} M$. In the particular case $\delta=d$ one obtains the $\mathbb R$-trilinear map $T(\mathbf{x},\mathbf{y},\mathbf{z})=\langle {[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]} ,\mathbf{z} \rangle$ introduced in \cite{Courant90}.
\end{rem}
\begin{rem}\label{rem:68}
The definition of the minimal operator depends on the particular lift of $\varphi$ to $\mathcal E_M$ obtained by imposing condition 3) in Proposition \ref{prop:9}. If a different lift $\widetilde \varphi + g$ is chosen, for some function $g\in \Omega_M^0$, then $P({\rm ad}_{\widetilde \varphi+g})(d)=\delta_{\varphi}-P'(\varphi)(dg)$. In particular, the definition of the minimal operator depends on the choice of lift only up to addition of a section of $\mathbb{T} M$. Hence, by Lemma \ref{lemma:7}, the definition of the minimal torsion (and thus of the Courant tensor) of $\varphi$ is independent on condition 3) in Proposition \ref{prop:9}.
\end{rem}
\begin{rem}
Combining Lemma \ref{lem:8} and Lemma \ref{lem:40}, we see that $\mathcal M_\varphi$ is indeed skew-symmetric and $\Omega_M^0$-bilinear. An analogous calculation shows that $\mathcal C_\varphi$ is totally anti-symmetric and $\Omega_M^0$-trilinear.
\end{rem}
\begin{prop}\label{prop:44}
Let $\varphi$ be a generalized polynomial structure on $M$. Then the following conditions are equivalent:
\begin{enumerate}[1)]
\item $\delta_\varphi\in \mathbb{T} M$;
\item the minimal torsion of $\varphi$ vanishes;
\item the Courant tensor of $\varphi$ vanishes.
\end{enumerate}
\end{prop}
\begin{proof}
The equivalence between 2) and 3) is a consequence of the non-degeneracy of the tautological inner product. The equivalence of 1) and 2) is a consequence of Lemma \ref{lemma:7}.
\end{proof}
\begin{mydef}
A generalized polynomial structure is {\it minimal} if the equivalent conditions 1)-3) in Proposition \ref{prop:44} are satisfied.
\end{mydef}
\begin{prop}\label{prop:46}
Let $\varphi$ be a generalized polynomial structure with minimal polynomial $P(x)=a_Nx^N+\cdots +a_1x+a_0$. Then
\begin{equation}\label{eq:7}
\mathcal C_\varphi(\mathbf{x},\mathbf{y},\mathbf{z})=\sum_{i=0}^N \sum_{i_1+i_2+i_3=i} a_i (-1)^i\binom{i}{i_1,i_2,i_3} T(\varphi^{i_1}(\mathbf{x}),\varphi^{i_2}(\mathbf{y}),\varphi^{i_3}(\mathbf{z}))
\end{equation}
for all $\mathbf{x},\mathbf{y},\mathbf{z}\in \mathbb{T} M$.
\end{prop}
\begin{proof}
Assume $\delta\in \mathcal E_M$ is odd and such that ${[}\!{[} \mathbb{T} M,\mathbb{T} M{]}\!{]}_\delta \subseteq \mathbb{T} M$. Then, for all $\mathbf{x},\mathbf{y},\mathbf{z}\in \mathbb{T} M$, the operator
\begin{equation}\label{eq:39}
({\rm ad}_\mathbf{z} {\rm ad}_\mathbf{y}{\rm ad}_\mathbf{x} {\rm ad}_{\widetilde \varphi} + {\rm ad}_{\varphi(\mathbf{z})}{\rm ad}_\mathbf{y}{\rm ad}_\mathbf{x} + {\rm ad}_\mathbf{z}{\rm ad}_{\varphi(\mathbf{y})}{\rm ad}_\mathbf{x} + {\rm ad}_\mathbf{z}{\rm ad}_\mathbf{y}{\rm ad}_{\varphi(\mathbf{x})}-{\rm ad}_{\widetilde \varphi} {\rm ad}_\mathbf{z} {\rm ad}_\mathbf{y}{\rm ad}_\mathbf{x})(\delta)
\end{equation}
is equal to zero. The last term of \eqref{eq:39} vanishes because $({\rm ad}_\mathbf{z} {\rm ad}_\mathbf{y}{\rm ad}_\mathbf{x}) (\delta)\in \Omega_M^0$ and $[\widetilde \varphi,\Omega_M^0]=0$. Hence
\begin{equation}\label{eq:8}
T_{[\widetilde \varphi,\delta]}(\mathbf{x},\mathbf{y},\mathbf{z})+T_\delta (\varphi(\mathbf{x}),\mathbf{y},\mathbf{z})+T_\delta (\mathbf{x},\varphi(\mathbf{y}),\mathbf{z})+ T_\delta(\mathbf{x},\mathbf{y},\varphi(\mathbf{z}))=0\,.
\end{equation}
Iterating this step, we find
\[
T_{{\rm ad}^i_{\widetilde \varphi}(d)}(\mathbf{x},\mathbf{y},\mathbf{z}) = (-1)^i\sum_{i_1+i_2+i_3=i} \binom{i}{i_1,i_2,i_3}T(\varphi^{i_1}(\mathbf{x}),\varphi^{i_2}(\mathbf{y}),\varphi^{i_3}(\mathbf{z}))\, ,
\]
from which the result easily follows.
\end{proof}
\begin{rem}\label{rem:PolynomialAction}
Let $\varphi$ be an endomorphism of $\mathbb{T} M$. Following \cite{Kosmann-Schwarzbach19}, we consider the action of the polynomial ring $\mathbb C[u,v,w]$ on the space of local linear operators $\zeta:\mathbb{T} M\otimes \mathbb{T} M\to \mathbb{T} M$ defined by $(u\cdot \zeta)(\mathbf{x},\mathbf{y})=\varphi(\zeta(\mathbf{x},\mathbf{y}))$, $(v\cdot \zeta)(\mathbf{x},\mathbf{y})=\zeta(\varphi(\mathbf{x}),\mathbf{y})$, and $(w\cdot \zeta)(\mathbf{x},\mathbf{y})=\zeta(\mathbf{x},\varphi(\mathbf{y}))$. If $\zeta_{CD}$ is the Dorfman bracket, then Proposition \ref{prop:46} can be equivalently restated as $\mathcal M_\varphi=P(u-v-w)\cdot \zeta_{CD}$.
Note that that the above action is equivalent to the action of $\mathbb{C}[\varphi_1, \varphi_2, \varphi_3]$ on the sheaf $\mathbb{T} M^{\otimes 3}$ determined by
$\varphi_1 \cdot = \varphi \otimes {\rm Id}\otimes {\rm Id}, \quad \varphi_2 \cdot = {\rm Id}\otimes\varphi \otimes {\rm Id}, \quad \varphi_3 \cdot = {\rm Id} \otimes {\rm Id}\otimes\varphi$
so that, for all $\zeta:\mathbb{T} M\otimes \mathbb{T} M\to \mathbb{T} M$,
$\langle u^iu^ju^k\zeta(\mathbf{x},\mathbf{y}),\mathbf{z}\rangle = f((-\varphi_3)^i\varphi_1^j\varphi_2^k(\mathbf{x}\otimes\mathbf{y}\otimes \mathbf{z}))$
where $f(\mathbf{x}\otimes\mathbf{y}\otimes\mathbf{z})=\langle\zeta(\mathbf{x},\mathbf{y}),\mathbf{z}\rangle$.
Using this latter action, Proposition \ref{prop:46} reads $\mathcal C_\varphi=T\circ P(-\varphi_1-\varphi_2-\varphi_3)$.
\end{rem}
\begin{example}\label{ex:47}
Let $\varphi$ be a generalized almost complex structure. Then
\begin{equation}
\mathcal C_\varphi(\mathbf{x},\mathbf{y},\mathbf{z})=2(-T(\mathbf{x},\mathbf{y},\mathbf{z})+T(\varphi(\mathbf{x}),\varphi(\mathbf{y}),\mathbf{z})+T(\varphi(\mathbf{x}),\mathbf{y},\varphi(\mathbf{z}))+T(\mathbf{x},\varphi(\mathbf{y}),\varphi(\mathbf{z})))
\end{equation}
for all $\mathbf{x},\mathbf{y},\mathbf{z}\in \mathbb{T} M$. Using the skew-symmetry of $\varphi$ and non-degeneracy of the tautological inner product this is equivalent to $\mathcal M_\varphi=2 \mathcal T_\varphi$, where $\mathcal T_\varphi$ is the Courant-Nijenhuis torsion. In particular $\varphi$ is minimal if and only if it is a generalized complex structure.
\end{example}
\begin{example}\label{ex:48}
Let $\varphi$ be a generalized F-structure. The Courant torsion of $\varphi$ can be calculated using Proposition \ref{prop:46}. From there, using skew-symmetry and non-degeneracy as in Example \ref{ex:47}, we obtain $\mathcal M_\varphi=-3\mathcal S_{\varphi}$, where $\mathcal S_\varphi$ is the shifted Courant-Nijenhuis torsion of $\varphi$. Hence, thanks to Example \ref{ex:58}, we conclude that $\varphi$ is minimal if and only if it is strongly integrable.
\end{example}
\begin{prop}\label{prop:52}
Let $\varphi$ be a generalized polynomial structure with minimal polynomial $P$ of degree $N$. Consider Jordan chains $\{\mathbf{x}_\alpha\}\subseteq L_\lambda$, $\{\mathbf{y}_\beta\}\subseteq L_\mu$, and $\{\mathbf{z}_\gamma\}\subseteq L_\nu$, so that $\varphi(\mathbf{x}_\alpha)=\lambda \mathbf{x}_\alpha+\mathbf{x}_{\alpha-1}$, $\varphi(\mathbf{y}_\beta)=\mu \mathbf{x}_\beta+\mathbf{x}_{\beta-1}$, and $\varphi(\mathbf{z}_\gamma)=\nu \mathbf{x}_\gamma+\mathbf{x}_{\gamma-1}$. Then
\begin{equation}\label{eq:11}
\mathcal C_\varphi (\mathbf{x}_\alpha,\mathbf{y}_\beta,\mathbf{z}_\gamma)=(-1)^NP(\lambda+\mu+\nu+S)T(\mathbf{x}_\alpha,\mathbf{y}_\beta,\mathbf{z}_\gamma)\, ,
\end{equation}
where $S$ is the shift operator defined by the formula
\begin{equation}\label{eq:43}
S T(\mathbf{x}_\alpha,\mathbf{y}_\beta,\mathbf{z}_\gamma)=T(\mathbf{x}_{\alpha-1},\mathbf{y}_\beta,\mathbf{z}_\gamma)+ T(\mathbf{x}_{\alpha}, \mathbf{y}_{\beta-1},\mathbf{z}_\gamma)+T(\mathbf{x}_\alpha,\mathbf{y}_\beta,\mathbf{z}_{\gamma-1})\,.
\end{equation}
\end{prop}
\begin{proof}
Consider a factorization $P(x)=\prod_{i=1}^N (x-\omega_i)$ with $\omega_i\in \Sigma(\varphi)$ not necessarily distinct. Using \eqref{eq:8} we have
\begin{equation}
T_{({\rm ad}_{\widetilde\varphi}-\omega_i)(\delta)}(\mathbf{x}_\alpha,\mathbf{y}_\beta\,\mathbf{z}_\gamma)=-(\lambda+\mu+\nu+\omega_i+S)T_\delta(\mathbf{x}_\alpha,\mathbf{y}_\beta,\mathbf{z}_\gamma)\, .
\end{equation}
Keeping into account that $\Sigma(\varphi)$ is symmetric with respect to $0\in \mathbb C$, and iterating over all factors of $P$, yields the result.
\end{proof}
\begin{cor}\label{cor:88}
Let $\varphi$ be a generalized polynomial structure with minimal polynomial $P(x)$. If $\mathbf{x}\in {\rm Ker}((\varphi-\lambda I)^p)$ and $\mathbf{y}\in{\rm Ker}((\varphi-\mu I)^q)$ then
\begin{equation}
P^{p+q-1}((\lambda+\mu) I-\varphi)({[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]})=0\,
\end{equation}
\end{cor}
\begin{proof}
Let $\mathbf S$ be the shift operator defined on derived brackets of Jordan chains by $\mathbf S {[}\!{[} \mathbf x_\alpha,\mathbf y_\beta{]}\!{]}_\delta={[}\!{[} \mathbf x_{\alpha-1},\mathbf y_\beta{]}\!{]}_\delta+{[}\!{[} \mathbf x_\alpha,\mathbf y_{\beta-1}{]}\!{]}_\delta$. Then we have the Taylor expansion
\begin{equation}
P^r((\lambda+\mu)I-\varphi+\mathbf S)=\sum_{s=0}^\infty (P^r)^{(s)}((\lambda+\mu) I -\varphi)\frac{\mathbf S^s}{s!}
\end{equation}
for every integer $r\ge 1$. The result then follows upon using induction on $p+q$ and the fact that $P$ divides $(P^r)^{(s)}$ for every $r>s$.
\end{proof}
\begin{prop}\label{prop:87}
Let $\varphi$ be a minimal generalized polynomial structure. Then the semisimple and nilpotent parts of $\varphi$ are both minimal.
\end{prop}
\begin{proof}
It follows from $\eqref{eq:44}$ that the $\varphi_s=Q_s(\varphi)$ for some polynomial $Q_s$, which also implies $\varphi_n=Q_n(\varphi)$ with $Q_n(x)=x-Q_s(x)$. Let $P_s$ be the minimal polynomial of $\varphi_s$. Then $0=P_s(\varphi_s)=P_s(Q_s(\varphi))$ and thus $P_s(Q_s(x))$ is divisible by the minimal polynomial of $\varphi$. Hence \begin{equation}
P_s({\rm ad}_{\widetilde {\varphi_s}})(d)=P_s(Q_s({\rm ad}_{\widetilde\varphi}))(d)
\end{equation}
is a section of $\mathbb T M$, proving that $\varphi_s$ is minimal. A completely analogous argument shows that $\varphi_n$ is minimal.
\end{proof}
\begin{cor}\label{cor:89}
Let $\varphi$ be a generalized polynomial structure such that all the eigenvalues have the same multiplicity. Then $\varphi$ is minimal if and only if its semisimple and nilpotent parts are both minimal.
\end{cor}
\begin{proof}
If $\varphi$ is minimal, then $\varphi_s$ and $\varphi_n$ are both minimal by Proposition \ref{prop:87}. For the converse, let $k$ be the common multiplicity of all eigenvalues and let $P_s(x)$ be the minimal polynomial of $\varphi_s$ so that $P(x)=P_s^k(x)$ is the minimal polynomial of $\varphi$. If $\varphi_n$ is minimal, then ${\rm ad}_{\widetilde{\varphi_n}}^j(d)\in \mathbb{T} M$ for all $j\ge k$ and thus, upon taking the Taylor expansion,
\begin{equation}
P({\rm ad}_{\widetilde\varphi})(d)-P({\rm ad}_{\widetilde{\varphi_s}})(d)-\cdots-P^{(k-1)}({\rm ad}_{\widetilde{\varphi_s}}){\rm ad}^{k-1}_{\widetilde{\varphi_n}}(d)\in \mathbb{T} M.
\end{equation}
On the other hand, $P_s(x)$ divides $P^{(i)}(x)$ for all $0\le i<k$ which concludes the proof.
\end{proof}
\begin{prop}\label{prop:88}
Let $\varphi$ be a semisimple generalized polynomial structure. If $\varphi$ is a weak generalized Nijenhuis operator, then $\varphi$ is minimal.
\end{prop}
\begin{proof}Assume $\varphi$ has minimal polynomial $P(x)$ of degree $N$.
Due to semisimplicity, the shift operator $S$ in \eqref{eq:11} acts trivially. For every $\mathbf{x}\in L_\lambda$, $\mathbf{y}\in L_\mu$, and $\mathbf{z}\in L_\nu$ we have
\begin{equation}\label{eq:47}
\mathcal C_{\varphi}(\mathbf{x},\mathbf{y},\mathbf{z})=(-1)^N P(\lambda +\mu+\nu)T(\mathbf{x},\mathbf{y},\mathbf{z})
\end{equation}
where $P(\lambda +\mu+\nu)T(\mathbf{x},\mathbf{y},\mathbf{z})$ is the ordinary product of real numbers. By Theorem \ref{theorem:56}, $T(\mathbf{x},\mathbf{y},\mathbf{z})=0$ unless $(\lambda+\mu)(\lambda+\nu)(\mu+\nu)=0$ in which case $P(\lambda +\mu+\nu)=0$. Hence the Courant tensor vanishes identically and $\varphi$ is minimal.
\end{proof}
\subsection{Higher Courant-Nijenhuis torsions}
Let $\varphi$ be an operator acting on $\mathbb T M$. Following \cite{Kosmann-Schwarzbach19} we define the {\it higher Courant-Nijenhuis torsions} $\mathcal T_\varphi^{(n)}$, for any positive integer $n$, as follows. If $n=1$ set $\mathcal T_\varphi^{(1)}=\mathcal T_\varphi$. If $n\ge 1$ we define $\mathcal T_\varphi^{(n)}$ recursively by the formula
\begin{equation}\label{eq:17}
\mathcal T_\varphi^{(n+1)}(\mathbf{x},\mathbf{y})=\varphi^2 \mathcal T_\varphi^{(n)}(\mathbf{x},\mathbf{y})+\mathcal T_\varphi^{(n)}(\varphi(\mathbf{x}),\varphi(\mathbf{y}))-\varphi \mathcal S_\varphi^{(n)}(\mathbf{x},\mathbf{y})
\end{equation}
for all $\mathbf{x},\mathbf{y}\in \mathbb{T} M$ where, for every $n\ge 1$,
\begin{equation}
\mathcal S_\varphi^{(n)}(\mathbf{x},\mathbf{y})=\mathcal T_\varphi^{(n)}(\varphi(\mathbf{x}),\mathbf{y})+\mathcal T_\varphi^{(n)}(\mathbf{x},\varphi(\mathbf{y}))
\end{equation}
is the $n$-th {\it shifted higher Courant-Nijenhuis torsion} of $\varphi$. It is convenient to extend this definition and set $\mathcal T_\varphi^{(0)}(\mathbf{x},\mathbf{y})={[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]}$ and $\mathcal S^{(0)}_\varphi (\mathbf{x},\mathbf{y})={[}\!{[} \varphi(\mathbf{x}),\mathbf{y}{]}\!{]}+{[}\!{[} \mathbf{x}, \varphi(\mathbf{y}){]}\!{]}$ for all $\mathbf{x},\mathbf{y}\in \mathbb{T} M$.
\begin{rem}\label{rem:65}
Let $f:TM\to TM$ be a polynomial structure and let $\varphi=f\oplus (-f^T)$ the associated generalized polynomial structure. Then $\mathcal T_\varphi^{(m)}(X,Y)=\tau_f^{(m)}(X,Y)$ for all $X,Y\in TM$, where $\tau_f^{(m)}$ is the $m$-th higher Nijenhuis torsion of $f$ as defined in \cite{TempestaTondo18a}. In particular, $\tau^{(1)}_f$ and $\tau^{(2)}_f$ are respectively the ordinary Nijenhuis torsion and the Haantjes torsion of $f$.
\end{rem}
\begin{example}
If $n=2$ we have a Courant algebroid analogue of the Haantjies tensor
\begin{align*}
\mathcal T_\varphi^{(2)}(\mathbf{x},\mathbf{y})=&\varphi^4({[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]}) + {[}\!{[} \varphi^2(\mathbf{x}),\varphi^2(\mathbf{y}){]}\!{]} + \varphi^2({[}\!{[} \varphi^2(\mathbf{x}),\mathbf{y}{]}\!{]} +{[}\!{[} \mathbf{x},\varphi^2(\mathbf{y}){]}\!{]}+4{[}\!{[} \varphi(x),\varphi(\mathbf{y}){]}\!{]}) \\
&-2\varphi(\varphi^2({[}\!{[} \varphi(\mathbf{x}),\mathbf{y}{]}\!{]}+{[}\!{[} \mathbf{x},\varphi(\mathbf{y}){]}\!{]})+{[}\!{[} \varphi^3(\mathbf{x}),\mathbf{y}{]}\!{]}+{[}\!{[} \mathbf{x},\varphi^3(\mathbf{y}){]}\!{]})
\end{align*}
for which we propose the terminology {\it Courant-Haantjies tensor}.
\end{example}
\begin{rem}\label{rem:56}
Using the notation introduced in Remark \ref{rem:PolynomialAction}, we have
\begin{equation}
\mathcal T_\varphi^{(n)}=(u^2-uv-uw+vw)^n\cdot\zeta_{CD}=(u-v)^n(u-w)^n\cdot \zeta_{CD}\, .
\end{equation}
Therefore,
\begin{equation}
\mathcal T_\varphi^{(n)}(\mathbf{x},\mathbf{y})=\sum_{i,j=0}^n (-1)^{i+j}\binom{n}{i}\binom{n}{j}\varphi^{i+j}{[}\!{[} \varphi^{n-i}(\mathbf{x}),\varphi^{n-j}(\mathbf{y}){]}\!{]}
\end{equation}
for all $\mathbf{x},\mathbf{y}\in \mathbb{T} M$. Similarly, from
\begin{equation}
\mathcal S_\varphi^{(n+1)}=(u-v)^n(u-w)^n(u^2-uv-uw+vw)(v+w)\cdot \zeta_{CD}
\end{equation}
we obtain
\begin{equation}
\mathcal S_\varphi^{(n+1)}(\mathbf{x},\mathbf{y})=\sum_{i,j=0}^{n}(-1)^{i+j}\binom{n}{i}\binom{n}{j}\varphi^{i+j}\mathcal S_\varphi(\varphi^{n-i}(\mathbf{x}),\varphi^{n-j}(\mathbf{y}))
\end{equation}
for all $\mathbf{x},\mathbf{y}\in \mathbb{T} M$.
\end{rem}
\begin{theorem}\label{theorem:57}
Let $\varphi$ be a generalized polynomial structure with minimal polynomial $P(x)$. If $P(x)=a_0+a_2x^2+\cdots+a_{2N}x^{2N}$ then
\begin{equation}\label{eq:20}
\mathcal M_\varphi(\mathbf{x},\mathbf{y})=2\sum_{n=1}^{N} \sum_{m=1}^n \sum_{r=0}^{2n-2m} a_{2n} C_{n,m,r}\mathcal T_\varphi^{(m)}(\varphi^r(\mathbf{x}),\varphi^{2n-2m-r}(\mathbf{y}))
\end{equation}
for all $\mathbf{x},\mathbf{y}\in \mathbb{T} M$, where
\begin{equation}\label{eq:21}
C_{n,m,r}=\sum_{k=m}^n\sum_{q=2n-2m-r}^{2n-2k-r}(-1)^{q+r}4^{m-n}\binom{2n}{2k}\binom{k}{m}\binom{2k-2m}{2n-2m-q-r}\binom{2n-2k}{q}\,.
\end{equation}
If $P(x)=a_1x+a_3x^3+\cdots+a_{2N+1}x^{2N+1}$, then
\begin{equation}\label{eq:22}
\mathcal M_\varphi(\mathbf{x},\mathbf{y})=-\sum_{n=1}^{N} \sum_{m=1}^n \sum_{r=0}^{2n-2m} a_{2n+1} D_{n,m,r}\mathcal S_\varphi^{(m)}(\varphi^r(\mathbf{x}),\varphi^{2n-2m-r}(\mathbf{y}))
\end{equation}
for all $\mathbf{x},\mathbf{y}\in \mathbb{T} M$, where
\begin{equation}\label{eq:23}
D_{n,m,r}=\sum_{k=m}^n\sum_{q=2n-2m-r}^{2n-2k-r}(-1)^{q+r}4^{m-n}\binom{2n+1}{2k}\binom{k}{m}\binom{2k-2m}{2n-2m-q-r}\binom{2n-2k}{q}\,.
\end{equation}
\end{theorem}
\begin{proof} Let $\mathcal M_\varphi^{(n)}$ denote the derived bracket of the operator $({\rm ad}_{\widetilde\varphi}^n+(-1)^n{\rm ad}_{\widetilde{\varphi^n}})(d)$. Assume $P(x)=a_0+a_2x^2+\cdots+a_{2N}x^{2N}$. Then, by Remark \ref{rem:PolynomialAction}, the quantity
\begin{equation}
\mathcal M_\varphi(\mathbf{x},\mathbf{y})-\sum_{n=1}^N a_{2n} \mathcal M^{(2n)}_\varphi(\mathbf{x},\mathbf{y})
\end{equation}
is equal to
\begin{equation}
a_0{[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]}+\sum_{n=1}^N a_{2n}( \varphi^{2n}({[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]})-{[}\!{[}\varphi^{2n}(\mathbf{x}),\mathbf{y} {]}\!{]} -{[}\!{[} \mathbf{x},\varphi^{2n}(\mathbf{y}){]}\!{]})=0\, .
\end{equation}
Using the notation of Remark \ref{rem:PolynomialAction} and setting $z=2u-v-w$, we have
\begin{align*}
\mathcal M_\varphi^{(2n)} &= ((u-v-w)^{2n}+u^{2n}-v^{2n}-w^{2n})\cdot \zeta_{CD}\\
& = \left( \left(\frac{z-v-w}{2} \right)^{2n}+\left(\frac{z+v+w}{2} \right)^{2n} - v^{2n}-w^{2n} \right) \cdot \zeta_{CD}\\
&= \left(2^{1-2n}\sum_{k=0}^n \binom{2n}{2k}z^{2k}(v+w)^{2n-2k}-v^{2n}-w^{2n} \right) \cdot \zeta_{CD}\,.
\end{align*}
Setting $t=u^2-u(v+w)+v w$ we have $\mathcal T_\varphi^{(m)}=t^m\cdot \zeta_{CD}$. On the other hand, $z^2=4t+(v-w)^2$ from which we obtain
\begin{align*}
\mathcal M_\varphi^{(2n)}&=\left(2^{1-2n}\sum_{k=0}^n\binom{2n}{2k}(v+w)^{2n-2k}\sum_{m=0}^k 4^mt^m(v-w)^{2k-2m}-v^{2n}-w^{2n}\right)\cdot \zeta_{CD}\\
&=\left(2\sum_{m=0}^n 4^{m-n}t^m\sum_{k=m}^n\binom{2n}{2k}\binom{k}{m}(v-w)^{2k-2m}(v+w)^{2n-2k}-v^{2n}-w^{2n} \right) \cdot \zeta_{CD}\\
&=\left(\sum_{m=1}^n 2^{2m-2n+1}t^m\sum_{k=m}^n\binom{2n}{2k}\binom{k}{m}(v-w)^{2k-2m}(v+w)^{2n-2k} \right) \cdot \zeta_{CD}\, ,
\end{align*}
where the last equality follows from the identity
\begin{equation}
v^{2n}+w^{2n}=2^{1-2n}\sum_{k=0}^n\binom{2n}{2k}(v-w)^{2k}(v+w)^{2n-2k}\,.
\end{equation}
By expanding $(v-w)^{2k-2m}(v+w)^{2n-2k}$ and then summing over $n$ we obtain \eqref{eq:20} and \eqref{eq:21}. The proof of \eqref{eq:22} and \eqref{eq:23} is analogous: starting with
\begin{equation}
\mathcal M_\varphi^{(2n+1)} = ((u-v-w)^{2n+1}-u^n+v^n+w^n)\cdot \zeta_{CD}
\end{equation}
one repeats the changes of variables as above and then, using the representation $\mathcal S_\varphi^{(m)}=t^m(v+w)\cdot \zeta_{CD}$, obtain the desired expansion.
\end{proof}
\begin{example}\label{ex:69}
By Remark \ref{rem:34}, if the minimal polynomial $P(x)$ is quadratic it has to be of the form $x^2+a_0$ for some real number $a_0$. Substituting into \eqref{eq:20} we obtain $\mathcal M_\varphi = 2\mathcal T_\varphi$, independently of $a_0$. We conclude that a generalized polynomial structure of degree 2 is minimal if and only if it is a generalized Nijenhuis operator. In particular, for $a_0=1$ we recover the the observation made in Example \ref{ex:47}.
\end{example}
\begin{example}\label{ex:87}
Similarly, if the minimal polynomial $P(x)$ is cubic it has to be of the form $x^3+a_1x$ for some real number $a_1$. Substituting into \eqref{eq:22} we obtain $\mathcal M_\varphi=-3\mathcal S_\varphi$ independently of $a_1$. We conclude that a generalized polynomial structure of degree 3 is minimal if and only if it is a weak generalized Nijenhuis operator.
\end{example}
\begin{cor}\label{cor:90}
Let $\varphi$ be generalized polynomial structure of degree $4$ with exactly two eigenvalues. Then $\varphi$ is minimal if and only its semisimple and nilpotent parts are both generalized Nijenhuis operators.
\end{cor}
\begin{proof}
By assumption the minimal polynomial of $\varphi$ is of the form $(x^2+c)^2$ for some non-zero $c$. Hence the minimal polynomials of the semisimple and nilpotent part are, respectively, $x^2+c$ and $x^2$. In particular they are both quadratic and the statement follows by combining Corollary \ref{cor:89} and Example \ref{ex:69}.
\end{proof}
\begin{cor}
There is a canonical bijection between the set of minimal generalized polynomial structures with minimal polynomial $P(x)=(x^2+1)^2$ and the set of pairs $(\mathcal J_1,\mathcal J_2)$ consisting of a generalized complex structure $\mathcal J_1$ and a weak generalized tangent structure $\mathcal J_2$ such that $[\mathcal J_1,\mathcal J_2]=0$.
\end{cor}
\begin{proof}
Given a generalized polynomial structure $\varphi$ with minimal polynomial $P(x)=(x^2+1)^2$, then $\mathcal J_1=\varphi_s$ is a generalized almost complex structure and $\mathcal J_2=\varphi_n$ is a weak generalized almost tangent structure. Conversely, if $\mathcal J_1$ and $\mathcal J_2$ are as above then $\varphi=\mathcal J_1+\mathcal J_2$ is a generalized polynomial structure with minimal polynomial $P(x)=(x^2+1)^2$. The result then follows from Corollary \ref{cor:90}.
\end{proof}
\begin{cor}\label{cor:71}
Let $\varphi$ be a generalized polynomial structure with minimal polynomial $P(x)$ of degree at least 4. If $P(x)=a_0+a_2 x^2+\cdots+a_{2N-2}x^{2N-2}+x^{2N}$ then $\varphi$ is minimal if and only if
\begin{equation}
\mathcal T_\varphi^{(N)}(\mathbf{x},\mathbf{y})=-\sum_{n=1}^{N-1}\sum_{m=1}^n\sum_{r=0}^{2n-2m}a_{2n}C_{n,m,r}\mathcal T_\varphi^{(m)}(\varphi^r(\mathbf{x}),\varphi^{2n-2m-r}(\mathbf{y}))
\end{equation}
for all $\mathbf{x},\mathbf{y}\in \mathbb{T} M$, where the coefficients $C_{n,m,r}$ are defined by \eqref{eq:21}. If $P(x)=a_1x+a_3x^3+\cdots+a_{2N}x^{2N}+x^{2N+1}$ then $\varphi$ is minimal if and only if
\begin{equation}\label{eq:35}
\mathcal S_\varphi^{(N)}(\mathbf{x},\mathbf{y})=-\frac{1}{2N+1}\sum_{n=1}^{N-1}\sum_{m=1}^n\sum_{r=0}^{2n-2m}a_{2n+1}D_{n,m,r}\mathcal S_\varphi^{(m)}(\varphi^r(\mathbf{x}),\varphi^{2n-2m-r}(\mathbf{y}))
\end{equation}
for all $\mathbf{x},\mathbf{y}\in \mathbb{T} M$, where the coefficients $D_{n,m,r}$ are defined by \eqref{eq:23}.
\end{cor}
\begin{example}
If $P(x)$ is quartic, then $P(x)=x^4+a_2x^2+a_0$ for some real numbers $a_2$ and $a_0$. In this case the minimal torsion depends on $a_2$ but not on $a_0$. Moreover, $\varphi$ is minimal if and only if the Courant-Haantjes torsion satisfies
\begin{equation}
\mathcal T^{(2)}_\varphi(\mathbf{x},\mathbf{y})=-2(\mathcal T_\varphi(\varphi(\mathbf{x}),\varphi(\mathbf{y}))+\mathcal T_\varphi(\varphi^2(\mathbf{x}),\mathbf{y})+\mathcal T_\varphi(\mathbf{x},\varphi^2(\mathbf{y})))-a_2\mathcal T_\varphi(\mathbf{x},\mathbf{y})\,
\end{equation}
for all $\mathbf{x},\mathbf{y}\in \mathbb{T} M$.
\end{example}
\begin{cor}\label{cor:74}
Let $\varphi$ be a generalized polynomial structure.
\begin{enumerate}[1)]
\item If $\varphi$ is generalized Nijenhuis operator, then $\varphi$ is minimal.
\item If $\varphi$ is of odd degree and a weak generalized Nijenhuis operator, then it is minimal.
\end{enumerate}
\end{cor}
\begin{proof}
It follows from \eqref{eq:17} that $\mathcal T_\varphi=0$ implies $\mathcal T_\varphi^{(m)}=0$ for every $m\ge 1$. Then 1) follows from Theorem \ref{theorem:57}. The proof of 2) is similar and left to the reader.
\end{proof}
\begin{example}
Let $f:TM\to TM$ be a polynomial structure and let $\varphi=f\oplus (-f^T)$ be the associated generalized polynomial structure. Since the Dorfman bracket vanishes on $T^*M$ and preserves $TM$, then (using the anti-symmetry of the Courant tensor) we conclude that $\varphi$ is minimal if and only if $\mathcal C_\varphi(TM,TM,T^*M)=0$. Let us further assume assume that $f$ is integrable or, equivalently, that the Nijenhuis torsion of $f$ vanishes i.e.\ $\mathcal T_\varphi(TM,TM)=0$. Hence, invoking Theorem \ref{theorem:57}, the integrability of $f$ implies the minimality of $\varphi$. More generally, combining Remark \ref{rem:65} and Corollary \ref{cor:71} we obtain a remarkable characterization of minimality for generalized polynomial structures of the form $\varphi=f\oplus (-f^T)$ purely in terms of (ordinary) higher Nijenhuis torsions. Namely, if the minimal polynomial of $\varphi$ is $P(x)=a_0+a_2 x^2+\cdots+a_{2N-2}x^{2N-2}+x^{2N}$, then minimality is equivalent to the $N$-th higher Nijenhuis torsion being expressed as linear combination with constant coefficients of the lower order torsions according to the formula
\begin{equation}\label{eq:33}
\tau_f^{(N)}(X,Y)=-\sum_{n=1}^{N-1}\sum_{m=1}^n\sum_{r=0}^{2n-2m}a_{2n}C_{n,m,r}\tau_f^{(m)}(f^r(X),f^{2n-2m-r}(Y))
\end{equation}
for all $X,Y\in TM$. Similarly, if the minimal polynomial of $\varphi$ is of the form $P(x)=a_1x+a_3x^3+\cdots+a_{2N}x^{2N}+x^{2N+1}$, then $\varphi$ is minimal if and only if the restriction of \eqref{eq:35} to sections of $TM$ holds. It is worthwhile to note that while \eqref{eq:33} expresses minimality purely in terms of $f$ and its higher Nijenhuis torsions, the $a_k$ are the coefficients of the minimal polynomial of $\varphi$ which, as pointed out in Remark \ref{rem:29}, is in general not equal to the minimal polynomial of $f$. Hence, in the case of generalized polynomial structures defined by classical polynomial structures, minimality amounts to a novel relationship between the higher Nijenhuis torsions that, to the best of our knowledge, appears to be natural only from the vantage viewpoint of the generalized tangent bundle.
\end{example}
\begin{cor}
Let $\varphi$ be a non-zero semisimple polynomial structure. Then $\varphi$ is a generalized Nijenhuis operator if and and only if $\varphi$ is minimal with exactly two eigenvalues.
\end{cor}
\begin{proof} One direction follows from Corollary \ref{cor:74} and Remark \ref{rem:54}. For the converse, let us assume that $\varphi$ is semisimple with exactly two eigenvalues. Then its minimal polynomial is of the form $P(x)=x^2-\lambda^2$ for some non-zero real number $\lambda$. Hence, by Example \ref{ex:47} we have $0=\mathcal M_\varphi= 2\mathcal T_\varphi$ forcing $\varphi$ to be a generalized Nijenhuis operator.
\end{proof}
\begin{rem}
Let $\varphi$ be a generalized polynomial structure and let $\{\mathbf{x}_\alpha\}\subseteq L_\lambda$, $\{\mathbf{y}_\beta\}\subseteq L_\mu$ be Jordan chains. In the notation of Remark \ref{rem:PolynomialAction}, using the identity $(u-uv-uw+vw)=(u-v)(u-w)$ as in \cite{Kosmann-Schwarzbach19} we immediately obtain
\begin{equation}\label{eq:30}
\mathcal T_\varphi^{(m)}(\mathbf{x}_\alpha,\mathbf{y}_\beta)=\sum_{i,j=0}^{m} (-1)^{i+j}\binom{m}{i}\binom{m}{j}(\varphi-\lambda I)^{m-i}(\varphi-\mu I)^{m-j}{[}\!{[} \mathbf{x}_{\alpha-i},\mathbf{y}_{\beta-j}{]}\!{]}
\end{equation}
which generalizes Proposition 29 in \cite{TempestaTondo18a} to sections of $TM$. Because of this, many of the results of \cite{TempestaTondo18a} still hold if $TM$ is extended to $\mathbb{T} M$ and the Lie bracket is extended to the Dorfman bracket.
\end{rem}
\begin{example} Let $\varphi$ be a generalized polynomial structure with minimal polynomial of the form $P(x)=x^r Q(x)$ for some real polynomial $Q$ such that $Q(0)\neq 0$. It follows from \eqref{eq:30} that if $\mathbf{x},\mathbf{y}\in {\rm Ker}(\varphi)$, then $\mathcal T_\varphi^{(m)}(\mathbf{x},\mathbf{y})=\varphi^{2m}{[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]}$. Replacing $\varphi$ with $\varphi^{r}$ we see that $L_0$ is closed under the Dorfman bracket if and only if the restriction of $\mathcal T_{\varphi^r}^{(m)}$ to $L_0$ vanishes.
\end{example}
\begin{prop}\label{prop:72}
Let $\varphi$ be a generalized polynomial structure and let $\lambda\in \Sigma(\varphi)$. If there exists $m\ge 1$ such that $\mathcal T_\varphi^{(m)}(L_\lambda,L_\lambda)=0$, then $L_\lambda$ is closed under the Dorfman bracket.
\end{prop}
\begin{proof}
Thanks to \eqref{eq:30}, the proof of the corresponding statement (Proposition 42) in \cite{TempestaTondo18a} can be adapted step-by-step without difficulty.
\end{proof}
\begin{cor}\label{cor:72}
Let $\varphi$ be a generalized polynomial structure that is neither invertible nor nilpotent. Then none of the higher Courant-Nijenhuis torsions of $\varphi$ vanishes identically. In particular, $\varphi$ is not a generalized Nijenhuis operator.
\end{cor}
\begin{proof}
Combine Proposition \ref{prop:72} and Corollary \ref{cor:42}.
\end{proof}
\begin{rem}
Proposition \ref{prop:72} can be generalized as follows (see Theorem 45 in \cite{TempestaTondo18a}). If $\lambda_1,\ldots,\lambda_k$ are eigenvalues of $\varphi$ such that the restriction of one of the higher Courant-Nijenhuis torsions of $\varphi$ to $L=L_{\lambda_1}\oplus \cdots\oplus L_{\lambda_k}$ vanishes, then $L$ is closed under the Dorfman bracket. In light of Remark \ref{rem:13}, we observe that this statement (and, in particular, Proposition \ref{prop:72}) is in general false if the Dorfman bracket is replaced by the (antisymmetric) Courant bracket.
\end{rem}
\subsection{Bivariate Courant-Nijenhuis torsion}
\begin{mydef}
The {\it bivariate Courant-Nijenhuis torsion} of two commuting operators $\varphi'$ and $\varphi''$ acting on $\mathbb{T} M$ is defined as $\mathcal T_{\varphi'+\varphi''}-\mathcal T_{\varphi'}-\mathcal T_{\varphi''}$.
\end{mydef}
\begin{prop}\label{prop:93}
Let $\varphi'$, $\varphi''$ be generalized polynomial structures. Then $\varphi=\varphi'+\varphi''$ is minimal provided that the following conditions hold:
\begin{enumerate}[1)]
\item $\varphi'\circ \varphi''=0=\varphi''\circ \varphi'$;
\item $\varphi'$ and $\varphi''$ are both minimal;
\item $\varphi'$ and $\varphi''$ both annihilate the minimal polynomial of $\varphi$;
\item the bivariate Courant-Nijenhuis torsion of $\varphi'$ and $\varphi''$ vanishes.
\end{enumerate}
\end{prop}
\begin{proof} Let $P(x)=a_0+a_1x+\cdots+a_Nx^N$ be the minimal polynomial of $\varphi$.
We begin by noticing that 1) implies $[\varphi',\varphi'']=0$ and thus their bivariate Courant-Nijenhuis torsion as defined above is well defined. Adapting the notation of Remark \ref{rem:56} we denote by $\{u,v,w\}$, $\{u',v',w'\}$, and $\{u'',v'',w''\}$ generators of the polynomial actions on tensors defined, respectively, by $\varphi$, $\varphi'$, and $\varphi''$. In particular, $u=u'+u''$, $v=v'+v''$, and $w=w'+w''$. From 1), we see that $u'u''=v'v''=w'w''=0$. Using this we have
\begin{align}
\mathcal T_{\varphi'+\varphi''}-\mathcal T_{\varphi'}-\mathcal T_{\varphi''}&=((u-v)(u-w)-(u'-v')(u'-w')-(u''-v'')(u''-w''))\cdot \zeta_{CD}\nonumber\\
& = (u'-v'-w')(u''-v''-w'')\cdot \zeta_{CD}\label{eq:51}
\end{align}
and thus
\begin{equation}
0=((u-v-w)^n-(u'-v'-w')^n-(u''-v''-w'')^n)\cdot \zeta_{CD}
\end{equation}
for all $n\ge 1$. Hence
\begin{equation*}
\mathcal M_\varphi^{(n)}-\mathcal M_{\varphi'}^{(n)}-\mathcal M_{\varphi''}^{(n)}=\pm (u^n-v^n-w^n-(u')^n+(v')^n+(w')^n-(u'')^n+(v'')^n+(w'')^n)\cdot \zeta_{CD}.
\end{equation*}
By 3), we have $P(u)=P(u')=\cdots=P(w'')=0$. Hence
\begin{equation}\label{eq:46}
\mathcal M_\varphi = \sum_{k=0}^n a_k \mathcal M_{\varphi'}^{(k)}+\sum_{k=0}^n a_k \mathcal M_{\varphi''}^{(k)}\,.
\end{equation}
Another consequence of 3) is that the minimal polynomial of $\varphi'$ divides $P(x)$. Together with 2), this implies
\begin{equation}\label{eq:47bis}
\sum_{k=0}^N a_k \mathcal M^{(k)}_{\varphi'} = P(u'-v'-w')\cdot \zeta_{CD}=0\,.
\end{equation}
Finally, substituting \eqref{eq:47bis} and the analogous result for $\varphi''$ into \eqref{eq:46} shows that $\varphi$ is indeed minimal.
\end{proof}
\begin{example}\label{ex:102}
Let $\varphi$ be a generalized polynomial structure with minimal polynomial of the form $P(x)=x^r(x^2+c_0)(x^2+c_2)\cdots(x^2+c_s)$, for non-negative integers $r,s$ and arbitrary real numbers $c_1<c_2<\cdots<c_s$. Then the semisimple and the nilpotent parts of $\varphi$ satisfy conditions 1) and 3) in Proposition \ref{prop:93}. Hence $\varphi$ is minimal provided that $\varphi_s$ and $\varphi_n$ are both minimal with vanishing bivariate Courant-Nijenhuis torsion.
\end{example}
\begin{cor}\label{cor:103}
Let $\varphi$ be a generalized polynomial structure of degree 4 with exactly 3 distinct eigenvalues. Then $\varphi$ is minimal if and only if
\begin{enumerate}[1)]
\item the semisimple part $\varphi_s$ of $\varphi$ is a weak generalized Nijenhuis operator;
\item the nilpotent part $\varphi_n$ of $\varphi$ is a generalized Nijenhuis operator;
\item the bivariate Courant-Nijenhuis torsion of $\varphi_s$ and $\varphi_n$ vanishes.
\end{enumerate}
\end{cor}
\begin{proof} By assumption the minimal polynomial of $\varphi$ is of the form $P(x)=x^2(x^2+c)$ for some non-zero real number $c$.
In one direction this is a particular case of Example \ref{ex:102}. For the converse, assume that $\varphi$ is minimal. Proposition \ref{prop:87} then implies that $\varphi_s$ and $\varphi_n$ are both minimal. Thanks to the observations made in Example \ref{ex:87} and in Example \ref{ex:69}, this is enough to conclude that $\varphi_s$ is a weak generalized Nijenhuis operator and $\varphi_n$ is a generalized Nijenhuis operator. To prove 3), we employ the notation introduced in the proof of Proposition \ref{prop:93} with $\varphi'=\varphi_s$ and $\varphi''=\varphi_n$. Since, by the minimality of $\varphi_s$ and $\varphi_n$ we have $((u'-v'-w')^3+c(u'-v'-w'))\cdot \zeta_{CD}=0$ and $(u''-v''-w'')^2\cdot \zeta_{CD}=0$. Imposing the minimality of $\varphi$ we then obtain
\begin{equation}\label{eq:73}
0=(u-v-w)^2((u-v-w)^2+c)\cdot \zeta_{CD}=-2c(u'-v'-w')(u''-v''-w'')\cdot \zeta_{CD}\,.
\end{equation}
By \eqref{eq:51}, \eqref{eq:73} implies that the bivariate Courant-Nijenhuis torsion of $\varphi_s$ and $\varphi_n$ vanishes.
\end{proof}
\begin{cor}
There is a canonical bijection between the set of minimal generalized polynomial structures with minimal polynomial $P(x)=x^2(x^2+1)$ and the set of pairs $(\mathcal J_1,\mathcal J_2)$ consisting of a strongly integrable generalized F-structure $\mathcal J_1$ and a weak generalized tangent structure $\mathcal J_2$ such that
\begin{enumerate}[1)]
\item $\mathcal J_1\circ \mathcal J_2=0=\mathcal J_2\circ \mathcal J_1$
\item the bivariate Courant-Nijenhuis torsion of $\mathcal J_1$ and $\mathcal J_2$ vanishes.
\end{enumerate}
\end{cor}
\begin{proof}
Given a generalized polynomial structure $\varphi$ with minimal polynomial $P(x)=x^2(x^2+1)$ we have that $\mathcal J_1=\varphi_s$ is a generalized F-structure and $\mathcal J_2=\varphi_n$ is a weak generalized almost tangent structure satisfying condition 1). Conversely if $\mathcal J_1$ and $\mathcal J_2$ are as above, then $\varphi=\mathcal J_1+\mathcal J_2$ is a generalized polynomial structure with minimal polynomial $P(x)=(x^2+1)^2$. The result then follows from Corollary \ref{cor:103}.
\end{proof}
\subsection{Decomposition of the de Rham operator}
In this subsection we show how any generalized polynomial structure $\varphi$ induces a canonical decomposition of the de Rham operator into components labeled by the spectrum of $\varphi$. The following theorem provides a characterization of minimality in terms of the compatibility of this decomposition with the adjoint action of $\varphi$.
\begin{theorem}\label{theorem:83}
Let $\varphi$ be a generalized polynomial structure with minimal polynomial
\begin{equation}
P(x)=\prod_{\lambda\in \Sigma(\varphi)}(x-\lambda)^{m(\lambda)}\,.
\end{equation}
Then $\varphi$ is minimal if and only if there exists a decomposition of the de Rham operator
\begin{equation}\label{eq:38}
d = \sum_{\lambda \in \Sigma(\varphi)} d_\lambda
\end{equation}
such that $({\rm ad}_{\widetilde \varphi}-\lambda I)^{m(\lambda)}(d_\lambda)\in\mathbb{T} M$ for every $\lambda\in \Sigma(\varphi)$.
\end{theorem}
\begin{proof}
Assume such a decomposition exists. Since $({\rm ad}_{\widetilde \varphi}-\lambda I)^{m(\lambda)}$ divides $P({\rm ad}_{\widetilde\varphi})$, then $P({\rm ad}_{\widetilde\varphi})(d_\lambda)\in \mathbb{T} M$ for every $\lambda\in \Sigma(\varphi)$. Hence $P({\rm ad}_{\widetilde\varphi})(d)\in \mathbb{T} M$ and thus $\varphi$ is minimal. Conversely, assume that $\varphi$ is minimal. let
\begin{equation}\label{eq:40}
d_\lambda = \sum_{i=1}^{m(\lambda)}a_{\lambda,i}Q_{\lambda,i}({\rm ad}_{\widetilde\varphi})(d)\,,
\end{equation}
where, $Q_{\lambda,i}$ is a polynomial defined by the relation $(x-\lambda)^i Q_{\lambda,i}(x)=P(x)$ for each $\lambda\in \Sigma(\varphi)$ and for each $i\in\{1,\ldots,m(\lambda)\}$. Then \eqref{eq:38} holds and, for each $\lambda\in \Sigma(\varphi)$
\begin{equation}\label{eq:41}
({\rm ad}_{\widetilde\varphi}-\lambda I)^{m(\lambda)}(d_\lambda)=\sum_{i=1}^{m(\lambda)}a_{\lambda,i}({\rm ad}_{\widetilde\varphi}-\lambda I)^{m(\lambda)-i}P({\rm ad}_{\widetilde\varphi})(d)\,.
\end{equation}
The minimality of $\varphi$ implies that $P({\rm ad}_{\widetilde\varphi})(d)$ and thus \eqref{eq:41} is a section of the generalized tangent bundle.
\end{proof}
\begin{rem}
A priori the definition of $d_\lambda$ given in \eqref{eq:41} depends on the particular lift of $\varphi$ to $\mathcal E_M$ obtained by imposing condition 3) in Proposition \ref{prop:9}. On the other hand, if $\varphi$ is minimal then choosing a different lift $\widetilde \varphi +g$ for some function $g\in \Omega_M^0$ would result in a a different decomposition $d=\sum_\lambda d_\lambda'$ such that $d_\lambda-d_\lambda'$ is a section of $\mathbb{T} M\otimes \mathbb C$ annihilated by $({\rm ad}_{\widetilde \varphi}-\lambda I)^{m(\lambda)}$ i.e.\ a section of $L_\lambda$. As $\lambda$ runs through $\Sigma(\varphi),$, the sum of these sections is $\sum_\lambda (d_\lambda-d'_\lambda)=0$ and yet, by Theorem \ref{theorem:40} no cancellation can occur between sections labelled by different eigenvalues. Hence $d_\lambda=d'_\lambda$ for all $\lambda\in \Sigma(\varphi)$ and the decomposition $\eqref{eq:40}$ induced by a minimal generalized polynomial structure does not depend on its lift to $\mathcal E_M$. This is consistent with the observation made in Remark \ref{rem:68} that the minimal torsion and thus the notion of minimality does not depend on such a lift.
\end{rem}
\begin{example}\label{ex:103}
Let $\varphi$ be a generalized almost complex structure. Then $a_{\pm \sqrt{-1},1}=\mp\frac{\sqrt{-1}}{2}$ and $Q_{\pm \sqrt{-1},1}(x)=x\pm \sqrt{-1}$. Substituting into \eqref{eq:40} we obtain
\begin{equation}
d_{\sqrt{-1}}=\frac{1}{2}(d-\sqrt{-1}[\varphi,d])= \overline \partial
\end{equation}
and, similarly, $d_{-\sqrt{-1}}=\partial$.
\end{example}
\begin{rem}
Since \eqref{eq:40} expresses $d_\lambda$ as the image of $d$ under a polynomial in ${\rm ad}_{\widetilde \varphi}$, then ${[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]}_{d_\lambda}\in \mathbb{T} M\otimes \mathbb C$ for all $\mathbf{x},\mathbf{y}\in \mathbb{T} M\otimes \mathbb C$.
\end{rem}
\begin{cor}\label{cor:104}
Let $\varphi$ be a minimal generalized polynomial structure and let $d_\lambda$ be as in \eqref{eq:40}. Then ${[}\!{[} L_\mu, L_\nu {]}\!{]}_{d_\lambda}\subseteq L_{\lambda+\mu+\nu}$ for all $\lambda,\mu,\nu\in \Sigma(\varphi)$.
\end{cor}
\begin{proof}
Adapting the notation of Remark \ref{rem:PolynomialAction}, let $\zeta_{CD,\lambda}$ denote the local linear operator representing the derived bracket with respect to $d_\lambda$. Since $\varphi$ is minimal, by Theorem \ref{theorem:83} we have $(u-v-w-\lambda)^{m(\lambda)}\cdot \zeta_{CD,\lambda}=0$. Recall the shift operator $\mathbf S$ introduced in the proof of Corollary \ref{cor:88}. Multiplying both sides by $(u-v-w-\lambda)^{m(\mu)+m(\nu)}$ and setting $m=m(\lambda)+m(\mu)+m(\nu)$ we obtain
\begin{align*}
0&=(\varphi-(\lambda+\mu+\nu)I+\mathbf S)^m {[}\!{[} \mathbf{x}_\alpha,\mathbf{y}_\beta{]}\!{]}_{d_\lambda}\\
&= \sum_{k=0}^{m(\mu)+m(\nu)}\binom{m}{k}(\varphi-(\lambda+\mu+\nu)I)^{m-k}\mathbf S^k{[}\!{[} \mathbf{x}_\alpha,\mathbf{y}_\beta{]}\!{]}_{d_\lambda}
\end{align*}
for all Jordan chains $\{\mathbf{x}_\alpha\}\subseteq {\rm Ker}((\varphi-\mu I)^p)$ and $\{\mathbf{y}_\beta\}\subseteq {\rm Ker}((\varphi-\nu I)^q)$. If $p=q=1$ this reduces to $(\varphi-(\lambda+\mu+\nu)I)^m{[}\!{[} \mathbf{x}_\alpha,\mathbf{y}_\beta {]}\!{]}_{d_\lambda}=0$ which implies ${[}\!{[} {\rm Ker}(\varphi-\mu I),{\rm Ker}(\varphi-\nu I){]}\!{]}_{d_\lambda}\subseteq L_{\lambda+\mu+\nu}$. The result then follows by induction on $p+q$.
\end{proof}
\begin{rem}
As a consequence of Corollary \ref{cor:104}, if $\varphi$ is a minimal generalized polynomial structure, then
\begin{equation}\label{eq:57}
{[}\!{[} \mathbf{x} ,\mathbf{y} {]}\!{]} = \sum_{\lambda\in \Sigma(\varphi)} {[}\!{[} \mathbf{x},\mathbf{y}{]}\!{]}_{d_\lambda} \subseteq L_\mu+L_\nu + \sum_{\lambda\in \Sigma(\varphi)\setminus\{-\mu,-\nu\}} L_{\mu+\nu+\lambda}
\end{equation}
for all $\mathbf{x}\in L_\mu$, $\mathbf{y}\in L_\nu$.
\end{rem}
\subsection{The non-resonance condition}
The following definition is motivated by \eqref{eq:47}.
\begin{mydef}
Let $\varphi$ be a generalized polynomial structure with minimal polynomial $P(x)$. We say that $\varphi$ is {\it non-resonant} if $P(\lambda+\mu+\nu)\neq 0$ for any $\lambda,\mu,\nu \in \Sigma(\varphi)$ such that $(\lambda+\mu)(\lambda+\nu)(\mu+\nu)\neq 0$.
\end{mydef}
\begin{rem}
Non-resonance is a generic condition requiring that no eigenvalue can be written as a non-trivial (i.e.\ when two of the summands cancel each other) sum of three eigenvalues.
\end{rem}
\begin{example}
Every generalized polynomial structure with at most three eigenvalues is non-resonant. In particular, generalized polynomial structure of degree at most 3 are always non-resonant.
\end{example}
\begin{example} Let $\varphi$ be a generalized polynomial structure with at most five distinct eigenvalues (in particular, this occurs if $\varphi$ has degree at most five). Then $\varphi$ is non-resonant if and only if $\lambda\notin \{2\mu, 3\mu\}$ for all $\lambda,\mu\in \Sigma(\varphi)$.
\end{example}
\begin{theorem}\label{theorem:93}
Let $\varphi$ be a minimal non-resonant generalized polynomial structure. Then the semisimple part of $\varphi$ is a weak generalized Nijenhuis operator.
\end{theorem}
\begin{proof}
The non-resonance assumption guarantees that if $\mu+\nu\neq 0$, then \eqref{eq:57} reduces to ${[}\!{[} L_\mu,L_\nu{]}\!{]}\subseteq L_\mu+L_\nu$. Hence the result follows from Theorem \ref{theorem:56}.
\end{proof}
\begin{prop}\label{prop:116}
Let $\varphi$ be a non-resonant generalized polynomial structure. Then its semisimple part $\varphi_s$ is minimal if and only if $\varphi_s$ is a weak generalized Nijenhuis operator.
\end{prop}
\begin{proof}
In one direction this statement 2) in Proposition \ref{prop:88} specialized to the non-resonant case. The converse follows from Theorem \ref{theorem:93} applied to $\varphi_s$ (which is equal to its own semisimple part), taking into account that $\varphi$ and $\varphi_s$ share the same generalized eigenbundles.
\end{proof}
\begin{example}\label{ex:117}
Let $\varphi$ a generalized polynomial structure of with minimal polynomial $P(x)=(x^2+c_1)^{m_1}(x^2+c_2)^{m_2}$ for some positive integers $m_1$, $m_2$ and real numbers $c_1$, $c_2$ such that $c_1c_2(c_1-c_2)(c_1-4c_2)(4c_1-c_2)(c_1-9c_2)(9c_1-c_2)\neq 0$. Then $\varphi_s$ is minimal if and only it is a weak generalized Nijenhuis operator. In particular, a generic generalized polynomial structure of degree $4$ is minimal if and only if it is a weak generalized Nijenhuis operator.
\end{example}
\begin{example}
To illustrate the fact that the non-resonance assumption in Theorem \ref{theorem:93} and in Proposition \ref{prop:116} cannot be completely removed, consider the following example. Let $M$ be the cartesian product of the 3-dimensional Heisenberg nilmanifold with $S^1$. Then $\mathbb T M$ can be globally trivialized by vector fields $\mathbf{e}_1,\ldots,\mathbf{e}_4$ and dual 1-forms $\mathbf{e}^1,\ldots,\mathbf{e}^4$ with Dorfman brackets encoded by the single (up to skew-symmetry) constraint $T(\mathbf{e}_1,\mathbf{e}_2,\mathbf{e}^3)=1$. Let $\lambda$, $\mu$ be real numbers and let $\varphi$ be the unique skew-symmetric endomorphism of $\mathbb{T} M$ such that $\varphi(\mathbf{e}_1)=\lambda \mathbf{e}_1$, $\varphi(\mathbf{e}_2)=\lambda \mathbf{e}_2$, $\varphi(\mathbf{e}_3)=\mu \mathbf{e}_3$ and $\varphi(\mathbf{e}_4)=\mu \mathbf{e}_4$. Then the minimal polynomial of $\varphi$ is $P(x)=(x^2-\lambda^2)(x^2-\mu^2)$. Let us assume $\lambda\neq \mu$ so that $\varphi$ is semisimple and $L_\lambda$ is not closed under the Dorfman bracket. Let us further assume that $\lambda\neq 0$ so that by Theorem \ref{theorem:56} we know that $\varphi$ cannot be a weak generalized Nijenhuis operator. From \eqref{eq:47} we then have that $\varphi$ is minimal if and only if
\begin{equation}
0=\mathcal C_\varphi(\mathbf{e}_1,\mathbf{e}_2,\mathbf{e}^3)=P(2\lambda-\mu)=4\lambda(3\lambda-\mu)(\lambda-\mu)^2
\end{equation}
if and only if $3\lambda=\mu$, which as shown in Example \ref{ex:117} is a case in which non-resonance fails to hold. Hence $3\lambda=\mu\neq 0$ implies that $\varphi$ is minimal and semisimple but not a weak generalized Nijenhuis operator.
\end{example}
\begin{rem}
Let $\varphi$ be a non-resonant minimal generalized polynomial structure so that, by Theorem \ref{theorem:93}, $\varphi_s$ is a weak generalized Nijenhuis operator. In principle, we have two decompositions of the de Rham operator, namely \eqref{eq:38} and the decomposition $d=\sum_\lambda \delta_\lambda$ whose existence is guaranteed by Theorem \ref{theorem:56}. By comparing Remark \ref{rem:61} and Example \ref{ex:103}, we see that in the case of generalized complex structures one has $\delta_{\pm\sqrt{-1}}=d_{\mp\sqrt{-1}}$. For the general case, we note that specializing Corollary \ref{cor:104} by imposing the non-resonance condition, we can repeat the proof of 2)$\Rightarrow$ 3) in Theorem \ref{theorem:56} with $\delta_\lambda$ replaced by $d_{-\lambda}$. Thus, $d_{\pm\lambda}$ has degree $\mp 1$ with respect to the $(L_\lambda, L_{-\lambda})$-grading and degree 0 with respect to the $(L_\mu,L_{-\mu})$-grading for every $\mu\in \Sigma(\varphi)\setminus\{0,\pm \lambda\}$. In particular, we conclude that if $\varphi$ is minimal and non-resonant, then $d_\lambda=\delta_{-\lambda}$ for every $\lambda\in \Sigma(\varphi)\setminus\{0\}$. \end{rem}
\section{Examples: invariant polynomial structures on Lie groups}\label{sec:6}
In this section, $G$ will denote a real Lie group and $\mathfrak{g}$ will be its Lie algebra. The invariant polynomial structures on $G$ coincide with all skew-symmetric endomorphisms of the space $D_1=D(\mathfrak{g})=\mathfrak{g}\oplus\mathfrak{g}^*$ (the \emph{Drinfeld double} of $\mathfrak{g}$), identified with the space of all left-invariant sections of $\mathbb{T} G$.
Note that $D_1$ is itself a Lie algebra with respect to the Dorfman bracket.
In the following examples we will consider elements of the vector space $D_3=D_1^{\otimes 3}$, associated to the Lie algebra structure and an invariant polynomial structure $\varphi:D_1\rightarrow D_1$.
Note that the tautological inner product defines an isomorphism $D_1^*\rightarrow D_1, \ \omega\mapsto \widehat{\omega}$
such that $\omega(\mathbf{x})=2\langle\widehat{\omega},\mathbf{x}\rangle$ for all $\mathbf{x}\in D_1$. Via tensor product, we obtain isomorphisms
$D_1^{*\otimes h}\otimes D_1^{\otimes k}\rightarrow D_1^{\otimes(h+k)}, {\mathcal T} \mapsto \widehat{{\mathcal T}}$.
In particular, any invariant polynomial structure $\varphi$ produces the elements
$\widehat{\mathcal{C}}_{\varphi},\ \widehat{\mathcal{T}}^{(n)}_{\varphi},\ \widehat{\mathcal{S}}^{(n)}_{\varphi}\in D_3 $.
In the following examples we calculate the elements $\varphi_1^{i_1}\varphi_2^{i_2}\varphi_3^{i_3}\cdot \widehat{\mathcal{T}}^{(0)}_{\varphi}$, where the action was defined in Remark \ref{rem:PolynomialAction}: in particular, we will represent these elements with respect to a Jordan basis and adopt the following abbreviation:
$$\mathbf{x}\mathbf{y}\mathbf{z}:=\mathbf{x}\otimes\mathbf{y}\otimes\mathbf{z}.$$
Note that the polynomial action of $\varphi$ is graded skew-symmetric with respect to the following non-degenerate inner product:
$$\langle\mathbf{x}_1\mathbf{x}_2\mathbf{x}_3, \mathbf{y}_1\mathbf{y}_2\mathbf{y}_3\rangle=\prod_{i=1}^3\langle\mathbf{x}_i,\mathbf{y}_i\rangle$$
In our examples, $\{v_i\}_{i=1}^n$ will denote a fixed basis of $\mathfrak{g}$, with dual basis given by $\{\alpha_i\}_{i=1}^n$.
Moreover, in our description of the tensors it will be convenient to work with the complexification of $\varphi$ and to use to use a basis built upon the block decomposition of $\varphi$.
More specifically, we will choose bases of $D_1\otimes\mathbb{C}$ of the form $$\left\{\mathbf{b}_j^{V,l}:j,V,l\right\}$$ where the indices $V,l$ and $j$ are described as follows:
\begin{itemize}
\item each index $V$ indicates a distinct indecomposable complex block;
\item for a fixed block $V$ of degree $k$ we have $1\leq j\leq k$ and the vectors $$\left\{\mathbf{b}_k^{V,l}:l\right\}$$ form a basis of the semisimple part $W$ as described in Section \ref{decompos}. Also, these vectors will be chosen to be real if the eigenvalues of $V$ are real;
\item for fixed $V$ (of degree $k$) and $l$, we have $\varphi_n\left(\mathbf{b}_j^{V,l}\right)=\mathbf{b}_{j-1}^{V,l}$ for all $j=2,3,\dots k$ while $\varphi_n\left(\mathbf{b}_1^{V,l}\right)=0$, that is we have the Jordan chain
$$\mathbf{b}_k^{V,l}\longmapsto \mathbf{b}_{k-1}^{V,l}\longmapsto \mathbf{b}_{k-2}^{V,l}\longmapsto \mathbf{b}_{k-3}^{V,l}\longmapsto\dots\longmapsto\mathbf{b}_1^{V,l}\longmapsto 0.$$
(We will simplify the indices $V,l$ to lighten the notation).
\end{itemize}
\subsection{The Heisenberg group.}
Let $\mathfrak{g}=\mathfrak{n}(3)$, the 3-dimensional Heisenberg Lie algebra, which is nilpotent and has structure equations
$$[v_1,v_2]=v_3, \quad [v_1,v_3]=[v_2,v_3]=0.$$
Therefore,
$$\widehat{\mathcal{T}}^{(0)}_{\varphi}=\alpha_1\alpha_2v_3\pm \text{perm.}$$
where \lq\lq $\pm\ \text{perm.}$\rq\rq refers to the signed sum of terms where the three factors are permuted.
\subsubsection{Example.} Define the endomorphism $\varphi:D_1\rightarrow D_1$ as follows:
$$\varphi(v_1)=v_2,\quad \varphi(v_2)=v_3+\alpha_3,\quad \varphi(v_3)=-\alpha_2,$$
$$\varphi(\alpha_1)=0,\quad \varphi(\alpha_2)=-\alpha_1, \quad \varphi(\alpha_3)=-\alpha_2$$
By construction, $\varphi$ is skewsymmetric, it has
minimal polynomial $P(x)=x^5$, and two Jordan chains
$$v_1\longmapsto v_2\longmapsto v_3+\alpha_3\longmapsto -2\alpha_2\longmapsto 2\alpha_1\longmapsto 0$$
$$v_3-\alpha_3\longmapsto 0$$
which span two real blocks, of types $\Delta_5^+(0)$ and $\Delta_1^-(0)$ respectively.
Defining
$$\mathbf{b}_1=2\alpha_1,\quad \mathbf{b}_2=-2\alpha_2,\quad \mathbf{b}_3=v_3+\alpha_3,\quad \mathbf{b}_4=v_2,\quad \mathbf{b}_5=v_1,\quad \mathbf{b}_1'=v_3-\alpha_3$$
we get
\begin{equation}
-8\widehat{\mathcal{T}}^{(0)}_{\varphi}=\mathbf{b}_1\mathbf{b}_2\mathbf{b}_3+\mathbf{b}_1\mathbf{b}_2\mathbf{b}_1'\pm \text{perm.}
\end{equation}
Upon inspection, we see that the polynomials that do not annihilate the summands have total degree at most $3$. in particular, $\widehat{\mathcal{C}}_{\varphi}=P(\varphi_1+\varphi_2+\varphi_3)\cdot \widehat{\mathcal{T}}^{(0)}_{\varphi}=0$, i.e. $\varphi$ is minimal. For the same reason, $\widehat{\mathcal{T}}^{(k)}_{\varphi}=0$ for $k\geq 2$. Moreover,
\begin{align*}
-8\widehat{\mathcal T}^{(1)}_{\varphi}=&(\varphi_3^2+\varphi_1\varphi_2+\varphi_1\varphi_3+\varphi_2\varphi_3)\cdot (-8\widehat{\mathcal T}^{(0)}_{\varphi})\\
=&(\mathbf{b}_1\mathbf{b}_2\mathbf{b}_1 -\mathbf{b}_2\mathbf{b}_1\mathbf{b}_1)+(\mathbf{b}_1\mathbf{b}_2\mathbf{b}_1-\mathbf{b}_2\mathbf{b}_1\mathbf{b}_1)\\
& +(-\mathbf{b}_1\mathbf{b}_1\mathbf{b}_2+\mathbf{b}_2\mathbf{b}_1\mathbf{b}_1)+(\mathbf{b}_1\mathbf{b}_1\mathbf{b}_2-\mathbf{b}_1\mathbf{b}_2\mathbf{b}_1)\\
=&-\mathbf{b}_2\mathbf{b}_1\mathbf{b}_1+\mathbf{b}_1\mathbf{b}_2\mathbf{b}_1\,,
\end{align*}
while
\begin{equation}
-8\widehat{{\mathcal S}}^{(1)}_{\varphi}=(-\varphi_1-\varphi_2)\cdot (-\mathbf{b}_2\mathbf{b}_1\mathbf{b}_1+\mathbf{b}_1\mathbf{b}_2\mathbf{b}_1)=\mathbf{b}_1\mathbf{b}_1\mathbf{b}_1-\mathbf{b}_1\mathbf{b}_1\mathbf{b}_1=0\,,
\end{equation}
and also $\widehat{{\mathcal S}}^{(n)}_{\varphi}=0$ for all $n\geq 1$.
\subsubsection{Example}
Let now $$\varphi(v_1)=0,\quad \varphi(v_2)=-v_1,\quad \varphi(v_3)=-v_2,$$
$$\varphi(\alpha_1)=-\alpha_2,\quad \varphi(\alpha_2)=v_3+\alpha_3, \quad \varphi(\alpha_3)=-v_2$$
As in the previous example, $\varphi$ is skew-symmetric, is has minimal polynomial $P(x)=x^5$ and real block decomposition
$\Delta_5^+(0)\oplus\Delta_1^-(0)$. Its
Jordan chains are
$$\alpha_1\longmapsto \alpha_2\longmapsto v_3+\alpha_3\longmapsto -2v_2\longmapsto 2v_1\longmapsto 0$$
$$v_3-\alpha_3\longmapsto 0$$
and block decomposition of same type as in the previous example. Defining
$$\mathbf{b}_1=2v_1,\quad \mathbf{b}_2=-2v_2,\quad \mathbf{b}_3=v_3+\alpha_3,\quad \mathbf{b}_4=\alpha_2,\quad \mathbf{b}_5=\alpha_1,\quad \mathbf{b}_1'=v_3-\alpha_3$$
we get
\begin{equation}
2\widehat{\mathcal{T}}^{(0)}=\mathbf{b}_5\mathbf{b}_4\mathbf{b}_3+\mathbf{b}_5\mathbf{b}_4\mathbf{b}_1'\pm \text{perm.}
\end{equation}
In this case, upon inspection we see that the polynomials that do not annihilate the summands have total degree at most $9$, so that $\widehat{\mathcal T}^{(k)}_{\varphi}=0$ and $\widehat{\mathcal S}^{(k)}_{\varphi}=0$ for $k\geq 5$. Now we prove that $\widehat{\mathcal{C}}_{\varphi}\neq 0$,
by showing that
$\langle\widehat{\mathcal{C}}_{\varphi}, \mathbf{x}\mathbf{y}\mathbf{z}\rangle\neq 0$ for some $\mathbf{x}\mathbf{y}\mathbf{z}$.
Note first that
\begin{equation}\label{eq:82}
\langle\widehat{\mathcal{C}}_{\varphi}, \mathbf{x}\mathbf{y}\mathbf{z}\rangle=\langle P(\varphi_1+\varphi_2+\varphi_3)\cdot\widehat{\mathcal{T}}^{(0)}_{\varphi}, \mathbf{x}\mathbf{y}\mathbf{z}\rangle=-\langle \widehat{\mathcal{T}}^{(0)}_{\varphi}, P(\varphi_1+\varphi_2+\varphi_3)\cdot\mathbf{x}\mathbf{y}\mathbf{z}\rangle
\end{equation}
Setting $\mathbf{x}\mathbf{y}\mathbf{z}=\mathbf{b}_1'\mathbf{b}_5\mathbf{b}_3$ in \eqref{eq:82}, we get
\begin{equation}
P(\varphi_1+\varphi_2+\varphi_3)\cdot \mathbf{b}_1'\mathbf{b}_5\mathbf{b}_3=5\mathbf{b}_1'\mathbf{b}_1\mathbf{b}_2+10\mathbf{b}_1'\mathbf{b}_2\mathbf{b}_1
\end{equation}
and
\begin{align*}
\langle2\widehat{\mathcal{T}}^{(0)}_{\varphi},5\mathbf{b}_1'\mathbf{b}_1\mathbf{b}_2+10\mathbf{b}_1'\mathbf{b}_2\mathbf{b}_1\rangle=&\langle\mathbf{b}_1'\mathbf{b}_5\mathbf{b}_4-\mathbf{b}_1'\mathbf{b}_4\mathbf{b}_5,5\mathbf{b}_1'\mathbf{b}_1\mathbf{b}_2+10\mathbf{b}_1'\mathbf{b}_2\mathbf{b}_1\rangle\\
=&5\langle\mathbf{b}_1'\mathbf{b}_5\mathbf{b}_4,\mathbf{b}_1'\mathbf{b}_1\mathbf{b}_2\rangle-10\langle\mathbf{b}_1'\mathbf{b}_4\mathbf{b}_5,\mathbf{b}_1'\mathbf{b}_2\mathbf{b}_1\rangle\\
=&-5\,,
\end{align*}
so that $\langle\widehat{\mathcal{C}}_{\varphi}, \mathbf{b}_1'\mathbf{b}_5\mathbf{b}_3\rangle \neq 0$. Therefore, $\varphi$ is not minimal. \\
Now, we are going to prove that $\widehat{{\mathcal T}}^{(4)}_{\varphi}\neq 0$, by showing that $\langle\widehat{{\mathcal T}}^{(4)}_{\varphi}, \mathbf{b}_5\mathbf{b}_4\mathbf{b}_5\rangle\neq 0$. In fact,
\begin{align*}
\langle 2\widehat{\mathcal T}^{(4)}_{\varphi}, \mathbf{b}_5\mathbf{b}_4\mathbf{b}_5\rangle=&\langle2\widehat{\mathcal T}^{(0)}_{\varphi}, (\varphi_1+\varphi_3)^4(\varphi_2+\varphi_3)^4\cdot\mathbf{b}_5\mathbf{b}_4\mathbf{b}_5\rangle\\
=&\sum_{i,j}\binom{4}{i}\binom{4}{j}\langle2\widehat{\mathcal T}^{(0)}_{\varphi}, \mathbf{b}_{1+i}\mathbf{b}_{j}\mathbf{b}_{5-i-j}\rangle\\
=& 6\langle\mathbf{b}_5\mathbf{b}_4\mathbf{b}_3,\mathbf{b}_1\mathbf{b}_2\mathbf{b}_3\rangle-4\langle\mathbf{b}_5\mathbf{b}_3\mathbf{b}_4,\mathbf{b}_1\mathbf{b}_3\mathbf{b}_2\rangle -16\langle\mathbf{b}_4\mathbf{b}_5\mathbf{b}_3,\mathbf{b}_2\mathbf{b}_1\mathbf{b}_3\rangle\\
&+16\langle\mathbf{b}_4\mathbf{b}_3\mathbf{b}_5,\mathbf{b}_2\mathbf{b}_3\mathbf{b}_1\rangle+24\langle\mathbf{b}_3\mathbf{b}_5\mathbf{b}_4,\mathbf{b}_3\mathbf{b}_1\mathbf{b}_2\rangle-36\langle\mathbf{b}_3\mathbf{b}_4\mathbf{b}_5,\mathbf{b}_3\mathbf{b}_2\mathbf{b}_1\rangle\\
=&10\,.
\end{align*}
However, a similar calculation shows that $\widehat{\mathcal{S}}^{(4)}_{\varphi}=0$.
\subsubsection{Example}
Define $$\varphi(v_1)=v_2-\alpha_2, \quad \varphi(v_2)=-v_1+\alpha_1,\quad \varphi(v_3)=0$$
$$\varphi(\alpha_1)=\alpha_2\quad \varphi(\alpha_2)=-\alpha_1,\quad\varphi(\alpha_3)=0$$
In this case, $\varphi$ has minimal polynomial $P(x)=x(x^2+1)^2$, block decomposition $\Delta_2^+(\sqrt{-1},-\sqrt{-1})\oplus \Delta_1^0(0,0)$
and Jordan chains
$$v_1\longmapsto -\alpha_2\longmapsto 0,\qquad v_2\longmapsto \alpha_1\longmapsto 0$$
$$v_3\longmapsto 0 \qquad \alpha_3\longmapsto 0$$
Define
$$\mathbf{b}_1^1=-\alpha_2-\sqrt{-1}\alpha_1,\quad \mathbf{b}_2^1=v_1-\sqrt{-1}v_2,\quad \mathbf{b}_i^2=\overline{\mathbf{b}_i^1},\quad \mathbf{b}_1^3=v_3, \quad \mathbf{b}_1^4=\alpha_3 $$
Then,
\begin{equation}
2\sqrt{-1}\widehat{\mathcal{T}}^{(0)}_{\varphi}=\mathbf{b}_1^1\mathbf{b}_1^2\mathbf{b}_1^3\pm \text{perm.}
\end{equation}
Therefore,
$\widehat{\mathcal{C}}_{\varphi}, \widehat{\mathcal{T}}^{(n)}_{\varphi}, \widehat{\mathcal{S}}^{(n)}_{\varphi}$ do not depend on the nilpotent part of $\varphi$, and
$\varphi_1^{i_1}\varphi_2^{i_2}\varphi_3^{i_3}\cdot \widehat{\mathcal{T}}^{(0)}_{\varphi}=0$ if $i_1i_2i_3\neq 0$.
Moreover, if $i_1i_2\neq 0
$,
\begin{equation}
2\varphi_1^{i_1}\varphi_2^{i_2}\cdot\widehat{\mathcal{T}}^{(0)}_{\varphi}=\left({\sqrt{-1}}\right)^{i_1+i_2-1}\left((-1)^{i_2}\mathbf{b}_1^1\mathbf{b}_1^2\mathbf{b}_1^3
-(-1)^{i_1}\mathbf{b}_1^2\mathbf{b}_1^1\mathbf{b}_1^3
\right)
\end{equation}
and analogously, if $i_1i_3\neq 0$,
\begin{equation}
2\varphi_1^{i_1}\varphi_3^{i_3}\cdot\widehat{\mathcal{T}}^{(0)}_{\varphi}=\left({\sqrt{-1}}\right)^{i_1+i_3-1}\left(-(-1)^{i_3}\mathbf{b}^1_1\mathbf{b}_1^3\mathbf{b}_1^2+(-1)^{i_1}\mathbf{b}_1^2\mathbf{b}_1^3\mathbf{b}_1^1\right)
\end{equation}
and if $i_2i_3\neq 0$,
\begin{equation}
2\varphi_2^{i_2}\varphi_3^{i_3}\cdot\widehat{\mathcal{T}}^{(0)}_{\varphi}=\left({\sqrt{-1}}\right)^{i_2+i_3-1}\left((-1)^{i_3}\mathbf{b}_1^3\mathbf{b}_1^1\mathbf{b}_1^2-(-1)^{i_2}\mathbf{b}_1^3\mathbf{b}_1^2\mathbf{b}_1^1\right)\,.
\end{equation}
Finally, for $i\neq 0$,
\begin{align*}
2\varphi_1^{i}\cdot\widehat{\mathcal{T}}^{(0)}_{\varphi}=&\left({\sqrt{-1}}\right)^{i-1}\left(\mathbf{b}_1^1\mathbf{b}_1^2\mathbf{b}_1^3-\mathbf{b}^1_1\mathbf{b}_1^3\mathbf{b}_1^2-(-1)^i\mathbf{b}_1^2\mathbf{b}_1^1\mathbf{b}_1^3+(-1)^i\mathbf{b}_1^2\mathbf{b}_1^3\mathbf{b}_1^1\right)\\
2\varphi_2^{i}\cdot\widehat{\mathcal{T}}^{(0)}_{\varphi}=&\left({\sqrt{-1}}\right)^{i-1}\left((-1)^i\mathbf{b}_1^1\mathbf{b}_1^2\mathbf{b}_1^3-\mathbf{b}_1^2\mathbf{b}_1^1\mathbf{b}_1^3+\mathbf{b}_1^3\mathbf{b}_1^1\mathbf{b}_1^2-(-1)^i\mathbf{b}_1^3\mathbf{b}_1^2\mathbf{b}_1^1\right)\\
2\varphi_3^{i}\cdot\widehat{\mathcal{T}}^{(0)}_{\varphi}=&\left({\sqrt{-1}}\right)^{i-1}\left(-(-1)^i\mathbf{b}^1_1\mathbf{b}_1^3\mathbf{b}_1^2+\mathbf{b}_1^2\mathbf{b}_1^3\mathbf{b}_1^1+(-1)^i\mathbf{b}_1^3\mathbf{b}_1^1\mathbf{b}_1^2-\mathbf{b}_1^3\mathbf{b}_1^2\mathbf{b}_1^1\right)\,.
\end{align*}
In particular, $(\varphi_1+\varphi_2+\varphi_3)\cdot\widehat{\mathcal{T}}^{(0)}_{\varphi}=0$, whence $\widehat{\mathcal{C}}_{\varphi}=0$. Also,
\begin{align*}
2\widehat{\mathcal{T}}^{(n)}_{\varphi}=&2(\varphi_1+\varphi_3)^n(\varphi_2+\varphi_3)^n\cdot \widehat{\mathcal{T}}^{(0)}_{\varphi}\\
=&2\varphi_1^n\varphi_2^n\cdot \widehat{\mathcal{T}}^{(0)}_{\varphi}\\
=&-\sqrt{-1}\left(\mathbf{b}_1^1\mathbf{b}_1^2\mathbf{b}_1^3
-\mathbf{b}_1^2\mathbf{b}_1^1\mathbf{b}_1^3\right)
\end{align*}
which is not vanishing for all integers $n>0$. This calculation also implies $\widehat{\mathcal{S}}^{(n)}_{\varphi}=0$ for all integers $n>0$.
\subsubsection{Example}
$$\varphi(v_1)=v_2, \quad \varphi(v_2)=-v_1,\quad \varphi(v_3)=0$$
$$\varphi(\alpha_1)=\alpha_2-v_2, \quad \varphi(\alpha_2)=-\alpha_1+v_1,\quad \varphi(\alpha_3)=0$$
$$\alpha_1\longmapsto -v_2\longmapsto 0, \quad \alpha_2\longmapsto v_1 \longmapsto 0$$
$$v_3\longmapsto 0, \quad \alpha_3\longmapsto 0$$
$$\Delta_2^+(\sqrt{-1},-\sqrt{-1})\oplus \Delta_1^0(0,0)$$
Let
$$\mathbf{b}_1^1=-v_2-\sqrt{-1}v_1, \quad \mathbf{b}_2^1=\alpha_1-\sqrt{-1}\alpha_2,\quad \mathbf{b}_i^2=\overline{\mathbf{b}_i^1},\quad \mathbf{b}_1^3=v_3, \quad \mathbf{b}_1^4=\alpha_3$$
Hence
\begin{equation}
2\sqrt{-1}\widehat{\mathcal{T}}^{(0)}_{\varphi}=\mathbf{b}_2^1\mathbf{b}_2^2\mathbf{b}_1^3\pm \text{perm.}
\end{equation}
so that, as before, $\varphi_1\varphi_2\varphi_3\cdot \widehat{\mathcal{T}}^{(0)}_{\varphi}=0$. In order to prove that $\widehat{{\mathcal C}}_{\varphi}\neq 0$, it is sufficient to show that $\langle \widehat{{\mathcal C}}_{\varphi}, \mathbf{b}_1^1\mathbf{b}_2^2\mathbf{b}_1^4\rangle\neq 0$.
First of all, note that
\begin{equation}
\langle \widehat{{\mathcal C}}_{\varphi}, \mathbf{b}_1^1\mathbf{b}_2^2\mathbf{b}_1^4\rangle=-\langle \widehat{{\mathcal T}}_{\varphi}^{(0)}, R(\varphi_1,\varphi_2)\cdot \mathbf{b}_1^1\mathbf{b}_2^2\mathbf{b}_1^4\rangle\,,
\end{equation}
where
\begin{equation}
R(\varphi_1,\varphi_2)=5(\varphi_1^4\varphi_2+\varphi_1\varphi_2^4)+10(\varphi_1^3\varphi^2_2+\varphi_1^2\varphi_2^3)+6(\varphi_1^2\varphi_2+\varphi_1\varphi_2^2)\,.
\end{equation}
Now, for all positive integers $a,b$ with $a+b$ odd,
\begin{equation}
(\varphi_1^a\varphi_2^b+\varphi_1^b\varphi_2^a)\cdot \mathbf{b}_1^1\mathbf{b}_2^2\mathbf{b}_1^4=(a-b)\left(\sqrt{-1}\right)^{a+b-1}(-1)^b\mathbf{b}_1^1\mathbf{b}_1^2\mathbf{b}_1^4
\end{equation}
whence
\begin{equation}
R(\varphi_1,\varphi_2)\cdot\mathbf{b}_1^1\mathbf{b}_2^2\mathbf{b}_1^4=\mathbf{b}_1^1\mathbf{b}_1^2\mathbf{b}_1^4
\end{equation}
and
\begin{align*}\langle \mathbf{b}_2^1\mathbf{b}_2^2\mathbf{b}_1^3-\mathbf{b}_2^1\mathbf{b}_1^3\mathbf{b}_2^2-&\mathbf{b}_2^2\mathbf{b}_2^1\mathbf{b}_1^3+\mathbf{b}_2^2\mathbf{b}_1^3\mathbf{b}_2^1+\mathbf{b}_1^3\mathbf{b}_2^1\mathbf{b}_2^2-\mathbf{b}_1^3\mathbf{b}_2^2\mathbf{b}_2^1 \ ,\ \mathbf{b}_1^1\mathbf{b}_1^2\mathbf{b}_1^4 \rangle=\\
=&\langle -\mathbf{b}_2^2\mathbf{b}_2^1\mathbf{b}_1^3\ ,\ \mathbf{b}_1^1\mathbf{b}_1^2\mathbf{b}_1^4 \rangle\neq 0,
\end{align*}
which implies $\widehat{{\mathcal C}}_{\varphi}\neq 0$. Similar calculations show that $\widehat{\mathcal{T}}^{(n)}_{\varphi}\neq 0 $ and $\widehat{\mathcal{S}}^{(n)}_{\varphi}\neq 0 $ for $n\geq 1$.
\subsection{A four-dimensional nilpotent group} Let $\mathfrak{g}$ be the four-dimensional nilpotent Lie algebra with structure equations
\begin{equation}
[v_1,v_2]=v_3, \quad [v_2,v_4]=-v_1
\end{equation}
and $[v_i,v_j]=0$ if $\{i,j\}\neq \{1,2\}$ or $\{2,4\}$. In this case,
\begin{equation}
\widehat{\mathcal{T}}^{(0)}_{\varphi}=\alpha_1\alpha_2v_3-\alpha_2\alpha_4v_1\pm \text{perm.}
\end{equation}
\subsubsection{Example} The endomorphism $\varphi: D_1\rightarrow D_1$ defined by
$$\varphi(v_1)=0, \quad \varphi(v_2)=v_4, \quad \varphi(v_3)=v_1, \quad \varphi(v_4)=v_3,$$
$$\varphi(\alpha_1)=-\alpha_3,\quad \varphi(\alpha_2)=0, \quad \varphi(\alpha_3)=-\alpha_4, \quad \varphi(\alpha_4)=-\alpha_2$$
is skew-symmetric, has single indecomposable real block of type $\Delta_4^0(0,0)$
and minimal polynomial $P(x)=x^4$. Moreover, its Jordan chains are
$$v_2\longmapsto v_4\longmapsto v_3\longmapsto v_1\longmapsto 0,$$
$$-\alpha_1\longmapsto\alpha_3\longmapsto-\alpha_4\longmapsto\alpha_2\longmapsto 0.$$
As before, setting
$$\mathbf{b}_1^1=v_1, \quad \mathbf{b}_2^1=v_3, \quad \mathbf{b}_3^1=v_4,\quad \mathbf{b}_4^1=v_2$$
$$\mathbf{b}_1^2=\alpha_2, \quad \mathbf{b}_2^2=-\alpha_4, \quad \mathbf{b}_3^2=\alpha_3,\quad \mathbf{b}_4^2=-\alpha_1$$
we get
\begin{equation}\widehat{\mathcal{T}}^{(0)}_{\varphi}=-\mathbf{b}_4^2\mathbf{b}_1^2\mathbf{b}_2^1+\mathbf{b}_1^2\mathbf{b}_2^2\mathbf{b}_1^1 \pm \text{perm.}
\end{equation}
and
\begin{equation}
\widehat{\mathcal{C}}_{\varphi}=(\varphi_1+\varphi_3)^4\cdot(-\mathbf{b}_4^2\mathbf{b}_1^2\mathbf{b}_2^1)\pm \text{perm.}=-4\mathbf{b}_1^2\mathbf{b}_1^2\mathbf{b}_1^1\pm \text{perm.}=0\,,
\end{equation}
so that $\varphi$ is minimal. Moreover, for $n\geq 1$,
\begin{equation}
\widehat{\mathcal T}^{(n)}_{\varphi}=(\varphi_2\varphi_3+\varphi_3^2)^n(\mathbf{b}_1^2\mathbf{b}_4^2\mathbf{b}_2^2-\mathbf{b}_1^2\mathbf{b}_2^2\mathbf{b}_4^2)+(\varphi_1\varphi_3+\varphi_3^2)^n(-\mathbf{b}_4^2\mathbf{b}_1^2\mathbf{b}_2^2+\mathbf{b}_2^2\mathbf{b}_1^2\mathbf{b}_4^2)
\end{equation}
whence
\begin{equation}
\widehat{\mathcal T}^{(1)}_{\varphi}=\mathbf{b}_1^2\mathbf{b}_3^2\mathbf{b}_1^2-\mathbf{b}_1^2\mathbf{b}_2^1\mathbf{b}_3^2-\mathbf{b}_1^2\mathbf{b}_2^2\mathbf{b}_2^2-\mathbf{b}_3^2\mathbf{b}_1^2\mathbf{b}_1^2+\mathbf{b}_1^2\mathbf{b}_1^2\mathbf{b}_3^2+\mathbf{b}_2^2\mathbf{b}_1^2\mathbf{b}_2^2
\end{equation}
and $\widehat{\mathcal T}^{(n)}_{\varphi}=0$ for $n\geq 2$. Finally,
\begin{equation}\widehat{\mathcal S}^{(1)}_{\varphi}=\mathbf{b}_1^2\mathbf{b}_2^2\mathbf{b}_1^2-\mathbf{b}_1^2\mathbf{b}_1^1\mathbf{b}_3^2-\mathbf{b}_2^2\mathbf{b}_1^2\mathbf{b}_1^2\end{equation}
and $\widehat{\mathcal S}^{(n)}_{\varphi}=0$ for $n\geq 2$.
\begin{bibdiv}
\begin{biblist}
\bib{AldiGrandini15}{article}{
author={Aldi, Marco},
author={Grandini, Daniele},
title={Generalized contact geometry and T-duality},
journal={J. Geom. Phys.},
volume={92},
date={2015},
pages={78--93}
}
\bib{AldiGrandini16}{article}{
author={Aldi, Marco},
author={Grandini, Daniele},
title={An abstract Morimoto theorem for generalized $F$-structures},
journal={Q. J. Math.},
volume={67},
date={2016},
number={2},
pages={161--182}
}
\bib{AldiGrandini17}{article}{
author={Aldi, Marco},
author={Grandini, Daniele},
title={Generalized almost product structures and generalized
CRF-structures},
journal={J. Geom. Phys.},
volume={121},
date={2017},
pages={93--107}
}
\bib{Batista14}{article}{
author={Batista, Carlos},
title={Pure subspaces, generalizing the concept of pure spinors},
journal={J. Geom. Phys.},
volume={81},
date={2014},
pages={117--127}
}
\bib{BenedettiLisca18}{article}{
author={Benedetti, Riccardo},
author={Lisca, Paolo},
title={Framing 3-manifolds with bare hands},
journal={Enseign. Math.},
volume={64},
date={2018},
number={3-4},
pages={395--413},
}
\bib{BlagaCrasmareanu14}{article}{
author={Blaga, A. M.},
author={Crasmareanu, M.},
title={A class of almost tangent structures in generalized geometry},
journal={Balkan J. Geom. Appl.},
volume={19},
date={2014},
number={2},
pages={23--35}
}
\bib{BlagaNannicini20}{article}{
author={Blaga, Adara M.},
author={Nannicini, Antonella},
title={Generalized metallic structures},
journal={Rev. Un. Mat. Argentina},
volume={61},
date={2020},
number={1},
pages={73--86}
}
\bib{BoualemBrouzet12}{article}{
author={Boualem, Hassan},
author={Brouzet, Robert},
title={Semi-simple generalized Nijenhuis operators},
journal={J. Geom. Mech.},
volume={4},
date={2012},
number={4},
pages={385--395}
}
\bib{Buttin74}{article}{
author={Buttin, Claudette},
title={Th\'{e}orie des op\'{e}rateurs diff\'{e}rentiels gradu\'{e}s sur les formes
diff\'{e}rentielles},
journal={Bull. Soc. Math. France},
volume={102},
date={1974},
pages={49--73}
}
\bib{BurgoyneCushman77}{article}{
author={Nicholas Burgoyne},
author={Richard Cushman},
title={Conjugacy Classes in Linear Groups },
journal={Journal of Algebra},
volume={44},
date={1977},
pages={339--362}
}
\bib{Cavalcanti06}{article}{
author={Cavalcanti, Gil R.},
title={The decomposition of forms and cohomology of generalized complex
manifolds},
journal={J. Geom. Phys.},
volume={57},
date={2006},
number={1},
pages={121--132}
}
\bib{Courant90}{article}{
author={Courant, Theodore James},
title={Dirac manifolds},
journal={Trans. Amer. Math. Soc.},
volume={319},
date={1990},
number={2},
pages={631--661}
}
\bib{Eliopoulos65}{article}{
author={Eliopoulos, Hermes A.},
title={On the general theory of differentiable manifolds with almost
tangent structure},
journal={Canad. Math. Bull.},
volume={8},
date={1965},
pages={721--748}
}
\bib{GoldbergPetridis73}{article}{
author={Goldberg, Samuel I.},
author={Petridis, Nicholas C.},
title={Differentiable solutions of algebraic equations on manifolds},
journal={Kodai Math. Sem. Rep.},
volume={25},
date={1973},
pages={111--128}
}
\bib{GoldbergYano70}{article}{
author={Goldberg, Samuel I.},
author={Yano, Kentaro},
title={Polynomial structures on manifolds},
journal={Kodai Math. Sem. Rep.},
volume={22},
date={1970},
pages={199--218}
}
\bib{Gualtieri11}{article}{
author={Gualtieri, Marco},
title={Generalized complex geometry},
journal={Ann. of Math. (2)},
volume={174},
date={2011},
number={1},
pages={75--123}
}
\bib{Guttenberg07}{article}{
author={Guttenberg, Sebastian},
title={Brackets, sigma models and integrability of generalized complex
structures},
journal={J. High Energy Phys.},
date={2007},
number={6},
pages={004, 67 pp. (electronic)}
}
\bib{Haantjes55}{article}{
author={Haantjes, J.},
title={On $X_m$-forming sets of eigenvectors},
journal={Nederl. Akad. Wetensch. Proc. Ser. A. {\bf 58} = Indag. Math.},
volume={17},
date={1955},
pages={158--162},
}
\bib{Hitchin03}{article}{
author={Hitchin, Nigel},
title={Generalized Calabi-Yau manifolds},
journal={Q. J. Math.},
volume={54},
date={2003},
number={3},
pages={281--308}
}
\bib{Humphreys78}{book}{
author={Humphreys, James E.},
title={Introduction to Lie algebras and representation theory},
series={Graduate Texts in Mathematics},
volume={9},
note={Second printing, revised},
publisher={Springer-Verlag, New York-Berlin},
date={1978}
}
\bib{IshiharaYano65}{article}{
author={Yano, Kentaro},
author={Ishihara, Shigeru},
title={Structure defined by $f$ satisfying $f^{3}+f=0$},
conference={
title={Proc. U.S.-Japan Seminar in Differential Geometry},
address={Kyoto},
date={1965},
},
book={
publisher={Nippon Hyoronsha, Tokyo},
},
date={1966},
pages={153--166}
}
\bib{Kosmann-Schwarzbach04}{article}{
author={Kosmann-Schwarzbach, Yvette},
title={Derived brackets},
journal={Lett. Math. Phys.},
volume={69},
date={2004},
pages={61--87},
}
\bib{Kosmann-Schwarzbach19}{article}{
author={Kosmann-Schwarzbach, Yvette},
title={Beyond recursion operators},
conference={
title={Geometric methods in physics XXXVI},
},
book={
series={Trends Math.},
publisher={Birkh\"{a}user/Springer, Cham},
},
date={2019},
pages={167--180}
}
\bib{Lehmann-Lejeune66}{article}{
author={Lehmann-Lejeune, J.},
title={Int\'{e}grabilit\'{e} des $G$-structures d\'{e}finies par une $1$-forme
$0$-d\'{e}formable \`a valeurs dans le fibre tangent},
language={French},
journal={Ann. Inst. Fourier (Grenoble)},
volume={16},
date={1966},
number={fasc. 2},
pages={329--387}
}
\bib{Meinrenken13}{book}{
author={Meinrenken, Eckhard},
title={Clifford algebras and Lie theory},
series={Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A
Series of Modern Surveys in Mathematics [Results in Mathematics and
Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics]},
volume={58},
publisher={Springer, Heidelberg},
date={2013},
}
\bib{PoonWade11}{article}{
author={Poon, Yat Sun},
author={Wade, A{\"{\i}}ssa},
title={Generalized contact structures},
journal={J. Lond. Math. Soc. (2)},
volume={83},
date={2011},
number={2},
pages={333--352}
}
\bib{Tomasiello08}{article}{
author={Tomasiello, Alessandro},
title={Reformulating supersymmetry with a generalized Dolbeault operator},
journal={J. High Energy Phys.},
date={2008},
number={2},
pages={010, 25}
}
\bib{TempestaTondo18a}{article}{
author={Tempesta, Piergiulio},
author={Tondo, Giorgio},
title={A new family of higher-order Haantjes tensors and integrability},
eprint={arXiv:1809.05908v3}
}
\bib{TempestaTondo21}{article}{
author={Tempesta, Piergiulio},
author={Tondo, Giorgio},
title={Haantjes algebras and diagonalization},
journal={J. Geom. Phys.},
volume={160},
date={2021},
pages={103968, 21}
}
\bib{Vaisman08}{article}{
author={Vaisman, Izu},
title={Generalized CRF-structures},
journal={Geom. Dedicata},
volume={133},
date={2008},
pages={129--154}
}
\bib{Vanzura76}{article}{
author={Vanzura, Jiri},
title={Integrability conditions for polynomial structures},
journal={Kodai Math. Sem. Rep.},
volume={27},
date={1976},
number={1-2},
pages={42--50}
}
\bib{Vanzurova98}{article}{
author={Vanzurov\'{a}, Alena},
title={Differential forms on manifolds with a polynomial structure},
journal={Math. Slovaca},
volume={48},
date={1998},
number={5},
pages={527--533}
}
\bib{Vinogradov90}{article}{
author={Vinogradov, A. M.},
title={The union of the Schouten and Nijenhuis brackets, cohomology, and
superdifferential operators},
language={Russian},
journal={Mat. Zametki},
volume={47},
date={1990},
number={6},
pages={138--140},
}
\bib{Yano63}{article}{
author={Yano, Kentaro},
title={On a structure defined by a tensor field $f$ of type $(1,\,1)$
satisfying $f^{3}+f=0$},
journal={Tensor (N.S.)},
volume={14},
date={1963},
pages={99--109}
}
\end{biblist}
\end{bibdiv}
\vskip.1in\noindent
\address{Marco Aldi\\
Department of Mathematics and Applied Mathematics\\
Virginia Commonwealth University\\
Richmond, VA 23284, USA\\
\email{<EMAIL>}}
\vskip.1in\noindent
\address{Daniele Grandini\\
Department of Mathematics and Economics\\
Virginia State University\\
Petersburg, VA 23806, USA\\
\email{<EMAIL>}}
\end{document}
|
\section{Missing Proofs} \label{app:proofs}
We provide the complete proofs of the theorems stated in the main paper.
We defer the proofs of the technical results to \cref{app:tech}.
\subsection{Proof of \cref{th:performance decomposition}}
\PerformanceDecomposition*
First we prove the equality using a new performance difference lemma that we will prove in \cref{app:tech}. This result may be of independent interest.
\begin{restatable}[General Performance Difference Lemma]{lemma}{GPDL} \label{lm:gpdl}
Consider the reshaped MDP $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}$ defined by some $f:\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\to\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$ and $\lambda\in[0,1]$.
For any policy $\pi$, any state distribution $d_0$ and any $V:\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\to\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$, it holds that
\begin{align*}
V(d_0) - V^\pi(d_0)
&= \frac{\gamma(1-\lambda)}{1-\gamma} \Ebb_{s,a\sim d^\pi} \Ebb_{s'|s,a} \left[ h(s') - V(s')\right] \\
&\quad
+ \lambda \left( V(d_0) - \widetilde{V}^\pi(d_0) \right) + \frac{1-\lambda}{1-\gamma} \left( V(d^{\pi}) - \widetilde{V}^\pi(d^{\pi}) \right)
\end{align*}
\end{restatable}
Now take $V$ as $\widetilde{V}^*$ in the equality above.
Then we can write
\begin{align*}
V^*(d_0) - V^\pi(d_0) &= \left( V^*(d_0) - \widetilde{V}^*(d_0) \right) + \frac{\gamma(1-\lambda)}{1-\gamma} \Ebb_{s,a\sim d^\pi} \Ebb_{s'|s,a} \left[ h(s') - \widetilde{V}^*(s')\right] \\
&\quad + \lambda \left( \widetilde{V}^*(d_0) - \widetilde{V}^\pi(d_0) \right) + \frac{1-\lambda}{1-\gamma} \left( \widetilde{V}^*(d^{\pi}) - \widetilde{V}^\pi(d^{\pi}) \right)
\end{align*}
which is the regret-bias decomposition.
Next we prove that these two terms are independent of constant offsets. For the regret term, this is obvious because shifting the heuristic by a constant would merely shift the reward by a constant.
For the bias term, we prove the invariance below.
\begin{restatable}{proposition}{BiasShift}
$\mathrm{Bias}(h,\lambda,\pi) = \mathrm{Bias}(h+b,\lambda,\pi)$ for any $b\in\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$.
\end{restatable}
\begin{proof}
Notice that any $b\in\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$ and $\pi$,
$
\widetilde{V}^\pi(s;f+b) - \widetilde{V}^\pi(s;f) = \sum_{t=0}^\infty (\lambda\gamma)^t (1-\lambda)\gamma b = \frac{(1-\lambda)\gamma }{1-\lambda\gamma}b
$.
Therefore, we can derive
\begin{align*}
\mathrm{Bias}(h+b,\lambda,\pi) - \mathrm{Bias}(h,\lambda,\pi)
&= - \frac{(1-\lambda)\gamma }{1-\gamma\lambda} b + \frac{\gamma(1-\lambda)}{1-\gamma} \Ebb_{s,a\sim d^\pi} \Ebb_{s'|s,a} \left[ b - \frac{(1-\lambda)\gamma }{1-\gamma\lambda} b \right] \\
&= \frac{\gamma(1-\lambda)}{1-\gamma} b - \left( 1+ \frac{\gamma(1-\lambda)}{1-\gamma} \right) \frac{(1-\lambda)\gamma }{1-\gamma\lambda} b
\end{align*}
Since
\begin{align*}
\left( 1+ \frac{\gamma(1-\lambda)}{1-\gamma} \right) \frac{(1-\lambda)\gamma }{1-\gamma\lambda} b
= \frac{1-\gamma + \gamma(1-\lambda)}{1-\gamma} \frac{(1-\lambda)\gamma }{1-\gamma\lambda} b
= \frac{ 1- \gamma\lambda}{1-\gamma} \frac{(1-\lambda)\gamma }{1-\gamma\lambda} b
= \frac{ (1-\lambda)\gamma }{1-\gamma} b
\end{align*}
we have $\mathrm{Bias}(h+b,\lambda,\pi) - \mathrm{Bias}(h,\lambda,\pi)=0$.
\end{proof}
\subsection{Proof of \cref{th:regret as action gap}}
\RegretAsActionGap*
Define the Bellman backup for the reshaped MDP:
\begin{align*}
(\widetilde{\BB} V)(s,a) \coloneqq \widetilde{r}(s,a) + \widetilde{\gamma} \Ebb_{s'|s,a}[V(s')]
\end{align*}
Then by \cref{lm:online value difference} in \cref{app:tech}, we can rewrite the regret as
\begin{align*}
\lambda \left( \widetilde{V}^*(d_0) - \widetilde{V}^\pi(d_0) \right) + \frac{1-\lambda}{1-\gamma} \left( \widetilde{V}^*(d^{\pi}) - \widetilde{V}^\pi(d^{\pi}) \right)
= \Ebb_{\rho^\pi(d_0)} \left[ \sum_{t=0}^\infty \gamma^t \left( \widetilde{V}^*(s_t) - (\widetilde{\BB} \widetilde{V}^*)(s_t,a_t) \right) \right]
\end{align*}
Notice the equivalence $ \widetilde{V}^*(s) - (\widetilde{\BB} \widetilde{V}^*)(s,a) = - \widetilde{A}^*(s,a)$. This concludes the proof.
\subsection{Proof of \cref{th:preserved MDP properties}}
\PreservedProperties*
For the first statement, notice $\widetilde{r}(s,a) \in [0, 1 + \frac{(1-\lambda)\gamma}{1-\gamma}]$. Therefore, we have $\widetilde{V}^\pi(s) \geq 0$ as well as
\begin{align*}
\widetilde{V}^\pi(s)
&\leq \frac{1}{1-\lambda\gamma} \left( 1 + \frac{(1-\lambda)\gamma}{1-\gamma} \right)\\
&= \frac{1}{1-\lambda\gamma} \frac{1-\gamma + (1-\lambda)\gamma}{1-\gamma} = \frac{1}{1-\gamma}
\end{align*}
For the second statement, we just need to show the reshaped reward $\widetilde{r}(s,a)$ is linear in $\phi(s,a)$. This is straightforward because $\Ebb_{s'|s,a}[h(s')]$ is linear in $\phi(s,a)$.
\subsection{Proof of \cref{th:l-inf bias bound}}
\LinfBiasBound*
By \cref{th:performance decomposition}, we know that
$\mathrm{Bias}(h,\lambda,\pi) = \mathrm{Bias}(h+b,\lambda,\pi)$ for any $b\in\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$. Now consider $b^*\in\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$ such that $\| h +b^* - V^* \|_\infty \leq \epsilon$. Then by \cref{lm:value change}, we have also $
\| h + b^* - \widetilde{V}^{\pi^*} \|_\infty
\leq \epsilon + \frac{(1-\lambda)\gamma \epsilon}{1-\lambda\gamma}
$.
Therefore, by \cref{th:bias bound}, we can derive with definition of the bias,
\begin{align*}
\mathrm{Bias}(h,\lambda,\pi) &=
%
\mathrm{Bias}(h+b^*,\lambda,\pi) \\
%
&\leq
(1-\lambda)\gamma \left( \CC(\pi^*,V^*-h-b^*, \lambda\gamma) +
\CC(\pi,h+b^*-\widetilde{V}^*, \gamma)
\right)
\\
&\leq
(1-\lambda)\gamma \left( \CC(\pi^*,V^*-h-b^*, \lambda\gamma) +
\CC(\pi,h+b^*-\widetilde{V}^{\pi^*}, \gamma)
\right)
\\
&\leq
(1-\lambda)\gamma
\left(\frac{\epsilon}{1-\lambda\gamma} + \frac{1}{1-\gamma}
(\epsilon +
\frac{(1-\lambda)\gamma \epsilon}{1-\lambda\gamma}) \right)\\
&\leq
(1-\lambda)\gamma
\left(\frac{\epsilon}{1-\gamma} + \frac{1}{1-\gamma}
(\epsilon +
\frac{(1-\lambda)\gamma \epsilon}{1-\gamma}) \right)\\
&= \frac{2(1-\lambda)\gamma \epsilon}{1-\gamma}
+ \frac{(1-\lambda)^2\gamma^2 \epsilon}{(1-\gamma)^2}
\leq \frac{(1-\lambda\gamma)^2}{(1-\gamma)^2} \epsilon
\end{align*}
\subsection{Proof of \cref{th:bias bound}}
\BiasBound*
Recall the definition of bias:
\begin{align*}
\mathrm{Bias}(h,\lambda,\pi) =
\left( V^*(d_0) - \widetilde{V}^*(d_0) \right) + \frac{\gamma(1-\lambda)}{1-\gamma} \Ebb_{s,a\sim d^\pi} \Ebb_{s'|s,a} \left[ h(s') - \widetilde{V}^*(s')\right]
\end{align*}
For the first term, we can derive by performance difference lemma (\cref{lm:pdl}) and \cref{lm:value difference}
\begin{align*}
V^*(d_0) - \widetilde{V}^*(d_0) &\leq V^*(d_0) - \widetilde{V}^{\pi^*}(d_0) \\
&= (1-\lambda)\gamma \Ebb_{\rho^{\pi^*}(d_0)} \left[ \sum_{t=1}^\infty (\lambda\gamma)^{t-1} ( V^*(s_t) -h(s_t)) \right]
= (1-\lambda)\gamma \CC(\pi, V^*-f, \lambda\gamma)
\end{align*}
For the second term, we can rewrite it as
\begin{align*}
\frac{\gamma(1-\lambda)}{1-\gamma} \Ebb_{s,a\sim d^\pi} \Ebb_{s'|s,a} \left[ h(s') - \widetilde{V}^*(s')\right]
&= \gamma(1-\lambda) \Ebb_{\rho^{\pi}(d_0)} \left[ \sum_{t=1}^\infty \gamma^{t-1} ( h(s_t) - \widetilde{V}^*(s_t) ) \right] \\
&= (1-\lambda)\gamma \CC(\pi^*, f-\widetilde{V}^*, \gamma)
\end{align*}
\subsection{Proof of \cref{th:improvable heuristic}}
\ImprovableHeuristic*
Let $d_t^\pi(s;s_0)$ denote the state distribution at the $t$th step after running $\pi$ starting from $s_0\in\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}$ in $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$ (i.e. $d_0^\pi(s;s_0) = {\mathbbm1}\{s=s_0\}$). Define the mixture
\begin{align} \label{eq:average state distribution in reshaped mdp}
\widetilde{d}_{s_0}^{\pi}(s) \coloneqq (1-\widetilde{\gamma})\sum_{t=0}^\infty \widetilde{\gamma}^t d_t^\pi(s ; s_0)
\end{align}
where we recall the new discount $\widetilde{\gamma}=\gamma\lambda$
By performance difference lemma (\cref{lm:pdl}), we can write for any policy $\pi$ and any $s_0 \in \smash{}$
\begin{align*}
\widetilde{V}^\pi(s_0) - h(s_0) = \frac{1}{1-\lambda\gamma} \Ebb_{\widetilde{d}_{s_0}^{\pi}}[ (\widetilde{\BB} h)(s,a) - h(s)]
\end{align*}
Notice that
\begin{align*}
(\widetilde{\BB} h)(s,a)
&=\widetilde{r}(s,a) + \widetilde{\gamma} \Ebb_{s'|s,a} [h(s')] \\
&= r(s,a) + (1-\lambda) \gamma \Ebb_{s'|s,a}[h(s')] + \lambda\gamma \Ebb_{s'|s,a} [h(s')] \\
&= r(s,a) + \gamma \Ebb_{s'|s,a}[h(s')] = (\BBh)(s,a)
\end{align*}
Let $\pi$ denote the greedy policy of $\argmax_a (\BB h) (s,a)$.
Then we have, by the improvability assumption we have $(\BBh)(s,\pi) - h(s)\geq 0$ and therefore,
\begin{align*}
\widetilde{V}^*(s_0) \geq \widetilde{V}^\pi(s_0)
&= h(s_0) + \frac{1}{1-\lambda\gamma} \Ebb_{\widetilde{d}_{s_0}^{\pi}}[ (\widetilde{\BB} h)(s,a) - h(s)] \\
&= h(s_0) + \frac{1}{1-\lambda\gamma} \Ebb_{\widetilde{d}_{s_0}^{\pi}}[ (\BBh)(s,a) - h(s)] \\
&\geq h(s_0)
\end{align*}
Since $s_0$ is arbitrary above, we have the desired statement.
\subsection{Proof of \cref{th:bellman pessimism and improvable heuristic}}
\BellmanPessimismAndImprovableHeuristic*
The proof is straightforward:
We have $\max_a (\BBh)(s,a) \geq (\BBh)(s,\pi) \geq Q(s,\pi) = h(s)$, which is the definition of $h$ being improvable. For the argument of uniform lower bound, we chain the assumption $Q(s,a)\leq (\BB h)(s,a)$:
\begin{align*}
h(s) = Q(s,\pi') &=
r(s,\pi') + \gamma \Ebb_{s'|s,\pi'}[h(s')] \\
&\leq r(s,\pi') + \gamma \left(r(s',\pi'), + \gamma \Ebb_{s''|s',\pi'}[h(s'')] \right) \\
&\leq V^{\pi'}(s)
\end{align*}
\section{Technical Lemmas} \label{app:tech}
\subsection{Lemmas of Performance Difference}
Here we prove a general performance difference for the $\lambda$-weighting used in the reshaped MDPs.
\GPDL*
Our new lemma includes the two below performance difference lemmas in the literature as special cases.
\cref{pr:lambda weighted pdf} can be obtained by setting $V=f$; \cref{lm:pdl} can be obtained by further setting $\lambda=0$ (that is, \cref{lm:pdl} is a special case of \cref{pr:lambda weighted pdf} with $\lambda=0$; and \cref{lm:gpdl} generalizes both). The proofs of these existing performance difference lemmas do not depend on the new generalization in \cref{lm:gpdl}, please refer to~\citep{kakade2002approximately,cheng2020policy} for details.
\begin{lemma}[Performance Difference Lemma~\citep{kakade2002approximately,cheng2020policy} ]\label{lm:pdl}
For any policy $\pi$, any state distribution $d_0$ and any $V:S\to\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$, it holds that
\begin{align*}
V(d_0) - V^\pi(d_0) = \frac{1}{1-\gamma} \Ebb_{d^\pi}[V(s) - (\BB V) (s,a)]
\end{align*}
\end{lemma}
\begin{restatable}[$\lambda$-weighted Performance Difference Lemma~\citep{cheng2020policy}]{lemma}{LambdaWeightedPDL} \label{pr:lambda weighted pdf}
For any policy $\pi$, $\lambda\in[0,1]$, and $f:\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\to\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$, it holds that
\begin{align*}
f(d_0) - V^\pi(d_0)
&= \lambda \left( f(d_0) - \widetilde{V}^\pi(d_0) \right) + \frac{1-\lambda}{1-\gamma} \left( f(d^{\pi}) - \widetilde{V}^\pi(d^{\pi}) \right)
\end{align*}
\end{restatable}
\subsubsection{Proof of \cref{lm:gpdl}}
First, we use the standard performance difference lemma (\cref{lm:pdl}) in the original MDP and \cref{lm:Bellman backup difference} for the first and the last steps below,
\begin{align*}
V(d_0) - V^\pi(d_0)
&= \frac{1}{1-\gamma} \Ebb_{d^{\pi}} \left[ V(s) - (\BB V)(s,a) \right]\\
&= \frac{1}{1-\gamma} \Ebb_{d^{\pi}} \left[ (\widetilde{\BB} V)(s,a) - (\BB V) (s,a) \right] + \frac{1}{1-\gamma} \Ebb_{d^{\pi}} \left[ V(s) - (\widetilde{\BB} V)(s,a) \right] \\
&= \frac{\gamma(1-\lambda)}{1-\gamma} \Ebb_{s,a\sim d^\pi} \Ebb_{s'|s,a} \left[ h(s') - V(s')\right]
+ \frac{1}{1-\gamma} \Ebb_{s,a\sim d^{\pi}} \left[ V(s) - (\widetilde{\BB} V)(s,a) \right]
\end{align*}
Finally, substituting the equality in \cref{lm:online value difference} into the above equality concludes the proof.
\subsection{Properties of reshaped MDP}
The first lemma is the difference of Bellman backups.
\begin{lemma} \label{lm:Bellman backup difference}
For any $V:\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\to\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$,
\begin{align*}
(\BB V)(s,a) - (\widetilde{\BB} V) (s,a)
&= (1-\lambda) \gamma \Ebb_{s'|s,a}[ V(s') -h(s')]
\end{align*}
\end{lemma}
\begin{proof}
The proof follows from the definition of the reshaped MDP:
\begin{align*}
&(\BB V)(s,a) - (\widetilde{\BB} V) (s,a)\\
&= r(s,a) + \gamma \Ebb_{s'|s,a}[ V(s') ] - r(s,a) - (1-\lambda)\gamma \mathbb{E}_{s'|s,a}[h(s')] - \gamma \lambda \Ebb_{s'|s,a}[ V(s') ] \\
&= (1-\lambda) \gamma \Ebb_{s'|s,a}[ V(s') -h(s')]
\end{align*}
\end{proof}
This lemma characterizes, for a policy, the difference in returns.
\begin{lemma} \label{lm:value difference}
For any policy $\pi$ and $h:\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\to\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$,
\begin{align*}
V^\pi(s) - \widetilde{V}^{\pi}(s)
&=(1-\lambda)\gamma \Ebb_{\rho^{\pi}(s)} \left[ \sum_{t=1}^\infty (\lambda\gamma)^{t-1} ( V^\pi(s_t) -h(s_t)) \right]
\end{align*}
\end{lemma}
\begin{proof}
The proof is based on performance difference lemma (\cref{lm:pdl}) applied in the reshaped MDP and \cref{lm:Bellman backup difference}.
Recall the definition $\widetilde{d}_{s_0}^\pi(s)$ in \eqref{eq:average state distribution in reshaped mdp} and define $\widetilde{d}_{s_0}^\pi(s,a) = \widetilde{d}_{s_0}^\pi(s)\pi(a|s)$.
For any $s_0\in\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}$,
\begin{align*}
V^\pi(s_0) - \widetilde{V}^{\pi}(s_0)
&= \frac{1}{1-\gamma\lambda} \Ebb_{s,a\sim\widetilde{d}_{s_0}^{\pi}} [ V^\pi(s) - \widetilde{\BB} V^\pi (s,a) ]\\
&= \frac{1}{1-\gamma\lambda} \Ebb_{s,a\sim\widetilde{d}_{s_0}^{\pi}} [ (\BB V^\pi)(s,a) - (\widetilde{\BB} V^\pi) (s, a) ] \\
&= \frac{(1-\lambda)\gamma}{1-\gamma\lambda} \Ebb_{s,a\sim\widetilde{d}_{s_0}^{\pi}} \Ebb_{s'|s,a}[ V^\pi(s') -h(s')]
\end{align*}
Finally, substituting the definition of $\widetilde{d}_{s_0}^{\pi}$ finishes the proof.
\end{proof}
A consequent lemma shows that $ h$ and $\widetilde{V}^{\pi}$ are close, when $h$ and $V^\pi$ are.
\begin{lemma} \label{lm:value change}
For a policy $\pi$, suppose $-\epsilon_l \leq h(s) - V^\pi(s) \leq \epsilon_u$. It holds
\begin{align*}
-\epsilon_l - \frac{(1-\lambda)\gamma \epsilon_u}{1-\lambda\gamma} \leq h(s) - \widetilde{V}^{\pi}(s)
\leq \epsilon_u + \frac{(1-\lambda)\gamma \epsilon_l}{1-\lambda\gamma}
\end{align*}
\end{lemma}
\begin{proof}
We prove the upper bound by \cref{lm:value difference}; the lower bound can be shown by symmetry.
\begin{align*}
h(s) - \widetilde{V}^{\pi}(s)
&\leq \epsilon_u + V^\pi(s) - \widetilde{V}^{\pi}(s) \\
&= \epsilon_u + (1-\lambda)\gamma \Ebb_{\rho^{\pi}(s)} \left[ \sum_{t=1}^\infty (\lambda\gamma)^{t-1} ( V^\pi(s_t) -h(s_t)) \right]\\
&\leq \epsilon_u + \frac{(1-\lambda)\gamma \epsilon_l}{1-\lambda\gamma}
\end{align*}
\end{proof}
The next lemma relates online Bellman error to value gaps.
\begin{lemma} \label{lm:online value difference}
For any $\pi$ and $V:\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\to\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$,
\begin{align*}
\frac{1}{1-\gamma} \left( \Ebb_{d^{\pi}} \left[ V(s) - (\widetilde{\BB} V)(s,a) \right] \right) = \lambda \left( V(d_0) - \widetilde{V}^\pi(d_0) \right) + \frac{1-\lambda}{1-\gamma} \left( V(d^{\pi}) - \widetilde{V}^\pi(d^{\pi}) \right)
\end{align*}
\end{lemma}
\begin{proof}
We use \cref{lm:Bellman backup difference} in the third step below.
\begin{align*}
&\Ebb_{d^{\pi}} \left[ V(s) - (\widetilde{\BB} V)(s,a) \right]\\
&=
\Ebb_{d^{\pi}} \left[ V(s) - (\widetilde{\BB} \widetilde{V}^\pi)(s,a) \right] + \Ebb_{d^{\pi}} \left[ \widetilde{\BB} \widetilde{V}^\pi(s,a) - (\widetilde{\BB} V)(s,a) \right] \\
&= \Ebb_{d^{\pi}} \left[ V(s) - \widetilde{V}^\pi(s) \right] + \Ebb_{d^{\pi}} \left[ (\widetilde{\BB} \widetilde{V}^\pi)(s,a) - (\widetilde{\BB} V)(s) \right] \\
&= \Ebb_{d^{\pi}} \left[ V(s) - \widetilde{V}^\pi(s) \right] - \lambda\gamma \Ebb_{s,a\sim d^{\pi}} \Ebb_{s'|s,a} \left[ \widetilde{V}^\pi(s') - V(s') \right]\\
&= (1-\gamma) \Ebb_{\rho^\pi(d_0)}\left[ \sum_{t=0}^\infty \gamma^t (V(s_t) - \widetilde{V}^\pi(s_t)) - \lambda \gamma^{t+1} ( \widetilde{V}^\pi(s_{t+1}) - V(s_{t+1}) ) \right]\\
&= (1-\gamma)\lambda( V(d_0) - \widetilde{V}^\pi(d_0) ) + (1-\gamma)(1-\lambda) \Ebb_{\rho_\pi(d_0)}\left[ \sum_{t=0}^\infty \gamma^t (V(s_t) - \widetilde{V}^\pi(s_t)) \right]
\end{align*}
\end{proof}
\section{Experiments} \label{app:experiments}
\subsection{Details of the MuJoCo Experiments} \label{app:details}
We consider four dense reward MuJoCo environments (Hopper-v2, HalfCheetah-v2, Humanoid-v2, and Swimmer-v2) and a sparse reward version of Reacher-v2.
\rev{
In the sparse reward Reacher-v2, the agent receives a reward of $0$ at the goal state (defined as $\norm{g(s)-e(s)}\leq 0.01$ and $-1$ elsewhere, where $g(s)$ and $e(s)$ denote the goal state and the robot's end-effector positions, respectively. We designed a heuristic $h(s) = r(s,a) - 100 \norm{e(s)-g(s)} $, as this is a goal reaching task. Here the policy is randomly initialized, as no prior batch data is available before interactions.
}
In the dense reward experiments, we suppose that a batch of data collected by multiple behavioral policies are available before learning, and a heuristic is constructed by an offline policy evaluation algorithm from the batch data; in the experiments, we generated these behavioral policies by running SAC\xspace from scratch and saved the intermediate policies generated in training.
We designed this heuristic generation experiment to simulate the typical scenario where offline data collected by multiple policies of various qualities is available before learning. In this case, a common method for inferring what values a good policy could get is to inspect the realized accumulated rewards in the dataset.
For simplicity, we use basic Monte Carlo regression to construct heuristics, where a least squares regression problem was used to fit a fully connected neural network to predict the empirical returns on the trajectories in the sampled batch of data.
\rev{Specifically, for each dense reward Mujoco experiment, we ran SAC for 200 iterations and logged the intermediate policies for every 4 iterations, resulting in a total of 50 behavior policies. In each random seed of the experiment, we performed the following: We used each behavior policy to collect trajectories of at most 10,000 transition tuples, which gave about 500,000 offline data points over these 50 policies. These data were used to construct the Monte-Carlo regression data, which was done by computing the accumulated discounted rewards along sampled trajectories. Then we generated the heuristic used in the experiment by fitting a fully connected NN with (256,256)-hidden layers using default ADAM with step size 0.001 and minibatch size 128 for 30 epochs over this randomly generated dataset of 50 behavior policies.}
For the dense reward Mujoco experiments, we also use behavior cloning (BC\xspace) with $\ell_2$ loss to warm start RL agents based on the same batch dataset of 500,000 offline data points.
The base RL algorithm here is SAC, which is based on the standard implementation of Garage (MIT License)~\cite{garage}.
The policy and the value networks are fully connected neural networks, independent of each other. The policy is Tanh-Gaussian and the value network has a linear head.
\paragraph{Algorithms.}
We compare the performance of different algorithms below.
\begin{enumerate*}[label=\textit{\arabic*)}]
\item BC\xspace
\item SAC
\item SAC with BC\xspace warm start (SAC w/ BC\xspace)
\item HuRL\xspace with a zero heuristic and BC\xspace warm start (HuRL\xspace-zero\xspace)
\item HuRL\xspace with the Monte-Carlo heuristic and BC\xspace warm start (HuRL\xspace-MC\xspace).
\end{enumerate*}
For the HuRL\xspace algorithms, the mixing coefficient $\lambda_n$ is scheduled as
\begin{align*}
\lambda_n &= \lambda_0 + (1-\lambda_0)\tanh\left( \frac{n-1}{\alpha N-1} \times \arctan(0.99)\right) / 0.99 \\
&\eqqcolon \lambda_0 + (1-\lambda_0) c_\omega
\tanh(\omega (n-1) )
\end{align*}
for $n=1,\dots,N$, where $\lambda_0 \in [0,1]$ is the initial $\lambda$ and $\alpha>0$ controls the increasing rate.
This schedule ensures that $\lambda_N=1$ when $\alpha=1$. Increasing $\alpha$ from $1$ makes $\lambda_n$ converge to $1$ slower.
We chose these algorithms to illustrate the effect of each additional warm-start component (BC\xspace and heuristics) added on top of the base algorithm SAC\xspace.
HuRL\xspace-zero\xspace is SAC w/ BC\xspace but with an extra $\lambda$ schedule described above that further lowers the discount, whereas SAC\xspace and SAC w/ BC\xspace keep a constant discount factor.
\paragraph{Evaluation and Hyperparameters.}
In each iteration, the RL agent has a fixed sample budget for environment interactions, and its performance is measured in terms of the undiscounted accumulated rewards (estimated by 10 rollouts) of the deterministic mean policy extracted from SAC\xspace.
The hyperparameters used in the algorithms above were selected as follows. The selection was done by uniformly random grid search\footnote{We ran 300 and 120 randomly chosen configurations from \cref{tb:hp grid search} with different random seeds to tune the base algorithm and the $\lambda$-scheduler, respectively. Then the best configuration was used in the following experiments.} over the range of hyperparameters in \cref{tb:hp grid search} to maximize the AUC of the training curve.
\begin{table}[ht]
\begin{center}
\begin{tabular}{ c| c }
Polcy step size & [0.00025, 0.0005, 0.001, 0.002] \\
Value step size & [0.00025, 0.0005, 0.001, 0.002] \\
Target step size & [0.005, 0.01, 0.02, 0.04] \\
$\gamma$ & [0.9, 0.99, 0.999] \\
$\lambda_0$ &[0.90, 0.95, 0.98, 0.99] \\
$\alpha$ & [$10^{-5}$, 1.0, $10^{5}$] \\
\end{tabular}
\end{center}
\caption{HuRL\xspace's hyperparameter value grid for the MuJoCo experiments.}
\label{tb:hp grid search}
\end{table}
First, the learning rates (policy step size, value step size, target step size) and the discount factor of the base RL algorithm, SAC\xspace, were tuned for each environment to maximize the performance. This tuned discount factor is used as the de facto discount factor $\gamma$ of the original MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$.
Fixing the hyperparameters above, $\lambda_0$ and $\alpha$ for the $\lambda$ schedule of HuRL\xspace were tuned for each environment and each heuristic.
The tuned hyperparameters and the environment specifications are given in \cref{tb:sparse mujoco exp configs,tb:mujoco exp configs} below. (The other hyperparameters, in addition to the ones tuned above, were selected manually and fixed throughout all the experiments).
\begin{table}[ht]
\begin{center}
\begin{tabular}{ c| c }
Environment & Sparse-Reacher-v2 \\\hline
Obs. Dim & 11 \\
Action Dim & 2 \\
Evaluation horizon & 500 \\
$\gamma$ & 0.9 \\
Batch Size & 10000 \\
Policy NN Architecture & (64,64) \\
Value NN Architecture & (256,256) \\
Polcy step size & 0.00025 \\
Value step size & 0.00025\\
Target step size & 0.02\\
Minibatch Size & 128 \\
Num. of Grad. Step per Iter. & 1024 \\ \hline
HuRL\xspace $\lambda_0$ & 0.5 \\
HuRL\xspace-MC $\alpha$ & $10^5$ \\
\end{tabular}
\end{center}
\caption{Sparse reward MuJoCo experiment configuration details. All the values other than $\lambda$-scheduler's (i.e. those used in SAC\xspace) are shared across different algorithms in the comparison. All the neural networks here fully connected and have $\tanh$ activation; the numbers of hidden nodes are documented above.
Note that when $\alpha=10^5$, effectively $\lambda_n=\lambda_0$ in the training iterations; when $\alpha=10^{-5}$, $\lambda_n \approx 1 $ throughout.
}
\label{tb:sparse mujoco exp configs}
\end{table}
\begin{table}[ht]
\begin{center}
\begin{tabular}{ c| c | c| c| c }
Environment & Hopper-v2 & HalfCheetah-v2 & Swimmer-v2 & Humanoid-v2 \\\hline
Obs. Dim & 11 & 17 & 8 &376\\
Action Dim & 3 & 6 & 2 & 17\\
Evaluation horizon & 1000 & 1000 & 1000 & 1000 \\
$\gamma$ & 0.999 & 0.99 & 0.999 & 0.99 \\
Batch Size & 4000 & 4000 & 4000 & 10000 \\
Policy NN Architecture & (64,64) & (64,64) & (64,64) & (256,256) \\
Value NN Architecture & (256,256) & (256,256) & (256,256) & (256,256) \\
Polcy step size & 0.00025 & 0.00025 & 0.0005 & 0.002 \\
Value step size & 0.0005 & 0.0005 & 0.0005 & 0.00025\\
Target step size & 0.02 & 0.04 & 0.0100 & 0.02\\
Num. of Behavioral Policies & 50 & 50 & 50 & 50 \\
Minibatch Size & 128 & 128 & 128 & 128 \\
Num. of Grad. Step per Iter. & 1024 & 1024 & 1024 & 1024 \\ \hline
HuRL\xspace-MC $\lambda_0$ & 0.95 & 0.99 & 0.95 & 0.9 \\
HuRL\xspace-MC $\alpha$ & $10^5$ & $10^5$ & 1.0 & 1.0 \\
HuRL\xspace-zero $\lambda_0$ & 0.98 & 0.99 & 0.99 & 0.95\\
HuRL\xspace-zero $\alpha$ & $10^{-5}$ & $10^5$ & 1.0 & $10^{-5}$\\
\end{tabular}
\end{center}
\caption{Dense reward MuJoCo experiment configuration details. All the values other than $\lambda$-scheduler's (i.e. those used in SAC\xspace) are shared across different algorithms in the comparison. All the neural networks here fully connected and have $\tanh$ activation; the numbers of hidden nodes are documented above.
Note that when $\alpha=10^5$, effectively $\lambda_n=\lambda_0$ in the training iterations; when $\alpha=10^{-5}$, $\lambda_n \approx 1 $ throughout.
}
\label{tb:mujoco exp configs}
\end{table}
Finally, after all these hyperparameters were decided, we conducted additional testing runs with 30 different random seeds and report their statistics here. The randomness include the data collection process of the behavioral policies, training the heuristics from batch data, BC, and online RL, but the behavioral policies are fixed.
\rev{
While this procedure takes more compute (the computation resources are reported below; tuning the base SAC takes the most compute),
it produces more reliable results
without (luckily or unluckily) using some hand-specified hyperparameters or a particular way of aggregating scores when tuning hyperparameters across environments.
Empirically, we also found using constant $\lambda$ around $0.95 \sim 0.98$ leads to
good performance, though it may not be the best environment-specific choice.
}
\paragraph{Resources.}
Each run of the experiment was done using an Azure Standard\_H8 machine (8 Intel Xeon E5 2667 CPUs; memory 56 GB; base frequency 3.2 GHz; all cores peak frequency 3.3 GHz; single core peak frequency 3.6 GHz). The Hopper-v2, HalfCheetah-v2, Swimmer-v2 experiments took about an hour per run. The Humanoid-v2 experiments took about 4 hours.
No GPU was used.
\paragraph{Extra Experiments with VAE-based Heuristics.}
We conduct additional experiments of HuRL\xspace using a VAE-filtered pessimistic heuristic. This heuristic is essentially the same as the Monte-Carlo regression-based heuristic we discussed, except that an extra VAE (variational auto-encoder) is used to classify states into known and unknown states in view of the batch dataset, and then the predicted values of unknown states are set to be the lowest empirical return seen in the dataset. In implementation, this is done by training a state VAE (with a latent dimension of 32) to model the states in the batch data, and then a new state classified as unknown if its VAE loss is higher than 99-th percentile of the VAE losses seen on the batch data. The implementation and hyperparameters are based on the code from \citet{liu2020provably}. We note, however, that this basic VAE-based heuristic does not satisfy the assumption of \cref{th:bellman pessimism and improvable heuristic}.
These results are shown in \cref{fig:extra results}, where HuRL\xspace-VAEMC denotes HuRL\xspace using this VAE-based heuristic. Overall, we see that such a basic pessimistic estimate does not improve the performance from the pure Monte-Carlo version (HuRL\xspace-MC\xspace); while it does improve the results slightly in HalfCheetah-v2, it gets worse results in Humanoid-v2 and Swimmer-v2 compared with HuRL\xspace-MC\xspace. Nonetheless, HuRL\xspace-VAEMC is still better than the base SAC\xspace.
\begin{figure*}[th]
\centering
\begin{subfigure}{0.245\textwidth}
\includegraphics[width=\textwidth]{figures/vae_plot_test-hopper.png}
\caption{Hopper-v2}
\label{fig:extra hopper}
\end{subfigure}
\begin{subfigure}{0.245\textwidth}
\includegraphics[width=\textwidth]{figures/vae_plot_test-humanoid.png}
\caption{Humanoid-v2}
\label{fig:extra humanoid}
\end{subfigure}
\begin{subfigure}{0.245\textwidth}
\includegraphics[width=\textwidth]{figures/vae_plot_test-swimmer.png}
\caption{Swimmer-v2}
\label{fig:extra swimmer}
\end{subfigure}
\begin{subfigure}{0.245\textwidth}
\includegraphics[width=\textwidth]{figures/vae_plot_test-halfcheetah.png}
\caption{HalfCheetah-v2}
\label{fig:extra halfcheetah}
\end{subfigure}
\caption{\small{Extra experimental results of different MuJoCo environments. The plots show the $25$th, $50$th, $75$th percentiles of each algorithm's performance over 30 random seeds.}}
\label{fig:extra results}
\end{figure*}
\subsection{Procgen Experiments}
In addition to MuJoCo environments, where the agent has direct access to the true low-dimensional system state, we conducted experiments on the Procgen benchmark suite \cite{cobbe2020leveraging, procgen}. The Procgen suite consists of 16 procedurally generated Atari-like game environments, whose main conceptual differences from MuJoCo environments are partial observability and much higher dimensionality of agents' observations (RGB images). The 16 games are very distinct structurally, but each game has an unlimited number of levels\footnote{In Procgen, levels aren't ordered by difficulty. They are merely game variations.} that share common characteristics. All levels of a given game are situated in the same underlying state space and have the same transition function but differ in terms of the regions of the state space reachable within each level and in their observation spaces. We focus on the \emph{sample efficiency} Procgen mode \cite{procgen}: in each RL episode the agent faces a new game level, and is expected to eventually learn a single policy that performs well across all levels of the given game.
Besides the differences in environment characteristics between MuJoCo and Procgen, the Procgen experiments are also dissimilar in their design:
\begin{itemize}
\item In contrast to the MuJoCo experiments, where we assumed to be given a batch of data from which we constructed a heuristic and a warm-start policy, in the Procgen experiments we simulate a scenario where we are given \emph{only} the heuristic function itself. Thus, we don't warm-start the base algorithm with a BC policy when running HuRL\xspace.
\item In the Procgen experiments, we share a single set of all hyperparameters' values -- those of the base algorithm, those of HuRL\xspace's $\lambda$-scheduling, and those used for generating heuristics -- across all 16 games. This is meant to simulate a scenario where HuRL\xspace is applied across a diverse set of problems using good but problem-independent hyperparameters.
\end{itemize}
\paragraph{Algorithms.}
We used PPO \cite{schulman2017proximal} implemented in RLlib (Apache License 2.0) \cite{pmlr-v80-liang18b} as the base algorithm. We generated a heuristic for each game as follows:
\begin{itemize}
\item We ran PPO for $8M$ environment interaction steps and saved the policy after every $500K$ steps, for a total of 16 checkpoint policies.
\item We ran the policies in a random order by executing 12000 environment interaction steps using each policy. For each rollout trajectory, we computed the discounted return for each observation in that trajectory, forming $\langle observation, return \rangle$ training pairs.
\item We used this data to learn a heuristic via regression. We mixed the data, divided it into batches of 5000 training pairs and took a gradient step w.r.t. MSE computed over each batch. The learning rate was $10^{-4}$.
\end{itemize}
Our main algorithm, a HuRL\xspace flavor denoted as PPO-HuRL\xspace, is identical to the base PPO but uses the Monte-Carlo heuristic computed as above.
\paragraph{Hyperparameters and evaluation} The base PPO's hyperparameters in RLlib were chosen to match PPO's performance reported in the original Procgen paper~\cite{cobbe2020leveraging} for the "easy" mode as closely as possible across all 16 games (\citet{cobbe2020leveraging} used a different PPO implementation with a different set of hyperparameters). As in that work, our agent used the IMPALA-CNN$ \times 4$ network architecture~\cite{pmlr-v80-espeholt18a,cobbe2020leveraging} without the LSTM. The heuristics employed the same architecture as well. We used a single set of hyperparameter values, listed in \cref{tb:ppo exp configs}, for all Procgen games, both for policy learning and for generating the checkpoints for computing the heuristics.
\begin{table}[h]
\begin{center}
\begin{tabular}{c|c}
\hline
Impala layer sizes & 16, 32, 32 \\
Rollout fragment length & 256 \\
Number of workers & 0 \emph{(in RLlib, this means 1 rollout worker)} \\
Number of environments per worker & 64 \\
Number of CPUs per worker & 5 \\
Number of GPUs per worker & 0 \\
Number of training GPUs & 1 \\
$\gamma$ & 0.99 \\
SGD minibatch size & 2048 \\
Train batch size & 4000 \\
Number of SGD iterations & 3 \\
SGD learning\ rate & 0.0005 \\
Framestacking & off \\
Batch mode & truncate\_episodes \\
Value function clip parameter &10.0 \\
Value function loss coefficient &0.5 \\
Value function share layers &true \\
KL coefficient &0.2 \\
KL target &0.01 \\
Entropy coefficient &0.01 \\
Clip parameter &0.1 \\
Gradient clip &null \\
Soft horizon & False \\
No done at end: & False \\
Normalize actions & False \\
Simple optimizer & False \\
Clip rewards & False \\
GAE $\lambda$ &0.95 \\ \hline
PPO-HuRL\xspace\ $\lambda_0$ & 0.99 \\
PPO-HuRL\xspace\ $\alpha$ & 0.5 \\
\end{tabular}
\end{center}
\caption{Procgen experiment configuration details: RLlib PPO's and HuRL\xspace's hyperparameter values. All the values were shared across all 16 Procgen games.
}
\label{tb:ppo exp configs}
\end{table}
\begin{table}[h]
\begin{center}
\begin{tabular}{ c| c }
$\lambda_0$ &[0.95, 0.97, 0.985, 0.98, 0.99] \\
$\alpha$ & [0.5, 0.75, 1.0, 3.0, 5.0] \\
\end{tabular}
\end{center}
\caption{HuRL\xspace's hyperparameter value grid for the Procgen experiments.}
\label{tb:ppo_hurl}
\end{table}
In order to choose values for PPO-HuRL\xspace's hyperparameters $\alpha$ and $\lambda_0$, we fixed all of PPO's hyperparameters, took the pre-computed heuristic for each game, and did a grid search over $\alpha$ and $\lambda_0$'s values listed in \cref{tb:ppo_hurl} to maximize the normalized average AUC across all games. To evaluate each hyperparameter value combination, we used 4 runs per game, each run using a random seed and lasting 8M environment interaction steps. The resulting values are listed in \cref{tb:ppo exp configs}. Like PPO's hyperparameters, they were kept fixed for all Procgen environments.
To obtain experimental results, we ran PPO and PPO-HuRL\xspace\ with the aforementioned hyperparameters on each of 16 games 20 times, each run using a random seed and lasting 8M steps as in \citet{procgen}. We report the 25th, 50th, and 75th-percentile training curves. Each of the reported training curves was computed by smoothing policy performance in terms of unnormalized game scores over the preceding 100 episodes.
\paragraph{Resources.}
Each policy learning run used a single Azure ND6s machine (6 Intel Xeon E5-2690 v4 CPUs with 112 GB memory and base core frequency of 2.6 GHz; 1 P40 GPU with 24 GB memory). A single PPO run took approximately 1.5 hours on average. A single PPO-HuRL\xspace\ run took approximately 1.75 hours.
\paragraph{Results.} The results are shown in \cref{fig:procgen}. They indicate that, HuRL\xspace helps despite the highly challenging setup of this experiment: a) environments with a high-dimensional observation space; a) the chosen hyperparameter values being likely suboptimal for individual environments; c) the heuristics naively generated using Monte-Carlo samples from a mixture of policies of wildly varying quality; and d) the lack of policy warm-starting.
We hypothesize that PPO-HuRL\xspace's performance can be improved further with environment-specific hyperparameter tuning and a scheme for heuristic-quality-dependent adjustment of HuRL\xspace's $\lambda$-schedules on the fly.
\begin{figure*}[ht!]
\centering
\makebox[\textwidth][c]{
\includegraphics[width=1.2\textwidth]{figures/final_results.png}
}
\caption{\small{PPO-HuRL\xspace's results on Procgen games. PPO-HuRL\xspace\ yields gains on half of the games and performs at par with PPO on most of the rest. Thus, on balance, PPO-HuRL\xspace\ helps despite the highly challenging setup of this experiment, but tuning HuRL\xspace's $\lambda$-schedule on the fly depending on the quality of the heuristic can potentially make HuRL\xspace's performance more robust in settings like this. }}
\label{fig:procgen}
\end{figure*}
\section{Preliminaries}
\vspace{-1mm}
\subsection{Notation}
\vspace{-1mm}
We focus on discounted infinite-horizon Markov Decision Processes (MDPs) for ease of exposition. The technique proposed here can be extended to other MDP settings.\footnote{The results here can be readily applied to finite-horizon MDPs; for other infinite-horizon MDPs, we need further, e.g., mixing assumptions for limits to exist.}
A discounted infinite-horizon MDP is denoted as a 5-tuple $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}=(\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}, \mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C}, P, r, \gamma)$,
where $\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}$ is the state space, $\mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C}$ is the action space, $P(s'|s,a)$ is the transition dynamics, $r(s,a)$ is the reward function, and $\gamma \in [0,1)$ is the discount factor.
Without loss of generality, we assume $r:\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\times\mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C}\to[0,1]$. We allow the state and action spaces $\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}$ and $\mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C}$ to be either discrete or continuous.
Let $\Delta(\cdot)$ denote the space of probability distributions. A decision-making policy $\pi$ is a conditional distribution $\pi:\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\to\Delta(\mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C})$, which can be deterministic.
We define some shorthand for writing expectations:
For a state distribution $d\in\Delta(\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U})$ and a function ${V : \mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\to\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb}$, we define $V(d) \coloneqq \Ebb_{s\sim d}[V(s)]$; similarly, for a policy $\pi$ and a function $Q :\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\times\mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C}\to\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$, we define ${Q(s, \pi) \coloneqq \Ebb_{a\sim\pi(\cdot|s)}[Q(s,a)]}$. Lastly, we define $\Ebb_{s'|s,a} \coloneqq \Ebb_{s'\sim P(\cdot|s,a)}$.
Central to solving MDPs are the concepts of value functions and average distributions.
For a policy $\pi$, we define its state value function $V^\pi$ as
$
V^\pi(s) \coloneqq \Ebb_{\rho_s^\pi} \left[ \sum_{t=0}^\infty \gamma^t r(s_t, a_t) \right],
$
where $\rho_s^\pi$ denotes the trajectory distribution of $s_0, a_0, s_1, \dots$
induced by running $\pi$ starting from $s_0 =s$.
We define the state-action value function (or the Q-function) as
$Q^\pi(s,a) \coloneqq r(s,a) + \gamma \Ebb_{s'|s,a} [V^\pi(s')]$.
We denote the optimal policy as $\pi^*$ and its state value function as $V^* \coloneqq V^{\pi^*}$. Under the assumption that rewards are in $[0,1]$, we have $V^\pi(s), Q^\pi(s,a)\in[0, \frac{1}{1-\gamma}]$ for all $\pi$, $s\in\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}$, and $a\in\mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C}$.
We denote the initial state distribution of interest as $d_0 \in \Delta(\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U})$ and the state distribution of policy $\pi$ at time $t$ as $d_t^\pi$, with $d_0^\pi = d_0$.
Given $d_0$, we define the average state distribution of a policy $\pi$ as
$
d^\pi \coloneqq (1-\gamma) \sum_{t=0}^\infty \gamma^t d_t^\pi
$.
With a slight abuse of notation, we also write $d^\pi(s,a) \coloneqq d^\pi(s)\pi(a|s)$.
\vspace{-1mm}
\subsection{Setup: Reinforcement Learning with Heuristics} \vspace{-1mm}
\label{sec:RL with heuristic setup}
We consider RL with prior knowledge expressed in the form of a \rev{heuristic value function}.
The goal is to find a policy $\pi$ that has high return through interactions with an unknown MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$, i.e.,
$\max_\pi V^\pi(d_0)$.
While the agent here does not fully know
$\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$, we suppose that, before interactions start the agent is provided with a heuristic $h:\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\to\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$ which the agent can query throughout learning.
The heuristic $h$ represents a prior guess of the optimal value function $V^*$ of $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$.
Common sources of heuristics are domain knowledge as typically employed in planning, and logged data collected by exploratory or by expert behavioral policies.
In the latter, a heuristic guess of $V^*$ can be computed from the data by offline RL algorithms. For instance, when we have trajectories of an expert behavioral policy,
Monte-Carlo regression estimate of the observed returns
may be a good guess of $V^*$.
Using heuristics to solve MDP problems has been popular in planning and control, but its usage is rather limited in RL. The closest provable technique in RL is PBRS~\citep{ng1999policy}, where the reward is modified into
$\overline{r}(s,a) \coloneqq r(s,a) + \gamma \Ebb_{s'|s,a}[h(s')] - h(s)$.
It can be shown that this transformation does not introduce bias into the policy ordering, and therefore solving the new MDP $\overline{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}} \coloneqq (\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U},\mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C},P,\overline{r},\gamma)$ would yield the same optimal policy $\pi^*$ of $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$.
Conceptually when the heuristic is the optimal value function $h = V^*$, the agent should be able to find the optimal policy $\pi^*$ of $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$ by acting myopically, as $V^*$ already contains all necessary long-term information for good decision making.
\rev{However, running an RL algorithm with the PBRS reward (i.e. solving $\overline{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}} \coloneqq (\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U},\mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C},P,\overline{r},\gamma)$) does not take advantage of this shortcut. To make learning efficient, we need to also let the base RL algorithm know that acting greedily (i.e., using a smaller discount) with the shaped reward can yield good policies.
An intuitive idea
is to run the RL algorithm to maximize $\overline{V}^\pi_\lambda(d_0)$, where $\overline{V}^\pi_\lambda$ denotes the value function of $\pi$ in an MDP $\overline{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}_\lambda \coloneqq (\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U},\mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C},P,\overline{r},\lambda\gamma)$ for some $\lambda\in[0,1]$.
However this does not always work. For example, when $\lambda=0$, $\max_\pi \overline{V}^\pi_\lambda(d_0)$ only optimizes for the initial states $d_0$, but obviously the agent is going to encounter other states in $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$.
We next propose a provably correct version, HuRL\xspace, to leverage this short-horizon insight.
}
\section*{Checklist}
\begin{enumerate}
\item For all authors...
\begin{enumerate}
\item Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
\answerYes{}
\item Did you describe the limitations of your work?
\answerYes{\Cref{sec:disclim}.}
\item Did you discuss any potential negative societal impacts of your work?
\answerYes{\Cref{sec:disclim}.} It is a conceptual work that doesn't have foreseeable societal impacts yet.
\item Have you read the ethics review guidelines and ensured that your paper conforms to them?
\answerYes{}
\end{enumerate}
\item If you are including theoretical results...
\begin{enumerate}
\item Did you state the full set of assumptions of all theoretical results?
\answerYes{The assumptions are in the theorem, proposition, and lemma statements throughout the paper.}
\item Did you include complete proofs of all theoretical results?
\answerYes{\Cref{app:proofs} and \Cref{app:tech}.}
\end{enumerate}
\item If you ran experiments...
\begin{enumerate}
\item Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? \answerYes{In the supplemental material.}
\item Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)?
\answerYes{\Cref{app:experiments}.}
\item Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)?
\answerYes{\Cref{sec:exps} and \Cref{app:experiments}.}
\item Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)?
\answerYes{\Cref{app:experiments}.}
\end{enumerate}
\item If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
\begin{enumerate}
\item If your work uses existing assets, did you cite the creators?
\answerYes{References \cite{garage}, \cite{todorov2012mujoco}, and \cite{procgen} in \Cref{ssec:setup}.}
\item Did you mention the license of the assets?
\answerYes{In \Cref{ssec:setup}.}
\item Did you include any new assets either in the supplemental material or as a URL?
\answerYes{Heuristic computation and scripts to run training.}
\item Did you discuss whether and how consent was obtained from people whose data you're using/curating?
\answerNA{}
\item Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content?
\answerNA{}
\end{enumerate}
\item If you used crowdsourcing or conducted research with human subjects...
\begin{enumerate}
\item Did you include the full text of instructions given to participants and screenshots, if applicable?
\answerNA{}
\item Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable?
\answerNA{}
\item Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation?
\answerNA{}
\end{enumerate}
\end{enumerate}
\section{Discussion and Limitations} \label{sec:disclim}
\vspace{-1mm}
This work is an early step towards theoretically understanding the role and potential of heuristics in guiding RL algorithms.
We propose a framework, HuRL\xspace, that can accelerate RL when an informative heuristic is provided.
HuRL\xspace induces a horizon-based regularization of RL, complementary to existing warm-starting schemes, and we provide theoretical and empirical analyses to support its effectiveness.
While this is a conceptual work without foreseeable societal impacts yet, we hope that it will help counter some of AI's risks by making RL more predictable via incorporating prior into learning.
We remark nonetheless that the effectiveness of HuRL\xspace depends on the available heuristic. While HuRL\xspace can eventually solve the original RL problem even with a non-ideal heuristic, using a bad heuristic can slow down learning.
Therefore, an important future research direction is to adaptively tune the mixing coefficient based on the heuristic quality with curriculum or meta-learning techniques.
In addition, while our theoretical analysis points out a strong connection between good heuristics for HuRL\xspace and pessimistic offline RL, techniques for the latter are not yet scalable and robust enough for high-dimensional problems.
Further research on offline RL can unlock the full potential of HuRL\xspace.
\section{Discussion and Limitations}
\section{Experiments} \label{sec:exps}
\vspace{-1mm}
\defSAC\xspace{SAC\xspace}
\defSAC w/ BC\xspace{SAC w/ BC\xspace}
\defHuRL\xspace-zero\xspace{HuRL\xspace-zero\xspace}
\defHuRL\xspace-MC\xspace{HuRL\xspace-MC\xspace}
\defBC\xspace{BC\xspace}
We validate our framework HuRL\xspace experimentally in MuJoCo (commercial license)~\citep{todorov2012mujoco} robotics control problems and Procgen games (MIT License)~\citep{procgen},
where soft actor critic (SAC)~\cite{haarnoja2018soft} and proximal policy optimization (PPO)~\cite{schulman2017proximal} were used as the base RL algorithms, respectively\footnote{Code to replicate all experiments is available at \href{https://github.com/microsoft/HuRL}{https://github.com/microsoft/HuRL}.}.
The goal is to study whether HuRL\xspace can accelerate learning by shortening the horizon with heuristics.
In particular, we conduct studies to investigate the effects of different heuristics and mixing coefficients.
Since the main focus here is on the possibility of leveraging a \emph{given} heuristic to accelerate RL algorithms, in these experiments we used vanilla techniques
to construct heuristics for HuRL\xspace.
Experimentally studying the design of heuristics for a domain or a batch of data is beyond the scope of the current paper but are important future research directions.
For space limitation, here we report only the results of the MuJoCo experiments. The results on Procgen games along with other experimental details can also be found in \cref{app:experiments}.
\vspace{-1mm}
\subsection{Setup} \label{ssec:setup}
\vspace{-1mm}
We consider four MuJoCo environments with dense rewards (Hopper-v2, HalfCheetah-v2, Humanoid-v2, and Swimmer-v2) \rev{and a sparse reward version of Reacher-v2 (denoted as Sparse-Reacher-v2)\footnote{The reward is zero at the goal and $-1$ otherwise.}.}
\rev{We design the experiments to simulate two learning scenarios.
First, we use Sparse-Reacher-v2 to simulate the setting where an engineered heuristic based on domain knowledge is available; since this is a goal reaching task, we designed a heuristic $h(s) = r(s,a) - 100 \norm{e(s)-g(s)} $, where $e(s)$ and $g(s)$ denote the robot's end-effector position and the goal position, respectively.
Second, we use the dense reward environments to model scenarios} where a batch of data collected by multiple behavioral policies is available before learning, and a heuristic is constructed by an offline policy evaluation algorithm from the batch data \rev{(see Appendix~\ref{app:details} for details). In brief}, we generated these behavioral policies by running SAC\xspace from scratch and saved the intermediate policies generated in training.
We then use least-squares regression to fit a
neural network to predict empirical Monte-Carlo returns of the trajectories in the sampled batch of data.
We also use behavior cloning (BC\xspace) to warm-start \rev{all} RL agents based on the same batch dataset \rev{in the dense reward experiments.}
The base RL algorithm here, SAC, is based on the standard implementation in Garage (MIT License)~\cite{garage}.
The policy and value networks are fully connected independent neural networks. The policy is Tanh-Gaussian and the value network has a linear head.
\vspace{-2mm}
\paragraph{Algorithms.}
We compare the performance of different algorithms below.
\begin{enumerate*}[label=\textit{\arabic*)}]
\item BC\xspace
\item SAC
\item SAC with BC\xspace warm start (SAC w/ BC\xspace)
\item \rev{HuRL\xspace with the engineered heuristic (HuRL\xspace)}
\item HuRL\xspace with a zero heuristic and BC\xspace warm start (HuRL\xspace-zero\xspace)
\item HuRL\xspace with the Monte-Carlo heuristic and BC\xspace warm start (HuRL\xspace-MC\xspace)
\item \rev{SAC with PBRS reward (and BC\xspace warm start, if applicable) (PBRS)}.
\end{enumerate*}
For the HuRL\xspace algorithms, the mixing coefficient was scheduled as $\lambda_n = \lambda_0 + (1-\lambda_0) c_\omega
\tanh(\omega (n-1) )$, for $n=1,\dots,N$, where $\lambda_0 \in [0,1]$, $\omega>0$ controls the increasing rate, and $c_\omega$ is a normalization constant such that $\lambda_\infty=1$ and $\lambda_n \in[0,1]$.
We chose these algorithms to study the effect of each additional warm-start component (BC\xspace and heuristics) added on top of vanilla SAC\xspace.
HuRL\xspace-zero\xspace is SAC w/ BC\xspace but with an extra $\lambda$ schedule above that further lowers the discount, whereas SAC\xspace and SAC w/ BC\xspace keep a constant discount factor.
\vspace{-2mm}
\paragraph{Evaluation and Hyperparameters.}
In each iteration, the RL agent has a fixed sample budget for environment interactions, and its performance is measured in terms of undiscounted cumulative returns of the deterministic mean policy extracted from SAC\xspace.
The hyperparameters used in the algorithms above were selected as follows. First, the learning rates and the discount factor of the base RL algorithm, SAC\xspace, were tuned for each environment. The tuned discount factor was used as the discount factor $\gamma$ of the original MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$.
Fixing the hyperparameters above, we additionally tune $\lambda_0$ and $\omega$ for the $\lambda$ schedule of HuRL\xspace for each environment and each heuristic.
Finally, after all these hyperparameters were fixed, we conducted additional testing runs with 30 different random seeds and report their statistics here.
Sources of randomness included the data collection process of the behavioral policies, training the heuristics from batch data, BC, and online RL. However, the behavioral policies were fixed across all testing runs.
\rev{
We chose this hyperparameter tuning procedure to make sure that the baselines (i.e. SAC) compared in these experiments are their best versions.
}
\begin{figure*}[t]
\centering
\begin{subfigure}{0.29\textwidth}
\includegraphics[width=\textwidth]{figures/sparse_reacher_plottest-sparsereacher-001.png}
\caption{Sparse-Reacher-v2}
\label{fig:sparse reacher}
\end{subfigure}
\begin{subfigure}{0.29\textwidth}
\includegraphics[width=\textwidth]{figures/paper_plot_test-humanoid.png}
\caption{Humanoid-v2}
\label{fig:main humanoid}
\end{subfigure}
\begin{subfigure}{0.29\textwidth}
\includegraphics[width=\textwidth]{figures/paper_plot_test-hopper.png}
\caption{Hopper-v2}
\label{fig:main hopper}
\end{subfigure}
\begin{subfigure}{0.29\textwidth}
\includegraphics[width=\textwidth]{figures/paper_plot_test-swimmer.png}
\caption{Swimmer-v2}
\label{fig:main swimmer}
\end{subfigure}
\begin{subfigure}{0.29\textwidth}
\includegraphics[width=\textwidth]{figures/paper_plot_test-halfcheetah.png}
\caption{HalfCheetah-v2}
\label{fig:main halfcheetah}
\end{subfigure}
\begin{subfigure}{0.29\textwidth}
\includegraphics[width=\textwidth]{figures/hopper_lambda.png}
\caption{ $\lambda_0$ ablation.}
\label{fig:lambda ablation (hopper)}
\end{subfigure}
\caption{\small{Experimental results.
(a) uses an engineered heuristic for a sparse reward problem;
(b)-(e) use heuristics learned from offline data and share the same legend; (e) shows ablation results of different initial $\lambda_0$ in Hopper-v2.
The plots show the $25$th, $50$th, $75$th percentiles of algorithm performance over 30 random seeds.}}
\label{fig:main results}
\vspace{-2mm}
\end{figure*}
\vspace{-1mm}
\subsection{Results Summary}
\vspace{-1mm}
\cref{fig:main results} shows the results on the MuJoCo environments. \rev{Overall, we see that HuRL\xspace is able to leverage engineered and learned heuristics to significantly improve the learning efficiency. This trend is consistent across all environments that we tested on.}
\rev{For the sparse-reward experiments, we see that SAC and PBRS struggle to learn, while HuRL\xspace~is able to converge to the optimal performance much faster.
For the dense reward experiments, similarly HuRL\xspace-MC\xspace converges much faster, though the gain in HalfCheetah-v2 is minor and it might have converged to a worse local maximum in Swimmer-v2.
}
In addition, we see that warm-starting SAC\xspace using BC (i.e. SAC w/ BC\xspace) can improve the learning efficiency compared with the vanilla SAC\xspace, but using BC\xspace alone does not result in a good policy. Lastly, we see that using the zero heuristic (HuRL\xspace-zero\xspace) with extra $\lambda$-scheduling does not further improve the performance of SAC w/ BC\xspace. This comparison verifies that the learned Monte-Carlo heuristic provides non-trivial information.
\rev{
Interestingly, we see that applying PBRS to SAC leads to even worse performance than running SAC with the original reward. There are two reasons why SAC+PBRS is less desirable than SAC+HuRL as we discussed before:
\begin{enumerate*}[label=\textit{\arabic*)}]
\item
PBRS changes the reward/value scales in the induced MDP, and popular RL algorithms like SAC are very sensitive to such changes. In contrast, HuRL\xspace~induces values on the same scale as we show in \cref{th:preserved MDP properties}.
\item In HuRL\xspace, we are effectively providing the algorithm some more side-information to let SAC shorten the horizon when the heuristic is good.
\end{enumerate*}
}
The results in \cref{fig:main results} also have another notable aspect.
Because the datasets used in the dense reward experiments contain trajectories collected by a range of policies, it is likely that BC\xspace suffers from disagreement in action selection among different policies.
Nonetheless, training a heuristic using a basic Monte-Carlo regression seems to be less sensitive to these conflicts and still results in a helpful heuristic for HuRL\xspace.
One explanation can be that heuristics are only functions of states, not of states and actions, and therefore the conflicts are minor.
Another plausible explanation is that HuRL\xspace only uses the heuristic to \emph{guide} learning, and does not completely rely on it to make decisions
Thus, HuRL\xspace can be more robust to the heuristic quality, or, equivalently, to the quality of prior knowledge.
\vspace{-1mm}
\subsection{Ablation: Effects of Horizon Shortening}
\vspace{-1mm}
To further verify that the acceleration in \cref{fig:main results} is indeed due to horizon shortening, we conducted an ablation study for HuRL\xspace-MC\xspace on Hopper-v2, whose results are presented in \cref{fig:lambda ablation (hopper)}.
HuRL\xspace-MC\xspace's best $\lambda$-schedule hyperparameters on Hopper-v2, which are reflected in its performance in the aforementioned \cref{fig:main hopper}, induced a near-constant schedule at $\lambda=0.95$; to obtain the curves in \cref{fig:lambda ablation (hopper)}, we ran HuRL\xspace-MC\xspace with constant-$\lambda$ schedules for several more $\lambda$ values.
\cref{fig:lambda ablation (hopper)} shows that increasing $\lambda$ above $0.98$ leads to a performance drop. Since using a large $\lambda$ decreases bias and makes the reshaped MDP more similar to the original MDP, we conclude that the increased learning speed on Hopper-v2 is due to HuRL\xspace's horizon shortening (coupled with the guidance provided by its heuristic).
\section{Introduction}
\label{sec:intro}
Many recent empirical successes of reinforcement learning (RL) require solving problems with very long decision-making horizons.
OpenAI Five~\citep{berner2019dota} used episodes that were $20000$ timesteps on average, while AlphaStar~\citep{vinyals2019grandmaster} used roughly $5000$ timesteps.
Long-term credit assignment is a very challenging statistical problem, with the sample complexity
growing quadratically (or worse) with the horizon~\citep{dann2015sample}. Long horizons (or, equivalently, large discount factors) also increase RL's computational burden, leading to slow optimization convergence~\citep{sidford2018near}.
This makes RL algorithms require prohibitively large amounts of interactions and compute: even with tuned hyperparameters, AlphaStar needed over $10^8$ samples and OpenAI Five needed over $10^7$ PFLOPS of compute.
A popular approach to mitigate the statistical and computational issues of tabula rasa RL methods is to warm-start or regularize learning with prior knowledge~\citep{tessler2020maximizing,NEURIPS2019_eba237ec,bhardwaj2020blending,farahmand2016truncated,vinyals2019grandmaster,berner2019dota,nair2020accelerating,hester2018deep}.
For instance, AlphaStar learned a policy and value function from human demonstrations and regularized the RL agent using imitation learning (IL).
AWAC~\citep{nair2020accelerating} warm-started a policy using batch policy optimization on exploratory datasets.
While these approaches have been effective in different domains, none of them explicitly address RL's complexity dependence on horizon.
In this paper, we propose a complementary regularization
technique that relies on heuristic value functions, or \emph{heuristics}\footnote{\rev{We borrow this terminology from the planning literature to refer to guesses of $V^*$ in an MDP~\citep{kolobov2012}.
}} for short, to effectively shorten the problem horizon faced by an online RL agent \rev{for fast learning}.
We call this approach Heuristic-Guided Reinforcement Learning\xspace (HuRL\xspace). The core idea is simple: given a Markov decision process (MDP) $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}=(\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}, \mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C}, P, r, \gamma)$ and a heuristic $h:\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\to\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$, we select a mixing coefficient $\lambda\in[0,1]$ and have the agent solve a new MDP $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}=(\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}, \mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C}, P, \widetilde{r}, \widetilde{\gamma})$ with a reshaped reward and a smaller discount \rev{(i.e. a shorter horizon)}:
\begin{align} \label{eq:short summary}
\widetilde{r}(s,a) \coloneqq r(s,a) + (1-\lambda)\gamma \Ebb_{s'\sim P(\cdot| s,a)}[h(s')]
\quad
\text{and}
\quad
\widetilde{\gamma} \coloneqq \lambda\gamma.
\end{align}
HuRL\xspace effectively introduces horizon-based regularization that determines whether long-term value information should come from collected experiences or the heuristic.
By modulating the effective horizon via $\lambda$, we trade off the bias and the complexity of solving the reshaped MDP. HuRL\xspace with $\lambda=1$ recovers the original problem and with $\lambda=0$ creates an easier contextual bandit problem~\citep{pmlr-v119-foster20a}.
A heuristic $h$ in HuRL\xspace represents a prior guess of the desired long-term return of states, which ideally is the optimal value function $V^*$ of the unknown MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$.
\rev{When the heuristic $h$ captures the state ordering of $V^*$ well, conceptually, it becomes possible to make good long-term decisions by short-horizon planning or even acting greedily.}
How do we construct a good heuristic? In the planning literature, this is typically achieved by solving a relaxation of the original problem \cite{Hoffmann_2001,Richter_2010,akolobov-aaai10}.
Alternatively, one can learn it from batch data collected by exploratory behavioral policies (as in offline RL~\citep{gulcehre2021regularized}) or from expert policies (as in IL~\citep{cheng2020policy}).\footnote{We consider the RL setting for imitation where we suppose the rewards of expert trajectories are available.}
\rev{For some dense reward problems,
a zero heuristic can be effective in reducing RL complexity, as exploited by the guidance discount framework~\citep{blackwell1962discrete,petrik2008biasing,jiang2015dependence,jiang2016structural,chen2018improving,amit2020discount}.}
In this paper, we view heuristics as a unified representation of various forms of prior knowledge, such as expert demonstrations, exploratory datasets, and engineered guidance.
Although the use of heuristics to accelerate search has been popular in planning and control algorithms, e.g., A*~\citep{hart1968formal}, MCTS~\citep{browne2012survey}, and MPC~\citep{zhong2013value,bhardwaj2020blending,hoeller2020deep,bejjani2018planning}, its theory is less developed for settings where the MDP is \emph{unknown}.
\rev{The closest work in RL is
potential-based reward shaping (PBRS)~\citep{ng1999policy}, which reshapes the reward into $\bar{r}(s,a) = r(s,a)+\gamma \Ebb_{s'|s,a}[h(s')]-h(s)$ while keeping the original discount.
PBRS can use any heuristic to reshape the reward while preserving the ordering of policies.
However, giving PBRS rewards to an RL algorithm does not necessarily lead to faster learning,
because the base RL algorithm
would still seek to explore to resolve long-term credit assignment.
HuRL\xspace~allows common RL algorithms to leverage the short-horizon potential provided by a heuristic to learn faster.
}
In this work, we provide a theoretical foundation of HuRL\xspace to enable adopting heuristics and horizon reduction for accelerating RL, \rev{combining advances from the PBRS and the guidance discount literatures.}
On the theoretical side, we derive a bias-variance decomposition of HuRL\xspace's horizon-based regularization in order to characterize the solution quality as a function of $\lambda$ and $h$.
Using this insight, we provide sufficient conditions for achieving an effective trade-off, including properties required of a base RL algorithm that solves the reshaped MDP $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}_\lambda$.
Furthermore, we define the novel concept of an \emph{improvable} heuristic and prove that good heuristics for HuRL\xspace can be constructed from data using existing \emph{pessimistic} offline RL algorithms (such as pessimistic value iteration~\citep{jin2020pessimism,liu2020provably}).
The effectiveness of HuRL\xspace depends on the heuristic quality, so we design HuRL\xspace to employ a sequence of mixing coefficients (i.e. $\lambda$s) that increases as the agent gathers more data from the environment.
Such a strategy induces a learning curriculum that enables HuRL\xspace to remain robust to non-ideal heuristics.
HuRL\xspace starts off by guiding the agent's search direction with a heuristic.
As the agent becomes more experienced, it gradually removes the guidance and lets the agent directly optimize the true long-term return.
We empirically validate HuRL\xspace in MuJoCo~\citep{todorov2012mujoco} robotics control problems and Procgen games~\citep{procgen}
with various heuristics and base RL algorithms.
The experimental results demonstrate the versatility and effectiveness of HuRL\xspace in accelerating RL algorithms.
\section{Heuristic-Guided Reinforcement Learning}
\vspace{-1mm}
We propose a general framework, HuRL\xspace, for leveraging heuristics to accelerate RL.
In contrast to tabula rasa RL algorithms that attempt to directly solve the long-horizon MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$, HuRL\xspace uses a heuristic to guide the agent in solving a sequence of short-horizon MDPs so as to amortize the complexity of long-term credit assignment.
In effect, HuRL\xspace creates a heuristic-based learning curriculum to help the agent learn faster.
\vspace{-1mm}
\subsection{Algorithm}
\vspace{-1mm}
HuRL\xspace takes a reduction-based approach to realize the idea of heuristic guidance.
As summarized in \cref{alg:ouralg}, HuRL\xspace takes a heuristic $h:\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\to\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$ and a base RL algorithm $\LL$ as input, and outputs an approximately optimal policy for the original MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$.
During training, HuRL\xspace iteratively runs the base algorithm $\LL$ to collect data from the MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$ and then uses the heuristic $h$ to modify the agent's collected experiences.
Namely, in iteration $n$, the agent interacts with the original MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$ and saves the raw transition tuples\footnote{If $\LL$ learns only with trajectories, we transform each tuple and assemble them to get the modified trajectory.} $\mathcal{D}}\def\EE{\mathcal{E}}\def\FF{\mathcal{F}_n = \{(s,a,r,s')\}$ (line \ref{line:data collection}). HuRL\xspace\ then defines a reshaped MDP $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}_n \coloneqq (\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U},\mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C},P,\widetilde{r}_n, \widetilde{\gamma}_n)$ (line \ref{line:reshape mdp}) by changing the rewards and lowering the discount factor:
\begin{align} \label{eq:reshaped MDP}
\widetilde{r}_n(s,a) \coloneqq r(s,a) + (1-\lambda_n)\gamma \Ebb_{s'|s,a}[h(s')] \qquad \text{and}\qquad
\widetilde{\gamma}_n \coloneqq \lambda_n \gamma,
\end{align}
where $\lambda_n\in[0,1]$ is the mixing coefficient.
The new discount $\widetilde{\gamma}_n$ effectively gives $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}_n$ a shorter horizon than $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$'s,
while the heuristic $h$ is blended into the new reward in \eqref{eq:reshaped MDP}
to account for the missing long-term information.
We call $\widetilde{\gamma}_n =\lambda_n \gamma $ in \eqref{eq:reshaped MDP} the \emph{guidance discount} to be consistent with prior literature~\citep{jiang2015dependence}, which can be viewed in terms of our framework as using a zero heuristic.
In the last step (line \ref{line:train}), HuRL\xspace calls the base algorithm $\LL$ to perform updates with respect to the reshaped MDP $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}_n$.
This is realized by
\begin{enumerate*}[label=\textit{\arabic*)}]
\item setting the discount factor used in $\LL$ to $\widetilde{\gamma}_n $, and
\item setting the sampled reward to $r + (\gamma - \widetilde{\gamma}_n) h(s')$ for every transition tuple $(s,a,r,s')$ collected from $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$.
\end{enumerate*}
We remark that the base algorithm $\LL$ in line \ref{line:data collection} always collects trajectories of lengths proportional to the original discount $\gamma$, while internally the optimization is done with a lower discount $\widetilde{\gamma}_n$ in line \ref{line:train}.
Over the course of training, HuRL\xspace repeats the above steps with a sequence of increasing mixing coefficients $\{\lambda_n\}$.
From \eqref{eq:reshaped MDP} we see that as the agent interacts with the environment, the effects of the heuristic in MDP reshaping decrease and the effective horizon of the reshaped MDP increases.
\begin{algorithm}[t]
\caption{Heuristic-Guided Reinforcement Learning\xspace (HuRL\xspace)} \label{alg:ouralg}
\begin{algorithmic}[1]
\REQUIRE MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}=(\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U},\mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C},P,r,\gamma)$, RL algorithm $\LL$, heuristic $h$, mixing coefficients $\{\lambda_n\}$.
\FOR{$n=1,\dots,N$}
\STATE $\mathcal{D}}\def\EE{\mathcal{E}}\def\FF{\mathcal{F}_n \gets \LL$.CollectData($\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$) \label{line:data collection}
\STATE Get $\lambda_n$ from $\{\lambda_n\}$ and construct $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}_n = (\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U},\mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C},P,\widetilde{r}_n, \widetilde{\gamma}_n)$ according to \eqref{eq:reshaped MDP} using $h$ and $\lambda_n$ \label{line:reshape mdp}
\STATE $\pi_n \gets \LL$.Train($\mathcal{D}}\def\EE{\mathcal{E}}\def\FF{\mathcal{F}_n$, $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}_n$) \label{line:train}
\ENDFOR
\RETURN $\pi_N$
\end{algorithmic}
\end{algorithm}
\vspace{-1mm}
\subsection{HuRL\xspace as Horizon-based Regularization}
\vspace{-1mm}
We can think of HuRL\xspace as introducing a horizon-based \emph{regularization} for RL, where the regularization center is defined by the heuristic and its strength diminishes as the mixing coefficient increases.
As the agent collects more experiences, HuRL\xspace gradually removes the effects of regularization and the agent eventually optimizes for the original MDP.
HuRL\xspace's regularization is designed to reduce learning variance, similar to the role of regularization in supervised learning.
Unlike the typical weight decay imposed on function approximators (such as the agent's policy or value networks), our proposed regularization leverages the structure of MDPs to regulate the complexity of the MDP the agent faces, which scales with the MDP's discount factor (or, equivalently, the horizon). When the guidance discount $\widetilde{\gamma}_n$ is lower than the original discount $\gamma$ (i.e. $\lambda_n<1$), the reshaped MDP $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}_n$ given by \eqref{eq:reshaped MDP} has a shorter horizon and requires fewer samples to solve.
However, the reduced complexity comes at the cost of bias, because the agent is now incentivized toward maximizing the performance with respect to the heuristic rather than the original long-term returns of $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$.
In the extreme case of $\lambda_n=0$,
HuRL\xspace would solve a zero-horizon contextual bandit problem with contexts (i.e. states) sampled from $d^\pi$ of $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$.
\begin{figure*}[t]
\centering
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[height=0.65\textwidth]{figures/toy_value_heatmap.png}
\caption{Heatmap of different values.}
\label{fig:toy_value_heatmap}
\end{subfigure}
\hspace{2mm}
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[height=0.65\textwidth]{figures/toy_policy_heatmap.png}
\caption{Different policy behaviors.}
\label{fig:toy_policy_heatmap}
\end{subfigure}
\hspace{2mm}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[height=0.65\textwidth]{figures/Fig3.png}
\caption{HuRL\xspace with different $h$ and $\lambda$.}
\label{fig:toy_plot}
\end{subfigure}
\caption{\small{\textbf{Example of HuRL\xspace in a chain MDP.} Each cell in a row in each diagram represents a state from $\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}=\{1,\dots,10\}$. The agent starts at state $3$ ($s_0$), and states $1$ and $10$ are absorbing ($Abs$ in subfigure a-(1)). Actions $\mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C}=\{\leftarrow, \rightarrow\}$ move the agent left or right in the chain unless the agent is in an absorbing state. \textbf{Subfig. a-(1)} shows the reward function: $r(2,\leftarrow) =0.1, r(4,\rightarrow) =-0.2,r(5,\rightarrow)=0.1$, and all state-action pairs not shown in a-(1) yield $r=0$. \textbf{Subfig. a-(2)} shows $V^*$ for $\gamma = 0.9$. \textbf{Subfig. a-(3)} shows a good heuristic $h$ --- $V(\text{random }\pi)$. \textbf{Subfig. a-(4)} shows a bad heuristic $h$ --- $V(\text{myopic }\pi)$. \textbf{Subfig. b-(1)}: $\pi^*$ for $V^*$ from a-(2). \textbf{Subfig. b-(2)}: $\tilde{\pi}^*$ from HuRL\xspace with $h=0,\lambda=0.5$. \textbf{Subfig. b-(3)}: $\tilde{\pi}^*$ from HuRL\xspace with the good $h$ from (a).(3) and $\lambda=0.5$. \textbf{Subfig. b-(4)}: $\tilde{\pi}^*$ from the bad $h$ from a-(4), $\lambda=0.5$. \textbf{Subfig. b-(5)}: $\tilde{\pi}^*$ from the bad $h$ and $\lambda=1$. \textbf{Subfig. (c) illustrates the takeaway message}: \emph{using HuRL\xspace with a good $h$ can find $\pi^*$ from $s_0$ even with a small $\lambda$ (see the $x$-axis), while HuRL\xspace with a bad $h$ requires a much higher $\lambda$ to discover $\pi^*$.}
}}
\label{fig:toy_example}
\end{figure*}
\vspace{-1mm}
\subsection{A Toy Example}
\vspace{-1mm}
We illustrate this idea in a chain MDP environment in \cref{fig:toy_example}. The optimal policy $\pi^*$ for this MDP's original $\gamma = 0.9$ always picks action $\rightarrow$, as shown in \cref{fig:toy_policy_heatmap}-(1), giving the optimal value $V^*$ in \cref{fig:toy_value_heatmap}-(2).
Suppose we used a smaller guidance discount $\widetilde{\gamma} = 0.5\gamma$ to accelerate learning. This is equivalent to HuRL\xspace with a zero heuristic $h=0$ and $\lambda=0.5$. Solving this reshaped MDP yields a policy $\widetilde{\pi}^*$ that acts very myopically in the original MDP, as shown in \cref{fig:toy_policy_heatmap}-(2); the value function of $\widetilde{\pi}^*$ in the original MDP is visualized in \cref{fig:toy_value_heatmap}-(4).
Now, suppose we use \cref{fig:toy_value_heatmap}-(4) as a heuristic in HuRL\xspace instead of $h=0$. This is a bad choice of heuristic (Bad $h$)
as it introduces a large bias with respect to $V^*$ (cf. \cref{fig:toy_value_heatmap}-(2)). On the other hand, we can roll out a random policy in the original MDP and use its value function as the heuristic (Good $h$), shown in \cref{fig:toy_value_heatmap}-(3).
Though the random policy has an even \emph{lower} return at the initial state $s=3$, it gives a \emph{better} heuristic because this heuristic shares the same trend as $V^*$ in \cref{fig:toy_value_heatmap}-(1).
HuRL\xspace run with Good $h$ and Bad $h$ yields policies in \cref{fig:toy_policy_heatmap}-(3,4), and the quality of the resulting solutions in the original MDP, $V_{\lambda}^{\widetilde{\pi}^*}(d_0)$, is reported in \cref{fig:toy_plot} for different $\lambda$. Observe that HuRL\xspace with a good heuristic can achieve $V^*(d_0)$ with a much smaller horizon $\lambda \le 0.5$. Using a bad $h$ does not lead to $\pi^*$ at all when $\lambda = 0.5$ (\cref{fig:toy_policy_heatmap}-(4)) but is guaranteed to do so when $\lambda$ converges to $1$. (\cref{fig:toy_policy_heatmap}-(5)).
\vspace{-1mm}
\section{Theoretical Analysis}
\vspace{-1mm}
When can HuRL\xspace accelerate learning?
Similar to typical regularization techniques, the horizon-based regularization of HuRL\xspace leads to a bias-variance decomposition that can be optimized for better finite-sample performance compared to directly solving the original MDP.
However, a non-trivial trade-off is possible only when the regularization can bias the learning toward a good direction.
In HuRL\xspace's case this is determined by the heuristic, which resembles a prior we encode into learning.
In this section we provide HuRL\xspace's theoretical foundation.
We first describe the bias-variance trade-off induced by HuRL\xspace. Then we show how suboptimality in solving the reshaped MDP translates into performance in the original MDP, and identify the assumptions HuRL\xspace needs the base RL algorithm to satisfy.
In addition, we
\rev{explain how HuRL\xspace relates to PBRS}, and characterize the quality of heuristics and sufficient conditions for constructing good heuristics from batch data using offline RL.
For clarity, we will focus on the reshaped MDP $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}=(\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}, \mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C}, P, \widetilde{r}, \widetilde{\gamma})$ for a fixed $\lambda\in[0,1]$, where $\widetilde{r}, \widetilde{\gamma}$ are defined in \eqref{eq:short summary}.
We can view this MDP as the one in a single iteration of HuRL\xspace.
For a policy $\pi$, we denote its state value function in $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}$ as $\widetilde{V}^\pi$, and the optimal policy and value function of $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}$ as $\widetilde{\pi}^*$ and $\widetilde{V}^*$, respectively.
The missing proofs of the results from this section can be found in \cref{app:proofs}.
\vspace{-1mm}
\subsection{Short-Horizon Reduction: Performance Decomposition}
\vspace{-1mm}
Our main result is a performance decomposition, which characterizes how a heuristic $h$ and suboptimality in solving the reshaped MDP $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}$ relate to performance in the original MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$.
\begin{restatable}{theorem}{PerformanceDecomposition}
\label{th:performance decomposition}
For any policy $\pi$, heuristic $f:\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\to\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$, and mixing coefficient $\lambda\in[0,1]$,
\begin{align*}
V^*(d_0) - V^\pi(d_0) &= \mathrm{Regret}(h,\lambda,\pi) + \mathrm{Bias}(h,\lambda,\pi)
\end{align*}
where we define
\begin{align}
\mathrm{Regret}(h,\lambda,\pi) &\coloneqq
\lambda \left( \widetilde{V}^*(d_0) - \widetilde{V}^\pi(d_0) \right) + \frac{1-\lambda}{1-\gamma} \left( \widetilde{V}^*(d^{\pi}) - \widetilde{V}^\pi(d^{\pi}) \right)
\label{eq:regret} \\
%
\mathrm{Bias}(h,\lambda,\pi) &\coloneqq
\left( V^*(d_0) - \widetilde{V}^*(d_0) \right) + \frac{\gamma(1-\lambda)}{1-\gamma} \Ebb_{s,a\sim d^\pi} \Ebb_{s'|s,a} \left[ h(s') - \widetilde{V}^*(s')\right] \label{eq:bias}
\end{align}
Furthermore, $\forall b\in\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$, $\mathrm{Bias}(h,\lambda,\pi) = \mathrm{Bias}(h+b,\lambda,\pi)$
and
$\mathrm{Regret}(h,\lambda,\pi) = \mathrm{Regret}(h+b,\lambda,\pi)$.
\end{restatable}
The theorem shows that suboptimality of a policy $\pi$ in the original MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$ can be decomposed into
\begin{enumerate*} [label=\textit{\arabic*)}]
\item a \emph{bias} term due to solving a reshaped MDP $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}$ instead of the original MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$, and
\item a \emph{regret} term (i.e. the learning variance) due to $\pi$ being suboptimal in the reshaped MDP $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}$.
\end{enumerate*}
Moreover, it shows that heuristics are equivalent up to constant offsets. In other words, only the relative ordering between states that a heuristic induces matters in learning, not the absolute values.
Balancing the two terms trades off bias and variance in learning.
Using a smaller $\lambda$
replaces the long-term information with the heuristic and make the horizon of the reshaped MDP $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}$ shorter.
Therefore, given a finite interaction budget, the regret term in \eqref{eq:regret} can be more easily minimized, though the bias term in \eqref{eq:bias} can potentially be large if the heuristic is bad.
On the contrary, with $\lambda=1$, the bias is completely removed, as the agent solves the original MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$ directly.
\vspace{-1mm}
\subsection{Regret, Algorithm Requirement, and Relationship with PBRS}
\vspace{-1mm}
The regret term in \eqref{eq:regret} characterizes the performance gap due to $\pi$ being suboptimal in the reshaped MDP $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}$, because $\mathrm{Regret}(h,\lambda,\widetilde{\pi}^*) = 0$ for any $h$ and $\lambda$.
For learning, we need the base RL algorithm $\LL$ to find a policy $\pi$ such that the regret term in \eqref{eq:regret} is small.
By the definition in \eqref{eq:regret}, the base RL algorithm $\LL$ is required not only to find a policy $\pi$ such that $\widetilde{V}^*(s) - \widetilde{V}^\pi(s)$ is small for states from $d_0$,
\emph{but also for states $\pi$ visits when rolling out in the original MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$}.
In other words, it is insufficient for the base RL algorithm to only optimize for $\widetilde{V}^\pi(d_0)$ (the performance in the reshaped MDP with respect to the initial state distribution; \rev{see \cref{sec:RL with heuristic setup}}).
For example, suppose $\lambda=0$ and $d_0$ concentrates on a single state $s_0$. Then maximizing $\widetilde{V}^\pi(d_0)$ alone would only optimize $\pi(\cdot|s_0)$ and the policy $\pi$ need not know how to act in other parts of the state space.
To use HuRL\xspace, we need the base algorithm to learn a policy $\pi$ that has small \emph{action gaps} in the reshaped MDP $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}$ \emph{but along trajectories in the original MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$}, as we show below. This property is satisfied by off-policy RL algorithms such as Q-learning~\citep{jin2018q}.
\begin{restatable}{proposition}{RegretAsActionGap} \label{th:regret as action gap}
For \emph{any} policy $\pi$, heuristic $f:\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\to\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$ and mixing coefficient $\lambda\in[0,1]$,
\begin{align*}
\textstyle
\mathrm{Regret}(h,\lambda,\pi)
= - \Ebb_{\rho^\pi(d_0)} \left[ \sum_{t=0}^\infty \gamma^t \widetilde{A}^*(s_t,a_t) \right]
\end{align*}
where $\rho^\pi(d_0)$ denotes the trajectory distribution of running $\pi$ from $d_0$, and
$\widetilde{A}^*(s,a) = \widetilde{r}(s,a)+\widetilde{\gamma}\Ebb_{s'|s,a}[\widetilde{V}^*(s')] - \widetilde{V}^*(s) \leq 0 $ is the action gap
with respect to the optimal policy $\widetilde{\pi}^*$ of $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}$.
\end{restatable}
Another way to comprehend the regret term is through studying its dependency on $\lambda$.
When $\lambda=1$, $\mathrm{Regret}(h,0,\pi) = V^*(d_0) - V^\pi(d_0)$, which is identical to the policy regret in $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$ for a \emph{fixed} initial distribution $d_0$.
On the other hand, when $\lambda=0$,
$
\mathrm{Regret}(h,0,\pi)=
\max_{\pi'} \frac{1}{1-\gamma} \Ebb_{s\sim d^{\pi}}[ \widetilde{r}(s,\pi') - \widetilde{r}(s,\pi)]
$, which is the regret of a \emph{non-stationary} contextual bandit problem where the context distribution is $d^\pi$ (the average state distribution of $\pi$).
In general, for $\lambda\in(0,1)$, the regret notion mixes a short-horizon non-stationary problem and a long-horizon stationary problem.
One natural question is whether the reshaped MDP $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}$ has a more complicated and larger value landscape than the original MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$, because these characteristics may affect the regret rate of a base algorithm.
We show that $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}$ preserves the value bounds and linearity of the original MDP.
\begin{restatable}{proposition}{PreservedProperties} \label{th:preserved MDP properties}
Reshaping the MDP as in \eqref{eq:short summary} preserves the following characteristics:
\begin{enumerate*}[label=\textit{\arabic*)}]
\item If $h(s) \in [0,\frac{1}{1-\gamma}]$, then $\widetilde{V}^\pi(s) \in [0,\frac{1}{1-\gamma}]$ for all $\pi$ and $s\in\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}$.
\item If $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}$ is a linear MDP with feature vector $\phi(s,a)$ (i.e. $r(s,a)$ and $\Ebb_{s'|s,a}[g(s')]$ for any $g$ can be linearly parametrized in $\phi(s,a)$), then $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}$ is also a linear MDP with feature vector $\phi(s,a)$.
\end{enumerate*}
\end{restatable}
\vspace{-1mm}
\rev{
On the contrary, the MDP $\overline{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}_\lambda \coloneqq (\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U},\mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C},P,\overline{r},\lambda\gamma)$ in \cref{sec:RL with heuristic setup} does not have these properties. We can show that $\overline{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}_\lambda$ is equivalent to $\widetilde{\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}}$ up to a PBRS transformation (i.e., $\bar{r}(s,a) = \tilde{r}(s,a)+ \tilde{\gamma} \Ebb_{s'|s,a}[h(s')]-h(s)$).
Thus, HuRL\xspace~incorporates guidance discount into PBRS
with nicer properties.
}
\subsection{Bias and Heuristic Quality}
\label{sec:heuristic_quality}
\vspace{-1mm}
The bias term in \eqref{eq:bias} characterizes suboptimality due to using a heuristic $h$ in place of long-term state values in $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$.
What is the best heuristic in this case?
From the definition of the bias term in \eqref{eq:bias}, we see that the ideal heuristic is the optimal value $V^*$, as $ \mathrm{Bias}(V^*,\lambda,\pi) = 0$ for all $\lambda\in[0,1]$.
By continuity, we can expect that if $h$ deviates from $V^*$ a little, then the bias is small.
\begin{restatable}{corollary}{LinfBiasBound}\label{th:l-inf bias bound}
If $\inf_{b\in\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb} \|h + b - V^*\|_\infty \leq \epsilon$, then $\mathrm{Bias}(h,\lambda,\pi)
\leq \frac{(1-\lambda\gamma)^2}{(1-\gamma)^2} \epsilon
$.
\end{restatable}
To better understand how the heuristic $h$ affects the bias, we derive an upper bound on the bias by replacing the first term in \eqref{eq:bias} with an upper bound that depends only on $\pi^*$.
\begin{restatable}{proposition}{BiasBound} \label{th:bias bound}
For $g:\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\to\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$ and $\eta\in[0,1]$, define $\CC(\pi,g,\eta) \coloneqq \Ebb_{\rho^{\pi}(d_0)} \left[ \sum_{t=1}^\infty \eta^{t-1} g(s_t) \right] $.
Then
$
\mathrm{Bias}(h,\lambda,\pi) \leq
(1-\lambda)\gamma ( \CC(\pi^*,V^*-h, \lambda\gamma) +
\CC(\pi,h-\widetilde{V}^*, \gamma)
)
$.
\end{restatable}
In \cref{th:bias bound}, the term $(1-\lambda)\gamma \CC(\pi^*,V^*-h, \lambda\gamma)$ is the underestimation error of the heuristic $h$ under the states visited by the optimal policy $\pi^*$ in the original MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$.
Therefore, to minimize the first term in the bias, we would want the heuristic $h$ to be large along the paths that $\pi^*$ generates.
However, \cref{th:bias bound} also discourages the heuristic from being arbitrarily large, because the second term in the bias in \eqref{eq:bias} (or, equivalently, the second term in \cref{th:bias bound}) incentivizes the heuristic to underestimate the optimal value of the reshaped MDP $\widetilde{V}^*$.
More precisely, the second term requires the heuristic to obey some form of spatial consistency. A quick intuition is the observation that if $h(s)=V^{\pi'}(s)$ for some $\pi'$ or $h(s)=0$, then $h(s) \leq \widetilde{V}^*(s)$ for all $s\in\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}$.
More generally, we show that if the heuristic is \emph{improvable} with respect to the original MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$ (i.e. the heuristic value is lower than that of the max of Bellman backup), then $h(s) \leq \widetilde{V}^*(s)$.
By \cref{th:bias bound}, learning with an improvable heuristic in HuRL\xspace has a much smaller bias.
\begin{definition}
Define the Bellman operator $(\BBh)(s,a) \coloneqq r(s,a) +\gamma \Ebb_{s'|s,a}[h(s')]$.
A heuristic function $h:\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\to\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$ is said to be \emph{improvable} with respect to an MDP $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$ if $\max_a (\BB h)(s,a) \geq h(s)$.
\end{definition}
\begin{restatable}{proposition}{ImprovableHeuristic} \label{th:improvable heuristic}
If $h$ is improvable with respect to $\mathcal{M}}\def\NN{\mathcal{N}}\def\OO{\mathcal{O}$, then $\widetilde{V}^*(s) \geq h(s) $, for all $\lambda\in[0,1]$.
\end{restatable}
\vspace{-1mm}
\subsection{Pessimistic Heuristics are Good Heuristics }
\vspace{-1mm}
While \cref{th:l-inf bias bound} shows that HuRL\xspace can handle an imperfect heuristic, this result is not ideal. The corollary depends on the $\ell_\infty$ approximation error, which can be difficult to control in large state spaces.
Here we provide a more refined sufficient condition of good heuristics. We show that the concept of \emph{pessimism} in the face of uncertainty provides a finer mechanism for controlling the approximation error of a heuristic and would allow us to remove the $\ell_\infty$-type error.
This result is useful for constructing heuristics from data that does not have sufficient support.
From \cref{th:bias bound} we see that the source of the $\ell_\infty$ error is the second term in the bias upper bound, as it depends on the states that the agent's policy visits which can change during learning.
To remove this dependency, we can use improvable heuristics (see \cref{th:improvable heuristic}), as they satisfy $h(s) \leq \widetilde{V}^*(s)$.
Below we show that Bellman-consistent pessimism
yields improvable heuristics.
\begin{restatable}{proposition}{BellmanPessimismAndImprovableHeuristic} \label{th:bellman pessimism and improvable heuristic}
Suppose $h(s) = Q(s,\pi')$ for some policy $\pi'$ and function $Q:\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}\times\mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C}\to\Rbb}\def\Q{\Qbb}\def\Z{\Zbb}\def\N{\Nbb}\def\C{\Cbb$ such that
$ Q(s,a) \leq (\BB h) (s,a) $,
$\forall s\in\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}$, $a\in\mathcal{A}}\def\BB{\mathcal{B}}\def\CC{\mathcal{C}$.
Then $h$ is improvable and $f(s)\leq V^{\pi'}(s)$ for all $s\in\mathcal{S}}\def\TT{\mathcal{T}}\def\UU{\mathcal{U}$.
\end{restatable}
The Bellman-consistent pessimism in \cref{th:bellman pessimism and improvable heuristic} essentially says that $h$ is pessimistic with respect to the Bellman backup.
This condition has been used as the foundation for designing pessimistic off-policy RL algorithms, such as pessimistic value iteration~\citep{jin2020pessimism} and algorithms based on pessimistic absorbing MDPs~\citep{liu2020provably}.
In other words, these pessimistic algorithms can be used to construct good heuristics with small bias in \cref{th:bias bound} from offline data.
With such a heuristic, the bias upper bound would be simply
$\mathrm{Bias}(h,\lambda,\pi) \leq
(1-\lambda)\gamma \CC(\pi^*,V^*-h, \lambda\gamma)$.
Therefore, as long as enough batch data are sampled from a distribution that covers states that $\pi^*$ visits,
these pessimistic algorithms can construct good heuristics with nearly zero bias for HuRL\xspace with high probability.
\section{Related Work}
\label{sec:related}
\vspace{-1mm}
\vspace{-2mm}
\paragraph{Discount regularization.}
The horizon-truncation idea can be traced back to Blackwell optimality in the known MDP setting~\citep{blackwell1962discrete}.
Reducing the discount factor amounts to running HuRL\xspace with a zero heuristic.
\citet{petrik2008biasing,jiang2015dependence,jiang2016structural} study the MDP setting; \citet{chen2018improving} study POMDPs. \citet{amit2020discount} focus on discount regularization for Temporal Difference (TD) methods, while \citet{NEURIPS2019_eba237ec} use a logarithmic mapping
to lower the discount for online RL.
\vspace{-2mm}
\paragraph{Reward shaping.}
Reward shaping has a long history in RL, from the seminal PBRS work~\citep{ng1999policy} to recent bilevel-optimization approaches~\citep{hu2020learning}.
\citet{tessler2020maximizing} consider a complementary problem to HuRL\xspace: given a discount $\gamma'$, they find a reward $r'$ that preserves trajectory ordering in the original MDP.
Meanwhile there is a vast literature on bias-variance trade-off for online RL with horizon truncation. TD($\lambda$)~\citep{seijen2014true,efroni2018beyond} and Generalized Advantage Estimates~\citep{schulman2015high} blend value estimates across discount factors, while~\citet{sherstan2020gamma} use the discount factor as an input to the value function estimator. TD($\Delta$)~\citep{romoff2019separating} computes differences between value functions across discount factors.
\paragraph{Heuristics in model-based methods.}
Classic uses of heuristics include A*~\citep{hart1968formal}, Monte-Carlo Tree Search (MCTS)~\citep{browne2012survey}, and Model Predictive Control (MPC)~\citep{richalet1978model}.
\citet{zhong2013value} shorten the horizon in MPC using a value function approximator.
\citet{hoeller2020deep} additionally use an estimate for the running cost to trade off solution quality and amount of computation.
\citet{bejjani2018planning} show heuristic-accelerated truncated-horizon MPC on actual robots and tune the value function throughout learning.
\citet{bhardwaj2020blending} augment MPC with a terminal value heuristic, which can be viewed as an instance of HuRL\xspace where the base algorithm is MPC.
\rev{\citet{asai2020} learn an MDP expressible in the STRIPS formalism that can benefit from relaxation-based planning heuristics.}
\rev{But HuRL\xspace is more general, as it does not assume model knowledge and can work in unknown environments.}
\vspace{-2mm}
\paragraph{Pessimistic extrapolation.}
Offline RL techniques employ pessimistic extrapolation for robustness~\citep{jin2020pessimism}, and their learned value functions can be used as heuristics in HuRL\xspace.
\citet{kumar2020conservative} penalize out-of-distribution actions in off-policy optimization while \citet{liu2020provably} additionally use a variational auto-encoder (VAE) to detect out-of-distribution states. We experimented with VAE-filtered pessimistic heuristics in \cref{app:experiments}.
Even pessimistic offline evaluation techniques~\citep{gulcehre2021regularized} can be useful in HuRL\xspace, since function approximation often induces extrapolation errors~\cite{lu2018non}.
\vspace{-2mm}
\paragraph{Heuristic pessimism vs. admissibility.}
\rev{Our concept of heuristic pessimism can be easily confused for the well-established notion of \emph{admissibility}~\citep{russell2020}, but in fact they are opposites. Namely, an admissible heuristic never \emph{underestimates} $V^*$ (in the return-maximization setting),
while a pessimistic one never \emph{overestimates} $V^*$. Similarly, our notion of improvability is distinct from \emph{consistency}: they express related ideas, but with regards to pessimistic and admissible value functions, respectively. Thus, counter-intuitively from the planning perspective, our work shows that for policy \emph{learning}, \emph{in}admissible heuristics are desirable. \citet{pearl1981} is one of the few works that has analyzed desirable implications of heuristic inadmissibility in planning.
}
\vspace{-2mm}
\paragraph{Other warm-starting techniques.}
HuRL\xspace is a new way to warm-start online RL methods.
\citet{bianchi2013heuristically} use a heuristic policy to initialize agents' policies.
\citet{hester2018deep,vinyals2019grandmaster} train a value function and policy using batch IL and then used them as regularization in online RL.
\citet{nair2020accelerating} use off-policy RL on batch data and fine-tune the learned policy.
Recent approaches of hybrid IL-RL have strong connections to HuRL\xspace~\citep{sun2017deeply,cheng2020policy,sun2018truncated}.
In particular, \citet{cheng2020policy} is a special case of HuRL\xspace with a max-aggregation heuristic.
\citet{farahmand2016truncated} use several related tasks to learn a task-dependent heuristic
and perform shorter-horizon planning or RL. \rev{Knowledge distillation approaches~\citep{hinton2015distilling} can also be used to warm-start learning, but in contrast to them, HuRL\xspace\ expects prior knowledge in the form of state value estimates, not features, and doesn't attempt to make the agent internalize this knowledge. A
HuRL\xspace\ agent learns from its own environment interactions, using prior knowledge only as guidance.
Reverse Curriculum approaches~\cite{florensa2017reverse} create short horizon RL problems by initializing the agent close to the goal, and moving it further away as the agent improves. This gradual increase in the horizon inspires the HuRL\xspace approach. However, HuRL\xspace does not require the agent to be initialized on expert states and can work with many different base RL algorithms.
}
|
\section{Introduction}
Task-oriented dialogue (ToD) systems are designed to assist humans in performing daily activities, such as ticket booking, travel planning, and online shopping. These systems are the core modules of virtual assistants (e.g., Apple Siri and Amazon Alexa), and they provide natural language interfaces for online services~\cite{rastogi2020towards}. Recently, there has been growing interest in developing deep learning-based end-to-end ToD systems \cite{bordes2016learning,wen2017network,eric2017key,qin2019entity,qin2020dynamic,banerjee2019graph,neelakantan2019neural,eric2017copy,madotto2018mem2seq,reddy2019multi,wu2019global,hosseini2020simple,peng2020soloist,lin2020mintl,byrne2020tickettalk} because they can handle complex dialogue patterns with minimal hand-crafted rules. To advance the existing state-of-the-art, large-scale datasets~\cite{budzianowski2018multiwoz,rastogi2020towards,byrne2020tickettalk} have been proposed for training and evaluating such data-driven systems.
However, existing datasets for end-to-end ToD modelling are limited to a single language, such as English~\cite{budzianowski2018multiwoz,mosig2020star}, or Chinese~\cite{zhu2020crosswoz,quan2020risawoz}. The absence of bilingual or multilingual datasets not only limits the research on cross-lingual transfer learning~\cite{razumovskaia2021crossing} but also hinders the development of robust end-to-end ToD systems for multilingual countries and regions.
To tackle the challenge mentioned above, we introduce {\bbfamily BiToD}, a bilingual multi-domain dataset for task-oriented dialogue modelling. {\bbfamily BiToD} has 7,232 bilingual dialogues (in English and Chinese), spanning seven services within five domains, where each dialogue is annotated with dialogue states, speech-acts, and service API calls. Therefore, {\bbfamily BiToD} can be used for building both end-to-end ToD systems and dialogue sub-modules (e.g., Dialogue State Tracking).
We propose three evaluation settings: 1) \textbf{\textit{monolingual}}, in which the models are trained and tested on either English or Chinese data, 2) \textbf{\textit{bilingual}}, where the models are trained with bilingual data and tested with English and Chinese dialogues simultaneously, and 3) \textbf{\textit{cross-lingual}}, where the models are first trained with the source language and then tested in a few-shot setting in the target language.
\begin{figure}%
\begin{center}
\begin{adjustbox}{width=0.48\textwidth}
\begin{tikzpicture}
\definecolor{chatcolor1}{HTML}{E2F0FF}
\definecolor{chatcolor2}{HTML}{FDF0FF}
\definecolor{chatcolor3}{HTML}{FFC9CF}
\definecolor{chatcolor4}{HTML}{E6E6E6}
{\small
\fontfamily{cmss}\selectfont
\node[align=left, text width=6cm, fill=chatcolor2, rounded corners=1mm, anchor=north west] at (2,0) {\includegraphics[height=0.29cm]{img/emoji/girl.png}: Hi, can you help me find a place to eat?};
\node[align=left, text width=6cm,fill=chatcolor1, rounded corners=1mm, anchor=north west] at (0,-.97) {\includegraphics[height=0.29cm]{img/emoji/robot.png}: Sure! How much do you want to spend and how high of a rating would you prefer?};
\node[align=left, text width=6cm, fill=chatcolor2, rounded corners=1mm, anchor=north west] at (2,-1.99) {\includegraphics[height=0.29cm]{img/emoji/girl.png}: I'd like to eat at an expensive restaurant rated at least 9.};
\node[align=left, text width=6cm,fill=chatcolor1, rounded corners=1mm, anchor=north west] at (0,-2.94) {\includegraphics[height=0.29cm]{img/emoji/robot.png}: Got it. What kind of food do you want?};
\node[align=left, text width=6cm, fill=chatcolor2, rounded corners=1mm, anchor=north west] at (2,-3.87) {\includegraphics[height=0.29cm]{img/emoji/girl.png}: Any type of food is fine, but I want a place with Vegan Options.};
\node[align=left, text width=6.5cm,fill=chatcolor3, rounded corners=1mm, anchor=north west] at (0,-4.87) {\includegraphics[height=0.29cm]{img/emoji/robot.png}\includegraphics[height=0.29cm]{img/emoji/glass.png}: \texttt{RESTAURANTS\_SEARCH(dietary=Vegan,\\ \quad\quad \ rating$\geq$9, price=expensive)}};
\node[align=left, text width=6.5cm,fill=chatcolor4, rounded corners=1mm, anchor=north west] at (0,-5.85) { \texttt{\textbf{name}: Sushi Tsubomi\\
\textbf{location}: \framebox[\width]{\textit{Wan Chai}}\\
\textbf{cuisine}: Japanese\\
\textbf{price}: expensive\\
\textbf{rating}: 10\\
\textbf{dietary}: Vegan\\
\textbf{available options}: 51\\}};
\node[align=left, text width=6cm,fill=chatcolor1, rounded corners=1mm, anchor=north west] at (0,-8.49) {\includegraphics[height=0.29cm]{img/emoji/robot.png}: Okay. Out of 51 restaurants, I recommend you go to Sushi Tsubomi. They are a Japanese restaurant rated at a 10.};
\node[align=center, text width=1cm, rounded corners=1mm, anchor=north west] at (2,-9.8) { {\Large$\cdots$} };
\node[align=left, text width=6cm, fill=chatcolor2, rounded corners=1mm, anchor=north west] at (2,-10.25) {\includegraphics[height=0.29cm]{img/emoji/girl.png}: I'd like to visit an attraction nearby as well, rated at an 8 minimum.};
\node[align=left, text width=6cm,fill=chatcolor1, rounded corners=1mm, anchor=north west] at (0,-11.25) {\includegraphics[height=0.29cm]{img/emoji/robot.png}: Sure, I can help with that. What type of attraction would you like?};
\node[align=left, text width=6cm, fill=chatcolor2, rounded corners=1mm, anchor=north west] at (2,-12.19) {\includegraphics[height=0.29cm]{img/emoji/girl.png}: I'd like to visit some famous landmarks.};
\node[align=left, text width=6cm,fill=chatcolor3, rounded corners=1mm, anchor=north west] at (0,-12.8) {\includegraphics[height=0.29cm]{img/emoji/robot.png}\includegraphics[height=0.29cm]{img/emoji/glass.png}: \texttt{ATTRACTIONS\_SEARCH(rating$\geq$8, type=Landmarks, location=\framebox[\width]{\textit{Wan Chai}})}};
\node[align=left, text width=6cm,fill=chatcolor4, rounded corners=1mm, anchor=north west] at (0,-13.96) { \texttt{\textbf{name}: Hong Kong Convention $\cdots$\\
\textbf{location}: Wan Chai\\
\textbf{rating}: 8\\
\textbf{type}: Sights \& Landmarks\\
\textbf{address}: 1 Expo Dr., Wanchai\\
\textbf{phone}: 2582 8888\\
\textbf{available options}: 5\\}};
\node[align=center, text width=1cm, rounded corners=1mm, anchor=north west] at (2,-16.47) {};
}
\end{tikzpicture}
\end{adjustbox}
\quad
\begin{adjustbox}{width=0.48\textwidth}
\begin{tikzpicture}
\definecolor{chatcolor1}{HTML}{E2F0FF}
\definecolor{chatcolor2}{HTML}{FDF0FF}
\definecolor{chatcolor3}{HTML}{FFC9CF}
\definecolor{chatcolor4}{HTML}{E6E6E6}
{\small
\fontfamily{cmss}\selectfont
\begin{CJK*}{UTF8}{gkai}
\node[align=left, text width=6cm, fill=chatcolor2, rounded corners=1mm, anchor=north west] at (2,0) {\includegraphics[height=0.29cm]{img/emoji/girl.png}: 你好，我想订一家餐厅吃饭，我不想吃法餐。};
\node[align=left, text width=6cm,fill=chatcolor1, rounded corners=1mm, anchor=north west] at (0,-0.97) {\includegraphics[height=0.29cm]{img/emoji/robot.png}: 好的，请问您是在哪里找餐厅？要什么价位的呢？};
\node[align=left, text width=6cm, fill=chatcolor2, rounded corners=1mm, anchor=north west] at (2,-1.96) {\includegraphics[height=0.29cm]{img/emoji/girl.png}: 位置和价位都不重要。但需要在9分以上。};
\node[align=left, text width=6cm,fill=chatcolor3, rounded corners=1mm, anchor=north west] at (0,-2.87) {\includegraphics[height=0.29cm]{img/emoji/robot.png}\includegraphics[height=0.29cm]{img/emoji/glass.png}: \texttt{餐馆查询(评分$\geq$9, 菜品≠法餐)}};
\node[align=left, text width=6cm,fill=chatcolor4, rounded corners=1mm, anchor=north west] at (0,-3.47) { \texttt{名字: ChocoDuck Bistro\\
位置: 新界\\
菜品: 意餐\\
价格范围: 适中\\
评分 :10\\
饮食限制: 适合素食主义者\\
地址: 3 Kin San Lane, Cheung Chau\\
电话: 9735 0049\\
可用选项: 463\\}};
\node[align=left, text width=6cm,fill=chatcolor1, rounded corners=1mm, anchor=north west] at (0,-6.82) {\includegraphics[height=0.29cm]{img/emoji/robot.png}: 有463家不错的餐厅，其中有一家评分为10分，名字叫ChocoDuck Bistro，他主营意餐。};
\node[align=left, text width=6cm, fill=chatcolor2, rounded corners=1mm, anchor=north west] at (2,-8.1) {\includegraphics[height=0.29cm]{img/emoji/girl.png}: 还有别的吗};
\node[align=left, text width=6cm,fill=chatcolor3, rounded corners=1mm, anchor=north west] at (0,-8.78) {\includegraphics[height=0.29cm]{img/emoji/robot.png}\includegraphics[height=0.29cm]{img/emoji/glass.png}: \texttt{餐馆查询(菜品$\neq$法餐, 评分$\geq$9, 名字$\neq$ChocoDuck Bistro)} };
\node[align=left, text width=6cm,fill=chatcolor4, rounded corners=1mm, anchor=north west] at (0,-9.77) { \texttt{名字: Hooked\\
位置: 中区\\
菜品: 海鲜\\
价格范围: 便宜\\
评分: 10\\
饮食限制: null\\
地址: 半山坚道86-88号\\
电话: 2915 1118\\
可用选项: 462\\}};
\node[align=left, text width=6cm,fill=chatcolor1, rounded corners=1mm, anchor=north west] at (0,-13.2) {\includegraphics[height=0.29cm]{img/emoji/robot.png}: 还有462家，推荐您可以试试Hooked，它以海鲜为主，评分是10分};
\node[align=left, text width=6cm, fill=chatcolor2, rounded corners=1mm, anchor=north west] at (2,-14.15) {\includegraphics[height=0.29cm]{img/emoji/girl.png}: 好，帮我订一下，星期二下午5:50的。};
\node[align=left, text width=6cm,fill=chatcolor1, rounded corners=1mm, anchor=north west] at (0,-14.73) {\includegraphics[height=0.29cm]{img/emoji/robot.png}: 好的，请问您几位用餐？怎么称呼您？};
\node[align=left, text width=6cm, fill=chatcolor2, rounded corners=1mm, anchor=north west] at (2,-15.3) {\includegraphics[height=0.29cm]{img/emoji/girl.png}: 马连良，5位。};
\node[align=left, text width=6cm,fill=chatcolor1, rounded corners=1mm, anchor=north west] at (0,-15.87) {\includegraphics[height=0.29cm]{img/emoji/robot.png}: 好的，跟您确认一下，为您预订星期二下午5:50，Hooked餐厅5位用餐。};
\end{CJK*}
}
\end{tikzpicture}
\end{adjustbox}
\end{center}
\caption{Dialogue examples in English (on the left) and Chinese (on the right) from the {\bbfamily BiToD} dataset (best viewed in colour). The user (in pink) asks for restaurant and attraction information. At the same time, the system (in blue) responds to the user utterance by calling APIs (in red) when needed and by incorporating the API results (in grey). }%
\label{fig:example}%
\end{figure}
The contribution of this work is three-fold. 1) We propose the first bilingual dataset ({\bbfamily BiToD}) with a total of 7, 232 dialogues for \textit{\textbf{end-to-end}} ToD modeling. {\bbfamily BiToD} serves as an effective benchmark for evaluating bilingual ToD systems and cross-lingual transfer learning approaches. 2) We provide novel baselines under the three evaluation settings, i.e., monolingual, bilingual, and cross-lingual. 3) We show the effectiveness of training a bilingual ToD system compared to two independent monolingual ToD systems as well as the potential of leveraging a bilingual knowledge base and cross-lingual transfer learning to improve the system performance under low resource condition.
The paper is organized as follows: We next describe the {\bbfamily BiToD} data collection methods in Section 2. We then describe our proposed tasks in section 3. Section 4 introducew our baselines, and we finally present and discuss results in Section 5.
\section{{\bbfamily BiToD} Dataset}
{\bbfamily BiToD} is designed to develop virtual assistants in multilingual cities, regions, or countries (e.g., Singapore, Hong Kong, India, Switzerland, etc.). For the {\bbfamily BiToD} data collection, we chose Hong Kong since it is home to plenty of attractions, restaurants and more, and is one of the most visited cities globally, especially by English and Chinese speakers. This section describes the knowledge base construction and provides detailed descriptions of the dialogue collection.
\subsection{Knowledge Base Collection}
We collect publicly available Hong Kong tourism information from the Web, to create a knowledge base that includes 98 metro stations, 305 attractions, 699 hotels, and 1,218 restaurants. For the weather domain, we synthetically generate the weather information on different dates. Then, we implement seven service APIs (\texttt{Restaurant\_Searching}, \texttt{Restaurant\_Booking}, \texttt{Hotel\_Searching}, \texttt{Hotel\_Booking}, \texttt{Attraction\_Searching}, \texttt{MTR\_info}, \texttt{Weather\_info}) to query our knowledge base. The knowledge base statistics are shown in Table \ref{tab:kb}. Although we aim to collect a fully parallel knowledge base, we observe that some items do not include bilingual information. For example, several traditional Cantonese restaurants do not have English names, and similarly, some restaurants do not provide addresses in Chinese. This lack of parallel information reflects the real-world challenges that databases are often incomplete and noisy.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{img/BITOD.png}
\caption{Illustration of the bilingual dialogues collection pipeline: a) Design a schema for each service API; b) Sample user goals from the bilingual knowledge base (KB) according to schema; c) Based on one of the user goals, the dialogue simulator generates the dialogue outlines while interacting with the APIs; d) Convert the dialogue outlines to natural conversations via crowdsourcing. Note that English and Chinese user goals are sampled independently.}
\label{fig:data_collection}
\end{figure}
\subsection{Dialogue Data Collection}
The dialogues are collected through a four-phase pipeline, as shown in Figure \ref{fig:data_collection}. We first design a schema, as a flowchart, for each service API, to specify the possible API queries and expected system actions after the API call. Then, user goals are sampled from the knowledge base according to the pre-defined schemas. Based on the user goals, the dialogue simulator interacts with the APIs to generate dialogue outlines.
Finally, the dialogue outlines are converted into natural conversations through crowdsourcing. Our data collection methodology extends the Machine-to-Machine (M2M) approaches~\cite{shah2018building,rastogi2020towards} to bilingual settings to minimize the annotation overhead (time and cost).
\begin{table}[t]
\centering
\begin{tabular}{@{}lcccccccccc@{}}
\toprule
\textbf{Domain} & \multicolumn{2}{c}{\textbf{Attraction}} & \multicolumn{2}{c}{\textbf{Hotel}} & \multicolumn{2}{c}{\textbf{Restaurant}} & \multicolumn{2}{c}{\textbf{Weather}} & \multicolumn{2}{c}{\textbf{Metro}} \\ \midrule
Language & EN & ZH & EN & ZH & EN & ZH & EN & ZH & EN & ZH \\ \midrule
\# Slots & 7 & 7 & 9 & 9 & 13 & 13 & 5 & 5 & 5 & 5 \\
\# Entities & 1,079 & 1,118 & 2,642 & 2,652 & 5,489 & 5,035 & 77 & 77 & 161 & 161 \\ \bottomrule
\end{tabular}
\caption{Bilingual knowledge base statistics.}
\label{tab:kb}
\end{table}
\paragraph{Schemas and APIs.}
The dialogue schema shown as a flowchart (\texttt{Restaurant\_Searching}) in Figure \ref{fig:data_collection}.a specifies the input and output options of the API and the desired system behaviours. To elaborate, the user searches a restaurant by name, location, cuisine, etc. Then the system calls the API and informs the user of the restaurant name and other requested information. If the user is not satisfied with the search results, the system continues searching and provides other options. To ensure the provided services are realistic, we impose a few restrictions, as in \cite{rastogi2020towards}. Firstly, each API has a list of required slots, and the system is not allowed to hit the API without specifying values for these slots. For example, the system needs to obtain departure and destination locations before calling the metro-info API. Secondly, the system must confirm the booking information with the user before making any reservations (e.g., restaurant booking).
\paragraph{User Goals.}
A user goal consists of a list of intents and a set of constraints under each intent. Figure \ref{fig:data_collection}.b shows a single domain (intent) example where the user's intent is \textit{Restaurant\_Search}. A constraint is defined with a triple (slot, relation, value) (e.g., \texttt{(Rating, at\_least, 4)}). Different from previous work, which defined user constraints as slot-value pairs, we impose slot-value relations~\cite{mosig2020star} (listed in Figure~\ref{fig:histo}.b) to promote more diverse user goals. To generate a user goal, we first sample a list of intents. We randomly sample a set of slot-relation-value combinations from the bilingual knowledge base for each intent, which includes non-existent combinations to create unsatisfiable user requests. In multi-domain scenarios, we set a certain probability to share the same values for some of the cross-domain slots (e.g., date and location) to make the transition among domains smooth. For example, users might want to book restaurants and hotels on the same date or take the metro from the location of the booked restaurant to their hotel. Note that the user goals for English and Chinese are sampled independently, as the real-world customer service conversations are often unparalleled.
\paragraph{Dialogue Outline Generation.}
Dialogue outlines are generated by a bilingual dialogue simulator that accepts user goals in both languages as inputs. The dialogue simulator consists of a user agent and a system agent. Both agents interact with each other using a finite set of actions specified by speech acts over a probabilistic automaton designed to capture varied dialogue trajectories \cite{rastogi2020towards}. Each speech act takes a slot or slot-relation-value triple as an argument. When the conversation starts, the user agent is assigned a goal, while the system agent is initialized with a set of requests related to the services. During the conversation, the user informs constraints according to the user goal, and the system responds to the user queries while interacting with the service APIs. For some services, the system needs to request all the required slots before querying the APIs. After the API call, the system either informs the search result or searches for other options until the user intents are fulfilled. Following \cite{rastogi2020towards}, we also augment the value entities during the dialogue outlines generation process, e.g., \textit{\textbf{Tsim Sha Tsui}} can be replaced with its abbreviation \textit{\textbf{TST}}, as shown in Figure \ref{fig:data_collection}.c. After the user goal is fulfilled by a series of user and system actions, we convert all the actions into natural language using templates. In this phase, we obtain the dialogue states annotations and speech acts automatically for both the user and system sides.
\paragraph{Dialogue Paraphrase.} The dialogue outlines are converted to natural dialogues via crowdsourcing. Figure \ref{fig:zh_interface} and \ref{fig:en_interface} show the interface for Chinese and English paraphrasing, where workers see the full dialogue and rewrite the dialogue turn by turn. Before the task, workers are asked to read the instructions, shown in Figure \ref{fig:en_instructions}. In the instructions, we specify that the paraphrased dialogue should retain the same meaning as the dialogue outline but sound like a real conversation between a user and a professional assistant. The user utterances are expected to be creative and diverse, while the system utterances are expected to be formal and correct. To ensure all the essential information is presented in the new dialogue, we highlight all the entities with bold text. In the user utterances, the highlighted entities are allowed to be paraphrased without losing their original meaning; e.g., \textit{``The restaurant should provide \textbf{Vegan Options}"} is allowed to be rewritten as \textit{``I would like to find a \textbf{vegan-friendly} restaurant"}. In contrast, all the entities in the system utterances are required to be unchanged.
\paragraph{Quality Verification.} After the dialogue paraphrasing, workers are asked to read through the new dialogue and answer the following questions, as in ~\cite{shah2018building}: 1) \textit{Does it seem like a conversation between a user that sounds like you and an assistant that sounds formal?} 2) \textit{Does it have the same meaning as the original conversation, while still making sense on its own?} The first question is for examining whether the new conversation is realistic, and the second question is for verifying whether the dialogue outline and the paraphrased dialogue are valid. Given the two answer options: 1) \textit{Yes}, 2) \textit{No, but I cannot make it better}, 97.56\% of annotators chose the first option for the first question and 98.89\% of them chose the first option for the second question.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{img/hist_plot.png}
\caption{Data statistic of {\bbfamily BiToD}: a) dialogue distribution of lengths of single and multi-domain dialogues, b) distribution of different relation types, and c) distribution of speech acts of users and systems.}
\label{fig:histo}
\end{figure}
\subsection{Dataset Statistics}
We collected 7,232 dialogues with 144,798 utterances, in which 3,689 dialogues are in English, and 3,543 dialogues are in Chinese. We split the data into 80\% training, 8\% validation, and 12\% testing, resulting in 5,787 training dialogues, 542 validation dialogues, and 902 testing dialogues. In Figure~\ref{fig:histo} we show the main data statistics of the BiToD corpus. As shown in Figure~\ref{fig:histo}.a, the lengths of the dialogues vary from 10 turns to more than 50 turns. Multi-domain dialogues, in both English and Chinese, have many more turns compared to single-domains. The most used relation in user goals is \texttt{equal\_to} (Figure~\ref{fig:histo}.b), and the most common speech-acts (Figure~\ref{fig:histo}.c) for users and systems are \texttt{inform} and \texttt{offer}, respectively. Finally, in Table~\ref{tab:ontology} in the Appendix, we list all the informable and requestable slots per domain.
\subsection{Dataset Features}
Table \ref{tab:comparison} shows the comparison of the {\bbfamily BiToD} training set to previous ToD datasets. Prior work for end-to-end ToD modelling only focuses on a single language. Our {\bbfamily BiToD} is the first \textbf{\textit{bilingual}} ToD corpus with comparable data size. In addition to its bilingualism, {\bbfamily BiToD} also provides the following unique features:
\paragraph{Deterministic API.} Given an API query for recommendation services (e.g., restaurant searching and hotel searching), there is typically more than one matched item. Previous works~\cite{budzianowski2018multiwoz,zhu2020crosswoz} have randomly sampled one or two items as API results and returned them to users. However, in real-world applications, the system should recommend items according to certain criteria (e.g., user rating).
Moreover, the randomness of the API also increases the difficulty of evaluating the models. Indeed, the evaluation metrics in \cite{budzianowski2018multiwoz,zhu2020crosswoz} rely on delexicalized response templates, which are not compatible with knowledge-grounded generation approaches~\cite{madotto2018mem2seq,wu2019global}.
To address these issues, we implement deterministic APIs by ranking the matched items according to user ratings.
\paragraph{Complex User Goal.} To simulate more diverse user goals, we impose different relations for slot-value pairs. For example, in the restaurant searching scenarios, a user might want to eat Chinese food \texttt{(cuisine, equal\_to, Chinese)}, or do not want Chinese food \texttt{(cuisine, not, Chinese)}. Figure~\ref{fig:histo}.b shows the distribution of different relations in user goals.
\begin{table}[t]
\resizebox{\linewidth}{!}{
\begin{tabular}{@{}r|cccccccc@{}}
\toprule
\multicolumn{1}{l|}{} & \textbf{MultiWoZ} & \textbf{FRAMES} & \textbf{TM-1} & \textbf{SGD} & \textbf{STAR} & \textbf{RiSAWOZ} & \textbf{CrossWoz} & \textbf{BiToD} \\ \midrule
\textit{Language(s)} & EN & EN & EN & EN & EN & ZH & ZH & EN, ZH \\
\textit{Number of dialogues} & 8,438 & 1,369 & 13,215 & 16,142 & 5,820 & 10,000 & 5,012 & 5,787 \\
\textit{Number of domains} & 7 & 1 & 6 & 16 & 13 & 12 & 5 & 5 \\
\textit{Number of APIs} & 7 & 1 & 6 & 45 & 24 & 12 & 5 & 7 \\
\textit{Total number of turns} & 115,434 & 19,986 & 274,647 & 329,964 & 127,833 & 134,580 & 84,692 & 115,638 \\
\textit{Average turns / dialogues} & 13.46 & 14.6 & 21.99 & 20.44 & 21.71 & 13.5 & 16.9 & 19.98 \\
\textit{Slots} & 25 & 61 & - & 214 & - & 159 & 72 & 68* \\
\textit{Values} & 4,510 & 3,871 & - & 14,139 & - & 4,061 & 7,871 & 8,206* \\
\textit{Deterministic API} & \xmark & \xmark & \xmark & \xmark & \xmark & \xmark & \xmark & \cmark \\
\textit{Complex User Goal} & \xmark & \xmark & \xmark & \xmark & \cmark & \xmark & \xmark & \cmark \\
\textit{Mixed-Language Context} & \xmark & \xmark & \xmark & \xmark & \xmark & \xmark & \xmark & \cmark \\
\textit{Provided KB} & \cmark & \xmark & \xmark & \xmark & \cmark & \cmark & \cmark & \cmark \\ \bottomrule
\end{tabular}
}
\caption{Comparison of BiToD to previous ToD datasets. The numbers are provided for the \textit{\textbf{training set}} except for FRAMES and STAR. *We consider entities in different language as different slots and values.}
\label{tab:comparison}
\end{table}
\paragraph{Mixed-Language Context.} Our corpus contains code-switching utterances as some of the items in the knowledge base have mixed-language information. In the example in Figure~\ref{fig:example}.b, the system first recommends a restaurant called \textit{ChocoDuck Bistro} and the user asks for other options. Then the system searches other restaurants with an additional constraint \texttt{(restaurant\_name, not, ChocoDuck Bistro)}. In this example, both restaurants only have English names, which is a common phenomenon in multilingual regions like Hong Kong. Thus, ToD systems need to handle the mixed-language context to make correct API calls.
\paragraph{Cross-API Entity Carry-Over} Our corpus includes scenarios where the value of a slot is not presented in the conversation, and the system needs to carry over values from previous API results. In the example in Figure \ref{fig:example}.a, the user first finds and books a restaurant without specifying the location; then she (\includegraphics[height=0.29cm]{img/emoji/girl.png}) wants an attraction nearby the restaurant. In this case, the system needs to infer the attraction location (\texttt{\framebox[\width]{\textit{Wan Chai}}}) from the restaurant search result.
\section{Tasks \& Evaluations}
\subsection{Dialogue State Tracking}
Dialogue state tracking (DST), an essential task for ToD modelling, tracks the users' requirements over multi-turn conversations. DST labels provide sufficient information for a ToD system to issue APIs and carry out dialogue policies. In this work, we formulate a dialogue state as a set of slot-relation-value triples.
We use Joint Goal Accuracy (\textbf{JGA}) to evaluate the performance of the DST. The model outputs are correct when all of the predicted slot-relation-value triples exactly match the oracle triples.
\subsection{End-to-End Task Completion}
A user's requests are fulfilled when the dialogue system makes correct API calls and correctly displays the requested information. We use the following automatic metrics to evaluate the performance of end-to-end task completion: 1) Task Success Rate (\textbf{TSR}): whether the system provides the correct entity and answers all the requested information of a given task, 2) Dialogue Success Rate (\textbf{DSR}): whether the system completes all the tasks in the dialogue, 3) API Call Accuracy ($\textbf{API}_{Acc}$): whether the system generates a correct API call, and 4) \textbf{BLEU}~\cite{papineni2002bleu}: measuring the fluency of the generated response.
\subsection{Evaluation Settings}
\paragraph{Monolingual.} Under the monolingual setting, models are trained and tested on either English or Chinese dialogues.
\paragraph{Bilingual.} Under the bilingual setting, models are trained on bilingual dialogues (full training set), and in the testing phase, the trained models are expected to handle dialogues in both languages simultaneously without any language identifiers.
\paragraph{Cross-lingual.} This setting simulates the condition of lacking data in a certain language, and we study how to transfer the knowledge from a high resource language to a low resource language. Models have full access to the source language in this setting but limited access to the target language (10\%).
\section{Proposed Baselines}
Our proposed baselines are based on the recent state-of-the-art end-to-end ToD modeling approach \textit{MinTL}~\cite{lin2020mintl} and cross-lingual transfer approach MTL~\cite{liu2020attention}. We report the hyper-parameters and training details in the Appendix~\ref{sec:infra}.
\begin{figure}[t]
\begin{minipage}[c]{0.57\textwidth}
\centering
\includegraphics[width=0.8\textwidth]{img/model.png}
\end{minipage}
\begin{minipage}[c]{0.4\textwidth}
\caption{Model response generation workflow. Given the dialogue history $\mathcal{H}$, the knowledge $K$ (which can also be empty), and the dialogue state $B$, the $\mathrm{mSeq2Seq}$ 1) updates the dialogue state by generating $Lev$, and 2) generate a textual output and checks if it is an $\mathrm{API}$ or a $\mathrm{Response}$. If the output is an $\mathrm{API}$, (2a) the system queries the KB and updates the knowledge $K$; otherwise (2b) the $\mathrm{Response}$ is shown to the user. See this \href{https://github.com/HLTCHKUST/BiToD/blob/main/bitod.gif}{GIF} for more details.}
\label{fig:model}
\end{minipage}
\end{figure}
\paragraph{Notations.}
We define a dialogue $\mathcal{D}=\{U_1,S_1, \dots, U_T, S_T\}$ as an alternating set of utterances from user and systems. At turn $t$, we denote a dialogue history as $\mathcal{H}_t=\{U_{t-w},S_{t-w}, \dots, S_{t-1},U_t\}$, where $w$ is the context window size. We denote the dialogue state and knowledge state at turn $t$ as $B_t$ and $K_t$, respectively.
\subsection{ToD Modeling}
Figure \ref{fig:model} describes the workflow of our baseline model. We initialize the dialogue state $B_0$ and knowledge state $K_0$ as empty strings. At turn $t$, the input of our model is the current dialogue history $H_t$, previous dialogue state $B_{t-1}$ and knowledge state $K_{t-1}$. Similar to the text-to-text transfer learning approach~\cite{raffel2020exploring}, we add a prompt $P_B = ``Track Dialogue State:"$ to indicate the generation task. Then, a multilingual sequence-to-sequence ($\mathrm{mSeq2Seq}$) model takes the flattened input sequence and outputs the \textit{Levenshtein Belief Spans} ($Lev_t$)~\cite{lin2020mintl}:
\begin{align*}
(1): \mathrm{Lev}_t = \mathrm{mSeq2Seq}(P_B, \mathcal{H}_t, B_{t-1}, K_{t-1}).
\end{align*}
The $Lev_t$ is a text span that contains the information for updating the dialogue state from $B_{t-1}$ to $B_t$.
The updated dialogue state $B_t$ and a response generation prompt, $P_R = "Response:"$, are used as input.
Then, the model will either generate an API name (2a) when an API call is needed at the current turn, or a plain text response directly returned to the user (2b). If the model generates an API name, it is
\begin{align*}
(2a): \mathrm{API} =\mathrm{mSeq2Seq}(P_R, \mathcal{H}_t, B_{t}, K_{t-1}),
\end{align*}
the system will query the API with the constraints in the dialogue state and update the knowledge state $K_{t-1} \rightarrow K_t$. The updated knowledge state and API name are incorporated into the model to generate the next turn response generation.
\begin{align*}
(2b): \mathrm{Response} = \mathrm{mSeq2Seq}(P_R, \mathcal{H}_t, B_{t}, K_{t}, \mathrm{API}).
\end{align*}
All the aforementioned generation process are based on a single $\mathrm{mSeq2Seq}$, and we initialized our model with two pre-trained models, \textbf{mT5}~\cite{raffel2020exploring} and \textbf{mBART}~\cite{liu2020multilingual}.
\subsection{Cross-lingual Transfer}
Based on the modelling strategy mentioned above, we propose three baselines for the cross-lingual setting.
\paragraph{mSeq2seq.} Directly finetune the pre-trained mSeq2seq models like mBART and mT5 on the 10\% dialogue data in the target language.
\paragraph{Cross-lingual Pre-training (CPT).} First, pre-train the mBART and mT5 models on the source language, then finetune the models on the 10\% target language data.
\paragraph{Mixed-Language Pre-training (MLT).} To leverage the fact that our knowledge base contains the bilingual parallel information for most of the entities, we replace the entities in the source language data (both input sequence and output sequence) with their target language counterpart in our parallel knowledge base to generate the mixed-language training data. We first pre-train the mSeq2seq models with the generated mixed-language data, then finetune the models on the 10\% target language data.
\begin{table}[]
\resizebox{\linewidth}{!}{
\begin{tabular}{r|cccccccccc}
\toprule
\multicolumn{1}{l|}{} & \multicolumn{5}{c|}{\textit{English (EN)}} & \multicolumn{5}{c}{\textit{Chinese (ZH)}} \\ \midrule
\multicolumn{1}{l|}{\textit{\textbf{Models}}} & \textit{\textbf{TSR}} & \textit{\textbf{DSR}} & \textit{\textbf{API$_{\mathrm{Acc}}$}} & \textit{\textbf{BLEU}} & \multicolumn{1}{c|}{\textit{\textbf{JGA}}} & \textit{\textbf{TSR}} & \textit{\textbf{DSR}} & \textit{\textbf{API$_{\mathrm{Acc}}$}} & \textit{\textbf{BLEU}} & \textit{\textbf{JGA}} \\ \midrule
\multicolumn{11}{c}{\textit{\textbf{Monolingual}}} \\ \midrule
\textit{mBART} & 56.00 & 33.71 & 57.03 & 35.34 & \multicolumn{1}{c|}{67.36} & 56.82 & 29.35 & 71.89 & 20.06 & \textbf{72.18} \\
\textit{mT5} & 69.13 & 47.51 & 67.92 & 38.48 & \multicolumn{1}{c|}{69.19} & 53.77 & 31.09 & 63.25 & 19.03 & 67.35 \\ \midrule
\multicolumn{11}{c}{\textit{\textbf{Bilingual}}} \\ \midrule
\textit{mBART} & 42.45 & 17.87 & 65.35 & 28.76 & \multicolumn{1}{c|}{69.37} & 40.39 & 16.96 & 65.37 & 5.23 & 69.50 \\
\textit{mT5} & \textbf{71.18} & \textbf{51.13} & \textbf{71.87} & \textbf{40.71} & \multicolumn{1}{c|}{\textbf{72.16}} & \textbf{57.24} & \textbf{34.78} & \textbf{65.54} & \textbf{22.45} & 68.70 \\ \midrule
\multicolumn{11}{c}{\textit{\textbf{Cross-lingual}}} \\ \midrule
\multicolumn{1}{l|}{} & \multicolumn{5}{c|}{ZH $\rightarrow$ EN (10\%)} & \multicolumn{5}{c}{EN $\rightarrow$ ZH (10\%)} \\ \midrule
\multicolumn{1}{l|}{\textit{\textbf{Models}}} & \textit{\textbf{TSR}} & \textit{\textbf{DSR}} & \textit{\textbf{API$_{\mathrm{Acc}}$}} & \textit{\textbf{BLEU}} & \multicolumn{1}{c|}{\textit{\textbf{JGA}}} & \textit{\textbf{TSR}} & \textit{\textbf{DSR}} & \textit{\textbf{API$_{\mathrm{Acc}}$}} & \textit{\textbf{BLEU}} & \textit{\textbf{JGA}} \\ \midrule
\textit{mBART} & 1.11 & 0.23 & 0.60 & 3.17 & \multicolumn{1}{c|}{4.64} & 0.00 & 0.00 & 0.00 & 0.01 & 2.14 \\
\textit{+ CPT} & 36.19 & 16.06 & 41.51 & 22.50 & \multicolumn{1}{c|}{42.84} & 24.64 & 11.96 & 29.04 & 8.29 & 28.57 \\
\multicolumn{1}{r|}{\textit{+ MLT}} & 33.62 & 11.99 & 41.08 & 20.01 & \multicolumn{1}{c|}{55.39} & 44.71 & 21.96 & \textbf{54.87} & 14.19 & \textbf{60.71} \\ \midrule
\textit{mT5} & 6.78 & 1.36 & 17.75 & 10.35 & \multicolumn{1}{c|}{19.86} & 4.16 & 2.20 & 6.67 & 3.30 & 12.63 \\
\textit{+ CPT} & 44.94 & 24.66 & 47.60 & 29.53 & \multicolumn{1}{c|}{48.77} & 43.27 & 23.70 & 49.70 & 13.89 & 51.40 \\
\multicolumn{1}{r|}{\textit{+ MLT}} & \textbf{56.78} & \textbf{33.71} & \textbf{56.78} & \textbf{32.43} & \multicolumn{1}{c|}{\textbf{58.31}} & \textbf{49.20} & \textbf{27.17} & 50.55 & \textbf{14.44} & 55.05 \\ \bottomrule
\end{tabular}
}
\caption{Dialogue state tracking and end-to-end task completion results in monolingual, bilingual, and cross-lingual settings.}
\label{tab:main_results}
\end{table}
\section{Results \& Discussion}
The main results for DST and end-to-end task completion are reported in Table \ref{tab:main_results}. Note that the $API_{Acc}$ is highly correlated with the \textit{JGA} because the dialogue states contain constraints for issuing APIs. And, the \textit{DSR} is a more challenging metric compared to \textit{TSR} because the dialogue might contain 2-5 tasks.
\paragraph{Monolingual vs Bilingual.} Comparing the models that are trained under monolingual and bilingual setting, the latter can leverage more training data and handle tasks in both languages simultaneously without a language identifier. We observe that \textit{mT5} achieves better results in the bilingual settings, while \textit{mBART} performs better with monolingual training. The underlying reason might be the different pre-training strategies of the two mSeq2seq models. \textit{mBART} is pre-trained with language tokens in both the encoder and decoder, but in our bilingual setting, we do not provide any language information. Such a discrepancy does not exist in the \textit{mT5} model, as it is pre-trained without language tokens.
\paragraph{Cross-lingual.} We observe that it is difficult for the baseline models to converge with minimal training data (10\%) due to the complex ontology and diverse user goals. Interestingly, pre-training the mSeq2seq models on the source language improves both DST and task completion performance. Such results indicate the excellent cross-lingual transferability of multilingual language models. Furthermore, the mixed-language training strategy further improves the cross-lingual few shot performance, especially the \textit{JGA}, which suggests that the bilingual knowledge base can facilitate the cross-lingual knowledge transfer in the low resource scenario.
\paragraph{Limitations and Future Work.} The main limitation of this work is the low number of languages in the corpus due to the difficulty of collecting the knowledge base in languages other than English and Chinese in Hong Kong. in future work, we plan to extend the dataset to more languages including low resource languages in dialogue research (e.g., Indonesian), to better examine the cross-lingual transferability of end-to-end ToD systems. Another limitation is that the M2M data collection might not cover rare and unexpected user behaviours (e.g., non-collaborative dialogues), as dialogue simulators generate the dialogue outlines. However, we see {\bbfamily BiToD} as a necessary step for building robust multilingual ToD systems before tackling even more complex scenarios.
\section{Related Work}
Many datasets have been proposed in the past to support various assistant scenarios. In English, Wen et al. \cite{wen2016conditional} collected a single domain dataset with a Wizard-of-Oz (Woz) setup, which was latter extended to multi-domain by many follow-up works~\cite{budzianowski2018multiwoz,mosig2020star,moon2020situated,chen2021action}. Despite its effectiveness, Woz data collection method is expensive since two annotators need to be synchronized to conduct a conversation, and the other set of annotators need to annotate speech-act and dialogue states further. To reduce the annotation overhead (time and cost), Byrne et al. \cite{shah2018building} proposed a Machines Talking To Machines (M2M) self-chat annotations schema. Similarly, Rastongi et al.\cite{rastogi2020towards,byrne2019taskmaster} applied M2M to collect a large-scale schema-guided ToD dataset, and Kottur et al.\cite{kottur2021simmc} extended it to multimodal setting. In languages other than English, only a handful of datasets have been proposed. In Chinese, Zhu et al. ~\cite{zhu2020crosswoz}, and Quan et al.\cite{quan2020risawoz} proposed WoZ style datasets, and in German, the WMT 2020 Chat translated the dataset from Byrne et al. \cite{byrne2019taskmaster}. To the best of our knowledge, all the above-mentioned datasets are monolingual, thus making our BiToD dataset unique since it includes a bilingual setting and all the annotations needed for training an end-to-end task-oriented dialogue system. In the chit-chat setting, XPersona~\cite{lin2020xpersona} has a translation corpus in seven languages, but it is limited in the chit-chat domain. Finally, Razumovskaia et al.\cite{razumovskaia2021crossing} made an excellent summarization of the existing corpus for task-oriented dialogue systems, highlighting the need for multilingual benchmarks, like {\bbfamily BiToD}, for evaluating the cross-lingual transferability of end-to-end systems.
\section{Conclusion}
We present {\bbfamily BiToD}, the first bilingual multi-domain dataset for end-to-end task-oriented dialogue modeling. {\bbfamily BiToD} contains over 7k multi-domain dialogues (144k utterances) with a large and realistic knowledge base. It serves as an effective benchmark for evaluating bilingual ToD systems and cross-lingual transfer learning approaches. We provide state-of-the-art baselines under three evaluation settings (monolingual, bilingual and cross-lingual). The analysis of our baselines in different settings highlights 1) the effectiveness of training a bilingual ToD system compared to two independent monolingual ToD systems, and 2) the potential of leveraging a bilingual knowledge base and cross-lingual transfer learning to improve the system performance under low resource conditions.
\bibliographystyle{unsrt}
|
\section{Introduction}
\label{sec:intro}
With the rapid rise in user-generated web content, the scale and complexity of online hate have reached unprecedented levels in recent years.
ADL (Anti-Defamation League) conducted a nationally representative survey of Americans in December 2018 and the report shows that
over half (53\%) of Americans experienced some type of online harassment.\footnote{https://www.adl.org/onlineharassment} This number is higher than the 41\% reported to a comparable question asked in 2017 by the Pew Research Center~\cite{pew2017online}. To address the growing online hate,
a great deal of research has focused on automatic hate speech classification.
Most of the previous work focuses on binary classification~\cite{warner2012detecting, zhong2016content, nobata2016abusive, gao2017recognizing, qian2018leveraging} or coarse-grained multi-class classification~\cite{waseem2016hateful, badjatiya2017deep, davidson2017automated}.
\begin{figure}[t]
\centering
\includegraphics[width=0.49\textwidth]{intro2.jpg}
\caption{An illustration of our proposed task. $hg_i$: the $i$th hate group. The model is trained on a sequence of sub-datasets, split by their hate ideologies, e.g., anti-Muslim and Kuklux Klan. The task on each sub-dataset is to identify the hate group given the tweet.}
\label{fig:intro}
\end{figure}
\citet{qian2018hierarchical} argue that fine-grained classification is necessary for fine-grained hate speech analysis. The Southern Poverty Law Center (SPLC) monitors hate groups throughout the United States by a variety of methodologies to determine the activities of groups and individuals, including reviewing hate group publications.\footnote{https://www.splcenter.org/fighting-hate/extremist-files/ideology} Therefore, instead of differentiating normal posts from the other offensive ones, ~\citet{qian2018hierarchical} propose a more fine-grained hate speech classification task that attributes hate groups to individual tweets.
However, a common limitation of all the research mentioned above is that they assume the dataset to be static and train the classifiers on each isolated dataset, i.e., isolate learning, ignoring the rapid increase of the amount of data in social media and the rapid change of the hot topic.
A report from L1ght\footnote{https://l1ght.com/Toxicity\_during\_coronavirus\_Report-L1ght.pdf}, a company that specializes in measuring online toxicity, suggests that amid the growing threat of the coronavirus, there has been a 900\% growth in hate speech towards China and Chinese people on Twitter since February 2020.
As a result of the rapid change of social media content, the hate speech classifiers are required to be able to continuously learn and accumulate knowledge from a stream of data, i.e., lifelong learning. Learning on each portion of the data is considered as a task, so a stream of tasks are joined to be trained sequentially. In this work, we propose a novel lifelong fine-grained hate speech classification task, as illustrated in Figure~\ref{fig:intro}.
The models trained by isolate learning tend to face catastrophic forgetting~\cite{mccloskey1989catastrophic,ratcliff1990connectionist,mcclelland1995there,french1999catastrophic} due to a non-stationary data distribution in lifelong learning. To address this problem, an extensive body of work has been proposed for various lifelong learning tasks. However, our experiments show that the commonly-used lifelong learning methods still exhibit catastrophic forgetting in our proposed tasks. One important difference between the Twitter hate group dataset and the other image datasets commonly used in lifelong learning study is that the similarity among the different tasks is unstable and relatively low, as indicated by the low average Jaccard Indexes of the topic words in Table~\ref{tab:data}. To alleviate this problem, we introduce VRL to distill the knowledge from each task into a latent variable distribution.
We also augment the model with a memory module and adapt the clustering algorithm, LB-SOINN, to select the most important samples from the training dataset of each task. Our implementation is publicly available\footnote{https://www.aclweb.org/anthology/2021.naacl-main.183/}.
Our contributions are three-fold:
\begin{itemize}
\item This is the first paper on lifelong learning of fine-grained hate speech classification.
\item We propose a novel method that utilizes VRL along with an LB-SOINN memory module to alleviate catastrophic forgetting resulted from a severe change of data distribution.
\item Experimental results show that our proposed method outperforms the state-of-the-art significantly on the average F1 scores.
\end{itemize}
\section{Related Work}
\label{sec:related}
Most research on lifelong learning alleviates catastrophic forgetting in the following three directions.
\noindent{\bf Regularization-based Methods:}
These methods impose constraints on the weight update. The goal of the constraints is to minimize deviation from trained weights when training on a new task. The constraints are generally modeled by additional regularization terms~\cite{ewc,Zenke2017,pathnet2017,liu2018rotate,ritter2018online}.
Elastic Weight Consolidation (EWC)~\cite{ewc} alleviates catastrophic forgetting by slowing down learning on the model parameters which are important to the previous task. The importance of the parameters is estimated by the Fisher information matrix.
Instead of the Fisher information matrix, PathNet~\cite{pathnet2017} uses agents embedded in the neural network to determine which parameters of the neural network can be reused for new tasks and the task-relevant pathways are frozen during training on new tasks.
\begin{table}[t!]
\small
\centering
\begin{tabular}{|l|c|c|}
\hline
\textbf{Ideology} & \textbf{Avg. JI} & \textbf{Keywords} \\\hline
Christian Identity &0.019 &Jesus, Yahuwshua\\\hline
Radical Tr. Catholic &0.031 &catholic, remnant\\\hline
Neo Confederate &0.039 &southern, Free Dixie \\\hline
Anti Semitism &0.047 &Israel, Trump\\\hline
Anti Catholic &0.049 &Texe Marrs, truth\\\hline
Hate Music &0.049 &death, radio\\\hline
Anti Muslim &0.064 &Muslim, Islam\\\hline
Black Separatist &0.071 &black, panther\\\hline
Racist Skinhead &0.074 &shirt, white\\\hline
Anti Immigration &0.075 &immigration, border\\\hline
Holocaust Identity &0.078 &Jewish, Trump\\\hline
Neo Nazi &0.091 &Hitler, white\\\hline
Kuklux Klan &0.100 &ni**a, f**king\\\hline
Anti LGBTQ &0.100 &family, marriage\\\hline
White Nationalist &0.105 &white, America\\\hline
\end{tabular}
\caption{Information about the 15 hate ideologies. Tr.: Traditional. Avg JI: the average of the Jaccard Index between the topic words of one ideology and those of another ideology. The topic words are extracted by Latent Dirichlet Allocation (LDA)~\cite{blei2003latent}. The top 2 most frequent topic words are selected as keywords.}
\label{tab:data}
\end{table}
\noindent{\bf Architecture-based Methods:}
The main idea of this approach is to change architectural properties to dynamically accommodating new tasks, such as assigning a dedicated capacity inside a model for each task.
~\citet{progressive2016} propose Progressive Neural Networks, where the model architecture is expanded by allocating a new column of neural network for each new task.
~\citet{part2016incremental, part2017incremental} combine Convolutional Neural Network with LB-SOINN for incremental online learning of object classes. Although they also use LB-SOINN in their work, the usage of LB-SOINN in this work is completely different. They use LB-SOINN to predict object class while our proposed method adapts the original LB-SOINN to calculate the importance of the training samples without making any prediction on the class.
A problem with the methods in this category is that the available computational resources are limited in practice. As a result, the model expansion will be prohibited when the number of tasks increases to a certain degree.
\noindent{\bf Data-based Methods:}
These methods alleviate catastrophic forgetting by utilizing a memory module, which either stores a small number of real samples from previous tasks or distills knowledge from previous tasks.
The main feature of Gradient Episodic Memory (GEM)~\cite{gem} is the episodic memory, storing a subset of the samples from the observed tasks. GEM computes the losses on the episodic memories and treats them as inequality constraints, avoiding them to increase. Averaged GEM~\cite{agem} is a more efficient version of GEM. ~\citet{epidosicmemory} propose a lifelong language learning model using a key-value memory module for sparse experience replay and local adaptation. ~\citet{lamol} formulate lifelong language learning as a language modeling task and replay the generated pseudo-samples of previous tasks during training.
There are also studies combining multiples methods above.~\citet{xia2017distantly}
combine the architecture-based method and the data-based method. ~\citet{wang2019sentence} combine the regularization method and the data-based method for lifelong learning on relation extraction. Our proposed method is also a combination of the regularization method and the data-based method but in a different way.
\section{Task Description}
\label{sec:desc}
We use the dataset as in~\citet{qian2018hierarchical}, where the tweet handles are collected based on the hate groups identified by SPLC. SPLC categorizes these hate groups according to their hate ideologies. For each hate ideology, the top three Twitter handles are selected in terms of the number of followers.
The dataset includes all the content (tweets, retweets, and replies) posted with each Twitter account from the group's inception date, as early as 2009, until 2017. Altogether, the dataset consists of 42 hate groups from 15 different ideologies. Table~\ref{tab:data} shows the 15 ideologies. Each instance in the dataset is a text tuple of (tweet, hate group name, hate ideology).
We separate the dataset by ideology. The reason is that various existing hate speech datasets collect data using keywords or hashtags~\cite{waseem2016hateful,davidson2017automated,golbeck2017large}, which have a strong relationship with hate ideologies or topics. We also observe that the hot spots of society can lead to a significant shift of major hate speech topics or the emergence of new hate ideologies on social media as mentioned in section~\ref{sec:intro}, indicating that the expansion of the hate speech dataset may be accompanied by the emergence of new hate ideologies.
Therefore, we separate the collected data into a sequence of 15 subsets according to their ideologies
and sort them by the date of the first tweet post in each subset, from the earliest to the latest.
The task on each subset is to identify the hate group given the tweet text. ~\citet{qian2018hierarchical} propose a hierarchical Conditional Variational Autoencoder model for the fine-grained hate speech classification task. The architecture and the training process of their model require the number of classes to be pre-defined. However, we do not pre-define the number of classes in our task since such kind of information is not available in the real-world application of lifelong learning. The model should be able to incorporate emerging hate groups at any time of training. In order to satisfy this condition, we formulate the task of identifying the group as a ranking task, instead of a classification task. For each tweet, we provide the model with a set of candidate groups, consisting of all the previously seen hate groups, including the ground truth group. The model takes each combination of the tweet and the candidate group as input and outputs a score. The corresponding loss function is:
\begin{equation}
\mathcal{L}_r\!=\!\!\sum_{(x,y_s)\in D}\sum_{y_{i}\in{Y\backslash\{y_{s}\}}}\!\!\!\!\!h(f_{\theta}(x, y_{s})-f_\theta(x, y_{i}))
\label{eq:lossmargin}
\end{equation}
where $x$ is the tweet text, $y_s$ is the ground truth group of $x$. $Y$ is candidate group set of $x$, which consists of all the seen hate groups until $x$ is observed by the model, including the ground truth group $y_s$ of $x$, so $y_i \in {Y\backslash\{y_{s}\}}$ is the negative candidate group of $x$. $f_{\theta}$ is the scoring model parameterized by $\theta$. $h(a)=\max(0, m-a)$, $m$ is the chosen margin.
Same as in other lifelong learning studies, we consider learning on each of the hate ideologies in the sequence as a task, so we have a sequence of 15 tasks.
As mentioned in section~\ref{sec:intro}, the similarity among our tasks is unstable and relatively low. Therefore, when the model is continuously trained on the tasks, it may encounter a sudden change of vocabulary, topic, and input data distribution. This makes our tasks more challenging compared to the other lifelong learning tasks because the abrupt change can make the catastrophic forgetting problem more severe. This is also the reason that some techniques achieving significant improvement in the image classification tasks do not perform well on our task (see section~\ref{sec:experiment}).
\begin{figure*}[t]
\centering
\includegraphics[width=0.68\textwidth]{model2.jpg}
\caption{An illustration of our method. The dotted arrows indicate the computation of the loss. The light-colored dashed arrows illustrate the update of the memory module. Note that the layers in the rounded rectangle share parameter weight. There is only one encoder for the group input, followed by two linear layers. We make a copy of it in the figure just for a clear illustration of loss computation. $\hat{x}$: the reconstructed tweet input. $s_1$, $s_2$: scores of ($x$, $y_s$) and ($x$, $y_i$) separately. $\mu_z^*$ and $\Sigma_z^*$ are the previously memorized distribution on the latent variable of $x$. $L_{rec}$ is the reconstruction loss, which is the first term in equation~\ref{equ:lossvae}. Please refer to section~\ref{sec:approach} for the meaning of other variables in the figure.}
\label{fig:model}
\end{figure*}
\section{Our Approach}
\label{sec:approach}
As mentioned in section~\ref{sec:related}, one way to alleviate catastrophic forgetting is to use a memory module, storing a small number of real samples from previous tasks and a simple way to utilize the memorized samples is to replay the memory when training on a new task, such as mixing them with the training samples from the current task.
The idea behind this approach is that the memorized samples should reflect the data distribution so that the replay of the memory can help the model make invariant predictions on the samples of the previous tasks. However, this approach may not work well when the size of the memory is small. The reason is that when there is only a small amount of data memorized, the memory is not able to reflect the data distribution of the previous task and thus the model can easily overfit on the memorized samples instead of generalizing to all the samples in the previous task.
We address this problem from two aspects. First, since the memory size is limited, it is beneficial to select the most representative training samples in the previous tasks to memorize. Second, simply storing the real training samples in the memory may not be sufficient to represent the knowledge of the previous tasks, so we need a better way to distill knowledge from the observed samples along with a method to utilize it when training on a new task. We combine two techniques: Variational Representation Learning (VRL) and Load-Balancing Self-Organizing Incremental Neural Network (LB-SOINN) to achieve these goals. We propose a supervised version of LB-SOINN to select the most important training samples in the current task. VRL not only distills the knowledge from the current training task but also provides an appropriate hidden representation as input for the LB-SOINN, so we introduce VRL first.
\subsection{Variational Representation Learning}
\label{subsec:vae}
The distilled knowledge of previous tasks can take various forms, but the key point is that it should be related to the data distribution of the corresponding task so that it can be utilized to alleviate catastrophic forgetting. Inspired by the Variational Autoencoder (VAE)~\cite{kingma2013auto}, we consider the distribution of the hidden representation of the input data as the distilled knowledge.
The original VAE model is proposed for data generation, so the objective of the original VAE is:
\begin{equation}
Obj=\sum_{x\in X}\log p(x)
\end{equation}
\begin{equation}
p(x)=\int_{z}p(x|z)p(z)dz
\end{equation}
$z$ is the latent variable, i.e., the hidden representation of the input. Since the integration over $z$ is intractable, we instead try to maximize the corresponding evidence lower bound (ELBO) and the corresponding loss function is as follows:
\begin{equation}
\begin{aligned}
\mathcal{L}_{vae}=\sum_{x\in X}&E_{z\sim p_\alpha(z|x)}[-\log p_\varphi(x|z)]+\\
&D_{KL}[q_\alpha(z|x)||p_\beta(z)]
\label{equ:lossvae}
\end{aligned}
\end{equation}
$p(x|z)$, $q(z|x)$, and $p(z)$ are the likelihood distribution, posterior distribution, and prior distribution. $\alpha$,$\varphi$, and $\beta$ indicate parameterization. The loss function can be separated into two parts. The first part $E[-\log p(x|z)]$ is the reconstruction loss, trying to reconstruct the input text from the latent variable. It pushes $z$ to reserve as much information of the input as possible. This is consistent with our goal to learn the knowledge of the data distribution.
The second part is $D_{KL}[q(z|x)||p(z)]$, where $D_{KL}$ is the Kullback–Leibler (KL) divergence. Minimizing it pushes the posterior and the prior distributions to be close to each other.
By assuming the posterior $p(z|x)$ to be a multivariate Gaussian distribution $\mathcal{N}(\mu_z, \Sigma_z)$, the latent variable $z$ is sampled from $\mathcal{N}(\mu_z, \Sigma_z)$.
In the original VAE, $p(z)$ is chosen to be a simple Gaussian distribution $\mathcal{N}(0, 1)$. However, this is over-simplified in our task because different from the unsupervised generation task of the original VAE, our ranking task is supervised. Our task not only requires $z$ to contain information of the tweet text itself but also requires it to indicate the group information of the tweet. In other words, the distilled distribution should be conditioned on both the tweet and its group label to reflect the data distribution in a supervised task. Setting the prior to be the same for all the hate groups pushes $z$ or the distribution of $z$ to ignore the label information. Instead, the prior should be different for each hate group, so we replace $p(z)$ with $p(u|y_s)$, where $y_s$ is the group label of $x$ and $u$ is the latent variable. $p(u|y_s)$ is assumed to be a multivariate Gaussian distribution $\mathcal{N}(\mu_u, \Sigma_u)$. Note that the replacement itself can not guarantee $p(u|y_s)$ to be different for each hate group because the loss function in equation~\ref{equ:lossvae} does not push $p(u|y_s)$ to satisfy this condition. However, the ranking loss function~\ref{eq:lossmargin} fills in the gap. Therefore, our loss function on the current training task is a combination of these two.
\begin{equation}
\begin{aligned}
\mathcal{L}_{cur}\!=\!\!\!\!\!\!\sum_{(x,y_s)\in D} & \sum_{y_{i}\in{Y\backslash\{y_{s}\}}}\!\!\!\!\!h(f_{\theta}(x, y_{s})\!-\!f_\theta(x, y_{i}))\\
&\!\!\!+\!E_{z\sim p_\alpha(z|x)}[-\log p_\varphi(x|z)]\\
&\!\!\!+D_{KL}[q_\alpha(z|x)||p_\beta(u|y_{s})]
\end{aligned}
\end{equation}
The right part of Figure~\ref{fig:model} illustrates the computation process of VRL.
\subsection{LB-SOINN Memory Module}
\label{subsec:memory}
VRL provides a way to summarize knowledge into latent variable distributions. However, we still need a method to utilize the learned distribution to alleviate catastrophic forgetting. We do this by incorporating a memory module $D_{mem}$ to store a small subset of important training samples along with their latent variable distributions, so each sample stored in the memory is a tuple of $(x, y_z, q_{\alpha^\prime}(z|x))$. Here $q_{\alpha^\prime}(z|x)$ is the distribution computed when the model completes training on the task that $(x, y_z)$ belongs to. The memorized samples are taken as anchor points when training on a new task. We introduce a memory KL divergence loss to push $q_\alpha(z|x)$ computed when training on a new task to be close to the memorized distribution $q_{\alpha^\prime}(z|x))$. Therefore, the complete loss function is:
\begin{equation}
\begin{aligned}
&\mathcal{L}=\mathcal{L}_{cur}+ D_{KLmem}\\
&=\mathcal{L}_{cur}+\!\!\!\!\!\!\!\sum_{(x,y_s)\in D_{mem}}\!\!\!\!\!\!\!\!D_{KL}[q_\alpha(z|x)||q_{\alpha^\prime}(z|x))]
\end{aligned}
\label{eq:lossfinal}
\end{equation}
Since the size of the memory is limited,
we introduce a supervised version of LB-SOINN to select the most important training samples in the current task. The input for the LB-SOINN is the hidden representation of the tweet text, which is $z$ in the case of Variational Representation Learning (see Figure~\ref{fig:model}). We refer readers to~\citet{zhang2013load} for the detailed explanation of LB-SOINN. The original LB-SOINN is an unsupervised clustering algorithm that clusters unlabeled data by topology learning. We utilize the topology learning of LB-SOINN instead of clustering since our task is supervised. Therefore, we make the following adjustments to the original LB-SOINN.
\noindent1) The criteria to add a new node: Add a new node to the node set if one of the following condition is satisfied:
a) The distance between the input and the winner is larger than the winner's threshold.
b) The distance between the input and the second winner is larger than the second winner's threshold.
c) The label of the input sample is not the same as the label of the winner.
\noindent2) Build connections between nodes: Connect the two nodes with an edge only if the winner and the second winner belong to the same class.
\noindent3) We disable the removal of edges whose ages are greater than a predefined parameter. We disable the deleting of nodes and the algorithm of updating the subclass labels of every node. The node label is the label of the instances assigned to it. Our adjusted algorithm guarantees that each node will only be assigned the samples from one class.
LB-SOINN keeps track of the density of each node, which is defined as the mean accumulated points of a node. A node gets points when there is an input sample assigned to it. If the mean distance of the node from its neighbors is large, we give low points to the node. In contrast, if the mean distance of the node from its neighbors is small, we give high points to the node. Therefore, the density of the node reflects the number of nodes close to it and also the number of samples assigned to it. We take the density of the node as a measurement of the importance of the samples assigned to the node. After the LB-SOINN finishes training on the samples from the current task, we sort the samples according to the density of the node they are assigned to and the top $K$ samples are selected to write to the memory. We divide the memory equally for each of the previous tasks, so $K=M/t$, where $M$ is the total memory size and $t$ is the number of observed tasks, including the current task. The old memory consists of samples from the previous $t-1$ tasks and each task keeps $M/(t-1)$ samples in the old memory. For each of the $t-1$ tasks, the $M/(t-1)-M/t$ samples with the lowest node densities are deleted, resulting in $K$ empty slots in the memory, which is then rewritten by the selected $K$ samples in the current task.
\begin {table*}[t!]
\centering
\small
\begin{tabular}{|l|c|c|c|c|c|c|}
\hline
Number of observed tasks & \multicolumn{2}{|c|}{t=5} &\multicolumn{2}{|c|}{t=10}& \multicolumn{2}{|c|}{t=15}\\
\hline
Avg F1 score (\%) & Macro & Micro & Macro & Micro & Macro & Micro \\
\hline
\hline
Multitask &15.26 &67.07 &5.05 &37.20 &3.57 &38.61 \\
\hline
\hline
Fine-tuning &6.02 &16.44 &4.35 &5.77 &3.96 &6.18 \\
\hline
Fine-tuning + BERT &6.02 &16.44 &4.06 &5.45 &3.03 &5.80 \\
\hline
Fine-tuning + RMR &11.15 &44.40 &2.56 &15.77 &3.51 &15.19 \\
\hline
EWC &8.57 &20.42 &2.42 &6.81 &1.95 &7.27 \\
\hline
GEM &\textbf{13.04} &30.95 &3.07 &12.51 &2.70 &15.07 \\
\hline
Ours &12.61 &\textbf{49.75} &\textbf{6.96} &\textbf{47.30} &\textbf{5.13} &\textbf{44.62} \\
\hline
\end{tabular}
\caption{Experimental results. RMR: random memory replay. The best results are in bold. }
\label{tab:results}
\end{table*}
\begin{table*}[t!]
\centering
\small
\begin{tabular}{|l|c|c|c|c|c|c|}
\hline
Number of observed tasks & \multicolumn{2}{|c|}{t=5} &\multicolumn{2}{|c|}{t=10}& \multicolumn{2}{|c|}{t=15}\\
\hline
Avg F1 score (\%) & Macro & Micro & Macro & Micro & Macro & Micro \\
\hline
Full Model &12.61 &{49.75} &\textbf{6.96} &\textbf{47.30} &{5.13} &\textbf{44.62} \\
\hline
w/o $D_{KLmem}$ &\textbf{15.00} &\textbf{58.64} &4.21 &36.36 &3.72 &40.87 \\
\hline
w/o VRL &11.05 &35.03 &4.53 &13.69 &3.65 &11.28 \\
\hline
w/o LB-SOINN &13.01 &50.99 &6.15 &44.42 &\textbf{5.59} &30.91 \\
\hline
\end{tabular}
\caption{Ablation study. w/o $D_{KLmem}$: $D_{KLmem}$ in the equation~\ref{eq:lossfinal} is removed. w/o VRL: VRL is replaced by the model used in the fine-tuning setting, i.e., fine-tuning + LB-SOINN memory replay. w/o LB-SOINN: LB-SOINN memory replay is replaced by random memory replay, i.e., VRL + RMR. The best results are in bold. }
\label{tab:ablation}
\end{table*}
\section{Experiments}
\label{sec:experiment}
\subsection{Experimental Settings}
For each task, we randomly sample 5000 tweets from the 80\% of the collected data for training, 10\% of the collected data for testing, and the rest 10\% for development. We allow the model to make more than one pass over the training samples in the current task or the current memory during training. We use average macro F1 score and average micro F1 score for evaluation.
\begin{equation}
\mbox{Average F1:} AvgF1(t)=\frac{1}{t}\sum_{i=1}^{t} F1_{t,i}
\end{equation}
where $F1_{t,i}$ is the F1 score, either macro F1 or micro F1, achieved by the model on the $i$th task after being trained on the $t$th task. The larger this metric, the better the model.
We compare our methods with the following methods:
\noindent{\bf Fine-tuning:}
The model contains two bidirectional LSTM encoders~\cite{hochreiter1997long, zhou2016attention, liu2016learning} to encode the tweet and the group separately. The score of the group is calculated as the cosine distance between the hidden state of the tweet encoder and that of the group encoder. This model is also the backbone model of all the methods described below, except Fine-tuning + BERT.
The model is directly fine-tuned on the stream of tasks, one after another, by the ranking loss function in~\ref{eq:lossmargin}.
\noindent{\bf Fine-tuning+BERT:} The training framework is the same as above, but each encoder is replaced by a pre-trained BERT model~\cite{devlin2019bert} followed by a linear layer.
The linear layers are fine-tuned during training.
\noindent{\bf Fine-tuning+RMR (Random Memory Replay):} We augment the fine-tuning method with an additional memory module. Same as in section~\ref{subsec:memory}, the memory is divided equally for each task, but instead of using LB-SOINN, the $K$ samples are randomly sampled from the current training data and then rewrite $K$ random slots in the old memory.
\noindent{\bf EWC:} EWC is a regularization-based method, adding a penalty term $\sum_i\frac{\lambda}{2}F_i(\theta_i-\theta^*_{i})^2$ to the ranking loss function~\ref{eq:lossmargin}. $F_i$ is the diagonal of the Fisher information matrix $F$, $\theta$ is the model parameter, and $i$ labels each parameter. $\theta^*$ is the model parameter when the model finishes training on the previous task. $\lambda$ is set to 2e6 in our experiments.
\noindent{\bf GEM:} We use the episodic memory in the original paper: the memory is populated with $m$ random samples from each task. $m$ is a predefined size of the episodic memory. We set $m=100$ in our experiments, so each task can add 100 tweets to the memory. By the end of the 15 tasks, the total memory of GEM contains 1500 tweets.
\noindent{\bf Multitask Learning:} The tasks are trained simultaneously. We mix the training data from multiple tasks to train the model. This setting does not follow the lifelong learning setting where the tasks are trained sequentially. We add this setting in our experiments to show the potential room for improvement concerning each lifelong learning method.
We do not compare our method with Support Vector Machine~\cite{suykens1999least} or Logistic Regression, because they require the number of classes to be fixed and to be known in advance, which is unrealistic in our tasks. We also do not compare our method with~\citet{qian2018hierarchical} since the latter also has this requirement, as mentioned in section~\ref{sec:desc}. Adapting their method for the lifelong learning setting requires modifying both the model architecture and the training algorithm, which is beyond the scope of this paper.
In all our experiments, we use 1-layer bi-LSTM as encoders except the fine-tuning + BERT setting and we use cosine distance to measure similarity. The input of the group encoder is the concatenation of the group name and its hate ideology.
We use 1-layer bidirectional GRU~\cite{cho2014learning} as the decoder in VRL. The hidden size of the encoders and the decoders is 64. The latent variable size in VRL is 128. We use 300-dimensional randomly initialized word embeddings. All the neural networks are optimized by Adam optimizer with the learning rate 1e-4. The batch size is 64. The loss margin $m=0.5$. The maximum number of training epochs for each task is set to 20. For LB-SOINN, $\lambda\!\!=\!\!1000$, $\eta\!\!=\!\!1.04$.
The memory size is limited to 1000 tweets for all the methods using a memory module except GEM. We do not set episodic memory size for each task as GEM because for lifelong hate speech classification, the number of tasks keeps increasing in the real world, and assuming unlimited total memory is unrealistic.
\subsection{Experimental Results}
The experimental results are shown in Table~\ref{tab:results}. We report the performance of each method after the model finishes training on the first 5 tasks, first 10 tasks, and all the 15 tasks. The average macro-F1 score is much lower than the average micro-F1 score due to the imbalanced data of each task.
The large performance gap between the multitask training and fine-tuning shows that there exists severe catastrophic forgetting and that the low average F1 scores in the fine-tuning setting are not due to the model capacity. Replacing the bi-LSTM encoder with the pre-trained BERT encoder does not improve the performance
This reconfirms that the low scores result from catastrophic forgetting, not model capacity. Actually fine-tuning and fine-tuning with BERT achieves the same average F1 scores at $t=5$ because both models completely forget the previous tasks after converging on the fifth task, so both models achieve the same F1 scores on the testing data of the fifth task while achieving 0 scores on the previous four tasks. Due to the large model capacity of BERT, fine-tuning with BERT tends to overfit on the training data more seriously, leading to slight performance decline at $t=10$ and $t=15$ compared to using bi-LSTM encoders. Since model capacity is not the key factor to solve catastrophic forgetting, we simply use bi-LSTM as encoders in our model instead of BERT, considering the computational cost.
Adding RMR to the fine-tuning setting achieves significant performance improvement, even better than EWC or GEM. This is related to the characteristic of our tasks mentioned at the end of section~\ref{sec:desc}. EWC remembers previous tasks by slowing down the update of the model parameters important to them, which is more suitable for the sequence of tasks that are similar to each other. However, significant changes in vocabulary, topic, or input data distribution are very common in our sequence of tasks, making memory replay more efficient than EWC. The performance of GEM during the second half of the training is close to that of fine-tuning with RMR, but there exists a gap in the first half. The reason is that GEM sets an episodic memory for each task, of which the size is 100 in our experiments, so before the 10th task in the sequence, the size of the total memory available for GEM is less than that of the memory module used in the fine-tuning with RMR setting.
Although RMR improves the performance, the average F1 scores still drop quickly when the number of tasks increases. In the late stage of sequential training, each task can only keep dozens of samples in the memory and the model is not able to generalize well based on the memory. Our method solves this problem by combining VRL and LB-SOINN memory replay. The performance of our model is better and more stable than the other methods when the number of tasks increases. Our method achieves higher scores than multitask training in the last four columns of Table~\ref{tab:ablation} because learning on one task is easier than learning on a mix of tasks simultaneously. Every model in our sequential training experiments can easily achieve high F1 scores on the current task, making a large contribution to the average F1 scores. However, when doing multitask training, the model loses this benefit.
To investigate the effect of our method, we conduct the ablation study as shown in Table~\ref{tab:ablation}.
Removing $D_{KLmem}$ from the final loss function in equation~\ref{eq:lossfinal} does not lower the performance when the number of observed tasks is small ($t\!\!=\!\!5$) because each task can store a few hundreds of samples in the memory at the early stage of sequential training, which is sufficient for the model to learn the previous tasks. However, when the number of tasks increases, $D_{KLmem}$ shows its effect on alleviating catastrophic forgetting.
Fine-tuning+LB-SOINN (Table~\ref{tab:ablation}) does not perform as well as fine-tuning+RMR (Table~\ref{tab:results}), while VRL+LB-SOINN (i.e., full model) performs better than VRL+RMR (Table~\ref{tab:ablation}). The reason lies in the input for LB-SOINN
Compared to the hidden representations spread evenly in the hidden space, the hidden representations which are well-organized in different group clusters make it easier for LB-SOINN to learn a reasonable topology structure of the training samples. VRL achieves this by explicitly pushing the hidden representation of tweets to follow a learned multivariate Gaussian distribution unique to each group. On the other hand, directly using the hidden state of the tweet encoder does not exhibit such kind of characteristics. VRL not only distills task knowledge but also provides an appropriate input for LB-SOINN, as stated in section~\ref{sec:approach}.
\begin{figure}[t]
\centering
\includegraphics[width=0.4\textwidth]{erroranalysis.jpg}
\caption{The testing results of the first 5 tasks in the sequence when our model is trained on the first 10 tasks.}
\label{fig:error}
\end{figure}
\subsection{Error Analysis}
Although our model achieves significant improvement over the baseline methods, we observe that our method does not perform well on the first task. As shown in Figure~\ref{fig:error}, there exists a large gap between the performance on the first task and the other tasks, and the micro-F1 score on the first task quickly drops to almost 0 when the number of observed tasks increases. We find the same results after we change the order of tasks in the sequence, so this is not the result of the task difficulty but is the result of our method.
We find this problem is due to the reconstruction loss, which is the first part in equation~\ref{equ:lossvae}.
The model observes a very limited number of tweets when training on the first task, making it difficult to learn the language model and reconstruct the tweet. As a result, the tweet representation learned on the first task may not contain the information we require, resulting in a large performance gap. When the number of observed tasks increases, this problem goes away quickly. We anticipate pre-training the VAE in our model (the left branch in Figure~\ref{fig:model}) on a large Twitter corpus can alleviate this problem at the beginning of training.
\section{Conclusion}
In this paper, we introduce the lifelong hate speech classification task and propose to use the VRL and LB-SOINN memory module to alleviate catastrophic forgetting.
Our proposed method has the potential to benefit other lifelong learning tasks where the similarity between the contiguous tasks can be low. We make our implementation freely available to facilitate more application and investigation of our method in the future\footnote{https://www.aclweb.org/anthology/2021.naacl-main.183/}.
\bibliographystyle{acl_natbib}
|
\section{Introduction}
Roughly speaking, the Carath\'eodory's procedure in constructing the Lebesgue measure in $\mathbb{R}$ is the following: first, it is built a outer measure on $\mathbb{R}$ from the length of the closed intervals of $\mathbb{R}$. Once the outer measure is constructed, the next step is to define the Lebesgue measurable sets. The collection of measurable subsets of $\mathbb{R}$ turns out to be a $\sigma$-algebra. Then the Lebesgue measure is defined as the outer measure restricted on the $\sigma$-algebra of Lebesgue measurable subsets of $\mathbb{R}$. Such a measure, say $m$, satisfies
\[
m \left( \bigcup_j E_j \right) = \sum_j m(E_j)
\]
for every countable collection $\{ E_j \}$ of pairwise disjoint measurable sets. This pro-perty of $m$ is called countable additivity
(see, e.g. \cite{zygmund}).
The purpose of this note is to construct a \textit{multiplicative measure} on the positive real axis $\mathbb{R}_{>0} = (0, +\infty)$. For them, we adapt the Carath\'eodory's procedure to our new setting. We consider the positive real axis $\mathbb{R}_{>0}$ with the usual topology and define the "length" of each closed interval $I=[a, b]$ in $\mathbb{R}_{>0}$ by $$\ell(I) = b \cdot a^{-1}.$$ Then, to apply the
Carath\'eodory's idea with these elements, we will obtain a measure $\mu$ and a $\sigma$-algebra $\mathcal{M}$ in $\mathbb{R}_{>0}$ with the following properties:
\[
\mu(\emptyset) = 1, \,\,\,\,\,\,\, \mu((0,1)) = +\infty, \,\,\,\,\,\,\, \mu(\mathbb{R}_{>0}) = +\infty,
\]
\[
\mu \left( \bigcup_j E_j \right) = \prod_j \mu(E_j)
\]
for every countable collection $\{ E_j \}$ of pairwise disjoint sets of $\mathcal{M}$. We call to this property:
\textit{countable multiplicativity}. By this reason, we say that $\mu$ is a \textit{multiplicative measure}.
Once constructed the measure $\mu$ and the $\sigma$-algebra $\mathcal{M}$, we will show that $\mathcal{M}$ coincides with the $\sigma$-algebra of the Lebesgue measurable subsets of $\mathbb{R}_{>0}$ and $\mu(E) = \lambda(E)$ for each $E \in \mathcal{M}$,
where
\[
\lambda(E) = \exp\left(\int_{E} \frac{1}{x} \, dx \right).
\]
\
In Section 2 we give some preliminaries about single and double infinite products of positive real numbers which are necessary to our construction. The multiplicative measure $\mu$ is constructed in Section 3. Finally, in Section 4, we show that $\mu(E) = \lambda(E)$ for each $E \in \mathcal{M}$.
\
{\bf Notation.} The positive real axis will be denoted by $\mathbb{R}_{>0} = (0, +\infty)$ and $\mathbb{R}_{\geq 1} = [1, +\infty)$.
\section{Preliminaries}
In this section we present some basic facts on certain single and double infinite products of real numbers (see, e.g. \cite{apostol}).
\subsection{Single infinite products} Since we will consider only infinite products with positive factors, the following
simplified definition will be adopted.
\
Given a sequence $\{ a_j \}$ of positive real numbers, let
\[
p_N = \prod_{j=1}^{N} a_j = a_1 \cdot a_2 \cdot \cdot \cdot a_N.
\]
The number $p_N$ is called the \textit{$N$th partial product} of the sequence $\{ a_j \}$.
\begin{definition} We say that an infinite product of positive real numbers $\displaystyle{\prod_{j=1}^{+\infty}} a_j$ converges if there exists a number $p >0$ such that
\[
\lim_{N \to +\infty} p_N = p.
\]
In this case, $p$ is called the value of the infinite product and we write $p = \displaystyle{\prod_{j=1}^{+\infty}} a_j$.
\end{definition}
\begin{remark} If $p = 0$ or $p = +\infty$ or the limit does not exist we say that the product diverges. A necessary
condition for convergence of an infinite product is $\displaystyle{\lim_{j \to +\infty}} a_j =1$, but this is not sufficient.
\end{remark}
\begin{remark}\label{suc mayor a 1} If $\{ a_j \}_{j=1}^{+\infty}$ is a sequence of real numbers greater than or equal to $1$, then
$1 \leq p_N \leq p_{N+1}$ for all $N \geq 1$. So, $\prod_{j=1}^{+\infty} a_j = \sup \{ p_N : N \in \mathbb{N}\} \in [1, +\infty]$. This is,
$\prod_{j=1}^{+\infty} a_j$ converges in $\mathbb{R}_{\geq 1}$ or $\prod_{j=1}^{+\infty} a_j = +\infty$.
\end{remark}
The following theorem on rearrangements is an immediate consequence of Theorem 3.3 in \cite{moorthy} and from remark \ref{suc mayor a 1}.
\begin{theorem}\label{theo rearr} If $\{ a_j \}_{j=1}^{+\infty}$ is a sequence of real numbers greater than or equal to $1$, then
\[
\prod_{k=1}^{+\infty} a_{\sigma(k)} = \prod_{j=1}^{+\infty} a_{j}
\]
for each bijection $\sigma : \mathbb{N} \to \mathbb{N}$.
\end{theorem}
\
The proof of Theorem 4 is simpler. Indeed, for $a_j \geq 1$, we have that
\[
\prod_j a_j = \exp\left( \sum_j \log(a_j) \right) = \exp\left( \sum_k \log(a_{\sigma(k)}) \right) = \prod_j a_{\sigma(k)}.
\]
\subsection{Double infinite products}
Formally, a double infinite product of positive real numbers is of the form
\[
\prod_{i,j=1}^{+\infty} a_{ij}
\]
where the $a_{i j}$'s are indexed by pair of natural numbers $i, j \in \mathbb{N}$.
\
A sequence of real numbers of the form $\{ a_{ij} \}$ is called a double sequence.
\
Our main interest is in giving a definition of a double infinite product of real numbers greater than or equal to $1$ that does not depend on the order of its factors.
\
If $F \subset \mathbb{N} \times \mathbb{N}$ is a finite subset of pairs of natural numbers, then we denote by
\[
\prod_{F} a_{ij} = \prod_{(i,j) \in F} a_{ij}
\]
the partial product of all $a_{ij}$'s whose indices $(i,j)$ belong to $F$.
\begin{definition} The unordered infinite product of real numbers $a_{ij} \geq 1$ is
\[
\prod_{\mathbb{N} \times \mathbb{N}} a_{ij} = \sup_{F \in \mathcal{F}} \left\{ \prod_{F} a_{ij} \right\}
\]
where the supremum is taken over the collection $\mathcal{F}$ of all finite subsets $F \subset \mathbb{N} \times \mathbb{N}$.
\end{definition}
Such an unordered infinite product converges if the supremum is finite and diverges to $+\infty$ if the supremum is $+\infty$.
Note that this supremum exists in $\mathbb{R}_{\geq 1}$ if and only if the finite partial products are bounded from above.
\
Next, we define rearrangements of a double infinite product into single infinite product and show that every rearrangement of a
convergent unordered double infinite product converges to the same product.
\
\begin{definition} A rearrangement of a double infinite product of positive real numbers
\[
\prod_{i,j}^{+\infty} a_{ij}
\]
is a single infinite product of the form
\[
\prod_{k=1}^{+\infty} b_k, \,\,\,\,\,\, b_k = a_{\sigma{(k)}}
\]
where $\sigma : \mathbb{N} \to \mathbb{N} \times \mathbb{N}$ is a one-to-one and onto map.
\end{definition}
The following result is an adaptation of Theorem 8.42 in \cite{apostol}, pp. 201, to the context of double products.
\begin{theorem}\label{theo rearr 2} If the unordered product of a double sequence of real numbers greater than or equal to $1$ converges,
then every rearrangement of the double infinite product into a single infinite product converges to the unordered product.
\end{theorem}
\begin{proof} Suppose that the unordered product $\prod a_{ij}$, $a_{ij} \geq 1$, converges with
\[
\prod_{\mathbb{N} \times \mathbb{N}} a_{ij} = s \in \mathbb{R}_{\geq 1},
\]
and let
\[
\prod_{k=1}^{+\infty} b_k
\]
be a rearrangement of the double product corresponding to a map $\sigma : \mathbb{N} \to \mathbb{N} \times \mathbb{N}$.
For $N \in \mathbb{N}$, let
\[
F_N = \{ \sigma(k) \in \mathbb{N} \times \mathbb{N} : 1 \leq k \leq N \},
\]
so that
\[
\prod_{k=1}^{N} b_k = \prod_{F_N} a_{ij}.
\]
Given $\epsilon > 0$, let $F \subset \mathbb{N} \times \mathbb{N}$ be a finite set such that
\[
s - \epsilon < \prod_{F} a_{ij} \leq s,
\]
and let $K \in \mathbb{N}$ be defined by $K = \max \{ \sigma^{-1}(i,j) : (i,j) \in F \}$.
If $N \geq K$, then $F_{N} \supseteq F$ and since $a_{ij} \geq 1$, we obtain
\[
s - \epsilon < \prod_{F} a_{ij} \leq \prod_{F_N} a_{ij} \leq s.
\]
This implies that
\[
\left| \prod_{k=1}^{N} b_k - s \right| \leq \epsilon, \,\,\,\,\,\,\,\, \forall \, N \geq K.
\]
Thus,
\[
\prod_{k=1}^{+\infty} b_k = \prod_{\mathbb{N} \times \mathbb{N}} a_{ij}.
\]
\end{proof}
The rearrangement of a double infinite product into a single infinite product is one natural way to
interpret a double infinite product in terms of single infinite products. Another way is to use iterated
products of single infinite product.
Given a double product $\prod a_{ij}$, one can define two iterated products, obtained by multiplying first over one index followed by the other:
\[
\prod_{i=1}^{+\infty} \left( \prod_{j=1}^{+\infty} a_{ij} \right) = \lim_{N \to +\infty} \prod_{i=1}^{N} \left(
\lim_{M \to +\infty} \prod_{j=1}^{M} a_{ij} \right),
\]
\[
\prod_{j=1}^{+\infty} \left( \prod_{i=1}^{+\infty} a_{ij} \right) = \lim_{M \to +\infty} \prod_{j=1}^{M} \left(
\lim_{N \to +\infty} \prod_{i=1}^{M} a_{ij} \right).
\]
\begin{theorem}\label{iter} A unordered product of real numbers $a_{ij} \geq 1$ converges if and only if either one of the iterated products
\[
\prod_{i=1}^{+\infty} \left( \prod_{j=1}^{+\infty} a_{ij} \right), \,\,\,\,\,\, \prod_{j=1}^{+\infty} \left( \prod_{i=1}^{+\infty} a_{ij} \right)
\]
converges. Moreover, both iterated products converge to the unordered product:
\[
\prod_{\mathbb{N} \times \mathbb{N}} a_{ij} = \prod_{i=1}^{+\infty} \left( \prod_{j=1}^{+\infty} a_{ij} \right) =
\prod_{j=1}^{+\infty} \left( \prod_{i=1}^{+\infty} a_{ij} \right)
\]
\end{theorem}
\begin{proof} Suppose that the unordered product converges. Since $a_{ij} \geq 1$ and $\sup_{I} \left( \sup_{J} c_{ij} \right) =
\sup_{I \times J} c_{ij}$ for each double sequence of reals, we obtain
\[
\prod_{i=1}^{+\infty} \left( \prod_{j=1}^{+\infty} a_{ij} \right) = \sup_{N} \left\{\prod_{i=1}^{N} \left( \sup_{M}\prod_{j=1}^{M} a_{ij} \right) \right\}
= \sup_{(N, M) \in \mathbb{N} \times \mathbb{N}} \left\{ \prod_{i=1}^{N} \prod_{j=1}^{M} a_{ij} \right\} \leq
\prod_{\mathbb{N} \times \mathbb{N}} a_{ij}.
\]
Conversely, suppose that one of the iterated products exists. Without loss of generality, we suppose that
\[
\prod_{i=1}^{+\infty} \left( \prod_{j=1}^{+\infty} a_{ij} \right) < +\infty.
\]
Let $F \subset \mathbb{N} \times \mathbb{N}$ be a finite subset. Then, there exist two natural numbers $N$ and $M$ such that
\[
F \subset \{1, 2, ..., N \} \times \{ 1, 2, ..., M \} =: R.
\]
So that,
\[
\prod_{F} a_{ij} \leq \prod_{R} a_{ij} = \prod_{i=1}^{N} \left( \prod_{j=1}^{M} a_{ij} \right) \leq
\prod_{i=1}^{+\infty} \left( \prod_{j=1}^{+\infty} a_{ij} \right).
\]
Therefore, the unordered product converges and
\[
\prod_{\mathbb{N} \times \mathbb{N}} a_{ij} \leq \prod_{i=1}^{+\infty} \left( \prod_{j=1}^{+\infty} a_{ij} \right).
\]
\end{proof}
\section{Main Results}
To construct our multiplicative measure $\mu$, we consider the positive real axis $\mathbb{R}_{>0}$ with the usual topology and for each closed interval $I=[a,b]$ in $\mathbb{R}_{>0}$, we define its "length", $\ell(I)$, by
\[
\ell(I) = b \cdot a^{-1},
\]
where $\cdot$ is the product of $\mathbb{R}$. It is easy to check that
\
1. $\ell(I) \geq 1$ for each closed interval $I = [a,b]$ in $\mathbb{R}_{>0}$.
\
2. If $I$, $J_1$ and $J_2$ are three closed intervals in $\mathbb{R}_{>0}$ such that $I \subseteq J_1 \bigcup J_2$, then
$\ell(I) \leq \ell(J_1) \cdot \ell(J_2)$.
\
In the sequel, a cover of a set $E \subseteq \mathbb{R}_{>0}$ is a countable collection $S$ of intervals $I =[a,b]$, with
$0 < a < b < +\infty$, such that $E \subset \bigcup_{I \in S} I$.
\
Now we shall construct the exterior measure of an arbitrary subset $E$ of $\mathbb{R}_{>0}$. Given a set $E \subseteq \mathbb{R}_{>0}$,
we cover $E$ by a countable collection of intervals $S= \{ I_j \}$ in $\mathbb{R}_{>0}$, and let
\[
\nu(S) = \prod_{I_j \in S} \ell(I_j).
\]
We point out that the value $\nu(S)$ is independent of the rearrangement of the cover $S=\{ I_j \}$ of $E$ (see Theorem \ref{theo rearr}). The exterior measure of $E$, denoted $\mu_{e}(E)$, is defined by
\[
\mu_e (E) = \inf_{S} \nu(S),
\]
where the infimum is taken over all such covers $S$ of $E$. Thus, $1 \leq \mu_e (E) \leq + \infty$. Moreover, if $E_1 \subset E_2 \subset
\mathbb{R}_{>0}$, then $\mu_e (E_1) \leq \mu_e(E_2)$.
\begin{theorem}\label{intervalo} Let $I =[a,b] \subset\mathbb{R}_{>0}$. Then $\mu_e (I) = \ell(I)$.
\end{theorem}
\begin{proof} That $\mu_e(I) \leq \ell(I)$ is obvious. To show the opposite inequality we choose a countable cover $S=\{ I_j \}$ of $I$.
Let $0 < \epsilon < 1$ be fixed. For each $j \in \mathbb{N}$, let $I_{j}^{\ast}$ be an interval of $\mathbb{R}_{>0}$ such that its interior
$(I_{j}^{\ast})^{\circ} \supset I_j$ and $\ell(I_{j}^{\ast}) \leq \frac{1 + \epsilon^{j}}{1 + \epsilon^{j+1}} \cdot \ell(I_j)$. Then
$I \subset \bigcup_j (I_{j}^{\ast})^{\circ}$. Since $I$ is closed and bounded, by the Heine-Borel Theorem, there exists $N \in \mathbb{N}$ such that $I \subset \bigcup_{j=1}^{N} I_{j}^{\ast}$. Now, the property 2. above implies that
$\ell(I) \leq \prod_{j=1}^{N} \ell(I_{j}^{\ast})$. Therefore,
\[
\ell(I) \leq \frac{1 + \epsilon}{1 + \epsilon^{N+1}} \prod_{j=1}^{N} \ell(I_j) \leq (1 + \epsilon) \cdot \nu(S).
\]
Then letting $\epsilon \to 0^{+}$, we obtain $\ell(I) \leq \nu(S)$ and hence $\ell(I) \leq \mu_e(I)$.
\end{proof}
\begin{theorem}\label{sub multi} If $E = \bigcup_j E_j$ is a countable union of subsets of $\mathbb{R}_{>0}$, then $$\mu_e (E) \leq \prod_j \mu_e(E_j).$$
\end{theorem}
\begin{proof} Without loss of generality, we may assume that $\prod_j \mu_e(E_j) < +\infty$. Now, we fix $0 < \epsilon < 1$.
Given $j \in \mathbb{N}$, we choose intervals $I_{ij}$ of $\mathbb{R}_{>0}$ such that $E_j \subset \bigcup_{i} I_{ij}$ and
\begin{equation}\label{ineq}
\prod_{i} \ell(I_{ij}) < \frac{1+\epsilon^{j}}{1+ \epsilon^{j+1}}\mu_{e}(E_j).
\end{equation}
Let $\sigma : \mathbb{N} \to \mathbb{N} \times \mathbb{N}$ be a bijection, then the collection $\{ J_k \}$ defined by $J_k = I_{\sigma(k)}$ is a cover of $E$. So that,
\begin{equation}\label{submulti 2}
\mu_{e}(E) \leq \prod_{k} \ell(J_k) = \prod_{\mathbb{N} \times \mathbb{N}} \ell(I_{ij}) = \prod_j \left( \prod_i \ell(I_{ij}) \right)
\leq (1+ \epsilon) \prod_{j} \mu_e(E_j),
\end{equation}
where the first equality follows from Theorem \ref{theo rearr 2}, the second from Theorem \ref{iter} and the last inequality from
(\ref{ineq}) and
\[
\prod_{j=1}^{N} \frac{1+\epsilon^{j}}{1+ \epsilon^{j+1}}\mu_{e}(E_j) = \frac{1+\epsilon}{1+ \epsilon^{N+1}} \prod_{j=1}^{N} \mu_e(E_j) \to
(1+ \epsilon) \prod_{j=1}^{+\infty} \mu_e(E_j).
\]
Finally, the result follows by letting $\epsilon \to 0$ in (\ref{submulti 2}).
\end{proof}
The next theorem relates the exterior measure of an arbitrary set of $\mathbb{R}_{>0}$ with the exterior measure of open sets and
$G_{\delta}$ sets of $\mathbb{R}_{>0}$. This theorem is an analogous of the theorems 3.6 and 3.8 in \cite{zygmund}, pp. 44.
\begin{theorem}\label{G delta} Let $E \subseteq \mathbb{R}_{>0}$, then
\begin{enumerate}
\item[(i)] given $0 < \epsilon < 1$, there exists and open set $G$ such that $E \subset G$ and $\mu_e(G) \leq (1+ \epsilon) \cdot \mu_e(E)$;
\item[(ii)] there exists a set $H$ of type $G_{\delta}$ in $\mathbb{R}_{>0}$ such that $E \subset H$ and $\mu_e(E) = \mu_e(H)$.
\end{enumerate}
\end{theorem}
\begin{proof} It is easy to check that $(i)$ implies $(ii)$. To prove $(i)$, we fix $0 < \epsilon < 1$ and choose a cover $\{ I_j \}$ of $E$ in $\mathbb{R}_{>0}$ such that $\prod_j \ell(I_j) \leq \sqrt{(1 + \epsilon)} \cdot \mu_e(E)$. For each $j \in \mathbb{N}$,
let $I_{j}^{\ast}$ be an interval of $\mathbb{R}_{>0}$ such that its interior $(I_{j}^{\ast})^{\circ} \supset I_j$ and
$\ell(I_{j}^{\ast}) \leq \frac{\sqrt{(1 + \epsilon^{j})}}{\sqrt{(1 + \epsilon^{j+1})}} \cdot \ell(I_j)$. If
$G = \bigcup_j (I_{j}^{\ast})^{\circ}$, then $G$ is an open set such that $E \subset G$ and
\[
\mu_e(G) \leq \prod_j \ell(I_{j}^{\ast}) \leq \sqrt{(1 + \epsilon)} \cdot \prod_j \ell(I_j) \leq (1+\epsilon) \cdot \mu_{e}(E).
\]
\end{proof}
The following corollary is an immediate consequence of Theorem \ref{G delta}, $(i)$.
\begin{corollary}
Let $E \subset \mathbb{R}_{>0}$. Then $$\mu_e(E) = \inf \mu_e(G),$$ where the infimum is taken over all open sets $G$ of $\mathbb{R}_{>0}$ containing $E$.
\end{corollary}
A set $E \subseteq \mathbb{R}_{>0}$ is said to be \textit{measurable} if given $\epsilon > 0$, there exists an open set $G$ of
$\mathbb{R}_{>0}$ such that $E \subset G$ and
\[
\mu_e (G \setminus E) < 1 + \epsilon.
\]
If $E \subseteq \mathbb{R}_{>0}$ is measurable, its exterior measure is called its measure, which we denote by $\mu(E)$, i.e.:
\[
\mu(E) = \mu_e(E), \,\,\, \text{for measurable} \,\, E.
\]
We say that a set $E$ is $\mu$-measurable if satisfies the above definition.
\
The next list of examples and properties of measurable sets of $\mathbb{R}_{>0}$ follows from the previous results and the definition of measurability.
\
$1)$ Every open set of $\mathbb{R}_{>0}$ is $\mu$-measurable. For instance, the empty set and the sets $\mathbb{R}_{>0}$ and $(0,1)$ are $\mu$-measurable with $\mu(\emptyset) = 1$, $\mu(\mathbb{R}_{>0}) = +\infty$ and $\mu((0,1)) = +\infty$.
\
$2)$ Every set of exterior measure $1$ is $\mu$-measurable. For instance, every point $\{ a \}$ of $\mathbb{R}_{>0}$ is $\mu$-measurable with
$\mu(\{ a \}) = 1$.
\
The claim $2)$ follows from Theorem \ref{G delta}, $(i)$.
\
$3)$ If $E = \bigcup_j E_j$ is a countable union of $\mu$-measurable sets of $\mathbb{R}_{>0}$, then $E$ is $\mu$-measurable and
\[
\mu(E) \leq \prod_j \mu(E_j).
\]
Indeed, given $0 < \epsilon < 1$, for each $j=1, 2, ...$, we take an open set $G_j$ such that $E_j \subset G_j$ and
$\mu_e(G_j \setminus E_j) < \frac{1 + \epsilon^{j}}{1 + \epsilon^{j+1}}$. Then $G = \bigcup_j G_j$ is an open such that $E \subset G$,
and since $G \setminus E \subset \bigcup_j (G_j \setminus E_j)$, we have
\[
\mu_e(G \setminus E) \leq \mu_e \left( \bigcup_j (G_j \setminus E_j) \right) \leq \prod_j \mu_e ((G_j \setminus E_j)) \leq (1+ \epsilon),
\]
where the second inequality follows from Theorem \ref{sub multi}. Thus $E$ is $\mu$-measurable and
$\mu(E) = \mu_e(E) \leq \prod_j \mu_e(E_j) = \prod_j \mu(E_j)$.
\
$4)$ Every interval $I \subset \mathbb{R}_{>0}$ is $\mu$-measurable. If $I$ is bounded and $\inf I > 0$, then $\mu(I) = \ell(I)$; opposite case
$\mu(I) = +\infty$.
\
The claim $4)$ follows from Theorem \ref{intervalo}, and from $1)$ and $2)$ above.
\
$5)$ If $\{ I_j \}_{j=1}^{N}$ is a finite collection of nonoverlapping intervals of $\mathbb{R}_{>0}$, then $\bigcup_j I_j$ is $\mu$-measurable and
$\mu\left( \bigcup_j I_j \right) = \prod_j \mu(I_j)$.
\
The claim $5)$ is an extension of Theorem \ref{intervalo}.
\
$6)$ If $E_1$ and $E_2$ are two subsets of $\mathbb{R}_{>0}$ such that
\[
d(E_1, E_2) := \inf \{ |x - y| : x \in E_1, y \in E_2 \} > 0,
\]
then $\mu_e(E_1 \bigcup E_2) = \mu_e(E_1) \cdot \mu_e(E_2)$.
\
The proof of the claim $6)$ is similar to that of Lemma 3.16, in \cite{zygmund}, but reformulated to our setting.
\
$7)$ Every closed set $F$ of $\mathbb{R}_{>0}$ is $\mu$-measurable.
\
To see $7)$, we assume first that $F$ is a compact set in $\mathbb{R}_{>0}$, so $\inf F > 0$ and hence $\mu_e(F) < +\infty$. Given $0 < \epsilon < 1$, by Theorem \ref{G delta} statement $(i)$, there exists an open set $G$ such that $F \subset G$ and
$\mu_e(G) \leq (1+ \epsilon) \cdot \mu_e(F)$. Since $G\setminus F$ is open in $\mathbb{R}_{>0}$, Theorem 1.11 in \cite{zygmund}, implies there exists a collection of nonoverlapping closed intervals $\{ I_j \}$, where $I_j =[a_j, b_j]$ with $0 < a_j < b_j$, such that
$G\setminus F = \bigcup_j I_j$. Thus, $\mu_e (G\setminus F) \leq \prod_j \mu_e(I_j)$. Then, to prove the $\mu$-measurability of $F$ it suffices
to show that $\prod_j \mu_e(I_j) \leq 1 + \epsilon$. We have $G = F \cup (\bigcup I_j) \supset F \cup (\bigcup_{j=1}^{N} I_j)$ for each $N$.
Since $F$ and $\bigcup_{j=1}^{N} I_j$ are disjoint and compact, by $6)$, we obtain
\[
\mu_e(G) \geq \mu_e \left( F \cup (\bigcup_{j=1}^{N} I_j) \right) = \mu_e(F) \cdot \mu_e \left( \bigcup_{j=1}^{N} I_j \right).
\]
By $5)$, we have $\prod_{j=1}^{N} \mu_e(I_j) = \mu_e \left( \bigcup_{j=1}^{N} I_j \right) \leq \displaystyle{\frac{\mu_e(G)}{\mu_e(F)}} \leq 1 + \epsilon$. This proves the result in the case when $F$ is compact.
To end, let $F$ be any closed subset of $\mathbb{R}_{>0}$ and write $F = \bigcup_j F_j$, where
$F_j = F \cap \{ x > 0 : j^{-1} \leq x \leq j \}$, $j=2, 3, ...$. Since, each $F_j$ is compact and, thus, $\mu$-measurable; then the $\mu$-measurability of $F$ follows from $3)$.
\
$8)$ The complement of a $\mu$-measurable set of $\mathbb{R}_{>0}$ is $\mu$-measurable.
\
The proof of $8)$ is similar to that of Theorem 3.17 in \cite{zygmund}.
\
$9)$ If $\{ E_j \}$ is a countable collection of $\mu$-measurable sets of $\mathbb{R}_{>0}$, then $\bigcap_j E_j$ is $\mu$-measurable.
\
The proof of $9)$ is similar to that of Theorem 3.18 in \cite{zygmund}.
\
$10)$ A set $E$ of $\mathbb{R}_{>0}$ is $\mu$-measurable if and only if given $\epsilon >0$, there exists a closed set $F \subset E$ such that
$\mu_e(E \setminus F) < 1 + \epsilon$.
\
The proof of $10)$ is similar to that of Lemma 3.22 in \cite{zygmund}.
\
The following result follows from $3)$, $8)$ and $9)$.
\begin{theorem} Let $\mathcal{M}$ be the collection of $\mu$-measurable subsets of $\mathbb{R}_{>0}$. Then, $\mathcal{M}$ is a $\sigma$-algebra.
\end{theorem}
From $1)$ and this theorem we obtain the following corollary.
\begin{corollary} Every Borel set of $\mathbb{R}_{>0}$ is $\mu$-measurable.
\end{corollary}
We are now in a position to prove the following theorem.
\begin{theorem} If $\{ E_j \}$ is a countable collection of disjoint $\mu$-measurable subsets of $\mathbb{R}_{>0}$, then
\[
\mu \left( \bigcup_j E_j \right) = \prod_j \mu(E_j).
\]
\end{theorem}
\begin{proof} First, we assume that each $E_j$ is bounded with $\inf E_j > 0$. Given $0 < \epsilon < 1$ and $j=1, 2,...$, by $10)$, there exists a closed $F_j \subset E_j$ such that $\mu(E_j \setminus F_j) < \frac{1+ \epsilon^{j}}{1+ \epsilon^{j+1}}$. Then,
$\mu (E_j) \leq \frac{1+ \epsilon^{j}}{1+ \epsilon^{j+1}} \cdot \mu(F_j)$. Since the $E_j$ are bounded with $\inf E_j >0$ and disjoint,
the $F_j$ are compact and disjoint. Then, By $6)$, we have $\mu \left( \bigcup_{j=1}^{N} F_j \right) = \prod_{j=1}^{N} \mu(F_j)$, for each
$N$. So,
\[
\mu \left(\bigcup_j E_j \right) \geq \prod_{j=1}^{N} \mu(F_j) \geq \frac{1 + \epsilon^{N+1}}{1 + \epsilon} \prod_{j=1}^{N} \mu(E_j) \geq
\frac{1}{1 + \epsilon} \prod_{j=1}^{N} \mu(E_j),
\]
for each $N$. By letting $N \to + \infty$ and since $0 < \epsilon < 1$ is arbitrary, we obtain $\mu \left(\bigcup_j E_j \right) \geq
\prod_{j=1}^{+\infty} \mu(E_j)$. The opposite inequality follows from Theorem \ref{sub multi}. This proves the theorem in the case when
the $E_j$ are bounded with $\inf E_j > 0$.
For the general case, let $I_k = [k^{-1}, k]$, $k =1, 2, ...$, and we define $J_1= I_1$ and $J_k = I_k \setminus I_{k-1}$ for $k \geq 2$.
Then the sets $A_{jk} = E_j \cap J_k$, with $j, k=1,2,...$, are bounded, disjoint, $\mu$-measurable and with $\inf A_{jk} >0$. Since
$E_j = \bigcup_{k} A_{jk}$ and $\bigcup_{j} E_j = \bigcup_{j, k} A_{jk}$, by the case already established and proceeding, with the double product, as in the proof of Theorem \ref{sub multi}, we have
\[
\mu \left( \bigcup_j E_j \right) = \mu \left( \bigcup_{j,k} A_{jk} \right) = \prod_{\mathbb{N} \times \mathbb{N}} \mu(A_{jk}) =
\prod_{j=1}^{+\infty} \left( \prod_{k=1}^{+\infty} \mu(A_{jk}) \right)= \prod_{j=1}^{+\infty} \mu(E_j).
\]
This completes the proof.
\end{proof}
We also have the Carath\'eodory's characterization for $\mu$-measurable sets of $\mathbb{R}_{>0}$. The proof of the following theorem is similar to that of Theorem 3.30, in \cite{zygmund}, but reformulated to our setting.
\begin{theorem} A set $E$ of $\mathbb{R}_{>0}$ is $\mu$-measurable if and only if for every set $A$ of $\mathbb{R}_{>0}$
\[
\mu_e(A) = \mu_e(A \cap E) \cdot \mu_e(A \setminus E).
\]
\end{theorem}
We summarize our main result in the following theorem.
\begin{theorem} Let $\mathbb{R}_{>0}$ with the usual topology. Then there exist a $\sigma$-algebra $\mathcal{M}$ containing every Borel set
of $\mathbb{R}_{>0}$ and a measure $\mu : \mathcal{M} \to [1, +\infty]$ such that
\begin{enumerate}
\item[(i)] $\mu(\emptyset) = 1$.
\item[(ii)] For each interval $I$ of $\mathbb{R}_{>0}$, we have that $\mu(I) = \ell(I)$ if $I$ is bounded and $\inf I >0$; opposite case $\mu(I)=+\infty$.
\item[(iii)]
\[
\mu \left( \bigcup_j E_j \right) = \prod_j \mu(E_j)
\]
for every countable collection $\{ E_j \}$ of pairwise disjoint sets of $\mathcal{M}$. We call to this property:
countable multiplicativity.
\end{enumerate}
\end{theorem}
By this Theorem, we say that $\mu$ is a \textit{multiplicative measure}. So, we have constructed a multiplicative measure on
$\mathbb{R}_{>0}$ using only the topology usual on $\mathbb{R}_{>0}$ and its multiplicative structure.
\
In the next section, we shall describe the connection between our measure $\mu$ and the Lebesgue measure.
\section{The measure $\lambda$}
Let $\lambda$ be now the (multiplicative) measure defined on the $\sigma$-algebra of Lebesgue measurable subsets of $\mathbb{R}_{>0}$ by
\[
\lambda(E) = \exp\left(\int_{E} \frac{1}{x} \, dx \right).
\]
\\
It is clear that for each interval $I=[a,b] \subset \mathbb{R}_{>0}$, $\mu(I) = \lambda(I)$. So, $\mu(G) = \lambda(G)$ for each open set
$G \subset \mathbb{R}_{>0}$. If we show that the $\sigma$-algebra $\mathcal{M}$ coincides with the $\sigma$-algebra of the Lebesgue measurable sets of $\mathbb{R}_{>0}$, then it will follow that $\mu(E) = \lambda(E)$ for all $E \in \mathcal{M}$.
From the definition of the $\mu$-measurability follows that $E \in \mathcal{M}$ (or $E$ is \textit{$\mu$-measurable}) if and only if $E = H \setminus U$ where $H$ is of type $G_{\delta}$ and $\mu(U) = 1$. Thus, to prove that $E$ is $\mu$-measurable if and only if $E$ is Lebesgue measurable it suffices to show that
\[
\mu(U) =1 \,\,\,\, \text{if and only if} \,\,\,\,\, \left| U \right|_e = 0,
\]
\\
here $| \cdot |_e$ denote the outer Lebesgue measure.
Suppose $\mu(U) = 1$. By Theorem \ref{G delta}, claim (i), given $\epsilon > 0$, there exists an open set $G$, such that $U \subset G$ and
\begin{equation}\label{la G}
\mu(G) \leq 1 + \epsilon.
\end{equation}
\\
Theorem 1.1, in \cite{zygmund}, implies there exists a collection of nonoverlapping closed interval $\{ I_j = [a_j, b_j] \}$ such that
$G = \bigcup_j I_j$. Then, $\log(G) \subset \bigcup_j [\log(a_j), \log(b_j)]$. Since $\lambda(G)=\mu(G)$, we obtain,
by (\ref{la G}), that
\[
| \log(G) |_{e} \leq \sum_{j} [\log(b_j) - \log(a_j) ] = \log\left(\lambda(G)\right) = \log (\mu(G)) \leq
\log(1+\epsilon).
\]
So, $| \log(G) |_{e} = 0$ and, hence, $| \log(U) |_{e} = 0$. Then, to apply the exponential map on the region $\log(U)$, we have
$|U|_e = 0$.
Similarly, it is proved that $|U|_e=0$ implies $\mu(U) =1$. Thus, the $\sigma$-algebra $\mathcal{M}$ coincides with the $\sigma$-algebra of the Lebesgue measurable sets of $\mathbb{R}_{>0}$, and $\mu \equiv \lambda$.
\
|
\subsection{Notations} Let the set of users and items be denoted as $\mathcal{U}=\{u_1, u_2,...,u_n\}$ and $\mathcal{I}=\{ i_1, i_2,...,i_m \}$, where $n$ is the number of users and $m$ is the number of items.
Let $\textbf{R} \in \{0,1\}^{n \times m}$ be the user-item interaction matrix,
where $r_{ui}=1$ if the user $u$ has an interaction with the item $i$, otherwise $r_{ui}=0$.
Lastly, for a user $u$, $\mathcal{I}_{u}^{+} = \{ i \in \mathcal{I} | r_{ui} = 1\}$ is the set of interacted items and $\mathcal{I}_{u}^{-} = \{ i \in \mathcal{I} | r_{ui} = 0\}$ is the set of unobserved items.
\subsection{Top-$K$ Recommender System}
A top-$K$ recommender system aims to find a ranked list of unobserved items for each user.
To make the ranked list, the top-$K$ RS predicts the score $\hat{r}_{ui} = P(r=1|u,i)$ for each item $i$ in $\mathcal{I}_{u}^{-}$ for each user $u$.
Generally, top-$K$ recommender systems use a point-wise loss or a pair-wise loss.
The point-wise loss usually takes the negative log-likelihood for the binary preferences \cite{ncf17, cdae16}.
For each interaction $(u,i)$, the point-wise loss is defined as follows:
\begin{equation}
\mathcal{L}_{point}(u,i) = - \Big( r_{ui} \text{log} \hat{r}_{ui} + (1-r_{ui}) \text{log} (1-\hat{r}_{ui}) \Big).
\end{equation}
Similarly, the pair-wise loss is defined as follows \cite{bpr09}:
\begin{equation}
\mathcal{L}_{pair}(u,i,j) = - \text{log} (\hat{r}_{uij}),
\end{equation}
where $i$ is a positive item and $j$ is a sampled negative item, and $\hat{r}_{uij} = P(r_{ui} > r_{uj})$.
With one of those two losses, the collaborative filtering loss can be formalized for each user $u$ as follows:
\begin{equation}
\begin{gathered}
\mathcal{L}_{CF} = \sum_{i \in \mathcal{I}_{u}^{+}} \mathcal{L}_{point}(u,i) + \sum_{j \in \mathcal{NS}(\mathcal{I}_{u}^{-})} \mathcal{L}_{point}(u,j) \\
\text{or} \\
\mathcal{L}_{CF} = \sum_{i \in \mathcal{I}_{u}^{+}} \sum_{j \in \mathcal{NS}(\mathcal{I}_{u}^{-})} \mathcal{L}_{pair}(u,i,j),
\end{gathered}
\end{equation}
where $\mathcal{NS}(\mathcal{I}_{u}^{-})$ is a set of randomly sampled negative items from $\mathcal{I}_{u}^{-}$.
\begin{figure*}[t]
\includegraphics[width=0.8\linewidth]{figure/method.png}
\caption{Illustration of Bidirectional Distillation (BD) for top-$K$ recommender systems.}
\end{figure*}
\subsection{Knowledge Distillation for Top-$K$ RS}
Knowledge Distillation (KD) is a model-agnostic framework to train a compact model (student) by using the knowledge transferred from a pretrained cumbersome model (teacher) \cite{kd15}.
The student model uses the class probabilities produced by the teacher as "soft target" along with the "hard target" from the training set.
For the multiclass classification, the loss function for KD can be defined as follows:
\begin{equation}
\mathcal{L} = \mathcal{L}_{CE}(\sigma(\textbf{z}_S), \textbf{y}) + \lambda \mathcal{L}_{CE}(\sigma(\frac{\textbf{z}_S}{T}), \sigma(\frac{\textbf{z}_T}{T})),
\end{equation}
where $\mathcal{L}_{CE}(\cdot)$ denotes the cross-entropy loss, $\sigma(\cdot)$ denotes the softmax function, $\textbf{z}_S$ and $\textbf{z}_T$ are the logits of the student and the teacher, respectively, $\textbf{y}$ is the hard target from the training set, $T$ is the temperature which controls the smoothness, and $\lambda$ is a weight that controls the effect of KD.
This KD framework is widely adopted in computer vision field \cite{kd-cv17, kd-cv17nips} and has achieved remarkable success.
For top-$K$ recommender systems, a few work \cite{rd18, cd19} adopted KD to train a compact neural recommender with the guidance of a cumbersome neural recommender.
Ranking Distillation (RD) \cite{rd18} firstly proposes a KD framework for RS:
\begin{equation}
\mathcal{L} = \mathcal{L}_{CF} + \lambda \mathcal{L}_{RD} \text{,}
\end{equation}
where $\mathcal{L}_{CF}$ is the collaborative filtering loss in Equation 3, and $\mathcal{L}_{RD}$ is the distillation loss.
RD uses the distillation loss $\mathcal{L}_{RD}$ to make the student give high scores on top-$N$ items of the teacher's recommendation list:
\begin{equation}
\mathcal{L}_{RD} = - \sum_{n=1}^N w_n \cdot \text{log} \hat{r}_{u,\pi_n},
\end{equation}
where $N$ is the sampling size, $w_n$ is a rank-aware weight, $\pi_n$ is the $n$-th item in the teacher's recommendation list and $\hat{r}_{u,\pi_n}$ denotes the model prediction $P(r=1|u,\pi_n)$.
Collaborative Distillation (CD) \cite{cd19} proposes a subsequent KD framework for RS with a sampling technique.
In RD, distillation process is done only with the top-$N$ items of the teacher's recommendation list.
Instead, CD uses a top-aware sampling technique to choose the negative items from the teacher's recommendation list:
\begin{equation}
\mathcal{L}_{CD} = - \sum_{j \in \mathcal{TS}(\mathcal{I}_{u}^{-})} q_{uj}\text{log}\hat{r}_{uj} + (1-q_{uj})\text{log}(1-\hat{r}_{uj}),
\end{equation}
where $\mathcal{TS}(\mathcal{I}_{u}^{-})$ is a set of top-aware-sampled items from the teacher's recommendation list, $|\mathcal{TS}(\mathcal{I}_{u}^{-})|$ is $N$, and $q_{uj}$ is the soft target produced by the teacher.
Overall, those two existing KD frameworks make the student model imitate the teacher's predictions on highly ranked unobserved items and successfully train a student model which is comparable to the teacher model.
Nevertheless, they still have room for improvement by the following reasons;
First, existing distillation methods are unidirectional, which means that only the student get the transferred knowledge from the teacher.
However, we argue that the teacher model also could be taught by the student model.
As we observed in Section 2, the student is not always inferior to the teacher, but rather performs better than the teacher on the significant proportion of the test set.
Therefore, existing unidirectional distillation frameworks are ineffective for top-$K$ RS and the student's performance is highly limited by the teacher.
Second, the distillation process puts emphasis only on the highly ranked unobserved items.
However, highly ranked items may not be very useful for the distillation, since most of items ranked highly by the teacher are also ranked highly by the student (Figure 1b).
Therefore, for the student, we need to focus on the items that are not only ranked highly by the teacher but also ranked lowly by the student, and vice versa.
\end{comment}
\subsection{Overview}
Figure 2 shows an overview of BD.
Unlike the existing KD methods for RS, both the teacher and the student are trained simultaneously by using each other's knowledge (i.e., recommendation list) along with the binary training set (i.e., user-item interactions).
First, the teacher and the student produce the recommendation list for each user.
Second, BD decides what knowledge to be transferred for each distillation direction based on the rank discrepancy-aware sampling scheme.
Lastly, the teacher and the student are trained with the distillation loss along with the original collaborative filtering loss.
To summarize, each of them is trained as follows:
\begin{equation}
\begin{split}
& \mathcal{L}_T(\theta_T) = \mathcal{L}_{CF}(\theta_T) + \lambda_{S \rightarrow T} \cdot \mathcal{L}_{BD}(\theta_T; \theta_S), \\
& \mathcal{L}_S(\theta_S) = \mathcal{L}_{CF}(\theta_S) + \lambda_{T \rightarrow S} \cdot \mathcal{L}_{BD}(\theta_S; \theta_T),
\end{split}
\end{equation}
where $T$ and $S$ denote the teacher and the student respectively, $\theta_*$ is the model parameters.
$\mathcal{L}_{CF}$ is the collaborative filtering loss depending on the base model which can be any existing recommender, and $\mathcal{L}_{BD}$ is the bidirectional distillation loss.
Lastly, $\lambda_{S \rightarrow T}$ and $\lambda_{T \rightarrow S}$ are the hyperparameters that control the effects of the distillation loss in each direction.
Within BD, the teacher and the student are collaboratively improved with each other based on their complementarity.
Also, during the training, the knowledge distilled between the teacher and the student gets gradually evolved along with the recommenders;
the improvement of the teacher leads to the acceleration of the student's learning, and the accelerated student again improves the teacher's learning.
Trained in the bidirectional way, both the teacher and the student are significantly improved compared to when being trained separately.
As a result, the student trained with BD outperforms the student model trained with the conventional distillation that relies on a pre-trained and fixed teacher.
\subsection{Distillation Loss}
We formalize the distillation loss that transfers the knowledge between the recommenders.
By following the original distillation loss that matches the class distributions of two classifiers for a given image \cite{kd15},
we design the distillation loss for the user $u$ as follows:
\begin{equation}
\begin{aligned}
\mathcal{L}_{BD}(\theta_T; \theta_S) &= \sum_{j \in \mathcal{RDS}_{S \rightarrow T}(\mathcal{I}_{u}^{-})} \mathcal{L}_{BCE}(\hat{r}^T_{uj}, \hat{r}^S_{uj})\\
\mathcal{L}_{BD}(\theta_S; \theta_T) &= \sum_{j \in \mathcal{RDS}_{T \rightarrow S}(\mathcal{I}_{u}^{-})} \mathcal{L}_{BCE}(\hat{r}^S_{uj}, \hat{r}^T_{uj}),
\end{aligned}
\end{equation}
where $\mathcal{L}_{BCE}(p, q) = q \log p + (1-q) \log (1-p)$ is the binary cross-entropy loss,
$\hat{r}_{uj} = P(r_{uj}=1 | u,j)$ is the prediction of a recommender and $\mathcal{RDS}_{*}(\mathcal{I}_{u}^{-})$ is a set of the unobserved items sampled by the rank discrepancy-aware sampling.
$\hat{r}_{uj}$ is computed by $\sigma(z_{uj}/T)$
where $\sigma(\cdot)$ is the sigmoid function, $z_{uj}$ is the logit, and $T$ is the temperature that controls the smoothness.
Our distillation loss is a binary version of the original KD loss function.
Similar to the original KD loss transferring the knowledge of class probabilities, our loss transfers a user's potential positive and negative preferences on the unobserved items.
Specifically, in the binary training set, the unobserved interaction $r_{uj}$ is only labeled as “0”.
However, as mentioned earlier, it is ambiguous whether the user actually dislikes the item or potentially likes the item.
Through the distillation, each recommender can get the other's opinion of how likely (and unlikely) the user would be interested in the item, and such information helps the recommenders to better cope with the ambiguous nature of RS.
\subsection{Rank Discrepancy-aware Sampling}
We propose the rank discrepancy-aware sampling scheme that decides what knowledge to be transferred for each
distillation direction.
As we observed in Section 2, most of the items ranked highly by the teacher are already ranked highly by the student and vice versa.
Thus, the existing methods \cite{cd19,rd18} that simply choose the high-ranked items cannot give enough information to the other recommender.
Moreover, for effective bidirectional distillation, the performance gap between the teacher and the student should be carefully considered in deciding what knowledge to be transferred, as it is obvious that not all knowledge of the student is helpful to improve the teacher.
In this regard, we develop a sampling scheme based on the rank discrepancy of the teacher and the student, and tailor it differently for each distillation direction with consideration of their different capacities.
The underlying idea of the scheme is that each recommender can get informative knowledge by focusing on the items ranked highly by the other recommender, but ranked lowly by itself.
The sampling strategy for each distillation direction is defined as follows:
\vspace{3pt}
\noindent
\textbf{Distillation from the teacher to the student.}
As the teacher has a much better overall performance than the student, the opinion of the teacher should be considered more reliable than that of the student in most cases.
Thus, for this direction of the distillation, we make the student follow the teacher's predictions on many rank-discrepant items.
Formally, for each user $u$, the probability of an item $i$ to be sampled is computed as follows:
\begin{equation}
p_{T \rightarrow S}(i) \propto tanh(\text{max}((rank_S(i) - rank_T(i)) \cdot \epsilon_t, 0)),
\end{equation}
where $rank_T(i)$ and $rank_S(i)$ denote the ranks assigned by the teacher and the student on the item $i$ for the user $u$ respectively, and $rank_{*}(i)=1$ is the highest ranking\footnote{we omit the superscript $u$ from $rank_{*}^{u}(i)$ for the simplicity.}.
We use a hyper-parameter $\epsilon_t$ $(> 0)$ to control the smoothness of the probability.
With this probability function, we sample the items ranked highly by the teacher but ranked lowly by the student.
Since $tanh(\cdot)$ is a saturated function, items with rank discrepancy above a particular threshold would be sampled almost uniformly.
As a result, the student learns the teacher's broad knowledge of most of the rank-discrepant items.
\vspace{3pt} \noindent \textbf{Distillation from the student to the teacher.}
As shown in Section 2, the teacher is not always superior to the student, especially for RS.
That is, the teacher can be also further improved by learning from the student.
However, at the same time, the large performance gap between the teacher and the student also needs to be considered for effective distillation.
For this direction of the distillation, we make the teacher follow the student's predictions on only a few selectively chosen rank-discrepant items.
The distinct probability function is defined as follows:
\begin{equation}
p_{S \rightarrow T}(i) \propto exp((rank_T(i) - rank_S(i)) \cdot \epsilon_e),
\end{equation}
where $\epsilon_e$ $(> 0)$ is a hyper-parameter to control the smoothness of the probability.
We use the exponential function to put particular emphasis on the items that have large rank discrepancies.
Therefore, this probability function enables the teacher to follow the student's predictions only on the rank-discrepant items that the student has very high confidence in.
\subsection{Model Training}
\begin{algorithm}[t]
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{Training data $\mathcal{D}$, the number of total epochs $e$, rank updating period $p$}
\Output{Teacher model $(\theta_T)$, Student model $(\theta_S)$}
Warm up $\theta_T$ and $\theta_S$ with only $\mathcal{L}_{CF}$\\
\For{$t=0,1,...,(e-1)$}{
\If{$t \text{ } \% \text{ } p == 0$}{
Teacher and Student update their recommendation lists
}
\For{$(u,i) \in \mathcal{D}$}{
\BlankLine
\tcc{Train Teacher}
Draw $\mathcal{RDS}_{S \rightarrow T}(\mathcal{I}_{u}^{-})$ with probability $p_{S \rightarrow T}(\cdot)$ \\
Compute $\mathcal{L}_{CF}(\theta_T)$ and $\mathcal{L}_{BD}(\theta_T; \theta_S)$ \\
Update $\theta_T$
\BlankLine
\tcc{Train Student}
Draw $\mathcal{RDS}_{T \rightarrow S}(\mathcal{I}_{u}^{-})$ with probability $p_{T \rightarrow S}(\cdot)$ \\
Compute $\mathcal{L}_{CF}(\theta_S)$ and $\mathcal{L}_{BD}(\theta_S; \theta_T)$ \\
Update $\theta_S$
}
}
\caption{Bidirectional Distillation Framework.}
\end{algorithm}
Algorithm 1 describes a pseudo code for the end-to-end training process within BD framework.
The training data $\mathcal{D}$ consists of observed interactions $(u,i)$.
First, the model parameters $\theta_T$ and $\theta_S$ are warmed up only with the collaborative filtering loss (line 1), as the predictions during the first few epochs are very unstable.
Second, we make the recommenders produce the recommendation lists for the subsequent sampling.
Since it is time-consuming to produce the recommendation lists every epoch, we conduct this step every $p$ epochs (line 3-4).
Next, we decide what knowledge to be transferred in each distillation direction via the rank discrepancy-aware sampling.
It is worth noting that the unobserved items sampled by the rank discrepancy-aware sampling can be used also for the collaborative filtering loss.
Finally, we compute the losses with the sampled items for the teacher and the student, respectively, and update the model parameters.
\subsection{Experimental Setup}
\subsubsection{Datasets}
We use three real-world datasets: CiteULike\footnote{https://github.com/changun/CollMetric/tree/master/citeulike-t} \cite{CUL13}, Foursquare\footnote{https://sites.google.com/site/yangdingqi/home/foursquare-dataset} (Tokyo Check-in) \cite{FS14} and Yelp\footnote{https://github.com/hexiangnan/sigir16-eals/blob/master/data/yelp.rating} \cite{yelp16}.
We only keep users who have at least five ratings for CiteULike and Foursquare, ten ratings for Yelp as done in \cite{ncf17, bpr09}.
Data statistics after the preprocessing are presented in Table 1.
We also report the experimental results on ML100K and AMusic, which are used for CD \cite{cd19}, in Appendix for the direct comparison.
\begin{table}[t]
\renewcommand{\arraystretch}{0.7}
\caption{Data Statistics}
\begin{tabular}{ccccc}
\toprule
Dataset & \#Users & \#Items & \#Ratings & Sparsity \\
\midrule
CiteULike & 5,219 & 25,187 & 130,788 & 99.90\% \\
Foursquare & 2,293 & 61,858 & 537,167 & 99.62\% \\
Yelp & 25,677 & 25,815 & 730,623 & 99.89\% \\
\bottomrule
\end{tabular}
\end{table}
\subsubsection{Evaluation Protocol and Metrics}
We adopt the widely used \textit{leave-one-out} evaluation protocol.
For each user, we hold out the last interacted item for testing and the second last interacted item for validation as done in \cite{cd19, ncf17}.
If there is no timestamp in the dataset, we randomly take two observed items for each user.
Then, we evaluate how well each method can rank the test item higher than all the unobserved items for each user (i.e., $\mathcal{I}_u^{-}$).
Note that instead of randomly choosing a predefined number of candidates (e.g., 99), we adopt the full-ranking evaluation that uses all the unobserved items as candidates.
Although it is time-consuming, it enables a more thorough evaluation compared to using random candidates \cite{fullrank20, cd19}.
As we focus on top-$K$ recommendation for implicit feedback, we employ two widely used metrics for evaluating the ranking performance of recommenders: Hit Ratio (H@$K$) \cite{hr16} and Normalized Discounted Cumulative Gain (N@$K$) \cite{NDCG02}.
H@$K$ measures whether the test item is present in the top-$K$ list and N@$K$ assigns a higher score
to the hits at higher rankings in the top-$K$ list.
We compute those two metrics for each user, then compute the average score.
Lastly, we report the average value of five independent runs for all methods.
\subsubsection{Base Models}
BD is a model-agnostic framework applicable for any top-$K$ RS.
We validate BD with three base models that have different model architectures and learning strategies.
Specifically, we choose two widely used deep learning models and one latent factor model as follows:
\begin{itemize}
\item \textbf{NeuMF \cite{ncf17}}: A deep recommender that adopts Matrix Factorization (MF) and Multi-Layer Perceptron (MLP) to capture complex and non-linear user-item relationships.
NeuMF uses the point-wise loss function for the optimization.
\item \textbf{CDAE \cite{cdae16}}:
A deep recommender that adopts Denoising Autoencoders (DAE) \cite{dae08} for the collaborative filtering.
CDAE uses the point-wise loss function for the optimization.
\item \textbf{BPR \cite{bpr09}}: A learing-to-rank recommender that adopts MF \cite{mf09} to model the user-item interaction.
BPR uses the pair-wise loss function for the optimization under the assumption that observed items are more preferred than unobserved items.
\end{itemize}
\input{5exzmaintable}
\subsubsection{Methods Compared}
The proposed framework is compared with the state-of-the-art KD methods for top-$K$ RS.
\begin{itemize}
\item \textbf{Ranking Distillation (RD) \cite{rd18}}: A pioneering KD method for top-$K$ RS.
RD makes the student give high scores on top-ranked items by the teacher.
\item \textbf{Collaborative Distillation (CD) \cite{cd19}}: A state-of-the-art KD method for top-$K$ RS.
CD makes the student imitate the teacher's scores on the items ranked highly by the teacher.
\end{itemize}
\subsubsection{Implementation Details}
For all the base models and baselines, we use PyTorch \cite{pytorch19} for the implementation.
For each dataset, hyperparameters are tuned by using grid searches on the validation set.
We use Adam optimizer \cite{adam14} with L2 regularization and the learning rate is chosen from $\{$0.00001, 0.0001, 0.001, 0.002$\}$, and we set the batch size as 128.
For NeuMF, we use 2-layer MLP for the network.
For CDAE, we use 2-layer MLP for the encoder and the decoder, and the dropout ratio is set to 0.5.
The number of negative samples is set to 1 for NeuMF and BPR, $5*|\mathcal{I}_{u}^{+}|$ for CDAE as suggested in the original paper \cite{cdae16}.
For the distillation, we adopt as many learning parameters as possible for the teacher model until the ranking performance is no longer increased on each dataset.
Then, we build the student model by employing only one-tenth of the learning parameters used by the teacher.
The number of model parameters of each base model is reported in Table 3.
For KD competitors (i.e., RD, CD), $\lambda_{KD}$ is chosen from $\{$0.01, 0.1, 0.5, 1$\}$, the number of items sampled for the distillation is chosen from $\{$10, 15, 20, 30, 40, 50$\}$, and the temperature $T$ for logits is chosen from $\{$1, 1.5, 2$\}$.
For other hyperparameters, we use the values recommended from the public implementation and the original papers \cite{rd18, cd19}.
For BD, $\lambda_{T \rightarrow S}$ and $\lambda_{S \rightarrow T}$ are set to 0.5, the number of items sampled by rank discrepancy-aware sampling is chosen from $\{$1, 5, 10$\}$, $\epsilon_t$ is chosen from $\{10^{-2}, 10^{-3}, 10^{-4}, 10^{-5}\}$, $\epsilon_e$ is chosen from $\{10^{-3}, 10^{-4}, 10^{-5}\}$, the temperature $T$ for logits is chosen from $\{$2, 5, 10$\}$ and the rank updating period $p$ (in Algorithm 1) is set to 10.
\subsection{Performance Comparison}
Table 2 shows the recommendation performance of each KD method on three real-world datasets and three different base models.
In Table 2, "Teacher" and "Student" indicate the base models trained separately without any distillation technique, "BD-Teacher" and "BD-Student" are the teacher and the student trained simultaneously with BD.
"CD" and "RD" denote the student trained with CD and RD, respectively.
We analyze the experimental results from various perspectives.
\input{5exzmodelparams}
\begin{figure*}[t]
\begin{subfigure}[t]{0.495\linewidth}
\includegraphics[width=0.5\linewidth]{figure/ModelSize_NeuMF_T.png}
\includegraphics[width=0.5\linewidth]{figure/ModelSize_BPR_T.png}
\caption{Performance of teacher with student of varying size}
\end{subfigure}
\begin{subfigure}[t]{0.495\linewidth}
\includegraphics[width=0.5\linewidth]{figure/ModelSize_NeuMF_S.png}
\includegraphics[width=0.5\linewidth]{figure/ModelSize_BPR_S.png}
\caption{Performance of student with teacher of varying size}
\end{subfigure}
\caption{Effect of the model size on BD. The blue dashed line indicates the performance of the model trained separately.}
\end{figure*}
We first observe that the teacher recommender is consistently improved by the knowledge transferred from the student with BD by up to 20.79\% (for H@50 on CiteULike).
This result verifies our claim that the knowledge of the student also could be useful for improving the teacher.
Also, this result strongly indicates that the distillation process of BD effectively resolves the challenge of a large performance gap, and successfully transfer the informative knowledge from the student.
In specific, the ranking discrepancy-aware sampling enables the teacher to focus on the items that the student has very high confidence in.
This strategy can minimize the adverse effects from the performance gap, making the teacher be further improved based on the complementary knowledge from the student.
Second, the student recommender is significantly improved by the knowledge transferred from the teacher with BD by up to 39.88\% (for N@50 on CiteULike).
Especially, the student trained with BD considerably outperforms the student trained with the existing KD methods (i.e., RD, CD) by up to 13.94\% (for H@100 on Yelp).
The superiority of BD comes from the two contributions;
BD makes the student follow the teacher’s predictions on the rank-discrepant items which are more informative than the merely high-ranked items.
Also, within BD, the improvement of the teacher leads to the acceleration of the student’s learning, and the accelerated student again improves the teacher’s learning.
With better guidance from the improved teacher, the student with BD can achieve superior performance than the student with the existing KD methods.
To verify the effectiveness of each contribution, we conduct in-depth analyses in the next sections.
Lastly, Table 3 shows the model size and online inference efficiency for each base model.
For making the inferences, we use PyTorch with CUDA on GTX Titan X GPU and Intel i7-7820X CPU.
As the student has only one-tenth of the learning parameters used by the teacher, the student requires less computational costs, thus achieves lower inference latency.
Moreover, the student trained with BD shows comparable or even better recommendation performance to the teacher (e.g., NeuMF and CDAE on Foursquare).
On CiteULike, we observe that the student achieves comparable performance by employing 20$\sim$30\% of learning parameters used by the teacher.
These results show that BD can be effectively adopted to train a small but powerful recommender.
Moreover, as mentioned earlier, BD is also applicable in setting where there is no constraint on the model size or inference time.
Specifically, BD can be adopted to maximize the performance of a large recommender with numerous learning parameters (i.e., the teacher).
\subsection{Model Size Analysis}
We control the size of the teacher and the student to analyze the effects of the capacity gap between two recommenders on BD.
We report the results of a deep model and a latent factor model (i.e., NeuMF and BPR) on CiteULike.
Figure 3a shows the performance of the teacher trained with the student of varying sizes and Figure 3b shows the performance of the student trained with the teacher of varying sizes.
“S” indicates the size of the student (10\% of the teacher), “M” indicates the medium size (50\% of the teacher), and “T” indicates the size of the teacher.
Also, “fixed S” refers to the pre-trained student recommender with the size of "S", “fixed T” refers to the pre-trained teacher recommender with the size of "T".
Note that "fixed S/T" are not updated during the training (i.e., unidirectional knowledge distillation).
First, we observe that both the teacher and the student achieves the greatest performance gain when the capacity gap between two recommenders is largest;
the teacher shows the best performance with the smallest student (i.e., S in Fig. 3a) and the student performs best with the largest teacher (i.e., T in Fig. 3b).
As mentioned in Section 2, the teacher and the student have complementarity as some user-item relationships can be better captured without expensive computations.
In this regard, such complementarity can be maximized when the capacity gap between the two recommenders is large enough.
This result supports our claim that the performance gain comes from the different but reciprocal knowledge of two recommenders.
Also, it is worth noting that there are still performance improvements when two recommenders have identical sizes (i.e., T in Fig. 3a, S in Fig. 3b).
This can be understood as a kind of self-distillation effect \cite{born18, self19} when the teacher has the same size as the student.
Although they have very similar kinds of knowledge, they can still regularize each other, preventing its counterpart from being overfitted to a few observed interactions.
Second, the teacher is more improved when it is trained along with the learning student than when it is trained with the fixed student (i.e., S vs. fixed S in Fig. 3a).
Similarly, the student is more improved when it is trained together with the learning teacher than when it is trained with the fixed teacher (i.e., T vs. fixed T in Fig. 3b).
In the unidirectional distillation, the pre-trained recommender (i.e., fixed S/T) is no longer improved, thus always conveys the same knowledge during the training.
On the contrary, within BD, both recommenders are trained together, thus the knowledge distilled between the recommenders gets gradually evolved as they are improved during the training.
As a result, each recommender can be further improved based on the evolved knowledge of its counterparts.
This result again shows the superiority of our bidirectional distillation over the unidirectional distillation of the existing methods.
\subsection{Synchronization Analysis}
\begin{table}[t]
\renewcommand{\arraystretch}{0.8}
\caption{Average Rank Difference before and after the training with BD.}
\begin{tabular}{ccccc}
\toprule
\multicolumn{2}{c}{Base Model} & CiteULike & Foursquare & Yelp\\
\midrule
\multirow{2}{*}{NeuMF} & before BD & 0.1944 & 0.1020 & 0.0918\\
& after BD & 0.1323 & 0.0957 & 0.0824\\
\midrule
\multirow{2}{*}{CDAE} & before BD & 0.1190 & 0.0963 & 0.0544\\
& after BD & 0.0871 & 0.0883 & 0.0472\\
\midrule
\multirow{2}{*}{BPR} & before BD & 0.1380 & 0.1180 & 0.0560\\
& after BD & 0.1064 & 0.1082 & 0.0511\\
\bottomrule
\end{tabular}
\end{table}
\begin{figure}[t]
\begin{subfigure}[t]{\linewidth}
\includegraphics[width=0.5\linewidth]{figure/After_NeuMF_cul.png}
\includegraphics[width=0.5\linewidth]{figure/After_NeuMF_fs.png}
\end{subfigure}
\caption{Rank difference after the training with BD.}
\end{figure}
We perform in-depth analysis to investigate how well the teacher and the student learn each other's complementary knowledge within the proposed framework.
Specifically, we evaluate how much synchronized the two recommenders are within BD, which shows that they are improved based on each other's complementarity.
To quantify the degree of the synchronization, we define (normalized) Average Rank Difference as follows:
\begin{equation}
\text{Average Rank Diff.} = \frac{1}{n \cdot m} \sum_{\mathcal{D}_{test}} \abs{\text{Rank Diff.}^{u}(i)},
\end{equation}
where $n$ and $m$ are the numbers of users and items, respectively, and $\mathcal{D}_{test}$ is the test set that contains the held-out observed interaction for each user.
The rank difference ($\text{Rank Diff.}^{u}(i)$) is defined in Equation 1.
Table 4 shows the change in the average rank difference after the training with BD.
We observe that the average rank difference gets consistently decreased on all the datasets.
This result indicates that the teacher and the student get synchronized by transferring their knowledge to each other.
Figure 4 shows the rank difference between the teacher and the student after the training with BD.
We adopt NeuMF as the base model as done in Section 2.
Note that the (normalized) average rank difference is proportional to the extent of the blue area.
We observe that the blue area in Figure 4 shrinks compared to that in Figure 1 after the training with BD.
Interestingly, we observe that the teacher is significantly improved on CiteULike dataset (by up to 20.79\%) which has the largest average rank difference change before and after the training with BD.
The large change indicates that the two recommenders get synchronized well during the training, which leads to significant improvements based on each other's knowledge.
\subsection{Sampling Scheme Analysis}
We examine the effects of diverse sampling schemes on the performance of BD to verify the superiority of the proposed sampling scheme.
Note that the schemes decide what knowledge to be distilled within BD.
We compare five different sampling schemes as follows:
1) Rank discrepancy-aware sampling,
2) Rank-aware sampling \cite{cd19},
3) Top-$N$ selection \cite{rd18},
4) Uniform sampling,
5) \textit{Swapped} rank discrepancy-aware sampling.
The rank discrepancy-aware sampling, which is our proposed scheme, focuses on the rank-discrepant items between the teacher and the student.
On the other hand, the rank-aware sampling (adopted in CD) and top-$N$ selection (adopted in RD) focus on the items ranked highly by one recommender.
The uniform sampling randomly chooses items from the entire recommendation list.
Finally, the swapped rank discrepancy-aware sampling, which is the ablation of the proposed scheme, swaps the sampling probability function of each distillation direction;
we make the teacher follow the student on most of the rank-discrepant items (with \textit{tanh}), and make the student imitate only a few predictions of the teacher (with \textit{exp}).
\begin{table}[t]
\small
\renewcommand{\arraystretch}{1}
\caption{Recommendation performance (H@50) of BD with different sampling schemes on CiteULike. Numbers in bold face are the best results.}
\begin{tabular}{cc cc}
\toprule
Base Model & Sampling Scheme & Teacher & Student\\
\midrule
\multirow{5}{*}{NeuMF} & Rank discrepancy-aware & \textbf{0.1722} & \textbf{0.0924} \\
& Rank-aware \cite{cd19} & 0.1617 & 0.0812 \\
& Top-$N$ selection \cite{rd18} & 0.1480 & 0.0766 \\
& Uniform & 0.1590 & 0.0726 \\
& Swapped rank discrepancy-aware & 0.1512 & 0.0747 \\
\midrule
\multirow{5}{*}{CDAE} & Rank discrepancy-aware & \textbf{0.1983} & \textbf{0.0943} \\
& Rank-aware \cite{cd19} & 0.1818 & 0.0891 \\
& Top-$N$ selection \cite{rd18} & 0.1757 & 0.0870 \\
& Uniform & 0.1788 & 0.0819 \\
& Swapped rank discrepancy-aware & 0.1733 & 0.0851 \\
\midrule
\multirow{5}{*}{BPR} & Rank discrepancy-aware & \textbf{0.1479} & \textbf{0.0853} \\
& Rank-aware \cite{cd19} & 0.1367 & 0.0805 \\
& Top-$N$ selection \cite{rd18} & 0.1264 & 0.0772 \\
& Uniform & 0.1306 & 0.0749 \\
& Swapped rank discrepancy-aware & 0.1281 & 0.0803 \\
\bottomrule
\end{tabular}
\end{table}
Table 5 shows the performance of BD with different sampling schemes.
We first observe that the proposed scheme achieves the best result both for the teacher and the student.
This result verifies the effectiveness of two components of the proposed scheme: 1) sampling based on the rank discrepancy, 2) probability functions differently designed for each distillation direction.
Specifically, we can see the effectiveness of rank discrepant-based sampling by comparing our sampling scheme with rank-aware sampling.
As observed in Section 2, the items merely ranked highly by the teacher may be not informative enough to fully enhance the student.
With the proposed scheme, BD focuses on the rank-discrepant items, thus can further improve the recommenders.
Also, we observe that swapping the probability function of each distillation direction (i.e., the swapped rank discrepancy-aware) significantly degrades the performance.
This result strongly indicates that each probability function is well designed to effectively cope with the large performance gap of two recommenders.
\subsection{Hyperparameter Analysis}
In this section, we provide thorough analyses that examine the effects of important hyperparameters on BD.
For the sake of the space, we report the results of NeuMF on CiteULike dataset.
We observe similar tendencies with other base models and datasets.
\begin{figure}[t]
\includegraphics[width=\linewidth, scale=1.7]{figure/lambdaTS.png}
\caption{Recommendation performance (H@50) of teacher and student with varying lambda.}
\end{figure}
\begin{figure}[t]
\begin{subfigure}[t]{\linewidth}
\includegraphics[width=0.5\linewidth]{figure/epsilonT.png}
\includegraphics[width=0.5\linewidth]{figure/epsilonS.png}
\end{subfigure}
\caption{Recommendation performance (H@50) of teacher and student with varying smoothing factor.}
\end{figure}
First, Figure 5 shows the performance of the teacher and the student with varying $\lambda_{T \rightarrow S}$ and $\lambda_{S \rightarrow T}$ which control the effects of the distillation losses.
The best performance is achieved when both $\lambda_{T \rightarrow S}$ and $\lambda_{S \rightarrow T}$ are around 0.5.
Also, we observe that both the recommenders are considerably improved when $\lambda_{T \rightarrow S}$ and $\lambda_{S \rightarrow T}$ have similar values (i.e., diagonal entries).
Moreover, we observe that the performance of the student is more robust with respect to $\lambda$ than that of the teacher.
We believe that this is because the student has a lower overall performance than the teacher, thus can easily take advantage of the teacher in broad settings.
\begin{figure}[t]
\begin{subfigure}[t]{\linewidth}
\includegraphics[width=0.5\linewidth]{figure/updateT.png}
\includegraphics[width=0.5\linewidth]{figure/updateS.png}
\end{subfigure}
\caption{Recommendation performance (H@50) of teacher and student with varying updating period.}
\end{figure}
Second, Figure 6 shows the performance of the teacher and the student with varying $\epsilon_t$ and $\epsilon_e$ which control the smoothness of the probability functions in rank discrepancy-aware sampling.
When $\epsilon_t$ and $\epsilon_e$ are small, the probability functions become smooth.
The best performance is achieved when both parameters are around $10^{-4}$.
When $\epsilon_e$ is bigger than $10^{-3}$, the probability $p_{S \rightarrow T}(\cdot)$ gets too sharp.
Thus, the teacher cannot fully learn the student's complementary knowledge, which leads to degraded performance.
Lastly, Figure 7 shows the performance of the teacher and the student with varying rank updating period $p$ (in Algorithm 1).
Since it is time-consuming to generate the recommendation list every epoch, we update the recommendation list of the two models every $p$ epoch.
In this paper, we use $p=10$ which shows the comparable performance to the upper-bound ($p=1$).
\section{Appendix}
In this section, we report the experimental results on ML100K and AMusic, which are used for CD \cite{cd19}, for the direct comparison with CD.
We do not include this result in our main table, because we consider those datasets are relatively small to simulate the real-world evaluation (ML100K has only 943 users and 1682 items).
We adopt CDAE as the base model and we employ 2-layer MLP for the encoder and the decoder of CDAE.
Experimental results on their experimental settings are as follows:
\begin{table}[h]
\small
\caption{Performance comparison with CD.}
\begin{tabular}{c cc cc}
\toprule
\multirow{2}{*}{Model} & \multicolumn{2}{c}{ML100K} & \multicolumn{2}{c}{AMusic} \\
& H@50 & N@50 & H@50 & N@50 \\
\midrule
Teacher & 0.3966 & 0.1259 & 0.1748 & 0.0533 \\
BD-Teacher & 0.4317 & 0.1442 & 0.1936 & 0.0607 \\
BD-Student & 0.4111 & 0.1321 & 0.1723 & 0.0545 \\
CD & 0.3786 & 0.1232 & 0.1650 & 0.0506 \\
Student & 0.3503 & 0.1078 & 0.1265 & 0.0416 \\
\midrule
\textit{Improv.T} & 8.85\% & 14.54\% & 10.76\% & 13.88\% \\
\textit{Improv.B} & 8.59\% & 7.22\% & 4.42\% & 7.71\% \\
\bottomrule
\end{tabular}
\end{table}
All results are the average of five iterations and statistically significant with p=0.01.
The proposed approach (BD) still outperforms the best competitor (CD) both on ML100K and AMusic.
Moreover, the ranking performance of the teacher also increases with BD.
\vspace{20cm}
\section{Introduction}
\input{1intro.tex}
\section{Analysis: Teacher vs Student}
\input{2Analysis}
\section{Problem Formulation}
\input{3pre}
\section{Method}
\input{4model.tex}
\section{Experiments}
\input{5ex.tex}
\section{Related Work}
\input{6rw}
\section{Conclusion}
\input{7con.tex}
\section*{Acknowledgements}
This work was supported by the NRF grant funded by the MSIT (South Korea, No. 2020R1A2B5B03097210), and the IITP grant funded by the MSIT (South Korea, No. 2018-0-00584, 2019-0-01906).
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
For a knot $K\subset S^3$ and a slope $\frac pq\in\mathbb Q\cup\{\infty\}$, let $S^3_{p/q}(K)$ be the manifold obtained by the $\frac pq$--surgery on $K$.
A slope $\frac pq$ is said to be {\it characterizing} for a given knot $K_0\subset S^3$ if whenever $S^3_{p/q}(K)\cong S^3_{p/q}(K_0)$ for a knot $K\subset S^3$, then $K=K_0$. Here, ``$\cong$'' stands for an orientation preserving homeomorphism. Obviously the trivial slope $\frac 10$ is never characterizing for any knot.
A long standing conjecture due to Gordon states that all nontrivial slopes are characterizing for the unknot. This conjecture was originally proved using Monopole Floer homology \cite{KMOSz}, and there were also proofs via Heegaard Floer homology \cite{OSzGenus,OSzRatSurg}. Based on work of Ghiggini \cite{Gh}, Ozsv\'ath and Szab\'o \cite{OSz3141} proved the same result for the trefoil knot and the figure--$8$ knot. To date, the unknot, the trefoil knot and the figure--$8$ knot are the only knots for which it is known that all but finitely many slopes are characterizing.
The next simplest knot is the torus knot $T_{5,2}$. It is reasonable to expect that all nontrivial slopes are characterizing for $T_{5,2}$. In \cite{NiZhang}, it is proved that a nontrivial slope $\frac pq$ is characterizing for $T_{5,2}$ unless $\frac pq$ is a negative integer or $-47\le p\le32$ and $1\le q\le8$. It is also known that the slopes \[8,9,10,11,12,\frac{17}2,\frac{19}2,\frac{21}2,\frac{23}2,\frac{28}3,\frac{29}3,\frac{31}3,\frac{32}3\] are characterizing \cite{BZ,Baker,GreeneCabling,NiZhangFinite}. Characterizing slopes for general torus knots have been studied in \cite{NiZhang,McCoySharp,McCoyTorus,NiCharSlope}.
In this paper, we further narrow down the range of possible non-characterizing slopes for $T_{5,2}$. Our main result is the
following theorem.
\begin{thm}\label{thm:T52}
A nontrivial slope $\frac pq$ is characterizing for $T_{5,2}$
if $\frac pq>-1$ and $\frac pq\notin\{0,1, \pm\frac12,\pm\frac13\}$.
\end{thm}
Recall that a rational homology $3$-sphere $Y$ is an {\it L-space} if the rank of $\widehat{HF}(Y)$ is equal to the order of $H_1(Y;\mathbb Z)$. For $T_{5,2}$, the surgery slopes which yield L-spaces are the ones greater than or equal to $3=2g(T_{5,2})-1$. Thus Theorem \ref{thm:T52} implies that all nontrivial L-space slopes of $T_{5,2}$ are characterizing for $T_{5,2}$.
\begin{cor}\label{cyclic-finite}If a nontrivial $p/q$-surgery, with $|p|\leq 9$, on a nontrivial knot $K$ in $S^3$ produces a manifold
of finite fundamental group, then $K$ is one of the torus knots
$T_{3,\pm2}$ and $T_{5,\pm2}$.
\end{cor}
\begin{proof}
By changing $K$ to its mirror image, we may also assume that $p/q$ is positive.
When $S^3_{p/q}(K)$ has cyclic fundamental group, i.e. when it is a lens space,
we may further assume that $p/q=p$ is an integer for otherwise $K$ is a torus knot \cite{CGLS}
and $S^3_{p/q}(K)$ is never a lens space when $p\leq 9$ and $q\geq 2$ \cite{Moser}.
Now \cite[Theorem 1.4]{Greene} implies that when $p\leq 9$,
the knot $K$ is a fibered knot of genus at most $2$, and $S^3_p(K)\cong S^3_p(T_{3,2})$ or $S^3_p(T_{5,2})$.
Now we use \cite{Gh} or Theorem~\ref{thm:T52} to get that $K=T_{3,2}$ or $T_{5,2}$.
When $S^3_{p/q}(K)$ has finite but non-cyclic fundamental group,
it is shown in \cite[Theorems~2, 3 and Table~1]{Doig} that either
$K$ is $T_{3,2}$ or $T_{5,2}$, or $p/q=7$ or $8$,
$S^3_{p/q}(K)\cong S^3_{p/q}(T_{5,2})$.
We may now apply Theorem~\ref{thm:T52} to conclude that $K=T_{5,2}$.
\end{proof}
The bound $9$ in Corollary~\ref{cyclic-finite} appears to be the best one could get for hyperbolic knots in $S^3$ with the current techniques. The $10$--surgery on $T_{4,3}$ is a spherical space form with non-cyclic fundamental group, and the current techniques could not rule out the possibility that the same surgery on a hyperbolic knot with the same knot Floer homology as $T_{4,3}$ yields the same manifold. Conjecturally on a hyperbolic knot $K$ in
$S^3$ if a nontrivial $\frac pq$-surgery yields a manifold of finite fundamental group, then $|p|\geq 17$, a bound which can be realized on the $(-2,3,7)$--pretzel knot.
Our proof of Theorem \ref{thm:T52}, given in the next section, is mainly based on
our earlier work \cite{NiZhang} and a recent paper of Baldwin--Hu--Sivek \cite{BHS}.
\vspace{5pt}\noindent{\bf Acknowledgements.}\quad The first author was
partially supported by NSF grant
number DMS-1811900.
\section{Proof of Theorem \ref{thm:T52}}\label{sect:proof}
For a knot $K$ in $S^3$, ${\Delta}_K(T)$
denotes the symmetric Alexander polynomial of $K$.
The following theorem and remark are \cite[Theorem~4.1]{NiZhang} and the remark after the proof.
\begin{thm}\label{thm:Fibered}
Suppose that $S^3_{p/q}(K)\cong S^3_{p/q}(T_{5,2})$ for a knot $K\subset S^3$ and a nontrivial slope $\frac pq$. Then one of the following two cases happens:
1) $K$ is a genus $(n+1)$ fibered knot for some $n\ge1$ with
\begin{equation}\label{eq:DeltaKn}
\Delta_K(T)=(T^{n+1}+T^{-(n+1)})-2(T^n+T^{-n})+(T^{n-1}+T^{1-n})+(T+T^{-1})-1.
\end{equation}
2) $K$ is a genus $1$ knot with $\Delta_K(T)=3T-5+3T^{-1}$.
Moreover, if $$\frac pq\in\left\{\frac pq>1\right\}\cup \left\{
\frac pq<-6, |q|\ge2\right\},$$ then the number $n$ in the first case must be $1$.
\end{thm}
\begin{rem}\label{addendum}
We have the following addendum to Theorem~\ref{thm:Fibered}:
\newline
(a) If $p$ is even, then case 2) of Theorem~\ref{thm:Fibered} cannot happen and in case 1) of Theorem~\ref{thm:Fibered}, the number $n$ must be odd.
\newline
(b) If $p$ is divisible by $3$, then case 2) cannot happen and in case 1), the number $n$ is not divisible by $3$.
\end{rem}
The following result is implicitly contained in \cite[Subsection~4.1]{NiZhang}. Background information about Heegaard Floer homology can be found in \cite[Section~3]{NiZhang}.
\begin{prop}\label{prop:SameHFK}
Suppose that $S^3_{p/q}(K)\cong S^3_{p/q}(T_{5,2})$ for a knot $K\subset S^3$ and a nontrivial slope $\frac pq>1$. Then
\begin{equation}\label{eq:HFKisom}
\widehat{HFK}(S^3,K)\cong\widehat{HFK}(S^3, T_{5,2})
\end{equation} as a bigraded group.
\end{prop}
\begin{proof}
In \cite[Subsection~4.1]{NiZhang}, it is proved that $\Delta_K=\Delta_{T_{5,2}}$ and $V_0(K)=V_1(K)=1$. By \cite[Proposition~4.2]{NiZhang},
\[
t_s(K)=V_s(K)+\mathrm{rank}H_{\mathrm{red}}(A_s^+).
\]
Since $t_0(K)=t_1(K)=1$ and $t_s(K)=0$ when $s>1$, we have
\[
HF_{\mathrm{red}}(A_s^+)=0\quad\text{whenever }s\ge0.
\]
So $K$ is an L-space knot. Since $\Delta_K=\Delta_{T_{5,2}}$, we get (\ref{eq:HFKisom}) by \cite{OSzLens}.
\end{proof}
\begin{rem}
In \cite[Proposition~4.7]{NiZhang}, it is proved that if $\frac pq<-6$ and $|q|\ge2$, then $\Delta_K=\Delta_{T_{5,2}}$. However, we cannot conclude that (\ref{eq:HFKisom}) holds. Algebraically, it is possible that $\widehat{HFK}(S^3,K,1)$ has rank $3$, while $HF^+(S^3_{p/q}(K))\cong HF^+(S^3_{p/q}(T_{5,2}))$.
\end{rem}
For any hyperbolic knot $K$ in $S^3$,
a result of Gabai and Mosher \cite{Mosher} states that the
complement of $K$ contains an essential lamination $\lambda$ which has an associated degeneracy locus $d(\lambda)$ in the form of $d(\lambda)=\frac mn$, $(m,n)\in\mathbb Z^2\setminus\{(0,0)\}$, such that if \[\Delta(\frac pq, d(\lambda)):=|pn-qm|\ge2,\] then $\lambda$
remains an essential lamination in $S^3_{p/q}(K)$.
Since $S^3_{1/0}(K)=S^3$ does not contain any essential lamination, it follows that
$d(\lambda)=m/0$ or $m/1$.
Furthermore we have
{\bf Fact (i)} If $\Delta(\frac pq, d(\lambda))\ge3$, then $S^3_{p/q}(K)$ is an irreducible, atoroidal and non-Seifert fibered manifold \cite[Theorem~2.5]{Wu}.
{\bf Fact (ii)} If $K$ is fibered and $d(\lambda)=m/1$ where $\lambda$ is the stable lamination of $K$, then $|m|\ge2$ \cite[Theorem~8.8]{GabaiProblems}. (Another proof of this was given in \cite{Roberts2}.)
\begin{prop}\label{prop:p/q<1}
Suppose that $S^3_{p/q}(K)\cong S^3_{p/q}(T_{5,2})$, $|\frac pq|<1$ and $\frac pq\notin\{0,\pm\frac12,\pm\frac13\}$. Then $K=T_{5,2}$.
\end{prop}
\begin{proof}
When $K$ is a torus knot, Proposition \ref{prop:p/q<1} holds by \cite[Lemma 4.2]{McCoyTorus}.
Next suppose that $K$ is a satellite knot.
Let $K'$ be a companion knot of $K$. Since $S^3_{p/q}(K)\cong S^3_{p/q}(T_{5,2})$ is atoroidal and irreducible (since $p/q\ne 10$) \cite{Moser}, the work of Gabai \cite{GSurg} implies that $K$ is a $0$--bridge or $1$--bridge braid in a tubular neighborhood $V$ of $K'$.
If $K$ is a $0$--bridge braid, then $K$ is a $(r, s)$-cable of $K'$
and the surgery slope $p/q$ must be $(qrs\pm 1)/q$ \cite{Go}.
If $K$ is a $1$--bridge braid, then it follows from \cite[Lemma 3.2]{G1bridge} that $\frac pq\in\mathbb Z$. In both cases we get a contradiction with the assumption that $|p/q|<1$.
So $K$ is hyperbolic.
By Theorem~\ref{thm:Fibered}, either $K$ is fibered or $g(K)=1$.
If $K$ is fibered and $d(\lambda)=\frac m0$, since $|q|\ge3$, $\Delta(d(\lambda),\frac pq)\ge3$. Hence $S^3_{p/q}(K)$
is not Seifert fibered by Fact (i) above, which contradicts the assumption $S^3_{p/q}(K)\cong S^3_{p/q}(T_{5,2})$.
So we may assume that when $K$ is fibered, $d(\lambda)=m/1$ for the stable lamination $\lambda$ of $K$, and therefore $|m|\geq 2$ by Fact (ii) above.
Since $|\frac pq|<1$ and $|m|\ge2$, $\Delta(d(\lambda),\frac pq)\ge3$. Again by Fact (i) we get a contradiction with the assumption
$S^3_{p/q}(K)\cong S^3_{p/q}(T_{5,2})$.
So $K$ is hyperbolic and $g(K)=1$. By \cite[Theorem 1.5]{BGZ} $0<|p|\leq 3$
since $S^3_{p/q}(K)\cong S^3_{p/q}(T_{5,2})$ is a Seifert fibered space.
By Remark~\ref{addendum}, $|p|=1$.
Since $\frac pq\notin\{0,\pm\frac12,\pm\frac13\}$, we again have $\Delta(d(\lambda),\frac pq)\ge3$ (whether $d(\lambda)=\frac m0$ or $\frac m1$), which leads to a contradiction as in the preceding paragraph.
\end{proof}
\begin{prop}\label{prop:p/q>1}
Suppose that $S^3_{p/q}(K)\cong S^3_{p/q}(T_{5,2})$ for a nontrivial slope $\frac pq$ with $\frac pq>1$. Then $K=T_{5,2}$.
\end{prop}
\begin{proof}To get a contradiction we assume that $K\ne T_{5,2}$.
By Proposition \ref{prop:SameHFK}, we have $\widehat{HFK}(S^3,K)\cong\widehat{HFK}(S^3, T_{5,2})$.
In \cite[Section 3]{BHS}, Baldwin--Hu--Sivek proved that if $\widehat{HFK}(S^3,K)\cong\widehat{HFK}(S^3, T_{5,2})$ and $K\ne T_{5,2}$, then $K$ is a hyperbolic, doubly-periodic, genus two fibered knot,
the degeneracy locus of the stable lamination of $K$ is $4$, and moreover
there exists a pseudo-Anosov $5$-braid $\beta$ whose closure $B=\hat{\beta}$ is an unknot with braid axis $A$, such that $K$ is the lift of $A$ in the branched double cover $\Sigma(S^3,B)\cong S^3$.
Let $V$ be the exterior of $A$ in $S^3$ and $M_K$ the exterior of $K$ in $S^3$. Then $V$ is a solid torus
and $M_K$ is a double branched cover of $V$ with
$B$ as the branched set in $V$.
Let ${\tau}$ be the corresponding covering involution on $M_K$
and $U$ the branched set in $M_K$. Then $U$ is the fixed point set of ${\tau}$ which is a knot disjoint from ${\partial} M_K$, and $(M_K,U)/{\tau}=(V, B)$.
The restriction of ${\tau}$ on ${\partial} M_K$ is a free action--an order two
rotation along the longitude factor of ${\partial} M_K$.
We also use $M_K(p/q)$ to denote the surgery manifold $S^3_{p/q}(K)$
and similarly $V(p/q)$ for $S^3_{p/q}(A)$.
Note that the involution ${\tau}$ on $M_K$ extends to an involution
${\tau}_{p/q}$ on $M_K(p/q)$.
In fact if we let $N_{p/q}$ denote the filling solid torus
in forming $M_K(p/q)=M_K\cup N_{p/q}$ and let $K_{p/q}$ be the center circle of $N_{p/q}$,
then the fixed point set of ${\tau}_{p/q}$ is
\begin{equation}\label{fixed point set for p/q}
Fix({\tau}_{p/q})=\left\{\begin{array}{l}U,\;\;\;\;\;\;\;\;\;\;\;\;\;\;\mbox{if $p$ is odd}
\\U\cup K_{p/q}, \;\;\;\mbox{if $p$ is even }\end{array}\right.
\end{equation}
and
\begin{equation}\label{branched set for p/q}
M_K(p/q)/{\tau}_{p/q}=\left\{\begin{array}{l}V(p/2q),\;\;\;\;\;\;\;\mbox{if $p$ is odd}
\\V((p/2)/q), \;\;\;\mbox{if $p$ is even.}\end{array}\right.
\end{equation}
Let $Y$ be the exterior of $U$ in $M_K$ and $W$ the exterior of $B$ in $V$. Then
$Y$ is a free double cover of $W$.
Since $B$ is the closure of a pseudo-Anosov braid in $V$,
$W$ is hyperbolic.
Hence $Y$ is also hyperbolic.
Note that $M_K(p/q)\cong M_{T_{5,2}}(p/q)$ is a Seifert fibered space whose base orbifold
is $S^2(2, 5, d)$ with $d=|p-10q|$ if $p/q\ne 10$, and is a connected sum of
two lens spaces of orders $2$ and $5$ if $p/q=10$ \cite{Moser}.
Since $K$ is a doubly periodic knot, by \cite{HS} $M_K(p/q)$ is irreducible and thus $p/q\ne 10$, by \cite{WZ} $M_K(p/q)$ is not a lens space, and by \cite{MM} $M_K(p/q)$ is not a prism manifold.
Thus $M_K(p/q)$ is a Seifert fibered space
whose base orbifold is $S^2(2,5, d)$ with $d=|p-10q|>2$.
So there is a unique Seifert structure on $M_K(p/q)$.
By Thurston's Orbifold Theorem \cite{BP,CHK},
we may assume that the unique Seifert structure on $M_K(p/q)$
is ${\tau}_{p/q}$-invariant, i.e. ${\tau}_{p/q}$ sends every Seifert fiber to a Seifert fiber preserving the order of singularity.
Since the base orbifold of the Seifert fibered space $M_K(p/q)$
is orientable, the Seifert fibers of $M_K(p/q)$ can be coherently oriented.
If ${\tau}_{p/q}$ preserves the orientations of the Seifert fibers of $M_K(p/q)$, then
$Fix({\tau}_{p/q})$ consists of Seifert fibers (see \cite[Lemma 4.3]{BGZ}).
Since $K$ is hyperbolic, $K_{p/q}$ cannot be a component of $Fix({\tau}_{p/q})$.
Thus by Formula (\ref{fixed point set for p/q}), $p$ is odd and $Fix({\tau}_{p/q})=U$, and by Formula (\ref{branched set for p/q}),
$M_K(p/q)/{\tau}_{p/q}=V(p/2q)$.
Moreover, if we let $Y({\partial} M_K, p/q)$ denote the Dehn filling of $Y$ along the component
${\partial} M_K$ of ${\partial} Y$ with the slope $p/q$, and similarly define $W({\partial} V, p/2q)$,
then $Y({\partial} M_K, p/q)$ is Seifert fibered
and is a free double cover of
$W({\partial} V, p/2q)$. So the latter manifold $W({\partial} V, p/2q)$ is also Seifert fibered.
But $W({\partial} V, 1/0)$ is a solid torus (it is the exterior of the unknot $B$ in $S^3$).
We get a contradiction with \cite[Corollary 15]{MZ}.
Hence ${\tau}_{p/q}$ reverses the orientations of the Seifert fibers of $M_K(p/q)$. Since $p/q>1$, we may assume that both $p$ and $q$ are positive.
Let $\mu$ be the meridian of $K$, then $[\mu]$ generates $H_1(M_K)\cong\mathbb Z$ and $H_1(M_K(p/q))\cong\mathbb Z/p\mathbb Z$. Clearly ${\tau}_*[\mu]=[\mu]$, so
\begin{equation}\label{eq:InvId}
{\tau}_*=\mathrm{id} \text{ on }H_1(M_K),\qquad ({\tau}_{p/q})_*=\mathrm{id} \text{ on }H_1(M_K(p/q)).
\end{equation}
To apply a homological argument, we describe the Seifert fibered structure of $M_{T_{5,2}}(p/q)$ explicitly as follows.
Let $V_0\cup V_1$ be a standard genus one Heegaard splitting of $S^3$. We may assume that $T_{5,2}$ is embedded in $\partial V_0$
and is homologous to $5\mathcal L+2\mathcal M$, where $\mathcal L$ is the canonical longitude of $V_0$, and $\mathcal M$ the meridian of $V_0$. Let $\mu_0\subset \partial M_{T_{5,2}}$ be the meridian of $T_{5,2}$, let $\lambda_0$ be the canonical longitude of $T_{5,2}$ and
let $\mathcal C_i$ be the core of $V_i$, $i=1,2$. Then
$M_{T_{5,2}}$ is Seifert fibered with $C_0$ and $C_1$ as two singular fibers of order $5$ and $2$ respectively.
A regular fiber $\cal F$ of $M_{T_{5,2}}$ in ${\partial} M_{T_{5,2}}$ has slope $10{\mu}_0+{\lambda}_0$.
If $p/q\ne 10$, the Seifert structure of $M_{T_{5,2}}$ extends to one on $M_{T_{5,2}}(p/q)$ such that
the core $\mathcal C'$ of the filling solid torus is an order $d=|p-10q|$ singular fiber if $d>1$.
In $H_1(M_{T_{5,2}}(p/q))$, we have
\begin{equation}\label{eq:SingFiber}
[\mathcal F]=10[\mu_0],\quad[\mathcal C_0]=2[\mu_0],\quad [\mathcal C_1]=5[\mu_0],\quad [\mathcal C']=\pm q'[\mu_0],
\end{equation}
where $q'\in\mathbb Z$ satisfies that $qq'\equiv1\pmod p$.
Since ${\tau}_{p/q}$ sends a regular fiber to a regular fiber reversing
its orientation, we see, using (\ref{eq:InvId}) and (\ref{eq:SingFiber}), that $10[\mu_0]=-10[\mu_0]$ in $\mathbb Z/p\mathbb Z$. So $p|20$.
We claim that $p=1$ or $2$. Suppose otherwise that $p>2$.
We already know that $M_K(p/q)$ has three singular fibers of orders
$2, 5, d=|p-10q|>2$. If $d\ne 5$, then ${\tau}_{p/q}$ sends the order $d$ singular fiber $\mathcal C'$ to itself with opposite orientation.
Hence $q'[\mu_0]=-q'[\mu_0]$ in $\mathbb Z/p\mathbb Z$. Since $\gcd(p,q')=1$, we get $p|2$, which is not possible. So we must have $d=5$, which, together with the condition $p|20$, implies that $p/q=5$. By (\ref{eq:SingFiber}) the two order $5$ singular fibers $\mathcal C_0,\mathcal C'$ are homologous to $2[\mu_0]$ and $\pm[\mu_0]$ respectively. By (\ref{eq:InvId}), ${\tau}_{p/q}$ must send $\mathcal C_0$ to itself, and $\mathcal C'$ to itself. So $2[\mu_0]=-2[\mu_0]$ in $\mathbb Z/5\mathbb Z$, which is not possible.
Recall that $K$ has a degeneracy locus $4$. Since $\frac pq\le2$, we have $\Delta(\frac pq, 4)\ge3$ unless $\frac pq=2$.
By Fact (i), we only need to consider the case $\frac pq=2$.
By Formula (\ref{fixed point set for p/q}) the fixed point set of
${\tau}_2$ is $U\cup K_2$ and by Formula (\ref{branched set for p/q}),
$M_K(2)/{\tau}_{2}=V(1)$ which is $S^3$. Hence the branched set $B\cup K_{2}^*$ in
$M_K(2)/{\tau}_{2}=V(1)=S^3$
is a Montesinos link of two components \cite{Mon}, where $K_2^*$ is the image of $K_2$ under the map $M_K(2){\rightarrow} V(1)$, which is also the core of the filling solid torus of $V(1)$.
Note that $K_{2}^*$ is an unknot in $S^3$
while $B$ is the closure of a $5$-braid in the exterior of $K_{2}^*$ which is a solid torus.
Hence the linking number between $B$ and $K^*_{2}$ is $5$.
\begin{figure}[!ht]
\centerline{\includegraphics{involution2.eps}} \caption{The strong involution on $(S^3, T_{5,2})$
and the quotient}\label{involution on (5,2)}
\end{figure}
On the other hand $M_{T_{5,2}}(2)$ is a double branched cover of $S^3$ whose branched set in $S^3$ can be explicitly constructed
as follows.
The knot $T_{5,2}$ is strongly invertible
and the quotient spaces under the strong involution of $S^3$, a regular neighborhood of $T_{5,2}$
and the exterior, as well as the branched set
are as shown in Figure~\ref{involution on (5,2)}.
Furthermore $M_{T_{5,2}}(2)$ is a double branched cover of
$S^3$ whose branch set can be obtained by replacing the rational $1/0$-tangle in Figure~\ref{involution on (5,2)}~(2)
with the rational $2$-tangle.
The resulting branched set is the two-component
Montesinos link shown in Figure~\ref{branched set for slope 2}.
Note that $M_K(2)=M_{T_{5,2}}(2)$ has base orbifold $S^2(2,5,8)$ and so it has a unique Seifert fibration structure.
Since the number of singular fibers is $3$, the classification of Montesinos links \cite{BurdeZieschang} implies that $M_{T_{5,2}}(2)$ is the double branched cover of $S^3$
over a unique Montesinos link with $3$ rational tangles.
Therefore the link
shown in Figure~\ref{branched set for slope 2}
should be the link $B\cup K_2^*$.
However the linking number between the two components of the link in Figure~\ref{branched set for slope 2} is $3$, yielding a final contradiction with the early conclusion that the linking number between $B$ and $K_2^*$ is $5$.
\end{proof}
\begin{figure}[!ht]
\centerline{\includegraphics{branched2.eps}} \caption{The branched sets for the $2$-surgery on $T_{5,2}$}\label{branched set for slope 2}
\end{figure}
\begin{rem}
Proposition \ref{prop:p/q>1} can be proved without using the
degeneracy locus condition.
\end{rem}
Now the combination of Propositions \ref{prop:p/q<1} and \ref{prop:p/q>1} gives Theorem \ref{thm:T52}.
|
\section{Introduction}
Topological string amplitude is the generating function of
Gromov-Witten invariants which are usually rational numbers
according to their definitions \cite{HKKPTVVZ}. In 1998, Gopakumar
and Vafa \cite{GV0,GV1} found that topological string amplitude is
also the generating function of a series of integer-valued
invariants related to BPS counting in M-theory. Later, Ooguri and
Vafa \cite{OV} extended the above result to open string case, the
corresponding integer-valued invariants are named as OV invariants.
Furthermore, the OV invariants are further refined by Labasitida,
Mari\~no and Vafa in \cite{LMV}, then the resulted invariants are
called LMOV invariants \cite{LP}. An expanded physicist's
reconsideration of the GV and LMOV can be found in \cite{DW}. We
refer to \cite{Zhu3} for a brief review of the applications of these
integrality structures of topological strings in mathematics.
The open string LMOV invariants have been studied in many papers,
such as
\cite{LM1,LM2,LMV,LP,CLPZ,GKS,KS1,LZ,MMMS,MMMRSS,KRSS1,KRSS2}. Based
on the large $N$ duality of topological string and Chern-Simons
theory \cite{W,GV2,OV}, the open string LMOV invariants can be
approached by investigating the colored HOMFLYPT invariants of the
dual knots. For a knot $\mathcal{K}$, we use the notation
$n_{\mu,g,Q}(\mathcal{K})$ to denote the LMOV invariants of genus
$g$ with a boundary type $\mu$ which is a partition of a positive
integer, where $Q$ is a parameter describing the dependence of the
relative homology class of the dual Calabi-Yau geometry of the knot
$\mathcal{K}$. We refer to Section \ref{section2} for more detailed
definition of $n_{\mu,g,Q}(\mathcal{K})$.
In particular, when the knot $\mathcal{K}$ is a framed unknot
$U_\tau$ with framing $\tau\in \mathbb{Z}$, we have different ways
to compute its LMOV invariants according to string dualities which
have been proved in this situation \cite{LLZ,Zhou,EO}.
Then one can obtain several
explicit formulae \cite{GKS,LZ} for the genus zero LMOV invariants
$n_{\mu,0,Q}(U_\tau)$ of framed unknot $U_\tau$. It turns out that
these explicit formulae are certain combinations of the M\"obius
function and binomial numbers. Based on the integrality conjecture
for LMOV invariants, these formulae will give integers. However,
such an argument is not so obvious, a rigorous proof is required.
In this paper, we present a straightforward way to prove the
integrality of these formulae. We use the notation $n_{m,l}(\tau)$
to denote the LMOV invariants $n_{(m),0,l-\frac{m}{2}}(U_\tau)$ of
the framed unknot $U_\tau$ of genus $0$, where $m\geq 1$ and $l\geq
0$ are two integers. We have the following explicit formula
\cite{MV,LZ} for $n_{m,l}(\tau)$.
For $b\geq 0$ and $a\in \mathbb{Z}$, we introduce the notation
$\binom{a}{b}$ which is defined as follows
\makeatletter
\let\@@@alph\@alph
\def\@alph#1{\ifcase#1\or \or $'$\or $''$\fi}\makeatother
\begin{subnumcases}
{\binom{a}{b}=} 1, & $b=0$, \nonumber \\\nonumber \binom{a}{b},
&$b\geq 1$ and $a\ge 0$,\\\nonumber (-1)^{b}\binom{-a+b-1}{b},
&$b\geq 1$ and $a<0$.
\end{subnumcases}
\makeatletter\let\@alph\@@@alph\makeatother
We define
\begin{align*}
c_{m,l}(\tau)=-\frac{(-1)^{m\tau+m+l}}{m^2}\binom{m}{l}\binom{m\tau+l-1}{m-1},
\end{align*}
then
\begin{align} \label{integralformula}
n_{m,l}(\tau)=\sum_{d|m,d|l}\frac{\mu(d)}{d^2}c_{\frac{m}{d},\frac{l}{d}}(\tau),
\end{align}
where $\mu(d)$ denotes the M\"{o}bius functions.
In Section \ref{section3}, we prove that
\begin{theorem} \label{Thm1}
For any $\tau\in \mathbb{Z}$, $m\geq 1, l\geq 0$, we have
$n_{m,l}(\tau) \in \mathbb{Z}$.
\end{theorem}
\begin{remark}
In fact, such form of the formula (\ref{integralformula}) is very
general. For example, if we take some special values of $l$ or
$\tau$, it will give the formulae in \cite{GKS} (cf. the formulae
(1.4) and (1.5) in \cite{GKS}):
\begin{equation} \label{GKS1}
b_{K_p,r}^{-}=-\frac{1}{r^2}\sum_{d\mid r}\mu(\frac{r}{d}) \binom{3d-1}{d-1}, \qquad
b_{K_p,r}^{+}=\frac{1}{r^2}\sum_{d\mid r}\mu(\frac{r}{d}) \binom{(2|p|+1)d-1}{d-1}
\end{equation}
for $p\leq -1$ and
\begin{equation} \label{GKS2}
b_{K_p,r}^{-}=-\frac{1}{r^2}\sum_{d\mid r}\mu(\frac{r}{d}) (-1)^{d+1}\binom{2d-1}{d-1}, \qquad
b_{K_p,r}^{+}=\frac{1}{r^2}\sum_{d\mid r}\mu(\frac{r}{d}) (-1)^{d} \binom{(2p+2)d-1}{d-1}
\end{equation}
for $p\geq 2$. The above formulae (\ref{GKS1}) and (\ref{GKS2}) are
referred as the extremal BPS invariants of twist knots in
\cite{GKS}. Therefore, Theorem \ref{Thm1} implies the integrality of
formulae (\ref{GKS1}) and (\ref{GKS2}) immediately. Moreover, the
integrality of another special case of the formula
(\ref{integralformula}) was also proved in \cite{Zhu2}.
\end{remark}
Then, denoted by $n_{(m_1,m_2)}(\tau)$ the LMOV invariants
$n_{(m_1,m_2),0,\frac{m_1+m_2}{2}}(U_\tau)$ of the framed unknot
$U_\tau$ with $\mu=(m_1,m_2)$, $g=0$ and $Q=\frac{m_1+m_2}{2}$,
where $m_1\geq m_2\geq 1$, we obtain the following formula
\begin{align} \label{integralformula2}
n_{(m_1,m_2)}(\tau)& = \frac{1}{m_1+m_2}\sum_{d\mid m_1,d\mid m_2} \mu(d) (-1)^{(m_1+m_2)(\tau+1)/d} \\
& \cdot
\binom{(m_1\tau+m_1)/d-1}{m_1/d}\binom{(m_2\tau+m_2)/d}{m_2/d}\nonumber.
\end{align}
From this expression, we know that
$n_{(m_1,m_2)}(\tau)=n_{(m_2,m_1)}(\tau)$. With the similar method,
in Section \ref{section4}, we prove that
\begin{theorem} \label{Thm2}
For any $\tau\in \mathbb{Z}$ and $m_1,m_2\geq 1$, then
$n_{(m_1,m_2)}(\tau)\in \mathbb{Z}$.
\end{theorem}
Next, let $n_{m,g,Q}(\tau)$ be the LMOV invariants
$n_{(m),g,Q}(U_\tau)$ of higher genus $g$ with boundary condition
$\mu=(m)$. We define the following generating function
\begin{align*}
g_m(q,a)=\sum_{g\geq 0}\sum_{Q\in
\mathbb{Z}/2}n_{m,g,Q}(\tau)z^{2g-2}a^Q
\end{align*}
where $z=q^{\frac{1}{2}}-q^{-\frac{1}{2}}$.
Let
\begin{align*}
\mathcal{Z}_m(q,a)=(-1)^{m\tau}\sum_{|\nu|=m}\frac{1}{\mathfrak{z}_\nu}\frac{\{m\nu\tau\}}{\{m\}\{m\tau\}}\frac{\{\nu\}_a}{\{\nu\}}
\end{align*}
where $\mathfrak{z}_{\nu}=|Aut(\nu)|\prod_{i=1}^{l(\nu)}\nu_i$ and
$\{m\}$ denotes the quantum integer, see Section \ref{section2} for
introduction of the above notations.
By the definition of LMOV invariants in Section \ref{section2}, we
obtain the following expression
\begin{align*}
g_m(q,a)=\sum_{d|m}\mu(d)\mathcal{Z}_{m/d}(q^d,a^d).
\end{align*}
In Section \ref{section5}, we prove that
\begin{theorem} \label{Thm3}
For any $m\geq 1$, we have
$
g_m(q,a)\in z^{-2}\mathbb{Z}[z^2,a^{\pm \frac{1}{2}}],
$
where $z=q^{\frac{1}{2}}-q^{-\frac{1}{2}}$.
\end{theorem}
Therefore, Theorem \ref{Thm3} implies that $ n_{m,g,Q}(\tau)\in
\mathbb{Z} $ and moreover $n_{m,g,Q}(\tau)$ vanishes for large $g$
and $Q$.
\textbf{Acknowledgements.}
The authors are grateful to the referees for careful reading of the paper and
valuable comments and suggestions
which greatly improved the presentation of the content.
\section{LMOV invariants} \label{section2}
\subsection{Basic notations}
We first introduce some basic notations. A partition $\lambda$ is a finite sequence of positive integers
(\lambda_1,\lambda_2,..)$ such that $\lambda_1\geq
\lambda_2\geq\cdots$. The length of $\lambda$ is the total number of
parts in $\lambda$ and denoted by
$l(\lambda)$. The weight of $\lambda$ is defined by $|\lambda|
\sum_{i=1}^{l(\lambda)}\lambda_i$. The automorphism group of
$\lambda$, denoted by Aut($\lambda$), contains all the permutations
that
permute parts of $\lambda$ by keeping it as a partition. Obviously, Aut(
\lambda$) has the order $|\text{Aut}(\lambda)|=\prod_{i=1}^{l(\lambda)}m_i
\lambda)! $ where $m_i(\lambda)$ denotes the number of times that
$i$ occurs in $\lambda$. Define
$\mathfrak{z}_{\lambda}=|\text{Aut}(\lambda)|\prod_{i=1}^{\lambda}\lambda_i$.
In the following, we will use the notation $\mathcal{P}_+$ to denote
the set of all the partitions of positive integers. Let $0$ be the
partition of $0$, i.e. the empty partition. Define
$\mathcal{P}=\mathcal{P}_+\cup \{0\}$.
The power sum symmetric function of infinite variables
$\mathbf{x}=(x_1,..,x_N,..)$ is defined by
$p_{n}(\mathbf{x})=\sum_{i}x_i^n. $ Given a partition $\lambda$, we
define
$p_\lambda(\mathbf{x})=\prod_{j=1}^{l(\lambda)}p_{\lambda_j}(\mathbf{x}).
$ The Schur function $s_{\lambda}(\mathbf{x})$ is determined by the
Frobenius formula
\begin{align*}
s_\lambda(\mathbf{x})=\sum_{\mu}\frac{\chi_{\lambda}(\mu)}{\mathfrak{z}_\mu}p_\mu(\mathbf{x}),
\end{align*}
where $\chi_\lambda$ is the character of the irreducible
representation of
the symmetric group $S_{|\lambda|}$ corresponding to $\lambda$, we have
\chi_{\lambda}(\mu)=0$ if $|\mu|\neq |\lambda|$. The orthogonality
of character formula gives
\begin{align*}
\sum_\lambda\frac{\chi_\lambda(\mu)
\chi_\lambda(\nu)}{\mathfrak{z}_\mu}=\delta_{\mu \nu}.
\end{align*}
Let $n\in \mathbb{Z}$ and $\lambda,\mu,\nu$ denote the partitions.
We introduce the following notations
\begin{align*}
\{n\}_x=x^{\frac{n}{2}}-x^{-\frac{n}{2}}, \
\{\mu\}_{x}=\prod_{i=1}^{l(\mu)}\{\mu_i\}_x.
\end{align*}
In particular, let $\{n\}=\{n\}_q$ and $\{\mu\}=\{\mu\}_q$.
\subsection{LMOV invariants for framed knots}
Although the LMOV invariants are determined by the integrality
structure of topological open string partition function. Based on
the large $N$ duality of topological string and Chern-Simons theory
\cite{GV2,OV}, one can also introduce the LMOV invariants through
Chern-Simons theory of links/knots \cite{LM1,LM2,LMV,LP}.
Given a partition $\lambda$, we let
$\kappa_\lambda=\sum_{i=1}^{l(\lambda)}\lambda_i(\lambda_i-2i+1)$.
Let $\mathcal{K}_{\tau}$ be a knot with framing $\tau \in
\mathbb{Z}$. The framed colored HOMFLYPT invariant of
$\mathcal{K}_\tau$ is defined as follows
\begin{align*}
\mathcal{H}_\lambda(\mathcal{K}_\tau;q,a)=(-1)^{|\lambda|\tau}q^{\frac{\kappa_\lambda\tau}{2}}W_{\lambda}(\mathcal{K}_\tau;q,a),
\end{align*}
where $W_{\lambda}(\mathcal{K}_\tau;q,a)$ is the ordinary
(framing-independent) colored HOMFLYPT invariant of
$\mathcal{K}_\tau$, we refer to \cite{Zhu1} for the concrete
definition of $W_{\lambda}(\mathcal{K}_\tau;q,a)$.
Let $
\mathcal{Z}_{\mu}(\mathcal{K}_\tau;q,a)=\sum_{\lambda}\chi_{\lambda}(\mu)\mathcal{H}_{\lambda}(\mathcal{K}_\tau;q,a),
$ the Chern-Simons partition function of $\mathcal{K}_\tau$ is
defined by
\begin{align} \label{CSpartition}
Z_{CS}^{(S^3,\mathcal{K}_\tau)}(q,a,\mathbf{x})=\sum_{\lambda\in
\mathcal{P}}\mathcal{H}_\lambda(\mathcal{K}_\tau;q,a)s_{\lambda}(\mathbf{x})
=\sum_{\mu\in
\mathcal{P}}\frac{\mathcal{Z}_{\mu}(\mathcal{K}_\tau;q,a)}{\mathfrak{z}_\mu}p_{\mu}(\mathbf{x}).
\end{align}
Then we define the functions
$f_{\lambda}(\mathcal{K}_\tau;q,a)$ by
\begin{align*}
Z_{CS}^{(S^3,\mathcal{K}_\tau)}(q,a,\mathbf{x})=\exp\left(\sum_{d=1}^\infty\frac{1}{d}\sum_{\lambda\in
\mathcal{P}_+}f_{\lambda}(\mathcal{K}_\tau;q^d,a^d)s_{\lambda}(\mathbf{x}^d)\right).
\end{align*}
Let $
\hat{f}_{\mu}(\mathcal{K}_\tau;q,a)=\sum_{\lambda}f_{\lambda}(\mathcal{K}_\tau;q,a)M_{\lambda\mu}(q)^{-1},
$ where
\begin{align*}
M_{\lambda\mu}(q)=\sum_{\mu}\frac{\chi_{\lambda}(C_{\nu})\chi_{\mu}(C_{\nu})}{\frak{z}_{\nu}}\prod_{j=1}^{l(\nu)}(q^{\nu_{j}/2}-q^{-\nu_{j}/2}).
\end{align*}
Denote $z=q^{\frac{1}{2}}-q^{-\frac{1}{2}}$, then the LMOV
conjecture for framed knot $\mathcal{K}_\tau$ stated that
\cite{MV}, for any $\mu\in \mathcal{P}_+$, there are integers
$N_{\mu,g,Q}(\mathcal{K}_\tau)$ such that
\begin{align*}
\hat{f}_{\mu}(\mathcal{K}_\tau;q,a)=\sum_{g\geq 0}\sum_{Q\in
\mathbb{Z}/2}N_{\mu,g,Q}(\mathcal{K}_\tau)z^{2g-2}a^Q\in
z^{-2}\mathbb{Z}[z^{2},a^{\pm \frac{1}{2}}].
\end{align*}
Therefore,
\begin{align*}
g_{\mu}(\mathcal{K}_\tau;q,a)&=\sum_{\nu}\chi_{\nu}(\mu)\hat{f}_\nu(\mathcal{K}_\tau;q,a)\\\nonumber
&=\sum_{g\geq 0}\sum_{Q\in
\mathbb{Z}/2}n_{\mu,g,Q}(\mathcal{K}_\tau)z^{2g-2}a^Q\in
z^{-2}\mathbb{Z}[z^{2},a^{\pm \frac{1}{2}}].
\end{align*}
where
$n_{\mu,g,Q}(\mathcal{K}_\tau)=\sum_{\nu}\chi_{\nu}(\mu)N_{\nu,g,Q}(\mathcal{K}_\tau)$.
These conjectural integers $n_{\mu,g,Q}(\mathcal{K}_\tau)$ ( and
$N_{\nu,g,Q}(\mathcal{K}_\tau)$) are referred to as the LMOV
invariants in this paper. We refer to \cite{GKS,KS1,KRSS1,KRSS2} for
another slightly different introduction of LMOV invariants which are
referred to as OV invariants in \cite{Zhu2}. The integrality of
certain OV invariants for a large family of knots/links have been
proved in \cite{KRSS1,KRSS2} recently by using the knots-quivers
correspondence.
In order to get an explicit expression for
$g_{\mu}(\mathcal{K}_\tau;q,a)$, we introduce
$F_{\mu}(\mathcal{K}_\tau;q,a)$ through the following expansion
formula
\begin{align*}
\log(Z_{CS}^{(S^3,\mathcal{K}_\tau)}(q,a,\mathbf{x}))=\sum_{\mu\in
\mathcal{P}_+}F_{\mu}(\mathcal{K}_\tau;q,a)p_{\mu}(\mathbf{x}).
\end{align*}
Then, by formula (\ref{CSpartition}), we have
\begin{align*}
F_{\mu}(\mathcal{K}_\tau;q,a)=\sum_{n\geq
1}\sum_{\cup_{i=1}^{n}\nu^i=\mu}\frac{(-1)^{n-1}}{n}\prod_{i=1}^{n}\frac{\mathcal{Z}_{\nu^i}(\mathcal{K}_\tau;q,a)}{\mathfrak{z}_{\nu^i}}.
\end{align*}
\begin{remark}
For two partitions $\nu^1$ and $\nu^2$, the notation $\nu^1\cup
\nu^2$ denotes the new partition obtained by combining all the parts
in $\nu^1, \nu^2$. For example, if $\mu=(2,2,1)$, then the list of
all the pairs $(\nu^1,\nu^2)$ such that $\nu^1 \cup \nu^2 =(2,2,1)$
is
\begin{align*}
(\nu^1=(2), \nu^2=(2,1)),\ (\nu^1=(2,1), \nu^2=(2)),\\\nonumber
\
(\nu^1=(1), \nu^2=(2,2)), \ (\nu^1=(2,2), \nu^2=(1)). \
\end{align*}
\end{remark}
Finally, by using the M\"obius inversion formula, we obtain
\begin{align} \label{formulagmu}
g_{\mu}(\mathcal{K}_\tau;q,a)=\mathfrak{z}_{\mu}\frac{1}{\{\mu\}}\sum_{d|\mu}\frac{\mu(d)}{d}F_{\mu/d}(\mathcal{K}_\tau;q^d,a^d).
\end{align}
\begin{remark}
In the above discussion, for the sake of brevity, we only consider
the case of a framed knot, actually, the LMOV invariants can be
defined for any framed link.
\end{remark}
\subsection{LMOV invariants for framed unknot $U_\tau$}
In the following, we only consider the case of a framed unknot
$U_\tau$. In this situation, the large $N$ duality of topological
string and Chern-Simons theory \cite{MV} has been proved in
\cite{LLZ,Zhou}. Therefore, we can also compute the LMOV invariants
for framed unknot $U_\tau$ through the open topological string
theory.
We denote by $n_{m,l}(\tau)$ the LMOV invariants
$n_{(m),0,l-\frac{m}{2}}(U_\tau)$ of the framed unknot $U_\tau$,
where $g=0$ and $m\geq 1$, $l\geq 0$. According to the computations
shown in \cite{MV} ( or cf. pages 15-16 in \cite{LZ}), the explicit
closed formula for $n_{m,l}(\tau)$ is given by formula
(\ref{integralformula}). The integrality of $n_{m,l}(\tau)$ is given
by Theorem \ref{Thm1}.
Then, denoted by $n_{(m_1,m_2)}(\tau)$ the LMOV invariants
$n_{(m_1,m_2),0,\frac{m_1+m_2}{2}}(U_\tau)$ of the framed unknot
$U_\tau$ with $\mu=(m_1,m_2)$, $g=0$ and $Q=\frac{m_1+m_2}{2}$,
where $m_1\geq m_2\geq 1$. By using the computations in open
topological string theory (cf. pages 18-19 in \cite{LZ}), we obtain
the explicit closed formula for $n_{(m_1,m_2)}(\tau)$ which is given
by formula (\ref{integralformula2}). The integrality of
$n_{(m_1,m_2)}(\tau)$ is given by Theorem \ref{Thm2}.
Let $n_{m,g,Q}(\tau)$ be the LMOV invariants $n_{(m),g,Q}(U_\tau)$
of the framed unknot $U_\tau$. We consider the following generating
function
\begin{align*}
g_m(q,a)=\sum_{g\geq 0}\sum_{Q\in
\mathbb{Z}/2}n_{m,g,Q}(\tau)z^{2g-2}a^Q
\end{align*}
which can be computed by using formula (\ref{formulagmu}).
Considering the following function
\begin{align*}
\phi_{\mu,\nu}(x)=\sum_{\lambda}\chi_{\lambda}(\mu)\chi_{\lambda}(\nu)x^{\kappa_\lambda}.
\end{align*}
By Lemma 5.1 in \cite{CLPZ}, for $d\in \mathbb{Z}_+$, we have
\begin{align} \label{formulaphi}
\phi_{(d),\nu}(x)=\frac{\{d\nu\}_{x^2}}{\{d\}_{x^2}}.
\end{align}
By the expression of colored HOMFLYPT invariant for unknot (cf.
formula (4.6) in \cite{LP})
\begin{align*}
W_{\lambda}(U;q,a)=\sum_{\nu}\frac{\chi_{\lambda}(\nu)}{\mathfrak{z}_\nu}\frac{\{\nu\}_a}{\{\nu\}},
\end{align*}
we obtain
\begin{align*}
\mathcal{Z}_{\mu}(U_\tau;q,a)&=\sum_{\lambda}\chi_{\lambda}(\mu)\mathcal{H}_{\lambda}(U_\tau;q,a)\\\nonumber
&=(-1)^{|\mu|\tau}\sum_{\lambda}\chi_{\lambda}(\mu)q^{\frac{\kappa_\lambda
\tau}{2}}\sum_{\nu}\frac{\chi_{\lambda}(\nu)}{\mathfrak{z}_\nu}\frac{\{\nu\}_a}{\{\nu\}}\\\nonumber
&=(-1)^{|\mu|\tau}\sum_{\nu}\frac{1}{\mathfrak{z}_\nu}\phi_{\mu,\nu}(q^\frac{\tau}{2})\frac{\{\nu\}_a}{\{\nu\}}.
\end{align*}
In particular, for $\mu=(m)$ with $m\geq 1$, formula
(\ref{formulaphi}) implies that
\begin{align*}
\mathcal{Z}_{(m)}(U_\tau;q,a)=(-1)^{m\tau}\sum_{|\nu|=m}\frac{1}{\mathfrak{z}_\nu}\frac{\{m\nu\tau\}}{\{m\tau\}}\frac{\{\nu\}_a}{\{\nu\}}.
\end{align*}
For brevity, if we let
\begin{align*}
\mathcal{Z}_m(q,a)=\frac{1}{\{m\}}\mathcal{Z}_{(m)}(U_\tau;q,a)=(-1)^{m\tau}\sum_{|\nu|=m}\frac{1}{\mathfrak{z}_\nu}
\frac{\{m\nu\tau\}}{\{m\}\{m\tau\}}\frac{\{\nu\}_a}{\{\nu\}},
\end{align*}
then formula (\ref{formulagmu}) gives
\begin{align} \label{gm}
g_m(q,a)=\sum_{d|m}\mu(d)\mathcal{Z}_{m/d}(q^d,a^d).
\end{align}
Theorem \ref{Thm3} shows that
$ g_m(q,a)\in z^{-2}\mathbb{Z}[z^2,a^{\pm \frac{1}{2}}]$ for any $m\geq
1$.
\section{Proof of the Theorem \ref{Thm1}} \label{section3}
For nonnegative integer $n$ and prime number $p$, we introduce the
following function
\begin{equation} \label{functionfp}
f_p(n)=\prod_{i=1,p\nmid i}^n i = \frac{n!}{p^{[n/p]}[n/p]!}.
\end{equation}
Given a positive integer $k$, throughout this paper, we use the
notation $p^k \mid\mid n$ to denote that $p^k$ divides $n$, but
$p^{k+1}$ does not.
Before giving the proof of Theorem \ref{Thm1}, we first establish
several useful lemmas.
\begin{lemma}
For odd prime numbers $p$ and $\alpha\geq 1$ or for $p=2$, $\alpha\geq 2$,
we have $p^{2\alpha}\mid f_p(p^{\alpha} n)-f_p(p^{\alpha})^n$. For $p=2, \alpha=1$, $f_2(2n)\equiv (-1)^{[n/2]}\pmod{4}$
\label{lemma1}.
\end{lemma}
\begin{proof}
With $\alpha\geq 2$ or $p>2$, $p^{\alpha-1}(p-1)$ is even,
\begin{align*}
&f_p(p^\alpha n)-f_p(p^\alpha(n-1))f_p(p^\alpha) \\
&= f_p(p^\alpha(n-1))\left(\prod_{i=1,p\nmid i}^{p^\alpha} (p^{\alpha}(n-1)+i)-f_p(p^\alpha)\right) \\
& \equiv p^{\alpha}(n-1) f_p(p^\alpha(n-1))f_p(p^\alpha)\left(\sum_{i=1,p\nmid i}^{p^\alpha}\frac{1}{i}\right) \pmod{p^{2\alpha}} \\
& \equiv p^{\alpha}(n-1) f_p(p^\alpha(n-1))f_p(p^\alpha)\left(\sum_{i=1,p\nmid i}^{[p^\alpha/2]} (\frac{1}{i}+\frac{1}{p^{\alpha}-i})\right) \pmod{p^{2\alpha}} \\
& \equiv p^{\alpha}(n-1) f_p(p^\alpha(n-1))f_p(p^\alpha)\left(\sum_{i=1,p\nmid i}^{[p^\alpha/2]} \frac{p^{\alpha}}{i(p^\alpha-i)} \right)\equiv 0, \pmod{p^{2\alpha}}
\end{align*}
Thus the first part of the Lemma \ref{lemma1} is proved by
induction. For $p=2, \alpha=1$, the formula is straightforward.
\end{proof}
\begin{lemma}
For odd prime number $p$ and $m=p^\alpha a, l=p^\beta b$, $p\nmid a, p\nmid b$, $\alpha\geq 1, \beta \geq 0$, we have
\begin{equation*}
p^{2\alpha} \mid \binom{m}{l}\binom{m\tau+l-1}{m-1}-\binom{\frac{m}{p}}{\frac{l}{p}}\binom{\frac{m\tau+l}{p}-1}{\frac{m}{p}-1}
\end{equation*}
where for $\beta=0$, the second term is defined to be zero. \label{luo:lemma2-0}
\end{lemma}
\begin{proof}
\begin{align}
& \binom{m}{l}\binom{m\tau+l-1}{m-1}-\binom{\frac{m}{p}}{\frac{l}{p}}\binom{\frac{m\tau+l}{p}-1}{\frac{m}{p}-1} \nonumber \\
& = \binom{\frac{m}{p}}{\frac{l}{p}}\binom{\frac{m\tau+l}{p}-1}{\frac{m}{p}-1} \left( \frac{f_p(m)}{f_p(l)f_p(m-l)}\cdot \frac{f_p(m\tau+l)}{f_p(m)f_p(m(\tau-1)+l)}-1\right) \label{luo:eq5-0}
\end{align}
Write $\binom{\frac{m}{p}}{\frac{l}{p}}=\frac{m}{l}\binom{\frac{m}{p}-1}{\frac{l}{p}-1}$ and $\binom{\frac{m\tau+l}{p}-1}{\frac{m}{p}-1}=\frac{m}{m\tau +l} \binom{\frac{m\tau+l}{p}}{\frac{m}{p}}$, both are divisible by $p^{\max(\alpha-\beta,0)}$.
Each element of $\{m, l, m-l, m\tau+l, m(\tau-1)+l\}$ is divisible by $p^{\min(\alpha,\beta)}$, so by Lemma~\ref{lemma1},
\begin{equation}
\frac{f_p(m)}{f_p(l)f_p(m-l)}\cdot \frac{f_p(m\tau+l)}{f_p(m)f_p(m(\tau-1)+l)}-1 \label{luo:term1-0}
\end{equation}
is divisible by $p^{2\min(\alpha,\beta)}$ (including the case
$\beta=0$) in $p$-adic number field. Thus (\ref{luo:eq5-0}) is
divisible by
$p^{2\max(\alpha-\beta,0)+2\min(\alpha,\beta)}=p^{2\alpha}$.
\end{proof}
\begin{lemma}
For $m=2^\alpha a, l=2^\beta b, \alpha\geq 1, \beta \geq 0$,
\[ 2^{2\alpha}\mid (-1)^{m\tau+m+l}\binom{m}{l}\binom{m\tau+l-1}{m-1}-(-1)^{\frac{m\tau+m+l}{2}}\binom{\frac{m}{2}}{\frac{l}{2}}\binom{\frac{m\tau+l}{2}-1}{\frac{m}{2}-1},\]
where the second term is set to zero for $\beta=0$. \label{luo:lemma3-0}
\end{lemma}
\begin{proof}
For the case $\alpha\geq 2, \beta\geq 2$, both $m\tau+m+l$ and $(m\tau+m+l)/2$ are even, the Lemma is proved as in Lemma~\ref{luo:lemma2-0}. For the case $\beta=0$, both $\binom{m}{l}$ and $\binom{m\tau+l-1}{m-1}$ are divisible by $2^\alpha$, and the Lemma is also proved.
For remaining cases $\alpha>\beta =1$ or $\beta\geq \alpha=1$, we compute similarly as (\ref{luo:eq5-0}),
\begin{align}
& (-1)^{m\tau+m+l}\binom{m}{l}\binom{m\tau+l-1}{m-1}-(-1)^{\frac{m\tau+m+l}{2}}\binom{\frac{m}{2}}{\frac{l}{2}}\binom{\frac{m\tau+l}{2}-1}{\frac{m}{2}-1} \nonumber \\
& = \binom{\frac{m}{2}}{\frac{l}{2}}\binom{\frac{m\tau+l}{2}-1}{\frac{m}{2}-1} \left( \frac{f_2(m)}{f_2(l)f_2(m-l)}\cdot \frac{f_2(m\tau+l)}{f_2(m(\tau-1)+l)f_2(m)}-(-1)^{\frac{m\tau+m+l}{2}}\right) \label{luo:eq6-0}
\end{align}
Both $\binom{\frac{m}{2}}{\frac{l}{2}}$ and $\binom{\frac{m\tau+l}{2}-1}{\frac{m}{2}-1}$ are divisible by $2^{\alpha-1}$, it suffices to prove that the third factor is divisible by $4$, which is, by Lemma~\ref{lemma1},
\begin{equation}
(-1)^{[\frac{l}{4}]+[\frac{m-l}{4}]+[\frac{m\tau+l}{4}]+[\frac{m(\tau-1)+l}{4}]}-(-1)^{\frac{m\tau+m+l}{2}}. \pmod{4} \nonumber
\end{equation}
It is divisible by $4$ if
\begin{equation}
[\frac{l}{4}]+[\frac{m-l}{4}]+[\frac{m\tau+l}{4}]+[\frac{m(\tau-1)+l}{4}]+\frac{m\tau+m+l}{2}\label{luo:eq7-0}
\end{equation}
is even.
Parity of (\ref{luo:eq7-0}) depends only on $\tau\pmod{2}$. For
$\tau=1$, (\ref{luo:eq7-0}) reduces to
$[l/4]+[(m-l)/4]+[(m+l)/4]+[l/4]+l/2$. For $\tau=0$, it reduces to
$[l/4]+[(m-l)/4]+[l/4]+[(l-m)/4]+(m+l)/2$. Both are obviously even.
\end{proof}
Now, we can finish the proof of Theorem \ref{Thm1}.
\begin{proof}
For a prime number $p\mid m$, write $m=p^\alpha a, p\nmid a$.
\begin{align}
n_{m,l}(\tau)&=\sum_{d\mid m, d\mid l} \frac{\mu(d)}{d^2} c_{\frac{m}{d},\frac{l}{d}}(\tau) \nonumber \\
&= \frac{1}{m^2} \sum_{d\mid m, d\mid l}\mu(d)(-1)^{\frac{m\tau+m+l}{d}} \binom{\frac{m}{d}}{\frac{l}{d}}\binom{\frac{m\tau+l}{d}-1}{\frac{m}{d}-1} \nonumber \\
&=\frac{1}{m^2} \sum_{d\mid m, d\mid l, p\nmid d} \mu(d) \left((-1)^{\frac{m\tau+m+l}{d}}\binom{\frac{m}{d}}{\frac{l}{d}}\binom{\frac{m\tau+l}{d}-1}{\frac{m}{d}-1}-(-1)^{\frac{m\tau+m+l}{dp}} \binom{\frac{m}{dp}}{\frac{l}{dp}}\binom{\frac{m\tau+l}{dp}-1}{\frac{m}{dp}-1}\right) \label{luo:eq9-0}
\end{align}
where for $pd\nmid l$, the second term of (\ref{luo:eq9-0}) is understood to be zero. For odd prime number $p$, $\frac{m\tau+m+l}{d}$ and $\frac{m\tau+m+l}{dp}$ have the same parity. Since $p^\alpha\mid\mid\frac{m}{d}$, $p^{2\alpha}$ divides the summand in (\ref{luo:eq9-0}) by Lemma~\ref{luo:lemma2-0}. For $p=2$, it is divisible by $2^{2\alpha}$ by Lemma~\ref{luo:lemma3-0}.
$p^{2\alpha}$ divides the sum in (\ref{luo:eq9-0}) for every $p^\alpha \mid\mid m$, thus $n_{m,l}$ is an integer.
\end{proof}
\section{Proof of the Theorem \ref{Thm2}} \label{section4}
We introduce the following lemma first.
\begin{lemma}
If $p^\beta\mid\mid (a,b), p^\alpha\mid a+b$, then $p^{\alpha-\beta}$ divides
$$ \binom{a\tau+a-1}{a}\binom{b\tau+b}{b}.$$
\label{luo:lemma3-1}
\end{lemma}
\begin{proof}
Power of prime $p$ in $n!$ is
$ \sum_{k=1}^\infty [\frac{n}{p^k}]. $
Apply this to the binomial coefficients to find that the power of $p$ in $\binom{a\tau+a-1}{a}\binom{b\tau+b}{b}$ is
\begin{align*}
& \sum_{i=1}^{\infty} \left([\frac{a\tau+a-1}{p^i}]+[\frac{b\tau+b}{p^i}]\right)-\left([\frac{a\tau-1}{p^i}]+[\frac{b\tau}{p^i}]\right)-\left([\frac{a}{p^i}]+[\frac{b}{p^i}]\right) \\
&\geq \sum_{i=1}^{\alpha} \left((\frac{(a+b)(\tau+1)}{p^i}-1)-(\frac{(a+b)\tau}{p^i}-1)\right) - \sum_{i=1}^{\beta}(\frac{a+b}{p^i})-\sum_{i=\beta+1}^{\alpha}(\frac{a+b}{p^i}-1) \\
&=\alpha-\beta
\end{align*}
where we use the fact that for $k\mid m+n+1, k>1$, $[m/k]+[n/k]=(m+n+1)/k-1$ and for $k\mid m+n, k\nmid m$, $[m/k]+[n/k]=(m+n)/k-1$.
\end{proof}
Recall the definition of the function $f_{p}(n)$ given by formula
(\ref{functionfp}). It is obvious that
\begin{equation}
f_p(p^\alpha k) \equiv f_p(p^{\alpha})^k \equiv (-1)^k \pmod{p^\alpha} \label{luo:eq3-1}
\end{equation}
Now, we can finish the proof of Theorem \ref{Thm2}:
\begin{proof}
By definition,
\begin{align}
n_{(m_1,m_2)}(\tau)& = \frac{1}{m_1+m_2}\sum_{d\mid m_1,d\mid m_2} \mu(d) (-1)^{(m_1+m_2)(\tau+1)/d} \nonumber \\
& \cdot \binom{(m_1\tau+m_1)/d-1}{m_1/d}\binom{(m_2\tau+m_2)/d}{m_2/d} \label{luo:eq1-1}
\end{align}
Let $p$ be any prime divisor of $m_1+m_2$, $p^\alpha\mid \mid m_1+m_2$. We will prove $p^\alpha$ divides the summation in (\ref{luo:eq1-1}), thus $m_1+m_2$ also divides and $n_{m_1,m_2}$ are integers.
If $p\nmid m_1$, each summand in (\ref{luo:eq1-1}) corresponds to $p\nmid d$, so $p^\alpha\mid (m_1+m_2)/d$ and $p\nmid m_1/d$. By Lemma~\ref{luo:lemma3-1} applied to $a=m_1/d, b=m_2/d$, $p^{\alpha}$ divides each summand and thus the summation.
If $p^\beta \mid\mid m_1, \beta\geq 1$, consider two summands in (\ref{luo:eq1-1}) corresponding to $d$ and $pd$ such that $pd\mid (m_1,m_2), \mu(pd)\neq 0$. When $p$ is an odd prime or $\alpha\geq 2$, the sign $(-1)^{(m_1+m_2)(\tau+1)/d}$ and $(-1)^{(m_1+m_2)(\tau+1)/(pd)}$ are equal. When $p=2, \alpha=1$, modulo 2 the sign is irrelevant. Write $a=m_1/d, b=m_2/d$, then $p^\alpha\mid a+b, p^\beta\mid\mid a$.
\begin{align}
& \binom{a\tau+a-1}{a}\binom{b\tau+b}{b}-\binom{(a\tau+a)/p-1}{a/p}\binom{(b\tau+b)/p}{b/p} \nonumber \\
&= \binom{(a\tau+a)/p-1}{a/p}\binom{(b\tau+b)/p}{b/p}\left( \frac{f_p(a\tau+a)f_p(b\tau+b)}{f_p(a\tau)f_p(a)f_p(b\tau)f_p(b)} - 1\right) \nonumber \\
&= \binom{(a\tau+a)/p-1}{a/p}\binom{(b\tau+b)/p}{b/p} \frac{f_p(a\tau+a)f_p(b\tau+b)-f_p(a\tau)f_p(a)f_p(b\tau)f_p(b)}{f_p(a\tau)f_p(a)f_p(b\tau)f_p(b)} \label{luo:eq2-1}
\end{align}
The term $\binom{(a\tau+a)/p-1}{a/p}\binom{(b\tau+b)/p}{b/p}$ is divisible by $p^{\alpha-\beta}$ by Lemma~\ref{luo:lemma3-1}. The numerator of the fraction term in (\ref{luo:eq2-1}) is divisible by $p^\beta$ by (\ref{luo:eq3-1}), and the denominator is not divisible by $p$. We proved that $p^\alpha$ divides (\ref{luo:eq2-1}), take summation over $d$, we get that $p^\alpha$ divides the summation in (\ref{luo:eq1-1}).
This is true for any $p\mid m_1+m_2$, thus $n_{(m_1,m_2)}(\tau)$ is an integer.
\end{proof}
\section{Proof of the Theorem \ref{Thm3}} \label{section5}
We establish several lemmas first.
\begin{lemma}
Suppose $k$ is a positive integer, then the number
\begin{equation*}
c_m(k,y)=\sum_{|\lambda|=m} \frac{1}{\mathfrak{z}_\lambda} k^{l(\lambda)}\{\lambda\}_{y^2}
\end{equation*}
is equal to the coefficient of $t^m$ in $(\frac{1-t/y}{1-ty})^k$.
\end{lemma}
\begin{proof}
Suppose the number of $i$'s in the partition $\lambda$ is $a_i, i=1,\cdots$. Then
\begin{align*}
c_m(k,y)&=\sum_{\sum ia_i=m} \prod_i \frac{1}{a_i!i^{a_i}} k^{a_i}(y^i-y^{-i})^{a_i} \\
&=\left[ \prod_{i=1}^{\infty} \left(\sum_{j=0}^\infty t^{ij}\frac{1}{j!i^j} k^j(y^i-y^{-i})^j\right) \right]_{t^m} \\
&=\left[ \prod_{i=1}^{\infty} \exp(t^ik(y^i-y^{-i})/i) \right]_{t^m} \\
&=\left[ \exp(k\ln(1-ty)^{-1}+k\ln(1-t/y))\right]_{t^m}\\
&=\left[ (\frac{1-t/y}{1-ty})^k \right]_{t^m}
\end{align*}
\end{proof}
\begin{lemma} \label{mmtaulemma}
Let $R=\mathbb{Q}[q^{\pm 1/2},a^{\pm 1/2}]$. Then
\begin{align} \label{mmtaugm}
\{m\}\{m\tau\}g_m(q,a)=\sum_{d\mid m}\sum_{|\mu|=m/d}\frac{\mu(d)(-1)^{m\tau/d}}{\mathfrak{z}_\mu}\frac{\{m\mu\tau\}}{\{d\mu\}}\{d\mu\}_a
\end{align}
is divisible by $\{m\tau\}\{m\}/\{1\}^2$ in $R$.
\end{lemma}
\begin{proof}
By the definition (\ref{gm}) of $g_m(q,a)$, we have the formula
(\ref{mmtaugm}). It is clear that
$$\{m\}\{m\tau\}g_m(q,a) \in R.$$
Denote $\Phi_n(q)=\prod_{d\mid n} (q^d-1)^{\mu(n/d)}$ to be the
$n$-th cyclotomic polynomial, which is irreducible over $R$. Then
$q^n-1=\prod_{d\mid n} \Phi_d(q)$, and
\begin{align} \label{cyc}
\{m\}\{m\tau\}&=q^{-\frac{m+m\tau}{2}} \prod_{{m_1}\mid m} \Phi_{m_1}(q)\prod_{{m_1}\mid m\tau} \Phi_{m_1}(q)\\\nonumber
&=q^{-\frac{m+m\tau}{2}}\prod_{{m_1}\mid m} \Phi_{m_1}(q)^2\prod_{{m_1}\mid m\tau,
{m_1}\nmid m} \Phi_{m_1}(q)
\end{align}
(i) For $m_1\mid m\tau, m_1\nmid m$, and any $|\mu|=m/d$, at least
one of $d\mu_i$'s are not divisible by $m_1$, thus
$\{m\mu_i\tau\}/\{d\mu_i\}$ is divisible by $\Phi_{m_1}(q)$. So
$\Phi_{m_1}(q)$ divides $\{m\}\{m\tau\}g_m(q,a)$.
(ii) For $m_1\mid m$ and any $|\mu|=m/d$, if not all $d\mu_i$ are
divisible by $m_1$, then at least two of them are not divisible.
Then two of corresponding $\{m\mu_i\tau\}/\{d\mu_i\}$ are divisible
by $\Phi_{m_1}(q)$.
We consider modulo $\{m_1\}^2$ in the ring $R$. It is easy to see,
for $a,b\geq 1 $,
\begin{align*}
\frac{\{abm_1\}}{\{bm_1\}} \equiv
a\left(\frac{q^{m_1/2}+q^{-m_1/2}}{2}\right)^{(a-1)b}
\pmod{\{m_1\}^2}
\end{align*}
We write $x=(q^{m_1/2}+q^{-m_1/2})/2$, then $x^2\equiv 1
\pmod{\{m_1\}^2}$.
Then modulo $\Phi_{m_1}(q)^2$, we have
\begin{align}
&\{m\}\{m\tau\}g_m(q,a)\nonumber\\
&\equiv \sum_{d\mid m}\sum_{|\mu|=m/d,m_1\mid d\mu} \frac{\mu(d)(-1)^{m\tau/d}}{\mathfrak{z}_\mu} \frac{\{m\mu\tau\}}{\{d\mu\}}\,\{d\mu\}_a \nonumber\\
&\equiv \sum_{d\mid m}\sum_{|\mu|=m/d,m_1\mid d\mu} \frac{\mu(d)(-1)^{m\tau/d}}{\mathfrak{z}_\mu} \left(\frac{m\tau}{d}\right)^{l(\mu)}
x^{(m|\mu|\tau-d|\mu|)/m_1} \{d\mu\}_a \nonumber\\
&\equiv \sum_{d\mid m}\sum_{|\lambda|=m/\mathrm{lcm}(d,m_1)} \mu(d)(-1)^{m\tau/d}x^{\frac{m}{m_1}(\frac{m\tau}{d}-1)}
\cdot\frac{1}{\mathfrak{z}_\lambda}\left(\frac{m\tau}{\mathrm{lcm}(d,m_1)}\right)^{l(\lambda)} \{\lambda\}_{a^{\mathrm{lcm}(d,m_1)}} \nonumber\\
&\equiv \sum_{d\mid m} \mu(d) (-1)^{m\tau/d} x^{\frac{m}{m_1}(\frac{m\tau}{d}-1)} \left[ \left(\frac{1-t^{\mathrm{lcm}(d,m_1)}a^{-\mathrm{lcm}(d,m_1)/2}}{1-t^{\mathrm{lcm}(d,m_1)}a^{\mathrm{lcm}(d,m_1)/2}}\right)^{m\tau/\mathrm{lcm}(d,m_1)} \right]_{t^m}
\label{luo:eq44}
\end{align}
\begin{itemize}
\item For the cases $m_1$ with an odd prime factor $p$, or $p=2$ divides $m_1$ and $4\mid m$, or $p=2$ divides $m_1$ and $2\mid \tau$: Consider those $d$ with $\mu(d)\neq 0$ and $p\nmid d$, we have $\mathrm{lcm}(d,m_1)=\mathrm{lcm}(pd,m_1)$ and parity of $m\tau/d$ equals parity of $m\tau/(pd)$, but $\mu(d)=-\mu(pd)$. Thus two terms in (\ref{luo:eq44}) corresponding to $d$ and $pd$ cancelled.
\item For the remaining case $2\mid\mid m, m_1=2, 2\nmid\tau$: $\Phi_{m_1}(q)^2=(q^{1/2}+q^{-1/2})^2=2x+2$. Coefficients of $x$ in (\ref{luo:eq44}) equals sum of terms corresponds to odd $d\mid m, \mu(d)\neq 0$, while constant term coefficient equals to sum of terms corresponds to $2d\mid m, \mu(2d)\neq 0$. The coefficients of term for $d$ and $2d$ match, so (\ref{luo:eq44}) is divisible by $x+1$.
\end{itemize}
In summary, we have proved that for $m_1\mid m\tau, m_1\nmid m$,
$\Phi_{m_1}(q)$ divides $\{m\}\{m\tau\}g_m(q,a)$; for $m_1\mid m,
m_1 \neq 1$, $\Phi_{m_1}(q)^2$ divides $\{m\}\{m\tau\}g_m(q,a)$. By
(\ref{cyc}), the lemma is proved.
\end{proof}
\begin{lemma} \label{lemmaprime}
For any integer $m\geq 1$, we have
\begin{align*}
g_m(q,a)\in z^{-2}\mathbb{Q}[z^2,a^{\pm \frac{1}{2}}].
\end{align*}
\end{lemma}
\begin{proof}
By Lemma \ref{mmtaulemma}, we have
\begin{align*}
f(q,a):=z^2g_m(q,a)=\frac{\{1\}^2}{\{m\}\{m\tau\}}\sum_{d\mid
m}\sum_{|\mu|=m/d}\frac{\mu(d)(-1)^{m\tau/d}}{\mathfrak{z}_\mu}\frac{\{m\mu\tau\}}{\{d\mu\}}\{d\mu\}_a
\in \mathbb{Q}[q^{\pm \frac{1}{2}},a^{\pm 1}].
\end{align*}
As a function of $q$, it is clear $f(q,a)$ admits $
f(q,a)=f(q^{-1},a). $ Furthermore, for any $d|m$ and $|\mu|=m/d$, we
have
\begin{align*}
m|\mu|\tau-d|\mu|-m\tau-m\equiv m^2\tau/d-m\tau=m\tau(m/d-1) \equiv 0
\pmod{2},
\end{align*}
which implies $ f(q,a)=f(-q,a). $ Therefore,
$f(q,a)=z^2g_{m}(q,a)\in \mathbb{Q}[z^2,a^{\pm \frac{1}{2}}]$. The
lemma is proved.
\end{proof}
\begin{lemma} \label{integralitylemma}
For any $\tau\in \mathbb{Z}$, we have
\begin{align} \label{integralityformula}
\{m\}\{m\tau\}\mathcal{Z}_m(q,a)\in \mathbb{Z}[q^{\pm
\frac{1}{2}},a^{\pm \frac{1}{2}}].
\end{align}
\end{lemma}
\begin{proof}
Since
\begin{align*}
(-1)^{m\tau}\{m\}\{m\tau\}\mathcal{Z}_m(q,a)&=\sum_{|\mu|=m}\frac{\{m\tau
\mu\}}{\mathfrak{z}_\mu\{\mu\}}\{\mu\}_a\\\nonumber
&=\sum_{\sum_{j\geq 1}jk_j=m}\frac{\prod_{j\geq 1}(\{m\tau
j\}\{j\}_a)^{k_j}}{\prod_{j\geq 1}j^{k_j}k_j!},
\end{align*}
we construct a generating function
\begin{align} \label{formulafm}
f(x)&=\sum_{n\geq 0}x^n\sum_{\sum_{j\geq 1}jk_j=n}\frac{\prod_{j\geq
1}(\{m\tau j\}\{j\}_a)^{k_j}}{\prod_{j\geq 1}j^{k_j}k_j!}\\\nonumber
&=\sum_{n\geq 0}\sum_{\sum_{j\geq 1}jk_j=n}\frac{\prod_{j\geq
1}(\{m\tau j\}\{j\}_ax^j)^{k_j}}{\prod_{j\geq
1}j^{k_j}k_j!}\\\nonumber &=\exp\left(\sum_{j\geq 1}\frac{\{m\tau
j\}\{j\}_ax^j}{j\{j\}}\right),
\end{align}
Then $
(-1)^{m\tau}\{m\}\{m\tau\}\mathcal{Z}_m(q,a)=\left[f(x)\right]_{x^m}.
$
For $\tau=0$, it is the trivial case.
For $\tau \geq 1$, we use the expansion $\frac{\{m\tau
j\}}{\{j\}}=\sum_{k=0}^{m\tau-1}q^{\frac{j(m\tau-2k-1)}{2}}$, then
\begin{align*}
f(x)&=\exp\left(\sum_{k\geq 0}^{m\tau-1}\sum_{j\geq
1}\left(\frac{(q^{\frac{m\tau-1-2k}{2}}a^{\frac{1}{2}}x)^j}{j}-\frac{(q^{\frac{m\tau-1-2k}{2}}a^{-\frac{1}{2}}x)^j}{j}\right)\right)\\\nonumber
&=\exp\left(\sum_{k\geq
0}^{m\tau-1}\log\frac{1+q^{\frac{m\tau-1-2k}{2}}a^{-\frac{1}{2}}x}{1+q^{\frac{m\tau-1-2k}{2}}a^{\frac{1}{2}}x}\right)\\\nonumber
&=\prod_{k=0}^{m\tau-1}\frac{1+q^{\frac{m\tau-1-2k}{2}}a^{-\frac{1}{2}}x}{1+q^{\frac{m\tau-1-2k}{2}}a^{\frac{1}{2}}x}.
\end{align*}
We introduce the $q$-binomial coefficients defined by
\begin{align*}
\binom{m}{r}_q=\frac{(1-q^m)(1-q^{m-1})\cdots
(1-q^{m-r+1})}{(1-q)(1-q^2)\cdots (1-q^r)}
\end{align*}
for $r\leq m$, and in particular $\binom{m}{0}_q=1$. The
$q$-binomial coefficients $\binom{m}{r}_q\in \mathbb{Z}[q]$ (see
Chapter 2 of \cite{KS0} for $q$-binomial coefficients). There are
analogs of the binomial formula, and of Newton's generalized version
of it for negative integer exponents,
\begin{align*}
\prod_{k=0}^{n-1}(1+q^kt)&=\sum_{k=0}^nq^{\frac{k(k-1)}{2}}\binom{n}{k}_qt^k
\\\nonumber
\prod_{k=0}^{n-1}\frac{1}{(1-q^kt)}&=\sum_{k=0}^\infty
\binom{n+k-1}{k}_qt^k.
\end{align*}
Therefore, the coefficient $\left[f(x)\right]_{x^m}$ of $x^m$ in
$f(x)$ is given by
\begin{align*}
\sum_{j+k=m}(-1)^{k}q^{\frac{j(j-1)-(m\tau-1)m}{2}}a^{\frac{k-j}{2}}\binom{m\tau}{j}_q\binom{m\tau+k-1}{k}_q,
\end{align*}
which lies in the ring $\mathbb{Z}[q^{\pm \frac{1}{2}},a^{\pm
\frac{1}{2}}]$ by the integrality of Gaussian binomial.
For the case $\tau \leq -1$, we write $\{m\tau j\}=-\{-m\tau j\}$ in
the formula (\ref{formulafm}), then the similar computations give
the formula (\ref{integralityformula}).
\end{proof}
Now, we can finish the proof of Theorem \ref{Thm3} as follow:
\begin{proof}
Lemma \ref{lemmaprime} implies that there exist
rational numbers $n_{m,g,Q}(\tau)$, such that
\begin{align*}
z^2g_{m}(q,a)=\sum_{g \geq 0}\sum_{Q}n_{m,g,Q}(\tau)z^{2g}a^{Q}\in
\mathbb{Q}[z^2,a^{\pm \frac{1}{2}}].
\end{align*}
So we only need to show $n_{m,g,Q}(\tau)$ are integers. By lemma
\ref{integralitylemma} and the formula (\ref{gm}) for $g_{m}(q,a)$,
we have
\begin{align*}
\{m\}\{m\tau\}z^2g_{m}(q,a)\in \mathbb{Z}[q^{\pm \frac{1}{2}},a^{\pm
\frac{1}{2}}],
\end{align*}
which is equivalent to
\begin{align*}
(q^{\frac{m}{2}}-q^{-\frac{m}{2}})(q^{\frac{m\tau}{2}}-q^{-\frac{m\tau}{2}})\sum_{g
\geq 0}\sum_{Q}n_{m,g,Q}(\tau)(q^{1/2}-q^{-1/2})^{2g}a^{Q}\in
\mathbb{Z}[q^{\pm \frac{1}{2}},a^{\pm \frac{1}{2}}].
\end{align*}
So it is easy to get the contradiction if we assume there exists
$n_{m,g,Q}(\tau)$ which is not an integer.
\end{proof}
\vskip 30pt
$$ \ \ \ \ $$
|
\section{Introduction}
Cohomological Hall algebra of a quiver (with potential) was introduced by Kontesivech and Soibelman in \cite{KS}, which became a central tool in the study of quantized Donaldson-Thomas invariants. The underlying space of the cohomological Hall algebra of a finite quiver is the equivariant cohomology group of representations of the quiver, while multiplication is defined by using pull-back and push-forward operations. The discovery of cohomological Hall algebras opens new directions for motivic Donaldson-Thomas invariants, mathematical physics, representation theory and so on. Different versions of cohomological Hall algebras have been introduced and studied, see \cite{FR, SS, SV, YZ} for instance. To find a simple algebraic description of a cohomological Hall algebra is a hot topic (cf. \cite{FR19,SV}).
Let $d$ be a positive integer and $Q$ the $d$-loop quiver. The cohomological Hall algebra $\mathcal{H}^d$ of $Q$ admits an algebraic description. Namely, it is isomorphic to the space of symmetric polynomials in all possible numbers of variables endowed with a shuffle-type algebra structure. In order to develop an explicit and combinatorial setup for the study of the quantized Donaldson-Thomas invariants of $Q$, Reineke \cite{R} introduced the quantized cohomological Hall algebra $\mathcal{H}^d_q$ and the degenerate cohomological Hall algebra $\mathcal{H}^d_0$ of $Q$. In particular, the cohomological Hall algebra $\mathcal{H}^d$ is the specialization of $\mathcal{H}^d_q$ at $q=1$, while the degenerate cohomological Hall algebra $\mathcal{H}^d_0$ is the specialization of $\mathcal{H}^d_q$ at $q=0$. Surprisingly, the degenerate cohomological Hall algebra $\mathcal{H}_0^d$ admits a combinatorial realization via partitions, which plays a key role in the study of quantized Donaldson-Thomas invariants for $Q$ (cf. \cite{R}).
The aim of this note is to pursue a similar combinatorial realization of $\mathcal{H}^d_q$ as $\mathcal{H}^d_0$ and to understand the associativity of the multiplication of $\mathcal{H}^d_q$. To do this, we introduce an algebra $\mathbb{A}^d_q(\Lambda)$ associated to partitions, which can be viewed as a quantization of the algebra $\mathbb{A}^d(\Lambda)$ introduced in \cite[Definition 5.2]{R}. The associativity of the multiplication of $\mathbb{A}^d_q(\Lambda)$ comes from combinatorial properties of a class of polynomials, which have already appeared in the algebraic description of (quantized) cohomological Hall algebra $\mathcal{H}^d_q$.
We show that
$\mathbb{A}^d_q(\Lambda)$ is isomorphic to the quantized cohomological Hall algebra $\mathcal{H}^d_q$ and the isomorphism specializes to the isomorphism between $\mathbb{A}^d(\Lambda)$ and the degenerate cohomological Hall algebra $\mathcal{H}^d_0$.
Throughout this note, we fix a positive integer $d$.
\section{Quantized cohomological Hall algebras}
In this section, we recall the algebraic descriptions of the cohomological Hall algebra $\mathcal{H}^d$, the quantized cohomological Hall algebra $\mathcal{H}^d_q$ and the degenerate cohomological Hall algebra $\mathcal{H}^d_0$ of the $d$-loop quiver.
\subsection{$\mathcal{H}^d$ of the $d$-loop quiver}
For a non-negative integer $n$, denote by $S_n$ the symmetric group on $n$ letters. Let $\mathbb{Q}[x_1,\ldots, x_n]$ be the ring of polynomials in variables $x_1, \ldots, x_n$ with rational coefficients. Denote by $\mathbb{Q}[x_1,\ldots, x_n]^{S_n}$ the subspace of $\mathbb{Q}[x_1,\ldots, x_n]$ consisting of symmetric polynomials.
Let
\[\mathcal{H}^d=\bigoplus_{n\geq 0}\mathbb{Q}[x_1, \ldots, x_n]^{S_n}.
\]
We endow $\mathcal{H}^d$ with a multiplication $\ast$ as follows.
Given $f_1\in \mathbb{Q}[x_1,\ldots, x_{n_1}]^{S_{n_1}}$, $f_2\in \mathbb{Q}[x_1,\ldots, x_{n_2}]^{S_{n_2}}$, let
\begin{align}
&f_1\ast f_2(x_1,\ldots,x_{n_1+n_2})\nonumber\\
&=\sum_{\{i_1,\ldots,i_{n_1}\}} f_1(x_{i_1},\ldots,x_{i_{n_1}})f_2(x_{j_1},\ldots,x_{j_{n_2}})
\prod_{l=1}^{n_2}\prod_{t=1}^{n_1}
(x_{j_l}-x_{i_t})^{d-1},\label{qe0}
\end{align}
where the sum ranges over all the subsets $\{i_1,\ldots,i_{n_1}\}$ of $\{1,2,\ldots,n_1+n_2\}$ and $\{j_1,\ldots,j_{n_2}\}=\{1,2,\ldots,n_1+n_2\}\setminus
\{i_1,\ldots,i_{n_1}\}$. According to \cite[Theorem 2.2]{KS}, the \textit{cohomological Hall algebra} of the $d$-loop quiver is isomorphic to $(\mathcal{H}^d, \ast)$.
\subsection{$\mathcal{H}^d_q$ and $\mathcal{H}^d_0$ of $d$-loop quiver}
Let $q$ be an indeterminate and $\mathbb{Q}[q]$ the ring of polynomials in $q$ with rational coefficients.
The (naive) {\it quantized cohomological Hall algebra} $\mathcal{H}_q^d$ is the graded $\mathbb{Q}[q]$-module
\[\bigoplus_{n\geq 0}\mathbb{Q}[q][x_1,\ldots,x_n]^{S_n}
\]
endowed with the multiplication
\begin{align}
&f_1\ast f_2(x_1,\ldots,x_{n_1+n_2})\nonumber\\
&=\sum_{\{i_1,\ldots,i_{n_1}\}} f_1(x_{i_1},\ldots,x_{i_{n_1}})f_2(x_{j_1},\ldots,x_{j_{n_2}})
\prod_{l=1}^{n_2}\prod_{t=1}^{n_1}
(x_{j_l}-qx_{i_t})^{d-1},\label{qe1}
\end{align}
where $f_1\in \mathbb{Q}[q][x_1,\ldots, x_{n_1}]^{S_{n_1}}$, $f_2\in \mathbb{Q}[q][x_1,\ldots, x_{n_2}]^{S_{n_2}}$ and the sum ranges over all the subsets $\{i_1,\ldots,i_{n_1}\}$ of $\{1,2,\ldots,n_1+n_2\}$ and $\{j_1,\ldots,j_{n_2}\}=\{1,2,\ldots,n_1+n_2\}\setminus
\{i_1,\ldots,i_{n_1}\}$, see \cite[Definition 5.3]{R}.
We can specialize the algebra $\mathcal{H}^d_q$ to any $q\in \mathbb{Q}$. The {\it degenerate cohomological Hall algebra} $\mathcal{H}^d_0$ of the $d$-loop quiver is the specialization $\mathcal{H}^d_q|_{q=0}$ of $\mathcal{H}^d_q$ at $q=0$. In particular, $\mathcal{H}^d_0$ has the same underlying space as $\mathcal{H}^d$ and its multiplication reduces to
\begin{align*}
&f_1\ast f_2(x_1,\ldots,x_{n_1+n_2})\nonumber\\
&=\sum_{\{i_1,\ldots,i_{n_1}\}} f_1(x_{i_1},\ldots,x_{i_{n_1}})f_2(x_{j_1},\ldots,x_{j_{n_2}})
\left(\prod_{l=1}^{n_2}
x_{j_l}\right)^{(d-1)n_1}.
\end{align*}
\subsection{A combinatorial realization of $\mathcal{H}^d_0$}
Let $\mathbb{N}$ be the set of non-negative integers.
Let $n$ be a non-negative integer. A partition of length $n$ is a vector $\lambda=(\lambda_1,\lambda_2,\ldots, \lambda_n)\in \mathbb{N}^n$ such that $\lambda_1\leq \lambda_2\leq \cdots\leq \lambda_n$. For a partition $\lambda$, we denote by $|\lambda|$ the length of $\lambda$. For an integer vector $\mu\in \mathbb{N}^n$, denote by $\mu_{\leq}$ the partition obtained by rearranging the sequence of $\mu$ increasingly.
Let $\Lambda_n$ be the set of partitions of length $n$. Denote by $\Lambda$ the disjoint union of all $\Lambda_n$ for $n\geq 0$. Reineke \cite[Definition 5.2]{R} introduced an associative algebra $\mathbb{A}^d(\Lambda)$, which is the $\mathbb{Q}$-vector space with basis elements $\lambda\in \Lambda$ endowed with a multiplication as follows
\begin{align}
\mu\ast \nu=(\mu_1,\ldots, \mu_m, \nu_1+m(d-1), \ldots, \nu_n+m(d-1))_{\leq},
\end{align}
where $|\mu|=m$ and $|\nu|=n$. The multiplication of $\mathbb{A}^d(\Lambda)$ is obviously associative. It has been proved by \cite[Proposition 5.4]{R} that $\mathbb{A}^d(\Lambda)$ is isomorphic to $\mathcal{H}^d_0$.
\section{The algebra $\mathbb{A}^d_q(\Lambda)$}
In this section, we will introduce a $\mathbb{Q}[q]$-algebra associated to partitions, which can be viewed as a quantization of $\mathbb{A}^d(\Lambda)$.
\subsection{The polynomial $g_{m,n}$}
Let $x_1,x_2,\ldots$ be infinitely many formal variables. For a non-negative integer vector ${\bf a}=(a_1,\ldots, a_n)\in \mathbb{N}^n$, we write
\[x^{\bf a}:=\prod_{i=1}^nx_{i}^{a_i}.
\]
For a pair of positive integers $(m,n)$, denote by $g_{m,n}\in \mathbb{Q}[q][x_1,\ldots, x_{m+n}]$ the following polynomial
\begin{eqnarray}\label{gmn}
g_{m,n}:=
\prod_{s=m+1}^{m+n}\prod_{t=1}^m(x_s-qx_t)^{d-1}
=\sum_{\tiny \begin{array}{l}{\bf a}\in \mathbb{N}^m \\
{\bf b}\in \mathbb{N}^n\end{array}}c_{{m,n}}^{{\bf a},
{\bf b}}x^{({\bf a}, {\bf b})},
\end{eqnarray}
where $c_{m,n}^{{\bf a}, {\bf b}}\in \mathbb{Q}[q]$ is the coefficient of the monomial $x^{({\bf a},{\bf b})}:=x^{(a_1,\ldots,a_m,b_1,\ldots,b_n)}$ in the expansion of $g_{m,n}$. Let $l$ be a positive integer, we also denote
\[g_{m,n}[l]:=\prod_{s=l+m+1}^{l+m+n}\prod_{t=l+1}^{l+m}(x_s-qx_t)^{d-1}.
\]
By definition, it is easy to see that $g_{m,n}$ is symmetric with respect to
$x_1,\ldots, x_m$ and $x_{m+1}, \ldots, x_{m+n}$, respectively.
Consequently, for any ${\bf a}\in \mathbb{N}^m$,
${\bf b}\in \mathbb{N}^n$
and any permutations $\tau\in S_m,\sigma\in S_n$,
\[c_{m,n}^{\tau({\bf a}), \sigma({\bf b})}=c_{m,n}^{{\bf a}, {\bf b}}.\]
\begin{lemma}\label{l:cofficient-equality}
Let $l,m,n$ be positive integers and ${\bf a}\in \mathbb{N}^l, {\bf b}\in \mathbb{N}^m,{\bf c}\in \mathbb{N}^n$. We have
\begin{eqnarray}\label{f:coefficient-equality}
\sum_{\tiny \begin{array}{l}{\bf a}_1+{\bf a}_2={\bf a}\\ {\bf b}_1+{\bf b}_2={\bf b}
\end{array}}c_{l,m}^{{\bf a}_1, {\bf b}_1}c_{l+m,n}^{({\bf a}_2,{\bf b}_2), {\bf c}}
=\sum_{\tiny \begin{array}{l}{\bf c}_1+{\bf c}_2={\bf c}\\ {\bf b}_1+{\bf b}_2={\bf b}
\end{array}}c_{l,m+n}^{{\bf a}, ({\bf b}_2, {\bf c}_2)}c_{m,n}^{{\bf b}_1, {\bf c}_1},
\end{eqnarray}
where the sums are taking over all non-negative integer vectors ${\bf a}_1$, ${\bf a}_2$, ${\bf b}_1$, ${\bf b}_2$, ${\bf c}_1$, ${\bf c}_2$.
\end{lemma}
\begin{proof}
By definition, we have
\begin{eqnarray*}
g_{l,m}g_{l+m,n}&=&\prod_{s=l+1}^{l+m}\prod_{t=1}^l(x_s-qx_t)^{d-1} \prod_{s=l+m+1}^{l+m+n}\prod_{t=1}^{l+m}(x_s-qx_t)^{d-1} \\[5pt]
&=&\left(\sum_{\tiny \begin{array}{l}{\bf a}_1\in\mathbb{N}^l \\
{\bf b}_1\in\mathbb{N}^m\end{array}}c_{{l,m}}^{{\bf a}_1,{\bf b}_1}
x^{({\bf a}_1, {\bf b}_1)}\right) \left(\sum_{\tiny \begin{array}{l}
{\bf a}_2\in \mathbb{N}^l \\ {\bf b}_2\in \mathbb{N}^m\\ {\bf c}
\in \mathbb{N}^n\end{array}}c_{{l+m,n}}^{({\bf a}_2,{\bf b}_2),
{\bf c}}x^{({\bf a}_2, {\bf b}_2, {\bf c})}\right)\\[5pt]
&=&\sum_{\tiny \begin{array}{l}{\bf a}_1,{\bf a}_2\in \mathbb{N}^l \\ {\bf b}_1, {\bf b}_2\in \mathbb{N}^m\\ {\bf c}\in \mathbb{N}^n\end{array}}c_{{l,m}}^{{\bf a}_1,{\bf b}_1}c_{{l+m,n}}^{({\bf a}_2,{\bf b}_2),{\bf c}}x^{({\bf a}_1+{\bf a}_2, {\bf b}_1+{\bf b}_2, {\bf c})}.
\end{eqnarray*}
Thus the left hand side of (\ref{f:coefficient-equality}) is the coefficient of the monomial $x^{({\bf a}, {\bf b}, {\bf c})}$ in the expansion of $g_{l,m}g_{l+m,n}$.
On the other hand, consider the following polynomial
\begin{eqnarray*}
g_{l,m+n}(g_{m,n}[l])&=&\prod_{s=l+1}^{l+m+n}\prod_{t=1}^l (x_s-qx_t)^{d-1} \prod_{s=l+m+1}^{l+m+n}\prod_{t=l+1}^{l+m}(x_s-qx_t)^{d-1}\\[5pt]
&=&\left(\sum_{\tiny \begin{array}{l}{\bf a}\in \mathbb{N}^l\\ {\bf b}_1\in \mathbb{N}^m\\ {\bf c}_1\in \mathbb{N}n\end{array}}c_{l,m+n}^{{\bf a}, ({\bf b}_1,{\bf c}_1)}x^{({\bf a}, {\bf b}_1,{\bf c}_1)}\right)\left(\sum_{\tiny \begin{array}{l}{\bf b}_2\in \mathbb{N}^m\\ {\bf c}_2\in \mathbb{N}n\end{array}}c_{m,n}^{{\bf b}_2,{\bf c}_2}x^{({\bf 0}, {\bf b}_2,{\bf c}_2)}\right)\\[5pt]
&=&\sum_{\tiny \begin{array}{l}{\bf a}\in \mathbb{N}^l\\ {\bf b}_1,{\bf b}_2\in \mathbb{N}^m\\ {\bf c}_1,{\bf c}_2\in \mathbb{N}n\end{array}}c_{l,m+n}^{{\bf a}, ({\bf b}_1,{\bf c}_1)}c_{m,n}^{{\bf b}_2,{\bf c}_2}x^{({\bf a}, {\bf b}_1+{\bf b}_2,{\bf c}_1+{\bf c}_2)}.
\end{eqnarray*}
Then the right hand side of (\ref{f:coefficient-equality}) is the coefficient of the monomial $x^{({\bf a}, {\bf b}, {\bf c})}$ in the expansion of $g_{l,m+n}(g_{m,n}[l])$. The equality (\ref{f:coefficient-equality}) follows from the following obvious relation
\[g_{l,m}g_{l+m,n}=g_{l,m+n}(g_{m,n}[l]).\]
\end{proof}
\subsection{An auxiliary algebra $\mathbb{A}^d_q(\hat{\Lambda})$}
Denote by $\hat{\Lambda}$ the set of disjoint union of all $\mathbb{N}^n$ for $n\geq 0$. Let $\mathbb{A}^d_q(\hat{\Lambda})$ be the free $\mathbb{Q}[q]$-module with basis elements $\lambda\in \hat{\Lambda}$. For an element $\lambda\in \hat{\Lambda}$, we denote by $|\lambda|$ the length of $\lambda$, that is, $\lambda\in \mathbb{N}^{|\lambda|}$. Denote by $\emptyset$ the unique element of $\mathbb{N}^0$.
Let $\mu\in \mathbb{N}^m, \nu\in \mathbb{N}^n$ with $m,n\geq 1$, we define
$
\mu\ast \emptyset=\mu=\emptyset \ast \mu
$ and
\begin{align}\label{auxi}
\mu\ast\nu=\sum_{\tiny\begin{array}{l}{\bf a}\in \mathbb{N}^m\\ {\bf b}\in \mathbb{N}^n \end{array}}c_{m,n}^{{\bf a}, {\bf b}}(\mu+{\bf a}, \nu+{\bf b}),
\end{align}
where $c_{m,n}^{{\bf a}, {\bf b}}$ is as defined in \eqref{gmn}. Extend the multiplication defined in \eqref{auxi} $\mathbb{Q}[q]$-linearly to a multiplication of $\mathbb{A}^d_q(\hat{\Lambda})$.
\begin{lemma}
$(\mathbb{A}^d_q(\hat{\Lambda}), \ast)$ is an associative algebra with unit.
\end{lemma}
\begin{proof}
By definition $\emptyset$ is the unit. It remains to prove the associativity, that is, for $\mu,\nu,w\in \hat{\Lambda}$, one has $(\mu\ast \nu)\ast w=\mu\ast(\nu\ast w)$. There is nothing to prove if one of $\mu,\nu,w$ is $\emptyset$. Without loss of generality, we may assume that $|\mu|=l, |\nu|=m,|w|=n \geq 1$. By definition,
\begin{align*}
(\mu\ast \nu)\ast w&=\sum_{\tiny \begin{array}{l} {\bf a}_1\in \mathbb{N}^l\\ {\bf b}_1\in \mathbb{N}^m\end{array}}c_{l,m}^{{\bf a}_1, {\bf b}_1}(\mu+{\bf a}_1, \nu+{\bf b}_1)\ast w\\
&=\sum_{\tiny \begin{array}{l} {\bf a}_1\in \mathbb{N}^l\\ {\bf b}_1\in \mathbb{N}^m\end{array}}c_{l,m}^{{\bf a}_1, {\bf b}_1} \sum_{\tiny \begin{array}{l} {\bf a}_2\in \mathbb{N}^l\\ {\bf b}_2\in \mathbb{N}^m\\ {\bf c}\in \mathbb{N}^n\end{array}}c_{l+m,n}^{({\bf a}_2, {\bf b}_2), {\bf c}}(\mu+{\bf a}_1+{\bf a_2}, \nu+{\bf b}_1+{\bf b}_2, w+{\bf c})\\
&=\sum_{\tiny \begin{array}{l} {\bf a}_1, {\bf a}_2\in \mathbb{N}^l\\ {\bf b}_1,{\bf b}_2\in \mathbb{N}^m\\ {\bf c}\in \mathbb{N}^n\end{array}}c_{l,m}^{{\bf a}_1, {\bf b}_1}
c_{l+m,n}^{({\bf a}_2, {\bf b}_2), {\bf c}}
(\mu+{\bf a}_1+{\bf a_2}, \nu+{\bf b}_1+{\bf b}_2, w+{\bf c}).
\end{align*}
Therefore, the coefficient of $(\mu+{\bf a}, \nu+{\bf b}, w+{\bf c})$ in $(\mu\ast\nu)\ast w$ is
\[\sum\limits_{\tiny \begin{array}{l}{\bf a}_1+{\bf a}_2={\bf a}\\ {\bf b}_1+{\bf b}_2={\bf b}
\end{array}}c_{l,m}^{{\bf a}_1, {\bf b}_1}c_{l+m,n}^{({\bf a}_2,{\bf b}_2), {\bf c}}.\]
On the other hand, we have
\begin{eqnarray*}
\mu\ast(\nu\ast w)&=&\mu\ast \sum_{\tiny \begin{array}{l}{\bf b}_2\in \mathbb{N}^n\\ {\bf c}_2\in \mathbb{N}^n\end{array}}c_{m,n}^{{\bf b}_2, {\bf c}_2}(\nu+{\bf b}_2, w+{\bf c}_2)\\
&=&\sum_{\tiny \begin{array}{l} {\bf a}\in \mathbb{N}^l\\ {\bf b}_1,{\bf b}_2\in \mathbb{N}^m\\ {\bf c}_1, {\bf c}_2\in \mathbb{N}^n\end{array}}c_{l,m+n}^{{\bf a}, ({\bf b}_2, {\bf c}_2)}
c_{m,n}^{{\bf b}_1, {\bf c}_1}
(\mu+{\bf a}, \nu+{\bf b}_1+{\bf b}_2, w+{\bf c}_1+{\bf c}_2).
\end{eqnarray*}
Thus the coefficient of $(\mu+{\bf a}, \nu+{\bf b}, w+{\bf c})$ in $\mu\ast(\nu\ast w)$ is
\[\sum\limits_{\tiny \begin{array}{l}{\bf c}_1+{\bf c}_2={\bf c}\\ {\bf b}_1+{\bf b}_2={\bf b}
\end{array}}c_{l,m+n}^{{\bf a}, ({\bf b}_2, {\bf c}_2)}c_{m,n}^{{\bf b}_1, {\bf c}_1}.\]
We conclude that $(\mu\ast\nu)\ast w=\mu\ast(\mu\ast w)$ by Lemma \ref{l:cofficient-equality}.
\end{proof}
\begin{remark}
In fact, $\mathbb{A}^d_q(\hat{\Lambda})$ is an $\mathbb{N}$-graded algebra
\[\mathbb{A}^d_q(\hat{\Lambda})=\bigoplus_{n\in \mathbb{N}}\mathbb{A}^d_q(\hat{\Lambda})_n,
\]
whose $n$-th component is the free $\mathbb{Q}[q]$-module with basis $\mathbb{N}^n$.
\end{remark}
\subsection{The algebra $\mathbb{A}^d_q(\Lambda)$ associated to partitions}
Let $\mu,\nu\in \hat{\Lambda}$, we denote by $\mu\sim \nu$ if $\mu$ is a permutation of $\nu$. Clearly, $\sim$ is an equivalence relation of $\hat{\Lambda}$ and
we may identify $\Lambda$ with the quotient $\hat{\Lambda}/\sim$
of $\hat{\Lambda}$ by $\sim$. Let $W$ be the
$\mathbb{Q}[q]$-subspace of $\mathbb{A}^d_q(\hat{\Lambda})$ spanned by
\[
\{\mu\circleddash\nu~|~ \text{for all $\mu,\nu\in \hat{\Lambda}$
such that $\mu\sim \nu$}\}.
\]
Here, to
distinguish with the minus of vectors, we use $\circleddash$ to represent the minus
of $\mathbb{A}^d_q(\hat{\Lambda})$.
\begin{lemma}\label{l:homogeneous}
The two-side ideal $( W)$ of
$\mathbb{A}^d_q(\hat{\Lambda})$ generated by $W$ is exactly $W$.
\end{lemma}
\begin{proof}
Let $\mu\in \hat{\Lambda}$ with $|\mu|=m$ and $\sigma\in S_m$. For any $w\in \hat{\Lambda}$ with $|w|=n$, we have
\begin{align*
&(\mu\circleddash\sigma(\mu))\ast w\\
&=
\sum_{\tiny\begin{array}{l}{\bf a}\in\mathbb{N}^m\\
{\bf b}\in\mathbb{N}^n\end{array}}c_{m,n}^{{\bf a}, {\bf b}}
(\mu+{\bf a}, w+{\bf b})\circleddash\sum_{\tiny \begin{array}{l}{\bf a}\in \mathbb{N}^m\\
{\bf b}\in \mathbb{N}^n\end{array}}c_{m,n}^{{\bf a}, {\bf b}}
(\sigma(\mu)+{\bf a}, w+{\bf b})\\
&=\sum_{\tiny\begin{array}{l}{\bf a}\in\mathbb{N}^m\\
{\bf b}\in\mathbb{N}^n\end{array}}c_{m,n}^{{\bf a}, {\bf b}}
(\mu+{\bf a}, w+{\bf b})\circleddash\sum_{\tiny \begin{array}{l}{\bf a}\in \mathbb{N}^m\\
{\bf b}\in \mathbb{N}^n\end{array}}c_{m,n}^{{\bf a}, {\bf b}}
(\sigma(\mu)+\sigma({\bf a}), w+{\bf b})\\
&=\sum_{\tiny\begin{array}{l}{\bf a}\in\mathbb{N}^m\\
{\bf b}\in\mathbb{N}^n\end{array}}c_{m,n}^{{\bf a}, {\bf b}}[(\mu+{\bf a},w+{\bf b})\circleddash (\sigma(\mu+{\bf a}), w+{\bf b})].
\end{align*}
As a consequence, $(\mu\circleddash\sigma(\mu))\ast w$ belongs to $W$. Similarly, one can show that $w\ast (\mu\circleddash\sigma(\mu))$ belongs to $W$ as well.
\end{proof}
\begin{definition}\label{d:quantized-alg}
The algebra $\mathbb{A}^d_q(\Lambda)$ associated to partitions is the quotient algebra $\mathbb{A}^d_q(\hat{\Lambda})/(W)$.
\end{definition}
The following result is a direct consequence of Lemma \ref{l:homogeneous}, Definition \ref{d:quantized-alg} and (\ref{auxi}).
\begin{proposition}
\begin{enumerate}
\item[(1)] $\mathbb{A}^d_q(\Lambda)$ is an $\mathbb{N}$-graded associative algebra with unit, whose $n$-th component is a free $\mathbb{Q}[q]$-module with basis $\Lambda_n$;
\item[(2)] Let $\mu\in \Lambda_m, \nu\in \Lambda_n$. The multiplication of $\mu$ with $\nu$ in $\mathbb{A}^d_q(\Lambda)$ is as follows
\begin{eqnarray}\label{e:multiplication}
\mu\ast\nu=\sum_{\tiny \begin{array}{l}{\bf a}\in \mathbb{N}^m\\ {\bf b}\in \mathbb{N}^n\end{array}}c_{m,n}^{{\bf a}, {\bf b}}(\mu+{\bf a}, \nu+{\bf b})_{\leq},
\end{eqnarray}
where $c_{m,n}^{{\bf a}, {\bf b}}$ is as defined in \eqref{gmn}.
\end{enumerate}
\end{proposition}
Note that $c_{m,n}^{{\bf a},{\bf b}}$ is a polynomial in $q$. Hence we can specialize the value of $c_{m,n}^{{\bf a},{\bf b}}$ at $q=0$. The following result suggests that $\mathbb{A}^d_q(\Lambda)$ is a quantization of $\mathbb{A}^d(\Lambda)$.
\begin{corollary}
By specializing $q$ to $0$, we obtain
\[\mathbb{A}^d_q(\Lambda)|_{q=0}\cong \mathbb{A}^d(\Lambda).
\]
\end{corollary}
\begin{example} Assume that $d=4$.
Let $\mu=(0,2),\nu=(1)$. Then
\[g_{2,1}=(x_3-qx_1)^3(x_3-qx_2)^3.\]
One can compute $\mu\ast\nu$ in $\mathbb{A}^4_q(\Lambda)$ as follows.
\begin{align*}
\mu\ast\nu&=(q^6+3q^4+9q^2)\cdot(1, 3, 5)+(3q^2-3q^5)\cdot(2, 2, 5)\\
&\quad+(9q^4-3q^5-10q^3)\cdot(2, 3, 4)
+3q^4\cdot(3, 3, 3)+(3q^2-q^3)\cdot(0, 4, 5)\\
&\quad -9q^3\cdot(1, 4, 4) -3q\cdot(0, 3, 6) -3q\cdot(1, 2, 6)+(0, 2, 7).
\end{align*}
\end{example}
Let $q^{-1}$ be the formal inverse of $q$ and
$\mathbb{Q}[q^{\pm}]$ the Laurent polynomials in $q$
with rational coefficients. Let us consider the tensor
product $\mathbb{Q}[q^{\pm}]\otimes_{\mathbb{Q}[q]}
\mathbb{A}^d_q(\Lambda)$. In particular,
$\mathbb{A}^d_q(\Lambda)$ is a subalgebra of
$\mathbb{Q}[q^{\pm}]\otimes_{\mathbb{Q}[q]}\mathbb{A}^d_q(\Lambda)$.
The bar involution $\bar{ } : \mathbb{Q}[q^\pm]\to \mathbb{Q}[q^\pm]$ given by $q^i\mapsto q^{-i}$($i\in \mathbb{Z}$), induces a bar involution of
\begin{eqnarray*}
\bar{}:&&\mathbb{Q}[q^{\pm}]\otimes_{\mathbb{Q}[q]}\mathbb{A}^d_q(\Lambda)\to \mathbb{Q}[q^{\pm}]\otimes_{\mathbb{Q}[q]}\mathbb{A}^d_q(\Lambda),\\
&&f=\sum a_\mu\mu\mapsto \bar{f}=\sum\bar{a}_\mu\mu,
\end{eqnarray*}
where $a_\mu\in \mathbb{Q}[q^\pm]$. The following is a direct consequence of the property of the polynomial $g_{m,n}$ and (\ref{e:multiplication}).
\begin{proposition}
Let $\mu\in \Lambda_m,\nu\in \Lambda_n$. We have
\[\mu\ast\nu=q^{(d-1)mn}\overline{\nu\ast\mu}.
\]
\end{proposition}
\section{Isomorphism from $\mathbb{A}^d_q(\Lambda)$ to $\mathcal{H}^d_q$}
For a non-negative integer sequence $\lambda=(\lambda_1,\ldots, \lambda_n)\in \mathbb{N}^n$, we denote by
\begin{align}\label{eq2}
M_{\lambda}(x_1,\ldots,x_n):=\sum_{\sigma\in S_n}x_{\sigma(1)}^{\lambda_1}\cdots x_{\sigma(n)}^{\lambda_n}\in \mathcal{H}^d_q.
\end{align}
Let \[c({\lambda})=|\{\sigma\in S_n~|~ \lambda=(\lambda_{\sigma(1)}, \dots, \lambda_{\sigma(n)})\}|.
\]
Denote by $S_\lambda$ the set of different permutations of the parts of $\lambda$, that is,
\[S_\lambda=\{(t_1,\dots, t_n)~|~(t_1,\dots, t_n)=(\lambda_{\sigma(1)},\dots, \lambda_{\sigma(n)}), \text{for some $\sigma\in S_n$}\}.
\]
Clearly,
\[c(\lambda)=\frac{n!}{|S_\lambda|}.
\]
Let $m_\lambda(x_1,\ldots,x_n)$ be the monomial symmetric function of $\lambda$, that is, the sum of all the monomials $x_1^{\pi(1)}\cdots x_n^{\pi(n)}$, where $\pi\in S_\lambda$.
It is obvious that
\[M_{\lambda}(x_1,\ldots,x_n)=c(\lambda)\cdot m_{\lambda}(x_1,\ldots,x_n).\]
It is also easy to see that $m_{\lambda}(x_1,\ldots,x_n)=m_{\lambda'}(x_1,\ldots,x_n)$ and $M_{\lambda}(x_1,\ldots,x_n)=M_{\lambda'}(x_1,\ldots,x_n)$ for any $\lambda'\in S_{\lambda}$.
The following proposition is already known, see, for example, Carvalho and D'Agostino \cite{CD}, we give a short proof here for the convenience of the readers.
\begin{proposition}\label{pmp}
Let $\lambda,\mu$ be two partitions of length $n$. Then
\[M_{\lambda}(x_1,\ldots,x_n)m_{\mu}(x_1,\ldots,x_n)
=\sum_{w\in S_\mu}M_{\lambda+w}(x_1,\ldots,x_n).\]
\end{proposition}
\begin{proof} Fix $w\in S_{\mu}$, choose a permutation $\tau\in S_n$ (not necessarily unique) such that $w_i=\mu_{\tau(i)}.$ Then we have
\begin{eqnarray*}M_{\lambda+w}(x_1,\cdots,x_n)&=&\sum_{\pi\in S_{n}}x_{\pi(1)}^{\lambda_1+w_1}\cdots x_{\pi(n)}^{\lambda_n+w_n}\\
&=&\sum_{\pi\in S_{n}}x_{\pi(1)}^{\lambda_1+\mu_{\tau(1)}}\cdots x_{\pi(n)}^{\lambda_n+\mu_{\tau(n)}}.
\end{eqnarray*}
On the other hand,
\begin{align*}
M_{\lambda}(x_1,\cdots,x_n)M_{\mu}(x_1,\cdots,x_n)
&=\sum_{\pi\in S_n}x_{\pi(1)}^{\lambda_1}\cdots x_{\pi(n)}^{\lambda_n}\sum_{\beta\in S_n}x_{\beta(1)}^{\mu_1}\cdots x_{\beta(n)}^{\mu_n}\\
&=\sum_{\pi\in S_n}\left(\sum_{\beta\in S_n}x_{\pi(1)}^{\lambda_1+
\mu_{\beta^{-1}(\pi(1))}}\cdots x_{\pi(n)}^{\lambda_n+\mu_{\beta^{-1}(\pi(n))}}\right),
\end{align*}
For a given monomial $x_{\pi(1)}^{\lambda_1+\mu_{\tau(1)}}\cdots x_{\pi(n)}^{\lambda_n+\mu_{\tau(n)}}$ of $M_{\lambda+w}(x_1,\ldots,x_n)$,
there are $c(\mu)$ permutations $\beta\in S_n$ in the summand
\[\sum_{\beta\in S_n}x_{\pi(1)}^{\lambda_1+
\mu_{\beta^{-1}(\pi(1))}}\cdots x_{\pi(n)}^{\lambda_n+\mu_{\beta^{-1}(\pi(n))}}\]
such that $\mu_{\beta^{-1}(\pi(i))}=\mu_{\tau(i)}$ for $i=1,\ldots,n$. Moreover, since there are $|S_{\mu}|$ different $\lambda+w$, and each $\lambda+w$ corresponds to $c(\mu)$ different permutations $\beta\in S_n$, we see that
\begin{align}
M_{\lambda}(x_1,\ldots,x_n)M_{\mu}(x_1,\ldots,x_n)
=c(\mu)\sum_{w\in S_\mu}M_{\lambda+w}(x_1,\ldots,x_n).
\end{align}
Therefore
\[M_{\lambda}(x_1,\ldots,x_n)m_{\mu}(x_1,\ldots,x_n)
=\sum_{w\in S_\mu}M_{\lambda+w}(x_1,\ldots,x_n).\]
This completes the proof.
\end{proof}
\begin{example}
For example, let $\lambda=(0,1,2),\mu=(1,1,3)$. Then there are 3 different permutations of $\mu$, i.e., $S_{\mu}=\{(3,1,1),(1,3,1),(1,1,3)\}$.
Then we have
\[M_{(0,1,2)}m_{(1,1,3)}=M_{(3,2,3)}+M_{(1,4,3)}+M_{(1,2,5)}=M_{(2,3,3)}
+M_{(1,3,4)}+M_{(1,2,5)}.\]
\end{example}
\begin{corollary}\label{pm}
Let $\lambda,\mu$ be two partitions of length $n$. Then
\[m_{\lambda}(x_1,\ldots,x_n)m_{\mu}(x_1,\ldots,x_n)
=\sum_{w\in S_\mu}\frac{c(\lambda+w)}{c(\lambda)} m_{\lambda+w}(x_1,\ldots,x_n).\]
\end{corollary}
\begin{proposition}\label{2pp}
Let $\lambda=(\lambda_1,\ldots,\lambda_n)\in\Lambda_n$ and $\mu=(\mu_1,\ldots,\mu_k)\in\Lambda_k$. Then
\[
\sum_{\{i_1,\ldots,i_n\}} M_{\lambda}(x_{i_1},\ldots,x_{i_n})
M_{\mu}(x_{j_1},\ldots,x_{j_k})=M_{(\lambda,\mu)_{\le}}(x_1,\ldots,x_{n+k}),
\]
where the sum takes over all the subsets $\{i_1,\ldots,i_n\}$ of $\{1,\ldots,n+k\}$ and $\{j_1,\ldots,j_k\}=\{1,\ldots,n+k\}\setminus\{i_1,\ldots,i_n\}$.
\end{proposition}
\begin{proof} By definition \eqref{eq2},
\begin{align*}
&\sum_{\{i_1,\ldots,i_n\}}M_{\lambda}(x_{i_1},\ldots,x_{i_n})
M_{\mu}(x_{j_1},\ldots,x_{j_k})\\
&=\sum_{\{i_1,\ldots,i_n\}}\left(\sum_{\sigma\in S_n}x_{\sigma(i_1)}^{\lambda_1}\cdots x_{\sigma(i_n)}^{\lambda_n}\right)\left(\sum_{\pi\in S_k}x_{\pi(j_1)}^{\mu_1}\cdots x_{\pi(j_k)}^{\mu_k}\right)\\
&=\sum_{(\{i_1,\ldots,i_n\},\sigma,\pi)}x_{\sigma(i_1)}^{\lambda_1}\cdots x_{\sigma(i_n)}^{\lambda_n}x_{\pi(j_1)}^{\mu_1}\cdots x_{\pi(j_k)}^{\mu_k}.
\end{align*}
It is easy to see that the set of triples $\{(\{i_1,\ldots,i_n\},\sigma,\pi)\}$
is in bijection with the set of permutations $S_{n+k}$. More precisely, a triple $(\{i_1,\ldots,i_n\},\sigma,\pi)$ can be associated to the permutation \[\tau=(\tau(1),\ldots,\tau(n+k))
:=(\sigma(i_1),\ldots,\sigma(i_n),\pi(j_1),\ldots,\pi(j_k)).\]
Thus we have
\begin{align*}
&\sum_{\{i_1,\ldots,i_n\}}M_{\lambda}(x_{i_1},\ldots,x_{i_n})
M_{\mu}(x_{j_1},\ldots,x_{j_k})\\
&=\sum_{\tau\in S_{n+k}}x_{\tau(1)}^{\lambda_1}\cdots x_{\tau(n)}^{\lambda_n}x_{\tau(n+1)}^{\mu_1}\cdots x_{\tau(n+k)}^{\mu_k}\\
&=M_{(\lambda,\mu)}
(x_1,\ldots,x_{n+k})\\
&=M_{(\lambda,\mu)_{\leq}}(x_1,\ldots,x_{n+k}).
\end{align*}
This completes the proof.
\end{proof}
\begin{theorem}\label{t:main-thm}
Let $\Phi:\mathbb{A}^d_q(\Lambda)\to \mathcal{H}^d_q$ be the $\mathbb{Q}[q]$-linear map determined by
\begin{align}\label{maph}
\Phi(\lambda)=M_\lambda(x_1,\ldots, x_n),
\end{align}
where $\lambda$ is a partition of length $n$.
Then $\Phi$ is an algebra isomorphism between $\mathbb{A}^d_q(\Lambda)$ and $\mathcal{H}^d_q$.
\end{theorem}
\begin{proof}
Since $M_{\lambda}=c(\lambda)\cdot m_{\lambda}$, $\{M_{\lambda}|\lambda\in\Lambda\}$ forms a basis of $\mathcal{H}_q^d$. It is easy to see that $\Phi$ is a vector space isomorphism. To show that $\Phi$ is an algebra isomorphism, we need to show $\Phi(\mu\ast\nu)=\Phi(\mu)\ast \Phi(\nu)$, that is,
\begin{eqnarray}\label{pp}
M_{\mu}\ast M_{\nu}=\sum_{\tiny\begin{array}{l}{\bf a}\in \mathbb{N}^n\\ {\bf b}\in \mathbb{N}^k\end{array}}c_{n,k}^{{\bf a}, {\bf b}}\cdot M_{(\mu+{\bf a}, \nu+{\bf b})_{\leq }},
\end{eqnarray}
for any two partitions $\mu=(\mu_1,\ldots,\mu_n)$ and $\nu=(\nu_1,\ldots,\nu_k)$.
By definition \eqref{qe1},
\begin{align*}
&M_{\mu}\ast M_{\nu}(x_1,\ldots,x_{n+k})\\
&=\sum_{\{i_1,\ldots,i_n\}} M_{\mu}(x_{i_1},\ldots,x_{i_{n}})M_{\nu}(x_{j_1},\ldots,x_{j_{k}})
\prod_{l=1}^{k}\prod_{t=1}^{n}
(x_{j_l}-qx_{i_t})^{d-1}.
\end{align*}
Denote by $I=\{i_1,\ldots, i_n\}\subset \{1,\ldots, n+k\}$ and $J=\{j_1,\ldots, j_k\}$ such that $I\cup J=\{1,\ldots, n+k\}$.
Let \[g_{n,k}(I,J):=\prod_{s=1}^k\prod_{t=1}^n(x_{j_s}-qx_{i_t})^{d-1}.\] Clearly, $g_{n,k}(I,J)$ is symmetric with respect to $x_{i_1}, \ldots, x_{i_n}$ and $x_{j_1}, \ldots, x_{j_k}$, respectively. Hence we have
\begin{align*}
g_{n,k}(I,J)&=\sum_{\tiny\begin{array}{l}{\bf a}\in \mathbb{N}^n\\ {\bf b}\in \mathbb{N}^k\end{array}}c_{n,k}^{{\bf a}, {\bf b}}x_{i_1}^{a_1}\cdots x_{i_n}^{a_n}x_{j_1}^{b_1}\cdots x_{j_k}^{b_k}\\
&=\sum_{\tiny\begin{array}{l}{\bf a}\in \Lambda_n\\ {\bf b}\in \Lambda_k\end{array}}c_{n,k}^{{\bf a}, {\bf b}}m_{\bf a}(x_{i_1}, \ldots, x_{i_n})m_{\bf b}(x_{j_1},\ldots, x_{j_k}).
\end{align*}
Therefore,
\begin{align*}
&M_{\mu}(x_{i_1},\ldots, x_{i_n})M_\nu(x_{j_1},\ldots, x_{j_k})g_{n,k}(I,J)\\[5pt]
&=\sum_{\tiny\begin{array}{l}{\bf a}\in \Lambda_n\\ {\bf b}\in \Lambda_k\end{array}}c_{n,k}^{{\bf a}, {\bf b}}M_\mu(x_{i_1}, \ldots, x_{i_n})m_{\bf a}(x_{i_1}, \ldots, x_{i_n})M_\nu(x_{j_1}, \ldots, x_{j_k})m_{\bf b}(x_{j_1},\ldots, x_{j_k})\\[5pt]
&=\sum_{\tiny\begin{array}{l}{\bf a}\in \Lambda_n\\ {\bf b}\in \Lambda_k\end{array}}c_{n,k}^{{\bf a}, {\bf b}}\sum_{\lambda\in S_{\bf a}}M_{(\mu+\lambda)_\leq}(x_{i_1}, \ldots, x_{i_n})\sum_{w\in S_{\bf b}}M_{(\nu+w)_\leq}(x_{j_1}, \ldots, x_{j_k}) \\[5pt]
&=\sum_{\tiny\begin{array}{l}{\bf a}\in \mathbb{N}^n\\ {\bf b}\in \mathbb{N}^k\end{array}}c_{n,k}^{{\bf a}, {\bf b}}M_{(\mu+{\bf a})_\leq}(x_{i_1}, \ldots, x_{i_n})M_{(\nu+{\bf b})_\leq}(x_{j_1}, \ldots, x_{j_k}),
\end{align*}
where the second equality follows from Proposition \ref{pmp}.
Consequently,
\begin{align*}
&M_\mu\ast M_\nu(x_1,\ldots, x_{n+k})\\
&=\sum_{\{i_1,\ldots,i_n\}} M_{\mu}(x_{i_1},\ldots,x_{i_{n}})M_{\nu}(x_{j_1},\ldots,x_{j_{k}})
g_{n,k}(I,J)\\[5pt]
&=\sum_{\tiny\begin{array}{l}{\bf a}\in \mathbb{N}^n\\ {\bf b}\in \mathbb{N}^k\end{array}}c_{n,k}^{{\bf a}, {\bf b}}\sum_{\{i_1,\ldots, i_n\}}M_{(\mu+{\bf a})_\leq}(x_{i_1},\ldots, x_{i_n})M_{(\nu+{\bf b})_\leq}(x_{j_1}, \ldots, x_{j_k})\\[5pt]
&=\sum_{\tiny\begin{array}{l}{\bf a}\in \mathbb{N}^n\\ {\bf b}\in \mathbb{N}^k\end{array}}c_{n,k}^{{\bf a}, {\bf b}}M_{(\mu+{\bf a},\nu+{\bf b})_{\leq}},
\end{align*}
where the last equality follows from Proposition \ref{2pp}.
\end{proof}
To conclude, we remark that in the special case $d=2$, the product (\ref{e:multiplication}) can be defined combinatorially.
Given $\mu\in \Lambda_n$ and $\nu\in \Lambda_k$, let $G_{\mu,\nu}$ denote the set of direct bipartite graphs between nodes set $U=\{u_1,\ldots,u_n\}$ and $V=\{v_1,\ldots,v_k\}$ with an orientation of each edge. Clearly, $G_{\mu,\nu}$ has $2^{n\cdot k}$ such direct bipartite graphs.
Given a bipartite graph $G\in G_{\mu,\nu}$,
for $1\le i\le n$ and $1\le j\le k$,
let
\begin{align*}
a_i&=\mu_i+ \#\{\text{directed edges pointed to the node $u_i$}\},\\
b_j&=\nu_j+ \#\{\text{directed edges pointed to the node $v_j$}\}.
\end{align*}
Construct a vector $h(G)$ of $\mathbb{A}^2_q(\Lambda)$ according to $G$ as follows. The partition of $h(G)$ is $(a_1,\ldots,a_n,b_1,\ldots,b_k)_{\le}$. The coefficient of $h(G)$ is
$(-q)^{m(G)}$,
where
\[m(G)=\sum_{i=1}^n\#\{\text{directed edges pointed to the node $u_i$}\}.\]
Now we can define
\begin{align*}\label{mp}
\mu\ast\nu=\sum_{G\in G_{\mu,\nu}}h(G).
\end{align*}
\begin{figure}[ht]
\begin{center}
\begin{tikzpicture}
\node at (0mm,0mm){$\bullet$}; \node [below] (0mm,0mm) {3};
\node at (10mm,0mm){$\bullet$}; \node at (10mm,-3mm) {2};
\node at (20mm,0mm){$\bullet$}; \node at (20mm,-3mm) {0};
\node at (5mm,10mm){$\bullet$}; \node at (5mm,13mm) {2};
\node at (15mm,10mm){$\bullet$};\node at (15mm,13mm) {1};
\draw[->](0mm,0mm)--(1mm,2mm);\draw(1mm,2mm)--(5mm,10mm);
\draw[->](0mm,0mm)--(3mm,2mm);\draw(3mm,2mm)--(15mm,10mm);
\draw[->](10mm,0mm)--(7.5mm,5mm);\draw(7.5mm,5mm)--(5mm,10mm);
\draw[->](15mm,10mm)--(12.5mm,5mm);\draw(12.5mm,5mm)--(10mm,0mm);
\draw[->](5mm,10mm)--(8mm,8mm);\draw(8mm,8mm)--(20mm,0mm);
\draw[->](20mm,0mm)--(16mm,8mm);\draw(16mm,8mm)--(15mm,10mm);
\end{tikzpicture}
\end{center}
\vspace{-.8cm}
\caption{A direct bipartite graph $G$ in $G_{\mu,\nu}$.}\label{PPP}
\end{figure}
For example, let $\mu=(1,2)$ and $\nu=(0,2,3)$. Then $G_{\mu,\nu}$ has $2^6$ direct bipartite graphs, we illustrate one of them as in Figure \ref{PPP}. Thus $a_1=\mu_1+2, a_2=\mu_2+2, b_1=\nu_1+1,b_2=\nu_2+1,b_3=\nu_3+0$ and $m(G)=4$. If $m=2$, then we have
$h(G)=(-q)^4\cdot(1,3,3,3,4).$
\bibliographystyle{plain}
|
\section{Introduction}\label{sec:intro}
Given a kernel $K$, domain $\mathcal{X} \subset \mathbb{R}^d$, and accuracy $\varepsilon > 0$, the low-rank approximation problem
is to find a kernel $K_r$ of rank $r$, for $r$ as small as possible, satisfying
\[
\sup_{x,y \in \mathcal{X}} \abs{K(x,y) - K_r(x,y)} \leq \varepsilon.
\]
In addition to being a fundamental mathematical problem in its own right, low-rank approximation has broad algorithmic implications in data science and applied mathematics. The connection is that low-rank kernel approximations enable rapid computation of core algorithmic tasks such as the Discrete Gauss Transform, heat equation solvers~\citep{GreStr91}, optimal transport solvers~\citep{AltBacRudWee18,SolGoePey15}, and kernel methods in machine learning~\citep{RifYeoTom03,YanDurDav05,RahRec08,WilRas06}, among many others. Designing better approximations---i.e., approximations that require smaller rank $r$ for the same accuracy $\varepsilon$---immediately translates into faster algorithms for these myriad applications.
\par This broad applicability has led to an extensive literature on low-rank approximation of kernels. Existing approaches can be roughly partitioned into two categories depending on how the rank $r$ required to achieve $\varepsilon$ approximation accuracy scales in the problem parameters.
\begin{itemize}
\item \underline{High-precision approaches} scale polylogarithmically in the approximation accuracy $\varepsilon$, but exponentially in the ambient dimension $d$. A typical rate for approximating a smooth isotropic kernel over the unit\footnote{$\mathcal{X} \subset \mathbb{B}^d$ without loss of generality because rescaling the domain $\mathcal{X}$ is equivalent to rescaling the kernel function.} ball $\mathcal{X} = \mathbb{B}^d$ is
\begin{align}
r = O\left( \log 1/\varepsilon \right)^d,
\label{eq:curse-of-dim}
\end{align}
achieved for instance by polynomial methods~\citep{GreStr91,YanDurDav05,YanDurGum03,CotKesSre11,WanLiDar18}. Details in Proposition~\ref{prop:poly-standard}.
\item \underline{High-dimensional approaches} scale exponentially better in $d$, but exponentially worse in $\varepsilon$. A typical rate for approximating a positive-definite, isotropic kernel $K$ over $\mathcal{X} = \mathbb{B}^d$ is
\begin{align}
r = O\left( d \,\frac{\log(\sigma_K / \varepsilon)}{\varepsilon^2} \right),
\label{eq:rate-high-dim}
\end{align}
achieved by the Random Fourier Features method~\citep{RahRec08}.
Above, $\sigma_K^2$ is the trace of the Hessian of $k(x-y) = K(x,y)$ at $0$; this is called the ``curvature of $K$''. Details in Proposition~\ref{prop:rff-standard}.
\end{itemize}
\paragraph*{Key issue: dimension dependence.}
Both approaches have severe limitations in practice. On one hand, high-precision approaches are limited to dimensions $d\leq5$ or $10$, say, at the most. On the other hand, high-dimensional approaches cannot approximate to accuracy $\varepsilon$ better than a couple digits of precision with ranks $r$ of practical size (typically in the hundreds or thousands)---especially if the dimension $d$ is in the hundreds or thousands.
\paragraph*{Better approximation over structured domains?} A pervasive phenomenon throughout data science is that real-world datasets often lie on ``low-dimensional domains'' $\mathcal{X}$ in a high-dimensional ambient space $\mathbb{R}^d$. This motivates the critical hypothesis:
\begin{align*}
\text{\emph{If $\mathcal{X} \subset \mathbb{R}^d$ has ``effective dimension'' $d^{\star}$, }}&\text{\emph{then the dependence on the ambient dimension $d$}}
\\ \text{\emph{in the rates~\eqref{eq:curse-of-dim} and~\eqref{eq:rate-high-dim} can be }}&\text{\emph{improved to the analogous dependence on $d^{\star}$.}}
\end{align*}
There are different ways to formalize this notion of ``effective dimension''. Previous work has focused on exploiting local differentiable structure: consider $\mathcal{X}$ to be (a bounded subset of) a low-dimensional real manifold. In contrast, this paper seeks to exploit global algebraic structure: we consider $\mathcal{X}$ to be (a bounded subset of) a low-dimensional real algebraic variety.
\paragraph*{Global algebraic structure vs local differentiable structure.} These two settings of varieties and manifolds are in general incomparable.
Our investigation is motivated by the opportunity that \emph{the variety setting handles many popular applications that the manifold setting cannot.}
\par Indeed, existing bounds from the manifold literature often do not apply to the variety setting because they require the domain to satisfy smoothness or curvature bounds, and do not allow for singular points, cusps, self-intersections, etc. A quintessential example is problems involving \emph{sparse data}~\citep{CotKesSre11}, in which case $\mathcal{X}$ is a low-dimensional variety: the union of low-dimensional coordinate subspaces, see \S\ref{ssec:ex:uas}. This cannot be handled by previous manifold results since $0$ is a singular point of $\mathcal{X}$. Another important example is problems involving \emph{low-rank matrices}, in which case the relevant domain $\mathcal{X}$ is again a low-dimensional variety which cannot be handled by previous manifold results since $\mathcal{X}$ is not smooth, see \S\ref{ssec:ex:lr}. These issues can be critical in practice, not merely a theoretical technicality; see Figure~\ref{fig:nystrom-sparse}.
\begin{figure}
\begin{center}
\includegraphics[width=0.4\textwidth]{figures/nystrom_sparse}
\end{center}
\caption{Methods that exploit manifold structure (e.g., Nystr\"om) often \emph{cannot} exploit variety structure. This is evident even in simple settings: here we consider approximating the Gaussian kernel $e^{-\|x-y\|^2/2}$ over $1$-sparse vectors $x,y \in \mathbb{R}^{20}$ with norm at most $1$. Details in \S\ref{ssec:ex:uas}.}
\label{fig:nystrom-sparse}
\end{figure}
\subsection{Contribution: better approximation over varieties}\label{ssec:intro:cont}
This paper initiates the study of kernel approximation in the setting that the approximation domain is a bounded subset of a real algebraic variety $V \subset \mathbb{R}^d$. We show that in this setting, the aforementioned hypothesis is true: for both high-precision and high-dimensional approaches, the dependence on the ambient dimension $d$ can be improved to dependence on the dimension of $V$.
\begin{theorem}[Informal version of main results]\label{thm:informal}
Suppose $K$ is a smooth isotropic kernel with curvature $\sigma_K$. Consider approximating $K$ to $L^{\infty}$ error $\varepsilon$ over a domain $\mathcal{X} = V \cap \mathbb{B}^d$, where $V$ is a variety in $\mathbb{R}^d$ with dimension $d^{\star}$.
\begin{itemize}
\item \underline{High-precision approach (Theorem~\ref{thm:main-var})} It suffices to have rank
\[
r = O\left( \log 1/\varepsilon \right)^{d^{\star}}.
\]
\item \underline{High-dimensional approach (Theorem~\ref{thm:high-dim})} It suffices to have rank
\[
r = O\left(d^{\star} \frac{\log(\sigma_K d^{\star} /\varepsilon)}{\varepsilon^2} \right).
\]
\end{itemize}
\end{theorem}
\paragraph*{Remarks.}
\begin{description}
\item \emph{Our result gives tighter bounds for the many domains that are both manifolds and varieties.} For instance, while~\citep[Corollary 4]{AltBacRudWee18} shows an analog to our high-precision result in the case that $\mathcal{X}$ is a bounded manifold with dimension $d^{\star}$, the exponent is $2.5 d^{\star}$ rather than $d^{\star}$. Because manifold dimension and variety dimension coincide when $\mathcal{X}$ is both a real manifold and a real algebraic variety, Theorem~\ref{thm:main-var} provides precisely the high-precision result sought in the manifold literature: exponential dependence in the effective dimension, with no fudge factors.
\item \emph{Our techniques extend to more general kernels.} For simplicity, we state our results for smooth isotropic kernels since on one hand this captures most popular kernels used in practice, and on the other hand this level of generality yields simple proofs. Neither result requires isotropy (recall this means rotation-invariance and translation-invariance). Indeed, our high-dimensional result requires only translation-invariance, and our high-precision result assumes isotropy solely so that the ``smoothness'' assumption on the multivariate kernel $K$ can be simply stated in terms of the univariate function $f$ satisfying $K(x,y) = f(\|x-y\|^2)$. There are many ways to quantify smoothness. Our high-precision approach works whenever $K$ is well-approximated by a low-degree polynomial; this occurs for instance if $f$ is analytic in a neighborhood around the relevant domain and is satisfied for many kernels used in practice, for example the Gaussian and Cauchy kernels (see \S\ref{ssec:prelim:standard}).
\item \emph{Approximation rates of course cannot depend on the domain solely through its effective dimension.} Indeed, approximation over a space-filling curve or over a union of hyperplanes in a tightly gridded formation, is effectively as difficult as approximation over a full-dimensional domain. A similar concern also applies if the domain contains arbitrary lower-dimensional components (e.g., a cloud of points), see the discussion following Lemma~\ref{lem:hf-equidim}. In the manifold setting, degenerate domains are excluded by assuming bounds on their reach or on high-order derivatives of an atlas, see e.g.,~\citep{AltBacRudWee18,baraniuk2009random}. In our algebraic setting, we make the natural assumption that the variety's \emph{degree} is not super-exponentially large: $\log \deg(V) = O(\dim V)$. This assumption is satisfied in common situations (see \S\ref{sec:ex}), and can be checked either using standard techniques if the variety is known in advance (see \S\ref{ssec:prelim:ag}) or otherwise via estimation from samples~\citep{breiding2018learning}.
\item \emph{Adaptivity.} Both our results are presented from an existential point of view. An interesting algorithmic feature of the high-dimensional approach is that it automatically adapts to the variety. In contrast, the high-precision approach depends on the description of the variety.
\item \emph{Kernels restricted to a variety vs kernels on a variety.}
Approximating an isotropic kernel $K(x,y) = f(\|x-y\|^2)$ over a domain $\mathcal{X}$ can refer either to restricting $K$ to $x,y \in \mathcal{X}$, or replacing $\|x-y\|$ by an intrinsic metric over $\mathcal{X}$~\citep{feragen2016open}. This paper considers the former.
\end{description}
\subsection{Techniques}\label{ssec:intro:tech}
In order to prove our results, we synthesize techniques from the traditionally disparate fields of approximation theory and algebraic geometry. Key to our proofs are certain structural properties of polynomial kernels that hold when they are restricted to algebraic varieties---this may be of independent interest. We detail these properties below and how they enable us to exploit the algebraic structure of the approximation domain in the context of low-rank kernel approximation.
\paragraph*{High-precision approach.} The standard such approach exploits smoothness in order to approximate a kernel over a compact domain by a low-degree polynomial kernel. Details in \S\ref{ssec:prelim:standard}.
However, the fundamental obstacle is that the rank of a polynomial kernel grows rapidly in its degree $n$, namely as
$\binom{n+d}{d} \approx n^d$.
This is because expressing a polynomial in $d$ variables of degree $n$ as the weighted sum of monomials (or any other basis) potentially requires all monomials in $d$ variables of degree at most $n$---of which there are precisely this many. This exponential growth in $d$ is called the ``curse of dimensionality'' for high-precision approaches.
\par The primary insight behind our high-precision approach is that if a polynomial kernel is restricted to a variety $V \subset \mathbb{R}^d$, then its rank drops. In fact, significantly so if $\dim(V) \ll d$. This is perhaps most easily seen in the case of sparse data~\citep{CotKesSre11}, i.e., in the case that the variety $V$ consists of all $k$-sparse points on $\mathbb{R}^d$. Observe that in this case, any monomial that depends on more than $k$ variables vanishes on $V$. Thus a polynomial of degree $n$ over this variety $V$ can be expressed as a weighted sum of such monomials---of which there are exactly $\sum_{i=0}^k \binom{d}{i} \binom{n}{i}$ by Proposition~\ref{prop:sparse:ip}. Critically, this grows at a much slower rate of roughly $n^k$ rather than $n^d$. Such an improvement is crucial even in small-scale settings.
\par For other varieties $V$, it is not true that monomials vanish over $V$. However, the monomials become linearly dependent. For example, if $V = \{(x_1,x_2) : x_1^2 + x_2^2 = 1\}$ is the unit circle, then although no monomials vanish on $V$, the monomials $1$, $x_1^2$, and $x_2^2$ are linearly independent when restricted to $V$. This linear dependence is the generic source of rank reduction since it enables us to use a refined description of the feature space beyond the standard one of all low-degree monomials.
The refined description essentially\footnote{Factorizing over the cooordinate ring fully exploits the domain's algebraic structure. In \S\ref{ssec:hp-rank}, we obtain even better rank bounds by further exploiting a second source of algebraic structure: the polynomial function itself.}
uses bounded-degree monomials in the \emph{coordinate ring} corresponding to $V$. Intuitively, this removes all ``redundancies'' in the standard description in order to produce factorizations with lower rank.
Quantitatively, we provide an exact formula in terms of the rank of a finite-dimensional matrix (Proposition~\ref{prop:rank-exact}), and show how to compute tight asymptotic rank bounds in terms of the associated Hilbert function (Proposition~\ref{prop:rank-asymp}).
\paragraph*{High-dimensional approach.} The standard such approach uses Bochner's Theorem to express a positive-definite, translation-invariant kernel as the expectation $K(x,y) = \mathbb{E}_{\omega} [ f_{\omega}(x) f_{\omega}(y)]$ where $f_{\omega}$ is a bounded function for any realization of the random variable $\omega$, and then approximates this by taking the empirical average of $r$ samples~\citep{RahRec08}. Details in \S\ref{ssec:prelim:standard}. Both the standard analysis and our analysis proceed in two steps:
\begin{enumerate}
\item[(i)] Bound the approximation error on a finite subset $S$ of the compact domain $\mathcal{X}$.
\item[(ii)] Extend this approximation bound over $S$ to all of $\mathcal{X}$.
\end{enumerate}
Step (i) is straightforward: a standard Chernoff bound ensures small error over any finite set $S$ with high probability if $r$ grows logarithmically in $|S|$. Step (ii) is the fundamental obstacle: in order to extend an approximation bound over $S$ to $\mathcal{X}$, the set $S$ must be large. Existing analyses take $S$ to be an $\varepsilon$-net of $\mathcal{X}$ and argue step (ii) via Lipschitz smoothness of the approximation error; however the straightforward analysis requires $|S| \approx 1/\varepsilon^d$, whereby the rank $r$ scales at least linearly in $d$.
\par We approach step (ii) from an algebraic perspective rather than an analytic one. Briefly, our primary insight is that if the approximation domain $\mathcal{X}$ is a subset of a variety $V$, then the rigidity of polynomials enables us to prove step (ii) for a certain subset $S \subset \mathcal{X}$ that is of size exponential in $\dim(V)$ rather than in $d$.
\par To describe our approach in more detail, it is insightful to rephrase step (ii) in the language of approximation theory. Note that although smooth isotropic kernels and the aforementioned approximate kernel are not polynomials, they are well-approximated by low-degree polynomials, thus the approximation error (their difference) is essentially a low-degree polynomial. The upshot is that then step (ii) amounts to the following central problem in approximation theory\footnote{In the jargon of approximation theory, this problem can be equivalently stated as: Find the smallest set of interpolation nodes for which the corresponding Lebesgue constant is at most $\lambda$. It can also be stated in terms of bounding the norm of an associated interpolation operator, or finding small norming sets. See e.g.,~\citep{bos2018fekete,Riv81}.}: given a compact domain $\mathcal{X}$, degree $n$, and slack $\lambda > 1$, find the smallest subset $S \subset \mathcal{X}$ satisfying
\begin{align}
\max_{x \in \mathcal{X}} \abs{p_n(x)} \leq \lambda \cdot \max_{x \in S} \abs{p_n(x)},
\qquad \text{ for all polynomials $p_n$ of degree $n$}.
\label{eq:intro-norming}
\end{align}
A classical result in interpolation theory is that there exists a set $S$ of size $|S| \approx n^d$ satisfying~\eqref{eq:intro-norming} for $\lambda \approx n^d$. However, this is insufficient for our purposes for two reasons: we need $|S|$ to be exponential in $\dim(V)$ rather than $d$, and $\lambda$ to be $O(1)$. As shown in Proposition~\ref{prop:norming}, the former issue is fixable by working over polynomials in the corresponding coordinate ring, and the latter issue is fixable by slightly increasing $|S|$ and appealing to a tensorization argument inspired by~\citep{bloom2012polynomial}.
\subsection{Related work}\label{ssec:intro:prev}
\input{sections/intro_prev}
\subsection{Outline}\label{ssec:intro:outline}
\S\ref{sec:prelim} introduces relevant preliminaries.
Our results on kernel approximation over varieties are presented in \S\ref{sec:high-prec} (for high-precision regimes) and in \S\ref{sec:hd} (for high-dimensional regimes).
We illustrate our results on a number of example varieties in \S\ref{sec:ex}. We conclude in \S\ref{sec:discussion} with potential future directions.
\section{Kernel approximation over a variety: high-precision regime}\label{sec:high-prec}
Here we provide an exponential improvement in the rank~\eqref{eq:curse-of-dim} required by high-precision approaches for kernel approximation. Specifically, we show that if the approximation domain is a low-dimensional algebraic variety $V$ in a high-dimensional ambient space $\mathbb{R}^d$, then the curse of dimensionality for high-precision approaches can be alleviated: the exponential dependence in the ambient dimension $d$ is improvable to exponential dependence in the variety's dimension $\dim V$.
\begin{theorem}[High-precision approximation over a variety]\label{thm:main-var}
Suppose $K(x,y) = f(\|x-y\|^2)$ where $f$ satisfies Assumption~\ref{assump:smooth}. Suppose also $\mathcal{X} = V \cap \mathbb{B}^d$, where $V \subset \mathbb{R}^d$ is an equidimensional real algebraic variety.
There is a universal constant $c$ such that for all $\varepsilon > 0$, there is a kernel of rank
\begin{align}
r \leq \deg (V) \left(c \frac{\log(\alpha/\varepsilon)}{\log (1/\beta)} \right)^{\dim V}.
\label{eq:main-var:r}
\end{align}
that approximates $K$ on $\mathcal{X} \times \mathcal{X}$ to $L^{\infty}$ error $\varepsilon$.
\end{theorem}
\par As overviewed in \S\ref{ssec:intro:tech}, our approach has two components. The first controls the approximation error and is standard: exploit smoothness in order to approximate the kernel by a low-degree polynomial. The second controls the rank of our approximate kernel and is the critical new ingredient: exploit the algebraic structure of the domain in order to factorize the polynomial kernel in a succinct way. A simple statement of this second ingredient that gives asymptotic bounds is as follows; this rank bound is generically tight (see Remark~\ref{rem:rank-asymp:tight}).
\begin{prop}[Rank bound for polynomial kernels over varieties]\label{prop:rank-asymp}
Let $K_n$ be the restriction of an (indefinite) degree-$n$ polynomial kernel
to $V \times V$, where $V$ is a variety in $\mathbb{R}^d$. Then
\begin{align}
\rank K_n \leq
\operatorname{HF}_{V}(n).
\label{eq:rank-comp-HF}
\end{align}
\end{prop}
\par With this rank bound, the proof of Theorem~\ref{thm:main-var} follows readily.
\begin{proof}[Proof of Theorem~\ref{thm:main-var}]
By Assumption~\ref{assump:smooth}, there is a polynomial $p_n$ of degree $n = \lceil \log(\alpha/\varepsilon)/\log(1/\beta) \rceil$ satisfying $\|f-p_n\|_{[0,4]} \leq \varepsilon$. Thus the kernel
$K_{n}(x,y) = p_n(\|x-y\|^2)$ satisfies
$
\|K - K_{n}\|_{\mathcal{X} \times \mathcal{X}} \leq \varepsilon.
$
By Proposition~\ref{prop:rank-asymp}, the fact that $p_n(\|x-y\|^2)$ is a polynomial of degree\footnote{Although this degree increase for $n$ to $2n$ is irrelevant for the asymptotics in Theorem~\ref{thm:main-var}, a more refined analysis of isotropic polynomial kernels yields rank bounds that are better in practice, see Remark~\ref{rem:rank-asymp:use}.} $2n$, and Lemma~\ref{lem:hf-equidim},
$
\rank K_{n} \leq \deg V \binom{2n+\dim V}{\dim V}$. This is at most $\deg V \left( c n \right)^{\dim V}$
for some universal constant $c$.
\end{proof}
The remainder of the section is devoted to proving Proposition~\ref{prop:rank-asymp}. Along the way, we develop a more general understanding of how the rank of a polynomial kernel drops when it is restricted to an algebraic variety, since this may be of independent interest. In particular, we provide several illustrative examples in \S\ref{sssec:rank-ex}, describe the correspondence between polynomials over varieties and bilinear forms over coordinate rings in \S\ref{sssec:rank-forms}, provide an exact rank formula in terms of a finite-dimensional matrix in \S\ref{sssec:rank-exact}, and prove the asymptotic rank bound in Proposition~\ref{prop:rank-asymp} as well as remark on its tightness and common use cases in \S\ref{sssec:rank-asymp}.
\input{sections/variety_rank}
\section{Examples}\label{sec:ex}
In this section we consider several example varieties. We demonstrate the improved rates for kernel approximation implied by our results by computing the dimension, degree, and Hilbert function for each of these varieties. See Table~\ref{tab:variety} for a summary. We briefly remind the reader of how these three characteristics of varieties arise in our results.
\begin{itemize}
\item The \emph{dimension} of the variety is the predominant characteristic for our purposes, as the main point of our kernel approximation results in both the high-precision regime (Theorem~\ref{thm:main-var}) and high-dimensional regime (Theorem~\ref{thm:high-dim}) is that asymptotic dependence on the ambient dimension can be improved to the analogous dependence on the variety's dimension.
\item The \emph{degree} of the variety is a quantitative measure of the variety's regularity, which is required for kernel approximation (see \S\ref{ssec:intro:cont}). Our high-precision rate depends linearly on it (Theorem~\ref{thm:main-var}), and our high-dimensional rate depends logarithmically on it (Remark~\ref{rem:high-dim-deg}).
\item The \emph{Hilbert function} provides tight rank bounds on polynomial kernels over varieties (Proposition~\ref{prop:rank-asymp}). Our high-precision result (Theorem~\ref{thm:main-var}) only uses the asymptotics of this Hilbert function; computing the lower-order terms enables numerical computations and comparisons.
\end{itemize}
We include explicit computations to illustrate a variety of different techniques for determining these three quantities.
\begin{table}[]
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
\textbf{Variety} & \textbf{Ambient dim} & \textbf{Variety dim} & \textbf{Hilbert function} \bm{$\mathrm{HF}_V(n)$} & \textbf{Where} \\ \hline
$\mathbb{R}^d$ & $d$ & $d$ & $\binom{n+d}{d}$ & Ex~\ref{ex:ip-bound} \\ \hline
sphere & $d$ & $d-1$ & $\binom{n+d-1}{d-1} + \binom{n+d-2}{d-1}$ & Ex~\ref{ex:ip-bound} \\ \hline
$k$-sparse vectors & $d$ & $k$ & $\sum_{j=0}^k \binom{d}{j}\binom{n}{j}$ & \S\ref{ssec:ex:uas} \\ \hline
rank-$1$ matrices & $d = m_1m_2$ & $m_1+m_2-1$ & $\sum_{k=0}^n \binom{k+m_1-1}{m_1-1}\binom{k+m_2-1}{m_2-1}$ & \S\ref{ssec:ex:lr} \\ \hline
sym. rank-$1$ matrices & $d = \binom{m+1}{2}$ & $m$ & $\sum_{k=0}^n \binom{2k+m-1}{m-1}$ & \S\ref{ssec:ex:lr} \\ \hline
trig. moment curve & $d$ & $1$ & $dn+1$ & \S\ref{ssec:ex:moment} \\ \hline
$\operatorname{SO}(3)$ & $9$ & $3$ & $(2n+3)(2n+1)(n+1)/3$ & \S\ref{ssec:ex:SO3} \\ \hline
\end{tabular}
\caption{Characteristics of several example varieties.
}
\label{tab:variety}
\end{table}
\subsection{Sparse data}\label{ssec:ex:uas}
Many data-science applications involve $k$-sparse points in a high-dimensional ambient space $\mathbb{R}^d$, where $k \ll d$. These points lie on the algebraic variety $V$ which is the union of all $\binom{k}{d}$ coordinate subspaces of dimension $k$, i.e.,
\begin{align}
V = \bigcup_{S \subset [d], \, |S| = k} \left\{ x \in \mathbb{R}^d : x_i = 0, \, \forall i \notin S \right\}.
\label{eq:var-sparse}
\end{align}
\noindent It is clear that $\dim(V) = k$ since each of these $k$-dimensional hyperplanes is an irreducible variety. Following, we also compute the degree and Hilbert function of $V$.
\begin{prop}[Sparse data]\label{prop:sparse:ip}
Let $V$ be the variety in~\eqref{eq:var-sparse}, and suppose $k < d$. Then
\begin{itemize}
\item \underline{Hilbert function.} $\operatorname{HF}_{V}(n) = \sum_{j=0}^k \binom{d}{j}\binom{n}{j}$.
\item \underline{Dimension.} $\dim V = k$.
\item \underline{Degree.} $\deg V = \binom{d}{k}$.
\end{itemize}
\end{prop}
\begin{proof}[Proof of Proposition~\ref{prop:sparse:ip}]
$I(V)$ is the monomial ideal generated by $\{\prod_{i \in S} x_i : S \subseteq [d], |S| = k+1 \}$. Thus $\operatorname{HF}_{V}(n)$ is the number of monomials of degree at most $n$ in $\mathbb{R}[x]$ that are divisible by at most $k$ of $x_1, \dots, x_d$. We count these monomials via casework on the number $j$ of factors.
For each $j \in \{0, \dots, k\}$, there are $\binom{d}{j}$ choices of the $j$ factors $x_{i_1}, \dots, x_{i_j}$. The corresponding monomials are of the form $\prod_{\ell=1}^j x_{i_\ell}$ times monomials of degree at most $n-j$ in the $j$ variables $x_{i_1}, \dots, x_{i_j}$, of which there are $\binom{n}{j}$ many.
Therefore $\operatorname{HF}_{V}(n) = \sum_{j=0}^k \binom{d}{j}\binom{n}{j}$. Since this is a degree-$k$ polynomial with leading coefficient $\binom{d}{k} / k!$ for $n \geq d$, we have $\dim(V) = k$ and $\deg(V) = \binom{d}{k}$ by Lemma~\ref{lem:hilbert}.
\end{proof}
By Proposition~\ref{prop:rank-asymp}, this Hilbert function computation answers an open question (see the discussion in~\citep[\S3]{ongie2017algebraic}) about tight rank bounds for bounded-degree polynomial kernels that are restricted to sparse data $x,y \in V$. Previously, the only bound which exploited sparsity was $\binom{d}{k}\binom{n+k}{k}$~\citep{CotKesSre11,ongie2017algebraic}.
In contrast, our bound is always at least as good, generically exact (Remark~\ref{rem:rank-asymp:tight}), and sometimes orders-of-magnitude better.
For example, even for a small-scale instance of sparsity $k=5$ and dimension $d = 100$, the previous rank bound is in the \emph{billions} for degree $n=2$, whereas ours is only about five \emph{thousand}.
\par Note also that this variety $V$ of sparse vectors is not a manifold. This is why kernel approximation methods that exploit manifold structure perform poorly on sparse data, see the discussion in \S\ref{ssec:intro:cont} and Figure~\ref{fig:nystrom-sparse}. In that figure, we run the standard Nystr\"om method using jitter factor $1\mathrm{e}{-10}$ and plot its average performance over $50$ runs. Because there is no closed formula for the $L^{\infty}$ error of Nystr\"om over $\mathcal{X} = V \cap \mathbb{B}^d$, we plot a generous underestimate which evaluates the approximation error at a large number of sampled points. The high-precision method we compare is Taylor Features using our exact rank bound (Remark~\ref{rem:rank-asymp:tight} plus Proposition~\ref{prop:sparse:ip}). We plot its $L^{\infty}$ error which can be computed in closed form.
We conclude this discussion with a numerical illustration in the high-dimensional regime. Since a primary message of this paper is that kernel approximation has a stronger dependence on the variety dimension $k$ than on the ambient dimension $d$, we empirically investigate this in Figure~\ref{fig:rff-sparse} by plotting, for varying $d$ and $k$, the error distribution of RFF over $k$-sparse data in $\mathbb{R}^d$. In this plot, we see qualitatively the mild dependence of the error distribution in $d$, but a stronger dependence in $k$---this is consistent with Theorem~\ref{thm:high-dim}. Note that we plot the error distribution rather than the $L^{\infty}$ error since there is no closed form for the $L^{\infty}$ error of RFF and it requires a prohibitive number of samples to estimate empirically.
\begin{figure}
\begin{center}
\includegraphics[width=0.9\textwidth]{figures/rff/rff_sparse_labeled}
\end{center}
\caption{Error distribution of the RFF approach (see \S\ref{sec:hd}) for approximating the Gaussian kernel $e^{-\|x-y\|^2/2}$ over the variety $V$ of $k$-sparse vectors in $\mathbb{R}^d$, for varying $k$ (rows) and $d$ (columns). In each plot, the $x$-axis is the rank of RFF, and the $y$-axis is a boxplot of the approximation error at $10^6$ pairs of points which are drawn uniformly at random from $V$. From this plot, we see qualitatively the mild dependence of the error distribution in the ambient dimension $d$, but a stronger dependence in the variety dimension $k$---this is consistent with Theorem~\ref{thm:high-dim}.}
\label{fig:rff-sparse}
\end{figure}
\subsection{Low-rank matrices}\label{ssec:ex:lr}
Here we consider varieties of low-rank matrices.
For simplicity, we restrict to rank-$1$ matrices; one can perform similar albeit more complicated computations for any fixed rank $r$ since every such set of matrices is a determinantal variety, see e.g.~\citep{harris1984symmetric}.
\par Let us begin with the symmetric case
\begin{align}
V &= \{xx^T : x \in \mathbb{R}^m\}. \label{eq:rank1-sym}
\end{align}
Note that $V$ is a variety in the set $S^m \cong \mathbb{R}^{\binom{m+1}{2}}$ of symmetric $m \times m$ matrices as it is the vanishing set of all $2 \times 2$ minors (which are quadratic polynomials). As detailed below, the dimension of this variety is $m$, which can be much smaller than the ambient dimension $\binom{m+1}{2}$.
\begin{prop}[Symmetric rank-1 matrices]\label{prop:rank1-sym}
Let $V$ be the variety in~\eqref{eq:rank1-sym}. Then
\begin{itemize}
\item \underline{Hilbert function.} $\operatorname{HF}_{V}(n) = \sum_{k=0}^n \binom{2k+m-1}{m-1}$.
\item \underline{Dimension.} $\dim V = m$.
\item \underline{Degree.} $\deg V = 2^{m-1}$.
\end{itemize}
\end{prop}
\begin{proof}
Observe that $I(V)$ is the homogeneous ideal corresponding to the order-$2$ Veronese variety $\nu$ over $(m-1)$-dimensional projective space. Since the projective Hilbert function of this homogeneous ideal is $\binom{2n+m-1}{m-1}$ at degree $n$~\citep[Example 13.4]{harris2013algebraic},
\[
\operatorname{HF}_{V}(n) - \operatorname{HF}_{V}(n-1)
=
\binom{2n+m-1}{m-1},
\]
for $n \geq 1$ by the relation between the affine and projective Hilbert functions of a homogeneous ideal~\citep[Chapter 9, Theorem 12]{Cox13}. Since $\operatorname{HF}_{V}(0) = 1$, telescoping gives the desired Hilbert function identity.
To compute the dimension and degree of $V$, note that the relation between the affine Hilbert function of $V$ and the projective Hilbert function of $\nu$ implies that the affine Hilbert series of $V$ in indeterminate $t$ is equal to the projective Hilbert series of $\nu$ in $t$, divided by $1-t$. Thus $\dim V = \dim \nu + 1$, and $\deg V = \deg \nu$ by Lemma~\ref{lem:hilbert}. Since $\binom{2n+m-1}{m-1}$ is a polynomial of degree $m-1$ in $n$ with leading coefficient $2^{m-1}/(m-1)!$, Lemma~\ref{lem:hilbert} implies $\dim \nu = m-1$ and $\deg \nu = 2^{m-1}$.
\end{proof}
Next, we consider the variety of non-symmetric rank-$1$ matrices
\begin{align}
V = \{xy^T : x \in \mathbb{R}^{m_1}, y \in \mathbb{R}^{m_2} \}. \label{eq:rank1-nonsym}
\end{align}
Note that $V$ is a variety in $\mathbb{R}^{m_1 \times m_2} \cong \mathbb{R}^{m_1m_2}$ as it is the vanishing set of all $2 \times 2$ minors. As detailed below, the dimension of this variety is $m_1+m_2-1$, which can be much smaller then the ambient dimension $m_1m_2$.
\begin{prop}[Rank-1 matrices]\label{prop:rank1}
Let $V$ be the variety in~\eqref{eq:rank1-nonsym}.
Then
\begin{itemize}
\item \underline{Hilbert function.} $\operatorname{HF}_{V}(n) = \sum_{k=0}^n \binom{k+m_1-1}{m_1-1}\binom{k+m_2-1}{m_2-1}$.
\item \underline{Dimension.} $\dim V = m_1 + m_2 - 1$.
\item \underline{Degree.} $\deg V = \binom{m_1+m_2-2}{m_1-1}$.
\end{itemize}
\end{prop}
\begin{proof}
The proof is identical to the proof of Proposition~\ref{prop:rank1-sym}, with the Veronese variety replaced by the order-$2$ Segre variety over the Cartesian product of $(m_1-1)$ and $(m_2-1)$ dimensional projective space. The facts about this Segre variety that are needed are that its projective Hilbert function is $\binom{n+m_1-1}{m_1-1}\binom{n+m_2-1}{m_2-1}$ at degree $n$~\citep[Exercise 13.6]{harris2013algebraic}, from which it is evident that its dimension is $(m_1-1) + (m_2-1)$ and its degree is $\binom{(m_1-1)+(m_2-1)}{m_1-1}$ by Lemma~\ref{lem:hilbert}.
\end{proof}
\subsection{Trigonometric moment curve}\label{ssec:ex:moment}
Here we consider $V$ to be the trigonometric moment curve
\[
M_d =
\left\{ \begin{bmatrix} \cos(\theta), \cos(2 \theta), \dots, \cos(\tfrac{d}{2} \theta), \sin(\theta), \sin(2\theta),
\dots, \sin(\tfrac{d}{2}\theta) \end{bmatrix}^T \in \mathbb{R}^{d} \; : \;
\theta \in [0,2\pi)
\right\}
\]
for $d$ even. (We drop the $0$-th moments since they are constant and thus do not affect the variety's dimension, degree, or Hilbert function.) Although this variety $M_d$ is in ambient dimension $d$, it has dimension $1$. By Proposition~\ref{prop:rank-asymp}, this lets us prove rank bounds for polynomial kernels over $M_d$ that are linear in $d$ rather exponential in $d$. See Figure~\ref{fig:trig} for numerics.
\begin{prop}[Trigonometric moment curve]\label{prop:trig}
Suppose $d$ is an even integer, and let $V = M_d$.
\begin{itemize}
\item \underline{Hilbert function.} $\operatorname{HF}_{V}(n) = nd+1$.
\item \underline{Dimension.} $\dim V = 1$.
\item \underline{Degree.} $\deg V = d$.
\end{itemize}
\end{prop}
\begin{proof}
Denote $k = d/2$.
We establish the Hilbert function since it implies the other properties by Lemma~\ref{lem:hilbert}.
Denote a point in $V$ by $[x,y]^T$ where $x,y \in \mathbb{R}^k$ satisfy $x_j = \cos(j \theta)$ and $y_j = \sin(j \theta)$ for some $\theta \in [0,2\pi)$ and all $j \in [k]$. We claim that the following $(2k-1)k$ quadratic generators form a Gr\"obner basis for $V$ w.r.t. the grlex ordering where $x_1 > \cdots > x_k > y_1 \cdots > y_k$:
\begin{enumerate}[(i)]
\item \underline{Square terms.} Take $2x_i^2 - x_{2i} - 1$ and $2y_i^2 + x_{2i} - 1$ for $i = 1, \dots, \lfloor k/2\rfloor$; $2x_i^2 + 2y_k y_{2i-k} - x_{2k-2i} - 1$ and $2y_i^2 - 2y_{k} y_{2i-k} + x_{2k-2i} - 1$ for $i = \lfloor k/2\rfloor + 1, \dots, k-1$; and $x_i^2 + y_i^2 = 1$ for $i=k$.
\item \underline{$x_ix_j$ cross terms.} Let $i < j$. Take $2x_ix_j - x_{i+j} - x_{j-i}$ for $i+j \leq k$, and $2x_i x_j + 2y_{k} y_{i+j-k} - x_{j-i} - x_{2k-j-i}$ for $i + j > k$.
\item \underline{$x_iy_j$ cross terms.} Take $2x_iy_j - y_{i+j} - \operatorname{sign}(i-j)y_{|i-j|}$ for $i \in [k]$, $j \in [k-1]$.
\item \underline{$y_iy_j$ cross terms.} Take $2y_iy_j - x_{|i-j|} + x_{i+j}$ for $i < j < k$.
\end{enumerate}
Above, $x_0$ and $y_0$ denote $1$ and $0$, respectively. That these polynomials form a Gr\"obner basis is readily checked by observing that each is in $I(V)$ (follows from trigonometric sum-to-product and product-to-sum identities), and that the S-pair criterion holds~\citep{Cox13}.
\par Thus $\operatorname{LT}(I(V))$ is generated by all quadratic monomials in $S := \{x_1, \dots, x_k, y_1, \dots, y_{k-1}\}$. The corresponding standard monomials of degree at most $n$ are of two types:
\begin{itemize}
\item \underline{No factors in $S$}. Then the monomial is in $\mathbb{R}_{\leq n}[y_k]$.
There are $n+1$ such monomials.
\item \underline{Single linear factor from $S$}. There are $2k-1$ choices of this factor.
The rest of the monomial is in $\mathbb{R}_{\leq n-1}[y_k]$.
There are $(2k-1)n$ such monomials total.
\end{itemize}
Summing yields $2kn+1$ standard monomials total.
\end{proof}
\begin{remark}[Interpretation via combinatorial algebraic geometry]
The leading term ideal computed above for the moment curve can be interpreted as the graphical ideal corresponding to the graph on $2k$ vertices $V = \{x_1, \dots, x_k, y_1, \dots, y_k\}$ that is the complete graph with self-loops on $V \setminus \{y_k\}$.
\end{remark}
\subsection{Rotation matrices}\label{ssec:ex:SO3}
\begin{figure}
\centering
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width=0.75\linewidth]{figures/rank/trig_nolegend}
\caption{$V$ is the trigonometric moment curve $M_{100} \subset \mathbb{R}^{100}$. The rank is $100n+1$ by Proposition~\ref{prop:trig}, not $\binom{n+100}{n} = O(n^{100})$.}
\label{fig:trig}
\end{subfigure}%
\hfill
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width=1\linewidth]{figures/rank/SO3}
\caption{$V$ is the special orthogonal group $\operatorname{SO}(3) \subset \mathbb{R}^9$. The rank is $(2n+3)(2n+1)(n+1)/3$ by Proposition~\ref{prop:SO3}, not $\binom{n+9}{n} = O(n^9)$.}
\label{fig:so3}
\end{subfigure}
\caption{
Proposition~\ref{prop:rank-asymp} provides improved rank bounds on degree-$n$ polynomial kernels over a variety $V$, namely $\operatorname{HF}_{V}(n) =
O(n^{\dim V})$ rather than the standard bound $\binom{n+d}{d} = O(n^d)$.
This directly translates into better high-precision rates for kernel approximation (see \S\ref{sec:high-prec}). This improvement is demonstrated here for several varieties $V$. Observe that the $y$-axis is in log-scale.
}
\label{fig:ex-rank}
\end{figure}
Here we consider $V$ to be the special orthogonal group
\[
\operatorname{SO}(3) := \{X \in \mathbb{R}^{3 \times 3} : \det(X) = 1, \; X^TX = 1 \}.
\]
This is a variety in $\mathbb{R}^{9} \cong \mathbb{R}^{3 \times 3}$ since $\det(X) = 1$ is a polynomial equation in the entries $X_{ij}$, and $X^TX = 1$ is given by $9$ polynomial equations in the entries of $X$. Although this variety is in ambient dimension $9$, its dimension as a variety is significantly smaller, as is intuitively evident by the $3$-dimensional re-parameterization of $\operatorname{SO}(3)$ in terms of the pitch, yaw, and roll scalars. The following proposition makes this precise and computes an tight rank bound for polynomial kernels over $\operatorname{SO}(3)$ that is cubic in their degree. See Figure~\ref{fig:so3} for numerics. See also~\citep{brandt2017degree} for degree computations for higher-order special orthogonal groups.
\begin{prop}[SO(3)]\label{prop:SO3}
Let $V = \operatorname{SO}(3)$.
\begin{itemize}
\item \underline{Hilbert function.} $\operatorname{HF}_{V}(n) = (2n+3)(2n+1)(n+1)/3$.
\item \underline{Dimension.} $\dim V = 3$.
\item \underline{Degree.} $\deg V = 8$.
\end{itemize}
\end{prop}
\section{Discussion}\label{sec:discussion}
We conclude with several interesting directions for future research.
\paragraph*{Exploiting algebraic structure in other problems?} Over the past few decades, exploiting manifold structure has been established as a powerful tool for overcoming the curse of dimensionality throughout machine learning and statistics. This paper shows that one can similarly exploit variety structure (or even approximate variety structure\footnote{Since the approximate kernels in this paper are smooth (they are polynomials of bounded degree or sinuisoids with bounded frequency), they have low error on a neighborhood of the variety.}) in the context of kernel approximation. Can one use the techniques we develop to exploit algebraic structure implicit in datasets in other problems?
Applications to Optimal Transport will be investigated in forthcoming work.
\paragraph*{Interpolation between high-dimensional and high-precision methods?} Previously, methods in these two categories have been studied in a remarkably disparate way. A first, partial attempt at understanding these two approaches through a common framework is given in \S\ref{ssec:prelim:standard}. However, an understanding of if and how one can gracefully interpolate between these two very different rates remains open. In fact, this tradeoff between better dependence on the error and dimension is poorly understood not just in kernel approximation, but also in other classical fields such as numerical integration (Gaussian vs Monte-Carlo quadrature).
\paragraph*{Exploiting group symmetry in rank bounds?} Many kernels arising in practice enjoy group symmetries such as invariance with respect to coordinate permutations or sign flips. Can this structure be exploited to obtain better rank bounds for polynomial kernels---and thereby better low-rank approximations \`a la our approach in \S\ref{sec:high-prec}?
\paragraph*{Algorithmic questions.} Since the focus of this paper is on theoretical aspects of kernel approximation, our results are primarily existential in nature. Algorithmic questions about how to form these approximations are very interesting and of practical importance---in particular: efficiency, numerical stability, and automatic adaptivity to the variety. While our high-dimensional approach in \S\ref{sec:hd} enjoys these properties since it is based on RFF, these algorithmic questions are more nuanced for the high-precision approach in \S\ref{sec:high-prec} and depend on how the variety is described as input.
\section{Polynomial approximation and Fourier decay}\label{ssec:hd-trunc}
Here we show that if a kernel has a compactly supported Fourier transform, then that kernel is well-approximated by a low-degree polynomial. This is a convenient quantitative version of the standard fact that rapid decay in the frequency domain implies smoothness in the natural domain, since given a kernel whose Fourier distribution decays rapidly, one can truncate this Fourier distribution without changing the kernel much, and then approximate this by low-degree polynomials.
\begin{lemma}[Compactly supported Fourier transform implies polynomial approximation]\label{lem:ft-smooth}
Suppose kernel $K$ satisfies Assumption~\ref{assump:rff}. If its Bochner measure $\mu$ is supported on the ball of radius $r$, then for any $\varepsilon > 0$, there exists a polynomial kernel $P_n$ of degree
\[
n = O\left( r^2 + \log 1/\varepsilon \right)
\]
satisfying $\|K - P_n\|_{\mathbb{B}^d \times \mathbb{B}^d} \leq \varepsilon$.
\end{lemma}
\begin{proof}
By definition of $\mu$ and then an elementary trigonometric identity,
\[
K(x,y) = \mathbb{E}_{\omega \sim \mu} \left[ \cos(\langle \omega, x - y \rangle) \right]
=
\mathbb{E}_{\omega \sim \mu, \theta \sim \operatorname{Unif}([0,2\pi))} \left[ f_{\omega,\theta}(x) f_{\omega,\theta}(y) \right],
\]
where we use here the shorthand $f_{\omega,\theta}(x) := \cos(\langle \omega,x \rangle + \theta)$.
Now for each $\omega$ in the support of $D$ and $\theta \in [0,2\pi)$, there exists a polynomial $p_{\omega,\theta}$ of degree $n = O(r^2 + \log 1/\varepsilon)$ satisfying
\[
\sup_{x \in \mathbb{B}^d} \abs{f_{\omega,\theta}(x) - p_{\omega,\theta}(x)} \leq \varepsilon/3.
\]
For example, truncating the Taylor series expansion of the cosine function suffices.
Since the cosine function is bounded in magnitude by $1$, it follows that for all $x,y \in \mathbb{B}^d$,
\begin{align*}
\abs{f_{\omega,\theta}(x)f_{\omega,\theta}(y) - p_{\omega,\theta}(x)p_{\omega,\theta}(y)}
\leq
\sup_{\substack{z_1,z_2 \in [-1,1] \\ \varepsilon_1, \varepsilon_2 \in [-\varepsilon/3,\varepsilon/3]}} \abs{(z_1 + \varepsilon_1)(z_2 + \varepsilon_2) - z_1z_2}
= (1+\varepsilon/3)^2 - 1
\leq \varepsilon.
\end{align*}
Thus the degree-$n$ polynomial kernel
$
P(x,y) := \mathbb{E}_{w \sim D, \theta \sim U} \left[ p_{\omega,\theta}(x)p_{\omega,\theta}(y) \right]
$
satisfies
\begin{align*}
\|K - P\|_{\mathbb{B}^d \times \mathbb{B}^d}
&=
\sup_{x,y \in \mathbb{B}^d} \big| \mathbb{E}_{\omega,\theta} \left[ f_{\omega,\theta}(x) f_{\omega,\theta}(y) - p_{\omega,\theta}(x)p_{\omega,\theta}(y)\right] \big| \nonumber
\leq \varepsilon.
\end{align*}
\end{proof}
\subsection{Rank of polynomial kernels over algebraic varieties}\label{ssec:hp-rank}
\subsubsection{Illustrative examples}\label{sssec:rank-ex}
\par We begin by illustrating the underlying phenomenon through several simple examples (see \S\ref{sec:ex} for examples with more involved varieties). For simplicity, we consider the rotation-invariant kernel
\[
R_n(x,y) = \sum_{k=0}^n \frac{\langle x,y \rangle^k}{k!}.
\]
The same ideas extend to isotropic kernels, see Remark~\ref{rem:rank-asymp:use}. The significance of this kernel $R_n(x,y)$ is that it is the ``Taylor Features'' approximation of the Gaussian kernel $e^{-\|x-y\|^2/2}$, see Remark~\ref{rem:tay}, modulo omitting the scalings $e^{-(\|x\|+\|y\|^2)/2}$ which does not change the rank. In what follows, we abuse notation slightly by writing $R_n(V)$ to denote the restriction of the kernel $R_n$ to $V \times V$.
\begin{example}\label{ex:toy}
We demonstrate that the rank of $R_2$ drops when restricted to a variety $V \subset \mathbb{R}^2$.
\begin{itemize}
\item \underline{Full space.} If $V = \mathbb{R}^2$, then $R_2(x,y) = \langle \phi(x),\phi(y)\rangle$ where $\phi(x) = [1, x_1, x_2, x_1^2/\sqrt{2},x_2^2/\sqrt{2}, x_1x_2]^T \in \mathbb{R}^6$. Thus $\rank R_2(V) \leq 6$.
\item \underline{1-sparse data.} If $V = \{x \in \mathbb{R}^2 : x_1x_2 = 0\}$, then $R_2(x,y) = \langle \phi(x),\phi(y)\rangle$ where $\phi(x) = [1, x_1, x_2, x_1^2/\sqrt{2},x_2^2/\sqrt{2}]^T \in \mathbb{R}^5$. Thus $\rank R_2(V) \leq 5$.
\item \underline{Spherical data.} If $V = \{x \in \mathbb{R}^2 : x_1^2 + x_2^2 = 1\}$, then $R_2(x,y) = \langle \phi(x),\phi(y) \rangle$ where $\phi(x) = [ \sqrt{5}/2, x_1, x_2, x_1^2 -1/2, x_1x_2]^T \in \mathbb{R}^5$. Thus $\rank R_2(V) \leq 5$.
\end{itemize}
Note that the rank bounds in all these examples are tight, as shown next.
\end{example}
\par While the computations are straightforward in this toy example, computing rank bounds is clearly much more involved for more complicated kernels and varieties. Proposition~\ref{prop:rank-asymp} provides a simple, systematic approach for computing tight rank bounds.
\begin{example}[Using Proposition~\ref{prop:rank-asymp}]\label{ex:ip-bound}
Let us demonstrate how to use Proposition~\ref{prop:rank-asymp} to compute the rank of $R_n$ when restricted to a variety. Since Proposition~\ref{prop:rank-asymp} is tight for the Taylor Features kernel (Remark~\ref{rem:rank-asymp:tight}), $\rank R_n(V)
= \operatorname{HF}_{V}(n)
$ for any variety $V \subset \mathbb{R}^d$.
\begin{itemize}
\item \underline{Full space.} If $V = \mathbb{R}^d$, then by Lemma~\ref{lem:poly-dim},
\[
\rank R_n(V) = \operatorname{HF}_{V}(n) = \binom{n+d}{d}.
\]
\item \underline{$1$-sparse data.} If $V = \{x \in \mathbb{R}^d : x_ix_j = 0, \forall i < j \}$, then $I(V) = \langle x_ix_j : i < j \rangle$ is a monomial ideal and the standard monomials of degree at most $n$ are $1$ and $\{x_i^k\}_{i \in [d], k \in [n]}$.
Thus
\[
\rank R_n(V) = \operatorname{HF}_{V}(n) = nd+1.
\]
\item \underline{Spherical data.} If $V = \{x \in \mathbb{R}^d : \|x\|^2 = 1 \}$, then $I(V)$ is not a monomial ideal. The polynomial $\sum_{i=1}^d x_i^2 - 1$ generates $I(V)$ and forms a Gr\"obner basis for it w.r.t. grevlex, say. Thus $\operatorname{LT}(I(V)) = \langle x_1^2 \rangle$.
The corresponding standard monomials of degree at most $n$ are (i) monomials in $\mathbb{R}_{\leq n}[x_2, \dots, x_d]$; and (ii) $x_1$ times monomials in $\mathbb{R}_{\leq n-1}[x_2, \dots, x_d]$.
Thus
\[
\rank R_n(V) = \operatorname{HF}_{V}(n) = \binom{n+d-1}{d-1} + \binom{n+d-2}{d-1}.
\]
\end{itemize}
This proves optimality of the bounds in Examples~\ref{ex:toy} for $n=d=2$. Moreover, it shows that $\rank R_n(V)$ grows as $O(n)^d$, $O(n)$, and $O(n)^{d-1}$, respectively, because $\dim(V)$ is $d$, $1$, and $d-1$ for these three varieties $V$.
\end{example}
\subsubsection{Polynomial kernels over varieties as bilinear forms over coordinate rings}\label{sssec:rank-forms}
\par Our starting point for developing rank bounds is to view polynomial kernels as symmetric bilinear forms. First consider a polynomial kernel $K$ on the full space $\mathbb{R}^d \times \mathbb{R}^d$.
Recall that the degree of $K$ is the total degree in either variable $x, y \in \mathbb{R}^d$ (these two numbers are the same by symmetry).
A basic fact is that polynomials kernels of degree at most $n$ are in
$1$-$1$ correspondence with symmetric bilinear forms $C$ over $\mathbb{R}_{\leq n}[x]$. This fact is perhaps most intuitively understood in the monomial basis, where it amounts to the identity
\begin{align}
K(x,y) = f(x)^T C f(y), \label{eq:sbf}
\end{align}
where $f(x)$ has entries $x^{\alpha} = \prod_{i=1}^d x_i^{\alpha_i}$ and $C$ has entries $c_{\alpha,\beta}$ for multi-indices $\alpha,\beta \in \mathbb{N}_0^d$ satisfying $\sum_{i=1}^d \alpha_i$, $\sum_{i=1}^d \beta_i\leq n$, and $K(x,y) = \sum_{\alpha,\beta} c_{\alpha,\beta} x^{\alpha}y^{\beta}$.
\par How does this change if $K$ is restricted to $V \times V$, where $V$ is a variety in $\mathbb{R}^d$? This corresponds to restricting the symmetric bilinear form~\eqref{eq:sbf} to the space $\mathbb{R}_{\leq n}[V]$ of bounded-degree polynomials in the coordinate ring. (Recall from \S\ref{ssec:prelim:ag} that $\mathbb{R}_{\leq n}[V] = \mathbb{R}_{\leq n}[x] / I_{\leq n}(V)$.) This restricted form over $\mathbb{R}_{\leq n}[V]$ is easily computed in terms of the unrestricted form over $\mathbb{R}_{\leq n}[x]$ and the linear restriction map $\Phi : \mathbb{R}_{\leq n}[x] \to \mathbb{R}_{\leq n}[V]$, which maps a polynomial over $\mathbb{R}^d$ to its restriction over $V$. Note that $\Phi^T$ corresponds to the map that embeds the coordinate ring into the polynomial ring.
\begin{lemma}[Polynomial kernels over varieties as symmetric bilinear forms over coordinate rings]\label{lem:ker-sbf}
Let $K$ be the restriction of an (indefinite) degree-$n$ polynomial to $V \times V$, where $V$ is a variety in $\mathbb{R}^d$.
Then $K$ is equal to the symmetric bilinear form
\begin{align*}
\Phi C \Phi^T
\end{align*}
over $\mathbb{R}_{\leq n}[V]$, where $\Phi$ is the restriction map from $\mathbb{R}_{\leq n}[x]$ to $\mathbb{R}_{\leq n}[V]$, and $C$ is the symmetric bilinear form over $\mathbb{R}_{\leq n}[x]$ corresponding to $K$.
\end{lemma}
\begin{proof}
Although the lemma statement is basis-free, the proof is perhaps most intuitive by choosing the following convenient bases. Since $\Phi$ is the restriction map from $\mathbb{R}_{\leq n}[x]$ to $\mathbb{R}_{\leq n}[V]$, there is a polynomial basis $f_1, \dots, f_N$ of $\mathbb{R}_{\leq n}[x]$ such that (the equivalence classes corresponding to) $f_1, \dots, f_M$ form a basis for $\mathbb{R}_{\leq n}[V]$, and moreover $\Phi f_i = f_i$ for $i \in [M]$ and $\Phi f_i = 0$ for $i \in [N] \setminus [M]$. (Such a basis can be computed e.g., using Gr\"obner bases.) Abusing notation slightly, let $C$ and $\Phi$ denote the matrices corresponding to the respective linear maps w.r.t. these bases. Then $\Phi = [I_{M \times M}, 0_{M \times (N-M)}]$, and
\begin{align}
K(x,y)
= f(x)^T C f(y)
= g(x)^T \Phi C\Phi^T g(y),
\qquad \forall x,y \in V,
\label{eq:rank-exact:1}
\end{align}
where $g(x) := [f_1(x), \dots, f_M(x)]^T$.
\end{proof}
\subsubsection{Exact rank formula}\label{sssec:rank-exact}
The correspondence between polynomial kernels over varieties and symmetric bilinear forms over coordinate rings in Lemma~\ref{lem:ker-sbf} gives an exact formula for the rank of the former in terms of the rank of a finite-dimensional matrix.
\begin{prop}[Exact rank formula for polynomial kernels over varieties]\label{prop:rank-exact}
Consider the setup in Lemma~\ref{lem:ker-sbf}. Then the rank of $K$ over $V \times V$ is
\begin{align}
\rank K =
\rank \Phi C \Phi^T.
\label{eq:rank-exact}
\end{align}
In particular, if $C$ is positive definite, then
\begin{align}
\rank K = \rank \Phi = \operatorname{HF}_{V}(n).
\label{eq:rank-exact-simp}
\end{align}
\end{prop}
The proof makes use of the following generalization of the ``Unisolvence Theorem'' from the standard setting of $\mathbb{R}$ to the present setting of real algebraic varieties in $\mathbb{R}^d$.
For convenience, we state this in terms of the invertibility of a generalized Vandermonde matrix.
\begin{lemma}[Unisolvence Theorem on varieties]\label{lem:rank-exact:interp}
Suppose $V$ is a variety in $\mathbb{R}^d$. Let $M$ denote $\dim(\mathbb{R}_{\leq n}[V])$. There exist points $x_1, \dots, x_M \in V$ such that the matrix $S \in \mathbb{R}^{M \times M}$ with entries $S_{ij} = f_i(x_j)$, is non-singular for any basis $f_1, \dots, f_M$ of $\mathbb{R}_{\leq n}[V]$.
\end{lemma}
\begin{proof}
It suffices to show the claim for any fixed basis $f_1, \dots, f_M$. Define $f(x) := [f_1(x),\dots,f_M(x)]^T$, and let $W := \textrm{span}(\{f(x)\}_{x \in V}) \subseteq \mathbb{R}^M$. Since $W$ is a finite-dimensional vector space, it admits a basis of the form $\{f(x_1), \dots, f(x_k)\}$ for some $x_1, \dots, x_k \in V$. Clearly $k \leq M$. Assume for contradiction that $k < M$; else the claim follows. We make two observations. First, the $M \times k$ matrix with $ij$-th entry $f_i(x_j)$ has deficient row rank, thus there exist $\alpha_1, \dots, \alpha_M$ not all zero such that
\[
\sum_{i=1}^M \alpha_i f_i(x_j) = 0, \qquad \forall j \in [k].
\]
Second, since $\{f(x_1), \dots, f(x_k)\}$ is a basis of $W$, there exist functions $\ell_1, \dots, \ell_k : V \to \mathbb{R}$ satisfying
\[
f(x) = \sum_{j=1}^k \ell_j(x) f(x_j), \qquad \forall x \in V.
\]
(These are the corresponding Lagrange interpolating polynomials.) From these two observations it follows that $\sum_{i=1}^M \alpha_i f_i \equiv 0$ on $V$. Indeed, for all $x \in V$,
\[
\sum_{i=1}^M \alpha_i f_i(x) = \sum_{j=1}^k \ell_j(x) \sum_{i=1}^M \alpha_i f_i(x_j) = 0.
\]
This contradicts the fact that $\{f_1,\dots,f_M\}$ is a basis of of $\mathbb{R}_{\leq n}[V]$.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop:rank-exact}]
Consider the basis choice in the proof of Lemma~\ref{lem:ker-sbf}.
\par \underline{Proof of ``$\leq$''.}
Let $A^TB$ be a factorization of $\Phi C\Phi^T$ where $A$ and $B$ have $\rank \Phi C \Phi^T$ rows. Denote $\phi(x) = Ag(x)$ and $\psi(y) = Bg(y)$. Then by~\eqref{eq:rank-exact:1},
\begin{align}
K(x,y) =
\langle \phi(x), \psi(y) \rangle
\label{eq:rank-exact:fact}
\end{align}
is an explicit factorization of $K$ over $V \times V$ of rank equal to $\rank \Phi C \Phi^T$.
\par \underline{Proof of ``$\geq$''.} Consider points $x_1, \dots, x_M \in V$ guaranteed by Lemma~\ref{lem:rank-exact:interp},
and let $H \in \mathbb{R}^{M \times M}$ be the corresponding kernel matrix with entries $H_{ij} = K(x_i,x_j)$. Note that
\[
\rank K \geq \rank H,
\]
since the rank of the kernel $K$ over $V \times V$ is at least the rank of the kernel $K$ restricted to $\{x_1, \dots, x_M\} \times \{x_1, \dots, x_M\} \subseteq V \times V$, which in turn is precisely the rank of the matrix $H$. Now to bound $\rank H$, use~\eqref{eq:rank-exact:1} to write $H = S^T \Phi C \Phi^T S$ where $S \in \mathbb{R}^{M \times M}$ is the generalized Vandermonde matrix in Lemma~\ref{lem:rank-exact:interp} with entries $S_{ij} = f_i(x_j)$. Since $S$ is invertible,
\[
\rank H = \rank S^T \Psi C \Psi^T S = \rank \Phi C \Phi^T.
\]
\par \underline{Corollary when $C$ is positive definite.} In this case, $C = LL^T$ for an invertible matrix $L$. Thus
\[
\rank \Phi C \Phi^T = \rank (\Phi L)(\Phi L )^T = \rank \Phi L = \rank \Phi.
\]
Since $\Phi$ is the linear projection map onto $\mathbb{R}_{\leq n}[V]$,
the rank of $\Phi$ is the dimension of this space---which is by definition the Hilbert function $\operatorname{HF}_{V}(n)$.
\end{proof}
Note that this proof does more than establish the rank of $K$. It also identifies the feature space (the image of $\Phi C \Phi^T$ viewed as a subspace of $\mathbb{R}_{\leq n}[V]$), and an optimal factorization~\eqref{eq:rank-exact:fact}. Since this factorization has polynomial entries, we obtain the following corollary.
\begin{cor}[Polynomial kernels over varieties have optimal polynomial factorizations]\label{cor:poly-factor}
Consider the setup in Lemma~\ref{lem:ker-sbf}. There exist polynomial functions $\phi, \psi : V \to \mathbb{R}^{\rank K}$ such that $K(x,y) = \langle \phi(x), \psi(y) \rangle$ for all $x,y \in V$. Moreover, if $K$ is PD, then this holds with $\phi = \psi$.
\end{cor}
\subsubsection{Asymptotic rank formula}\label{sssec:rank-asymp}
While Proposition~\ref{prop:rank-exact} provides an exact formula for the rank of an arbitrary polynomial kernel over an arbitrary variety, it involves a matrix that is large even for moderate degree $n$ and dimension $d$. Proposition~\ref{prop:rank-asymp} provides a bound whose computation does not involve large matrices. The price to pay is that this bound is oblivious to the structure of $p$ beyond its degree. Nevertheless, this bound is tight for generic kernels. We now show how Proposition~\ref{prop:rank-asymp} follows from Proposition~\ref{prop:rank-exact}.
\begin{proof}[Proof of Proposition~\ref{prop:rank-asymp}]
By Proposition~\ref{prop:rank-exact}, a dimension bound, and the definition of the Hilbert function,
$\rank K = \rank(\Phi C \Phi^T) \leq \rank \Phi = \dim(\mathbb{R}_{\leq n}[V])
= \operatorname{HF}_{V}(n)$.
\end{proof}
We conclude with two remarks about Proposition~\ref{prop:rank-asymp}: tightness and common use cases.
\begin{remark}[Proposition~\ref{prop:rank-asymp} is generically tight]\label{rem:rank-asymp:tight}
The rank bound~\eqref{eq:rank-comp-HF} is exact for ``generic'' polynomial kernels. Indeed, the only inequality in the proof was $\rank \Phi C \Phi^T \leq \rank \Phi$, and this holds with equality by Proposition~\ref{prop:rank-exact} if the bilinear form $C$ is PD (e.g., for Taylor Features\footnote{Ignoring the exponential scalings which do not affect the rank, see Remark~\ref{rem:tay}.}).
\end{remark}
\begin{remark}[Proposition~\ref{prop:rank-asymp} for common polynomial kernels]\label{rem:rank-asymp:use}
Polynomial kernels of interest are typically rotation-invariant or isotropic; that is, of the form $R_n(x,y) = p_n(\langle x,y\rangle)$ or $K_n(x,y) = p_n(\|x-y\|^2)$, respectively, where $p_n$ is a univariate polynomial of degree $n$. Abuse notation slightly to denote the restrictions of these kernels to $V \times V$ by $R_n(V)$ and $K_n(V)$, respectively.
\begin{itemize}
\item \underline{Rotation-invariant kernels.}
In this case, Proposition~\ref{prop:rank-asymp} combined with the asymptotics in Lemma~\ref{lem:hilbert} gives the tight bound
\[
\rank R_n(V) \leq \operatorname{HF}_{V}(n) = \deg(V) n^{\dim V} + O(n^{\dim V - 1})
\]
which holds with equality for generic polynomials $p_n$ (and e.g., Taylor Features).
\item \underline{Isotropic kernels.}
In this case, a direct application of Proposition~\ref{prop:rank-asymp} gives a loose bound since it treats $p_n(\|x-y\|^2)$ as a generic polynomial of degree $2n$, leading to asymptotics of order
\[
\rank K_n(V) \leq \operatorname{HF}_{V}(2n) = \deg(V) (2n)^{\dim V} + O(n^{\dim V - 1}).
\]
A more refined analysis can essentially improve the $2n$ to $n$ by capturing the structure of the polynomial $p_n(\|x-y\|^2)$ beyond its degree.
This has an important effect in practice. However, it does not change the asymptotic bounds for kernel approximation (c.f., Theorem~\ref{thm:main-var}) since $n$ is only specified up to a constant anyways. As such, we do not investigate this further here.
\end{itemize}
\end{remark}
\section{Preliminaries}\label{sec:prelim}
In this section, we briefly recall relevant background about kernels in \S\ref{ssec:prelim:kernels} and algebraic geometry in \S\ref{ssec:prelim:ag}. Further details can be found in, e.g., the standard texts~\citep{WilRas06,SchSmo02,ShaCri04} for kernels and~\citep{Cox06,Cox13,Sha94} for algebraic geometry. In \S\ref{ssec:prelim:standard}, we describe two popular low-rank approximation approaches since we build upon them in the sequel. Readers familiar with any of these three topics should feel free to skip the corresponding sections. However, we note that the way we introduce previous approaches in \S\ref{ssec:prelim:standard} is non-standard: while the literature casts high-precision and high-dimensional approaches as fundamentally different, we attempt to introduce them in a somewhat unified manner.
\par \underline{Notation.} We denote the Euclidean norm by $\|\cdot\|$,
the $L^{\infty}$ norm over a domain $S$ by $\|\cdot\|_{S}$, the set $\{1, \dots, n\}$ by $[n]$,
and the set of positive (resp., non-negative) integers by $\mathbb{N}$ (resp., $\mathbb{N}_0)$.
We emphasize one non-standard notation: throughout a kernel is \emph{not} necessarily PSD unless explicitly specified.
We do this because our results in \S\ref{sec:high-prec} apply to polynomial kernels regardless of whether they are PSD, and this extension to indefinite kernels may be of interest since they are used in applications (e.g., the Fast Gauss Transform and Optimal Transport).
\subsection{Kernels}\label{ssec:prelim:kernels}
\paragraph*{Kernels and kernel matrices.} A \emph{kernel} $K : \mathcal{X} \times \mathcal{X} \to \mathbb{R}$ is a symmetric function of two arguments (not necessarily PSD, see above).
The \textit{kernel matrix} associated to a kernel $K$ and set of points $\{x_1, \dots, x_N\} \subset \mathcal{X}$ is the $N \times N$ symmetric matrix with $ij$-th entry $K(x_i,x_j)$.
\paragraph*{Types of kernels.} A kernel is \textit{positive semidefinite} (PSD) if for any $N \in \mathbb{N}$ and any set $\{x_1, \dots, x_N\}$, the corresponding $N \times N$ kernel matrix is a PSD matrix. Similarly, a kernel is \emph{positive definite} (PD) if all corresponding kernel matrices are PD matrices.
A \textit{polynomial kernel} is a kernel that is also a polynomial. The \emph{degree} of a polynomial kernel is the total degree in either of its arguments $x$ or $y$ (these two numbers are the same by symmetry).
A kernel is \textit{isotropic} if it is translation-invariant and rotation-invariant, i.e., $K(x,y)$ is a function only of $\|x-y\|$. An \textit{isotropic polynomial kernel} is a kernel of the form $p(\|x-y\|^2)$ where $p$ is a polynomial.
\paragraph*{Rank of kernels.} A \emph{rank-$r$ factorization} of a kernel $K : \mathcal{X} \times \mathcal{X} \to \mathbb{R}$ is a pair of ``feature maps'' $\phi, \psi : \mathcal{X} \to \mathbb{R}^r$ such that $K(x,y) = \langle \phi(x), \psi(y) \rangle$ for all $x,y$. The \textit{rank} of a kernel is the minimum $r$ for which there exists a rank-$r$ factorization. A PSD kernel admits a minimal-rank factorization with $\phi = \psi$. Observe that the rank of a kernel cannot increase upon restriction of the kernel's domain.
If a kernel $K$ admits a rank $r$ factorization $\langle \phi(x), \psi(y)\rangle$, then the $N \times N$ kernel matrix corresponding to any set of points $\{x_1, \dots, x_N\} \subset \mathcal{X}$ admits the factorization $A^TB$ where $A,B \in \mathbb{R}^{r \times N}$, and the $i$-th columns of $A$ and $B$ are respectively $\phi(x_i)$ and $\psi(x_j)$.
\subsection{Algebraic geometry}\label{ssec:prelim:ag}
\paragraph*{Polynomial spaces.}
We write $\mathbb{R}[x]$ to denote the polynomial ring of $d$-variate polynomials with real coefficients in the variable $x\in \mathbb{R}^d$. We write $\mathbb{R}_{n}[x]$ (resp., $\mathbb{R}_{\leq n}[x]$) to denote the linear space of polynomials in $\mathbb{R}[x]$ of degree $n$ (resp., at most $n$).
\begin{lemma}[Dimension of polynomial spaces]\label{lem:poly-dim}
Let $x \in \mathbb{R}^d$.
\begin{itemize}
\item \underline{Bounded-degree polynomials:} $\dim(\mathbb{R}_{\leq n}[x]) = \binom{n+d}{n}$.
\item \underline{Homogeneous polynomials:} $\dim(\mathbb{R}_n[x]) = \binom{n+d-1}{n}$.
\end{itemize}
\end{lemma}
\paragraph*{Varieties and ideals.} Throughout, $V$ is an (affine) \emph{real algebraic variety} in $\mathbb{R}^d$---or, variety for short---meaning that it is the set of points in $\mathbb{R}^d$ on which a set of polynomials in $\mathbb{R}[x]$ vanishes. The associated ideal of all vanishing polynomials is $I(V) = \{p \in \mathbb{R}[x] : p(x) = 0, \forall x \in V \}$. This is always a real radical ideal; in what follows, every ideal $I$ considered is real radical. The vanishing set for an ideal $I \subset \mathbb{R}[x]$ is the variety $V(I) = \{x \in \mathbb{R}^d : p(x) = 0, \forall p \in I\}$. Let $\mathbb{R}[V]$ denote the set of equivalence classes of polynomials in $\mathbb{R}[x]$, where two polynomials are identified if their restrictions to $V$ are identical. By the Real Nullstellensatz Theorem, $\mathbb{R}[V]$ is isomorphic to the coordinate ring $\mathbb{R}[x]/I(V)$. The spaces $\mathbb{R}[x]$ and $\mathbb{R}[V]$ form real vector spaces in the natural way.
\paragraph*{Dimension.} There are several equivalent definitions of the \emph{dimension} $\dim V$ of a variety $V$. An intuitive geometric definition is the maximum $d^{\star}$ for which there exists a sequence $V_0 \subsetneq V_1 \subsetneq \dots \subsetneq V_{d^{\star}}$ of irreducible subvarieties of $V$. Note that unlike manifolds, the dimension of a variety $V$ is not simply the dimension of the tangent space at any point $x \in V$---in fact, these tangent spaces might have dimension different from $\dim V$, or even be undefined altogether if the variety is not smooth at that point. For instance, the union of a disjoint plane and line is a $2$-dimensional variety. A variety is \emph{equidimensional} if each irreducible component has the same dimension.
\paragraph*{Degree.} The \emph{degree} $\deg V$ is the number of intersections over $\mathbb{C}^d$ (counted with intersection multiplicity) of $V$ with a subspace of co-dimension $\dim(V)$ in general position.
\paragraph*{Hilbertian quantities.} Let $I_{\leq n}(V)$ denote the set of polynomials of degree at most $n$ in the vanishing ideal $I(V)$. This is a vector subspace of $\mathbb{R}_{\leq n}[x]$. We denote the quotient space $\mathbb{R}_{\leq n}[x] / I_{\leq n}(V)$ by $\mathbb{R}_{\leq n}[V]$.
The \emph{Hilbert function}\footnote{This is sometimes called the affine Hilbert function to distinguish it from the projective Hilbert function. Similarly for the Hilbert series. We drop the word ``affine'' throughout since there is no confusion.} of $V$ is the function
$\operatorname{HF}_{V} : \mathbb{N}_0 \to \mathbb{N}_0$ defined by
\begin{align}
\operatorname{HF}_V(n)
= \dim\left( \mathbb{R}_{\leq n}[V] \right),
\label{eq:HF}
\end{align}
where the notion of dimension here is the one for vector spaces.
The \emph{Hilbert series} (a.k.a., Hilbert-Poincar\'e series) of $V$ is the generating function
\begin{align}
\operatorname{HS}_{V}(t) = \sum_{n=0}^{\infty} \operatorname{HF}_{V}(n) t^n,
\label{eq:HS}
\end{align}
viewed as a formal power series. The Hilbert function and series of a variety $V$ are identical to the corresponding Hilbert function and series for its vanishing ideal $I(V)$.
\begin{lemma}[Dimension and degree in Hilbertian quantities]\label{lem:hilbert}
Let $V$ be a variety.
\begin{itemize}
\item The Hilbert function $\operatorname{HF}_V(n)$ is a polynomial in $n$ for all sufficiently large $n$. This polynomial has degree $\dim V$ and leading coefficient $\deg V / \dim V!$.
\item The Hilbert series $\operatorname{HS}_V(t)$ can be expressed as a rational function in $t$ of the form $p_V(t)/(1-t)^{\dim V + 1}$, where $p_V$ is a polynomial with integer coefficients satisfying $p_V(1) = \deg V$.
\end{itemize}
\end{lemma}
\begin{lemma}~\citep{chardin1989majoration}\label{lem:hf-equidim}
If $V$ is an equidimensional variety, then
$\operatorname{HF}_V(n) \leq \deg V \binom{n + \dim V}{\dim V}$.
\end{lemma}
Throughout we present our results for equidimensional varieties; this assumption holds in all example varieties in this paper, and can be removed without changing the asymptotics in our results. Equidimensionality let us non-asymptotically bound the Hilbert Function via Lemma~\ref{lem:hf-equidim}; nevertheless, the same asymptotics $\operatorname{HF}_V(n) = \deg(V) n^{\dim V} + O(n^{\dim V-1})$ hold in general by Lemma~\ref{lem:hilbert}. The difference is that without equidimensionality, the non-asymptotic (a.k.a. transient) behavior of $\operatorname{HF}_V(n)$ can change since the lower-order terms
can depend on low-dimensional components of $V$. For instance, if $V$ is a line unioned with many points, then the dimension and degree of $V$ are dictated by the line (i.e., $\dim V = 1$ and $\deg V = 1$), whereby $\operatorname{HF}_V(n) = n + c_V$ for all sufficiently large $n$. However, the constant $c_V$ grows in the number of additional points in $V$.
Clearly some control on this effect is necessary since if the number of unioned points is sufficiently large, then the dimensionality of the coordinate ring increases.
\paragraph*{Computing properties of a variety.} A \emph{monomial ideal} is an ideal that is generated by monomials. If $I(V)$ is a monomial ideal, then the Hilbert function $\operatorname{HF}_{V}(n)$ is equal to the number of monomials in $\mathbb{R}_{\leq n}[x]$ that are not in $I$. These monomials are called \emph{standard monomials} and can be counted via inclusion-exclusion given a list of monomial generators for $I$. The dimension and degree of $V$ can then be read off from the Hilbert function via Lemma~\ref{lem:hilbert}.
\par Of course, $I(V)$ is not always a monomial ideal. For general varieties $V$, the Hilbert function can be computed algorithmically using Gr\"obner bases. Fix a graded monomial ordering on $\mathbb{R}[x]$.
The \emph{leading term} of a polynomial is the largest monomial w.r.t. that ordering.
The \emph{leading term ideal} $\operatorname{LT}(I)$ of $I$ is the monomial ideal generated by the leading term of each element of $I$. A \emph{Gr\"obner basis} for $I$ is a finite set of generators $\{g_i\}$ such that $\operatorname{LT}(I)$ equals the ideal generated by $\{\operatorname{LT}(g_i)\}$. The reason that a Gr\"obner basis helps to compute Hilbert functions is the following lemma, which reduces the general case of arbitrary ideals $I$ to the simpler case of monomial ideals.
\begin{lemma}[Reduction from general ideals to monomial ideals]\label{lem:LT}
Let $I$ be an ideal in $\mathbb{R}[x]$. For any graded monomial ordering on $\mathbb{R}[x]$, the Hilbert functions of $V(I)$ and $V(\operatorname{LT}(I))$ are identical.
\end{lemma}
Thus, given a Gr\"obner basis $\{g_i\}$ of an arbitrary ideal $I$, one can form a description of the leading term ideal $\operatorname{LT}(I)$ as the monomial ideal generated by $\{\operatorname{LT}(g_i)\}$, and then use this to compute the Hilbert function of $V(I)$ by counting the number of standard monomials in $\operatorname{LT}(I)$. This machinery is demonstrated through several concrete examples when we use it in \S\ref{sec:high-prec} and \S\ref{sec:ex}.
\input{sections/prelim_standard}
\subsection{Standard approaches for kernel approximation}\label{ssec:prelim:standard}
Here we describe two of the most popular low-rank approximation approaches since we build upon them in the sequel. These are polynomial-based approaches~\citep{YanDurDav05,CotKesSre11,WanLiDar18,zwicknagl2009power, schaback2008limit, wathen2015spectral} and Random Fourier Features (RFF)~\citep{RahRec08,liu2020random}. The former is suited for high-precision regimes, whereas the latter is suited for high-dimensional regimes.
Each of these two approaches consists of two steps:
\begin{enumerate}
\item Expand the relevant kernel
\begin{align}
K(x,y) = \int f_{\omega}(x) f_{\omega}(y) d\nu(\omega) \label{eq:prelim:int-rep}
\end{align}
as a convex combination of rank-$1$ functions. Here $\nu$ is a probability distribution---continuous if the representation~\eqref{eq:prelim:int-rep} is an integral, or discrete if~\eqref{eq:prelim:int-rep} is a sum.
\item Form a rank-$r$ approximation by taking $r$ of these infinitely many rank-$1$ functions.
\end{enumerate}
\par To explain the difference between polynomial-based approaches and RFF, let us begin with how they perform step (2). On one hand, polynomial-based approaches greedily choose the $r$ rank-$1$ functions $f_{\omega}(x)f_{\omega}(y)$ with largest weight $d\nu(\omega)$. On the other hand, RFF independently samples $r$ random rank-$1$ functions $f_{\omega}(x)f_{\omega}(y)$ according to the distribution $d\nu(\omega)$.
\par This difference in step (2) necessitates strikingly different kinds of representations~\eqref{eq:prelim:int-rep} in step (1). Intuitively, the greedy truncation scheme performs well on representations in which $\nu$ is a discrete distribution with rapidly decaying tails. In contrast, the random sampling scheme performs well on representations in which the functions $f_{\omega}$ have small magnitude. In particular, the representations~\eqref{eq:prelim:int-rep} used in step (1) are as follows.
\begin{itemize}
\item \underline{Polynomial-based representation}:
\begin{align}
K(x,y)
=
\sum_{n=0}^{\infty} c_n p_n(x,y)
=
\sum_{n=0}^{\infty} c_n \sum_{\alpha=1}^{\binom{n+d-1}{n}} u_{n,\alpha}(x)v_{n,\alpha}(y).
\label{eq:tay-rep}
\end{align}
This representation first expands $K$ into the sum of polynomial kernels $p_n$ of degree at most $n$ (typically via monomial expansions or Chebyshev expansions), and then factorizes each $p_n$. The inner sum is over $\binom{n+d-1}{n}$ polynomials because this is the dimension of the space of degree-$n$ homogeneous polynomials on $\mathbb{R}^d$.
\item \underline{RFF representation}:
\begin{align}
K(x,y) = \mathbb{E}_{\omega,\theta}
\left[ \cos(\langle \omega, x \rangle + \theta) \cos(\langle \omega, y \rangle + \theta) \right].
\label{eq:rff-int-rep}
\end{align}
Here $\omega$ is sampled from the Fourier transform $\mu$ of $k(x-y) = K(x,y)$; this is a probability distribution by Bochner's Theorem if $\mu$ is a continuous, PD, translation-invariant kernel normalized so that $K(0,0)=1$, see e.g.,~\citep{WilRas06}. Independently, $\theta$ is sampled from the uniform distribution over $[0,2\pi)$. This representation is obtained by simple trigonometric manipulation of the Fourier transform identity.
\end{itemize}
Given that both approaches seek to optimize the $L^{\infty}$ error metric, a natural question is why use one representation and not the other? The answer is based on the parameter regime.
\par On one hand, the fact that the representation~\eqref{eq:tay-rep} is a finite sum with coefficients that decay exponentially fast if $K$ is smooth, means that exponentially small error is obtained by truncating. This is critical for high-precision regimes. However, the issue with this approach is that the rank grows as $\Omega(n)^d$ in the truncation degree $n$, and this prohibitive beyond low dimensions $d$.
\par On the other hand, the fact that the integrand in the RFF representation~\eqref{eq:rff-int-rep} is bounded in magnitude by $1$, means that sampling-based quadrature converges at standard statistical rates which scale well in the dimension $d$. This is critical for high-dimensional regimes. However, the issue with this approach is that statistical rates require roughly $\Omega(1/\varepsilon^2)$ samples in order to obtain $\varepsilon$ accuracy, which is prohibitive for accuracies $\varepsilon$ beyond a few digits of precision.
\par Details on each of these methods and their formal guarantees follow.
\subsubsection{High-precision approximation via polynomial features}
As discussed in \S\ref{ssec:intro:cont}, there are many ways to quantify smoothness of a kernel. For simplicity, we assume (1) isotropy, meaning that the kernel admits a representation $K(x,y) = f(\|x-y\|^2)$; and (2) the univariate function $f$ satisfies the following smoothness condition. Note that we write $K(x,y) = f(\|x-y\|^2)$ rather than $f(\|x-y\|)$ since in the latter case, $K$ might not be smooth even if $f$ is; e.g., take $f$ to be the identity.
\begin{assumption}\label{assump:smooth}
There exist constants $\alpha > 0$ and $\beta \in (0,1)$ such that for all $n \in \mathbb{N}$, there is a polynomial $p_n$ of degree $n$ satisfying $\|f - p_n\|_{[0,4]} \leq \alpha \cdot \beta^{n}$.
\end{assumption}
A classical result of Bernstein from over a century ago shows that this assumption is essentially equivalent to analyticity of $f$ in a complex neighborhood around the approximation domain.
See also~\citep{Tre13} for other smoothness conditions that lead to fast rates for polynomial approximation.
\begin{lemma}[Analyticity implies approximation~\citep{bernstein1912ordre}]\label{lem:smooth}
Suppose $f$ is analytically continuable to the Bernstein ellipse $E_{\rho} = \mathrm{Interior}(\{z + z^{-1} + 2 : z \in \mathbb{C}, |z| = \rho\})$ of parameter $\rho > 1$ around $[0,4]$. Then $f$ satisfies Assumption~\ref{assump:smooth} with $\alpha = 2 \|f\|_{E_{\rho}}/(\rho-1)$ and $\beta = 1/\rho$.
\end{lemma}
This lemma ensures that Assumption~\ref{assump:smooth} is satisfied for popular kernels such as the Gaussian kernel, in which case $f(t) = e^{-t/2}$ is entire, and the Cauchy kernel, in which case $f(t) = (1+t/2)^{-1}$ has a pole at $-3/2$ and thus is analytically continuable to any Bernstein ellipse $E_{\rho}$ with parameter $\rho < \rho_{\max} = 7/4 + \sqrt{33}/4 \approx 3.186$.
Standard rates for approximating smooth isotropic kernels are immediate from combining this implication of smoothness with simple rank bounds on bounded-degree polynomial kernels, see e.g.,~\citep{WanLiDar18}. For completeness, we provide a short proof.
\begin{prop}[Standard rates for high-precision approximation]\label{prop:poly-standard}
Suppose $K(x,y) = f(\|x-y\|^2)$ where $f$ satisfies Assumption~\ref{assump:smooth}. There is a universal constant $c$ such that for all $\varepsilon > 0$, there is a kernel of rank
\[
r \leq
\left(c \frac{\log(\alpha/\varepsilon)}{\log (1/\beta)} \right)^{d}
= O\left(\log1/\varepsilon \right)^d.
\]
that approximates $K$ on $\mathbb{B}^d \times \mathbb{B}^d$ to $L^{\infty}$ error $\varepsilon$.
\end{prop}
\begin{proof}
By Assumption~\ref{assump:smooth}, there is a polynomial $p_n$ of degree $n = \lceil \log(\alpha/\varepsilon)/\log(1/\beta) \rceil$ satisfying $\|f-p_n\|_{[0,4]} \leq \varepsilon$. Thus the kernel $K_n(x,y) = p_n(\|x-y\|^2)$ satisfies $\|K - K_n\|_{\mathbb{B}^d \times \mathbb{B}^d} \leq \varepsilon$. Since $K_n$ has degree at most $2n$ in each argument, it follows from Lemma~\ref{lem:poly-dim} and a crude bound that $\rank K_n \leq \dim(\mathbb{R}_{\leq 2n}[x]) = \binom{2n+d}{d} = O(n)^d$.
\end{proof}
\begin{remark}[Taylor Features]\label{rem:tay}
For the Gaussian kernel $G(x,y) = e^{-\|x-y\|^2/(2\sigma^2)}$, one can obtain rank bounds which are slightly better in practice albeit the same asymptotically. The trick is to factor out the scalings $e^{-(\|x\|^2+\|y\|^2)/(2\sigma^2)}$ and then approximate the remainder $e^{\langle x,y \rangle/\sigma^2}$ via a rotation-invariant polynomial kernel $p_n(\langle x, y\rangle)$. The point is that the scaling factors do not affect the rank, and a rotation-invariant polynomial kernel $p_n(\langle x,y \rangle)$ generically has lower rank than an isotropic polynomial kernel $q_n(\|x-y\|^2)$ for $p_n$ and $q_n$ of the same degree (although both ranks are asympotically the same $O(n)^d$). Specifically, the popular Taylor Features kernel is
$
T_n(x,y) = e^{-(\|x\|^2+\|y\|^2)/(2\sigma^2)} \sum_{k=0}^n \tfrac{\langle x,y \rangle^k}{\sigma^{2k} k!}$, see e.g.,~\citep{YanDurDav05,YanDurGum03,CotKesSre11}.
\end{remark}
\subsubsection{High-dimensional approximation via Random Fourier Features}
The RFF kernel of rank $r$ is the empirical mean of $r$ samples of the integral representation~\eqref{eq:rff-int-rep}; that is,
\begin{align}
K_r(x,y) := \frac{1}{r} \sum_{i=1}^r \cos(\langle \omega_i, x_i \rangle + \theta_i) \cos(\langle \omega_i, y \rangle + \theta_i)
\label{eq:rff}
\end{align}
where $\omega_1, \dots, \omega_r$ are sampled from the Fourier transform $\mu$ of $k(x-y) = K(x,y)$, and $\theta_1, \dots, \theta_r \sim \operatorname{Unif}([0,2\pi))$ are all sampled independently. The guarantees of this approach are summarized as follows; see~\citep{RahRec08} for a proof.
\begin{assumption}\label{assump:rff}
$K$ is a continuous, positive-definite, translation-invariant kernel on $\mathbb{R}^d$ with normalization $K(0,0)=1$ and curvature $\sigma_K^2$.
\end{assumption}
\begin{prop}[Standard rates for high-dimensional approximation]\label{prop:rff-standard}
Suppose $K$ satisfies Assumption~\ref{assump:rff}. Then the kernel $K_r$ in~\eqref{eq:rff} has rank at most $r$ and satisfies $\| K - K_r \|_{\mathbb{B}^d \times \mathbb{B}^d} \leq \varepsilon$ with any constant probability for
\[
r = O\left( d \,\frac{\log(\sigma_K / \varepsilon)}{\varepsilon^2} \right).
\]
\end{prop}
Note that for a kernel $K$ satisfying Assumption~\ref{assump:rff}, its curvature $\sigma_K^2$, defined as the trace of the Hessian of $k(x-y) = K(x,y)$ at $0$, is equal to $\mathbb{E}_{\omega \sim \mu} \|\omega\|^2$~\citep{RahRec08}.
\section{Kernel approximation over a variety: high-dimensional regime}\label{sec:hd}
Here we show that high-dimensional approaches for kernel approximation perform substantially better if the approximation domain is a low-dimensional algebraic variety $V$ in a high-dimensional ambient space $\mathbb{R}^d$. Specifically, we improve the rank bound~\eqref{eq:rate-high-dim} of standard approaches by showing that their dependence on the ambient dimension $d$ can be improved to dependence on $\dim V$.
\begin{theorem}[High-dimensional approximation over a variety]\label{thm:high-dim}
Suppose $K$ is a kernel satisfying Assumption~\ref{assump:rff}. Suppose also $\mathcal{X} = V \cap \mathbb{B}^d$, where $V \subset \mathbb{R}^d$ is an equidimensional real algebraic variety satisfying $\deg V \leq e^{\dim V}$. For any $\varepsilon > 0$, there exists a kernel of rank
\begin{align}
r =
O\left( \frac{\dim V \, \log (\sigma_K \dim V /\varepsilon)}{\varepsilon^2}
\right)
\label{thm:high-dim-rank}
\end{align}
that approximates $K$ on $\mathcal{X} \times \mathcal{X}$ to $L^{\infty}$ error $\varepsilon$.
\end{theorem}
We prove this existential result in an algorithmic way. We use the Random Fourier Features kernel $K_r$ defined in~\eqref{eq:rff}, except with one minor technical modification: rather than sampling frequencies $\omega_1, \dots, \omega_r$ from the Fourier distribution, we sample them from a truncated version of it. That is, we re-sample a frequency if its squared norm is large (roughly $\mathbb{E}\|\omega\|^2/\varepsilon = \sigma_K^2/\varepsilon$). We prove that this construction works with high probability: if the rank is
\begin{align}
r = O\left( \frac{\dim V \, \log (\sigma_K \dim V /\varepsilon) + \logdel}{\varepsilon^2} \right),
\label{thm:high-dim-rank-hp}
\end{align}
then this RFF kernel approximates $K$ on $\mathcal{X} \times \mathcal{X}$ to $L^{\infty}$ error $\varepsilon$ with probability at least $\delta$. This implies the existential result in Theorem~\ref{thm:high-dim} by taking $\delta$ to be any constant.
\begin{remark}[Degree]\label{rem:high-dim-deg}
Dependence on $\deg V$ is unavoidable, see \S\ref{ssec:intro:cont}. In Theorem~\ref{thm:high-dim}, we assume for simplicity of presentation that $\deg V$ is not exponentially large, since in this case the contribution of $\deg V$ to the rank bound~\eqref{thm:high-dim-rank} is negligible. The proof extends to arbitrary degree essentially without change\footnote{
The only difference is that the size $\log |\mathcal{X}_n|$ of the norming sets in Proposition~\ref{prop:norming} increases by $\log \deg V + \dim V \log \log \deg V$
in order to balance terms in the tensoring proof. This slight increase in $\log |\mathcal{X}_n|$ results in an analogous increase in the final rank bound in Theorem~\ref{thm:high-dim-rank} since $r$ scales linearly in $\log |\mathcal{X}_n|$, see~\eqref{eq:rff-pf:r}.}
if the rank $r$ is increased by $\varepsilon^{-2} (\log \deg V + \dim V \log \log \deg V)$.
\end{remark}
\begin{remark}[Logarithmic dependence]
For the Gaussian kernel $G(x,y) = e^{-\|x-y\|^2/(2\sigma^2)}$ and the Cauchy kernel $C(x,y) = (1+\|x\|^2/(2\sigma^2))^{-1}$, Theorem~\ref{thm:high-dim} depends logarithmically on $d$ since $\log \sigma_K = \log(d/\sigma^2)$. By a similar proof technique (in particular using the key Proposition~\ref{prop:norming} about norming sets over varieties), we can show that an alternative approach based on polynomial approximation and then Johnson-Lindenstrauss projection achieves similar guarantees to Theorem~\ref{thm:high-dim}, with $\log d$ improved to $\log \dim V$. This generalizes to smooth isotropic kernels.
However, we focus on RFF since its much better algorithmic efficiency outweighs this lower-order term.
\end{remark}
See \S\ref{ssec:intro:tech} for an overview of the proof of Theorem~\ref{thm:high-dim}. As explained there, the key ingredient beyond the standard RFF analysis is to exploit the rigidity of polynomials over varieties to ensure the existence of ``norming sets'' of small size. We develop this ingredient in \S\ref{ssec:hd-norming}, and then use it to prove Theorem~\ref{thm:high-dim} in \S\ref{ssec:hd-pf}.
\subsection{Constant norming sets for varieties}\label{ssec:hd-norming}
Given a compact set $\mathcal{X} \subset \mathbb{R}^d$, slack $\lambda > 1$, and degree $n \in \mathbb{N}$, a \emph{norming set} is a subset $\mathcal{X}_n \subseteq \mathcal{X}$ satisfying $\|p\|_{\mathcal{X}} \leq \lambda\|p\|_{\mathcal{X}_n}$ for all polynomials $p$ of degree $n$. That norming sets of small cardinality exist is a classical result in approximation theory with many applications, for example bounding the operator norm of the interpolation projection operator~\citep{bos2018fekete}.
\par In order to prove Theorem~\ref{thm:high-dim}, we require a version of this standard existential result that is strengthened in two important ways. One is that we require the slack $\lambda$ to be constant (say $2$), rather than the large $\binom{n+d}{d}$ that is sufficient for standard applications in interpolation theory. The other is that we require the cardinality of the norming set to not grow exponentially in the ambient dimension $d$; the standard bound is $|\mathcal{X}_n| = \binom{n+d}{d} = O(n)^d$. This requires exploiting the fact that in the setup of this paper, $\mathcal{X}$ is a compact subset of an algebraic variety of dimension $\dim V \ll d$.
\begin{prop}[Constant norming set for variety]\label{prop:norming}
Let $\mathcal{X}$ be a compact subset of an equidimensional variety $V \subset \mathbb{R}^d$ satisfying $\deg V \leq e^{\dim V}$. For any integer $n \in \mathbb{N}$, there is a set $\mathcal{X}_n \subset \mathcal{X}$ of size $\log |\mathcal{X}_n| = O( \dim V \log(n \dim V))$ satisfying $\|p\|_{\mathcal{X}} \leq 2 \|p\|_{\mathcal{X}_n}$ for all polynomials $p \in \mathbb{R}_{\leq n}[x]$.
\end{prop}
These two ``tightenings'' of the standard result are achieved by combining the classical argument (based on Fekete sets) with a tensoring trick inspired by Proposition 23 of~\citep{bloom2012polynomial}.
Let us first introduce this classical argument. Our exposition is based on~\citep{bos2018fekete}; see that nice survey for further background.
\par Fix a compact domain $\mathcal{X} \subset \mathbb{R}^d$ and degree $n \in \mathbb{N}$. Let $N$ denote the dimension of the space $\mathbb{R}_{\leq n}[\mathcal{X}]$ of polynomials of degree at most $n$ restricted to $\mathcal{X}$, and let $\{p_1, \dots, p_N\}$
denote any basis of this space.
Define $\mathrm{vdm}(x_1, \dots, x_N)$ to be the determinant of the $N \times N$ Vandermonde matrix with $ij$-th entry $p_i(x_j)$. A \emph{Fekete set} for $\mathcal{X}$ of degree $n$ is a maximizer of
\[
\max_{\mathcal{F}_n \subset \mathcal{X}, \, |\mathcal{F}_n| = N} \abs{\mathrm{vdm}(\mathcal{F}_n)}.
\]
(Note that this definition is independent of the choice of basis $\{p_1, \dots, p_N\}$.) A basic fact about Fekete sets is that they are norming sets, albeit of large size $|\mathcal{F}_n| = N$ and for large slack $\lambda = N$.
\begin{lemma}[Fekete sets are norming sets; folklore]\label{lem:fekete}
Consider any compact domain $\mathcal{X} \subset \mathbb{R}^d$ and degree $n \in \mathbb{N}$.
Set $N = \dim(\mathbb{R}_{\leq n}[\mathcal{X}])$.
Then any Fekete set $\mathcal{F}_n$ satisfies $\|p\|_{\mathcal{X}} \leq N \|p\|_{\mathcal{F}_n}$.
\end{lemma}
\begin{proof}
This proof is folklore; we sketch it for completeness and refer to e.g.,~\citep{bos2018fekete} for details. Let $P$ denote the interpolation projection operator which given a continuous function $f$ over $\mathcal{X}$, outputs a polynomial $P(f)$ of degree at most $n$ which interpolates $f$ at $\mathcal{F}_n$. Then
\[
P(f)(x) = \sum_{i=1}^N f(x_i) \ell_i(x),
\]
where $\ell_i(x)$ is the $i$-th Lagrange interpolating polynomial for $\mathcal{F}_n$. Because of the classical identity
$\ell_i(x) = \mathrm{vdm}(x_1, \dots, x_{i-1},x, x_{i+1}, \dots, x_N) / \mathrm{vdm}(x_1, \dots, x_{i-1},x_i, x_{i+1}, \dots, x_N)$, it follows by definition of $\mathcal{F}_n$ being a Fekete set that
\[
\|\ell_i\|_{\mathcal{X}} = 1.
\]
Thus, for any polynomial $p \in \mathbb{R}_{\leq n}[x]$, we have
\[
\|p\|_{\mathcal{X}}
=
\left\|P(p)\right\|_{\mathcal{X}}
=
\left\|\sum_{i=1}^N p(x_i) \ell_i\right\|_{\mathcal{X}}
\leq
\sum_{i=1}^N |p(x_i)| \cdot \left\|\ell_i\right\|_{\mathcal{X}}
\leq
N\|p\|_{\mathcal{F}_n}.
\]
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop:norming}]
For shorthand, denote $\dim V$ by $d^{\star}$. For integer $a \in \mathbb{N}$ chosen shortly, let $\mathcal{X}_n = \mathcal{F}_{an}$ be a Fekete set for $\mathcal{X}$ of degree $an$. Since $|\mathcal{F}_{an}|$ is the dimension of the space $\mathbb{R}_{\leq an}[V]$,
\[
|\mathcal{X}_n|
=
|\mathcal{F}_{an}|
=
\operatorname{HF}_V(an)
\leq
\deg V \binom{an + d^*}{d^*}
\leq
\deg V \cdot (2ean)^{d^{\star}}
\]
by Lemma~\ref{lem:hf-equidim} and a crude bound.
By the assumption $\deg V \leq e^{d^{\star}}$, this implies
\begin{align}
|\mathcal{X}_n| \leq (15 an)^{d^{\star}}.
\label{eq:norming-proof}
\end{align}
Now observe that if
$\deg(p) \leq n$, then $\deg(p^{a}) \leq an$, thus
\[
\|p\|_{\mathcal{X}} = \|p^a\|_{\mathcal{X}}^{1/a} \leq \left( |\mathcal{F}_{an}| \cdot \|p^a\|_{\mathcal{F}_{an}} \right)^{1/a} = |\mathcal{F}_{an}|^{1/a} \cdot \|p\|_{\mathcal{X}_n},
\]
by the norming property of Fekete sets (Lemma~\ref{lem:fekete}). Choosing $a \geq 1$ strengthens the tensoring bound $|\mathcal{F}_{an}|^{1/a}$. In particular, choosing $a = \lceil c d^* \log(n d^*) \rceil$ for an appropriate constant $c$ ensures that
$|\mathcal{F}_{an}|^{1/a}
\leq
(15an)^{d^{\star}/a}
=
\exp( \frac{d^{\star}}{a} \log (15an) )
\leq 2$
as well as the desired guarantee on $\log|\mathcal{X}_n|$.
\end{proof}
\subsection{Proof of Theorem~\ref{thm:high-dim}}\label{ssec:hd-pf}
\paragraph*{Construction: RFF on truncated Fourier distribution.} By Bochner's Theorem,
\[
K(x,y) = \mathbb{E}_{\omega \sim \mu} \left[ e^{i \langle \omega, x - y\rangle} \right].
\]
Define
\begin{align}
\tilde{K}(x,y) = \mathbb{E}_{\tilde{\omega} \sim \tilde{\mu}} \left[e^{i \langle \tilde{\omega}, x - y \rangle} \right],
\label{eq:rff-int-rep-tilde}
\end{align}
where $\tilde{\mu}(\tilde{\omega}) = \frac{1}{1-p} \mu(\tilde{\omega}) \cdot \mathds{1}[\|\tilde{\omega}\|^2 \leq t]$
is $\mu$ truncated to the ball of squared norm $t = 2\mathbb{E}_{\omega \sim \mu} \|\omega\|^2 / \varepsilon = 2\sigma_K^2/\varepsilon$. Here, $p = \mathbb{P}_{\omega \sim \mu} [\|\omega\|^2 > t]$ so that $\tilde{\mu}$ is rescaled to a probability distribution.
\par The approximation we construct is the RFF approximation of $\tilde{K}$; that is,
\[
\tilde{K}_r(x,y) = \frac{1}{r}\sum_{i=1}^r e^{i \langle \omega, x - y\rangle},
\]
where $\tilde{\omega_1}, \dots, \tilde{\omega_r} \sim \tilde{\mu}$ are drawn independently. Note that $\tilde{K}$ has rank $2r$ by expanding the complex exponential into sinusoids~\citep{RahRec08}. We show that $\tilde{K}_r$ approximates $K$ well with high probability.
\paragraph*{Analysis step 1: Error bound for a fixed pair of points}
As in~\citep{RahRec08}, applying Hoeffding's inequality for sampling the integral representation~\eqref{eq:rff-int-rep-tilde} implies that for any fixed $x,y \in \mathcal{X}$,
\begin{align}
\mathbb{P}\left[ \abs{\tilde{K}(x,y) - {\tilde{K}}_r(x,y)} \leq \varepsilon \right]
\geq
1 - e^{- r\varepsilon^2/2 }.
\label{eq:rff-pf:hoeffding}
\end{align}
\paragraph*{Analysis step 2: Extending the error bound to the whole domain}
This is where the proof critically deviates from~\citep{RahRec08}: rather than union bound over an $\varepsilon$-net of $\mathcal{X}$, we union bound over a norming set for $\mathcal{X}$ and exploit algebraic properties of the domain.
\underline{Fourier distribution truncation.} By Markov's inequality,
\[
p = \mathbb{P}_{\omega \sim \mu}\left[ \|\omega\|^2 > t \right] \leq \frac{\mathbb{E}_{\omega \sim \mu}\|\omega\|^2}{t} = \frac{\varepsilon}{2}.
\]
Thus the kernel $\tilde{K}$ is uniformly close to $K$ because for all $x,y$,
\begin{align}
\abs{K(x,y) - \tilde{K}(x,y)}
=
\abs{
\mathbb{E}_{\omega \sim \mu} \left[ e^{i \langle \omega, x - y\rangle} \right]
-
E_{\tilde{\omega} \sim \tilde{\mu}} \left[ e^{i \langle \tilde{\omega}, x - y\rangle} \right]
}
\leq 2p
\leq \varepsilon.\label{eq:rff-pf:K-Ktilde}
\end{align}
Above, the second step is by conditioning on $\|w\|^2 \leq t$ and bounding the integrands by $1$.
\underline{Polynomial approximation.} We approximate $\tilde{K}$ and $\tilde{K_r}$ by low-degree polynomials. Since their Fourier Transforms are compactly supported, it can be shown (see Lemma~\ref{lem:ft-smooth}) that there exist polynomial kernels $P_n$ and $Q_n$ of degree $n = O(t + \log 1/\varepsilon) = O(\sigma_K^2/\varepsilon)$ that satisfy
\begin{align}
\|\tilde{K} - P_n\|_{\mathbb{B}^d \times \mathbb{B}^d} \leq \varepsilon \quad \text{ and } \quad
\|\tilde{K_r} - Q_n\|_{\mathbb{B}^d \times \mathbb{B}^d} \leq \varepsilon.
\label{eq:rff-pf:poly}
\end{align}
\underline{Using the norming set.} Let $\mathcal{X}_n \subset \mathcal{X}$ be the norming set guaranteed by Proposition~\ref{prop:norming}. Then by a union bound over~\eqref{eq:rff-pf:hoeffding} for all $(x,y) \in \mathcal{X}_n \times \mathcal{X}_n$, we have
\begin{align}
\mathbb{P}\left[ \|\tilde{K} - \tilde{K}_r\|_{\mathcal{X}_n \times \mathcal{X}_n} \leq \varepsilon \right] \geq 1 - \delta
\label{eq:rff-pf:union-bound}
\end{align}
if the rank $r$ is at least
\begin{align}
r \geq \frac{2}{\varepsilon^2} \log \left( \frac{|\mathcal{X}_n|^2}{\delta} \right).
\label{eq:rff-pf:r}
\end{align}
This gives the desired rank bound~\eqref{thm:high-dim-rank} by plugging in the bound $\log |\mathcal{X}_n| = O(\dim(V) \log (n \dim(V)))$ by Proposition~\ref{prop:norming}, and the definition of $n = O(\sigma_K^2/\varepsilon)$.
Moreover, in this success event of~\eqref{eq:rff-pf:union-bound},
\begin{align*}
\|K - \tilde{K}_r\|_{\mathcal{X} \times \mathcal{X}}
&\leq
\|\tilde{K} - \tilde{K}_r\|_{\mathcal{X} \times \mathcal{X}} + \varepsilon
\leq
\| P_n - Q_n\|_{\mathcal{X} \times \mathcal{X}} + 3\varepsilon
\leq
2\|P_n - Q_n\|_{\mathcal{X}_n \times \mathcal{X}} + 3\varepsilon
\\&\leq
4\|P_n - Q_n\|_{\mathcal{X}_n \times \mathcal{X}_n} + 3\varepsilon
\leq
4\|\tilde{K} - \tilde{K}_r\|_{\mathcal{X}_n \times \mathcal{X}_n} + 11\varepsilon
\leq
15\varepsilon.
\end{align*}
Above, the first inequality is by the uniform approximation~\eqref{eq:rff-pf:K-Ktilde} of $K$ by $\tilde{K}$. The second and penultimate inequalities are by replacing $\tilde{K}$ and $\tilde{K}_r$ with the respective polynomial approximations $P_n$ and $Q_n$, see~\eqref{eq:rff-pf:poly}. The third inequality is because $\mathcal{X}_n$ is a constant norming set (cf. Proposition~\ref{prop:norming}) and the fact that $P_n(x,y) - Q_n(x,y)$ is a degree $n$ polynomial in $x$ for fixed $y$; and vice versa for the fourth inequality. The final inequality is by the error bound~\eqref{eq:rff-pf:union-bound} on $\mathcal{X}_n \times \mathcal{X}_n$. Rescaling $\varepsilon$ by a constant factor $1/15$ concludes the proof.
|
\section{Nearly-Properly Learning GMMs: Full Version}\label{sec:GMM-full}
In this section, we complete the proof of our main result for learning GMMs, Theorem \ref{thm:main-gmm}. The high-level outline of the proof is similar to the proof of Theorem \ref{thm:fourier-main}, We localize the distribution by multiplying by a Gaussian multiplier $M_{\mu, \sigma^2}$. Note that the product of two Gaussians is still a Gaussian so multiplying a GMM by a Gaussian multiplier results in a re-weighted mixture of Gaussians. Roughly, we argue that the new weights on components of the mixture that are far away from the multiplier $M_{\mu, \sigma^2}$ are negligible so the resulting mixture is well-conditioned and we can then use Corollary \ref{coro:learn-tight-mixture} to reconstruct the localized distribution. To reconstruct the entire distribution, we show that it suffices to sum together $\wt{O}(k)$ different localized reconstructions.
\subsection{Localizing with Gaussian Multipliers}
First, we explicitly compute what happens when we have a Gaussian $G_1 = N(\mu_1, \sigma_1^2)$ and we multiply it by a Gaussian multiplier $M_{\mu, \sigma^2}(x) $.
\begin{claim}\label{claim:mult-by-gaussian}
We have the identity
\[
M_{\mu, \sigma^2}(x)N(\mu_1, \sigma_1^2) = \frac{1}{\sqrt{1 + \frac{\sigma_1^2}{\sigma^2}}}e^{-\frac{(\mu_1 - \mu)^2}{2(\sigma_1^2 + \sigma^2)}} N\left( \frac{\mu \sigma_1^2 + \mu_1 \sigma^2}{\sigma_1^2 + \sigma^2} , \frac{\sigma_1^2 \sigma^2}{\sigma_1^2 + \sigma^2} \right) \,.
\]
\end{claim}
\begin{proof}
We prove the above through direct computation.
\begin{align*}
M_{\mu, \sigma^2}(x)G_1(x) = e^{-\frac{(x - \mu)^2}{2\sigma^2} -\frac{(x - \mu_1)^2}{2\sigma_1^2}} \frac{1}{\sigma_1 \sqrt{2\pi }} = \frac{1}{\sigma_1 \sqrt{2\pi }} \cdot e^{-\frac{1}{2}\left( \left(\frac{1}{\sigma^2} + \frac{1}{\sigma_1^2} \right)x^2 - 2 \left( \frac{\mu}{\sigma^2} + \frac{\mu_1}{\sigma_1^2} \right)x + \frac{\mu^2}{\sigma^2} + \frac{\mu_1^2}{\sigma_1^2}\right)} \\ = \frac{1}{\sigma_1 \sqrt{2\pi }} \cdot \exp \left( -\frac{1}{2}\left( \sqrt{\frac{1}{\sigma^2} + \frac{1}{\sigma_1^2}}x - \frac{\frac{\mu}{\sigma^2} + \frac{\mu_1}{\sigma_1^2}}{\sqrt{\frac{1}{\sigma^2} + \frac{1}{\sigma_1^2}}} \right)^2 - \frac{1}{2} \cdot \frac{(\mu_1 - \mu)^2}{\sigma_1^2 + \sigma^2} \right) \\ = \frac{1}{\sqrt{1 + \frac{\sigma_1^2}{\sigma^2}}}e^{-\frac{(\mu_1 - \mu)^2}{2(\sigma_1^2 + \sigma^2)}} N\left( \frac{\mu \sigma_1^2 + \mu_1 \sigma^2}{\sigma_1^2 + \sigma^2} , \frac{\sigma_1^2 \sigma^2}{\sigma_1^2 + \sigma^2} \right) \,.
\end{align*}
\end{proof}
\subsection{ Building Blocks}
We first consider reconstructing a GMM $\mcl{M} = w_1G_1 + \dots w_kG_k$ after multiplying by a Gaussian multiplier $M_{\mu, \sigma^2}$. As a corollary of Claim \ref{claim:mult-by-gaussian}, we know that when the $C$-significant intervals (recall Definition \ref{def:sig-interval}) of a Gaussian $G_j$ and the multiplier $M_{\mu, \sigma^2}(x) $ are disjoint for large $C$, then the $L^1$ norm of their product is $e^{-\Omega(C^2)}$. In particular this means that after multiplying by $M_{\mu, \sigma^2}$, the only components that remain relevant are those that have nontrivial overlap with the multiplier $M_{\mu, \sigma^2}$. The only way these components will not form a well-conditioned mixture is if there is some $G_j$ that is very thin (i.e. $\sigma_j << \sigma$) and overlaps with $M_{\mu, \sigma^2}$. As long as this doesn't happen, we can apply Corollary \ref{coro:learn-tight-mixture}. We formalize this below.
\begin{corollary}\label{coro:reconstruct-multiplier}
Let $\mcl{M} = w_1G_1 + \dots + w_kG_k$ be an arbitrary mixture of Gaussians where $G_i = N(\mu_i, \sigma_i^2)$. Let $\eps > 0$ be some parameter and let $l = \lceil \sqrt{\log (1/\eps)} \rceil $. Assume we are given access to a distribution $f$. Let $M_{\mu, \sigma^2}$ be a Gaussian multiplier. Assume that for all $i \in [k]$, either $\sigma_i \geq 4l\sigma$ or the $10 l$-significant intervals of $G_i$ and $M_{\mu, \sigma^2}$ do not intersect. Then in $\poly(1/\eps)$ time and with high probability, we can compute a weighted sum $\wt{M}$ of at most $O(\log (1/\eps)) $ Gaussians such that
\[
\norm{\wt{M} - M_{\mu, \sigma^2}f }_1 \leq \eps + \poly(\log (1/\eps)) \norm{ M_{\mu, \sigma^2} (\mcl{M} - f)}_1 \,.
\]
\end{corollary}
\begin{proof}
We compute $M_{\mu, \sigma^2}f $ and let $C = \norm{M_{\mu, \sigma^2}f}_1$. If $C \leq \eps$ then we may simple output $0$. Otherwise, we will apply Corollary \ref{coro:learn-tight-mixture} on $M_{\mu, \sigma^2}f /C$ and multiply the result by $C$. We must first verify the conditions of Corollary \ref{coro:learn-tight-mixture}. Let $S \subset [k]$ be the indices such that the $10l$-significant intervals of $G_i$ and $M_{\mu, \sigma^2}$ intersect. First for $i \notin S$, by Claim \ref{claim:mult-by-gaussian},
\[
\norm{G_i M_{\mu, \sigma^2}}_1 \leq e^{-\frac{(\mu_i - \mu)^2}{2(\sigma_i^2 + \sigma^2)}} \leq e^{-10l^2} \leq \eps^{10} \,.
\]
Let
\[
\mcl{M}' = \sum_{i \in S} w_iG_i \,.
\]
Then we know
\[
\norm{ \frac{M_{\mu, \sigma^2}f}{C} - \frac{M_{\mu, \sigma^2} \mcl{M}'}{C} }_1 \leq \eps^{9} + \frac{\norm{M_{\mu, \sigma^2}(f - \mcl{M} )}_1}{C}
\]
Next, for $i \in S$,
\[
G_i M_{\mu, \sigma^2} = w_i' N\left( \frac{\mu \sigma_i^2 + \mu_i \sigma^2}{\sigma_i^2 + \sigma^2} , \frac{\sigma_i^2 \sigma^2}{\sigma_i^2 + \sigma^2} \right)
\]
for some weight $w_i'$ and since we must have $\sigma_i \geq 4l\sigma$, then
\begin{align*}
&\frac{\sigma^2}{2} \leq \frac{\sigma_i^2 \sigma^2}{\sigma_i^2 + \sigma^2} \leq \sigma^2 \\
&\left\lvert \frac{\mu \sigma_i^2 + \mu_i \sigma^2}{\sigma_i^2 + \sigma^2} - \mu \right \rvert = \left \lvert \frac{(\mu_i - \mu) \sigma^2}{\sigma_i^2 + \sigma^2} \right \rvert \leq \frac{l(\sigma + \sigma_i) \sigma^2}{\sigma_i^2 + \sigma^2} \leq \frac{\sigma}{2}
\end{align*}
Let
\[
\mcl{M}'' = \sum_{i \in S}\frac{w_i'}{\sum_{i \in S} w_i'} N\left( \frac{\mu \sigma_i^2 + \mu_i \sigma^2}{\sigma_i^2 + \sigma^2} , \frac{\sigma_i^2 \sigma^2}{\sigma_i^2 + \sigma^2} \right) \,.
\]
Then we deduce, since $\norm{M_{\mu, \sigma^2}f/C}_1 = 1$, that
\begin{align*}
\norm{ \frac{M_{\mu, \sigma^2}f}{C} - \mcl{M}'' }_1 \leq \norm{\mcl{M}'' - \frac{M_{\mu, \sigma^2} \mcl{M}'}{C} }_1 + \norm{ \frac{M_{\mu, \sigma^2}f}{C} - \frac{M_{\mu, \sigma^2} \mcl{M}'}{C} }_1 \leq 2 \norm{ \frac{M_{\mu, \sigma^2}f}{C} - \frac{M_{\mu, \sigma^2} \mcl{M}'}{C} }_1 \\ \leq \eps^{8} + 2\frac{\norm{M_{\mu, \sigma^2}(f - \mcl{M} )}_1}{C}
\end{align*}
and further, after applying a suitable linear transformation (taking $(\mu, \sigma^2) \rightarrow (0,1)$) that the mixture $\mcl{M}''$ is $0.5$-well-conditioned. Thus, we can apply Corollary \ref{coro:learn-tight-mixture} and compute a weighted sum of $O(\log (1/\eps))$ Gaussians, $\wt{\mcl{M}}$ such that
\[
\norm{\frac{M_{\mu, \sigma^2}f}{C} - \wt{\mcl{M}}}_1 \leq \poly(\log (1/\eps))\left( \eps^{8} + \frac{\norm{M_{\mu, \sigma^2}(f - \mcl{M} )}_1}{C} \right) \,.
\]
Now we can simply output $C \wt{\mcl{M}}$ (which is still a weighted sum of $O(\log (1/\eps))$ Gaussians) and we are done.
\end{proof}
Recall that Theorem \ref{thm:approx-interval} shows how to express an interval as a sum of Gaussian multipliers. Combining Theorem \ref{thm:approx-interval} with Corollary \ref{coro:reconstruct-multiplier}, we show that we can approximate a GMM over an interval as long as the interval does not overlap with a component that is much thinner than it.
\begin{lemma}\label{lem:reconstruct-interval}
Let $\mcl{M} = w_1G_1 + \dots + w_kG_k$ be an arbitrary mixture of Gaussians where $G_i = N(\mu_i, \sigma_i^2)$. Let $\eps > 0$ be some parameter and let $l = \lceil \sqrt{\log (1/\eps)} \rceil $. Assume we are given access to a distribution $f$. Let $I = [a,b]$ be an interval. Assume that for all $i \in [k]$, either $\sigma_i \geq (b - a)$ or the $10 l$-significant interval of $G_i$ does not intersect $I$. Then in $\poly(1/\eps)$ time and with high probability, we can compute a weighted sum $\wt{\mcl{M}}$ of at most $\poly(\log (1/\eps)) $ Gaussians such that
\[
\norm{\wt{\mcl{M}} - f \cdot 1_{I} }_1 \leq \poly(\log (1/\eps)) \left( \eps + \norm{ 1_{I} (\mcl{M} - f)}_1 \right)
\]
where $1_{I}$ denotes the indicator function of $I$.
\end{lemma}
\begin{proof}
Consider the function $\mcl{I} = \mcl{I}_{\eps , (b-a)/2}^{(a + b)/2}$ (recall Definition \ref{def:interval-function}). Now note that by Theorem \ref{thm:approx-interval}, $\mcl{I}$ can be written in the form
\[
\mcl{I} = \wt{w_1} M_{\wt{\mu_1}, \wt{\sigma_1}^2} + \dots + \wt{w_n} M_{\wt{\mu_n}, \wt{\sigma_n}^2}
\]
where $n = O(\log^2 1/\eps)$. Furthermore, for all $i \in [n]$, we have $0 \leq \wt{w_i} \leq 1$ and $\wt{\sigma_i} \leq (b - a)/(4l)$ and the $10l$-significant intervals of $M_{\wt{\mu_i}, \wt{\sigma_i}^2}$ are all contained in the interval $[a,b]$. Thus we can apply Corollary \ref{coro:reconstruct-multiplier} on $M_{\wt{\mu_i}, \wt{\sigma_i}^2} f$ for all $i \in [n]$. Adding the results with the corresponding weights $\wt{w_1}, \dots , \wt{w_n}$, we obtain a function $\wt{\mcl{M}}$ that is a weighted sum of at most $\poly(\log (1/\eps)) $ Gaussians such that
\[
\norm{\wt{\mcl{M}} - f \mcl{I}}_1 \leq \poly(\log (1/\eps)) \left( \eps + \sum_{i = 1}^n \wt{w_i}\norm{ M_{\wt{\mu_i}, \wt{\sigma_i}^2} (\mcl{M} - f)}_1 \right) = \poly(\log (1/\eps)) \left( \eps + \norm{ \mcl{I} (\mcl{M} - f)}_1 \right) \,.
\]
Thus,
\begin{equation}\label{eq:l1bound1}
\norm{\wt{\mcl{M}} - \mcl{M} \mcl{I}}_1 \leq \poly(\log (1/\eps)) \left( \eps + \norm{ \mcl{I} (\mcl{M} - f)}_1 \right) \,.
\end{equation}
Next, by the properties in Theorem \ref{thm:approx-interval},
\[
\norm{\mcl{M}(\mcl{I} - 1_I )}_1 \leq \eps \int_{-\infty}^{\infty} \mcl{M} + \int_{a}^{a + \eps(b- a)} \mcl{M} + \int_{b - \eps(b - a)}^b \mcl{M} \leq \eps + \left( \sum_{j=1}^k w_j \int_{a}^{a + \eps(b- a)} G_j + w_j\int_{b - \eps(b - a)}^b G_j \right)\,.
\]
Consider one of the component Gaussians $G_j$ where $j \in [k]$. If the $10l$-significant interval of $G_j$ does not intersect $[a,b]$ then it is clear that the total mass of $G_j$ on the interval $[a,b]$ is at most $\eps$. Otherwise, we know that the standard deviation of $G_j$ is at least $b-a$ which means that its mass on the set $[a, a+ \eps(b - a)] \cup [b - \eps(b - a), b]$ is at most $O(\eps)$. Thus we conclude that
\begin{equation}\label{eq:l1bound2}
\norm{\mcl{M}(\mcl{I} - 1_I )}_1 \leq O(\eps) \,.
\end{equation}
Also note that by the properties in Theorem \ref{thm:approx-interval}
\begin{equation}\label{eq:l1bound3}
\norm{ \mcl{I} (\mcl{M} - f)}_1 \leq \eps \int_{-\infty}^{\infty} |\mcl{M} - f| + 2\int_{a}^b |\mcl{M} - f| \leq 2(\eps + \norm{1_I(\mcl{M} - f)}_1) \,.
\end{equation}
Putting together (\ref{eq:l1bound1}, \ref{eq:l1bound2}, \ref{eq:l1bound3}) , we conclude
\[
\norm{\wt{\mcl{M}} - f \cdot 1_I}_1 \leq \norm{ 1_{I} (\mcl{M} - f)}_1 + \norm{\wt{\mcl{M}} - \mcl{M}1_I}_1 \leq \poly(\log (1/\eps)) \left( \eps + \norm{ 1_{I} (\mcl{M} - f)}_1 \right)
\]
and we are done.
\end{proof}
\subsection{Structural Properties}
Lemma \ref{lem:reconstruct-interval} allows us to reconstruct the unknown GMM $\mcl{M}$ over certain intervals. However, it cannot be applied to an arbitrary interval (because an interval may overlap with a component that is too thin). We will now prove several structural results that will imply that there exist $\wt{O}(k)$ intervals for which the conditions of Lemma \ref{lem:reconstruct-interval} are satisfied (i.e. these intervals do not overlap with components that are much thinner than themselves) and such that the union of these intervals contains most of the mass of $\mcl{M}$. Then, to complete the proof of Theorem \ref{thm:main-gmm}, we show how to find such a set of $\wt{O}(k)$ intervals using a dynamic program.
First, we define a modified density function for a GMM $\mcl{M} = w_1G_1 + \dots + w_kG_k$ where we modify each component Gaussian by restricting it to its $10l$-significant interval (and making it $0$ outside). It is clear that this modified function is close to $\mcl{M}$ in $L^1$-distance but it will be convenient to use in the analysis later on.
\begin{definition}
For a mixture of Gaussians $\mcl{M} = w_1G_1 + \dots + w_kG_k$ where $G_j = N(\mu_j, \sigma_j^2)$ and a parameter $l$, define the function $\mcl{M}_{\text{sig},l}(x)$ to be, at each point $x \in \R$, equal to the weighted sum of the components $G_j$ of $\mcl{M}$ such that $x$ is in the $10l$-significant interval of $G_j$. Formally,
\[
\mcl{M}_{\text{sig},l}(x) = \sum_{\substack{j \text{ such that}\\ |x - \mu_j| \leq 10l\sigma_j}} w_j G_j(x) \,.
\]
\end{definition}
The following claim is immediate from the definition.
\begin{claim}\label{claim:significant-function}
Let $\eps > 0$ be some parameter and let $l = \lceil \sqrt{\log 1/\eps} \rceil$. Then
\[
\norm{\mcl{M} - \mcl{M}_{\text{sig}, l}}_1 \leq \eps
\]
\end{claim}
\begin{proof}
The inequality holds because the total mass of a Gaussian outside of its $10l$-significant interval is at most $\eps$.
\end{proof}
We now present our first structural result.
\begin{claim}\label{claim:structural-characterization}
Let $\mcl{M} = w_1G_1 + \dots + w_kG_k$ be an arbitrary mixture of Gaussians where $G_i = N(\mu_i, \sigma_i^2)$. Let $\eps > 0$ be some parameter and let $l = \lceil \sqrt{\log (1/\eps)} \rceil $. There exist disjoint intervals $I_1, \dots , I_n$ with lengths, say $t_1, \dots , t_n$, where $n \leq 50kl$ with the following property:
\begin{itemize}
\item For each interval $I_i$, for all $j \in [k]$ either the the $10l$-significant interval of $G_j$ is disjoint from $I_i$ or $\sigma_j \geq t_i$
\item We have
\[
\norm{\mcl{M} - (1_{I_1} + \dots + 1_{I_n})\mcl{M}}_1 \leq \eps
\]
\end{itemize}
\end{claim}
\begin{proof}
Sort the Gaussians by their standard deviations, WLOG $\sigma_1 \leq \dots \leq \sigma_k$. Now we will create several intervals $A_1, A_2, \dots $ and we will also associate each interval with one of the Gaussians $G_1, \dots , G_k$ which we will call its parent.
First, set $A_1$ to be the $10l$-significant interval of $G_1$. Next, we will process the Gaussians $G_2, \dots , G_k$ in order. For $G_j$, assume that the intervals we have created so far are $A_1, \dots , A_m$ (which will be disjoint by construction). Now consider the $10l$-significant interval of $G_j$, say $L_j$. Note that removing the union of the intervals $A_1, \dots , A_m$ from $L_j$ divides $L_j$ into several (at most $m+1$) disjoint intervals. We label these intervals $A_{m+1}, A_{m+2}, \dots $ and set all of their parents to be $G_j$. We then move onto $G_{j+1}$ and repeat the above process. The following properties are immediate from the construction:
\begin{enumerate}
\item If the parent of $A_i$ is $G_j$ then the length of $A_i$ is at most $20l\sigma_j$
\item The union of all of the $A_i$ whose parent is among $G_1, \dots , G_j$ contains the $10l$-significant intervals of all of $G_1, \dots , G_j$
\item If the parent of $A_i$ is $G_j$, then $A_i$ is disjoint from the $10l$-significant intervals of $G_1, \dots , G_{j-1}$
\end{enumerate}
Now we claim that at the end of the algorithm, the total number of intervals is at most $2k$. To see this, say that after processing $G_{j-1}$, the intervals we have created are $A_1, \dots A_m$. Now consider the potential that is the number of intervals $m$ plus the number of connected components in $A_1 \cup \dots \cup A_m$. This potential can increase by at most $2$ when processing $G_j$ so thus the total number of intervals at the end of the execution is at most $2k$. We will now assume that the intervals at the end of the execution are $A_1, \dots , A_{2k}$ (if there are less than $2k$ intervals then add a bunch of dummy intervals of length $0$).
We now describe a post-processing step. For each of $A_1, \dots , A_{2k}$, if its parent is $G_j$ then divide it into intervals of length at most $\sigma_j$ and assign $G_j$ as the parent of all of these intervals. By property 1, we can ensure that this creates a total of at most $50kl$ intervals, say $I_1, \dots , I_{n}$ where $n \leq 50kl$. We use $t_1, \dots , t_n$ to denote their lengths. We now prove that this set of intervals satisfies the desired properties. First, note that the following properties are immediate from the construction:
\begin{enumerate}
\item If the parent of $I_i$ is $G_j$ then $I_i$ is contained in the $10l$-significant interval of $G_j$ and $t_i \leq \sigma_j$
\item The union of all of $I_1, \dots , I_{n}$ contains the $10l$-significant intervals of all of $G_1, \dots , G_k$
\item If the parent of $I_i$ is $G_j$, then $I_i$ is disjoint from the $10l$-significant intervals of $G_1, \dots , G_{j-1}$
\end{enumerate}
The first of the desired properties is clear since by construction if the parent of $I_i$ is $G_j$, then $t_i \leq \sigma_j$ and it must be disjoint from the $10l$-significant intervals of all of $G_1, \dots , G_{j-1}$ (where recall $G_1, \dots , G_k$ are sorted in increasing order of their standard deviation). Now it remains to verify the second property. Consider the function $\mcl{M}_{\text{sig},l}(x)$. Recall by Claim \ref{claim:significant-function},
\begin{equation}\label{eq:approxbound1}
\norm{\mcl{M} - \mcl{M}_{\text{sig}, l}}_1 \leq \eps
\end{equation}
Next observe that
\[
\mcl{M}_{\text{sig}, l} =(1_{I_1} + \dots + 1_{I_{n}}) \mcl{M}_{\text{sig}, l} \,.
\]
Combining the above, we have
\[
\norm{\mcl{M} - (1_{I_1} + \dots + 1_{I_{n}}) \mcl{M}_{\text{sig},l}}_1 \leq \eps \,.
\]
However, note that we have
\[
(1_{I_1} + \dots + 1_{I_{n}}) \mcl{M}_{\text{sig},l} \leq (1_{I_1} + \dots + 1_{I_{n}}) \mcl{M} \leq \mcl{M}
\]
everywhere along the real line. Thus, we immediately get the desired inequality.
\end{proof}
The next structural result shows that the intervals $I_1, \dots , I_n$ obtained in Claim \ref{claim:structural-characterization} are ``findable" in the sense that if we draw many samples from $\mcl{M}$ (or from a distribution $f$ that is close to $\mcl{M}$), then with high probability, there will be samples close to the endpoints of each of $I_1, \dots , I_n$. This will mean that algorithmically, it suffices to draw sufficiently many samples and then only consider intervals whose endpoints are given by a pair of samples.
\begin{claim}\label{claim:structural-part2}
Let $\mcl{M} = w_1G_1 + \dots + w_kG_k$ be an arbitrary mixture of Gaussians where $G_i = N(\mu_i, \sigma_i^2)$. Let $\eps > 0$ be some parameter and let $l = \lceil \sqrt{\log (1/\eps)} \rceil $. Let $f$ be a distribution. Assume we are given samples from $f$, say $x_1, \dots , x_Q$ for some sufficiently large $Q = \poly(k/\eps)$. Then with high probability, there exists pairs $\{x_{a_1}, x_{b_1} \} ,\dots , \{x_{a_n}, x_{b_n} \} $ such that
\begin{itemize}
\item The intervals $ J_1 = [x_{a_1}, x_{b_1}], \dots , J_n = [x_{a_n}, x_{b_n}] $ are disjoint \item $n \leq 50kl$
\item For each interval $J_i$, for all $j \in [k]$ either the the $10l$-significant interval of $G_j$ is disjoint from $J_i$ or $\sigma_j \geq |x_{b_i} - x_{a_i}|$
\item
\[
\norm{\mcl{M} -(1_{J_1} + \dots + 1_{J_n}) \mcl{M} }_1 \leq 4(\eps + \norm{\mcl{M} - f}_1) \,.
\]
\end{itemize}
\end{claim}
\begin{proof}
Let $I_1, \dots , I_n$ be the intervals computed in Claim \ref{claim:structural-characterization} applied to the mixture $\mcl{M}$ and assume that their lengths are $t_1, \dots , t_n$. Let $C = \lceil (k/\eps)^2 \rceil$. For each interval $I_i$, divide it into $C$ subintervals $I_i^1, \dots , I_i^C$ of length $t_i/C$ and assume that these subintervals are sorted in order. We say one of these subintervals is good if
\[
\int_{I_i^j} f \geq (\eps/k)^{10} \,.
\]
For an index $i$, let $c_i, d_i$ be the smallest and largest index such that $I_i^{c_i}, I_i^{d_i}$ are good respectively. Then with high probability for all $i$, there will be samples, say $x_{a_i}, x_{b_i}$ in $I_i^{c_i}$ and $I_i^{d_i}$. Now we will form the intervals $J_i = [x_{a_i}, x_{b_i}]$. The first two of the desired properties are clear. The third follows from the statement of Claim \ref{claim:structural-characterization}. It remains to verify the last. Similar to the proof of Claim \ref{claim:structural-characterization}, we consider the function $\mcl{M}_{\text{sig},l}$. Note that
\begin{align*}
&\int_{I_i \backslash J_i} \mcl{M}_{\text{sig},l} \leq \int_{I_i^1} \mcl{M}_{\text{sig},l} + \dots + \int_{I_i^{c_i}} \mcl{M}_{\text{sig},l} + \int_{I_i^{d_i}} \mcl{M}_{\text{sig},l} + \dots + \int_{I_i^C} \mcl{M}_{\text{sig},l} \\ &\leq \norm{1_{I_i}(\mcl{M}_{\text{sig},l} - f)}_1 + \int_{I_i^1} f + \dots + \int_{I_i^{c_i - 1}}f + \int_{I_i^{d_i + 1}} f + \dots + \int_{I_i^C} f + \int_{I_i^{c_i}} \mcl{M}_{\text{sig},l} + \int_{I_i^{d_i}} \mcl{M}_{\text{sig},l} \\ & \leq \norm{1_{I_i}(\mcl{M}_{\text{sig},l} - f)}_1 +(\eps/k)^2 + \int_{I_i^{c_i}} \mcl{M}_{\text{sig},l} + \int_{I_i^{d_i}} \mcl{M}_{\text{sig},l}
\end{align*}
where the last step follows by the minimality and maximality of $c_i, d_i$. Now by construction, the only Gaussians among $G_1, \dots , G_k$ whose $10l$-significant intervals intersect $I_i$ must have standard deviation at least $t_i$. Since $I_i^{c_i}, I_i^{d_i}$ each have length $(\eps/k)^2t_i$, we conclude
\[
\int_{I_i \backslash J_i} \mcl{M}_{\text{sig},l} \leq 3(\eps/k)^2 + \norm{1_{I_i}(\mcl{M}_{\text{sig},l} - f)}_1\,.
\]
Thus, we have
\[
\norm{(1_{I_1} + \dots + 1_{I_n}) \mcl{M}_{\text{sig},l} - (1_{J_1} + \dots + 1_{J_n}) \mcl{M}_{\text{sig},l}}_1 \leq \eps + \norm{(\mcl{M}_{\text{sig},l} - f)}_1 \leq 2\eps + \norm{(\mcl{M} - f)}_1\,.
\]
where we used Claim \ref{claim:significant-function} in the last step. Now, using the statement of Claim \ref{claim:structural-characterization}, we deduce
\[
\norm{ \mcl{M} - (1_{J_1} + \dots + 1_{J_n}) \mcl{M}_{\text{sig},l}}_1 \leq 4\eps + \norm{(\mcl{M} - f)}_1 \,.
\]
Since
\[
(1_{J_1} + \dots + 1_{J_{n}}) \mcl{M}_{\text{sig},l} \leq (1_{J_1} + \dots + 1_{J_{n}}) \mcl{M} \leq \mcl{M}
\]
everywhere along the real line, we immediately get the desired inequality.
\end{proof}
\subsection{Finishing the Proof}
We can now prove the key lemma and then Theorem \ref{thm:main-gmm} will follow as an immediate consequence since we can use the improper learner in Corollary \ref{coro:improper-learner}. The lemma states that given explicit access to a distribution $f$ that is $\eps$-close to a GMM, $\mcl{M}$, with $k$ components, we can output a GMM, $\wt{\mcl{M}}$, with $\wt{O}(k)$ components that is $\wt{O}(\eps)$-close to $f$. At a high-level, the proof involves attempting to reconstruct $f$ over various intervals using Lemma \ref{lem:reconstruct-interval} and then using a dynamic program to find a union of $\wt{O}(k)$ such intervals that approximates the entire function. We use Claim \ref{claim:structural-part2} to argue that such a solution exists so our dynamic program must find it.
\begin{lemma}\label{lem:reconstruct-Gaussian}
Let $\mcl{M} = w_1G_1 + \dots + w_kG_k$ be an arbitrary mixture of Gaussians where $G_i = N(\mu_i, \sigma_i^2)$. Let $\eps > 0$ be some parameter. Assume we are given access to a distribution $f$. Then in $\poly(k/\eps)$ time and with high probability, we can compute a mixture $\wt{\mcl{M}}$ of at most $k\poly(\log (k/\eps)) $ Gaussians such that
\[
\norm{\wt{\mcl{M}} - f }_1 \leq \poly(\log (k/\eps)) \left( \eps + \norm{ \mcl{M} - f}_1 \right) \,.
\]
\end{lemma}
\begin{proof}
Note that it suffices to compute a weighted sum of Gaussians that satisfies the desired inequality since rescaling such a weighted sum to a mixture will at most increase the $L^1$ error by a factor of $2$. Thus, from now on, we will not worry about ensuring the mixing weights sum to $1$.
Let $\gamma = \eps/k$ and $l = \lceil \sqrt{\log 1/\gamma}\rceil $. First draw $Q = \poly(k/\eps)$ samples $x_1, \dots , x_Q$ from $f$ for sufficiently large $Q$ that we can apply Claim \ref{claim:structural-part2} with $\eps \leftarrow \gamma$. While we do not know what the intervals $J_1, \dots , J_n$ are, we will set up a dynamic program to find a set of at most $50kl $ intervals that we can reconstruct $f$ over each one using Lemma \ref{lem:reconstruct-interval} and such that these intervals contain essentially all of the mass of $f$.
\\\\
For each pair $x_a, x_b $ with $a,b \in \{1,2, \dots , Q \}$, apply Lemma \ref{lem:reconstruct-interval} with parameter $\eps \leftarrow \gamma$ to attempt to approximate $f$ on the interval $[x_a,x_b]$. Let the output obtained be $\wt{\mcl{M}}_{x_a,x_b}$. Note that sometimes the algorithm will fail to output a good approximation to $f$ (because the assumptions of Lemma \ref{lem:reconstruct-interval} fail) but we can ensure that the output is a weighted sum of at most $\poly(\log 1/\gamma)$ Gaussians. In the proof we will only use the fact that when the assumptions of Lemma \ref{lem:reconstruct-interval} hold, then our approximation to $f$ restricted to the interval will be accurate.
\\\\
Now we show how to set up the dynamic program. WLOG the points $x_1, \dots , x_Q$ are sorted in nondecreasing order. We also use the convention that $x_0 = -\infty, x_{Q+1} = \infty$. Now, we maintain the following state for each index $0 \leq j \leq Q + 1$, and integer $c \leq 50kl$: the best approximation to $f$ from $(-\infty, x_j]$ using a sum of $\wt{\mcl{M}}_{x_a,x_b}$ over at most $c$ intervals. Formally,
\paragraph{Dynamic Program:} Let $DP_{j,c}$ be the minimum over all sets $S$ of $c$ pairs $(a^{(1)},b^{(1)}), \dots , (a^{(c)},b^{(c)}) \in [j] \times [j]$ such that $a^{(1)} < b^{(1)} \leq a^{(2)} < \dots < b^{(c)}$ of
\begin{align*}
&\norm{f\cdot 1_{(-\infty, x_j]} - \left(1_{[x_{a^{(1)}},x_{b^{(1)}}]}\wt{\mcl{M}}_{x_{a^{(1)}},x_{b^{(1)}}} + \dots + 1_{[x_{a^{(c)}},x_{b^{(c)}}]} \wt{\mcl{M}}_{x_{a^{(c)}},x_{b^{(c)}}} \right) }_1 \\ &+ \norm{\wt{\mcl{M}}_{x_{a^{(1)}},x_{b^{(1)}}} - 1_{[x_{a^{(1)}},x_{b^{(1)}}]}\wt{\mcl{M}}_{x_{a^{(1)}},x_{b^{(1)}}}}_1 + \dots + \norm{\wt{\mcl{M}}_{x_{a^{(c)}},x_{b^{(c)}}} - 1_{[x_{a^{(c)}},x_{b^{(c)}}]}\wt{\mcl{M}}_{x_{a^{(c)}},x_{b^{(c)}}}}_1 \,.
\end{align*}
Note that the first term represents the approximation error compared to $f$. We must truncate each function $\wt{\mcl{M}}_{x_{a^{(1)}},x_{b^{(1)}}}$ to its corresponding interval $[x_{a^{(1)}},x_{b^{(1)}}]$ in order for the problem to be solvable via dynamic programming because otherwise previous choices would affect later ones. Thus, we also need to add the additonal terms that represent the error from truncation. Note that the $L^1$ distance can be estimated using Claim \ref{claim:testL1}. We can solve the dynamic program in polynomial time because from each state $DP_{j,c}$, we simply consider adding all possible intervals among $x_j, x_{j+1}, \dots , x_Q$ as the next one.
\\\\
Now we prove that there is a good solution to the dynamic program for which the objective (for $j = Q+1$) is small. Let $a_1,b_1, \dots , a_n, b_n$ be the indices obtained in Claim \ref{claim:structural-part2}. We claim that setting
\[
(a^{(1)},b^{(1)}) = (a_1, b_1) ,\dots , (a^{(n)},b^{(n)}) = (a_n, b_n)
\]
results in the objective function being small. Let $J_i = [a_i,b_i]$ for all $i$. Let
\[
\wt{\mcl{M}}^{\text{good}} = 1_{[x_{a_1}, x_{b_1}]} \wt{\mcl{M}}_{x_{a_1},x_{b_1}} + \dots + 1_{[x_{a_n}, x_{b_n}]} \wt{\mcl{M}}_{x_{a_n},x_{b_n}} \,.
\]
The guarantee from Lemma \ref{lem:reconstruct-interval} implies that for all $i$,
\begin{equation}\label{eq:trunc-bound}
\norm{ \wt{\mcl{M}}_{x_{a_i},x_{b_i}} - 1_{[x_{a_i}, x_{b_i}]} \wt{\mcl{M}}_{x_{a_i},x_{b_i}}}_1 \leq \poly(\log 1/\gamma)(\gamma + \norm{1_{J_i}(\mcl{M} - f)}_1)
\end{equation}
and thus, using the guarantee from Lemma \ref{lem:reconstruct-interval} again, we have
\begin{align*}
\norm{\wt{\mcl{M}}^{\text{good}} - (1_{J_1} + \dots + 1_{J_n} )f } \leq \norm{\wt{\mcl{M}}_{x_{a_1},x_{b_1}} - 1_{J_1} \cdot f}_1 + \dots + \norm{\wt{\mcl{M}}_{x_{a_n},x_{b_n}} - 1_{J_n} \cdot f}_1 \\ \quad + \norm{ \wt{\mcl{M}}_{x_{a_1},x_{b_1}} - 1_{[x_{a_1}, x_{b_1}]} \wt{\mcl{M}}_{x_{a_1},x_{b_1}}}_1 + \dots + \norm{ \wt{\mcl{M}}_{x_{a_n},x_{b_n}} - 1_{[x_{a_n}, x_{b_n}]} \wt{\mcl{M}}_{x_{a_n},x_{b_n}}}_1 \\ \leq n \gamma \poly(\log 1/\gamma) + \poly(\log 1/\gamma) \left( \norm{1_{J_1}(\mcl{M} - f)}_1 + \dots + \norm{1_{J_n}(\mcl{M} - f)}_1\right) \\ \leq \poly(\log 1/\gamma)\left( \eps + \norm{\mcl{M} - f }_1 \right) \,.
\end{align*}
However also recall that by Claim \ref{claim:structural-part2},
\[
\norm{\mcl{M} -(1_{J_1} + \dots + 1_{J_n}) \mcl{M} }_1 \leq 4(\gamma +\norm{\mcl{M} - f }_1) \,.
\]
Combining the above two inequalities, we get
\[
\norm{\wt{\mcl{M}}^{\text{good}} - f}_1 \leq \poly(\log 1/\gamma)\left( \eps + \norm{\mcl{M} - f }_1 \right) \,.
\]
Finally, combining the above with (\ref{eq:trunc-bound}) implies that the objective value of the dynamic program is at most $ \poly(\log 1/\gamma)\left( \eps + \norm{\mcl{M} - f }_1 \right)$. Finally it remains to note that the objective of the dynamic program (for $j = Q+1$) is an upper bound on
\[
\norm{f - \left(\wt{\mcl{M}}_{x_{a^{(1)}},x_{b^{(1)}}} + \dots + \wt{\mcl{M}}_{x_{a^{(c)}},x_{b^{(c)}}} \right) }_1
\]
so thus, we can simply output the solution $\wt{\mcl{M}} = \wt{\mcl{M}}_{x_{a^{(1)}},x_{b^{(1)}}} + \dots + \wt{\mcl{M}}_{x_{a^{(c)}},x_{b^{(c)}}} $ and we are guaranteed to have
\[
\norm{f -\wt{\mcl{M}} }_1 \leq \poly(\log 1/\gamma)\left( \eps + \norm{\mcl{M} - f }_1 \right) \,.
\]
It is clear that $\wt{\mcl{M}}$ is a weighted sum of at most $k \poly(\log 1/\gamma)$ Gaussians (since for each interval there are $\poly(\log 1/\gamma)$ Gaussians and there are at most $50kl$ total intervals). It is clear that all of the steps run in $\poly(k/\eps)$ time and we are done.
\end{proof}
Now we can complete the proof of our main theorem, Theorem \ref{thm:main-gmm}.
\begin{proof}[Proof of Theorem \ref{thm:main-gmm}]
We can apply Corollary \ref{coro:improper-learner} to learn a distribution $f$ such that $d_{\TV}(\mcl{M} , f) \leq O(\eps)$. We can then apply Lemma \ref{lem:reconstruct-Gaussian} using $f$. Note that since $f$ is a piecewise polynomial, we can perform all of the explicit computations with the density function that are used in the proof of Lemma \ref{lem:reconstruct-Gaussian}. It is immediate that the output of Lemma \ref{lem:reconstruct-Gaussian} must satisfy
\[
d_{\TV}(\wt{\mcl{M}}, \mcl{M} ) \leq \wt{O}(\eps)
\]
so we are done.
\end{proof}
\section{Sparse Fourier Reconstruction: Full Version}\label{sec:sparse-fourier-full}
In this section, we complete the proof of our main result on sparse Fourier reconstruction, Theorem \ref{thm:fourier-main}. The key lemma that goes into the proof is stated below. At a high level, the lemma states that if we know roughly where the Fourier support of the unknown Fourier-sparse signal $\mcl{M}$ is located, then we can successfully reconstruct it.
\begin{lemma}\label{lem:fourier-manycluster}
Assume we are given $N, k, \eps , c$ with $0 < \eps < 0.1$. Let $l = \lceil \log kN / (\eps c) \rceil$ be some parameter. Let $\mcl{M}$ be a function that is $(k,1)$-simple. Also, assume that we are given a set $T \subset \R$ of size $N$ such that all of the support of $\wh{\mcl{M}}$ is within distance $1$ of $N$. Further, assume we are given access to a function $f$ such that
\[
\int_{-1}^1 |f(x) - \mcl{M}(x)|^2 dx \leq \eps^2 \,.
\]
There is an algorithm that runs in $\poly(N,k,l , 1/c)$ time and outputs a function $\wt{\mcl{M}}$ that is $(k\poly(l/c), k\poly(l/c) )$-simple and
\[
\int_{-1 + c}^{1 - c} |\wt{\mcl{M}}(x) - f(x)|^2 dx \leq \eps^2 \poly(l) \,.
\]
\end{lemma}
The proof of Lemma \ref{lem:fourier-manycluster} will involve localizing the frequencies and then using Corollary \ref{coro:fourier-simplecase} to reconstruct after localizing. We will do this for $\poly(N,k, \log 1/\eps)$ different localizations (based on the set $T$ that we are given). We will then select at most $\wt{O}(k)$ of these localized reconstructions to add together and output. The intuition behind why we can find such a set of $\wt{O}(k)$ localized reconstructions and ignore the rest is that $\mcl{M}$ is $k$-Fourier sparse so localizations that are far away from the frequencies of $\mcl{M}$ can essentially be ignored.
The localization procedure will involve convolving $f$ by a Gaussian times an exponential (technically we will convolve by a function that approximates a Gaussian times an exponential). Note that this is equivalent to multiplying the Fourier transform by a Gaussian multiplier. This will ensure that frequencies too far away from a certain target frequency will only contribute negligibly and we only need to worry about reconstructing the frequencies that are close to the target frequency.
\subsection{Properties of Localization}\label{sec:local-fourier-properties}
In this section, we formalize the localization step and prove several inequalities that will be used in the proof of Lemma \ref{lem:fourier-manycluster}. The way we would like to localize the frequencies is by multiplying by a Gaussian multiplier in Fourier space since afterwards, we would be able to essentially neglect any frequencies that are far away from the center of the Gaussian multiplier. This is equivalent to convolving by a Gaussian times an exponential, i.e. a function of the form $G(x)e^{2 \pi i \theta x}$, in real space. For technical reasons, we will actually define two types of functions, which we call kernels, that approximate functions of the form $G(x)e^{2 \pi i \theta x}$. The reason that we will need to work with both is that the first type can be computed efficiently while the second is easier to use in the analysis of our algorithm.
\\\\
We begin with a few definitions.
\begin{definition}
For a function $f: \R \rightarrow \C$ and any $l > 0$, define $f^{\trunc(l)}$ to be the function that is equal to $f$ on $[-l,l]$ and $0$ otherwise.
\end{definition}
\begin{definition}\label{def:localization-type2}
For parameters $\mu, l , c$. we define the function
\[
\mcl{K}_{\mu, l ,c} = (1/c)\mcl{P}_{l}^{\trunc(l)}(x/c) \cdot e^{2 \pi i \mu x}
\]
where $\mcl{P}_l$ is as defined in Definition \ref{def:poly-approx-Gaussian}.
\end{definition}
\begin{remark}
We call functions of the above form truncated polynomial kernels. The fact that such functions are truncated polynomials will make it easy to explicitly compute convolutions.
\end{remark}
\begin{definition}\label{def:localization-type1}
For parameters $\mu, l,a$, we define the function $\mcl{T}_{\mu,l ,a}$ as follows. First define
\[
\mcl{S}_{ l , a} = M_{0,1}^{\trunc(l)} ( x /a )
\]
(recall Definition \ref{def:gaussian-multiplier}) and then define
\[
\mcl{T}_{\mu,l ,a}(x) = \wh{\mcl{S}_{l, a}}(x) e^{2 \pi i \mu x} \,.
\]
\end{definition}
\begin{remark}
We call functions of the above form truncated Gaussian kernels. Note that truncated Gaussian kernels are compactly supported in Fourier space, which will be a convenient property in the analysis of our algorithm later on.
\end{remark}
Note that both $\mcl{K}_{0, l ,c} $ and $\mcl{T}_{0,l , 2\pi / c}$ are meant to approximate the Gaussian $N(0, c^2)$. The fact that $\mcl{K}_{0, l ,c} $ approximates $N(0,c^2)$ is clear from the definition (and Lemma \ref{lem:Gaussian-taylorseries}). To see why $\mcl{T}_{0,l , 2\pi / c}$ approximates the Gaussian, note that if in the definition of $\mcl{T}$, we did not truncate before taking the Fourier transform, then we would get exactly $N(0, c^2)$.
\\\\
We will now prove several inequalities relating to how convolving with the kernels $\mcl{K}$ and $\mcl{T}$ affect a function. The first set of bounds are an immediate consequence of Lemma \ref{lem:Gaussian-taylorseries}.
\begin{claim} \label{claim:easyL2bound}
Let $l, \eps > 0$ be parameters such that $l \geq \lceil \log 1/\eps \rceil$. Let $0 < c < 1$ be some constant. Then
\begin{align*}
&\norm{\mcl{K}_{0,l,c}(x) - N(0, c^2)(x) }_1 \leq O(\eps l) \\
&\norm{\mcl{K}_{0,l,c}(x) - N(0, c^2)(x) }_2^2 \leq O( l \eps^2 / c) \,.
\end{align*}
\end{claim}
\begin{proof}
We know by Lemma \ref{lem:Gaussian-taylorseries} that
\[
|(1/c) \mcl{P}_{l}(x/c) - N(0,c^2)(x)| \leq \eps /c
\]
for all $x \in [-lc, lc]$. Thus, since $G$ decays rapidly outside the interval $[-lc, lc]$ we have
\[
\norm{\mcl{K}_{0,l,c}(x) - N(0,c^2)(x) }_1 \leq O(\eps l) \,.
\]
The second inequality follows by a similar argument.
\end{proof}
The next claim formalizes the intuition that $\mcl{K}_{\mu, l ,c} $ and $\mcl{T}_{\mu,l , 2\pi / c}$ must be close because they both approximate the same function (the function $N(0,c^2) e^{2\pi i \mu x}$).
\begin{claim}\label{claim:fourier-localization}
Let $l$ be some parameter and let $\eps > 0$ be such that $l \geq \lceil \log 1/\eps \rceil$. Let $g$ be a function such that $\norm{\wh{g}}_1 \leq 1$. Then for any $\mu, c $,
\[
\norm{\mcl{K}_{\mu,l,c} * g - \mcl{T}_{\mu,l,2\pi/c} * g}_{\infty} \leq O( \eps l ) \,,
\]
where $*$ denotes convolution.
\end{claim}
\begin{proof}
First, note that it suffices to prove the above for $\mu = 0$. Now let $G = N(0, c^2)$. By Claim \ref{claim:easyL2bound}, we have
\[
\norm{\mcl{K}_{0,l,c}(x) - G(x) }_1 \leq O(\eps l) \,.
\]
Since $\norm{\wh{g}}_1 \leq 1$, we know that $\norm{g}_{\infty} \leq 1$ and thus for all $x$,
\begin{equation}\label{eq:convolutionbound}
|\mcl{K}_{0,l,c} * g(x) - G * g(x)| \leq O(\eps l) \,.
\end{equation}
Finally, observe that the Fourier transform of $G * g$ is equal to $\wh{G} \wh{g} $. Note that
\[
\wh{G}(x) = M_{0,(2\pi/ c)^2}(x) = M_{0,1}(cx/(2 \pi )) \,.
\]
By construction,
\[
\wh{\mcl{T}_{0,l,2\pi/c} * g} = M_{0,1}^{\trunc(l)} ( c x /(2\pi) ) \wh{g}
\]
which is equal to $\wh{G} \wh{g} $ restricted to the interval $[-2 \pi l/c, 2 \pi l/c]$. Using the fact that $\wh{G} = M_{0,(2\pi/ c)^2}$ decays rapidly outside $[-2 \pi l/c, 2 \pi l/c]$, we have that
\[
\norm{\wh{\mcl{T}_{0,l,2\pi/c} * g} - \wh{G} \wh{g} }_1 \leq \eps \,.
\]
Thus, $|\mcl{T}_{0,l,2\pi/c} * g(x) - G * g(x)| \leq \eps$ for all $x$ and combining with (\ref{eq:convolutionbound}), we are done.
\end{proof}
\subsection{Decoupling}
In our full algorithm, we will reconstruct frequency-localized versions of a function independently for different frequencies $\theta$ that we localize around. We will then combine our localized reconstructions by adding them. In this section, we prove several inequalities that will allow us to analyze what happens to our estimation error when we add different localized reconstructions together. Recall that convolving by a truncated polynomial kernel $\mcl{K}_{\mu, l ,c}$ or truncated Gaussian kernel $\mcl{T}_{\mu, l ,a}$ is approximately equivalent to multiplying the Fourier transform by a Gaussian multiplier centered around $\mu$. Lemma \ref{lem:approx-constant} implies that adding up evenly spaced Gaussian multipliers approximates the constant function. Thus, we expect that convolving by an expression of the form $\sum_{\mu} \mcl{K}_{\mu, l ,c}$ or $\sum_{\mu} \mcl{T}_{\mu, l ,a}$ where the sum is over evenly spaced $\mu$ should roughly recover the original function. The first two claims here formalize this intuition.
In the first claim, we analyze what happens when we add several localizations obtained by convolving with various truncated polynomial kernels.
\begin{claim}\label{claim:convolution-bound}
Let $l$ be some parameter and let $0 < \eps < 0.1$ be such that $l \geq \lceil \log 1/\eps \rceil$. Let $g$ be a function. Let $c$ be some constant. Let $S$ be a set of integer multiples of $2 \pi/(cl)$. Then
\[
\int_{-1}^1 \left \lvert \frac{1}{l\sqrt{2\pi}}\sum_{\mu \in S }\mcl{K}_{\mu, l ,c} * g \right \rvert^2 \leq (1 + O(\eps |S|))^2 \int_{-(1 + lc)}^{1 + lc} |g|^2 \,,
\]
\end{claim}
\begin{proof}
Note that $\mcl{K}_{\mu,l,c}$ is supported on the interval $[-lc, lc]$ so we can restrict $g$ to be supported on $[-(1 + lc), 1 + lc]$ and $0$ outside the interval. Define the function
\[
V(x) = \frac{1}{l\sqrt{2\pi}}\sum_{\mu \in S } \mcl{K}_{\mu,l,c} \,.
\]
Then
\[
\wh{V}(x) = \frac{1}{l\sqrt{2\pi}}\sum_{\mu \in S} \wh{\mcl{K}_{0,l,c}}(x - \mu) \,.
\]
Next, let $G$ denote the Gaussian $G = N(0, c^2)$. Recall by Claim \ref{claim:easyL2bound}, $\norm{\mcl{K}_{0,l,c}(x) - G(x) }_1 \leq O(\eps l)$ so $\norm{\wh{\mcl{K}_{0,l,c}} - \wh{G}}_{\infty} \leq O(\eps l)$. Let
\[
U(x) = \frac{1}{l\sqrt{2\pi}}\sum_{\mu \in S} \wh{G}(x - \mu) \,.
\]
Note that since $\wh{G} = M_{0, (2\pi/c)^2}$, for any $x$,
\[
|U(x)| = \frac{1}{l\sqrt{2\pi}}\sum_{\mu \in S} M_{0, (2\pi/c)^2}(x - \mu) \leq \frac{1}{l\sqrt{2\pi}}\sum_{\mu \in (2\pi)/(cl) \Z} M_{0, (2\pi/c)^2}(x - \mu) \leq 1 + \eps
\]
where the last inequality follows from Lemma \ref{lem:approx-constant}. Also $\norm{\wh{K_{0,l,c}} - \wh{G}}_{\infty} \leq O(\eps l)$ so for all $x$,
\[
|\wh{V}(x)| \leq 1 + O(\eps |S|) \,.
\]
We conclude
\[
\int_{-1}^1 |V * g(x)|^2 dx \leq \norm{V * g}_2^2 = \norm{\wh{V}\wh{g}}_2^2 \leq ( 1 + O(\eps |S|))^2\norm{\wh{g}}_2^2 =( 1 + O(\eps |S|))^2\norm{g}_2^2
\]
To complete the proof recall that we restricted $g$ to be supported on $[-(1 + lc), 1 + lc]$ and we are done.
\end{proof}
The next claim is similar to the previous one except we analyze what happens when we add several localizations obtained by convolving with various truncated Gaussian kernels.
\begin{claim}\label{claim:sum-approximation}
Let $a.l, \eps$ be parameters such that $0 < \eps < 0.1$ and $ l \geq \lceil \log 1/\eps \rceil$. Let $g$ be a function whose Fourier support is contained in a set $S_0 \subset \R$ and such that $\norm{\wh{g}}_1 \leq 1$. Let $S$ be a set of integer multiples of $a/l$ that contains all multiples within a distance $l \cdot a$ of $S_0$. Then
\[
\norm{g - \frac{1}{l \sqrt{2\pi}} \sum_{\mu \in S} \mcl{T}_{\mu, l, a} * g }_{\infty} \leq 2\eps \,.
\]
\end{claim}
\begin{proof}
Consider the function
\[
A(x) = \sum_{\mu \in S} \frac{1}{l\sqrt{2\pi}} M_{\mu, a^2}(x) \,.
\]
For all $x \in S_0$, we claim that
\[
|A(x) - 1| \leq \eps \,.
\]
To see this, note that by Lemma \ref{lem:approx-constant},
\[
\left \lvert \sum_{\mu \in (a/l)\Z} \frac{1}{l\sqrt{2\pi}} M_{\mu,a^2}(x) - 1 \right \rvert \leq 0.1\eps
\]
for all $x$. By assumption, the set $S$ contains all integers that are within $la$ of the set $S_0$ so for any $x \in S_0$,
\[
\sum_{\mu \in (a/l)\Z \backslash S } \frac{1}{l\sqrt{2\pi}} M_{\mu, a^2}(x) \leq 0.1 \eps \,,
\]
and we conclude that we must have $|A(x) - 1| \leq \eps$. Next, we claim that if we define
\[
B(x) = \sum_{\mu \in S} \frac{1}{l\sqrt{2\pi}} M_{0, 1}^{\textsf{trunc}(l)}((x - \mu)/a) \,,
\]
then we have for all $x$,
\[
|B(x) - A(x) | \leq \eps \,.
\]
To see this, first note that $M_{0, 1}((x - \mu)/a) = M_{\mu, a^2}(x)$. Next, using Gaussian tail decay, we have for all $x$,
\[
|B(x) - A(x) | \leq \sum_{\substack{\mu \in (a/l)\Z \\ |x - \mu| \geq la}} \frac{1}{l\sqrt{2\pi}} M_{\mu, a^2}(x) \leq \eps \,.
\]
Thus, we have
\[
|B(x) - 1| \leq 2\eps
\]
for $x \in S_0$. Note that by definition,
\[
\wh{g}(x) - \frac{1}{l\sqrt{2\pi}} \sum_{\mu \in S} \wh{\mcl{T}_{\mu, l, a} * g }(x) = \wh{g}(x) - \wh{g}(x)\sum_{\mu \in S} \frac{1}{l\sqrt{2\pi}} M_{0, 1}^{\textsf{trunc}(l)}((x - \mu)/a) = (1 - B(x)) \wh{g}(x)
\]
so therefore
\[
\norm{\wh{g} - \frac{1}{l\sqrt{2\pi}} \sum_{\mu \in S} \wh{\mcl{T}_{\mu, l, a} * g }}_{1} \leq 2 \eps
\]
and we conclude
\[
\norm{ g - \frac{1}{l \sqrt{2\pi}} \sum_{\mu \in S}\mcl{T}_{\mu, l, a} * g }_{\infty} \leq 2\eps
\]
as desired.
\end{proof}
The last result in this section will allow us to decouple errors from summing over different localizations. Note that naively, if we add together $n$ estimates with $L^2$ errors $\eps_1, \dots , \eps_n$, then the resulting $L^2$ error of the sum could be as large as $\eps_1 + \dots + \eps_n$. If the estimates were ``independent" on the other hand, we would expect the $L^2$ error of the sum to only be $\sqrt{\eps_1^2 + \dots + \eps_n^2}$. We prove that when adding together functions that are frequency-localized at different locations, the error essentially matches the latter bound (up to logarithmic factors). This tighter bound will be necessary in the proof of Lemma \ref{lem:fourier-manycluster}.
Note that if we have functions $g_1, \dots , g_n$ whose Fourier supports are disjoint, then it is immediate that
\[
\norm{g_1 + \dots + g_n}_2^2 = \norm{g_1}_2^2 + \dots + \norm{g_n}_2^2 \,.
\]
However, in our setting, we need to restrict the functions to the interval $[-1,1]$ first, which causes the Fourier supports to no longer be disjoint. Through a few additional arguments we are able to prove an analogous statement for bounding $\int_{-1}^1|g_1 + \dots + g_n|^2$. We do pay some additional losses both in the inequality itself and the bounds of the integral i.e. we need to integrate the individual functions over a slightly larger interval.
\begin{claim}\label{claim:decoupling-pre}
Let $\alpha, l, \eps$ be parameters such that $\alpha > 1$ and $ l \geq \lceil \log \alpha n/\eps \rceil$. Let $I_1, \dots , I_n$ be intervals of length at least $\alpha$ and assume that for any $x \in \R$, at most $l$ of the intervals contain $x$. Let $g_1, \dots , g_n$ be functions such that for all $j \in [n]$, $\norm{\wh{g_j}}_1 \leq 1$ and $\wh{g_j}$ is supported on $I_j$. Then
\[
\int_{-1}^1 |g_1 + \dots + g_n|^2 \leq \poly(l) \left( \eps^5 + \int_{-(1 + \alpha^{-1})}^{(1 + \alpha^{-1})} |g_1|^2 + \dots + \int_{-(1 + \alpha^{-1})}^{(1 + \alpha^{-1})} |g_n|^2 \right)\,.
\]
\end{claim}
\begin{proof}
Consider the Gaussian multiplier $M = M_{\mu, \alpha^{-2} l^{-100}}$ for some $\mu \in [-1,1]$. Now first, we bound
\[
\int_{-\infty}^{\infty} M(x)^2 |g_1(x) + \dots + g_n(x)|^2 dx \,.
\]
Define the functions $h_j = \wh{M g_j}$ for all $j$. Then by Plancherel,
\begin{equation}\label{eq:bound1}
\int_{-\infty}^{\infty} M(x)^2 |g_1(x) + \dots + g_n(x)|^2 dx = \int_{-\infty}^{\infty}|h_1 + \dots + h_n|^2 dx
\end{equation}
On the other hand, note that $h_j = \wh{M} * \wh{g_j}$. Let $J_j$ denote the interval containing all points within distance at most $10 \alpha l^{60}$ of the interval $I_j$. Let $h_j'$ be the function $h_j$ restricted to $J_j$ (and equal to $0$ outside). Recall that the support of $\wh{g_j}$ is contained within $I_j$. Then we claim that
\[
\int_{-\infty}^{\infty} |h_j - h_j'|^2 \leq (\eps/(\alpha n))^{100} \,.
\]
This follows because $|\wh{M}| = N(0, (2 \pi \alpha l^{50})^2)$ and $\norm{\wh{g_j}}_1 \leq 1$ so for a point $x$ that is distance $d$ away from the interval $I_j$, we have
\[
|h_j(x)| \leq \max_{ y \in I_j} |\wh{M}(x - y)| \leq N(0, (2 \pi \alpha l^{50})^2)(d) \,.
\]
Also note that$\norm{\wh{g_j}}_1 \leq 1$, implies $\norm{g_j}_{\infty} \leq 1$ so
\[
\norm{h_j}_2^2 = \norm{Mg_j}_2^2 \leq \norm{M}_2^2 \leq 1 \,.
\]
Combining the previous two inequalities over all $j$, we have
\begin{align*}
\left \lvert \norm{h_1 + \dots + h_n}_2 - \norm{h_1' + \dots + h_n'}_2 \right \rvert \leq (\eps/(\alpha n))^{49} \\
\norm{h_1 + \dots + h_n}_2 + \norm{h_1' + \dots + h_n'}_2 \leq 3n
\end{align*}
which implies
\begin{equation}\label{eq:bound2}
\int_{-\infty}^{\infty}|h_1 + \dots + h_n|^2 dx \leq 0.1 (\eps/(\alpha n))^{10} + \int_{-\infty}^{\infty}|h_1' + \dots + h_n'|^2 dx \,.
\end{equation}
Now note that since not too many of the intervals $I_j$ may contain the same point $x \in \R$, not too many of the extended intervals $J_j$ can contain the same point $x \in \R$. In particular, at most $O(l^{70})$ of the extended intervals can contain the same point $x \in \R$. In other words, each point $x \in \R$ is in the support of at most $O(l^{70})$ of the $h_1', \dots , h_n'$. Thus, by Cauchy Schwarz,
\begin{equation}\label{eq:bound3}
\int_{-\infty}^{\infty}|h_1' + \dots + h_n'|^2 \leq O(l^{70}) \left( \int_{-\infty}^{\infty}|h_1'|^2 + \dots + \int_{-\infty}^{\infty}|h_n'|^2 \right) \,.
\end{equation}
Now we bound
\begin{equation}\label{eq:bound4}
\int_{-\infty}^{\infty}|h_j'|^2 \leq \int_{-\infty}^{\infty} |h_j|^2 = \int_{-\infty}^{\infty} M(x)^2 |g_j(x)|^2 dx \leq 0.1 (\eps/(\alpha n))^{10} + \int_{- ( 1 + \alpha^{-1})}^{1 + \alpha^{-1}} M(x)^2 |g_j(x)|^2 dx
\end{equation}
where the last step holds because $\norm{g_j}_{\infty} \leq 1$ and the multiplier $M(x)$ is always at most $1$ and decays rapidly outside the interval $[- ( 1 + \alpha^{-1}), 1 + \alpha^{-1}]$ since $\mu \in [-1,1]$. Putting everything together (\ref{eq:bound1}, \ref{eq:bound2}, \ref{eq:bound3}, \ref{eq:bound4}), we get
\[
\int_{-\infty}^{\infty} M(x)^2 |g_1(x) + \dots + g_n(x)|^2 dx \leq \poly(l) \left( (\eps/(\alpha n))^5 + \sum_{j = 1}^n \int_{- ( 1 + \alpha^{-1})}^{1 + \alpha^{-1}} M(x)^2 |g_j(x)|^2 dx \right)\,.
\]
Now summing the above over different multipliers $M = M_{\mu, \alpha^{-2} l^{-100}}$ i.e. with $\mu$ uniformly spaced on $[-1,1]$ with spacing $\alpha^{-1} l^{-50}$, we conclude
\begin{align*}
\int_{-1}^{1} |g_1(x) + \dots + g_n(x)|^2 dx \leq 10 \sum_{\mu}\int_{-1}^{1} M_{\mu, \alpha^{-2} l^{-100}}(x)^2 |g_1(x) + \dots + g_n(x)|^2 dx \\ \leq \poly(l) \eps^5 + \poly(l) \left(\sum_{\mu}\int_{-(1 + \alpha^{-1})}^{(1 + \alpha^{-1})} M_{\mu, \alpha^{-2} l^{-100}}(x)^2(|g_1|^2 + \dots + |g_n|^2) \right) \\ \leq \poly(l) \eps^5 + \poly(l) \left(\int_{-(1 + \alpha^{-1})}^{(1 + \alpha^{-1})} |g_1|^2 + \dots + \int_{-(1 + \alpha^{-1})}^{(1 + \alpha^{-1})}|g_n|^2) \right)\,.
\end{align*}
\end{proof}
\subsubsection{Completing the Proof of Lemma \ref{lem:fourier-manycluster}}
In this section, we will complete the proof of Lemma \ref{lem:fourier-manycluster}. First, we need to introduce some notation. We will carry over all of the notation from the statement of Lemma \ref{lem:fourier-manycluster}. We also use the following conventions:
\begin{itemize}
\item Let $S_0 = \{\theta_1, \dots , \theta_k \}$ be the frequencies in the Fourier support of $\mcl{M}$
\item Let $\gamma > 0$ be parameter to be chosen later and let $l' = \lceil \log 1/ \gamma \rceil$ (we will ensure $\gamma$ is sufficiently small i.e. $\gamma < (\eps c/(k N))^{K}$ for some sufficiently large absolute constant $K$ )
\item Let the function $r(x)$ be defined as $r(x) = f(x) - \mcl{M}(x)$ on the interval $[-1,1]$ and $r(x) = 0$ outside the interval.
\end{itemize}
Recall that the way we will reconstruct the function is by attempting to localize around each of the points in the given set $T$ and reconstructing the localized function using Corollary \ref{coro:fourier-simplecase}. We then find $k \poly(l'/c)$ of these localized reconstructions that we can combine to approximate the entire function.
In the first claim, we bound the error of our reconstruction using Corollary \ref{coro:fourier-simplecase} for a given localization. Recall the two types of kernels, the truncated polynomial kernel and the truncated Gaussian kernel, defined in Section \ref{sec:local-fourier-properties}. Consider the kernels $\mcl{K}_{\mu,l',c/l'}$ and $\mcl{T}_{\mu, l',2\pi l'/c } $ (which, recall, are approximately the same). In the next lemma, we will bound the distance between $\mcl{K}_{\mu,l',c/l'} * f$ and $\mcl{T}_{\mu, l',2\pi l'/c } * \mcl{M}$ in terms of $r(x)$. The reason we care about these two functions is that the first is something that we can compute since we are given explicit access to $f$. On the other hand, the second is Fourier sparse and has bounded Fourier support so it can be plugged into Corollary \ref{coro:fourier-simplecase} (as the unknown function $\mcl{M}$).
\begin{claim}\label{claim:contraction-bound}
For any real number $\mu$,
\[
\int_{-1 + c}^{1 - c} | \mcl{K}_{\mu,l',c/l'} * f - \mcl{T}_{\mu, l',2\pi l'/c } * \mcl{M}|^2 \leq \poly(l'/c)\gamma^2 + 4\int_{-\infty}^{\infty} |M_{0, (2\pi l'/c)^2}(x - \mu) \wh{r}(x)|^2 dx \,.
\]
\end{claim}
\begin{proof}
Note that since $\mcl{K}_{\mu,l',c/ l'}$ is supported on $[-c, c]$,
\[
\int_{-1 + c}^{1 - c} \left \lvert \left(\mcl{K}_{\mu,l',c/l'} * f \right)(x) - \left(\mcl{K}_{\mu,l',c/l'} * \mcl{M} \right)(x) \right \rvert^2 dx \leq \int_{-1}^1 \lvert \left(\mcl{K}_{\mu,l',c/l'} * r \right)(x) \rvert^2 dx
\]
Now the Fourier transform of $\mcl{K}_{\mu,l',c/l'} * r$ is $\wh{\mcl{K}_{0,l',c/l'}}(x - \mu) \wh{r}(x )$ so
\[
\int_{-1}^1 \lvert \left(\mcl{K}_{\mu,l',c/l'} * r \right)(x) \rvert^2 dx \leq \int_{-\infty}^{\infty} |\wh{\mcl{K}_{0,l',c/l'}}(x - \mu) \wh{r}(x )|^2 dx
\]
We deduce that
\begin{align*}
\int_{-1 + c}^{1 - c} |\mcl{K}_{\mu,l',c/l'} * f - \mcl{T}_{\mu, l',2\pi l'/c } * \mcl{M}|^2 \leq 2 \int_{-1 + c}^{1 - c} \left\lvert \mcl{K}_{\mu,l',c/l'} * \mcl{M} - \mcl{T}_{\mu, l',2\pi l'/c } * \mcl{M}\right\rvert^2 \\ + 2\int_{-\infty}^{\infty} |\wh{\mcl{K}_{0,l',c/l'}}(x - \mu) \wh{r}(x )|^2 dx \\ \leq \poly(l'/c) \gamma^2 + 2\int_{-\infty}^{\infty} |\wh{\mcl{K}_{0,l',c/l'}}(x - \mu) \wh{r}(x )|^2 dx
\end{align*}
where the last inequality follows from Claim \ref{claim:fourier-localization}.
\\\\
Note since $r$ is supported on $[-1,1]$ and $\norm{r}_2^2 \leq \eps^2 \leq 0.1$, we must have $\norm{r}_1 \leq 1$ which then implies $\norm{\wh{r}}_{\infty} \leq 1$. Together with Claim \ref{claim:easyL2bound}, if we let $G = N(0, (c/l')^2)$ then we have
\begin{align*}
\int_{-\infty}^{\infty} |\wh{\mcl{K}_{0,l',c/l'}}(x - \mu) \wh{r}(x)|^2 dx \leq 2\int_{-\infty}^{\infty} |\wh{G}(x - \mu) \wh{r}(x )|^2 dx + 2 \norm{\wh{r}}_{\infty}^2 \norm{\mcl{K}_{0,l',c/l'} - G}_2^2 \\ \leq \poly(l'/c)\gamma^2 + 2\int_{-\infty}^{\infty} |M_{0, (2\pi l'/c)^2}(x - \mu) \wh{r}(x )|^2 dx \,.
\end{align*}
and combining with the previous inequality, we get the desired result.
\end{proof}
We are now ready to complete the proof of Lemma \ref{lem:fourier-manycluster}.
\begin{proof}[Proof of Lemma \ref{lem:fourier-manycluster}]
First, let $T'$ be the set of integer multiples of $2\pi/c$ that are within distance $(10l')^2/c$ of the set $T$. For all $\mu \in T'$, do the following. We compute the function $f^{(\mu)} = \mcl{K}_{\mu, l', c/l'} * f$. Next we apply Corollary \ref{coro:fourier-simplecase} (with appropriate rescaling) to compute a function $h^{(\mu)}$ in $\poly(l'/c)$ time that has Fourier support in $[ \mu - 2\pi l'^2/c, \mu + 2\pi l'^2/c]$, is $(\poly(l'/c), \poly(l'/c))$-simple and such that
\begin{equation}\label{eq:piecewise-estimates}
\int_{-1 + c}^{1 - c} |f^{(\mu)} - h^{(\mu)}|^2 \leq 20\left( \gamma^2 + \int_{-1 + c}^{1 - c} |f^{(\mu)} - \mcl{T}_{\mu, l' , 2\pi l'/c} * \mcl{M}|^2 \right) \,.
\end{equation}
To see why we can do this, note that
\[
\wh{ \mcl{T}_{\mu, l' , 2\pi l'/c} * \mcl{M}} = M_{0,1}^{\trunc(l')} ( c(x - \mu) /(2 \pi l')) \wh{\mcl{M}}
\]
is supported on $[ \mu - 2\pi l'^2/c, \mu + 2\pi l'^2/c]$. Also it is clear that $\norm{\wh{ \mcl{T}_{\mu, l' , 2\pi l'/c} * \mcl{M}}}_1 \leq \norm{\wh{\mcl{M}}}_1 \leq 1$.
\\\\
Now we choose a set $U \subset T'$ with $|U| \leq k(10l')^2$ such that the following quantity is minimized:
\[
E_{U} = \sum_{\mu \in U} \int_{-1 + c}^{1 - c} |f^{(\mu)} - h^{(\mu)}|^2 + \sum_{\mu \in T' \backslash U} \int_{-1 + c}^{1 - c} |f^{(\mu)} |^2 \,.
\]
Note that this can be done using a simple greedy procedure. First, we obtain a bound on the value $E_U$ that we compute. Let $U_0$ be the set of all integer multiples of $2\pi/c$ that are within distance $10l'^2/c$ of $S_0$ (recall that $S_0$ is the Fourier support of $\mcl{M}$ which consist of $k$ points). By assumption, we know that $U_0 \subset T'$ and it is clear that $|U_0| \leq k(10l')^2$. Note that by definition, for any $\mu \notin U_0$, the function $\mcl{T}_{\mu, l' , 2\pi l'/c} * \mcl{M}$ is identically $0$. Now using (\ref{eq:piecewise-estimates}), then Claim \ref{claim:contraction-bound},
\begin{align*}
E_{U_0} &\leq 20\sum_{\mu \in U_0}\left( \gamma^2 + \int_{-1 + c}^{1 - c} |f^{(\mu)} - \mcl{T}_{\mu, l' , 2\pi l'/c} * \mcl{M}|^2 \right) + \sum_{\mu \in T'\backslash U_0} \int_{-1 + c}^{1 - c} |f^{(\mu)} |^2 \\
& \leq 80 \sum_{\mu \in T'} \left( \poly(l'/c)\gamma^2 + \int_{-\infty}^{\infty} |M_{0, (2\pi l'/c)^2}(x - \mu) \wh{r}(x)|^2 dx\right)
\\ & \leq |T'| \poly(l'/c) \gamma^2 + 80\int_{-\infty}^{\infty} | \wh{r}(x )|^2 \sum_{\mu \in (2\pi/c)\Z} M_{0, (2\pi l'/c)^2}(x - \mu)^2dx
\\ & \leq \gamma + \poly(l') \norm{r}_2^2 \,.
\end{align*}
Note that we used the fact that $\gamma$ is sufficiently small and the tail decay properties of the Gaussian multipliers in the last step. Thus, we can ensure that the error that we compute satisfies $E_U \leq \gamma + \poly(l') \norm{r}_2^2$. Now we output the function
\[
\wt{\mcl{M}} = \sum_{\mu \in U} \frac{1}{l'\sqrt{2 \pi}}h^{(\mu)} \,.
\]
It remains to bound the error between $\wt{\mcl{M}}$ and $f$. First we apply Claim \ref{claim:decoupling-pre} to decouple over all $\mu \in T'$. Note that $h^{(\mu)}$ and $\mcl{T}_{\mu, l' , 2\pi l'/c} * \mcl{M}$ both have Fourier support contained in the interval $[ \mu - 2\pi l'^2/c, \mu +2\pi l'^2/c]$. For distinct $\mu$ that are integer multiples of $2\pi/c$, there are at most $O(l'^2)$ intervals that contain any point. Also, note that for all $\mu$, $\norm{h^{(\mu)}}_1 \leq \poly(l'/c)$ and $\norm{\wh{ \mcl{T}_{\mu, l' , 2\pi l'/c} * \mcl{M}}}_1 \leq 1$. Thus, by Claim \ref{claim:decoupling-pre} (with appropriate rescaling of the functions and the interval),
\begin{align*}
&\int_{-1 + 2c}^{1 - 2c} \left \lvert \wt{\mcl{M}} - \sum_{\mu \in T'} \frac{1}{l'\sqrt{2 \pi}} \mcl{T}_{\mu, l' , 2\pi l'/c} * \mcl{M} \right \rvert \\ &\leq \poly(l'/c) \gamma + \poly(l') \left( \sum_{\mu \in U} \int_{-1 + c}^{1 - c} |h^{(\mu)} - \mcl{T}_{\mu, l' , 2\pi l'/c} * \mcl{M} |^2 + \sum_{\mu \in T'\backslash U} \int_{-1 + c}^{1 - c} |\mcl{T}_{\mu, l' , 2\pi l'/c} * \mcl{M} |^2 \right)
\\ &\leq \poly(l'/c)\gamma + \poly(l') \left( \sum_{\mu \in U} \int_{-1 + c}^{1 - c} |h^{(\mu)} - f^{(\mu)}|^2 + \sum_{\mu \in T'\backslash U} \int_{-1 + c}^{1 - c} |f^{(\mu)}|^2 \right)
\\ &\quad + \poly(l') \sum_{\mu \in T'}\int_{-1 + c}^{1 - c} |f^{(\mu)} - \mcl{T}_{\mu, l' , 2\pi l'/c} * \mcl{M} |^2
\\ &\leq \poly(l'/c) \gamma + \poly(l')\left(E_U + \sum_{\mu \in T'} \left( \poly(l'/c)\gamma^2 + \int_{-\infty}^{\infty} |M_{0, (2\pi l'/c)^2}(x - \mu) \wh{r}(x )|^2 dx\right)\right)
\\ &\leq \poly(l'/c) \gamma + \poly(l')( E_U + \norm{r}_2^2)
\\& \leq \poly(l') \eps^2
\end{align*}
where we used that $\gamma = (\eps c/(kN))^{O(1)}$ is sufficiently small and Claim \ref{claim:contraction-bound} and the last two inequalities follow from the same argument as in the bound for $E_{U_0}$. Next, by Claim \ref{claim:sum-approximation} (and the definition of $T'$), we have that
\[
\norm{\mcl{M} - \sum_{\mu \in T'} \frac{1}{l'\sqrt{2 \pi}} \mcl{T}_{\mu, l' , 2\pi l'/c} * \mcl{M} }_{\infty} \leq O(\eps)
\]
so overall, we conclude
\[
\int_{-1 + 2c}^{1 - 2c} |\wt{\mcl{M}} - \mcl{M}|^2 \leq \poly(l') \eps^2
\]
from which we immediately deduce
\[
\int_{-1 + 2c}^{1 - 2c} |\wt{\mcl{M}} - f|^2 \leq \poly(l') \eps^2 \,.
\]
It is also clear that $\wt{\mcl{M}}$ is $(k \poly(l'/c), k \poly(l'/c))$-simple (since it is a sum of at most $k(10l')^2$ functions that are $( \poly(l'/c), \poly(l'/c))$-simple). Now we are done because $l' = O(l)$.
\end{proof}
Similar to obtaining Corollary \ref{coro:fourier-simplecase} from Lemma \ref{lem:Fourier-simplecase}, we can extend Lemma \ref{lem:fourier-manycluster} to work even when we do not know the target accuracy but only a lower bound on it.
\begin{corollary}\label{coro:fourier-manycluster}
Assume we are given $N, k, \eps , c$ with $0 < \eps < 0.1$. Let $l = \lceil \log kN / (\eps c) \rceil$ be some parameter. Let $\mcl{M}$ be a function that is $(k,1)$-simple. Also, assume that we are given a set $T \subset \R$ of size $N$ such that all of the support of $\wh{\mcl{M}}$ is within distance $1$ of $N$. Further, assume we are given access to a function $f$. There is an algorithm that runs in $\poly(N,k,l , 1/c)$ time and outputs a function $\wt{\mcl{M}}$ that is $(k\poly(l/c), k\poly(l/c) )$-simple and such that
\[
\int_{-1 + c}^{1 - c} |\wt{\mcl{M}}(x) - f(x)|^2 dx \leq \eps^2 + \poly(l) \int_{-1}^1 |f(x) - \mcl{M}(x)|^2 dx\,.
\]
\end{corollary}
\begin{proof}
This will be the exact same argument as the proof of Corollary \ref{coro:fourier-simplecase}. For a target accuracy $\eps' > \eps$ we run the algorithm in Lemma \ref{lem:fourier-manycluster} to get a function $\wt{\mcl{M}}_{\eps'}(x)$. We then check whether
\[
\int_{-1 + c}^{1 - c} |\wt{\mcl{M}}_{\eps'}(x) - f(x)|^2 dx \leq \poly(l) \eps'^2 \,.
\]
If the check passes, we take $\eps' \leftarrow 0.99 \eps'$ and repeat the above until we find the smallest $\eps'$ (up to a constant factor) for which the check passes. The guarantee of Lemma \ref{lem:fourier-manycluster} implies that for this $\eps'$, we can just output $\wt{\mcl{M}}_{\eps'}(x)$ and it is guaranteed to satisfy the desired inequality.
\end{proof}
\subsection{Proof of Main Theorem}
Using Theorem \ref{thm:sparse-fourier-old} and Corollary \ref{coro:fourier-manycluster}, we can prove our main theorem, Theorem \ref{thm:fourier-main}. The main thing that we need to prove is that the frequencies in the function $f'$ computed by Theorem \ref{thm:sparse-fourier-old} cover (within distance $\poly(k, \log 1/\eps)$) all of the frequencies in $\mcl{M}$. This will then let us use the frequencies in $f'$ to construct a set $T$ of size $\poly(k, \log 1/\eps)$ that covers all frequencies in $\mcl{M}$ to within distance $1$ that we can then plug into Corollary \ref{coro:fourier-manycluster}. We will need the following technical lemma from \cite{chen2016fourier}.
\begin{lemma}\label{lem:extension-bound}[Lemma 5.1 in \cite{chen2016fourier}]
For any $k$-Fourier sparse signal $g: \R \rightarrow \C$,
\[
\max_{x \in [-1,1]} |g(x)|^2 \leq O(k^4 \log^3 k) \int_{-1}^1 |g(x)|^2 dx \,.
\]
\end{lemma}
The above lemma roughly says that the mass of a $k$-Fourier sparse function cannot be too concentrated. We now finish the proof of Theorem \ref{thm:fourier-main}.
\begin{proof}[Proof of Theorem \ref{thm:fourier-main}]
We first apply Theorem \ref{thm:sparse-fourier-old} to compute a function $f'$. such that
\begin{enumerate}
\item $f'$ is $(\poly(k, \log 1/\eps), \exp( \poly(k, \log 1/\eps)) )$- simple
\item
\[
\int_{-1}^1 |f' - f|^2 \leq O\left(\eps^2 + \int_{-1}^1 |f - \mcl{M}|^2 \right) \,.
\]
\end{enumerate}
Let $L = (k \log 1/\eps)^{K}$ for some sufficiently large absolute constant $K$. Let $\gamma = e^{-L}$. Now apply Claim \ref{claim:sum-approximation} on the function $f'$ with parameters $a \leftarrow L, l \leftarrow L, \eps \leftarrow \gamma $. Let $S \subset L \Z$ be the set of all integer multiples of $L$ that are within distance $L^3$ of the Fourier support of $f'$. We have
\begin{equation}\label{eq:infbound1}
\norm{f' - \frac{1}{L\sqrt{2\pi}} \sum_{\mu \in S} \mcl{T}_{\mu, L , L^2} * f' }_{\infty} \leq 2 \norm{\wh{f'}}_1 \gamma
\end{equation}
Next, we apply Claim \ref{claim:fourier-localization} on the function $ (\mcl{M} - f')$. We deduce that for any $\mu$,
\begin{equation}\label{eq:infbound2}
\norm{\mcl{K}_{\mu, L, 2\pi/L^2} * (\mcl{M} - f')- \mcl{T}_{\mu, L , L^2} *(\mcl{M} - f')}_{\infty} \leq O\left(\gamma L \left(\norm{\wh{f'}}_1 +\norm{ \wh{\mcl{M}}}_1 \right) \right) \,.
\end{equation}
Finally, by Claim \ref{claim:convolution-bound} applied to the function $ (\mcl{M} - f')$ (with parameters $\eps \leftarrow \gamma ,l \leftarrow L, c \leftarrow (2\pi)/L^2$), we have
\begin{align*}
\int_{-1}^1 \left \lvert \frac{1}{L\sqrt{2\pi}} \sum_{\mu \in S}\mcl{K}_{\mu, L, 2\pi/L^2} * (\mcl{M} - f')\right \rvert^2 \leq (1 + O\left(\gamma |S|\right) )^2\int_{-1 - 2\pi/L}^{1 + 2\pi /L} | \mcl{M} - f'|^2 \leq 2 \int_{-1}^1 | \mcl{M} - f'|^2 \,.
\end{align*}
Note that in the last step we use Lemma \ref{lem:extension-bound} and the fact that $ \mcl{M} - f'$ is $\poly(k, \log 1/\eps)$-Fourier sparse so choosing $L = (k \log 1/\eps)^{O(1)}$ sufficiently large ensures that
\[
\int_{-1 - 2\pi/L}^{1 +2\pi /L} | \mcl{M} - f'|^2 \leq 1.1 \int_{-1}^1 | \mcl{M} - f'|^2 \,.
\]
Define the functions
\begin{align*}
&A(x) = f' - \frac{1}{L\sqrt{2\pi}} \sum_{\mu \in S}\mcl{T}_{\mu,L,L^2} * \mcl{M} \\
&B(x) = \frac{1}{L\sqrt{2\pi}} \sum_{\mu \in S}\mcl{K}_{\mu, L, 2\pi/L^2} * ( \mcl{M} - f')
\end{align*}
Note $\norm{A}_{\infty}, \norm{B}_{\infty} \leq |S|\left(\norm{\wh{f'}}_1 + \norm{\wh{\mcl{M}}}_1 \right) $. Combining (\ref{eq:infbound1}, \ref{eq:infbound2}) we have
\[
\int_{-1}^1 |A(x)|^2 - \int_{-1}^1 |B(x)|^2 \leq \gamma \poly\left(L, |S|, \norm{\wh{f'}}_1 + \norm{\wh{\mcl{M}}}_1 \right)
\]
However, we proved that $\int_{-1}^1 |B(x)|^2 \leq 2 \int_{-1}^1 |\mcl{M} - f'|^2 $ so choosing $L = (k \log 1/\eps)^{O(1)}$ sufficiently large and since $\gamma = e^{-L}$, we conclude
\begin{equation}\label{eq:true-errorbound}
\int_{-1}^1 |A(x)|^2 \leq \eps^2 + 2\int_{-1}^1 |\mcl{M} - f'|^2 \leq O\left(\eps^2 + \int_{-1}^1 |\mcl{M} - f|^2 \right)
\end{equation}
where we are using the guarantee from Theorem \ref{thm:sparse-fourier-old}. Now note that the function
\[
\mcl{M}' = \frac{1}{L\sqrt{2\pi}} \sum_{\mu \in S}\mcl{T}_{\mu,L,L^2} *\mcl{M}
\]
is $k$-Fourier sparse and has $\norm{\wh{\mcl{M}'}}_1 \leq 2\norm{\wh{\mcl{M}}}_1$ by Lemma \ref{lem:approx-constant}. Furthermore, all of its Fourier support is within distance $\poly(L)$ of the Fourier support of $f'$ (by the construction of the set $S$). Thus, we can apply Corollary \ref{coro:fourier-manycluster} on $f'$ (where we treat the unknown function as $\mcl{M}'$) with $N = \poly(L) = \poly(k , \log 1/\eps)$ and recover a function $\wt{\mcl{M}}$ such that
\begin{align*}
\int_{-1 + c}^{1 - c} |\wt{\mcl{M}} - f'|^2 \leq \eps^2 + \poly(\log k/(\eps c)) \int_{-1}^1 |f' - \mcl{M}'| = \eps^2 + \poly(\log k/(\eps c)) \int_{-1}^1 |A(x)|^2 \\ = \poly(\log k/(\eps c)) \left(\eps^2 + \int_{-1}^1 |\mcl{M} - f|^2 \right)
\end{align*}
where the last step is from (\ref{eq:true-errorbound}). Since $\int_{-1}^1 |f' - f|^2 \leq O\left(\eps^2 + \int_{-1}^1 |f - \mcl{M}|^2 \right)$, the above implies
\[
\int_{-1 + c}^{1 - c} |\wt{\mcl{M}} - f|^2 \leq \poly(\log k/(\eps c)) \left(\eps^2 + \int_{-1}^1 |\mcl{M} - f|^2 \right)
\]
and we are done.
\end{proof}
\subsection{Implementation of Computations}\label{sec:implementation}
In the proof of Theorem \ref{thm:fourier-main}, we use the result from \cite{chen2016fourier} to obtain an approximation $f'$ that is written as a sum of $\poly(k, \log 1/\eps)$ exponentials and has coefficients bounded by $\exp(\poly (k , \log 1/\eps))$. We then perform explicit computations using this function in our algorithm to eventually compute a sparser approximation with smaller coefficients. Here we briefly explain why these explicit computations can all be implemented efficiently. Note that all of the functions that we perform computations on can be written as sums of polynomials multiplied by exponentials i.e.
\begin{equation}\label{eq:func-form}
P_1(x)e^{2\pi i \theta_1 x} + \dots + P_n(x)e^{2\pi i \theta_n x}
\end{equation}
where there are at most $\poly(k, \log 1/\eps)$ terms in the sum, all of the polynomials have degree at most $\poly(k, \log 1/\eps)$ and all of the coefficients are bounded by $\exp(\poly (k , \log 1/\eps))$. To see this, note that convolving by a polynomial $P(x)$ truncated to an interval (recall the truncated polynomial kernel in Definition \ref{def:localization-type2} ) preserves a function of the form in (\ref{eq:func-form}) (only increasing the degrees of the polynomials by $\deg(P)$). All other computations that we need such as computing the exact value, adding and multiplying and integrating over some interval can clearly be done explicitly in $\poly (k , \log 1/\eps)$ time and to $\exp(\poly (k , \log 1/\eps))^{-1}$ accuracy for functions of the form specified in (\ref{eq:func-form}).
\section{Basic Tools}\label{appendix:basic}
In this section, we have a few basic tools that are used repeatedly throughout the paper.
\subsection{Chebyshev Polynomials}
Here we will introduce several basic results about the Chebyshev polynomials, which have algorithmic applications in a wide variety of settings \cite{rivlin2020chebyshev, guruswami2016robust}.
\begin{definition}[Chebyshev Polynomials]
The Chebyshev Polynomials are a family of polynomials defined as follows: $T_0(x) = 1, T_1(x) = x$ and for $n \geq 2$,
\[
T_n(x) = 2xT_{n-1}(x) - T_{n-2}(x) \,.
\]
\end{definition}
\begin{fact}\label{fact:chebyshev-basic}
The Chebyshev polynomials satisfy the following property:
\[
T_n(\cos \theta) = \cos n \theta \,.
\]
\end{fact}
As an immediate consequence of the above, we have a few additional properties.
\begin{fact}\label{fact:chebyshev-basic2}
The Chebyshev polynomials satisfy the following properties:
\begin{enumerate}
\item $T_n(x)$ has degree $n$ and leading coefficient $2^{n-1}$
\item For $x \in [-1,1]$, $T_n(x) \in [-1,1 ]$
\item $T_n(x)$ has $n$ zeros all in the interval $[-1,1]$
\item There are $n+1$ values of $x$ for which $T_n(x) = \pm 1$, all in the interval $[-1,1]$
\end{enumerate}
\end{fact}
\noindent In light of the previous properties, we make the following definition.
\begin{definition}
For an integer $n$, we define the Chebyshev points of degree $n$, say $t_0, \dots , t_n$, as the points in the interval $[-1,1]$ where the Chebyshev polynomial satisfies $T_n(t_j) = \pm 1$. Note that the Chebyshev points are exactly
\[
\{ \cos 0, \cos \frac{\pi}{n}, \dots \cos \frac{(n-1)\pi}{n}, \cos \pi \} \,.
\]
\end{definition}
Next, we have a result saying that if we have a bound on the value of a degree-$n$ polynomial at all of the degree $n$ Chebyshev points, then we can bound the value over the entire interval $[-1,1]$. Similar results are used in \cite{rivlin2020chebyshev, guruswami2016robust}, but there does not appear to be a directly usable reference.
\begin{claim}\label{claim:chebyshev-bound}
Let $P(x)$ be a polynomial of degree at most $n$ with real coefficients. Let $t_0, \dots , t_n$ be the Chebyshev points of degree $n$. Assume that $|P(t_j)| \leq 1 $ for $j = 0,1, \dots , n$. Then $|P(x) | \leq 2n$ for all $x \in [-1,1]$.
\end{claim}
\begin{proof}
By Lagrange interpolation, we may write
\[
P(x) = \frac{P(t_0) (x - t_1) \cdots (x - t_n)}{(t_0 - t_1) \cdots (t_0 - t_n)} + \dots + \frac{P(t_n) (x - t_0) \cdots (x - t_{n-1})}{(t_n - t_0) \cdots (t_n - t_{n-1})} \,.
\]
Thus, it suffices to upper bound the quantity
\[
F(x) = \left\lvert \frac{(x - t_1) \cdots (x - t_n)}{(t_0 - t_1) \cdots (t_0 - t_n)} \right\rvert + \dots + \left\lvert \frac{ (x - t_0) \cdots (x - t_{n-1})}{(t_n - t_0) \cdots (t_n - t_{n-1})} \right\rvert
\]
on the interval $[-1,1]$. Note that by Lagrange interpolation on $T_n(x)$, we have
\[
T_n(x) = \frac{T_n(t_0) (x - t_1) \cdots (x - t_n)}{(t_0 - t_1) \cdots (t_0 - t_n)} + \dots + \frac{T_n(t_n) (x - t_0) \cdots (x - t_{n-1})}{(t_n - t_0) \cdots (t_n - t_{n-1})} \,.
\]
Also note that $T_n(t_j) = (-1)^{n - j} $ which has the same sign as $(t_j - t_0) \cdots (t_j - t_{j-1}) (t_j - t_{j+1}) \cdots (t_j - t_n)$. Thus,
\[
\left\lvert \frac{1}{(t_0 - t_1) \cdots (t_0 - t_n)} \right\rvert + \dots + \left\lvert \frac{ 1}{(t_n - t_0) \cdots (t_n - t_{n-1})} \right\rvert = 2^{n-1} \,,
\]
since the leading coefficient of $T_n(x)$ is $2^{n-1}$. Now we will upper bound
\[
M = \max\left( |(x - t_1) \cdots (x - t_n)|, \dots , |(x - t_0) \cdots (x - t_{n-1})| \right)
\]
and once we do this, we will have a bound on $F(x)$ since $F(x) \leq 2^{n-1}M$. Define the polynomial
\[
Q(x) = (x-t_0)(x-t_1) \cdots (x - t_n) = \frac{\sqrt{x^2 - 1}}{2^n} \left( (x+ \sqrt{x^2 - 1})^n - (x - \sqrt{x^2 - 1})^n \right) \,.
\]
To see why the last equality is true, note that the RHS has roots at $t_0, \dots , t_n$ and is a monic polynomial of degree $n+1$ so it must be equal to $(x - t_0) \cdots (x - t_n)$. Now,
\[
M = \max\left( \left \lvert \frac{Q(x)}{x - t_0} \right \rvert, \dots , \left \lvert \frac{Q(x)}{x - t_n} \right \rvert \right) \leq \max |Q'(x)|
\]
where the last step holds by the mean value theorem (because $Q(t_j) = 0$ for all $j$). Now note that
\[
Q(\cos \theta) = -\frac{\sin \theta \sin(n\theta)}{2^{n-1}}
\]
so
\[
Q'(\cos \theta) = \frac{n\cos n \theta}{2^{n-1}} + \frac{\cos \theta \sin(n \theta)}{\sin(\theta) 2^{n-1}}
\]
and from the above it is clear that
\[
|Q'(\cos \theta)| \leq \frac{n}{2^{n-1}} + \frac{n}{2^{n-1}} = \frac{n}{2^{n-2}} \,.
\]
Now we are done because
\[
\max_{x \in [-1,1]} |P(x)| \leq F(x) \leq 2^{n-1}M \leq 2n \,.
\]
\end{proof}
It turns out that we can restate the above result in terms of convex hulls of points on the moment curve. This reformulation is the version that is useful in our algorithms.
\begin{definition}
For a real number $x$, we define the moment vector $\mcl{V}_n(x) = (1,x, \dots , x^{n})$.
\end{definition}
\begin{corollary}\label{coro:convexhull}
Let $t_0, t_1, \dots , t_n$ be the Chebyshev points of degree $n$. Then for any $x \in [-1,1]$, the point $\mcl{V}_n(x)$ is contained in the convex hull of the points
\[
\{ \pm 2n\mcl{V}_n(t_0), \dots , \pm 2n\mcl{V}_n(t_n) \} \,.
\]
\end{corollary}
\begin{proof}
Assume for the sake of contradiction that the above is not true. Then there must be a separating hyperplane. Assume that this hyperplane is given by $a \cdot x = b$ where $a$ is a vector and $b$ is a real number. Now WLOG $b \geq 0$ and we must have
\begin{align*}
& a \cdot \mcl{V}_n(x) \geq b \\
& |a \cdot \mcl{V}_n(t_j) | \leq \frac{b}{2n} \quad \forall j
\end{align*}
However, applying Claim \ref{claim:chebyshev-bound} with $P(x) = \frac{2n (a \cdot \mcl{V}_n(x)))}{b}$ gives a contradiction. Thus, no separating hyperplane can exist and we are done.
\end{proof}
\subsection{Approximating a Gaussian with a Polynomial}
We will also need to approximate Gaussians with polynomials. This is a somewhat standard result which we state below.
\begin{lemma}\label{lem:Gaussian-taylorseries}
Let $G = N(0, 1)$ be the standard Gaussian. Let $l$ be some parameter. Then we can compute a polynomial $P(x)$ of degree $(10l)^2$ such that for all $x \in [-2l, 2l]$,
\[
|G(x) - P(x)| \leq e^{-l} \,.
\]
\end{lemma}
\begin{proof}
Write
\[
G(x) = \frac{1}{ \sqrt{2\pi}} e^{-x^2/2} \,.
\]
and now we can write the Taylor expansion
\[
e^{-\frac{x^2}{2} } = \sum_{m = 0}^\infty \frac{\left(-\frac{x^2}{2}\right)^m }{m!} = \sum_{m =0}^{\infty}\frac{(-1)^m x^{2m}}{2^m m!}
\]
Now define
\[
P(x) = \sum_{m = 0}^{(10l)^2} \frac{(-1)^m x^{2m}}{2^m m!} \,.
\]
For $x \in [-2l,2l]$, we have
\begin{align*}
|G(x) - P(x)| \leq \left \lvert \sum_{m =(10l)^2 + 1}^\infty \frac{(-1)^m x^{2m}}{2^m m!} \right \rvert \leq \sum_{m =10^2l + 1}^\infty \frac{ (2l)^{2m}}{2^m m!} \leq \sum_{m =10^2l + 1}^\infty \left(\frac{(2l)^2}{2m/3} \right)^{m} \\ \leq \sum_{m =10^2l + 1}^\infty \frac{1}{2^{m}}\leq e^{-l} \,.
\end{align*}
\end{proof}
In light of the above, we use the following notation.
\begin{definition}\label{def:poly-approx-Gaussian}
We will use $\mcl{P}_{l}(x)$ to denote the polynomial computed in Lemma \ref{lem:Gaussian-taylorseries} for parameter $l$. Note that $\mcl{P}_l$ is a polynomial of degree $(10l)^2$ and for $G = N(0,1)$, we have
\[
|G(x) - \mcl{P}_l(x)| \leq e^{-l}
\]
for $x \in [-2l, 2l]$.
\end{definition}
\subsection{Linear Regression}
Recall that at the core of the problems we are studying, we are given some function $f$ and want to approximate it as a weighted sum $a_1 f_1 + \dots + a_nf_n$ of some functions $f_1, \dots , f_n \in \mcl{F}$ for some family of functions $\mcl{F}$. The result below allows us to solve the problem of computing the coefficients if we already know the components $f_1, \dots , f_n$ that we want to use. The precise technical statement is slightly more complicated in order to incorporate the various types of additional constraints that we may want to impose on the coefficients $a_1, \dots , a_n$.
\begin{lemma}\label{lem:linear-regression}
Let $\mcl{D}$ be a distribution on $\R$ that we are given. Also assume that we are given functions $f ,f_1, \dots , f_n, g, g_1, \dots , g_n : \R \rightarrow \R$.
Assume that there are nonnegative coefficients $a_1, \dots , a_n$ such that $a_1 + \dots + a_n \leq 1$ and
\[
\int_{-\infty}^{\infty} (f(x) - a_1f_1(x) - \dots - a_nf_n(x))^2 \mcl{D}(x)dx + \int_{-\infty}^{\infty} (g(x) - a_1g_1(x) - \dots - a_ng_n(x))^2 \mcl{D}(x)dx\leq \eps^2
\]
for some parameter $\eps > 0$. Then there is an algorithm that runs in $\poly(n, \log 1/\eps)$ time and outputs nonnegative coefficients $b_1, \dots , b_n$ such that $b_1 + \dots + b_n \leq 1$ and
\[
\int_{-\infty}^{\infty} (f(x) - b_1f_1(x) - \dots - b_nf_n(x))^2 \mcl{D}(x)dx + \int_{-\infty}^{\infty} (g(x) - b_1g_1(x) - \dots - b_ng_n(x))^2 \mcl{D}(x)dx \leq 2\eps^2 \,.
\]
\end{lemma}
\begin{proof}
Let $v = (1,b_1, \dots , b_n)$. Note that we can write
\[
\int_{-\infty}^{\infty} (f(x) - b_1f_1(x) - \dots - b_nf_n(x))^2 \mcl{D}(x)dx + \int_{-\infty}^{\infty} (g(x) - b_1g_1(x) - \dots - b_ng_n(x))^2 \mcl{D}(x)dx = v^TMv
\]
where $M$ is a matrix whose entries are $\int_{-\infty}^{\infty} \left(f(x) f_j(x) + g(x)g_j(x)\right) \mcl{D}(x)dx$ in the first row and column and the other entries are $\int_{-\infty}^{\infty} \left( f_i(x) f_j(x) + g_i(x)g_j(x) \right) \mcl{D}(x)dx$. Since all of these functions are given to us, we can explicitly compute $M$. Also note that clearly $M$ is positive semidefinite. Thus, we can compute its positive semidefinite square root, say $N$. Now
\[
v^TMv = \norm{Nv}_2^2
\]
so it remains to solve $\min_{v} \norm{Nv}_2^2 $ which is a convex optimization problem that we can solve efficiently (the size of the problem is $\poly(n)$).
\end{proof}
\section{Function Approximations Using Gaussians}\label{sec:approx-with-Gaussians}
In this section, we present several results about approximating functions as a sum of Gaussians. These results will be key building blocks in the localization steps of both of our algorithms. The main result of this section, Theorem \ref{thm:approx-interval}, allows us to $\eps$-approximate the indicator function of an interval as a sum of $\poly(\log 1/\eps)$-Gaussians.
First, it will be convenient to renormalize Gaussians so that their maximum value is $1$. After renormalization, we call them Gaussian multipliers.
\begin{definition}[Gaussian Multiplier]\label{def:gaussian-multiplier}
For parameters $\mu, \sigma$, we define
\[
M_{\mu, \sigma^2}(x) = e^{-\frac{(x - \mu)^2}{2 \sigma^2}}
\]
i.e. it is a Gaussian scaled so that its maximum value is $1$.
\end{definition}
We also introduce the some additional terminology.
\begin{definition}[Significant Interval]\label{def:sig-interval}
For a Gaussian multiplier $M_{\mu, \sigma^2}$, we say the $C$-significant interval of $M$ is $[\mu - C \sigma, \mu + C \sigma ]$. We will use the same terminology for a Gaussian $N(\mu, \sigma^2)$.
\end{definition}
It will be used repeatedly that for a Gaussian (or Gaussian multiplier), $1- \eps$-fraction of its mass is contained in its $O(\sqrt{\log 1/\eps})$-significant interval. We now state the main result of this section about approximating the indicator function of an interval as a weighted sum of Gaussian multipliers.
\begin{theorem}\label{thm:approx-interval}
Let $l$ be a positive real number and $0 < \eps < 0.1 $ be a parameter. There is a function $f$ with the following properties
\begin{enumerate}
\item $f$ can be written a linear combination of Gaussian multipliers
\[
f(x) = w_1M_{\mu_1, \sigma_1^2}(x) + \dots + w_nM_{\mu_n, \sigma_n^2}(x)
\]
where $n = O((\log 1/\eps)^2)$ and $0 \leq w_1, \dots , w_n \leq 1$
\item The $10 \sqrt{\log 1/\eps}$-significant intervals of all of the $M_{\mu_i, \sigma_i^2}$ are contained in the interval $[-(1 + \eps)l, (1 + \eps)l]$
\item $0 \leq f(x) \leq 1 + \eps$ for all $x$
\item $1 - \eps \leq f(x) \leq 1 + \eps$ for all $x$ in the interval $[-l, l]$
\item $0 \leq f(x) \leq \eps$ for $x \geq (1+ \eps)l $ and $x \leq -(1 + \eps)l$
\end{enumerate}
\end{theorem}
\subsection{Approximating a Constant Function}
As a preliminary lemma, we first show how to approximate a constant function using an infinite sum of Gaussian multipliers. In particular for standard deviation $\sigma$, it suffices to add evenly spaced multipliers with spacing $\sigma (\log 1/\eps)^{-1/2}$ (or less) to get an $\eps$-approximation to a constant function. The intuition behind this observation is that the Fourier transform of a Gaussian is also a Gaussian, which has exponential tail decay.
\begin{lemma}\label{lem:approx-constant}
Let $0 < \eps < 0.1$ be a parameter. Let $c$ be a real number such that $0 < c \leq (\log 1/\eps)^{-1/2}$. Define
\[
f(x) = \sum_{j = - \infty}^{\infty} \frac{c}{\sqrt{2\pi}} M_{ c j \sigma , \sigma^2}(x) \,.
\]
Then $1 - \eps \leq f(x) \leq 1 + \eps$ for all $x$.
\end{lemma}
\begin{proof}
WLOG $\sigma = 1$. Now the function $f$ is $c$-periodic and even, so we may consider its Fourier expansion
\[
f(x) = a_0 + 2 a_1 \cos \left( \frac{2\pi x}{c}\right) + 2a_2 \cos \left( \frac{4\pi x}{c}\right)+ \dots
\]
and we will now compute the Fourier coefficients. First note that
\[
a_0 = \frac{1}{c}\int_0^c f(x) dx = \frac{1}{\sqrt{2\pi}}\sum_{j = -\infty}^{\infty} \int_{c(j+1)}^{cj} M_{0, 1}(x) dx = 1 \,.
\]
Next, for any $j \geq 1$,
\begin{align*}
a_j = \frac{1}{c} \int_0^c f(x) \cos \left( \frac{2\pi j x}{c}\right) dx = \frac{1}{\sqrt{2\pi}} \sum_{j = -\infty}^{\infty} \int_{c(j+1)}^{cj} M_{0, 1}(x) \cos \left( \frac{2\pi j x}{c}\right) dx \\ = \frac{1}{\sqrt{2\pi}} \int_{-\infty}^\infty \frac{1}{2}\left( e^{\frac{-x^2}{2} + \frac{2\pi i j x}{c}} + e^{\frac{-x^2}{2} - \frac{2\pi i j x}{c}} \right) dx = e^{-\frac{2\pi^2 j^2}{c^2}} \,.
\end{align*}
It is clear that
\[
\sum_{j=1}^{\infty} e^{-\frac{2\pi^2 j^2}{c^2}} \leq \eps
\]
so we deduce that the function $f$ is between $1 - \eps$ and $1+\eps$ everywhere and we are done.
\end{proof}
\subsection{Approximating an Interval}
The next step in the proof of Theorem \ref{thm:approx-interval} is to show how to approximate an interval using a finite number of Gaussian multipliers i.e. we need to show how to create the sharp transitions at the ends of the interval. In light of Lemma \ref{lem:approx-constant}, we can create a function satisfying the last four properties by taking $\wt{O}((1/\eps)^2)$ evenly spaced Gaussians multipliers with standard deviation $\eps^2 l$. However, this is too many components and we must reduce the number of components to $O(\log^2 1/\eps)$. The way we do this is by merging most of these components (all but the ones on the ends) into fewer components with larger standard deviation. We keep iterating this merging process and prove that we can eventually reduce the number of components to $O(\log^2 1/\eps)$.
\\\\
First, the following result is an immediate consequence of Lemma \ref{lem:approx-constant}. It allows us to approximate a Gaussian with standard deviation $2\sigma$ as a weighted sum of Gaussians with standard deviation $\sigma$.
\begin{corollary}\label{coro:approx-wider-gaussian}
Let $\eps$ be a parameter. Let $c$ be a real number such that $0 < c \leq 0.5 (\log 1/\eps)^{-1/2}$. Let
\[
g(x) = \sum_{j = - \infty}^{\infty} \frac{\sqrt{2}c}{\sqrt{3\pi}} e^{-\frac{c^2j^2}{6}} M_{c j \sigma, \sigma^2}(x) \,.
\]
Then for all $x$,
\[
(1-\eps) M_{0, 4\sigma^2}(x) \leq g(x) \leq (1 +\eps)M_{0, 4\sigma^2}(x) \,.
\]
\end{corollary}
\begin{proof}
Lemma \ref{lem:approx-constant} (with $c \leftarrow \frac{2}{\sqrt{3}}c, \sigma \leftarrow \frac{2}{\sqrt{3}} \sigma$) implies that the function
\[
f(x) = \sum_{j = - \infty}^{\infty} \frac{\sqrt{2}c}{\sqrt{3\pi}} M_{ \frac{4}{3}c j \sigma , \frac{4}{3}\sigma^2}(x)
\]
is between $1 - \eps$ and $1+ \eps$ everywhere. Now consider
\begin{align*}
f(x) \cdot M_{0, 4\sigma^2}(x) = \sum_{j = - \infty}^{\infty} \frac{\sqrt{2}c}{\sqrt{3\pi}} M_{ \frac{4}{3}c j \sigma , \frac{4}{3}\sigma^2}(x) M_{0, 4\sigma^2}(x) = \sum_{j = - \infty}^{\infty} \frac{\sqrt{2}c}{\sqrt{3\pi}} e^{ -\frac{x^2 + 3(x - \frac{4}{3}cj \sigma)^2}{8\sigma^2}} \\ = \sum_{j = - \infty}^{\infty} \frac{\sqrt{2}c}{\sqrt{3\pi}} e^{-\frac{c^2j^2}{6}} e^{-\frac{(x - cj \sigma)^2}{2\sigma^2} } = \sum_{j = - \infty}^{\infty} \frac{\sqrt{2}c}{\sqrt{3\pi}} e^{-\frac{c^2j^2}{6}} M_{c j \sigma, \sigma^2}(x) \,.
\end{align*}
\end{proof}
In the next lemma, we show when given a sum of evenly spaced Gaussians with standard deviation $\sigma$, we can replace almost all of them (except for ones on the ends) with a sum of fewer evenly spaced Gaussians with standard deviation $2\sigma$.
\begin{lemma}\label{lem:widergaussian}
Let $\eps$ be a parameter. Let $c$ be a real number such that $0 < c \leq 0.01 (\log 1/\eps)^{-1/2}$. Let $b$ be a positive integer. Consider the function
\[
f(x) = \sum_{j = 0}^{b} \frac{c}{\sqrt{2 \pi}} M_{cj\sigma, \sigma^2}(x) \,.
\]
Let $C = \lceil 10^2 c^{-1} \log (1/\eps)^{1/2} \rceil$. There is a function $g$ of the form
\[
g(x) = \sum_{j = 0}^{2C} \frac{w_j c}{\sqrt{ 2\pi}} M_{cj\sigma, \sigma^2}(x) + \sum_{j = b - 2C }^b \frac{w_j c}{\sqrt{2\pi}} M_{cj\sigma, \sigma^2}(x) + \sum_{j = \lfloor C/2 \rfloor}^{\lceil (b-C) / 2 \rceil } \frac{c}{\sqrt{2\pi}} M_{2cj\sigma, 4\sigma^2}(x)
\]
where the $0 \leq w_0, \dots , w_{2C}, w_{b - 2C}, \dots , w_{k} \leq 1$ are weights and
\[
\norm{f - g}_{\infty} \leq \eps^{10} \,.
\]
\end{lemma}
\begin{proof}
Let $\eps' = \eps^{100}$. By Corollary \ref{coro:approx-wider-gaussian}, for any real numbers $j, x$,
\[
\left \lvert M_{c j \sigma, 4\sigma^2}(x) - \sum_{k = -\infty}^{\infty} \frac{\sqrt{2}c}{\sqrt{3\pi}} e^{-\frac{c^2k^2}{6}} M_{c(k + j) \sigma , \sigma^2}(x) \right \rvert \leq \eps' M_{c j \sigma, 4\sigma^2}(x) \,.
\]
Now we use the above inequality on each term of the last sum in the expression for $g(x)$.
\begin{align*}
\left \lvert \sum_{j = \lfloor C/2 \rfloor}^{\lceil (b-C) / 2 \rceil } \frac{c M_{2cj\sigma, 4\sigma^2}(x)}{\sqrt{2\pi}} - \sum_{j = \lfloor C/2 \rfloor}^{\lceil (b-C) / 2 \rceil }\sum_{k = -\infty}^{\infty} \frac{c^2}{\pi \sqrt{3}} e^{-\frac{c^2k^2}{6}} M_{c(k + 2j) \sigma , \sigma^2}(x) \right \rvert &\leq \eps' \sum_{j = \lfloor C/2 \rfloor}^{\lceil (b-C) / 2 \rceil } \frac{c M_{2c j \sigma, 4\sigma^2}(x)}{\sqrt{2\pi}} \\ &\leq 2\eps'
\end{align*}
where the last step follows from Lemma \ref{lem:approx-constant}. Now we rewrite the second sum in the LHS above. Let
\begin{align*}
S(x) = \sum_{j = \lfloor C/2 \rfloor}^{\lceil (b-C) / 2 \rceil }\sum_{k = -\infty}^{\infty} \frac{c^2}{\pi \sqrt{3}} e^{-\frac{c^2k^2}{6}} M_{c(k + 2j) \sigma , \sigma^2}(x) = \sum_{l = -\infty}^{\infty} \frac{c^2}{\pi \sqrt{3}}M_{c l\sigma, \sigma^2}(x) \sum_{j = \lfloor C/2 \rfloor}^{\lceil (b-C) / 2 \rceil } e^{-\frac{c^2(l - 2j)^2}{6}} \,.
\end{align*}
Define
\[
a_l = \sum_{j = \lfloor C/2 \rfloor}^{\lceil (b-C) / 2 \rceil } e^{-\frac{c^2(l - 2j)^2}{6}} \,.
\]
First, by applying Lemma \ref{lem:approx-constant} with parameters $c \leftarrow \frac{2}{\sqrt{3}}c, \sigma \leftarrow \sqrt{3}c^{-1}$, we have that for all real numbers $l$,
\[
\left \lvert \frac{\sqrt{3\pi}}{\sqrt{2}c} - \sum_{-\infty}^{\infty} e^{-\frac{c^2(l - 2j)^2}{6}} \right \rvert \leq \frac{\eps' \sqrt{3\pi}}{\sqrt{2}c} \,.
\]
By the way we chose $C$, we deduce that for all integers $l$ with $2C \leq l \leq b - 2C $,
\begin{equation}\label{eq:coeffbound1}
\left \lvert \frac{\sqrt{3\pi}}{\sqrt{2}c} - a_l \right \rvert \leq (2\eps') \frac{\sqrt{3\pi}}{\sqrt{2}c}
\end{equation}
for all integers $0 \leq l \leq 2C$, or $b - 2C \leq l \leq b$,
\begin{equation}\label{eq:coeffbound2}
a_l \leq \frac{\sqrt{3\pi}}{\sqrt{2}c} (1 + 2\eps')
\end{equation}
and finally for all integers $l < 0$ or $l > b$,
\begin{equation}\label{eq:coeffbound3}
a_l \leq \frac{\sqrt{3\pi}}{\sqrt{2}c} (2\eps') \,.
\end{equation}
To obtain these inequalities, we simply use the fact that the terms in the sum
\[
\sum_{-\infty}^{\infty} e^{-\frac{c^2(l - 2j)^2}{6}}
\]
decay exponentially when $j$ is far from $l/2$ so their total contribution is small.
\\\\
Now we can set $w_0, \dots , w_{2C}, w_{b - 2C}, \dots , w_{k}$ in the expresion for $g(x)$ as follows:
\[
w_j = \max\left(0, 1 - \frac{\sqrt{2}c}{\sqrt{3\pi}}a_j \right) \,.
\]
It is clear that all of these weights are between $0$ and $1$. We now have that
\begin{align*}
\norm{f - g}_{\infty} &\leq 2\eps' + \norm{f(x) - \left(S(x) + \sum_{j = 0}^{2C} \frac{w_j c}{\sqrt{ 2\pi}} M_{cj\sigma, \sigma^2}(x) + \sum_{j = b - 2C }^b \frac{w_j c}{\sqrt{2\pi}} M_{cj\sigma, \sigma^2}(x) \right) }_{\infty}
\end{align*}
The expression inside the norm on the RHS can be rewritten as
\begin{align*}
&\sum_{l = 2C + 1}^{b - 2C - 1} \left( \frac{c}{\sqrt{2\pi}} - \frac{c^2}{\pi\sqrt{3}} a_l \right)M_{c l\sigma, \sigma^2}(x) + \sum_{l = 0}^{2C} \left( \frac{c}{\sqrt{2\pi}} (1 - w_l) - \frac{c^2}{\pi\sqrt{3}} a_l \right)M_{c l\sigma, \sigma^2}(x) \\ &+ \sum_{l = b - 2C}^{b} \left( \frac{c}{\sqrt{2\pi}} (1 - w_l) - \frac{c^2}{\pi\sqrt{3}} a_l \right)M_{c l\sigma, \sigma^2}(x) + \sum_{l = -\infty}^{-1} - \frac{c^2}{\pi\sqrt{3}} a_l M_{c l\sigma, \sigma^2}(x) + \sum_{l = b + 1}^{\infty} - \frac{c^2}{\pi\sqrt{3}} a_l M_{c l\sigma, \sigma^2}(x)
\end{align*}
and combining (\ref{eq:coeffbound1},\ref{eq:coeffbound2}, \ref{eq:coeffbound3}), we deduce that the above has $L^{\infty}$ norm at most
\[
\norm{ (10\eps') \sum_{l = -\infty}^{\infty} \frac{c}{\sqrt{2\pi}} M_{c l\sigma, \sigma^2}(x) }_{\infty} \leq 20\eps' \,.
\]
where we used Lemma \ref{lem:approx-constant}. Thus, $\norm{f - g}_{\infty} \leq 22\eps'$ and we are done.
\end{proof}
We can now prove Theorem \ref{thm:approx-interval} by repeatedly applying Lemma \ref{lem:widergaussian}.
\begin{proof}[Proof of Theorem \ref{thm:approx-interval}]
Let $c = 0.01 (\log 1/\eps)^{-1/2}$. Let $K = \lceil \frac{1 + 0.5\eps }{c\eps^2} \rceil$
\[
f_0(x) = \sum_{j = - K}^{K} \frac{c}{\sqrt{2\pi}} M_{c j \eps^2 l, \eps^4 l^2}(x) \,.
\]
Let $\eps' = \eps^{10}$. Using Lemma \ref{lem:approx-constant}, (and basic tail decay properties of a Gaussian) we get that
\begin{itemize}
\item $0 \leq f_0(x) \leq 1 + \eps'$ for all $x$
\item $1 - \eps' \leq f_0(x) \leq 1 + \eps'$ for all $x$ in the interval $[-l, l]$
\item $0 \leq f_0(x) \leq \eps'$ for $x \geq (1+ \eps)l $ and $x \leq -(1 + \eps)l$
\end{itemize}
Now we can apply Lemma \ref{lem:widergaussian} to $f_0(x)$ to obtain
\[
f_1(x) = \sum_{j = -K}^{-K + 2C} \frac{w_j c}{\sqrt{ 2\pi}} M_{cj\eps^2l, \eps^4l^2}(x) + \sum_{j = K - 2C }^{K } \frac{w_j c}{\sqrt{2\pi}} M_{cj\eps^2l, \eps^4l^2}(x) + \sum_{j = -\lceil (K - C)/2 \rceil}^{\lceil (K - C)/2 \rceil} \frac{c}{\sqrt{2\pi}}M_{2cj\eps^2l, 4\eps^4l^2}(x)
\]
where $C = \lceil 10^2 c^{-1} \log (1/\eps)^{1/2} \rceil$, the $w_j$ are weights between $0$ and $1$, and
\[
\norm{f_1 - f_0} \leq \eps' \,.
\]
Now we can apply Lemma \ref{lem:widergaussian} again on the last sum in the expression for $f_1$. We have to do this at most $10 \log 1/\eps$ times before there are at most $O((\log 1/\eps)^2)$ components remaining. It is clear that in this procedure, the $10 \sqrt{\log 1/\eps}$-significant intervals of all of the Gaussian multipliers always remains in $[-(1 + \eps)l, (1 + \eps)l]$. Also, the total $L^{\infty}$ error incurred over all of the applications of Lemma \ref{lem:widergaussian} is at most $10\eps' \log 1/\eps \leq \eps^9$. It is clear that all of the weights are always nonnegative and in the interval $[0,1]$. Thus, the final function $f$ satisfies
\begin{itemize}
\item $0 \leq f(x) \leq 1 + \eps$ for all $x$
\item $1 - \eps \leq f(x) \leq 1 + \eps$ for all $x$ in the interval $[-l, l]$
\item $0 \leq f(x) \leq \eps$ for $x \geq (1+ \eps)l $ and $x \leq -(1 + \eps)l$
\end{itemize}
and we are done.
\end{proof}
In light of Theorem \ref{thm:approx-interval}, we may make the following definition.
\begin{definition}\label{def:interval-function}
For parameters $\eps, l$, let $\mcl{I}_{\eps, l}$ denote the function computed in Theorem \ref{thm:approx-interval} for parameters $\eps, l/(1 + \eps)$ . We will also use $\mcl{I}_{\eps, l}^{(a)}$ to denote the function $\mcl{I}_{\eps, l}(x - a)$.
\end{definition}
\begin{remark}
We define $\mcl{I}_{\eps, l}$ as above because it will be convenient later to be able to say that the significant part of $\mcl{I}_{\eps, l}$ is contained in the interval $[-l , l]$.
\end{remark}
\section{Introduction}
Many works in signal processing and learning theory operate under the assumption that the underlying model is {\em simple}. We will be interested in two particular settings:
\paragraph{Fourier-Sparse Interpolation}
We would like to interpolate a signal based on noisy measurements of it at a few points \cite{chen2016fourier}. In particular, we assume we can measure a signal
$f(t) = \mcl{M}(t) + \eta(t)$
at any point in the interval $[-1, 1]$. Here $\mcl{M}(t)$ is a structured signal that has a $k$-sparse Fourier representation \--- i.e.
$$ \mcl{M}(t) = \sum_{j=1}^k a_j e^{2 \pi i \theta_j t}$$
Furthermore we assume that $\sum_j |a_j| \leq 1$ and each frequency $\theta_j$ is in the interval $[-F, F]$. Finally $\eta(t)$ is noise that we will assume is bounded in $L_2$ norm. The goal is to compute a Fourier-sparse approximation $\widetilde{\mcl{M}}(t)$ that is close to $f(t)$, in the sense that its error is comparable to that of $\mcl{M}(t)$.
\paragraph{Robustly Learning One-dimensional Gaussian Mixture Models}
We would like to estimate the components of a one-dimensional mixture model from samples. In particular, we are given random samples from a distribution $f$ with the property that it is well-approximated by a Gaussian mixture model with at most $k$ components \--- i.e. we can write
$f(x) = \mathcal{M}(x) + \eta(x)$
where
$$\mathcal{M}(x) = \sum_{i =1}^k w_i G_i(x)$$
and the $w_i$'s are the mixing weights and the $G_i$'s are Gaussians. Moreover $\eta(x)$ is the noise, and we will assume that its $L_1$ norm is at most $2\epsilon$ and that it integrates to zero so that $f(x)$ is a distribution which satisfies $d_{TV}(f, \mathcal{M}) \leq \epsilon$. The goal is to compute an approximation $\widetilde{\mathcal{M}}(x)$ to $f(x)$ that is a GMM and that is $O(\epsilon)$-close in total variation distance.
\\
Both of these problems exhibit the same structured plus noise decomposition, and share the feature that the structured portion is a sum/mixture of at most $k$ simple building blocks. These problems have been intensively investigated. Many existing results focus on the {\em unique recovery problem}, where the goal is not merely to compete with the best $k$-sparse approximation but to uniquely recover its parameters (when the noise is sufficiently small, or even zero). The difficulty is for both of these problems you usually need to make a separation assumption to do this. In particular, Moitra \cite{moitra2015super} showed that for Fourier-sparse interpolation if the minimum separation between the frequencies is beneath some critical threshold, you need the noise to be exponentially small even to approximate the constituent frequencies. See also \cite{chen2020algorithmic}. Similarly, Moitra and Valiant \cite{moitra2010settling} showed that for Gaussian mixture models you need the number of samples to grow exponentially in $k$ to estimate the mixture on a component-by-component basis.
But what happens if we do not impose a separation condition? There are exponential time algorithms for both Fourier-sparse interpolation \cite{chen2019active, avron2019universal} and learning Gaussian mixture models \cite{daskalakis2014faster, acharya2014near} that perform a grid search over the parameters. Alternatively we can relax the requirement that we output a model with exactly $k$ frequencies/components polynomial time algorithms are known. Chen, Kane, Price and Song \cite{chen2016fourier} studied Fourier-sparse interpolation without a frequency gap and showed how to construct an approximation $\widetilde{\mcl{M}}(t)$ that satisfies
$$\|f(t) - \widetilde{\mcl{M}}(t)\|_2 \leq \| \eta(t) \|_2 + \eps \|\mcl{M}(t)\|_2$$
where the $L_2$ norm is taken over the interval $[-1, 1]$. Their algorithm works for any $\eps > 0$ and uses $\mbox{poly}(k, \log 1/\eps) \log F$ measurements. Moreover the $\widetilde{\mcl{M}}(t)$ that they output is $\mbox{poly}(k, \log 1/\eps)$-Fourier sparse. For Gaussian mixture models, kernel density estimators output an estimate $\widetilde{\mathcal{M}}$ with $O(k/\epsilon^2)$ components that satisfies
$$d_{\TV}(f, \widetilde{\mathcal{M}}) = O(\epsilon)$$
There are also approaches based on piece-wise polynomial approximation \cite{chan2013efficient, acharya2017sample}.
For us, the key point is that these estimators have many more frequencies/components than we started off with. Our main question is:
\begin{question}
Can we find sparser approximations for Fourier-sparse interpolation and learning in Gaussian mixture models?
\end{question}
\noindent An important motivation for studying this question comes from {\em model selection}. In particular:
\begin{question}
What happens when we don't know $k$ in advance?
\end{question}
In fact, in many scientific applications where we expect our signals to be Fourier-sparse or where we expect our distribution to be well-approximated by a Gaussian mixture model, choosing the number of frequencies/components is the most important issue. Consider the motivation given by Chen et al. \cite{chen2004testing}: In genetics, there are many situations where we have a continuous-valued trait that we can measure across a population and we want to understand its genetic basis. In fact it is often not clear whether the underlying genetic mechanism is simple or complex \--- i.e. is it controlled by just a few genes or are there many more genes waiting to be discovered that each have a small effect on the trait? These problems are often modeled by one-dimensional Gaussian mixture models \cite{chen2004testing}.
In summary, when we model a signal as Fourier-sparse or a distribution as being well-approximated by a mixture model, the frequencies/components usually represent some features of a natural process that we would like to extract and understand from data. Determining the number of components amounts to deciding whether we already have an accurate scientific model, or if there is more complexity to be discovered. So what can we say algorithmically about model selection? We could always run our interpolation/learning algorithms for different values of $k$ and decide afterwards which one among them is a better balance of fit vs. complexity. However therein lies the problem: The models that existing algorithms output are much more complex than what we started out with. So we don't know whether the fact that we found many frequencies/components is because no simple model fits our data, or because it's a failing of our algorithm.
Model selection, particularly choosing the number of components in a Gaussian mixture model, has been intensively studied in statistics for over fifty years \cite{neyman1966use}. From a statistical perspective what makes the problem challenging is that standard analysis of the likelihood ratio test breaks down because of lack of regularity and non-identifiability \cite{hartigan1985failure}. Despite many attempts \cite{ghosh1984asymptotic, lo2001testing, huang2017model} and rejoinders \cite{jeffries2003note}, even understanding the asymptotic distribution of the likelihood ratio statistics has been described as a long-standing challenge in the field \cite{kasahara2015testing}. In fact the algorithmic problems run even deeper: Even if we could analyze the likelihood ratio test, ultimately we would need efficient algorithms for computing it and non-asymptotic guarantees. Nevertheless it turns out we will be able to solve an approximate (and robust) version of model selection as an immediate by-product of algorithms. For example:
\begin{question}
Can we test whether a given distribution is close to a Gaussian mixture model with $k$ components, or far from every such model with at most $k' = \widetilde{O}(k)$ components?
\end{question}
Despite a long line of work on the problem of learning a Gaussian mixture model, the only existing guarantees for this problem either:
\begin{enumerate}
\item[(1)] Need to assume some strong separation condition on the components, e.g. so that they are clusterable
\item[(2)] Work by first solving the parameter learning problem \cite{moitra2010settling}, in which case we need an exponential in $k$ number of samples even information-theoretically
\item[(3)] Perform a grid search \cite{daskalakis2014faster, acharya2014near} or solve a system of polynomial equations \cite{li2017robust}, which runs in time exponential in $k$
\item[(4)] Improperly learn an approximation, e.g. through piece-wise polynomial approximations \cite{chan2013efficient, acharya2017sample}. In such approximations, we can bound the number of pieces based on $k$ and the target accuracy $\epsilon$. However the number of pieces can always be much smaller, in which case we cannot directly use the output of the algorithm to bound the number of components.
\end{enumerate}
In contrast, our algorithms work without any separation condition and run in polynomial time by slightly relaxing the number of components. We also obtain similar results for model selection in the context of Fourier-sparse interpolation. We remark that the model selection problem we study is conceptually related to tolerant property testing \cite{parnas2006tolerant}. In particular, there is a natural parallel to recent work on relaxed tolerant testing for juntas \cite{diakonikolas2007testing, iyer2021junta} where we would like to distinguish between functions that are $\gamma$-close to $k$-juntas, or $\gamma+\epsilon$-far from every $k' = O(k/\epsilon^2)$-junta through query access.
\subsection{Our Results and Techniques}
Our first main result is an improved algorithm for Fourier-sparse interpolation. Throughout this paper, for any function $f$, we will use $\widehat{f}$ to denote the Fourier transform of $f$.
\begin{theorem}\label{thm:fourier-main}
Let $f$ be a function defined on $[-1,1]$ and assume we are given query access to $f$. Let $\mcl{M}$ be a function that is $k$-Fourier-sparse, has $\norm{\wh{\mcl{M}}}_1 \leq 1$, and has frequencies in the interval $[-F, F]$. Then for any desired accuracy $\eps > 0$ and constant $c > 0$, in $\poly(k, \log 1/\eps) \log F $ queries and $\poly(k /c , \log 1/ \eps) \log^2 F$ time, we can output a function $\wt{\mcl{M}}$ such that with probability $1 - 2^{-\Omega(k)}$,
\begin{enumerate}
\item $\wt{\mcl{M}}$ is $k \poly( 1/c, \log k/\eps)$-Fourier sparse with $\norm{\wh{\wt{\mcl{M}}}}_1 \leq k \poly( 1/c, \log k/\eps )$
\item
\[
\int_{-1 + c}^{1 - c} |\wt{\mcl{M}} - f|^2 \leq \poly(\log k/(\eps c))\left(\eps^2 + \int_{-1}^1 |f - \mcl{M}|^2 \right) \,.
\]
\end{enumerate}
\end{theorem}
\begin{remark}
Note the constraints $\norm{\wh{\mcl{M}}}_1 $ and $\norm{\wh{\wt{\mcl{M}}}}_1$ translate into bounds on the sizes of the coefficients of the exponentials in $\mcl{M}$ and $\wt{\mcl{M}}$ respectively.
\end{remark}
Compared to the results of Chen, Kane, Price and Song \cite{chen2016fourier}, the main differences are that we output a $\widetilde{O}(k)$-Fourier sparse approximation, which is optimal up to polylogarithmic factors. However we can only bound the error on the interval $[-1+c, 1-c]$ for any $c > 0$, for technical reasons. The number of measurements is the same because, in fact, we are able to directly sparsify the output of their algorithm, again invoking the assumption that it is close to a $k$-Fourier sparse signal.
We obtain the same sort of improvement for fitting a Gaussian mixture model:
\begin{theorem}\label{thm:main-gmm}
Let $\mcl{M} = w_1 G_1 + \dots + w_k G_k$ be an unknown mixture of Gaussians. Let $k, \eps > 0$ be parameters. Assume that we are given samples from a distribution $f$ with $d_{\TV}(\mcl{M}, f) \leq \eps$. Then given $\wt{O}(k /\eps^2)$ samples, there is an algorithm that runs in $\poly(k /\eps)$ time and with probability $0.9$ (over the random samples), outputs a mixture of $\wt{O}(k)$ Gaussians, $\wt{\mcl{M}}$, such that
\[
d_{\TV}(\wt{\mcl{M}}, f ) \leq \wt{O}(\eps) \,.
\]
\end{theorem}
Earlier methods get an approximation with $O(k/\epsilon^2)$ components, which is particularly bad for small values of $\epsilon$. Other algorithms run in exponential in $k$ time. Again, we can take a weak estimator such as a kernel density estimator or a piece-wise polynomial approximation and find a sparse approximation to it, under the assumption that there is one. We remark that in high-dimensions this is already what goes wrong: There is evidence in the form of statistical query lower bounds that approximating a Gaussian mixture model with $k$ components requires exponential time \cite{diakonikolas2017statistical}.
We take a parallel approach to both of these problems. First, we study the well-conditioned case, where all the frequencies/components are close to each other. These are the hard cases for unique recovery. By reasoning about the Taylor expansion, we show that in the well-conditioned case $O(\log 1/\epsilon)$ terms suffice to get an $O(\epsilon)$-approximation. The key idea is that this structural result allows us to map the building blocks (i.e. complex exponentials/Gaussians) into $O(\log 1/\epsilon)$ dimensional vectors. From there a straightforward application of Caratheodory's theorem tell us that any well-conditioned signal/mixture always has a $O(\log 1/\epsilon)$-sparse $O(\epsilon)$-approximation. Moreover we can make these arguments constructive. In the case of Fourier-sparse interpolation, we show that we can obliviously choose a set of frequencies corresponding to the zeros of a Chebyshev polynomial. Finally, we show how to localize any signal/mixture into well-conditioned parts by multiplying by various Gaussians which effectively zero out distant frequencies/components. To reconstruct the entire signal/mixture from localized parts, we rely on a result where we approximate the indicator function of an interval as a weighted sum of Gaussians.
Lastly, we give applications to an approximate version of model selection. More formally, we can hypothesis test to distinguish between a distribution that is $\eps$-close to a mixture of $k$ Gaussians and a distribution that is not $\wt{O}(\eps)$-close to a mixture of $\wt{O}(k)$ Gaussians.
\begin{corollary}\label{coro:gmm-hypothesis-test}
Let $k, \eps >0$ be parameters we are given. Let $\mcl{F}_1$ be the family of distributions that are $\eps$-close to a GMM with $k$ components (in TV distance). Let $\mcl{F}_2$ be the family of distributions that are not $\wt{O}(\eps)$-close to any GMM with $\wt{O}(k)$ components. There is an algorithm that given $\poly(k /\eps)$ samples from a known distribution $\mcl{D}$, runs in $\poly(k /\eps)$ time, and outputs $1$ if $\mcl{D} \in \mcl{F}_1$ and outputs $2$ if $\mcl{D} \in \mcl{F}_2$ both with failure probability at most $0.2$.
\end{corollary}
\begin{remark}
Even if the distribution $\mcl{D}$ is completely unknown and we are only given samples from it, the above result still holds as long as $\mcl{D}$ is somewhat well behaved (note that such an assumption is necessary as hypothesis testing with respect to total variation distance without any assumptions on $\mcl{D}$ is impossible). In particular we can use piecewise polynomial approximation \cite{chan2013efficient} or kernel density estimates \cite{terrell1992variable} to learn a distribution $\mcl{D}'$ that is close to $\mcl{D}$ that we have an explicit form for and then run the hypothesis test using $\mcl{D}'$.
\end{remark}
The natural open question in our work is to improve our sparsity bounds, both for interpolation/learning and model selection. In principle it could be possible that there are efficient algorithms for these problems, however it is starting to seem less and less likely. Even without noise, learning a Gaussian mixture model with $k$ components without a separation condition in time $\mbox{poly}(k, 1/\epsilon)$ is open. From our work, we see that even in the well-conditioned case this is equivalent to being able to solve a structured system of polynomial equations with $O(k)$ variables representing the parameters of the mixture and $O(\log 1/\epsilon)$ constraints representing the moments. In fact, we conjecture that both the learning and model selection problems are computationally hard if we are not allowed to relax the number of components.
\input{related_work.tex}
\section{Problem Setup}
\subsection{Sparse Fourier Reconstruction}
We first set up the problem of sparse Fourier reconstruction. First, we introduce some notation and terminology. We say a signal is $k$-Fourier sparse if it can be written as a sum of $k$ exponentials.
\begin{definition}
We say a function $f$ is $k$-Fourier sparse if $f$ can be written in the form
\[
f(x) = \sum_{j = 1}^k a_j e^{2 \pi i \theta_j x}
\]
for some $a_1, \dots , a_k \in \C$ and frequencies $\theta_1, \dots , \theta_k \in \R$
\end{definition}
We will also care about bounds on the coefficients $a_1, \dots , a_k$. To this end, we introduce the following terminology.
\begin{definition}
We say a function $f$ is $(k,C)$ simple if $f$ can be written in the form
\[
f(x) = \sum_{j = 1}^k a_j e^{2 \pi i \theta_j x}
\]
where the coefficients $a_1, \dots , a_k$ satisfy $|a_1| + \dots + |a_k| \leq C$.
\end{definition}
Our main theorem, Theorem \ref{thm:fourier-main}, allows us to reconstruct a Fourier-sparse signal with bounded coefficients over an interval using a polynomial number of queries, as long as we are able to query on a slightly larger interval. The output of our algorithm matches the Fourier sparsity of the true signal up to logarithmic factors.
Unlike most previous results, our result does not require a gap between the true frequencies and is also computationally efficient. The previous paper that achieves the most comparable guarantees is \cite{chen2016fourier}. Theorem \ref{thm:fourier-main} is similar to the main result in \cite{chen2016fourier}, which is stated below, but there are a few key differences.
\begin{theorem}[Theorem 1.1 in \cite{chen2016fourier}]\label{thm:sparse-fourier-old}
Let $f$ be a function defined on $[-1,1]$ and assume we are given query access to $f$. Let $\mcl{M}$ be a function that is $(k,1)$-simple and has frequencies in the interval $[-F, F]$. Then for any desired accuracy $\eps$, in $\poly(k, \log 1/\eps) \log F $ samples and $\poly(k, \log 1/\eps) \log^2 F$ time, we can output a function $f'$ such that with probability $1 - 2^{-\Omega(k)}$,
\begin{enumerate}
\item $f'$ is $(\poly(k, \log 1/\eps), \exp( \poly(k, \log 1/\eps)) )$- simple
\item
\[
\int_{-1}^1 | f' - f|^2 \leq O\left( \eps^2 + \int_{-1}^1 |f - \mcl{M}|^2 \right) \,.
\]
\end{enumerate}
\end{theorem}
\begin{remark}
While the bound on the coefficients of $f'$ is not explicitly stated in Theorem 1.1 in \cite{chen2016fourier}, it immediately follows from the proof.
\end{remark}
Firstly, for constant $c$, our output is $\wt{O}(k)$-Fourier sparse (instead of $\poly(k)$-Fourier sparse in \cite{chen2016fourier}). Also, the coefficients in our output are also bounded by $\wt{O}(k)$ whereas the coefficients of the output of their algorithm may be exponentially large. On the other hand, their result is stronger in that they reconstruct the function over the entire interval (instead of the subinterval $[-(1 - c), 1 - c]$). Also, their result works even if they only assume a bound on $\int_{-1}^1 |\mcl{M}|^2$ instead of a bound on the Fourier coefficients of $\mcl{M}$.
\subsubsection{Abstracting Away the Query Model}
The proof of Theorem \ref{thm:fourier-main} will involve using Theorem \ref{thm:sparse-fourier-old} as a black-box to compute a function $f'$ that approximates $f$. We will then show how to sparsify $f'$. While the form of their output is not particularly important, the main reason that we use their reconstruction algorithm as a first step is that we can then perform explicit computations with the function, such as integrating over an interval, that are difficult to perform efficiently through direct queries. With this in mind, we can eliminate the need for queries and instead assume we have explicit access to $f$ and can perform any explicit computations that we want. We discuss in Section \ref{sec:implementation} why all of these computations can be implemented efficiently using the output of Theorem \ref{thm:sparse-fourier-old}.
\subsection{Learning GMMs}
In this setting, there is an unknown GMM $\mcl{M} = w_1G_1 + \dots + w_kG_k$. We receive $\poly(k/\eps)$ samples from a distribution $f$ that is $\eps$-close to $\mcl{M}$ in TV distance and our goal is to output a GMM, say $\wt{\mcl{M}} = \wt{w_1} \wt{G_1} + \dots + \wt{w_{k'}} \wt{G_{k'}}$, that is close to $f$. Our main result, Theorem \ref{thm:main-gmm}, achieves nearly-proper learning, outputting a GMM with $\wt{O}(k)$ components (compared to $k$ components).
Our result for model order selection, Corollary \ref{coro:gmm-hypothesis-test}, follows immediately from combining Theorem \ref{thm:main-gmm} with a standard procedure for testing the TV-distance between two distributions from samples (see \cite{yatracos1985rates}).
\begin{claim}\label{claim:testL1}
Let $\mcl{D}_1, \mcl{D}_2$ be two distributions for which we have explicitly computable density functions. Let $\eps, \tau > 0$ be parameters. Assume that we are given $O(1/\eps^2 \cdot \log 1/\tau)$ samples from $\mcl{D}_1$ and can efficiently sample from $\mcl{D}_2$. Then in $\poly(1/\eps \log 1/\tau)$ time, we can compute $d$ such that with probability $1 - \tau$,
\[
| d - d_{\TV}(\mcl{D}_1, \mcl{D}_2) | \leq \eps \,.
\]
\end{claim}
\begin{proof}[Proof of of Corollary \ref{coro:gmm-hypothesis-test}]
We can run the algorithm in Theorem \ref{thm:main-gmm} with parameters $k, \eps$ to obtain an output distribution $\wt{\mcl{M}}$ that is a mixture of $\wt{O}(k)$ Gaussians. We can then use Claim \ref{claim:testL1} with parameters $\eps, 0.01$ to measure the TV-distance between $\wt{\mcl{M}}$ and $\mcl{D}$ (note that we have explicit access to the pdf of $\mcl{D}$) and output $1$ or $2$ depending on if our estimate of the TV distance is less than $\wt{O}(\eps)$. Combining the guarantees of Theorem \ref{thm:main-gmm} and Claim \ref{claim:testL1} ensures that our output satisfies the desired properties.
\end{proof}
\subsubsection{Abstracting Away the Samples}
Similar to our algorithm for sparse Fourier reconstruction, we can abstract away the sampling model by using an improper learning algorithm first. We will use the improper learner in \cite{chan2013efficient} (see Theorem 37) whose output is a piecewise polynomial. We can then only work with this piecewise polynomial, which is an explicit function that we can then perform explicit computations with.
\begin{definition}
A function $f$ is $t$-piecewise degree $d$ if there is a partition of the real line into intervals $I_1, \dots , I_t$ and polynomials $q_1(x), \dots , q_t(x)$ of degree at most $d$ such that for all $i \in [t]$, $f(x) = q_i(x)$ on the interval $I_i$.
\end{definition}
The work in \cite{chan2013efficient} guarantees to learn a piecewise polynomial $f'$ that is close to $\mcl{M}$ in $L^1$ distance when given $\wt{O}(k/\eps^2)$ samples (and they also show that this sample complexity is essentially optimal).
\begin{theorem}[\cite{chan2013efficient}]\label{thm:improper-GMMs}
Let $\mcl{M} = w_1G_1 + \dots + w_kG_k$ be an unknown mixture of Gaussians and $f$ a distribution such that $d_{\TV}(f, \mcl{M}) \leq \eps$. There is an algorithm that, given $\wt{O}(k/\eps^2)$ samples from $f$, runs in $\poly(k/\eps)$ time and returns an $O(k)$-piecewise degree $O(\log 1/\eps)$ function $f'$ such that with $0.9$ probability (over the random samples),
\[
\norm{f' - f}_1 \le O(\eps) \,.
\]
\end{theorem}
For technical reasons, we will need a few simple post-processing steps after using Theorem \ref{thm:improper-GMMs}. We can ensure that the output hypothesis $f'$ is always nonnegative by splitting each polynomial into positive and negative parts and zeroing out the negative parts (since this will not increase the $L^1$ error). Finally, we can re-normalize so that the output $f'$ is actually a distribution. This renormalization at most doubles the $L^1$ error. Thus we have:
\begin{corollary}\label{coro:improper-learner}
Let $\mcl{M} = w_1G_1 + \dots + w_kG_k$ be an unknown mixture of Gaussians and $f$ a distribution such that $d_{\TV}(f, \mcl{M}) \leq \eps$. There is an algorithm that, given $\wt{O}(k/\eps^2)$ samples from $\mcl{D}$, runs in $\poly(k/\eps)$ time and returns an $O(k \log 1/\eps)$-piecewise degree $O(\log 1/\eps)$ function $f'$ such that $f'$ is a distribution and with $0.9$ probability (over the random samples),
\[
d_{\TV}(f, f') \le O(\eps) \,.
\]
\end{corollary}
\subsection{High-Level Approach}
Our algorithms for sparse Fourier reconstruction and learning GMMs follow the same high-level approach (even though the details will need to be done separately). Note that we can abstract the problem as follows: we have explicit access to some function $f$ and we want to find a sparse approximation to $f$ in terms of functions from some family $\mcl{F}$ i.e. we want to write
\[
f \sim a_1 f_1 + \dots + a_n f_n
\]
for $f_1, \dots , f_n \in \mcl{F}$ with $n$ small. The case of sparse Fourier interpolation corresponds to when $\mcl{F}$ is the family of exponentials $\{e^{2 \pi i \theta x} \}_{\theta \in \R}$ and the case of learning GMMs corresponds to when $\mcl{F}$ is the family of Gaussians $\{ N(\mu, \sigma^2) \}_{\mu, \sigma \in \R}$.
\subsubsection{Well-Conditioned Case}
We first solve the ``well-conditioned" case where all of the components $f_1, \dots , f_n$ are not too far from each other. Concretely this means that the frequencies of the exponentials are not too far from each other or the component Gaussians have comparable means and standard deviations. The key result that we prove is that in the ``well-conditioned" case, we can always compute an $\eps$-approximation using $O(\log 1/\eps)$ components. Note that in the general case, we need to assume that $f$ is close to some $k$-sparse combination of functions in $\mcl{F}$ and then prove that we output a $\wt{O}(k)$-sparse combination. However, in the well-conditioned case, our result does not depend on $k$ at all and always outputs a $O(\log 1/\eps)$-sparse combination.
We now discuss the intuition for why $O(\log 1/\eps)$ components suffices. For each function $f_j \in \mcl{F}$, we can expand it as a Taylor series
\[
f_j(x) = c_{f_j}^{(0)} + \frac{c_{f_j}^{(1)}x}{1!} + \frac{c_{f_j}^{(2)}x^2}{2!} + \dots
\]
for some coefficients $c_{f_j}^{(0)}, c_{f_j}^{(1)}, \dots $. Now, because we are in the well-conditioned case, we can argue that it suffices to keep $O(\log 1/\eps)$ terms of the Taylor series and discard the rest. The intuition is that we can ensure that the coefficients $c_{f_j}^{(l)}$ are bounded for all $l$ so because there is $(l!)$ in the denominator, the contribution of terms with $l > O(\log 1/\eps)$ becomes negligible. Thus, we can represent each function $f_j \in \mcl{F}$ as a vector $(c_{f_j}^{(0)}, \dots , c_{f_j}^{(l-1)} ) $ of length $l = O(\log 1/\eps)$. Now we have essentially embedded $\mcl{F}$ in a $O(\log 1/\eps)$-dimensional space so we can use, e.g. Caratheodory, to argue that any function in the convex hull of $\mcl{F}$ can be approximated as a $O(\log 1/\eps)$-sparse combination.
\subsubsection{Localization}
After solving the well-conditioned case, the next step is to reduce the general version to the well-conditioned version via localization. While we will explain in more detail in the individual sections for sparse Fourier reconstruction and learning GMMs (Sections \ref{sec:sparse-fourier-full} and \ref{sec:GMM-full} respectively), we give a very high-level overview here. If $f$ were a $k$-sparse combination, say
\[
f = a_1f_1 + \dots + a_k f_k \text{ where } f_1, \dots , f_k \in \mcl{F}
\]
we can then modify $f$, e.g. by multiplying by a Gaussian, in a way that adjusts the coefficients $a_1, \dots , a_k$. To localize around some function $f_0 \in \mcl{F}$, we can ensure that after the localizing operation, the coefficients $a_j$ of functions $f_j$ that are far from $f_0$ are exponentially small. Thus, these components can be eliminated. This will leave us with only components that are not too far from each other \--- exactly the well-conditioned case which we already know how to solve. We then sum our solutions over different localizations (i.e. different choices of $f_0$) to reconstruct the entire function. The rough reason that our results need $\wt{O}(k)$ components is that we need $O(\log 1/\eps)$ components to reconstruct each localization and we may need up to $k$ localizations.
\subsection{Organization}
In Section \ref{sec:well-conditioned}, we solve the well-conditioned case for both sparse Fourier reconstruction (Section \ref{sec:fourier-well-conditioned}) and learning GMMs (Section \ref{sec:GMM-well-conditioned}). Then in Section \ref{sec:approx-with-Gaussians}, we introduce some tools for localization that will be used in solving the full versions of both problems. We finish proving our full result for sparse Fourier reconstruction in Section \ref{sec:sparse-fourier-full} and finish proving our full result for learning GMMs in Section \ref{sec:GMM-full}. Appendix \ref{appendix:basic}, contains several basic tools that will be used throughout the paper.
\subsection{Related Work}
There is a vast literature on the three problems we consider. Here we will give a more detailed review of related work.
\paragraph{Continuous Time Sparse Fourier Transforms} Sparse Fourier transforms in the continuous setting, also known as sparse Fourier transforms off the grid, has been the subject of intensive study.
Indeed, the first algorithm for this problem dates back to Prony in 1795~\cite{pronyessai}.
Modern algorithms include MUSIC~\cite{schmidt1982signal}, ESPRIT~\cite{roy1986esprit}, maximum likelihood estimators \cite{bresler1986exact}, convex programming based methods \cite{candes2014towards} and the matrix pencil method \cite{moitra2015super}.
Most of these works, especially those that work in a noisy setting, require a frequency gap. Moreover they require more than $k$ samples (their bound usually depends on the frequency gap), even if the underlying signal is $k$-sparse in the Fourier domain.
A recent line of work has focused on the problem of improving the sample complexity \--- in particular getting bounds which only depend on $k$ with runtimes that are polynomial in $k$ \cite{fannjiang2012coherence,duarte2013spectral,tang2013compressed,tang2014near,boufounos2015s,huang2015super,price2015robust}. The setting where there is no gap and there is noise is particularly challenging. One approach is to relax the definition of a frequency gap, and require it only between ``clusters" of frequencies \cite{batenkov2020conditioning}.
Another line of work~\cite{avron2019universal,chen2019active} shows how to output a hypothesis which is $k$-sparse without any gap assumptions and with sample complexity which is polynomial in $k$. However these methods run in exponential time.
As we previously discussed, the most relevant work to us is~\cite{chen2016fourier,chen2019active}, which gives an algorithm whose running time and sample complexity are polynomial in $k$ that works without any gap assumptions, but for a relaxation where we are allowed to output a $\widetilde{O} (k^2)$-Fourier sparse signal.
\paragraph{Learning Mixtures of Gaussians and Model Selection} Since the pioneering work of Pearson in 1895~\cite{pearson1894contributions}, mixtures of Gaussians have become one of the most ubiquitous and well-studied generative models in both theory and practice.
Numerous problems have been studied on the context of learning mixtures of Gaussians, including clustering~\cite{dasgupta1999learning,vempala2004spectral,achlioptas2005spectral,dasgupta2007probabilistic,AroraK07,kumar2010clustering,awasthi2012improved,mixon2017clustering,hopkins2018mixture,kothari2018robust,DiakonikolasKS18}, learning in the presence of adversarial noise in high dimensional settings~\cite{DiakonikolasKS18,hopkins2018mixture,kothari2018robust,bakshi2020robustly,diakonikolas2020robustly,kane2021robust,liu2020settling,liu2021learning}, parameter estimation~\cite{kalai2010efficiently,belkin2015polynomial,moitra2010settling,hardt2015tight}, learning in smoothed settings~\cite{hsu2013learning,anderson2014more,bhaskara2014smoothed,ge2015learning}, and density estimation~\cite{devroye2012combinatorial,chan2014efficient,acharya2017sample}.
Of particular interest to us is the line of work on proper learning~\cite{feldman2006pac,acharya2014near,li2017robust,ashtiani2018nearly}, where the goal is to output a mixture of $k$-Gaussians which is close in total variation to the underlying ground truth.
Unfortunately, while the sample complexity of these algorithms is usually polynomial, the runtime for all known approaches is exponential in $k$.
In contrast, our runtimes are polynomial, albeit for a relaxed version of the problem, where the output is allowed to be a mixture of $k'$ Gaussians, for $k' > k$.
Such a ``semi-proper'' notion of learning has been considered before, but with considerably worse quantitative guarantees.
For instance, by using kernel density estimates, one can achieve $\eps$ approximation using $k' = O(k / \eps^C)$ for some constant $C$~\cite{devroye2012combinatorial}.
Similarly~\cite{bhaskara2015sparse} achieves $\eps$ error using $k' = O(k / \eps^3)$ pieces.
That is, for both of these approaches, they require a number of pieces which scales polynomially with $1 / \eps$.
In comparison, our dependence on $\eps$ in terms of the number of pieces is logarithmic.
As discussed previously, there are strong connections between proper learning and model selection
~\cite{neyman1966use,hartigan1985failure,ghosh1984asymptotic, lo2001testing,jeffries2003note,kasahara2015testing,huang2017model}.
In particular, efficient algorithms for proper learning would immediately imply fast algorithms for model selection.
Our algorithms yield new polynomial time algorithms for a natural approximate version of model selection, where we can test if the algorithm is either close to a mixture of $k$ Gaussians, or far from any mixture of $k'$ Gaussians.
Note that in this setting, minimizing the number of components output by the algorithm is crucial to improving the approximation guarantee.
Related notions have been considered in distribution testing~\cite{parnas2006tolerant,ValiantV10a,ValiantV10b,ValiantV11a,JiaoHW16,JiaoVHW17,HanJW16} and testing properties of boolean functions~\cite{diakonikolas2007testing, iyer2021junta}.
\section{The Well-Conditioned Case}\label{sec:well-conditioned}
In this section, we deal with the well-conditioned case for both sparse Fourier reconstruction and learning GMMs.
\subsection{Well-Conditioned Case: Sparse Fourier Reconstruction}\label{sec:fourier-well-conditioned}
Here, we will consider a function that has its Fourier support contained in one interval that is not too long i.e., all of its Fourier mass is not too spread out. Note that WLOG, we may assume that this interval is centered at $0$ since otherwise we can multiply by a suitable exponential to shift the Fourier support to be around $0$. We prove the following statement:
\begin{lemma}\label{lem:Fourier-simplecase}
Let $0 < \eps < 0.1$ be a parameter and let $l \geq \lceil \log 1/\eps \rceil$ be some parameter. Let $\mcl{M}$ be a function such that $\wh{\mcl{M}}$ is supported on $[-l, l]$ and such that $\norm{\wh{\mcl{M}}}_1 \leq 1$. Also assume that we have access to a function $f$ such that
\[
\int_{-1}^1 |f(x) - \mcl{M}(x)|^2 dx \leq \eps^2 \,.
\]
There is an algorithm that runs in $\poly(l)$ time and outputs a function $\wt{\mcl{M}}$ such that $\wt{\mcl{M}}$ is $(O(l), O(l))$-simple, has Fourier support contained in $[-l, l]$, and
\[
\int_{-1}^1 |\wt{\mcl{M}}(x) - f(x)|^2 dx \leq 16\eps^2 \,.
\]
\end{lemma}
\begin{remark}
Note that in this case, we do not need any constraint on the Fourier sparsity of $\mcl{M}$ to guarantee that the output of our algorithm is $O(\log 1/\eps)$-Fourier sparse. Also, unlike our full result, Theorem \ref{thm:fourier-main}, our output in this case is guaranteed to be a good approximation over the entire interval (instead of a subinterval).
\end{remark}
Our proof will be separated into two parts. The first step will be proving the existence of a function $\wt{\mcl{M}}$ of the desired form. The second step will be developing an algorithm to actually compute it.
\subsubsection{Existence of a Sparse Approximation}
First, we will prove that under the assumptions of Lemma \ref{lem:Fourier-simplecase}, an approximation $\wt{\mcl{M}}$ satisfying the desired properties exists. We will also prove that independent of the problem instance, it suffices to only consider a fixed set of $O(l)$ distinct frequencies given by the Chebyshev points (with suitable rescaling).
The proof relies on first taking the Taylor series of an exponential $e^{2\pi i \zeta x}$ and arguing that we only need to keep the first $O(\log 1/\eps)$ terms. This essentially lets us represent such an exponential with the coefficients of its Taylor series, which are (up to rescaling) $(1, \zeta, \zeta^2 , \dots )$. We then use Corollary \ref{coro:convexhull} to argue that an arbitary linear combination of such vectors can be replaced with a sparse combination with similarly sized coefficients.
\begin{lemma}\label{lem:Fourier-existence-simplecase}
Let $0 < \eps < 0.1$ be a parameter and let $l \geq \lceil \log 1/\eps \rceil$ be some parameter. Let $t_0, \dots , t_{10^2l}$ be the degree-$10^2l$ Chebyshev points. Let $\mcl{M}$ be a function such that $\wh{\mcl{M}}$ is supported on $[-l, l]$ and such that $\norm{\wh{\mcl{M}}}_1 \leq 1$. Then there is a function
\[
h(x) = \sum_{j = 0}^{10^2l} c_j e^{2\pi i (lt_j) x}
\]
where $c_0, \dots , c_{10^2l}$ are complex numbers such that $\sum_{j = 0}^{10^2l} |c_j| \leq 200l$ and
\[
\int_{-1}^1 (h(x) - \mcl{M}(x))^2 dx \leq \eps^2 \,.
\]
\end{lemma}
\begin{proof}
Note that $\mcl{M}$ can be written as $\mcl{M}(x) = \int_{-l}^l \wh{\mcl{M}}(\zeta) e^{2 \pi i x \zeta} d\zeta $. Now consider the Taylor expansion of
\[
e^{2 \pi i x \zeta} = \sum_{j = 0}^{\infty} \frac{(2 \pi i x \zeta)^j}{j!}
\]
Note that since $ -l \leq \zeta \leq l$, we have
\[
\sum_{j = 10^2l + 1}^{\infty} \left \lvert \frac{(2 \pi i \zeta)^j}{j!} \right \rvert \leq (\eps/l)^3 \,.
\]
In particular, if we define
\[
g_{\zeta}(x) = \sum_{j = 0}^{10^2l} \frac{(2 \pi i x \zeta)^j}{j!}
\]
then over the interval $x \in [-1,1]$
\begin{equation}\label{eq:taylor-bound}
|e^{2 \pi i x \zeta} - g_{\zeta}(x)| \leq (\eps/l)^3 \,.
\end{equation}
Next, for each $\zeta \in [-l,l]$, by Corollary \ref{coro:convexhull}, we can write the vector
\[
\mcl{V}_{10^2l}(\zeta) = w_0(\zeta) \mcl{V}_{10^2l}(t_0 l) + \dots + w_{10^2l}(\zeta) \mcl{V}_{10^2l}(t_{10^2l} l)
\]
for some real numbers (depending on $\zeta$) $w_0(\zeta), \dots , w_{10l}(\zeta)$ with $\sum |w_j(\zeta)| \leq 200l$. Thus,
\[
g_{\zeta}(x) = w_0(\zeta) g_{t_0 l}(x) + \dots + w_{10^2l}(\zeta) g_{t_{10^2l} l}(x)
\]
for the same weights. Now note that by (\ref{eq:taylor-bound}), for all $x \in [-1,1]$,
\[
\left \lvert \mcl{M}(x) - \sum_{j = 0}^{10^2l} g_{t_j l}(x) \left(\int_{-l}^l \wh{\mcl{M}}(\zeta) w_j(\zeta) d\zeta \right) \right \rvert = \left \lvert \mcl{M}(x) - \int_{-l}^l \wh{\mcl{M}}(\zeta) g_{\zeta}(x) d\zeta \right \rvert \leq 2l \cdot (\eps/l)^3 \,.
\]
Note that
\[
\sum_{j = 0}^{10^2l}\left \lvert \int_{-l}^l \wh{\mcl{M}}(\zeta) w_j(\zeta) d\zeta \right \rvert \leq 200l
\]
so by (\ref{eq:taylor-bound}), for all $x \in [-1,1]$,
\[
\left \lvert \sum_{j = 0}^{10^2l} \left(g_{t_j l}(x) - e^{2\pi i x (t_j l)} \right) \left(\int_{-l}^l \wh{\mcl{M}}(\zeta) w_j(\zeta) d\zeta \right) \right \rvert \leq 200l (\eps/l)^3
\]
so therefore for all $x \in [-1,1]$, we have
\[
\left \lvert \mcl{M}(x) - \sum_{j = 0}^{10^2l} e^{2\pi i x (t_j l)} \left(\int_{-l}^l \wh{\mcl{M}}(\zeta) w_j(\zeta) d\zeta \right) \right \rvert \leq 202l \cdot (\eps/l)^3
\]
and setting
\[
h(x) = \sum_{j = 0}^{10^2l} e^{2\pi i x (t_j l)} \left(\int_{-l}^l \wh{\mcl{M}}(\zeta) w_j(\zeta) d\zeta \right)
\]
immediately leads to the desired conclusion.
\end{proof}
\subsubsection{Completing the Proof of Lemma \ref{lem:Fourier-simplecase}}
By combining Lemma \ref{lem:Fourier-existence-simplecase} and Lemma \ref{lem:linear-regression}, we can complete the proof of Lemma \ref{lem:Fourier-simplecase}.
\begin{proof}[Proof of Lemma \ref{lem:Fourier-simplecase}]
We can separate $f$ into its real and imaginary parts, say $f_{\textsf{re}},f_{\textsf{im}}$ and we can separate $\mcl{M}$ into its real and imaginary parts $\mcl{M}_{\textsf{re}},\mcl{M}_{\textsf{im}}$. Now consider the Chebyshev points of degree $10^2l$, say $t_0, \dots , t_{10^2l}$. We will now apply Lemma \ref{lem:linear-regression} where we consider the set of functions
\[
\{f_1, \dots , f_n \} = \{ \pm \cos(2\pi t_0 x), \pm \sin(2\pi t_0 x), \dots , \pm \cos(2\pi t_{10^2l} x), \pm \sin(2\pi t_{10^2l} x) \} \,.
\]
The distribution $\mcl{D}$ is the uniform distribution on $[-1,1]$ and by Lemma \ref{lem:Fourier-existence-simplecase}, there are coefficients $a_1, \dots , a_n \geq 0$ with $a_1 + \dots + a_n \leq O(l)$ (note we can split the complex coefficients $c_j$ into their real and imaginary parts and then split into positive and negative parts) such that
\begin{align*}
\int_{-1}^1 (f_{\textsf{re}}(x) - (a_1f_1(x) + \dots + a_n f_n(x)))^2 dx \leq 2\int_{-1}^1 (\mcl{M}_{\textsf{re}}(x) - (a_1f_1(x) + \dots + a_n f_n(x)))^2 dx \\ + 2\int_{-1}^1 (f_{\textsf{re}}(x) - \mcl{M}_{\textsf{re}}(x))^2 dx \leq 4\eps^2
\end{align*}
and similar for the imaginary part of $f$. Applying Lemma \ref{lem:linear-regression} to both the real and imaginary part (after rescaling by $1/(O(l))$), adding the results, and rewriting the trigonometric functions using complex exponentials (note the set $\{t_0, \dots , t_{10^2l} \}$ is symmetric around $0$ so we can do this) completes the proof.
\end{proof}
We can slightly extend Lemma \ref{lem:Fourier-simplecase} to work even if we do not know the desired accuracy $\eps$ but only a lower bound on it. It suffices to run the algorithm for Lemma \ref{lem:Fourier-simplecase} and repeatedly decrease the target accuracy until our algorithm fails to find the optimal accuracy within a constant factor.
\begin{corollary}\label{coro:fourier-simplecase}
Let $l, \eps$ be parameters given to us such that $l \geq \lceil \log 1/\eps \rceil$. Let $\mcl{M}$ be a function such that $\wh{\mcl{M}}$ is supported on $[-l, l]$ and $\norm{\wh{\mcl{M}}}_1 \leq 1$. Assume that we have access to a function $g$ defined on $[-1,1]$. There is an algorithm that runs in $\poly(l)$ time and outputs a function $\wt{\mcl{M}}$ such that $\wt{\mcl{M}}$ is $(O(l), O(l))$-simple, has Fourier support contained in $[-l, l]$, and
\begin{align*}
\int_{-1}^1 |\wt{\mcl{M}}(x) - f(x)|^2 dx \leq 20\left( \eps^2 + \int_{-1}^1 |f(x) - \mcl{M}(x)|^2 dx \right) \,.
\end{align*}
\end{corollary}
\begin{proof}
For a target accuracy $\gamma > \eps$ we run the algorithm in Lemma \ref{lem:Fourier-simplecase} to get a function $\wt{\mcl{M}}_{\gamma}(x)$. We then check whether
\[
\int_{-1}^1 |\wt{\mcl{M}}_{\gamma}(x) - f(x)|^2 dx \leq 16\gamma^2 \,.
\]
Note that the above can be explicitly computed. If the above check passes, we then take $\gamma \leftarrow 0.99 \gamma $. Taking the smallest $\gamma$ for which the above succeeds, the guarantee from Lemma \ref{lem:Fourier-simplecase} ensures that we have a function $\wt{\mcl{M}}$ such that
\[
\int_{-1}^1 |\wt{\mcl{M}}(x) - f(x)|^2 dx \leq 20 \left(\eps^2 + \int_{-1}^1 |f(x) - \mcl{M}(x)|^2 dx \right) \,.
\]
It is clear that we run the routine from Lemma \ref{lem:Fourier-simplecase} at most $O(l)$ times so we are done.
\end{proof}
\subsection{Well-Conditioned Case: Learning GMMs}\label{sec:GMM-well-conditioned}
We now deal with learning well-conditioned GMMs. We begin by formally specifying the properties that we want the components of the mixture to have. Roughly, we want the components to have comparable variances and the separation between their means cannot be too large compared to the variances. This means that after applying a suitable linear transformation, the components are all not too far from the standard Gaussian $N(0,1)$.
\begin{definition}
We say a Gaussian where $G = N(\mu, \sigma^2)$ is $\delta$-well-conditioned if
\begin{itemize}
\item $\abs{\mu} \leq \delta$
\item $|\sigma^2 - 1| \leq \delta$
\end{itemize}
We say a mixture of Gaussians $\mcl{M} = w_1 G_1 + \dots + w_kG_k$ is $\delta$-well-conditioned if all of the components $G_1, \dots , G_k$ are $\delta$-well-conditioned.
\end{definition}
We now state our learning result for well-conditioned mixtures.
\begin{lemma}\label{lem:learn-tight-mixture}
Let $\eps > 0$ be a parameter. Assume we are given access to a distribution $f$ such that $d_{\TV}(f , \mcl{M}) \leq \eps$ where $\mcl{M} = w_1G_1 + \dots + w_kG_k$ is a $0.5$-well-conditioned mixture of Gaussians. Then we can compute, in $\poly(1/\eps)$ time, a mixture $\wt{\mcl{M}}$ of at most $O(\log 1/\eps)$ Gaussians such that $d_{\TV}(\mcl{M}, \wt{\mcl{M}}) \leq \wt{O}(\eps)$.
\end{lemma}
\begin{remark}
Note that in the well-conditioned case, the number of components in the mixture that we compute does not depend on $k$.
\end{remark}
Our algorithm for proving Lemma \ref{lem:learn-tight-mixture} can be broken down into two parts. In the first part, we find a mixture of $\poly(1/\eps)$ Gaussians that approximates $f$. We then show how to reduce this mixture of $\poly(1/\eps)$ Gaussians to $O(\log 1/\eps)$ Gaussians by using the Taylor series approximation to a Gaussian.
\begin{lemma}\label{lem:gaussian-learnmanycomponents}
Let $\eps > 0$ be a parameter. Assume we are given access to a distribution $f$ such that $d_{\TV}(f , \mcl{M}) \leq \eps$ where $\mcl{M} = w_1G_1 + \dots + w_kG_k$ is a $0.5$-well-conditioned mixture of Gaussians. Then we can compute, in $\poly(1/\eps)$ time, a mixture of at most $O(1/\eps^2)$ Gaussians that is $\wt{O}(\eps)$-close to $\mcl{M}$ in TV distance.
\end{lemma}
\begin{proof}
First, let $\mcl{T}$ be the set of all $0.5$-well-conditioned Gaussians such that $\mu$ and $\sigma^2$ are integer multiples of $0.1\eps$. Note $|\mcl{T}| = O(1/\eps^2)$.
\\\\
By rounding all of the Gaussians $G_1, \dots , G_k$ to the nearest element of $\mcl{T}$ (this increases our $L^1$ error by at most $\eps$), we may assume that all of the components $G_1, \dots , G_k$ are actually in $\mcl{T}$. Now note that since $\norm{f - w_1G_1 - \dots - w_kG_k}_1 \leq 2\eps$, we have for all $x$,
\begin{equation}\label{eq:L1bound}
|\wh{f}(x) - w_1\wh{G_1}(x) - \dots - w_k\wh{G_k}(x) | \leq 2\eps
\end{equation}
where $\wh{G_j}$ denotes taking the Fourier transform of the pdf of the Gaussian $G_j$. Let $l = \lceil \log 1/\eps \rceil $. We now have,
\[
\int_{-l}^l |\wh{f}(x) - w_1\wh{G_1}(x) - \dots - w_k\wh{G_k}(x) |^2 dx \leq O(l \eps^2) \,.
\]
Now let all of the Gaussians in $\mcl{T}$ be $G_1, \dots , G_m$ where $m = |\mcl{T}|$. By Lemma \ref{lem:linear-regression} (and splitting into real and imaginary parts), we can compute in $\poly(1/\eps)$ time, nonnegative weights $\wt{w_1}, \dots , \wt{w_{m}}$ with $\wt{w_1} + \dots + \wt{w_{m}} \leq 1 $ such that
\[
\int_{-l}^l |\wh{f}(x) - \wt{w_1}\wh{G_1}(x) - \dots - \wt{w_m}\wh{G_m}(x) |^2 dx \leq O(l \eps^2)
\]
which by Cauchy Schwarz implies that
\[
\int_{-l}^l |\wh{f}(x) - \wt{w_1}\wh{G_1}(x) - \dots - \wt{w_m}\wh{G_m}(x) | dx \leq O(l\eps) \,.
\]
Now note that since all of the Gaussians $G_1, \dots , G_m$ are $0.5$-well-conditioned, their Fourier transforms $\wh{G_j}$ also decay rapidly away from $[-l,l]$ so combining the above with (\ref{eq:L1bound}), we deduce that
\[
\int_{-\infty}^{\infty} |(\wt{w_1}\wh{G_1}(x) - \dots - \wt{w_m}\wh{G_m}(x) ) - w_1\wh{G_1}(x) - \dots - w_k\wh{G_k}(x) | \leq O(l\eps) \,.
\]
From the Fourier transform of the above we then get for all $x$
\[
|\wt{w_1}G_1(x) + \dots +\wt{w_m}G_m(x) - w_1G_1(x) - \dots - w_mG_m(x)| \leq O(l\eps)
\]
and since all of the Gaussians involved are $0.5$-well-conditioned, they all decay rapidly outside the interval $[-l,l]$ and we conclude
\[
\int_{-\infty}^{\infty} |\wt{w_1}G_1(x) + \dots +\wt{w_m}G_m(x) - w_1G_1(x) - \dots - w_mG_m(x)| dx \leq O(l^2 \eps) \,.
\]
Finally, note that by the above, we must have $1 - O(l^2 \eps) \leq \wt{w_1} + \dots + \wt{w_m} \leq 1 + O(l^2 \eps)$ so rescaling to an actual mixture i.e. so that the weights $\wt{w_1} + \dots +\wt{w_m} = 1$, will affect the above error by at most $O(l^2 \eps)$. Thus, we can output this mixture and we are done.
\end{proof}
Next, as an immediate consequence of Lemma \ref{lem:Gaussian-taylorseries}, a $0.5$-well-conditioned Gaussian can be well approximated by its Taylor expansion.
\begin{corollary}\label{coro:gaussian-taylorseries}
Let $G = N(\mu, \sigma^2)$ be a $0.5$-well-conditioned Gaussian. Let $\eps > 0$ be some parameter and let $l = \lceil \log 1/\eps \rceil$. Then we can compute a polynomial $P_G(x)$ of degree $(10l)^2$ such that for all $x \in [-l, l]$,
\[
|G(x) - P_G(x)| \leq O(\eps) \,.
\]
\end{corollary}
\begin{proof}
This follows immediately from using Lemma \ref{lem:Gaussian-taylorseries} and applying the appropriate linear transformation to the polynomial.
\end{proof}
We can now complete the proof of Lemma \ref{lem:learn-tight-mixture} by using Lemma \ref{lem:gaussian-learnmanycomponents} and then using Corollary \ref{coro:gaussian-taylorseries} and Caratheodory to reduce the number of components.
\begin{proof}[Proof of Lemma \ref{lem:learn-tight-mixture}]
By Lemma \ref{lem:gaussian-learnmanycomponents}, we can compute a mixture
\[
\wt{\mcl{M}} = \wt{w_1} \wt{G_1} + \dots + \wt{w_m} \wt{G_m}
\]
such that $m = O(1/\eps^2)$ and
\[
\norm{\wt{\mcl{M}} - \mcl{M}}_1 \leq \wt{O}(\eps) \,.
\]
For each Gaussian $\wt{G_j}$, let $P_{\wt{G_j}}(x)$ be the polynomial computed in Lemma \ref{lem:Gaussian-taylorseries}. Write
\[
P_{\wt{G_j}}(x) = a_{j,0} + a_{j,1}x + \dots + a_{j,(10l)^2} x^{(10l)^2} \,.
\]
Define the vector
\[
v_j = ( a_{j,0} , a_{j,1}, \dots , a_{j,(10l)^2}) \,.
\]
Now the point $\wt{w_1}v_1 + \dots + \wt{w_m}v_m$ is in the convex hull of $v_1, \dots , v_m$. By Caratheodory (since the space is $(10l)^2 + 1$-dimensional), it must be in the convex hull of some $(10l)^2 + 1$ of the vertices. Thus, we can compute indices $i_0, \dots , i_{(10l)^2}$ and nonnegative weights $w_0', \dots , w_{(10l)^2}'$ summing to $1$ such that
\[
\wt{w_1}v_1 + \dots + \wt{w_m}v_m = w_0'v_{i_0} + \dots +w_{(10l)^2}'v_{i_{(10l)^2}} \,.
\]
The above implies that for all $x$,
\[
\wt{w_1} P_{\wt{G_1}}(x) + \dots + \wt{w_m} P_{\wt{G_m}}(x) = w_0' P_{\wt{G_{i_0}}}(x) + \dots + w_{(10l)^2}' P_{\wt{G_{i_{(10l)^2}}}}(x) \,.
\]
Now by Corollary \ref{coro:gaussian-taylorseries} and the fact that all of the Gaussians are $0.5$-well-conditioned, meaning that they decay rapidly outside of $[-l,l]$, we conclude that if we set
\[
\mcl{M}' = w_0'\wt{G_{i_0}} + \dots + w_{(10l)^2}'\wt{G_{i_{(10l)^2}}}
\]
then
\[
\norm{\wt{\mcl{M}} - \mcl{M}'}_1 \leq \wt{O}(\eps)
\]
and then we have
\[
\norm{\mcl{M} - \mcl{M}'}_1 \leq \norm{\wt{\mcl{M}} - \mcl{M}}_1 + \norm{\wt{\mcl{M}} - \mcl{M}'}_1 \leq \wt{O}(\eps)
\]
as desired.
\end{proof}
We can slightly improve Lemma \ref{lem:learn-tight-mixture} to work even when we do not have a precise estimate of $d_{\TV}(f ,\mcl{M})$ since we can just repeatedly decrease our target accuracy until we cannot improve our accuracy further. Recall that we can use Claim \ref{claim:testL1} to test the $L^1$ distance between two distributions. We now have the following (slight) improvement of Lemma \ref{lem:learn-tight-mixture}.
\begin{corollary}\label{coro:learn-tight-mixture}
Let $\eps > 0$ be a parameter. Let $\mcl{M} = w_1G_1 + \dots + w_kG_k$ be an unknown $0.5$-well-conditioned mixture of Gaussians. Assume we are given access to a distribution $f$. Then we can compute, in $\poly(1/\eps)$ time, a mixture $\wt{\mcl{M}}$ of at most $O(\log 1/\eps)$ Gaussians such that with high probability,
\[
d_{\TV}(f, \wt{\mcl{M}}) \leq \eps^2 + \poly(\log 1/\eps) d_{\TV}(f, \mcl{M}) \,.
\]
\end{corollary}
\begin{proof}
We can simply start from $\eps' = 1$ and run the algorithm in Lemma \ref{lem:learn-tight-mixture} with parameter $\eps'$ and then estimate $d_{\TV}(f, \wt{\mcl{M}})$ using Claim \ref{claim:testL1}. If $d_{\TV}(f, \wt{\mcl{M}}) \leq \eps' \poly(\log 1/\eps) $ then we can decrease $\eps'$ by a factor of $0.9$ and repeat. Repeating this process and taking the smallest accuracy $\eps' \geq \eps^3$ for which the above check succeeds, we get (from the guarantee of Lemma \ref{lem:learn-tight-mixture}) that
\[
d_{\TV}(f, \wt{\mcl{M}}) \leq \eps^2 + \poly(\log 1/\eps) d_{\TV}(f, \mcl{M})
\]
and we are done.
\end{proof} |
\section{Introduction}
Many real-world problems involve the cooperation of multiple agents, such as unmanned aerial vehicles~\citep{pham2018cooperative, zhao2018multi} and sensor networks~\citep{stranders2009decentralised}. Like in single-agent settings, learning control policies for multi-agent teams largely relies on the estimation of action-value functions, no matter in value-based~\citep{sunehag2018value, rashid2018qmix, rashid2020weighted} or policy-based approaches~\citep{lowe2017multi, foerster2018counterfactual, wang2021off}. However, learning action-value functions for complex multi-agent tasks remains a major challenge. Learning individual action-value functions~\citep{tan1993multi} is scalable but suffers from learning non-stationarity because it treats other learning agents as part of its environment. Joint action-value learning~\citep{claus1998dynamics} is free from learning non-stationarity but requires access to global information that is often unavailable during execution due to partial observability and communication constraints.
Factored Q-learning~\citep{guestrin2002multiagent} combines the advantages of these two methods. Learning the global action-value function as a combination of local utilities, factored Q functions maintain learning scalability while avoiding non-stationarity. Enjoying these advantages, fully decomposed Q functions significantly contribute to the recent progress of multi-agent reinforcement learning~\citep{samvelyan2019starcraft, wang2021rode}. However, when fully decomposed, local utility functions only depend on local observations and actions, which may lead to miscoordination problems in partially observable environments with stochastic transition functions~\citep{wang2020learning, wang2020qplex} and a game-theoretical pathology called relative overgeneralization~\citep{panait2006biasing, bohmer2020deep}. Relative overgeneralization renders optimal decentralized policies unlearnable when the employed value function does not have enough representational capacity to distinguish other agents' effects on local utility functions.
Coordination graphs~\citep{guestrin2002coordinated} provide a promising approach to solving these problems. Using vertices to represent agents and (hyper-) edges to represent payoff functions defined over the joint action-observation space of the connected agents, a coordination graph expresses a higher-order value decomposition among agents. Finding actions with the maximum value in a coordination graph can be achieved by distributed constraint optimization (DCOP) algorithms~\citep{cheng2012coordinating}, which consists of multiple rounds of message passing along the edges. Recently, DCG~\citep{bohmer2020deep} scales coordination graphs to large state-action spaces, shows its ability to solve the problem of relative overgeneralization, and obtains competitive results on StarCraft II micromanagement tasks. However, DCG focuses on predefined static and dense topologies, which largely lack flexibility for dynamic environments and induce intensive and inefficient message passing.
The question is how to learn dynamic and sparse coordination graphs sufficient for coordinated action selection. This is a long-standing problem in multi-agent learning. Sparse cooperative Q-learning~\citep{kok2006collaborative} learns value functions for sparse coordination graphs, but the graph topology is static and predefined by prior knowledge. \citet{zhang2013coordinating} propose to learn minimized dynamic coordination sets for each agent, but the computational complexity grows exponentially with the neighborhood size of an agent. Recently,~\citet{castellini2019representational} study the representational capability of several sparse graphs but focus on random topologies and stateless games. In this paper, we push these previous works further by proposing a novel deep method that learns context-aware sparse coordination graphs adaptive to the dynamic coordination requirements.
For learning sparse coordination graphs, we propose to use the variance of pairwise payoff functions as an indicator to select edges. Sparse graphs are used when selecting greedy joint actions for execution and the update of Q-function. We provide a theoretical insight into our method by proving that the probability of greedy action selection changing after an edge is removed decreases with the variance of the corresponding payoff function. Despite the advantages of sparse topologies, they raise the concern of learning instability. To solve this problem, we further equip our method with network structures based on action representations for utility and payoff learning to reduce the influence of estimation errors on sparse topologies learning. We call the overall learning framework Context-Aware SparsE Coordination graphs (CASEC).
For evaluation, we present the Multi-Agent COordination (MACO) benchmark. This benchmark collects classic coordination problems raised in the literature of multi-agent learning, increases their difficulty, and classifies them into 6 classes. Each task in the benchmark represents a type of problem. We carry out a case study on the MACO~benchmark to show that~CASEC~can discover the coordination dependence among agents under different situations and to analyze how the graph sparsity influences action coordination. We further show that CASEC~can largely reduce the communication cost (typically by 50\%) and perform significantly better than dense, static graphs and several alternative methods for building sparse graphs. We then test CASEC~on the StarCraft II micromanagement benchmark~\citep{samvelyan2019starcraft} to demonstrate its scalability and effectiveness.
\section{Background}
In this paper, we focus on fully cooperative multi-agent tasks that can be modelled as a \textbf{Dec-POMDP}~\citep{oliehoek2016concise} consisting of a tuple $G\textup{\texttt{=}}\langle I, S, A, P, R, \Omega, O, n, \gamma\rangle$, where $I$ is the finite set of $n$ agents, $\gamma\in[0, 1)$ is the discount factor, and $s\in S$ is the true state of the environment. At each timestep, each agent $i$ receives an observation $o_i\in \Omega$ drawn according to the observation function $O(s, i)$ and selects an action $a_i\in A$. Individual actions form a joint action ${\bm{a}}$ $\in A^n$, which leads to a next state $s'$ according to the transition function $P(s'|s, {\bm{a}})$, a reward $r=R(s,{\bm{a}})$ shared by all agents. Each agent has local action-observation history $\tau_i\in \mathrm{T}\equiv(\Omega\times A)^*$. Agents learn to collectively maximize the global return $Q_{tot}(s, {\bm{a}})=\mathbb{E}_{s_{0:\infty}, a_{0:\infty}}[\sum_{t=0}^{\infty} \gamma^t R(s_t, {\bm{a}}_t) | s_0=s, {\bm{a}}_0={\bm{a}}]$.
In a \textbf{coordination graph}~\citep{guestrin2002coordinated} $\mathcal{G}=\langle\mathcal{V}, \mathcal{E}\rangle$, each vertex $v_i \in \mathcal{V}$ represents an agent $i$, and (hyper-) edges in $\mathcal{E}$ represent coordination dependencies among agents. In this paper, we consider pairwise edges, and such a coordination graph induces a factorization of the global Q function:
\begin{equation}
Q_{tot}(\bm\tau, {\bm{a}}) = \frac{1}{|\mathcal{V}|}\sum_i{q_i(\tau_i, a_i)} + \frac{1}{|\mathcal{E}|}\sum_{\{i,j\}\in\mathcal{E}} q_{ij}(\bm\tau_{ij},{\bm{a}}_{ij}),
\label{equ:q_tot}
\end{equation}
where $q_i$ and $q_{ij}$ is \emph{utility} functions for individual agents and pairwise \emph{payoff} functions, respectively. $\bm\tau_{ij}=\langle\tau_i,\tau_j\rangle$ and ${\bm{a}}_{ij}=\langle a_i,a_j\rangle$ is the joint action-observation history and action of agent $i$ and $j$, respectively.
Within a coordination graph, the greedy action selection required by Q-learning can not be completed by simply computing the maximum of individual utility and payoff functions. Instead, distributed constraint optimization (DCOP)~\citep{cheng2012coordinating} techniques can be used. \textbf{Max-Sum}~\citep{stranders2009decentralised} is a popular implementation of DCOP, which finds optimal actions on a coordination graph $\mathcal{G}=\langle\mathcal{V}, \mathcal{E}\rangle$ via multi-round message passing on a bipartite graph $\mathcal{G}_m=\langle \mathcal{V}_a, \mathcal{V}_q, \mathcal{E}_m \rangle$. Each node $i\in\mathcal{V}_a$ represents an agent, and each node $g\in\mathcal{V}_q$ represents a utility ($q_i$) or payoff ($q_{ij}$) function. Edges in $\mathcal{E}_m$ connect $g$ with the corresponding agent node(s). Message passing on this bipartite graph starts with sending messages from node $i\in\mathcal{V}_a$ to node $g\in\mathcal{V}_q$:
\begin{equation}
m_{i \rightarrow g}\left(a_{i}\right)=\sum_{h \in \mathcal{F}_{i} \backslash g} m_{h \rightarrow i}\left(a_{i}\right)+c_{i g},
\label{equ:3}
\end{equation}
where $\mathcal{F}_{i}$ is the set of nodes connected to node $i$ in $\mathcal{V}_q$, and $c_{i g}$ is a normalizing factor preventing the value of messages from growing arbitrarily large. The message from node $g$ to node $i$ is:
\begin{equation}
m_{g \rightarrow i}\left(a_{i}\right)=\max _{{\bm{a}}_{g} \backslash a_{i}}\left[q\left({\bm{a}}_{g}\right)+\sum_{h \in \mathcal{V}_{g} \backslash i} m_{h \rightarrow g}\left(a_{h}\right)\right],
\label{equ:4}
\end{equation}
where $\mathcal{V}_{g}$ is the set of nodes connected to node $g$ in $\mathcal{V}_a$, ${\bm{a}}_{g}\textup{\texttt{=}}\left\{a_{h}| h \in \mathcal{V}_{g}\right\}$, ${\bm{a}}_{g} \backslash a_{i}\textup{\texttt{=}}\left\{a_{h}| h \in \mathcal{V}_{g} \backslash \{i\}\right\}$, and $q$ represents utility or payoff functions. After several iterations of message passing, each agent $i$ can find its optimal action by calculating $a_{i}^{*}=\operatorname{argmax}_{a_{i}}\sum_{h \in \mathcal{F}_{i}} m_{h \rightarrow i}\left(a_{i}\right)$.
A drawback of Max-Sum or other message passing methods (e.g., max-plus~\citep{pearl2014probabilistic}) is that running them for each action selection through the whole system results in intensive computation and communication among agents, which is impractical for most applications with limited computational resources and communication bandwidth. In the following sections, we discuss how to solve this problem by learning sparse coordination graphs.
\section{Learning Context-Aware Sparse Graphs}\label{sec:graph}
In this section, we introduce our methods for learning context-aware sparse graphs. We first introduce how we construct a sparse graph for effective action selection in Sec.~\ref{sec:q-based_graph}. After that, we introduce our learning framework in Sec.~\ref{sec:framework}. Although sparse graphs can reduce communication overhead, they raise the concern of learning instability. We discuss this problem and how to alleviate it in Sec.~\ref{sec:method-ar}.
\subsection{Construct Sparse Graphs}\label{sec:q-based_graph}
Action values, especially the pairwise payoff functions, contain much information about mutual influence between agents. Let's consider two agents $i$ and $j$. Intuitively, agent $i$ needs to coordinate its action selection with agent $j$ if agent $j$'s action exerts significant influence on the expected utility of agent $i$. For a fixed action $a_i$, $\text{Var}_{a_j}\left[q_{ij}(\bm\tau_{ij}, {\bm{a}}_{ij})\right]$ can measure the influence of agent $j$ on the expected payoff. This intuition motivates us to use the \textbf{variance of payoff functions}
\begin{equation}
\zeta_{ij}^{q_{\text{var}}} = \max_{a_i} \text{Var}_{a_j}\left[q_{ij}(\bm\tau_{ij}, {\bm{a}}_{ij})\right],
\end{equation}
as an indicator to construct sparse graphs. The maximization operator guarantees that the most affected action is considered. When $\zeta_{ij}^{q_{\text{var}}}$ is large, the expected utility of agent $i$ fluctuates dramatically with the action of agent $j$, and they need to coordinate their actions. Therefore, with this measurement, to construct sparse coordination graphs, we can set a sparseness controlling constant $\lambda\in(0,1)$ and select $\lambda n(n-1)$ edges with the largest $\zeta_{ij}$ values.
To justify this approach, we theoretically prove that, the smaller the value of $\zeta_{ij}$ is, the more likely that the Max-Sum algorithm will select the same actions after removing the edge $(i,j)$.
\begin{prop}\label{prop}
For any two agents $i$, $j$ and the edge $e_{ij}$ connecting them in the coordination graph, after removing edge $e_{ij}$, greedy actions of agent $i$ and $j$ selected by the Max-Sum algorithm keep unchanged with a probability larger than
\begin{equation}
\frac{2}{n}\left[\frac{(\bar{m}-\min_{a_j}m(a_j))(\max_{a_j}m(a_j)-\bar{m})}{\left[\zeta_{ij} \left[q(a_i,a_j)\right] + 2A^2 + 2\sqrt{A^2\left(A^2 + \zeta_{ij} \left[q(a_i,a_j)\right]\right)}\right]^2} - 1\right] ,
\end{equation}
where $m(a_j) = m_{e_{ij}\rightarrow j}(a_j)$, $\bar m$ is the average of $m(a_j)$, $A = \max_{a_j}\left[\max_{a_i} r\left(a_i, a_j\right) - r\left(a_i, a_j\right)\right]$, and $r(a_i, a_j) = q(a_i, a_j) + m_{i\rightarrow e_{ij}}(a_j)$.
\end{prop}
The lower bound in Proposition~\ref{prop} increases with a decreasing $\zeta_{ij}$. Therefore, edges with a smaller $\zeta_{ij}$ are less likely to influence the results of Max-Sum, justifying the way we construct sparse graphs.
\subsection{Learning Framework}\label{sec:framework}
Like conventional Q-learning, CASEC~consists of two main components -- learning value functions and selecting greedy actions. The difference is that these two steps are now carried out on dynamic and sparse coordination graphs.
In CASEC, agents learn a shared utility function $q_{\xi_u}(\cdot | \tau_i)$, parameterized by $\xi_u$, and a shared pairwise payoff function $q_{\xi_p}(\cdot | \bm\tau_{ij})$, parameterized by $\xi_p$. The global Q value function is estimated as:
\begin{equation}
Q_{tot}(\bm\tau,{\bm{a}}) = \frac{1}{|\mathcal{V}|}\sum_i q_{\xi_u}(a_i | \tau_i) + \frac{1}{|\mathcal{V}|(|\mathcal{V}|-1)} \sum_{i\neq j} q_{\xi_p}({\bm{a}}_{ij}|\bm\tau_{ij}),
\end{equation}
which is updated by the TD loss:
\begin{equation}
\mathcal{L}_{TD}(\xi_u, \xi_p) = \mathbb{E}_{\mathcal{D}}\left[\left(r + \gamma \hat{Q}_{tot}(\bm\tau',\text{Max-Sum}(q_{\hat{\xi}_u}, q_{\hat{\xi}_p}))-Q_{tot}(\bm\tau,{\bm{a}})\right)^2\right].\label{eq:q-learning}
\end{equation}
$\text{Max-Sum}(\cdot,\cdot)$ is the greedy joint action selected by Max-Sum, $\hat{Q}_{tot}$ is a target network with parameters $\hat{\xi}_u$, $\hat{\xi}_p$ periodically copied from $Q_{tot}$, and the expectation is estimated with uniform samples from a replay buffer $\mathcal{D}$. Meanwhile, we also minimize a sparseness loss
\begin{equation}\label{equ:sparse_loss}
\mathcal{L}_{\text{sparse}}(\xi_p) = \frac{1}{n(n-1) |A|} \sum_{i\neq j} \sum_{a_i} \text{Var}_{a_j}\left[q_{ij}(\bm\tau_{ij}, {\bm{a}}_{ij})\right],
\end{equation}
which is a regularization on $\zeta_{ij}^{q_{\text{var}}}$. Introducing a scaling factor $\lambda_{\text{sparse}} \in (0,1]$ and minimizing $\mathcal{L}_{TD}(\xi_u, \xi_p)$ $+\lambda_{\text{sparse}}\mathcal{L}_{\text{sparse}}(\xi_p)$ builds in inductive biases which favor minimized coordination graphs that would not sacrifice global returns.
Actions with the maximized value are selected for Q-learning and execution. In our framework, such action selections are finished by running Max-Sum on sparse graphs induced by $\zeta_{ij}^{q_{\text{var}}}$. Running Max-Sum requires the message passing through each node and edge. To speed up action selections, similar to previous work~\citep{bohmer2020deep}, we use multi-layer graph neural networks without parameters in the message passing module to process messages in a parallel manner.
\subsection{Stabilize Learning}\label{sec:method-ar}
One question with estimating $q_{ij}$ is that there are $|A|\times|A|$ action-pairs, each of which requires an output head in conventional deep Q networks. As only executed action-pairs are updated during Q-learning, parameters of many output heads remain unchanged for long stretches of time, resulting in estimation errors. Previous work~\citep{bohmer2020deep} uses a low-rank approximation to reduce the number of output heads. However, it is largely unclear how to choose the best rank K for different tasks, and still only $\frac{1}{|A|}$ of the output heads are selected in one Q-learning update.
This problem of estimation errors becomes especially problematic in CASEC, where building coordination graphs relies on the estimation of $q_{ij}$. A negative feedback loop is created because the built coordination graphs also affect the learning of $q_{ij}$. This loop renders learning unstable (Fig.~\ref{fig:lc-smac}).
We propose to solve this question and stabilize training by 1) periodically fixing the way we construct graphs via using the target payoff function to build graphs; and 2) accelerating the training of payoff function between target network updates to reduce the estimation errors.
Specifically, for 2), we propose to condition the utility and payoff functions on action representations to improve sample efficiency. We train an action encoder $f_{\xi_a}(a)$, whose input is the one-hot encoding of an action $a$ and output is its representation $z_a$. We adopt the technique introduced by~\citet{wang2021rode} to train an effect-based action encoder. Specifically, action representation $z_a$, together with the current local observations, is used to predict the reward and observations at the next timestep. The prediction loss is minimized to update the action encoder $f_{\xi_a}(a)$. For more details, we refer readers to Appendix~\ref{appx:action_repr}. The action encoder is trained with few samples when learning begins and remains unchanged for the rest of the training process.
Using action representations, the utility and payoff functions can now be estimated as:
\begin{equation}
\begin{aligned}
& q_{\xi_u}(\tau_i,a_i) = h_{\xi_u}(\tau_i)^{\mathrm{T}}z_{a_i}; \\
& q_{\xi_p}(\bm\tau_{ij},{\bm{a}}_{ij}) = h_{\xi_p}(\bm\tau_{ij})^{\mathrm{T}}[z_{a_i}, z_{a_j}], \label{equ:q_action_repr}\\
\end{aligned}
\end{equation}
where $h$ includes a GRU~\citep{cho2014learning} to process sequential input and output a vector with the same dimension as the corresponding action representation. $[\cdot,\cdot]$ means vector concatenation. Using Eq.~\ref{equ:q_action_repr}, no matter which action is selected for execution, all parameters in the framework ($\xi_u$ and $\xi_p$) would be updated. The detailed network structure can be found in Appendix~\ref{appx:hyper}.
\section{Multi-Agent Coordination Benchmark}
To evaluate our sparse graph learning algorithm, we collect classic coordination problems from the cooperative multi-agent learning literature, improve their difficulty, and classify them into different types. Then, 6 representative problems are selected and presented as a new benchmark called Multi-Agent COordination (MACO) challenge (Table~\ref{tab:maco}).
\begin{wraptable}[14]{r}{0.6\linewidth}
\vspace{-1em}
\caption{Multi-agent coordination benchmark.}
\label{tab:maco}
\centering
\begin{tabular}{CRCRCRCRCR}
\toprule
\multicolumn{2}{c}{Task} &
\multicolumn{2}{l}{Factored} &
\multicolumn{2}{l}{\makecell{Pairwise}} &
\multicolumn{2}{l}{\makecell{Dynamic}} &
\multicolumn{2}{l}{\# Agents}\\
\cmidrule(lr){1-2}
\cmidrule(lr){3-4}
\cmidrule(lr){5-6}
\cmidrule(lr){7-8}
\cmidrule(lr){9-10}
\multicolumn{2}{c}{Aloha} & \multicolumn{2}{c}{$\surd$} & \multicolumn{2}{c}{$\surd$} & \multicolumn{2}{c}{} & \multicolumn{2}{c}{10} \\
\cmidrule(lr){1-2}
\cmidrule(lr){3-4}
\cmidrule(lr){5-6}
\cmidrule(lr){7-8}
\cmidrule(lr){9-10}
\multicolumn{2}{c}{Pursuit} & \multicolumn{2}{c}{$\surd$} & \multicolumn{2}{c}{$\surd$} & \multicolumn{2}{c}{$\surd$} & \multicolumn{2}{c}{10} \\
\cmidrule(lr){1-2}
\cmidrule(lr){3-4}
\cmidrule(lr){5-10}
\multicolumn{2}{c}{Hallway} & \multicolumn{2}{c}{$\surd$} & \multicolumn{2}{c}{} & \multicolumn{2}{c}{} & \multicolumn{2}{c}{12} \\
\cmidrule(lr){1-2}
\cmidrule(lr){3-4}
\cmidrule(lr){5-6}
\cmidrule(lr){7-8}
\cmidrule(lr){9-10}
\multicolumn{2}{c}{Sensor} & \multicolumn{2}{c}{$\surd$} & \multicolumn{2}{c}{} & \multicolumn{2}{c}{$\surd$} & \multicolumn{2}{c}{15} \\
\toprule
\multicolumn{2}{c}{Gather} & \multicolumn{2}{c}{} & \multicolumn{2}{c}{--} & \multicolumn{2}{c}{} & \multicolumn{2}{c}{5} \\
\cmidrule(lr){1-2}
\cmidrule(lr){3-4}
\cmidrule(lr){5-6}
\cmidrule(lr){7-8}
\cmidrule(lr){9-10}
\multicolumn{2}{c}{Disperse} & \multicolumn{2}{c}{} & \multicolumn{2}{c}{--} & \multicolumn{2}{c}{$\surd$} & \multicolumn{2}{c}{12} \\
\toprule
\end{tabular}
\end{wraptable}
At the first level, tasks are classified as factored and non-factored games, where factored games present an explicit decomposition of global rewards. Factored games are further categorized according to two properties -- whether the task requires pairwise or higher-order coordination, and whether the underlying coordination relationships change temporally. For non-factored games, we divide them into two classes by whether the task characterizes static coordination relationships among agents. To better test the performance of different methods, we increase the difficulty of the included problems by extending stateless games to temporally extended settings ($\mathtt{Gather}$ and $\mathtt{Disperse}$), complicating the reward function ($\mathtt{Pursuit}$), or increasing the number of agents ($\mathtt{Aloha}$ and $\mathtt{Hallway}$). We now briefly describe the setting of each game. For detailed description, we refer readers to Appendix~\ref{appx:maco}.
$\mathtt{Aloha}$~\citep{hansen2004dynamic,oliehoek2010value} consists of $10$ islands in a $2\times 5$ array. Each island has a backlog of messages to send. They send one message or not at each timestep. When two neighboring agents send simultaneously, messages collide and have to be resent. Islands start with $1$ package in the backlog. At each timestep, with a probability $0.6$ a new packet arrives if the maximum backlog ($5$) has not been reached. Each agent observes its position and the number of packages in its backlog. Agents receive a global reward of $0.1$ for each successful transmission, and $-10$ for a collision.
$\mathtt{Pursuit}$, also called $\mathtt{Predator}$ $\mathtt{and}$ $\mathtt{Prey}$, is a classic coordination problem~\citep{benda1986optimal, stone2000multiagent, son2019qtran}. Ten agents (predators) roam a $10\times 10$ map populated with 5 random walking preys for 50 environment steps. One prey is captured if two agents catch it simultaneously, after which the catching agents and the prey are removed from the map, resulting in a team reward of $1$. If only one agent tries to catch the prey, the prey would not be captured and the agents will be punished. We consider a challenging version of $\mathtt{Pursuit}$ by setting the punishment to $1$, which is the same as the reward obtained by a successful catch.
$\mathtt{Hallway}$~\citep{wang2020learning} is a multi-chain Dec-POMDP. We increase the difficulty of $\mathtt{Hallway}$ by introducing more agents and grouping them (Fig.~\ref{fig:hallway-env}). One agent randomly spawns at a state in each chain. Agents can observe its own position and choose to move left, move right, or keep still at each timestep. Agents win with a global reward of 1 if they arrive at state $g$ simultaneously with other agents in the same group. If $n_g > 1$ groups attempt to move to $g$ at the same timestep, they keep motionless and agents receive a global punishment of $-0.5 * n_g$.
$\mathtt{Sensor}$ has been extensively studied~\citep{lesser2012distributed, zhang2011coordinated}. We consider 15 sensors in a $3\times 5$ matrix. Sensors can scan the eight nearby points. Each scan induces a cost of -1, and agents can do $\mathtt{noop}$ to save the cost. Three targets wander randomly in the gird. If $k>2$ sensors scan a target simultaneously, the system gets a reward of $1.5*k$. Agents can observe the id and position of targets nearby.
$\mathtt{Gather}$ is an extension of the $\mathtt{Climb}$ Game~\citep{wei2016lenient}. In $\mathtt{Climb}$ Game, each agent has three actions: $A_i=\{a_0, a_1, a_2\}$. Action $a_0$ yields no reward (0) if only some agents choose it, but a high reward (10) if all choose it. Otherwise, if no agent chooses action $a_0$, a reward $5$ is obtained. We increase the difficulty of this game by making it temporally extended and introducing stochasticity. Actions are no longer atomic, and agents need to learn policies to realize these actions by navigating to goals $g_1$, $g_2$ and $g_3$ (Fig.~\ref{fig:gather-env}). Moreover, for each episode, one of $g_1$, $g_2$ and $g_3$ is randomly selected as the optimal goal (corresponding to $a_0$ in the original game).
$\mathtt{Disperse}$ consists of $12$ agents. At each timestep, agents can choose to work at one of 4 hospitals by selecting an action in $A_i=\{a_0, a_1, a_2,a_3\}$. At timestep $t$, hospital $j$ needs $x^j_t$ agents for the next timestep. One hospital is randomly selected and its $x^j_t$ is a positive number, while the need of other hospitals is $0$. If $y^j_{t+1}<x^j_t$ agents go to the selected hospital, the whole team would be punished $y^j_{t+1}-x^j_t$. Agents observe the local hospital's id and its need for the next timestep.
\section{Case Study: Learning Sparse Graphs on Sensors}
We are particularly interested in the dynamics and results of sparse graph learning. Therefore, we carry out a case study on $\mathtt{Sensors}$. When training CASEC~on this task, we select edges with $10\%$ edges with largest $\zeta_{ij}^{q_\text{var}}$ values to construct sparse graphs.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig-case_study/three.pdf}
\caption{Left: Learning curves (the number of successfully scanned targets) of CASEC~and DCG on $\mathtt{Sensors}$. Middle: The influence of graph sparseness on performance (reward and the number of scanned targets). Right: An example coordination graph learned by our method on $\mathtt{Sensor}$.}
\label{fig:three}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig-case_study/case-study.pdf}
\caption{Coordination graphs learned by our method on $\mathtt{Sensor}$.}
\label{fig:case}
\end{figure}
\textbf{Interpretable sparse coordination graphs.} In Fig.~\ref{fig:three} right, we show a screenshot of the game with the learned coordination graph at this timestep. We can observe that all edges in the learned graph involve agents around the targets. The action proposed by the individual utility function of $\mathtt{agent\ 8}$ is to scan $\mathtt{target\ 1}$. After coordinating its action with other agents, $\mathtt{agent\ 8}$ changes its action selection and scans target $\mathtt{target\ 2}$, resulting in an optimal solution for the given configuration. This result is in line with our theoretical analysis in Sec.~\ref{sec:q-based_graph}. The most important edges can be characterized by a large $\zeta$ value.
\textbf{Influence of graph sparseness on performance.} It is worth noting that with fewer edges in the coordination graph, CASEC~have better performance than DCG on sensors (Fig.~\ref{fig:three} left). This observation may be counter-intuitive at the first glance. To study this problem, we load the model after convergence learned by CASEC~and DCG, gradually remove edges from the full coordination graph, and check the change of successfully scanned targets and the obtained reward. Edges are removed in the descending order of $\zeta_{ij}^{q_{\text{var}}}$. Results are shown in Fig.~\ref{fig:three} middle.
It can be observed that the performance of DCG (both the number of scanned targets and return) does not change with the number of edges in the coordination graph. In another word, only the individual utility function contributes to action selection. Screenshots shown in Fig.~\ref{fig:case} (right column) are in line with this observation. With more edges, DCG each makes a less optimal decision: agent 14 no longer scans target 2.
In contrast, the performance of CASEC~grows with more edges in the coordination graph. Another interesting observation is that the number of scanned targets grows faster than the return. By referring to Fig.~\ref{fig:case} (left column), we can conclude that CASEC~first selects edges that help agents get more reward by scanning targets, and then selects edges can eliminate useless scan action. These results prove that our method can distinguish the most important edges.
We also study the influence of the sparseness loss (Eq.~\ref{equ:sparse_loss}). As shown in Fig.~\ref{fig:three} middle, the gap between return and the number of scanned targets is larger without the sparseness loss. The lower return when the number of edges is small indicates that CASEC~without the sparseness loss cannot effectively select edges that can bring more rewards.
\section{Experiments}\label{sec:exp}
In this section, we design experiments to answer the following questions: (1) How much communication can be saved by our method? How does communication threshold influence performance on factored and non-factored games? (Sec.~\ref{sec:exp-comm}) (2) How does our method compare to state-of-the-art multi-agent learning methods? (Sec.~\ref{sec:exp-maco},~\ref{sec:exp-smac}) (3) Is our method efficient in settings with larger action-observation spaces? (Sec.~\ref{sec:exp-smac}) For results in this section, we show the median performance with 8 random seeds as well as the 25-75\% percentiles.
\subsection{Graph Sparseness}\label{sec:exp-comm}
\begin{wraptable}[9]{r}{0.32\linewidth}
\vspace{-1em}
\caption{Percentage of communication saved for each task.}
\label{tab:comm}
\centering
\begin{tabular}{CRCRCR}
\toprule
\multicolumn{2}{c}{Aloha} &
\multicolumn{2}{l}{Pursuit} &
\multicolumn{2}{l}{Hallway} \\
\cmidrule(lr){1-2}
\cmidrule(lr){3-4}
\cmidrule(lr){5-6}
\multicolumn{2}{c}{80.0\%} & \multicolumn{2}{c}{70.0\%} & \multicolumn{2}{c}{50.0\%} \\
\toprule
\multicolumn{2}{l}{Sensor} &
\multicolumn{2}{l}{Gather} &
\multicolumn{2}{l}{Disperse} \\
\cmidrule(lr){1-2}
\cmidrule(lr){3-4}
\cmidrule(lr){5-6}
\multicolumn{2}{c}{90.0\%} & \multicolumn{2}{c}{30.0\%} & \multicolumn{2}{c}{60.0\%}\\
\toprule
\end{tabular}
\vspace{-1em}
\end{wraptable}
An important advantage of learning sparse coordination graphs is reduced communication costs. The complexity of running Max-Sum for each action selection is $O(k(|\mathcal{V}|+|\mathcal{E}|))$, where $k$ is the number of iterations of message passing. Sparse graphs cut down communication costs by reducing the number of edges.
We carry out a grid search to find the communication threshold under which sparse graphs have the best performance. We find that most implementations of dynamically sparse graphs require similar numbers of edges to prevent performance from dropping significantly. In Table~\ref{tab:comm}, we show the communication cut rates we use when benchmarking our method. Generally speaking, non-factored games typically require more messages than factored games, while, for most tasks, at least $50\%$ messages can be saved without sacrificing the learning performance.
\begin{wrapfigure}[13]{r}{0.5\linewidth}
\vspace{-1em}
\includegraphics[width=\linewidth]{fig-lc/threshold_test.pdf}
\caption{The influence of graph sparseness (1.0 represents complete graphs) on the performance on factored games ($\mathtt{Sensor}$, left) and non-factored games ($\mathtt{Gather}$, right).}
\label{fig:threshold}
\end{wrapfigure}
\textbf{Communication threshold vs. performance} In Fig.~\ref{fig:threshold}, we show the performance of our method under different communication thresholds. The threshold controls the sparseness of edges. We can observe that, on the factored game $\mathtt{Sensor}$, performance first grows then drops when more edges are included in the coordination graphs. These observations are in line with the fact that sparse coordination graphs can outperform complete graphs and fully-decomposed value functions on this task. In contrast, for the non-factored game $\mathtt{Gather}$, performance stabilizes beyond a certain threshold. Non-factored games usually involve complex coordination relationships, and denser topologies are suitable for this type of questions.
\subsection{MACO: Multi-Agent Coordination Benchmark}\label{sec:exp-maco}
We compare our method with state-of-the-art fully-decomposed multi-agent value-based methods (VDN \citep{sunehag2018value}, QMIX \citep{rashid2018qmix}, and Weighted QMIX \citep{rashid2020weighted}) and coordination graph learning method (DCG \citep{bohmer2020deep}) on the MACO~benchmark (Fig.~\ref{fig:lc-maco}). Since the number of actions is not very large in the MACO~benchmark, we do not use action representations when estimating the utility difference function for CASEC.
We can see that our method significantly outperforms fully-decomposed value-based methods. The reason is that fully-decomposed methods suffer from the \emph{relative overgeneralization} issue and miscoordination problems in partially observable environments with stochasticity. For example, on task $\mathtt{Pursuit}$~\citep{benda1986optimal}, if more than one agent catches one prey simultaneously, these agents will be rewarded 1. However, if only one agent catches prey, it fails and gets a punishment of -1. For an agent with a limited sight range, the reward it obtains when taking the same action (catching a prey) under the same local observation depends on the actions of other agents and changes dramatically. This is the relative overgeneralization problem. Another example is $\mathtt{Hallway}$~\citep{wang2020learning}, where several agents need to reach a goal state simultaneously but without the knowledge of each other's location. Fully-decomposed methods cannot solve this problem if the initial positions of agents are stochastic.
For DCG, we use its default settings of complete graphs and no low-rank approximation. We observe that DCG outperforms CASEC~on $\mathtt{Hallway}$ but is less effective on tasks characterized by sparse coordination interdependence like $\mathtt{Sensor}$. We hypothesize this is because coordinating actions with all other agents requires the shared estimator to express payoff functions of most agent pairs accurately enough, which is beyond the representational capacity of the network or needs more samples to learn, hurting the performance of DCG on loosely coupled tasks.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{fig-lc/smac.pdf}
\caption{Performance and TD errors compared to baselines and ablations on the SMAC benchmark.}
\label{fig:lc-smac}
\vspace{-1.5em}
\end{figure}
\subsection{StarCraft II Micromanagement Benchmark}\label{sec:exp-smac}
We compare our method against the state-of-the-art coordination graph learning method (DCG~\citep{bohmer2020deep}) and fully decomposed value-based MARL algorithms (VDN \citep{sunehag2018value}, QMIX \citep{rashid2018qmix}). For CASEC, we use action representations to estimate the payoff function. We train the action encoder for 50$k$ samples and keep action representations unchanged afterward. In Fig.~\ref{fig:lc-smac}, we show results on $\mathtt{5m\_vs\_6m}$ and $\mathtt{MMM2}$. Detailed hyperparameter settings of our method can be found in Appendix~\ref{appx:hyper}.
For DCG, we use its default settings, including low-rank approximation for learning the payoff function. We can see that CASEC~outperforms DCG by a large margin. The result proves that sparse coordination graphs provide better scalability to large action-observation spaces than dense and static graphs. In DCG's defense, low-rank approximation still induces large estimation errors. We replace low-rank approximation with action representations and find that DCG (\emph{Full (action repr.)}) achieves similar performance to CASEC~after 5M steps, but CASEC~is still more sample-efficient. Moreover, taking advantage of higher-order value decomposition, CASEC~is able to represent more complex coordination dynamics than fully decomposed value functions and thus performs better.
\textbf{Ablation study} Our method is characterized by two contributions: context-aware sparse topologies and action representations for learning the utility difference function. In this section, we design ablations to show their contributions.
The effect of sparse topologies can be observed by comparing CASEC~to \emph{Full (action repr.)}, which is the same as CASEC~other than using complete coordination graphs. We observe that sparse graphs enjoy better sample efficiency than full graphs, and the advantage becomes less obvious as more samples are collected. This observation indicates that sparse graphs introduce inductive biases that can accelerate training, and their representational capacity is similar to that of full graphs.
From the comparison between CASEC~to CASEC~using conventional Q networks (\emph{w/o action repr.}), we can see that using action representations can significantly stabilize learning. For example, learning diverges on $\mathtt{5m\_vs\_6m}$ without action representations. As analyzed before, this is because a negative feedback loop is created between the inaccurate utility difference function and coordination graphs.
To further consolidate that action representations can reduce the estimation errors and thus alleviate learning oscillation as discussed in Sec.~\ref{sec:method-ar}, we visualize the TD errors of CASEC~and ablations during training in Fig.~\ref{fig:lc-smac} right. We can see that action representations can dramatically reduce the TD errors. For comparison, the low-rank approximation can also reduce the TD errors, but much less significantly. Smaller TD errors prove that action representations provide better estimations of the value function, and learning with sparse graphs can thus be stabilized (Fig.~\ref{fig:lc-smac} left).
\section{Conclusion}
We study how to learn dynamic sparse coordination graphs, which is a long-standing problem in cooperative MARL. We propose a specific implementation and theoretically justify it. Empirically, we evaluate the proposed method on a new multi-agent coordination benchmark. Moreover, we equip our method with action representations to improve the sample efficiency of payoff learning and stabilize training. We show that sparse and adaptive topologies can largely reduce communication overhead as well as improve the performance of coordination graphs. We expect our work to extend MARL to more realistic tasks with complex coordination dynamics.
One limitation of our method is that the learned sparse graphs are not always cycle-free. Since the Max-Sum algorithm guarantees optimality only on acyclic graphs, our method may select sub-optimal actions. We plan to study how to solve this problem in future work.
Another limitation is that we set a fixed communication threshold then train and test our method. For future work, we plan to study how to adaptively determine the communication threshold during learning, automatically and accurately find the minimum threshold that can guarantee the learning performance.
\paragraph{Reproducibility} The source code for all the experiments along with a README file with instructions on how to run these experiments is attached in the supplementary material. In addition, the settings and parameters for all models and algorithms mentioned in the experiment section are detailed in Appendix~\ref{appx:hyper}.
\section{MACO: Multi-Agent Coordination Benchmark}\label{appx:maco}
In this paper, we study how to learn context-aware sparse coordination graphs. For this purpose, we propose a new Multi-Agent COordination (MACO) benchmark (Table~\ref{tab:maco}) to evaluate different implementations and benchmark our method. This benchmark collects classic coordination problems in the literature of cooperative multi-agent learning, increases their difficulty, and classifies them into different types. We now describe the detailed settings of tasks in the MACO~benchmark.
\subsection{Task Settings}
\textbf{Factored Games} are characterized by a clear factorization of global rewards. We further classify factored games into 4 categories according to whether coordination dependency is pairwise and whether the underlying coordination graph is dynamic (Table~\ref{tab:maco}).
$\mathtt{Aloha}$ (\citet{oliehoek2010value}, also similar to the Broadcast Channel benchmark problem proposed by \citet{hansen2004dynamic}) consists of $10$ islands, each equipped with a radio tower to transmit messages to its residents. Each island has a backlog of messages that it needs to send, and agents can choose to send one message or not at each timestep. Due to the proximity of islands, communications from adjacent islands interfere. This means that when two neighboring agents attempt to send simultaneously, a collision occurs and the messages have to be resent. Each island starts with $1$ package in its backlog. At each timestep, with probability $0.6$ a new packet arrives if the maximum backlog (set to $5$) has not been reached. Each agent observes its position and the number of packages in its backlog. A global reward of $0.1$ is received by the system for each successful transmission, while punishment of $-10$ is induced if the transmission leads to a collision.
$\mathtt{Pursuit}$, also called $\mathtt{Predator}$ $\mathtt{and}$ $\mathtt{Prey}$, is a classic coordination problem~\citep{benda1986optimal, stone2000multiagent, son2019qtran}. In this game, ten agents (predators) roam a $10\times 10$ map populated with 5 random walking preys for 50 environment steps. Based on the partial observation of any adjacent prey and other predators, agents choose to move in four directions, keep motionless, or catch prey (specified by its id). One prey is captured if two agents catch it simultaneously, after which the catching agents and the prey are removed from the map, resulting in a team reward of $1$. However, if only one agent tries to catch the prey, the prey would not be captured and the agents will be punished. The difficulty of $\mathtt{Pursuit}$ is largely decided by the relative scale of the punishment compared to the catching reward~\citep{bohmer2020deep}, because a large punishment exacerbates the relative overgeneralization pathology. In the MACO~benchmark, we consider a challenging version of $\mathtt{Pursuit}$ by setting the punishment to $1$, which is the same as the reward obtained by a successful catch.
$\mathtt{Hallway}$~\citep{wang2020learning} is a multi-chain Dec-POMDP whose stochasticity and partial observability lead to fully-decomposed value functions learning sub-optimal strategies. In the MACO~benchmark, we increase the difficulty of $\mathtt{Hallway}$ by introducing more agents and grouping them (Fig.~\ref{fig:hallway-env}). One agent randomly spawns at a state in each chain. Agents can observe its own position and choose to move left, move right, or keep still at each timestep. Agents win if they arrive at state $g$ simultaneously with other agents in the same group. In Fig.~\ref{fig:hallway-env}, different groups are drawn in different colors.
\begin{wrapfigure}[11]{r}{0.4\linewidth}
\vspace{-1em}
\centering
\includegraphics[width=\linewidth]{fig-maco/hallway-env.pdf}
\caption{Task $\mathtt{Hallway}$~\citep{wang2020learning}. To increase the difficulty of the game, we consider a multi-group version. Different colors represent different groups.} \label{fig:hallway-env}
\end{wrapfigure}
Each winning group induces a global reward of 1. Otherwise, if any agent arrives at $g$ earlier than others, the system receives no reward and all agents in that group would be removed from the game. If $n_g > 1$ groups attempt to move to $g$ at the same timestep, they keep motionless and agents receive a global punishment of $-0.5 * n_g$. The horizon is set to $max_i\ l_i + 10$ to avoid an infinite loop, where $l_i$ is the length of chain $i$.
$\mathtt{Sensor}$ (Fig. 3 in the main text) has been extensively studied in cooperative multi-agent learning~\citep{lesser2012distributed, zhang2011coordinated}. We consider 15 sensors arranged in a 3 by 5 matrix. Each sensor is controlled by an agent and can scan the eight nearby points. Each scan induces a cost of -1, and agents can choose $\mathtt{noop}$ to save the cost. Three targets wander randomly in the gird. If two sensors scan a target simultaneously, the system gets a reward of 3. The reward increases linearly with the number of agents who scan a target simultaneously -- all agents are jointly rewarded by $4.5$ and $6$ if three and four agents scan a target, respectively. Agents can observe the id and position of targets nearby.
\textbf{Non-factored games} do not present an explicit decomposition of global rewards. We classify non-factored games according to whether the game can be solved by a static (sparse) coordination graph in a single episode.
$\mathtt{Gather}$ is an extension of the $\mathtt{Climb}$ Game~\citep{wei2016lenient}. In $\mathtt{Climb}$ Game, each agent has three actions: $A_i=\{a_0, a_1, a_2\}$. Action $a_0$ yields no reward if only some agents choose it, but a high reward if all choose it. The other two actions are sub-optimal actions but can induce a positive reward without requiring precise coordination:
\begin{equation}
r({\bm{a}}) = \begin{cases}
10 & \# a_0 = n,\\
0 & 0< \# a_0< n, \\
5 & \text{otherwise}.
\end{cases}\label{equ:climb_env}
\end{equation}
We increase the difficulty of this game by making it temporally extended and introducing stochasticity. We consider three actions. Actions are no longer atomic, and agents need to learn policies to realize these actions by navigating to goals $g_1$, $g_2$ and $g_3$ (Fig.~\ref{fig:gather-env}).
\begin{wrapfigure}[13]{r}{0.35\linewidth}
\vspace{-1em}
\centering
\includegraphics[width=\linewidth]{fig-maco/gather-env.pdf}
\caption{Task $\mathtt{Gather}$. To increase the difficulty of this game, we consider a temporally extended version and introduce stochasticity.}
\label{fig:gather-env}
\end{wrapfigure}
Moreover, for each episode, one of $g_1$, $g_2$ and $g_3$ is randomly selected as the optimal goal (corresponding to $a_0$ in Eq.~\ref{equ:climb_env}). Agents spawn randomly, and only agents initialized near the optimal goal know which goal is optimal. Agents need to simultaneously arrive at a goal state to get any reward. If all agents are at the optimal goal state, they get a high reward of $10$. If all of them are at other goal states, they would be rewarded $5$. The minimum reward would be received if only some agents gather at the optimal goal. We further increase the difficulty by setting this reward to $-5$. It is worth noting that, for any single episode, $\mathtt{Gather}$ can be solved using a static and sparse coordination graph -- for example, agents can collectively coordinate with an agent who knows the optimal goal.
\begin{table*} [ht]
\caption{Temporally average performance on the multi-agent coordination benchmark.}
\label{tab:maco_avg_perf}
\centering\scriptsize
\begin{tabular}{c| P{0.25cm} P{0.25cm} c|P{0.25cm} P{0.25cm} c|P{0.25cm} P{0.25cm} c|P{0.25cm} P{0.25cm} c|P{0.25cm} P{0.25cm} c | c}
\Xhline{2\arrayrulewidth}
Topology& \multicolumn{3}{c|}{Full} & \multicolumn{3}{c|}{Rand.} & \multicolumn{3}{c|}{$\delta_{\max}$} & \multicolumn{3}{c|}{$q_{\text{var}}$} & \multicolumn{3}{c|}{$\delta_{\text{var}}$} & Attn. \\
\hline
Loss & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & -- \\
\hline
Aloha & 11.7& 36.8& 40.8& 39.1& 36.0& 37.2& 34.6& 32.0& 41.4& \textbf{44.8}& \textbf{41.9}& 32.2& 33.8& 29.7& \textbf{42.1} & 15.9\\
Pursuit & 3.93& 3.96& 3.98& 3.89& 3.88& 3.85& 4.05& \textbf{4.11}& \textbf{4.12}& 4.05& 4.08& 4.08& 4.02& 4.07& \textbf{4.09} & 3.87 \\
Hallway & \textbf{0.51}& 0.46& \textbf{0.51}& 0.47& 0.44& 0.47& 0.29& 0.38& 0.41& 0.33& 0.47& 0.42& \textbf{0.54}& 0.50& 0.46 & 0.00 \\
Sensor & 6.32& 0.00& 0.00& 6.80& 8.96& 6.77& \textbf{21.0}& \textbf{20.6}& 20.4& 20.4& 19.9& \textbf{20.7}& 20.1& 20.3& 19.7& 5.67\\
Gather & 0.52& \textbf{0.73}& 0.72& 0.54& \textbf{0.73}& 0.61& 0.67& \textbf{0.77}& \textbf{0.73}& 0.58& 0.72& 0.67& 0.64& 0.68& 0.65& 0.34\\
Disperse & 6.94& 6.67& 6.68& 7.06& 7.07& 7.00& 7.60& 7.68& 7.86& 7.56& 7.77& 7.70& \textbf{8.05}& \textbf{8.20}& 7.88& \textbf{8.49}\\
\Xhline{2\arrayrulewidth}
\end{tabular}
\end{table*}
\begin{table*} [t]
\caption{Stability of each implementation. The averaged distance between raw and highly smoothed curves is shown. The smaller the values are, the more stable the learning curves are. Stability scores shown in the main text have been scaled reversely to $[0,1]$.}
\label{tab:maco_fluc}
\centering\scriptsize
\begin{tabular}{c| P{0.25cm} P{0.25cm} c|P{0.25cm} P{0.25cm} c|P{0.25cm} P{0.25cm} c|P{0.25cm} P{0.25cm} c|P{0.25cm} P{0.25cm} c | c}
\Xhline{2\arrayrulewidth}
Topology& \multicolumn{3}{c|}{Full} & \multicolumn{3}{c|}{Rand.} & \multicolumn{3}{c|}{$\delta_{\max}$} & \multicolumn{3}{c|}{$q_{\text{var}}$} & \multicolumn{3}{c|}{$\delta_{\text{var}}$} & Attn. \\
\hline
Loss & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & -- \\
\hline
Aloha & 2.94& 1.18& 0.96& 0.25& 0.22& 0.22& 1.28& 0.93& 1.03& 0.66& 0.81& 2.20& 0.85& 1.79& 0.75& 1.33 \\
Pursuit & 0.35& 0.31& 0.41& 0.23& 0.21& 0.25& 0.20& 0.20& 0.26& 0.20& 0.13& 0.21& 0.19& 0.18& 0.18& 0.19\\
Hallway & 1.78& 1.85& 0.65& 0.42& 0.81& 1.19& 0.92& 0.84& 0.72& 0.67& 0.92& 0.78& 0.47& 1.79& 2.13& 1.63\\
Sensor & 0.76& 0.63& 0.89& 0.49& 0.18& 0.44& 0.11& 0.15& 0.18& 0.14& 0.11& 0.12& 0.12& 0.11& 0.11& 0.23\\
Gather & 0.29& 0.21& 0.21& 0.31& 0.27& 0.28& 0.14& 0.19& 0.21& 0.34& 0.14& 0.21& 0.21& 0.19& 0.21& 0.59\\
Disperse & 0.35& 0.53& 0.41& 0.26& 0.28& 0.39& 0.16& 0.23& 0.19& 0.25& 0.46& 0.33& 0.17& 0.16& 0.24& 0.06\\
\Xhline{2\arrayrulewidth}
\end{tabular}
\end{table*}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig-maco/aloha.pdf}
\caption{Performance of different implementations on $\mathtt{Aloha}$. Different colors indicate different topologies. Performance of different losses is shown in different sub-figures.}
\label{fig:aloha-lc}
\end{figure}
\begin{figure*}[ht]
\centering
\includegraphics[width=\linewidth]{fig-maco/pursuit.pdf}
\caption{Performance of different implementations on $\mathtt{Pursuit}$. Different colors indicate different topologies. Performance of different losses is shown in different sub-figures.}
\label{fig:prey-lc}
\end{figure*}
\begin{figure*}[ht]
\centering
\includegraphics[width=\linewidth]{fig-maco/hallway.pdf}
\caption{Performance of different implementations on $\mathtt{Hallway}$. Different colors indicate different topologies. Performance of different losses is shown in different sub-figures.}
\label{fig:hallway-lc}
\end{figure*}
\begin{figure*}[ht]
\centering
\includegraphics[width=\linewidth]{fig-maco/sensor.pdf}
\caption{Performance of different implementations on $\mathtt{Sensor}$. Different colors indicate different topologies. Performance of different losses is shown in different sub-figures.}
\label{fig:sensor-lc}
\end{figure*}
\begin{figure*}[ht]
\centering
\includegraphics[width=\linewidth]{fig-maco/gather.pdf}
\caption{Performance of different implementations on $\mathtt{Gather}$. Different colors indicate different topologies. Performance of different losses is shown in different sub-figures.}
\label{fig:gather-lc}
\end{figure*}
\begin{figure*}[ht]
\centering
\includegraphics[width=\linewidth]{fig-maco/disperse.pdf}
\caption{Performance of different implementations on $\mathtt{Disperse}$. Different colors indicate different topologies. Performance of different losses is shown in different sub-figures.}
\label{fig:disperse-lc}
\end{figure*}
$\mathtt{Disperse}$ consists of $12$ agents. At each timestep, agents can choose to work at one of 4 hospitals by selecting an action in $A_i=\{a_0, a_1, a_2,a_3\}$. At timestep $t$, hospital $j$ needs $x^j_t$ agents for the next timestep. One hospital is randomly selected and its $x^j_t$ is a positive number, while the need of other hospitals is $0$. If $y^j_{t+1}<x^j_t$ agents go to the selected hospital, the whole team would be punished $y^j_{t+1}-x^j_t$. Agents observe the local hospital's id and its need for the next timestep.
\subsection{Performance}
With this benchmark in hand, we are now able to evaluate our method for constructing sparse graphs. We compare our method with the following approaches.
\textbf{Maximum utility difference}\ \ $q_i$ (or $q_j$) is the expected utility agent $i$ (or $j$) can get without the awareness of actions of other agents. After specifying the action of agent $j$ or $i$, the joint expected utility changes to $q_{ij}$. Thus the measurement
\begin{align}
\zeta_{ij}^{\delta_{\max}} & = \max_{{\bm{a}}_{ij}} |\delta_{ij}(\bm\tau_{ij}, {\bm{a}}_{ij})|
\end{align}
can describe the mutual influence between agent $i$ and $j$. Here
\begin{equation}
\delta_{ij}(\bm\tau_{ij},{\bm{a}}_{ij}) = q_{ij}(\bm\tau_{ij},{\bm{a}}_{ij}) - q_i(\tau_i, a_i) - q_j(\tau_j, a_j)
\label{equ:2}
\end{equation}
is the \textbf{\emph{utility difference function}}.
We use a maximization operator here because two agents need to coordinate with each other if such coordination significantly affects the probability of selecting at least one action pair.
\textbf{Variance of utility difference}\ \ As discussed before, the value of utility difference $\delta_{ij}$ and variance of payoff functions can measure the mutual influence between agent $i$ and $j$. In this way, the variance of $\delta_{ij}$ serves as a second-order measurement, and we can use
\begin{equation}
\zeta_{ij}^{\delta_{\text{var}}} = \max_{a_i} \text{Var}_{a_j}\left[\delta_{ij}(\bm\tau_{ij}, {\bm{a}}_{ij})\right]\label{equ:zeta_delta_var}
\end{equation}
to rank the necessity of coordination relationships between agents. Again we use the maximization operation to base the measurement on the most influenced action.
For these three measurements, the larger the value of $\zeta_{ij}$ is, the more important the edge $\{i,j\}$ is. For example, when $\max_{a_i}\text{Var}_{a_j}\left[q_{ij}(\bm\tau_{ij}, {\bm{a}}_{ij})\right]$ is large, the expected utility of agent $i$ fluctuates dramatically with the action of agent $j$, and they need to coordinate their actions. Therefore, with these measurements, to learn sparse coordination graphs, we can set a sparseness controlling constant $\lambda\in(0,1)$ and select $\lambda n(n-1)$ edges with the largest $\zeta_{ij}$ values. To make the measurements more accurate in edge selection, we minimize the following losses for the two measurements, respectively:
\begin{align}
\mathcal{L}^{|\delta|}_{\text{sparse}} &= \frac{1}{n(n-1) |A|^2} \sum_{i\neq j}\sum_{a_i, a_j} |\delta_{ij}(\bm\tau_{ij}, {\bm{a}}_{ij})|; \nonumber \\
\mathcal{L}^{\delta_{\text{var}}}_{\text{sparse}} &= \frac{1}{n(n-1) |A|} \sum_{i\neq j} \sum_{a_i} \text{Var}_{a_j}\left[\delta_{ij}(\bm\tau_{ij}, {\bm{a}}_{ij})\right].
\end{align}
We scale these losses with a factor $\lambda_{\text{sparse}}$ and optimize them together with the TD loss. It is worth noting that these measurements and losses are not independent. For example, minimizing $\mathcal{L}^{\delta_{\text{var}}}_{\text{sparse}}$ would also reduce the variance of $q_{ij}$. Thus, in the next section, we consider all possible combinations between these measurements and losses.
\textbf{Observation-Based Approaches}
In partial observable environments, agents sometimes need to coordinate with each other to share their observations and reduce their uncertainty about the true state~\citep{wang2020learning}. We can build our coordination graphs according to this intuition.
Agents use an attention model~\citep{vaswani2017attention} to select the information they need. Specifically, we train fully connected networks $f_k$ and $f_q$ and estimate the importance of agent $j$’s observations to agent $i$ by:
\begin{equation}
\alpha_{ij} = f_k(\tau_i)^{\mathrm{T}} f_q(\tau_j).
\end{equation}
Then we calculate the global Q function as:
\begin{equation}
Q_{tot}(s, {\bm{a}}) = \frac{1}{|\mathcal{V}|}\sum_i{q_i(\tau_i, a_i)} + \sum_{i\neq j} \bar{\alpha}_{ij}q_{ij}(\bm\tau_{ij},{\bm{a}}_{ij}),
\end{equation}
where $\bar{\alpha}_{ij} = e^{\alpha_{ij}} / \sum_{i\neq j}e^{\alpha_{ij}}$. Then both $f_k$ and $f_q$ can be trained end-to-end with the standard TD loss. When building coordination graphs, given a sparseness controlling factor $\lambda$, we select $\lambda n(n-1)$ edges with the largest $\bar{\alpha}_{ij}$ values.
\subsection{Which method is better for learning dynamically sparse coordination graphs?}\label{sec:whichone}
We test all combinations of measurements and losses described above and our method on the MACO~benchmark and use the following three metrics to quantify their performance. (1) \emph{Performance after convergence}. This value reflects the representational capacity of each candidate. (2) \emph{Temporally average performance}. Final performance alone can not depict sample efficiency. To make up for this drawback, we use a metric called \emph{temporally average performance}, which is the area between a learning curve and the $x$-axis. (3) \emph{Stability}. As described in Sec.~\ref{sec:graph}, alternating between utility/payoff function learning and graph structure learning can lead to oscillations. To measure the stability of learning processes, we use the following metric. For a learning curve ${\bm{x}}=[x_1, x_2, \dots, x_T]$, we use Kalman Filtering~\citep{musoff2009fundamentals} to get a highly smoothed curve $\hat{{\bm{x}}}=[\hat{x}_1, \hat{x}_2, \dots, \hat{x}_T]$. Then we compute the distance $d=\sqrt{\sum_{t=1}^T (x_t - \hat{x}_t)^2}$ and average $d$ across different random seeds. The averaged distance is the stability measure. In Appendix~\ref{appx:stability_measure}, we discuss the influence of different choices of stability measurement.
\begin{table*} [t]
\caption{Performance after convergence on the multi-agent coordination benchmark.}
\vspace{-0.2em}
\label{tab:maco_perf}
\centering\scriptsize
\begin{tabular}{c| P{0.25cm} P{0.25cm} c|P{0.25cm} P{0.25cm} c|P{0.25cm} P{0.25cm} c|P{0.25cm} P{0.25cm} c|P{0.25cm} P{0.25cm} c | c}
\Xhline{2\arrayrulewidth}
Topology& \multicolumn{3}{c|}{Full} & \multicolumn{3}{c|}{Rand.} & \multicolumn{3}{c|}{$\delta_{\max}$} & \multicolumn{3}{c|}{$q_{\text{var}}$} & \multicolumn{3}{c|}{$\delta_{\text{var}}$} & Attn.\\
\hline
Loss & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & --\\
\hline
Aloha & 35.7& 49.9& 50.0& 45.8& 37.1& 44.0& 51.1& 41.1& \textbf{60.7}& \textbf{52.5}& 49.4& 49.7& 41.8& 35.2& \textbf{55.1} & 25.3 \\
Pursuit & 4.74& 4.76& 4.8& 4.71& 4.72& 4.74& 4.77& \textbf{4.83}& 4.78& 4.74& \textbf{4.83}& 4.79& 4.74& 4.81& \textbf{4.89} & 4.75 \\
Hallway & \textbf{1.00}& \textbf{1.00}& \textbf{1.00}& 0.98& 0.95& 0.92& 0.97& \textbf{1.00}& 0.99& 0.96& \textbf{1.00}& 0.99& \textbf{1.00}& 0.99& 0.99 & 0.01\\
Sensor & 7.38& 0.0& 0.0& 7.51& 15.0& 7.39& 24.0& 25.2& 24.6& \textbf{26.2}& 25.8& \textbf{26.5}& 25.7& 26.0& \textbf{26.1} & 7.42\\
Gather & 0.80& 0.99& 0.99& 0.84& 0.99& 0.81& 0.99& 0.99& 0.99& 0.83& 0.99& 0.98& \textbf{1.00}& 0.99& 0.99& 0.52 \\
Disperse & 8.07& 8.02& 8.14& 8.29& 7.71& 7.97& 8.32& 8.28& 8.38& 8.50& \textbf{8.57}& \textbf{8.58}& 8.74& 8.51& 8.53& \textbf{9.45}\\
\Xhline{2\arrayrulewidth}
\end{tabular}
\vspace{-2.5em}
\end{table*}
\textbf{Performance after Convergence}\ \ Table~\ref{tab:maco_perf} shows the averaged final performance of different implementations over 8 random seeds. For each task, we highlight the top 3 scores. Generally speaking, performance of two implementations is impressive across different types of problems: (1) Build graphs using the variance of payoff functions and minimize $\mathcal{L}_{\text{sparse}}^{q_{\text{var}}}$; (2) Construct graphs according to the variance of utility difference functions and minimize $\mathcal{L}_{\text{sparse}}^{\delta_{\text{var}}}$. We denote these two implementations by $q_{\text{var}}$ \& $\mathcal{L}_{\text{sparse}}^{q_{\text{var}}}$ and $\delta_{\text{var}}$ \& $\mathcal{L}_{\text{sparse}}^{\delta_{\text{var}}}$, respectively. In contrast, using the maximum absolute value of the utility difference function has the worst performance among value-based methods. We hypothesize that this is because the maximum value, compared to variance, is more vulnerable to noise and estimation errors in utility and payoff functions.
We also observe that, most dynamically sparse coordination graphs perform much better than complete coordination graphs, demonstrating the power of the adaptability and flexibility provided by sparse graphs. One exception is task $\mathtt{Hallway}$ that is characterized by static and high-order coordination relationships among agents. Such relationships can be easier captured by dense and static graphs.
\textbf{Temporally Average Performance}\ \ We show the temporally average performance in Table~\ref{tab:maco_avg_perf}. Again, we observe a performance gap between complete coordination graphs and most implementations of context-aware sparse graphs. \citet{castellini2019representational} finds that (randomly) sparse coordination graphs perform much worse than full graphs. This is aligned with our experimental results. Importantly, we show that carefully learned sparse graphs can significantly outperform complete graphs. We expect these observations to be a good supplementation to \citet{castellini2019representational} and can eliminate any possible misunderstanding about sparse coordination graphs.
\textbf{Stability and Recommendation}\ \ Table~\ref{tab:maco_fluc} shows the stability of each implementation on all tasks across the MACO~benchmark. Generally speaking, random graphs are the most stable while sparse graphs oscillate on some tasks like $\mathtt{Aloha}$, which is in line with our analyses in Sec.~\ref{sec:graph}.
In Fig.~\ref{fig:radar}, for the two implementations with outstanding final performance, we scale each score to $[0,1]$ on each task and show the averaged scaled scores over tasks. We can see that $q_{\text{var}}$ \& $\mathcal{L}_{\text{sparse}}^{q_{\text{var}}}$ has the best temporally average performance and stability. We thus recommend it for learning dynamically sparse graphs.
\begin{wrapfigure}[16]{r}{0.3\linewidth}
\centering
\includegraphics[width=\linewidth]{fig-maco/maco.png}
\caption{Normalized scores of two implementations with outstanding \emph{performance after convergence}.}
\label{fig:radar}
\end{wrapfigure}
Table~\ref{tab:maco_perf} shows the \emph{performance after convergence} of different implementations on the MACO~benchmark. This measurement reflects the representational capacity of different methods, but conveys less information about their sample efficiency and stability. To this end, we propose evaluations \emph{temporally average performance} and \emph{stability}. Table~\ref{tab:maco_avg_perf} and~\ref{tab:maco_fluc} show the scores of all the considered implementations evaluated by these two measurements.
The three measurements we use are just some statistics of the learning curves. We further show the complete learning curves of value-based implementations in Fig.~\ref{fig:aloha-lc}-\ref{fig:disperse-lc} and compare the best value-based method ($\delta_{\text{var}}$ \& $\mathcal{L}_{\text{sparse}}^{\delta_{\text{var}}}$) against the (attentional) observation-based method in Fig.~\ref{fig:attn-lc}. We can see that value-based methods generally perform better than the observation-based method, except for the task $\mathtt{Disperse}$. Compared to other games, observations provided by $\mathtt{Disperse}$ can reveal all the game information. In this case, the observation-based method can make better use of local observations and can easily learn an accurate coordination graph.
\begin{figure*}[ht]
\centering
\includegraphics[width=\linewidth]{fig-maco/attention.pdf}
\caption{Performance comparison between the best value-based method and the (attentional) observation-based approach on the MACO~benchmark.}
\label{fig:attn-lc}
\end{figure*}
\section{Mathematical Proof}
In this section, we provide proof to Proposition~\ref{prop}.
Without loss of generality, we consider two agents $1$ and $2$ and the edge between them, $(1, 2)$.
We prove our idea by comparing the action selection of agent $2$ before and after removing edge $(1, 2)$. In the following proof, we use $i$ to denote agent $i$ and $e$ to denote edge $(1, 2)$.
Action of agent $2$ is determined by
\begin{align}
a_2^* &= \arg\max_{a_2} \sum_{h\in \mathcal{F}_2} m_{h\rightarrow 2}(a_2) \\
& = \arg\max_{a_2} \left[m_{e\rightarrow 2}(a_2) + \sum_{h\in \mathcal{F}_2 / \{e\}} m_{h\rightarrow 2}(a_2)\right],\label{equ:1}
\end{align}
and we first see the influence of $m_{e\rightarrow 2}(a_2)$ on $a_2^*$. For clarity, we use $m(a_2)$ to denote $m_{e\rightarrow 2}(a_2)$ and $l(a_2)$ to denote $\sum_{h\in \mathcal{F}_2 / \{e\}} m_{h\rightarrow 2}(a_2)$. We are interested in whether $\arg\max_{a_2} l(a_2) = \arg\max_{a_2} \left[m(a_2)+l(a_2)\right]$. Let $\text{Range}({\bm{x}})$ denote the largest values of vector ${\bm{x}}$ minus the smallest value one and $a_2^j = \arg\max l(a_k)$ denote the maximum value of $a_k$.
The probability of this event holds if the following inequality holds:
\begin{equation}
\text{Range}\left[m(a_2)\right] \le \min_{a_2\neq a_2^j}(l(a_2^j)-l(a_2)).
\end{equation}
We use $\bar{m}$ to denote the average of $m(a_k)$. We rewrite this condition and obtain
\begin{align}
&Pr\left(\text{Range}\left[m(a_2)\right] \le \min_{a_2\neq a_2^j}(l(a_2^j)-l(a_2))\right) \\
=& Pr\left(\min_{a_2}m(a_2) < m(a_2) < \min_{a_2}m(a_2) + \min_{a_2\neq a_2^j}(l(a_2^j)-l(a_2))\right)
\end{align}
According to the Asymmetric two-sided Chebyshev's inequality, we get a lower bound of this probability:
\begin{equation}
4\frac{(\bar{m}-\min_{a_2}m(a_2))(\max_{a_2}m(a_2)-\bar{m})-\sigma^2}{\left[\min_{a_2\neq a_2^j}(l(a_2^j)-l(a_2))\right]^2}
\end{equation}
where $\sigma$ is the variance of $m(a_2)$, and $\bar m$ is the average of $m(a_2)$.
Suppose that we take $n$ actions independently. According to the von Szokefalvi Nagy inequality, we can further get the lower bound as follows:
\begin{equation}
\begin{aligned}
4\frac{(\bar{m}-\min_{a_2}m(a_2))(\max_{a_2}m(a_2)-\bar{m})-\sigma^2}{\left[\min_{a_2\neq a_2^j}(l(a_2^j)-l(a_2))\right]^2} &\geq 4\frac{(\bar{m}-\min_{a_2}m(a_2))(\max_{a_2}m(a_2)-\bar{m})-\sigma^2}{2n\sigma^2} \\
&= \frac{2}{n}\left[\frac{(\bar{m}-\min_{a_2}m(a_2))(\max_{a_2}m(a_2)-\bar{m})}{\sigma^2} - 1\right] \label{equ:lower1}
\end{aligned}
\end{equation}
Note that
\begin{align}
m_{e\rightarrow 2}(a_2) = \max_{a_1} \left[q(a_1, a_2) + m_{1\rightarrow e}(a_1) \right]
\end{align}
and we are interested in $q(a_1, a_2)$. We now study the relationship between $m_{e\rightarrow 2}(a_2)$ and $\max_{a_1} \left[q(a_1, a_2) \right]$. For clarity, we use $r(a_1,a_2)$ to denote $q(a_1,a_2) + m_{1\rightarrow e}(a_1)$, and $r(a_1^{i_2}, a_k)$ to denote $\max_{a_1} r(a_1, a_2)$. Then we have
\begin{equation}
\text{Var}_{a_2} \max_{a_1} r(a_1, a_2) = \text{Var}_{a_2} r(a_1^{i_2}, a_2)
\end{equation}
Since $i_2=\arg\max_{i} r(a_1,a_2)$ for a given $a_2$, we have
\begin{equation}
\text{Var}_{a_2} r(a_1,a_2) = \text{Var}_{a_2} \left[ r(a^{i_2}_1,a_2) - s_2 \right],
\end{equation}
where $s_2 > 0$, for a given $a_1$.
Since
\begin{align}
& \text{Var}_{a_2} \left[ r(a^{i_2}_1,a_2) - s_2 \right] \\
=& \text{Var}_{a_2} \left[ r(a^{i_2}_1,a_2)\right] + \text{Var} \left[ s_2\right] - 2\text{Cov}( r(a^{i_2}_1,a_2), s_2)
\end{align}
and
\begin{align}
\text{Cov}( r(a^{i_2}_1,a_2), s_2) \le \sqrt{\text{Var}_{a_2} \left[ r(a^{i_2}_1,a_2)\right]\text{Var} \left[s_2\right]},
\end{align}
it follows that
\begin{align}
& \text{Var}_{a_2} \left[ r(a^{i_2}_1,a_2) - s_2 \right] \\
\ge& \text{Var}_{a_2} \left[ r(a^{i_2}_1,a_2)\right] - 2\sqrt{\text{Var}_{a_2} \left[ r(a^{i_2}_1,a_2)\right]\text{Var} \left[s_2\right]}
\end{align}
Thus,
\begin{align}
\zeta_{12} \left[r(a_1,a_2)\right] = \max_{a_1}\text{Var}_{a_2} \left[r(a_1,a_2)\right] \ge \text{Var}_{a_2}\max_{a_1}\left[r(a_1,a_2)\right] - 2\sqrt{\text{Var}_{a_2} \left[ r(a^{i_2}_1,a_2)\right]\text{Var} \left[ s_2\right]}.
\end{align}
Observing that $\zeta_{12} \left[r(a_1,a_2)\right] = \max_{a_1}\text{Var}_{a_2} \left[r(a_1,a_2)\right] = \max_{a_1}\text{Var}_{a_2} \left[q(a_1,a_2)\right] = \zeta_{12} \left[q(a_1,a_2)\right]$, we have
\begin{equation}
\begin{aligned}
\sigma &\le \zeta_{12} \left[q(a_1,a_2)\right] + 2\sqrt{\text{Var}_{a_2} \left[ r(a^{i_2}_1,a_2)\right]\text{Var} \left[ s_2\right]}\\
&= \zeta_{12} \left[q(a_1,a_2)\right] + 2\sqrt{\sigma S}\\
\end{aligned}
\end{equation}
where $\sigma=\text{Var}_{a_2}\max_{a_1}\left[r(a_1,a_2)\right]$ and $\text{Var} \left[ s_2\right] = S$. According to the fixed-point theorem, the term $\sigma$ satisfies $\zeta_{12} \left[q(a_1,a_2)\right] + 2\sqrt{\sigma S} = \sigma$. We can solve this quadratic form and get $\sigma = \zeta_{12} \left[q(a_1,a_2)\right] + 2S \pm 2\sqrt{S\left(S + \zeta_{12} \left[q(a_1,a_2)\right]\right)}$. Because the $\sigma$ term is larger than $\zeta_{12} \left[q(a_1,a_2)\right] + 2S$, we get $\sigma = \zeta_{12} \left[q(a_1,a_2)\right] + 2S + 2\sqrt{S\left(S + \zeta_{12} \left[q(a_1,a_2)\right]\right)}$.
By inserting this inequality to the lower bound (Eq.~\ref{equ:lower1}), we get a lower bound related to $q(a_1,a_2)$:
\begin{equation}
\frac{2}{n}\left[\frac{(\bar{m}-\min_{a_2}m(a_2))(\max_{a_2}m(a_2)-\bar{m})}{\left[\zeta_{12} \left[q(a_1,a_2)\right] + 2S + 2\sqrt{S\left(S + \zeta_{12} \left[q(a_1,a_2)\right]\right)}\right]^2} - 1\right]
\end{equation}
if the vector ${\bm{x}}$ is larger than 0, we have: $\text{Var}(x) = \frac{1}{n} \sum_i (x_i - \bar{m}) \leq \frac{1}{n} \sum max_i x^2 \leq max_i x^2$. Thus we can further get the following bound:
\begin{equation}
\begin{aligned}
S &= \text{Var}_{a_2}\left[ \max_{a_1} r\left(a_1, a_2\right) - r\left(a_1, a_2\right) \right] \\
& \leq max_{a_2}^2 \left[ \max_{a_1} r\left(a_1, a_2\right) - r\left(a_1, a_2\right) \right] \\
\end{aligned}
\end{equation}
Let $A = \max_{a_2}\left[\max_{a_1} r\left(a_1, a_2\right) - r\left(a_1, a_2\right)\right]$, we have $S \leq A^2$. We can get the final lower bound:
\begin{equation}
\frac{2}{n}\left[\frac{(\bar{m}-\min_{a_2}m(a_2))(\max_{a_2}m(a_2)-\bar{m})}{\left[\zeta_{12} \left[q(a_1,a_2)\right] + 2A^2 + 2\sqrt{A^2\left(A^2 + \zeta_{12} \left[q(a_1,a_2)\right]\right)}\right]^2} - 1\right]
\end{equation}
\section{Stability Measurement}\label{appx:stability_measure}
For evaluating the stability of different implementations, we use the following stability measurement in our paper. For a learning curve ${\bm{x}}=[x_1, x_2, \dots, x_T]$, we use Kalman Filtering~\citep{musoff2009fundamentals} to get a highly smoothed curve $\hat{{\bm{x}}}=[\hat{x}_1, \hat{x}_2, \dots, \hat{x}_T]$. Then we compute the distance $d=\sqrt{\sum_{t=1}^T (x_t - \hat{x}_t)^2}$ and average $d$ across learning curves obtained using different random seeds. The averaged distance is the stability measure. However, different stability measurements may lead to different conclusions. In this section, we discuss the influence of the choice of stability measurements.
We propose some other stability measurements and compare the induced rankings. Specifically, we test the following methods for quantifying the volatility of a learning curve ${\bm{x}}=[x_1, x_2, \dots, x_T]$. We use (1) Exponential Moving Average (EMA) introduced by \citet{klinker2011exponential} and (2) MidPoint over a period by calculating the average of maximum and minimum within a rolling window to get the smoothed curve, and calculate the distance between the original and smoothed curve. Moreover, we use (3) Double Exponential Moving Average (DEMA) introduced by \citet{stanley1988digital} to smooth the curve and then calculate the distance similarly. In Table~\ref{tab:maco_fluc_measure}, we highlight the implementations that are the most stable on $\mathtt{Pursuit}$ as measured by each method. It can be observed that the results are similar. We thus conclude that the stability ranking reported in the paper is robust over different measurements.
\iffalse
\begin{table*} [t]
\caption{Influence of stability measurements. On $\mathtt{Aloha}$, we compare the stability rankings of different implementations as measured by 4 different measurements. For each measurement, we highlight the three stablest observation- or value-based methods.}
\label{tab:maco_fluc_measure}
\centering
\begin{tabular}{c| P{0.455cm} P{0.455cm} c|P{0.455cm} P{0.455cm} c|P{0.455cm} P{0.455cm} c|P{0.455cm} P{0.455cm} c|P{0.455cm} P{0.455cm} c | c}
\Xhline{2\arrayrulewidth}
Topology& \multicolumn{3}{c|}{Full} & \multicolumn{3}{c|}{Rand.} & \multicolumn{3}{c|}{$\delta_{\max}$} & \multicolumn{3}{c|}{$q_{\text{var}}$} & \multicolumn{3}{c|}{$\delta_{\text{var}}$} & Attn. \\
\hline
Loss & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & -- \\
\hline
Kalman & 0.67& 0.62& 2.54& 0.29& 0.58& 0.39& 0.63& 1.34& 2.28& 1.29& 1.25& \textbf{0.60}& \textbf{0.59}& 0.69& \textbf{0.62}& 2.04 \\
EMA & 0.98& 0.89& 3.20& 0.58& 0.96& 0.65& 0.89& 1.48& 2.27& 1.53& 1.93& \textbf{0.85}& \textbf{0.70}& 1.17& \textbf{0.79}& 2.49 \\
MidPoint & 0.81& 0.88& 3.17& 0.43& 0.82& 0.54& 0.86& 1.51& 2.36& 1.40& 1.62& \textbf{0.75}& \textbf{0.57}& 0.92& \textbf{0.70}& 2.39\\
PolyFit & 0.72& 0.41& 2.70& 0.29& 0.60& 0.36& 0.70& 1.65& 2.39& 1.04& 2.21& \textbf{0.57}& \textbf{0.48}& 0.77& \textbf{0.68}& 1.98\\
\Xhline{2\arrayrulewidth}
\end{tabular}
\end{table*}
\fi
\begin{table*} [h!]
\caption{Influence of stability measurements. On $\mathtt{Pursuit}$, we compare the stability rankings of different implementations as measured by 4 different measurements. For each measurement, we highlight the three stablest observation- or value-based methods. The values are multiplied by $10$.}
\label{tab:maco_fluc_measure}
\centering\scriptsize
\begin{tabular}{c| P{0.25cm} P{0.25cm} c|P{0.25cm} P{0.25cm} c|P{0.25cm} P{0.25cm} c|P{0.25cm} P{0.25cm} c|P{0.25cm} P{0.25cm} c | c}
\Xhline{2\arrayrulewidth}
Topology& \multicolumn{3}{c|}{Full} & \multicolumn{3}{c|}{Rand.} & \multicolumn{3}{c|}{$\delta_{\max}$} & \multicolumn{3}{c|}{$q_{\text{var}}$} & \multicolumn{3}{c|}{$\delta_{\text{var}}$} & Attn. \\
\hline
Loss & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & $|\delta|$ & $q_{\text{var}}$ & $\delta_{\text{var}}$ & -- \\
\hline
Kalman & 3.55& 3.11& 4.07& 2.32& 2.08& 2.49& 2.04& 1.96& 2.58& 1.99& \textbf{1.28}& 2.07& 1.86& \textbf{1.82}& \textbf{1.77}& 1.87 \\
EMA & 5.20& 4.87& 5.27& 3.59& 3.46& 3.76& 3.42& 3.70& 4.06& 3.48& \textbf{2.46}& 3.78& 3.51& \textbf{3.33}& \textbf{3.04}& 3.55 \\
MidPoint & 4.49& 4.16& 4.46& 3.04& 2.81& 3.30& \textbf{2.69}& 2.95& 3.37& 2.77& \textbf{1.91}& 2.96& 2.73& 2.72& \textbf{2.40}& 2.55\\
DEMA & 2.08& 1.90& 2.20& 1.43& 1.33& 1.50& 1.29& 1.35& 1.59& 1.32& \textbf{0.92}& 1.39& 1.30& 1.22& \textbf{1.19}& \textbf{1.20}\\
\Xhline{2\arrayrulewidth}
\end{tabular}
\end{table*}
\section{The SMAC benchmark}
On the SMAC benchmark, we compare our method against fully decomposed value function learning methods (VDN~\citep{sunehag2018value} \& QMIX~\citep{rashid2018qmix}) and a deep coordination graph learning method (DCG~\citep{bohmer2020deep}). Experiments are carried out on a hard map $\mathtt{5m\_vs\_6m}$ and a super hard map $\mathtt{MMM2}$. For the baselines, we use the code provided by the authors and their default hyper-parameters settings that have been fine-tuned on the SMAC benchmark. We also notice that both our method and all the considered baselines are implemented based on the open-sourced codebase PyMARL\footnote{https://github.com/oxwhirl/pymarl}, which further guarantees the fairness of the comparisons.
\section{Action Representation Learning}\label{appx:action_repr}
As discussed in Sec. 5.1 of the main text, we use action representations to reduce the influence of utility difference function's estimation errors on graph structure learning. In this section, we describe the details of action representation learning (the related network structure is shown in Fig.~\ref{fig:action_repr_learner}). We
\begin{wrapfigure}[18]{r}{0.35\linewidth}
\vspace{-0.5em}
\includegraphics[width=\linewidth]{fig-demo/action_repr_learner.pdf}
\caption{Framework for learning action representations, reproduced from~\citet{wang2021rode}.}
\label{fig:action_repr_learner}
\end{wrapfigure}
use the technique proposed by~\citet{wang2021rode} and learn an action encoder $f_e(\cdot; \theta_e)$: $\mathbb{R}^{|A|}\rightarrow\mathbb{R}^d$, parameterized by $\theta_e$, to map one-hot actions to a $d$-dimensional representation space. With the encoder, each action $a$ has a latent representation ${\bm{z}}_{a}$, \textit{i}.\textit{e}., ${\bm{z}}_{a}=f_e(a; \theta_e)$. The representation ${\bm{z}}_{a}$ is then used to predict the next observation $o_i'$ and the global reward $r$, given the current observation $o_i$ of an agent $i$, and the one-hot actions of other agents, ${\bm{a}}_{\textup{\texttt{-}} i}$. This model is a forward model, which is trained by minimizing the following loss:
\begin{equation}\label{equ:ar_learning}
\begin{aligned}
\mathcal{L}_e(\theta_e, \xi_e) & = \mathbb{E}_{({\bm{o}}, {\bm{a}}, r, {\bm{o}}')\sim \mathcal{D}}\big[\sum_i\|p_o({\bm{z}}_{a_i},o_i, {\bm{a}}_{\textup{\texttt{-}} i}) - o_i'\|^2_2 \\
& + \lambda_e\sum_i\left(p_r({\bm{z}}_{a_i},o_i, {\bm{a}}_{\textup{\texttt{-}} i}) - r\right)^2\big],
\end{aligned}
\end{equation}
where $p_o$ and $p_r$ is the predictor for observations and rewards, respectively. We use $\xi_e$ to denote the parameters of $p_o$ and $p_r$. $\lambda_e$ is a scaling factor, $\mathcal{D}$ is a replay buffer, and the sum is carried out over all agents.
In the beginning, we collect samples and train the predictive model shown in Fig.~\ref{fig:action_repr_learner} for 50$K$ timesteps. Then policy learning begins and action representations are kept fixed during training. Since tasks in the MACO~benchmark typically do not involve many actions, we do not use action representations when benchmarking our method. In contrast, StarCraft II micromanagement tasks usually have a large action space. For example, the map $\mathtt{MMM2}$ involves $16$ actions, and a conventional deep Q-network requires $256$ output heads for learning utility difference. Therefore, we equip our method with action representations to estimate the utility difference function when testing it on the SMAC benchmark.
\section{Architecture, Hyperparameters, and Infrastructure}\label{appx:hyper}
\subsection{CASEC}
In CASEC, each agent has a neural network to estimate its local utility. The local utility network consists of three layers---a fully-connected layer, a 64 bit GRU, and another fully-connected layer---and outputs an estimated utility for each action. The utility difference function is also a 3-layer network, with the first two layers shared with the local utility function to process local action-observation history. The input to the third layer (a fully-connected layer) is the concatenation of the output of two agents' GRU layer. The local utilities and pairwise utility differences are summed to estimate the global action value (Eq. 11 in the paper).
For all experiments, the optimization is conducted using RMSprop with a learning rate of $5\times10^{-4}$, $\alpha$ of 0.99, RMSProp epsilon of 0.00001, and with no momentum or weight decay. For exploration, we use $\epsilon$-greedy with $\epsilon$ annealed linearly from 1.0 to 0.05 over 50$K$ time steps and kept constant for the rest of the training. Batches of 32 episodes are sampled from the replay buffer. The default iteration number of the Max-Sum algorithm is set to 5. The communication threshold depends on the number of agents and the task, and we set it to $0.3$ on the map $\mathtt{5m\_vs\_6m}$ and $\mathtt{MMM2}$. We test the performance with different values ($1e\textup{\texttt{-}} 3$, $1e\textup{\texttt{-}} 4$, and $1e\textup{\texttt{-}} 5$) of the scaling weight of the sparseness loss $\mathcal{L}_{\text{sparse}}^{\delta_{\text{var}}}$ on $\mathtt{Pursuit}$, and set it to $1e\textup{\texttt{-}} 4$ for both the MACO and SMAC benchmark. The whole framework is trained end-to-end on fully unrolled episodes. All experiments on StarCraft II use the default reward and observation settings of the SMAC benchmark.
All the experiments are carried out on NVIDIA Tesla P100 GPU. We show the estimated running time of our method on different tasks in Table~\ref{tab:time-maco} and~\ref{tab:time-smac}. Typically, CASEC~can finish 1M training steps within 8 hours on MACO~tasks and in about 10 hours on SMAC tasks. In Table~\ref{tab:time-comparision}, we compare the computational complexity of action selection for CASEC and DCG, which is the bottleneck of both algorithms. CASEC is slightly faster than DCG by virtue of graph sparsity.
\begin{table*}[h!]
\caption{Approximate running time of CASEC~on tasks from the MACO~benchmark.}
\centering
\begin{tabular}{CRCRCRCRCRCR}
\toprule
\multicolumn{2}{c}{Aloha} &
\multicolumn{2}{l}{Pursuit} &
\multicolumn{2}{l}{Hallway} &
\multicolumn{2}{l}{Sensor} &
\multicolumn{2}{l}{Gather} &
\multicolumn{2}{l}{Disperse} \\
\cmidrule(lr){1-2}
\cmidrule(lr){3-4}
\cmidrule(lr){5-6}
\cmidrule(lr){7-8}
\cmidrule(lr){9-10}
\cmidrule(lr){11-12}
\multicolumn{2}{c}{13h (2M)} & \multicolumn{2}{c}{17h (2M)} & \multicolumn{2}{c}{7h (1M)} & \multicolumn{2}{c}{4.5h (0.5M)} & \multicolumn{2}{l}{6.5h (1M)} & \multicolumn{2}{c}{8h (1M)}\\
\toprule
\end{tabular}
\label{tab:time-maco}
\caption{Approximate running time of CASEC~on tasks from the SMAC benchmark.}
\centering
\begin{tabular}{CRCRCRCRCR}
\toprule
\multicolumn{2}{l}{5m\_vs\_6m} &
\multicolumn{2}{l}{MMM2}\\
\cmidrule(lr){1-2}
\cmidrule(lr){3-4}
\multicolumn{2}{c}{18h (2M)} & \multicolumn{2}{c}{21h (2M)} \\
\toprule
\end{tabular}
\label{tab:time-smac}
\end{table*}
\begin{table}[h!]
\centering
\caption{Average time (milliseconds) for 1000 action selection phases of CASEC/DCG. CASEC uses a graph with sparseness 0.2 while DCG uses the full graph. To ensure a fair comparison, both Max-Sum/Max-Plus algorithms pass messages for 8 iterations. The batch size is set to 10.}
\vspace{0.5em}
\begin{tabular}{|c|c|c|c|}
\hline
& 5 actions & 10 actions & 15 actions \\ \hline
5 agents & 2.90/3.11 & 3.15/3.39 & 3.42/3.67\\ \hline
10 agents & 3.17/3.45 & 3.82/4.20 & 5.05/5.27 \\ \hline
15 agents & 3.41/3.67 & 5.14/5.4 & 7.75/8.02 \\ \hline
\end{tabular}
\label{tab:time-comparision}
\end{table}
|
\section{Introduction}
\label{intro}
Cluster analysis seeks to partition data into distinct subsets, or clusters,
such that observations in the same cluster are more similar than observations
from different clusters. There are numerous methods for cluster analysis with
applications in many fields \citep{dubes1976, jain1999}. As an unsupervised
learning technique, there is little consensus on how to validate the clustering
obtained. With many available clustering techniques, subjective choices must be
made which influence the outcome of a clustering procedure
\citep{islr}. It is also difficult to quantify the uncertainty associated with
clustering estimates.
Among the various algorithms for cluster analysis, agglomerative hierarchical
clustering remains one of the most used \citep{hca}. Agglomerative hierarchical
clustering is a heuristic method that takes as its input a matrix of pairwise
distances among items. Beginning with every item in its own cluster, the algorithm
sequentially merges the most similar clusters (based on the pairwise distances)
until all observations are in a single cluster. These nested clusters can be
visualized as a dendrogram. There are subjective decisions involved in
hierarchical clustering; namely, the linkage type used to build the tree
and the tree-cutting method used to obtain a clustering estimate.
Another commonly used class of clustering methods includes k-means and k-medoids,
which both seek to minimize the within-cluster distances from each point to a
point specified as its cluster center. k-means minimizes squared Euclidean
distances to each cluster centroid. In contrast, k-medoids uses an actual
data point for each cluster center and thus can take any dissimilarity
measure as input \citep{kaufman1990}. The same input of a pairwise distance
matrix makes k-medoids a comparable procedure to hierarchical clustering. The most
influential choice a user must make in k-medoids clustering is the value of
$k$, that is, how many clusters the estimate should contain.
We propose CaviarPD: Cluster Analysis via Random Partition Distributions. Like
hierarchical clustering and k-medoids, CaviarPD is based on pairwise distances, yet
it provides a unique way to assess clustering uncertainty.
The CaviarPD method relies on sampling from a random partition
distribution that is based on pairwise distances, e.g., the Ewens-Pitman
Attraction (EPA) distribution \citep{dahl2017}. Thus, like the other two methods,
the EPA distribution uses pairwise distances as input. Unlike hierarchical and
k-medoids clustering, this form of cluster analysis allows us to make probability
statements about clustering relationships, thereby quantifying the uncertainty of the
estimate. In doing so, CaviarPD provides an alternate way to visualize clusterings by
showing the pairwise probabilities that items are clustered together. An
implementation of the method is provided in the \texttt{caviarpd} package \citep{caviarpdlib} in
\texttt{R}, which is available on CRAN.
To compare hierarchical and k-medoids clustering with our proposed CaviarPD
method, we evaluate how well each method performs in eight different case
studies where the true partition of the data is known. We compare methods
using two partition loss functions, namely, Binder loss \citep{binder1978} and
VI loss \citep{meila2007,wade2018}. Through the case studies, we show the advantages of
CaviarPD over hierarchical and k-medoids clustering. All methods
tend to estimate the true partition of the data well; however, the hierarchical
clustering results are highly varied between linkages and choices of cutting the
tree. There is little statistical reasoning to guide the choice of linkage and
tree cutting. Furthermore, because CaviarPD gives the probabilities that
items are clustered together, CaviarPD provides additional
information about the clustering relationships beyond what
hierarchical or k-medoids clustering provide.
\section{Existing Distance-Based Clustering Methods}
\subsection{Clustering Concepts and Terminology}
We introduce common concepts used in both traditional clustering methods and
CaviarPD.
For a more thorough description of current clustering practices, we
suggest \citet{hca}. A clustering $\mathbfit{c} = (c_1, ..., c_n)$ gives labels for $n$
items in which items $i$ and $j$ are in the same cluster if and only if $c_i =
c_j$. Equivalently, a partition $\pi = \{ S_1, \ldots, S_q \}$ of integers $1,
\ldots, n$ is composed of mutually exclusive, non-empty, and exhaustive subsets
such that $i, j \in S$ implies that $c_i = c_j$. We use the terms `clustering'
and `partition' interchangeably and note that the term `cluster' is synonymous
with `subset'.
In cluster analysis, we seek to cluster $n$ observations of a dataset $\mathcal{D}
= \{\mathbf{x}_1 ,\dots, \mathbf{x}_n\}$ into distinct groups so that
observations within a group are more similar than observations from different
groups. CaviarPD, like the other two methods described, relies on distance
information between observations in order to partition items into subsets.
The pairwise distances between observations $\mathbf{x}_i$ and $\mathbf{x}_j$
are calculated from a specified distance function $d(\mathbf{x}_i,\mathbf{x}_j)$.
One common distance function is Euclidean distance:
$d(\mathbf{x}_i,\mathbf{x}_j)=\sqrt{(\mathbf{x}_i -
\mathbf{x}_j)'(\mathbf{x}_i-\mathbf{x}_j)}$. The pairwise distances between
all items can be stored in an $n \times n$ distance matrix $\mathbfit{D}$, where
$d_{ij}=d(\mathbf{x}_i,\mathbf{x}_j)$. The choice of distance metric is
also an important consideration in the analysis, but our task here is to
compare distance-based clustering methods given the user's chosen distance
matrix.
\subsection{Hierarchical Clustering}
In this section we highlight the decisions a user must make with hierarchical
clustering. For a thorough introduction to hierarchical clustering, see
\citet{mma} or \citet{islr}.
Recall that in agglomerative hierarchical clustering, each observation begins
in its own cluster and the most similar clusters are sequentially merged until
all data points are in a single cluster. The criteria used to define similarity
between clusters is called the linkage and is computed from the pairwise
distance between items in each cluster. Hierarchical clustering requires that
a user decide which linkage to use and how to cut the dendrogram to obtain a
partition estimate. For a more detailed explanation of the
subset distance computations for each linkage type, see \citet{nielsen2016}.
\begin{figure}[t]
\centering
\includegraphics[width=0.49\textwidth]{figures/avg.pdf}
\hfill
\includegraphics[width=0.49\textwidth]{figures/com.pdf}\\
\includegraphics[width=0.49\textwidth]{figures/ward.pdf}
\hfill
\includegraphics[width=0.49\textwidth]{figures/sing.pdf}
\caption{Clustering dendrograms for the wine dataset using complete, average, Ward and single linkage.}
\label{fig:wine-linkage}
\end{figure}
We demonstrate how the choice of linkage leads to highly varied clustering
estimates. Figure~\ref{fig:wine-linkage} shows the resulting dendrograms
from four common linkages applied to the wine recognition dataset using the
function \texttt{hclust} in \textsf{R} \citep{r2018}. The wine dataset
contains many different chemical attributes for samples of wine from three
different cultivars. The distance matrix was computed with Euclidean distance
using all 13 chemical attributes. The clustering structure represented in the
dendrograms is very different for all four linkages. Single linkage produces
long chains of connected clusters. Ward linkage seeks to create compact
spherical clusters such that the dendrogram shows more distinct clusters
\citep{ward1963}. For the wine dataset, the dendrograms for average and complete
linkage show clustering structures between the long chains of single linkage and compact clusters of Ward linkage.
After choosing the type of linkage, the structure of the dendrogram is
used to cut the tree to obtain a clustering estimate. From the dendrogram
using Ward linkage, there appear to be three main clusters. Inspection of
the other dendrograms leads to less definite conclusions about the number
of clusters in the dataset as each tree varies drastically. Using complete
linkage, one could reasonably argue for a cut of the tree that gives
anywhere from 2 to 7 clusters. \citet{langfelder2008} introduce a solution
to the tree cutting problem with the dynamic tree cut (DTC) algorithm.
This procedure takes into account the structure of the tree to
detect clusters and also allows for the tree to be cut at variable heights,
providing greater flexibility in clustering estimation.
The lack of agreement between the linkages and tree cutting estimates is
prominent and concerning. There appeared to be three main clusters for ward
linkage, yet the DTC produced a default estimate of 5 clusters. The DTC applied
to the complete linkage dendrogram resulted in an estimate with 3 clusters.
Likewise, DTC applied to the average linkage dendrogram also resulted in 3 main
clusters with a single observation in a fourth cluster by itself. Lastly, the
single linkage tree resulted in a clustering estimate with all but 3
observations in a single cluster.
\subsection{K-Medoids Clustering}
k-means clustering remains one of the most common and popular clustering approaches
used today. The closely related k-medoids clustering, like k-means, seeks to
minimize the sum of distances between points in a cluster and its cluster center.
Unlike k-means, k-medoids uses actual data points (exemplars or medoids) for the
cluster centers. Thus, k-medoids only uses the pairwise distance matrix as input,
making it a comparable clustering approach to hierarchical clustering and CaviarPD.
Since there are $\binom{n}{k}$ combinations of medoids that can be tested to
minimize the total distance, finding the exact solution to the equation is
difficult. \citet{kaufman1990} proposed the PAM algorithm (partitioning
around medoids) to conduct a nonexhaustive but greedy search through the
combination space. The PAM algorithm was then modified by
\citet{schubert2019} to reduce computation time even further at the expense
of some accuracy. Since the datasets in our case studies are all relatively
small, we can easily use the original PAM algorithm for our k-medoids comparison.
In our case studies we use the silhouette method, a straightforward procedure
for selecting $k$ in k-means and k-medoids clustering \citep{rous1987}. The
silhouette method is most reliable when the choices for $k$ are narrowed down
to a reasonable range using some prior intuition about the data. As with
hierarchical clustering, the k-medoids approach offers no way to assess
uncertainty in the results. We use the \texttt{pam} function in the
\texttt{cluster} package \citep{cluster} since it has a built-in calculation for silhouette
width.
\section{Cluster Analysis Via Random Partition Distributions}
\label{caviarpd}
We propose the CaviarPD method, a novel approach to clustering based on random
partition distributions. A reference implementation is available on CRAN in the
\texttt{caviarpd} package \citep{caviarpdlib}.
CaviarPD is based on the Ewens-Pitman Attraction (EPA) distribution,
originally proposed as a prior distribution for partitions in a Bayesian
framework \citep{dahl2017}. In the proceeding subsections, we explain how the
EPA distribution can be used directly for cluster analysis along with the additional
components of CaviarPD. Since the EPA distribution is a probability distribution over
partitions, our simulation-based method provides as output probabilities of pairs of items being
clustered together.
\subsection{Ewens-Pitman Attraction Distribution}
In the EPA distribution, observations are sequentially allocated to subsets of a
partition with probability proportional to the attraction of the item to that
subset. We use $\boldsymbol{\sigma} = (\sigma_1, ..., \sigma_n)$ to denote the order of the
$n$ observations that are sequentially allocated into subsets in order to form a
partition. Thus $\sigma_t$ is the $t$th observation allocated in the partition.
This sequential allocation of items yields a sequence of partitions for each
step of the allocation. Let $\pi(\sigma_1,...,\sigma_{t-1})$ denote the
partition of the first $t-1$ observations at time $t-1$. At each time $t$, let
$K_t$ be the number of subsets in the partition $\pi(\sigma_1,...,\sigma_{t})$.
When $t = 1$, the first item is allocated to a new subset by itself with
probability 1 as shown in equation \ref{epa2}. For $t > 1$, $\sigma_t$ can
either be allocated to one of the existing $K_{t-1}$ subsets or it can be
allocated to a new subset in the partition. The probability mass function is
conveniently expressed as the product of increasingly conditional probabilities:
\begin{equation}
\label{epa1}
p(\pi_n|\alpha,\lambda,\boldsymbol{\sigma}) =
\prod_{t=1}^{n} p_t(\alpha,\lambda,\pi(\sigma_1,...,\sigma_{t-1}))
\end{equation}
\begin{multline}
\label{epa2}
p_t(\alpha,\lambda,\pi(\sigma_1,...,\sigma_{t-1})) =
\mathrm{Pr}(\sigma_t \in A \,| \, \alpha, \lambda,
\pi(\sigma_1,...,\sigma_{t-1})) \\ =
\begin{cases}
\frac{t - 1}{\alpha+t-1}\cdot \frac{\sum_{\sigma_s \in
A}\lambda(\mathbf{x}_{\sigma_t},\mathbf{x}_{\sigma_s})}{\sum_{s=1}^{t-1}\lambda(\mathbf{x}_{\sigma_t},\mathbf{x}_{\sigma_s})}
&\text{for} \; S \in \pi(\sigma_1,...,\sigma_{t-1}) \\
\frac{\alpha}{\alpha + t -1} &\text{for} \; S \;
\text{being a new subset}
\end{cases}
\end{multline}
The probability that $\sigma_t$ is allocated into subset $S$ is a function of
the similarity function $\lambda$ and the mass parameter $\alpha$.\footnote{
\citeauthor{dahl2017} also include a discount parameter
$\delta$ that further influences sampling from the EPA distribution;
For simplicity, we set $\delta=0$, yielding (\ref{epa2}).}
The similarity function $\lambda(\mathbf{x}_i,\mathbf{x}_j)$ gives
pairwise similarity between observations $\mathbf{x}_i$ and $\mathbf{x}_j$ for any
$\mathbf{x}_i,\mathbf{x}_j \in \mathcal{D}$. \citet{dahl2017} propose a general
class of similarity functions $\lambda(\mathbf{x}_i,\mathbf{x}_j)=f(d_{ij})$,
where $f$ is a non-increasing function of the pairwise distance $d_{ij}$ between
observations $\mathbf{x}_i$ and $\mathbf{x}_j$. $d_{ij}$ is the same distance
matrix $\mathbfit{D}$ used as input for hierarchical and k-medoids clustering. Two common
similarity functions are reciprocal similarity $f(d) = d^{-\tau}$ and
exponential similarity $f(d) = \mathrm{exp}(-\tau d)$. The parameter $\tau \ge
0$ is the temperature, which has the effect of accentuating or dampening the
distance between items.
\subsection{Sampling from the EPA Distribution}
Sampling from the EPA distribution is a straight forward process
\citep{dahl2017}. Because items are allocated with probability proportional to
the similarity to items in an existing subset, similar items are more likely to
be clustered together in simulation. In order to sample a partition $\pi_n$ from
the EPA distribution, we begin with a permutation (some ordering) of the data
and fixed $\alpha$, $\lambda$. The first item $\sigma_1$ is allocated to a subset
by itself with probability 1. The next item, $\sigma_2$ can either the assigned
to the subset with $\sigma_1$ or to a new subset by itself. The probability of
each allocation is given in equation \ref{epa2}. Let $\sigma_2$ be randomly
assigned to the existing subset or new subset with the respective probabilities.
For each subsequent item, the item $\sigma_{t+1}$ is randomly assigned to an
existing subset or new subset with probability respective to being assigned to
$S_1, S_2, ..., S_{K}, S_{K+1}$. Here $S_1, ... , S_{K}$
represent the existing subsets of the partition and $S_{K+1}$ is a new
subset. Continue to sequentially assign the items until a partition of the data
is obtained. The resulting partition is a single draw from the EPA distribution.
Sampling can be parallelized over many cores to simultaneously obtain many draws
from the EPA distribution.
The order by which the data are sampled affects the resulting probabilities of
obtaining particular partitions. To remove this dependence on sampling order,
randomize the order by which the items are allocated into subsets for each draw.
This has the effect of making the probability of each partition independent of
any particular permutation of the data.
\subsection{Visualizing Pairwise Probabilities}
A key advantage of CaviarPD over traditional clustering is its ability to
quantify and visualize uncertainty in clustering estimates. This is done using a
heat map from a summary of the samples from the EPA distribution.
Each of the partitions $\pi$ can be represented as an $n \times n$ association
matrix denoted $\gamma(\pi)$, where the $(i,j)$ element of the association matrix
is an indicator that observations $\mathbf{x}_i$ and $\mathbf{x}_j$ are in the
same cluster. In short, $ \gamma_{ij}(\pi) = \mathrm{I}(c_i = c_j) $. For $B$
samples from the EPA distribution, there are $B$ $\gamma(\pi)$ matrices. These
matrices $\gamma(\pi_1),...,\gamma(\pi_B)$ can then be averaged together
element-wise to create a pairwise similarity matrix, which contains the
estimated pairwise probabilities that items appear in the same subset for a
given $\alpha$ and $\lambda$. The pairwise similarity matrix is an $n \times n$
matrix denoted $\Psi$, where the $i$th, $j$th element is the relative frequency
with which items $i$ and $j$ are clustered together among the samples.
$$\mathrm{Pr}(c_i = c_j) \approx \Psi_{ij} = \frac{1}{B} \sum_{k=1}^{B} \gamma_{ij}(\pi_k)$$
Each element of the $\Psi$ matrix is the estimated probability that observations
$\mathbf{x}_i$ and $\mathbf{x}_j$ are in the same cluster for a given $\lambda$
and $\alpha$. $\Psi$ can then be conveniently visualized in the form of a heat map.
\begin{figure}[tb]
\centering
\includegraphics[width=2.5in]{figures/wine-conf-3clust.pdf}
\caption{Heat map of the pairwise probabilities for the wine dataset with $\alpha=0.90$, $\delta=0$, and $\tau=10$.}
\label{fig:wine-conf-3clust_1}
\end{figure}
In the heat map, the color of a cell represents the probability that two
items are clustered together while also highlighting the actual clustering
estimate. In Figure~\ref{fig:wine-conf-3clust_1} we can see clearly that, for
$\alpha=0.90$, there appear to be 3 distinct clusters. In this visualization, the
observations are ordered to group similar items together, creating the block
diagonal. The heat map becomes even more useful when the observations are
ordered by an estimated partition, thereby showing the probability relationships
within and between clusters. For example, clusters 1 and 2 appear much more
similar than clusters 1 and 3 because there is a darker shade of yellow in the
off-diagonal blocks for clusters 1 and 2, and a near-white shade of yellow in
the off-diagonals for clusters 1 and 3.
\subsection{Partition Estimation From Samples}
The Bayesian literature provides multiple methods for obtaining a point estimate
of a random partition based on samples from a posterior partition distribution
\citep{dahlsalsopaper}. Again, our paper here is not Bayesian, but we can draw
upon the Bayesian literature on random partitions to obtain a representative point
estimate of a partition based on samples from the EPA distribution.
In decision theory, a loss function is specified in order to pick an optimal
estimate that incurs minimal expected loss. For partitions, loss functions
evaluate how distant the estimated partition is from the true partition of the
data. Binder loss is a function of the number of disagreements
between the estimated and true partition for all possible pairs of observations
\citep{binder1978}. The function is a weighted sum of the two types of
disagreements: observations are in different clusters when they should be in the
same cluster, and observations are in the same cluster when they should be in
different clusters. \citet{wade2018} demonstrate that when the weights of the
two errors are equal, $w_1=w_2$, then the partition that minimizes Binder loss
is given by:
$$ \hat{\pi}_{\text{binder}} = \argmin_{\pi \in \Pi} \sum_{i<j} (\gamma_{ij}(\pi) -
\Psi_{ij})^2. $$
\citet{wade2018} also propose using the variation of information, introduced by
\citet{meila2007} as a loss function. The variation of information is developed
from information theory and is the information present in both clusters
(partition entropy) minus the information shared between the two clusters.
Minimization by enumeration over the whole partition space is
unfeasible, so we use the SALSO method \citep{dahlsalsopaper}
as implemented in the \textsf{R} package \texttt{salso}
\citep{dahl_salso}.
\subsection{Selecting the Mass Parameter}
Most parameters in the EPA distribution have a default value. However,
the mass parameter $\alpha$ does not and it is highly influential in
determining the number of subsets in a partition estimate.
An objective algorithm is needed that can select a default mass
parameter for any given dataset. Ideally, the heat maps generated from
estimates with this algorithm should show that items in the same cluster
are clustered together with high probability. They should also show
clear distinction between subsets of the partition in the heat map (i.e.
low probability regions between clusters). Figure \ref{fig:wine-conf-plots}
shows two heat maps for the pairwise probabilities of the wine dataset
with the mass parameter set at 0.9 and 0.7, respectively. When the mass
is 0.9, the resulting heat map shows overall higher pairwise
probabilities within clusters as compared to the mass set at 0.7.
Likewise, there is less variance of the pairwise probabilities
within clusters when the mass is 0.9. The estimated three clusters
appear more distinct because of the greater within-cluster pairwise
probabilities. The goal of the mass selection algorithm, in the case
of the wine dataset, is to pinpoint $\alpha$ around 0.9 instead of 0.7.
\begin{figure}[tb]
\centering
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width=1.75in]{figures/wine-conf-3clust.pdf}
\caption{$\alpha=0.90$}
\label{fig:wine-conf-3clust}
\end{subfigure}
\quad
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width=1.75in]{figures/wine-conf-2clust.pdf}
\caption{$\alpha=0.70$}
\label{fig:wine-conf-2clust}
\end{subfigure}
\caption{Heat maps of the wine dataset for two different mass parameters}
\label{fig:wine-conf-plots}
\end{figure}
We propose an objective mass selection procedure based on the
silhouette method, which is also used to select $k$ in k-means and
k-medoids clustering \citep{rous1987}. The `silhouette width' of a
given observation in any clustering estimate (not just k-medoids) is a
measure of compactness. It is proportional to the difference between the
observation's average within-cluster distance and its average distance
to points in the nearest cluster. The average silhouette width is the mean
silhouette width for all points in a dataset; seeking to maximize the
average width is intended to result in more compact and distinct clusterings.
Thus, the average silhouette width is a numerical summary that corresponds well
to the properties outlined in the preceeding paragraph of a ``ideal'' heat map.
In our mass selection algorithm, the user proposes a range of cluster
counts to consider in estimation. Boundary mass values are obtained for
the minimum and maximum of this range, and a grid search is performed
between these boundary masses. The mass from the clustering estimate
that results in the maximum average silhouette width is then used
for $\alpha$.
\comment{
$$ \hat{\pi}_{\text{sil}} = \frac{1}{n} \sum_{i=1}^n \frac{b(i) - a(i)}{\max\{a(i), b(i)\}} $$
$$ a(i) = \frac{1}{C_i - 1} \sum_{j \in C_i, i \neq j} d(i,j) \hspace{.2cm} \text{and} \hspace{.2cm} b(i) = \min_{k \neq i} \frac{1}{|C_k|} \sum_{j \in C_k} d(i,j) $$
}
\section{Case Studies}
To compare CaviarPD with the other distance-based clustering methods, we evaluate
how well each method clusters data where the true partition of the data is known.
Datasets were obtained from the UC Irvine machine learning repository
\citep{dua2017}. In order to evaluate the quality of the estimate we use
Binder and VI loss, which both measure the similarity between two
partitions. Both of these metrics are bounded to be nonnegative, where lower values
are indicative of a more accurate partition. We compare average, complete, and Ward
linkages to CaviarPD. Because the possibilities for tree cutting are numerous,
we cut the dendrograms using the default settings for the \texttt{cutreeDynamic}
function in the \texttt{dynamicTreeCut} package in \textsf{R} to obtain a
partition estimate for each of the linkages \citep{langfelder2016r}.
For each of the datasets with numeric attributes, we centered and scaled
the data and computed the Euclidean pairwise distances. For
datasets with categorical attributes, we computed the Jaccard pairwise
distances. Centering and scaling the data coerces the mean and standard
deviation of each predictor to be 0 and 1, respectively. Though not always
necessary, this helps safeguard against attributes with large variance carrying
more weight in the distance computation. We use all three linkage types from
hierarchical clustering to give it the best chance at competing with CaviarPD.
We also choose $k$ in k-medoids using the silhouette method. Despite these
advantages given to these other methods, CaviarPD still remains highly
competitive.
In CaviarPD, we fixed the temperature $\tau$ at 10 for the exponential
similarity function $\lambda$. These defaults seem to be sufficient for the
majority of applications of CaviarPD. We used our mass selection
algorithm to obtain a default value for the mass $\alpha$ and corresponding
partition estimate. When using the mass selection algorithm, Binder loss and
VI gave very similar clustering estimates. For this reason, and for the sake of
simplicity, we use only the Binder loss CaviarPD estimates in our comparisons.
Again, we use the \texttt{caviarpd} function from the \texttt{caviarpd} package
in \texttt{R} to carry out the analysis.
\subsection{Wine Dataset}
For the wine dataset, CaviarPD estimates the true partition on par with the main
combinations of tree cutting and linkages from hierarchical clustering. However,
the partition estimates from hierarchical clustering are inconsistent in
their results. Without knowing the true partition of the data, we would not know
that average linkage produces the best estimate.
\begin{table}[tb]
\begin{tabular}{lrrrrrrrrrrr}
\toprule[1.5pt]
& \multicolumn{3}{c}{Wine} && \multicolumn{3}{c}{House Votes} \\
\cmidrule{2-4} \cmidrule{6-8}
& K & Binder & VI & & K & Binder & VI \\
\midrule
CaviarPD & 3 & \textbf{0.09} & \textbf{0.68} & & 2 & \textbf{0.22} & \textbf{0.99} \\
Average: Default DTC & 4 & \textbf{0.09} & 0.69 & & 3 & \textbf{0.22} & 1.03 \\
Complete: Default DTC & 3 & 0.19 & 1.21 & & 10 & 0.43 & 2.82 \\
Ward: Default DTC & 5 & 0.17 & 1.23 & & 6 & 0.37 & 2.20 \\
K-Medoids & 3 & 0.12 & \textbf{0.68} & & 2 & 0.23 & 1.10 \\
\bottomrule[1.5pt]
\end{tabular}
\caption{Clustering results for the wine and house votes datasets}
\label{tab:wine-house}
\end{table}
Table~\ref{tab:wine-house} gives the clustering results for the wine dataset,
which was used for demonstration in previous sections and contains 13 chemical
attributes on wines from 3 different cultivars. The number of clusters for a
particular estimate is given by $K$. The DTC default gives roughly
the correct number of clusters for average and complete linkage (the fourth
cluster for average linkage is a singleton cluster), but not for ward linkage.
Overall, average linkage and CaviarPD perform the best when
compared to the true partition under Binder loss; however, when using a VI
comparison, CaviarPD and k-medoids slightly outperform all of the DTC cuts.
These differences are minimal, but the estimates for Ward and complete linkage
fall severely short of the estimates produced by CaviarPD and average linkage.
In any case, it is worth noting the lack of consistency in the resulting
estimates from hierarchical clustering, and the ability of CaviarPD to cluster
just as accurately as the best DTC cut and k-medoids.
\subsection{House Dataset}
The house votes dataset contains voting records for the 1984 House of
Representatives. The class attribute is party affiliation, Republican or
Democrat. Having only two clusters yet nearly 500 total observations, this
data provides a rigorous test for all clustering methods. Results are displayed
in Table \ref{tab:wine-house}. CaviarPD and average linkage produce estimates
with nearly identical loss metrics. However, note that CaviarPD also identifies
the correct number of clusters at 2 (along with k-medoids).
\begin{figure}[tb]
\centering
\includegraphics[width=0.75\textwidth]{figures/house-losses.pdf}
\caption{Plot of Binder and VI loss across a grid of mass values for the house votes dataset.}
\label{fig:house-losses}
\end{figure}
The house dataset is also an excellent example for visualizing each loss
function as a function of the mass $\alpha$. Recall that the true partition is
only 2 clusters. Viewing Figure \ref{fig:house-losses}, note how the Binder
loss starts out high and immediately declines. This dip corresponds to the
clustering estimate moving from 1 cluster to 2 clusters as the mass increases.
Since the true partition is 2 clusters, this 2-cluster range of mass values
(between about $\alpha=.65$ and $\alpha=1.0$) results in the lowest Binder
loss. Moving up to the VI loss function, we see that VI loss penalizes less for
a small underestimation of the cluster count. As such, the 1-cluster estimate at
$\alpha=0.5$ is still deemed somewhat `accurate'. However, once the estimate moves
past the 2-cluster range, it quickly increases (particularly for VI loss),
making it clear that mass values any higher than 1.0 will not yield adequate
estimates. Thus, it becomes clear the importance of selecting viable mass
parameters.
\subsection{Flea-Beetle Dataset}
For smaller datasets, it is not uncommon to estimate the true partition of
the data. As an example, we take the flea-beetle dataset, which contains
measurements on three different species of beetles. It contains only 74
observations total. Both CaviarPD and Ward linkage (with the DTC cut) are
able to estimate the exact partition of the data. This results in the
comparative loss functions, both Binder and VI, being equal to 0.
\begin{table}
\begin{tabular}{lrrrrrrrrrrr}
\toprule[1.5pt]
& \multicolumn{3}{c}{Flea-Beetle} && \multicolumn{3}{c}{Olive} \\
\cmidrule{2-4} \cmidrule{6-8}
& K & Binder & VI & & K & Binder & VI \\
\midrule
CaviarPD & 3 & \textbf{0.0} & \textbf{0.0} & & 6 & 0.06 & \textbf{0.89} \\
Average: Default DTC & 2 & 0.19 & 0.83 & & 10 & 0.07 & 1.05 \\
Complete: Default DTC & 2 & 0.27 & 1.08 & & 14 & 0.12 & 1.69 \\
Ward: Default DTC & 3 & \textbf{0.0} & \textbf{0.0} & & 9 & 0.11 & 1.25 \\
K-Medoids & 3 & 0.02 & 0.17 & & 7 & \textbf{0.05} & 0.94 \\
\bottomrule[1.5pt]
\end{tabular}
\caption{Clustering results for the flea and olive datasets}
\label{tab:flea-olive}
\end{table}
\subsection{Olive Dataset}
With the olive dataset we demonstrate the use of the pairwise similarity matrix
to detect clusters within subsets of the estimated partition. The olive dataset
contains measurements on the levels of different oils in olives from nine
different regions of Italy.
The CaviarPD estimate, despite having the lowest VI loss, has
difficulty separating some pairs of regions, resulting in an estimate with
only 6 clusters. k-medoids encounters a similar problem and only detects
7 clusters. On the other hand, the average-linkage DTC estimate
concentrates far too many observations in the first 3 clusters, while also
creating a 10th cluster with only a few observations. The only hierarchical
clustering linkage that correctly estimates 9 clusters is Ward linkage;
however, Ward linkage misclassifies too many observations to be considered
as accurate as the other estimates.
\begin{figure}
\centering
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width=1.75in]{figures/olive-conf.pdf}
\caption{Olive dataset, $\alpha = 2$.}
\label{fig:olive-conf}
\end{subfigure}
\quad
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width=1.75in]{figures/lymph-conf.pdf}
\caption{Lymphography dataset, $\alpha = 1.1$.}
\label{fig:lymph-conf}
\end{subfigure}
\label{olive-lymph}
\caption{Heat maps of the olive and lymphography datasets}
\end{figure}
Heat maps from CaviarPD can be used to examine subcluster structure,
as demonstrated by clusters 4 and 5 in Figure~\ref{fig:olive-conf}.
In cluster 4 of the heat map, there appear to be two higher probability
regions of items being grouped together within each cluster. The same
is observed in cluster 5. In short, those clusters merged olives from
two different regions. Of course, post partition processing is also
possible in hierarchical clustering; however, in hierarchical clustering,
these decisions have no basis on probabilities as they do in CaviarPD.
\subsection{Additional Datasets}
The datasets in Table~\ref{tab:results_all} further demonstrate the
effectiveness of CaviarPD for partition estimation. The yeast dataset, which
contains 10 subsets in the true partition, is an example of a larger dataset
(approximately 1,500 observations) that makes running the entire mass selection
algorithm computationally intensive. Though the algorithm is still feasible
for a dataset this size, it may not be practical for significantly larger
sets. In such cases, we recommend building heat maps for 3 to 5
different values of $\alpha$ and selecting the mass for the plot with the
most concentrated pairwise probabilities. The lymphography
dataset is an example of data for which the 4 cluster distinctions are not
well-represented by the attributes. This leads to poor estimates by all
three methods, and a relatively uninterpretable
heat map as given in Figure \ref{fig:lymph-conf}. The E. coli dataset
contains attributes for 8 different localization sites of proteins in E. coli
bacteria. Chemical composition measurements are taken on 6 different glass
products in the glass dataset.
\begin{table}[tb]
\begin{tabular}{lrrrrrrrrrrr}
\toprule[1.5pt]
& \multicolumn{3}{c}{Yeast} && \multicolumn{3}{c}{Lymphography} \\
\cmidrule{2-4} \cmidrule{6-8}
& K & Binder & VI & & K & Binder & VI \\
\midrule
CaviarPD & 12 & 0.32 & 3.06 & & 6 & 0.49 & \textbf{1.44} \\
Average: Default DTC & 5 & 0.51 & \textbf{3.00} & & 4 & \textbf{0.41} & 2.25 \\
Complete: Default DTC & 19 & \textbf{0.24} & 4.83 & & 4 & 0.44 & 2.53 \\
Ward: Default DTC & 16 & \textbf{0.24} & 4.70 & & 4 & \textbf{0.41} & 2.56 \\
K-Medoids & 8 & 0.26 & 3.48 & & 2 & 0.43 & 1.96 \\
\midrule
& \multicolumn{3}{c}{E. Coli} && \multicolumn{3}{c}{Glass Products} \\
\cmidrule{2-4} \cmidrule{6-8}
& K & Binder & VI & & K & Binder & VI \\
\midrule
CaviarPD & 6 & 0.12 & \textbf{1.22} & & 6 & 0.35 & 2.92 \\
Average: Default DTC & 2 & 0.72 & 2.17 & & 3 & 0.50 & \textbf{1.90} \\
Complete: Default DTC & 6 & \textbf{0.11} & 1.25 & & 4 & 0.42 & 2.42 \\
Ward: Default DTC & 6 & 0.18 & 1.65 & & 4 & 0.35 & 3.05 \\
K-Medoids & 6 & 0.17 & 1.80 & & 6 & \textbf{0.34} & 2.93 \\
\bottomrule[1.5pt]
\end{tabular}
\caption{Clustering results for additional datasets}
\label{tab:results_all}
\end{table}
\subsection{Discussion}
We do not claim that CaviarPD
vastly outperforms all hierarchical clustering and k-medoids methods, as
that was not usually the case in these eight case studies. Rather, we have
shown that CaviarPD performs comparably to (and in some cases, slightly
better than) the other methods. The benefit of CaviarPD, however, is the
ability to assess clustering uncertainty in the
form of the heat map plots from the pairwise similarity matrix.
In hierarchical clustering, there is no one linkage type or cutting technique
that consistently produces the best cut of the tree. We compared the best
possible DTC cut of the dendrogram for each linkage, but there is no cutting
rule that will consistently guide the user to that result. In k-medoids
clustering, there is less variability, but still no way to express uncertainty.
In contrast to these techniques, CaviarPD has a validating method to select
the mass parameter $\alpha$. It also produces a pairwise similarity matrix which
provides a clear and consistent representation of how the data are grouped.
\subsection{Other Distributions}
The EPA distribution is suited for the CaviarPD method since it is a random
partition distribution taking pairwise distances as input. Most other
random partition distrtibutions in the Bayesian literature do not take pairwise
probabilities and would therefore not be suitable for CaviarPD. There is, however,
one other partition distribution which does, called the distance-dependent Chinese Restaurant
Process (ddCRP) \citep{blei2011}. The \texttt{caviarpd} function
allows users to specify the ddCRP distribution rather
than the default EPA distribution. Hence, clustering with the ddCRP distribution
from a coding standpoint is merely one additional argument.
However, default parameters for clustering with the ddCRP distribution have not
been investigated, and we have not found a good method for selecting optimal mass
for the ddCRP.
In addition, while the ddCRP distribution does sometimes generate estimates that
are comparable with CaviarPD, it does not appear to lead to effective visualizations.
As an example, we take the flea-beetle dataset in Figure
\ref{fig:flea-conf-plots}, for which CaviarPD (and Ward linkage) were able to
estimate precisely the true partition. Plotting the estimate from the EPA
distribution shows clear distinction between the 3 clusters. All items in the
same cluster have high probabilities of being clustered together, and we can see
that elements in clusters 2 and 3 are extremely unlikely to be grouped together.
The plot from the ddCRP estimate shows none of this - it is virtually monochromatic
(besides the diagonal 1.0 probabilities of each item being clustered with
itself).
Estimates with the ddCRP distribution do not seem to be any more effective than
those from the EPA. They also are unable to provide the key insights that the
EPA estimates do, and have far more subjectivity in tuning the parameters. For
these reasons, we do not currently recommend using the ddCRP for the CaviarPD method.
\begin{figure}
\centering
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width=1.75in]{figures/flea-epa.pdf}
\caption{EPA Distribution}
\label{fig:flea-epa}
\end{subfigure}
\quad
\begin{subfigure}{.48\textwidth}
\centering
\includegraphics[width=1.75in]{figures/flea-ddcrp.pdf}
\caption{ddCRP Distribution}
\label{fig:flea-ddcrp}
\end{subfigure}
\caption{Heat maps of the flea-beetle dataset. The EPA clustering estimates the exact partition, while the ddCRP clustering estimates a near-exact partition.}
\label{fig:flea-conf-plots}
\end{figure}
\section{Conclusion} \label{conclusion}
Cluster analysis via random partition distributions simplifies the user's
dilemma for clustering. In hierarchical clustering, the different linkages and
lack of consistent tree cutting rules present a user with many subjective
choices. These different choices in hierarchical clustering lead to highly
varied partition estimates. For both hierarchical and k-medoids clustering,
there are no probability statements one can make regarding the clustering.
In contrast, data driven statistics help guide the
user in clustering estimation for CaviarPD. The pairwise similarity matrix
$\Psi$ gives a probabilistic understanding for how items are clustered together
in a dataset. Algorithmic clustering methods are not based on a
probability distribution and therefore quantify clustering uncertainty is difficult.
The central weakness of CaviarPD is the computational cost to select
a mass for large datasets. Hierarchical clustering, k-medoids (and the
corresponding k-means method) all have the potential to
run more quickly in these cases. It is also worth noting that while the
heat maps can be very insightful, they do not always visualize
overall clustering structure as comprehensively as a dendrogram would
in hierarchical clustering \citep{sander}. In summary, CaviarPD is not
the undisputed best clustering method in every instance. Rather, it
provides unique advantages over traditional approaches in the majority
of clustering problems.
\pagebreak
\bibliographystyle{elsarticle-harv} |
\section{Introduction}
\IEEEPARstart{C}oNFIDENCE-based motion planning~\cite{nakka2019nsoc,blackmore2010probabilistic,blackmore2011chance,toit2012robot} and control algorithms~\cite{tsukamoto2020,zhu2019chance}, that incorporate uncertainties in the dynamic model and environment to guarantee safety and performance with high probability, enable safe operation of robots and autonomous systems in partially-known and dynamic environments. A probabilistic approach can allow for integration with a higher-level discrete decision-making algorithm for information gathering~\cite{kaelbling2013integrated,nakka2021information}, and for safe exploration~\cite{nakka2020chance,wabersich2021,Nakka2021SpacecraftLearning} to learn the interaction with an unknown environment. Examples of autonomous systems that require safety guarantees under uncertainty include spacecraft with thrusters as actuators during proximity operations~\cite{nakka2018six,nakka2021information,Nakka2021SpacecraftLearning}, powered descent on Mars~\cite{ridderhof2018uncertainty}, and quadrotors flying in turbulent winds~\cite{shi2018neural,shi2020neural}.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{figures_tro/main_algo_flow_v2.pdf}
\caption{\changed{Caltech's M-STAR~\cite{nakka2018six} (Multi-Spacecraft Testbed for Autonomy Research) planning a safe trajectory to ensure safety under uncertainty in actuation during a proximity maneuver. The motion planning problem is formulated as a chance-constrained stochastic optimal control problem solved in two steps. Step 1: Project the stochastic problem to a deterministic problem by using generalized polynomial chaos approach and distributional robustness (the different colors correspond to trajectories with different values of the cost function); and Step 2: Use deterministic solvers to compute an optimal solution to the deterministic problem.}}\label{fig:space_lab_experiment}
\end{figure}
A motion planning problem considering safety in conjunction with optimality under uncertainty can be formulated as a continuous-time continuous-space stochastic nonlinear optimal control problem (SNOC) with chance constraints. In this paper, we present the generalized polynomial chaos-based sequential convex programming (gPC-SCP) method, as depicted in~\cref{fig:space_lab_experiment}, to solve a chance-constrained SNOC problem. The method involves deriving a deterministic nonlinear optimal control (DNOC) problem with convex constraints that are a surrogate to the SNOC problem with linear and quadratic chance constraints. We derive the DNOC problem by accounting for nonlinear stochastic dynamics using generalized polynomial chaos expansions (gPC)~\cite{xiu2002wiener,xiu2009fast,ghanem2003stochastic} and obtaining deterministic convex approximations of linear and quadratic chance constraints using distributional robustness~\cite{nemirovski2007convex,calafiore2006distributionally,zymler2013distributionally}. \changed{The DNOC problem is then solved using sequential convex programming (SCP)~\cite{morgan2014model,morgan2012spacecraft,morgan2016swarm} for trajectory optimization. The gPC-SCP method computes the full probability distribution of the state as a function of time. In order to compute only the mean of the state as a function of time, we derive a predictor-corrector gPC-SCP (gPC-SCP$^\mathrm{PC}$) method, where we decouple the prediction step (uncertainty propagation) and the optimization step (trajectory correction), thereby deriving an iterative algorithm for fast real-time planning.}
The main contributions of the paper are as follows:
\begin{enumerate}[label=(\alph*)]
\item We present a systematic sequence of approximations for the chance-constrained SNOC problem to compute a convex-constrained DNOC problem using gPC projection. We analyze the gPC projection of the stochastic dynamics for existence and uniqueness~\cite{kushner1967stochastic,khasminskii2011stochastic} of a solution in the gPC space. We provide examples to study the effect of the projection on both the controllability of surrogate dynamics and the feasibility of the DNOC problem. We prove the convexity of the distributionally-robust linear and quadratic chance constraints in the gPC space.
\item To characterize the deterministic approximation obtained using gPC projection, we present a rigorous analysis on the convergence of the DNOC problem to the SNOC problem for the unconstrained case. Then, we prove that any feasible solution to the constrained DNOC problem is feasible for the chance-constrained SNOC problem with an appropriate gPC transformation step applied.
\item We derive convex surrogates for collision checking with deterministic and stochastic obstacle state models. We integrate this collision constraint with a sampling-based planning method~\cite{hauser2016,lavalle_2006} to derive an algorithm that computes safe and optimal motion plans under uncertainty.
\item \changed{We derive a predictor-corrector algorithm using the gPC-SCP method to compute a mean state trajectory that is safe and optimal under uncertainty. We compare gPC-SCP$^\mathrm{PC}$ with gPC-SCP and in terms of problem complexity, computation cost, and the ability to perform covariance control.}
\item We validate the convergence and the safety provided by the convex constraints in simulation on a three-degree-of-freedom robot dynamics and six-degree-of-freedom spacecraft simulator robot. We show empirically that the gPC-SCP and the gPC-SCP$^\mathrm{PC}$ methods compute a safer trajectory, thereby having a higher success rate in comparison to the Gaussian approximation~\cite{zhu2019chance,du2011probabilistic} of the collision chance constraints. We demonstrate the efficacy of the gPC-SCP and the gPC-SCP$^\mathrm{PC}$ methods by computing a safe trajectory for a spacecraft proximity maneuver under uncertainty in the environment (obstacles) on the robotic spacecraft dynamics simulator hardware~\cite{nakka2018six} and by executing the trajectory in real-time closed-loop experiments.
\end{enumerate}
\subsection{Related Work}
Existing methods to solve a chance-constrained stochastic optimal control problem use moment space propagation~\cite{todorov2005generalized,van2012motion,ridderhof2019nonlinear,zhu2019chance}, unscented transformation-based propagation~\cite{calafiore2013stochastic}, Monte Carlo sample propagation~\cite{janson2018monte,blackmore2010probabilistic,blackmore2011chance}, and scenario-based~\cite{calafiore2013,calafiore2013stochastic} approaches to construct a deterministic surrogate problem. Although these methods alleviate the curse of dimensionality, they do not provide asymptotic convergence guarantees for a DNOC problem. Monte Carlo methods provide asymptotic convergence guarantees but often require large samples to estimate the constraint satisfaction for nonlinear systems and use mixed-integer programming~\cite{blackmore2010probabilistic} solvers for computing a solution. We use gPC propagation~\cite{ghanem2003stochastic} to construct a DNOC problem that converges to the SNOC problem asymptotically. The gPC projection transforms the chance constraints from being a non-convex constraint in moment space to a convex constraint in the gPC space. This enables the use of sequential convex programming~\cite{morgan2014model,morgan2016swarm} method for computing a solution. Additionally, we study the existence and uniqueness~\cite{arnold1974stochastic} of a solution and the controllability of the deterministic surrogate dynamics of the stochastic dynamics.
Earlier work~\cite{castillo2020real,zhu2019chance,blackmore2011chance} uses a Gaussian approximation of the linear and the quadratic chance constraint for collision checking and for terminal constraint satisfaction. While this avoids multi-dimensional integration of chance constraints for feasibility checking, Gaussian approximation might not be an equivalent representation (or) even a subset of the feasible set in the presence of stochastic process noise in dynamics. We use distributional robustness~\cite{calafiore2006distributionally,zymler2013distributionally} property to propose a new deterministic second-order cone constraint and a quadratic constraint approximation of the linear and quadratic chance constraints. We prove that the deterministic approximations are a subset of the respective chance constraints.
In~\cite{blackmore2010probabilistic,blackmore2011chance}, linear chance constraints were considered for probabilistic optimal planning for linear systems. The literature on chance-constrained programming focuses on problems with deterministic decision variables and uncertain system parameters for both linear~\cite{calafiore2006distributionally} and nonlinear~\cite{zymler2013distributionally} cases. The results~\cite{calafiore2006distributionally,nemirovski2007convex} on distributional robust subset and convex approximations of the chance constraints can be readily transformed to the case with a random decision variable for an unknown measure. The quadratic chance constraint would lead to an inner semi-definite program~\cite{vandenberghe2007generalized} that adds complexity to the SNOC problem considered in this paper. The linear chance constraint for collision checking was first presented in~\cite{nakka2019nsoc}. In~\cite{tlew2020}, authors show that linearized chance constraint is a subset of the original nonlinear chance constraint for a Gaussian confidence-based constraint. Since the local Gaussian assumption might not be valid for nonlinear systems, we present proof for the distributionally robust convex constraint formulation that includes uncertainty in obstacle state for a nonlinear stochastic differential equation.
The gPC expansion approach was used for stability analysis and control design of uncertain systems~\cite{mesbah2016stochastic,mesbah2014stochastic,bavdekar2016stochastic,hover2006application,fisher2008stability,kim2013wiener}. For trajectory optimization, recent work focused on nonlinear systems with parametric uncertainty~\cite{boutselis2017stochastic,fisher2011optimal} with no constraints on the state or linear systems with linear chance-constraints that do not extend to the SNOC problem considered here and lack analysis on the deterministic approximation of the uncertain system. The gPC approach was used to compute a moment-space receding horizon approximation~\cite{buehler2017efficient}, which was solved using nonlinear programming methods. We extend prior work to incorporate nonlinear dynamics and include analysis on the deterministic approximation. We formulate convex constraints for linear and quadratic constraints in gPC space and use this formulation to design algorithms for motion planning and control of a nonlinear stochastic dynamic system.
\changed{Furthermore, we present a new predictor-corrector formulation (gPC-SCP$^\mathrm{PC}$). The gPC-SCP$^\mathrm{PC}$ method decouples the stochastic propagation using gPC and planning using SCP to optimize a nominal trajectory and ensure safety under uncertainty. In~\cite{foust2020optimal}, a similar approach was used to correct for the linearization and discretization errors via nonlinear propagation in the SCP. To the best of our knowledge, no other work used such an approach for planning under uncertainty for nonlinear stochastic systems with chance constraints.}
We generalize and extend our prior conference paper~\cite{nakka2019nsoc} significantly as follows: i) we derive a sequence of approximations from a SNOC problem to the DNOC problem, which provides a modular architecture to understand trajectory optimization under uncertainty; ii) we include examples discussing the effect of gPC projection on the controllability of stochastic dynamics; iii) we design a motion planning algorithm to handle uncertainty in both dynamics and obstacle location, which takes advantage of the state-of-the-art sampling-based~\cite{hauser2016} planning algorithms; iv) \changed{we formulate a predictor-corrector extension of gPC-SCP and provide conditions for real-time trajectory generation; and v) we validate the gPC-SCP and the gPC-SCP$^\mathrm{PC}$ approach empirically on three and six-degree-of-freedom robot dynamics and the spacecraft simulator hardware testbed.}
\subsubsection*{Organization} We discuss the stochastic nonlinear optimal control (SNOC) problem with results on deterministic approximations of chance constraints along with preliminaries on gPC expansions in~\cref{sec:nsoc_prelim}. In~\cref{sec:det_opt_prob}, we present the deterministic surrogate of the SNOC problem in terms of the gPC coefficients and the SCP formulation of the DNOC problem. In~\cref{sec:motion_planning_gpc_scp}, we apply the gPC-SCP method under uncertainty in dynamics and constraints for motion planning using SNOC solutions. Furthermore, we derive the gPC-SCP$^\mathrm{PC}$ method and compare with the original gPC-SCP method. In~\cref{sec:simulations_and_experiments}, we validate the gPC-SCP method via simulations on a 3DOF and 6DOF spacecraft simulator robot and via experiments on the robotic spacecraft simulator testbed. Finally, we conclude the paper in~\cref{sec:conclusion} with a brief discussion on the approach and impact of the method.
\subsubsection*{Notation}
For a random variable $x \in \mathbb{R}^{d_x}$, $\mu_x$ is the mean, $\Sigma_x$ is the covariance matrix, $\mathbb{R}$ is real line, $d_{x}$ is the dimension of $x$. $\mathbb{E}$ and $\mathrm{Pr}$ are the expectation operation and probability measure, respectively. We define a deterministic vector as $\bar{x}$. The $p-$norm of a vector $\bar{u} \in \mathbb{R}^{d_u}$ is defined as $\|\bar{u}\|_{p} = (\sum_{1}^{d_{u}}|\bar{u}_i|^p)^{\frac{1}{p}}$. The risk measure for constraint violation is $\epsilon$. We define the gPC state using $\mathrm{X}$ and the indicator function as $I$. We use $\mathbb{I}$ for an identity matrix and $\mathds{1}$ for a matrix with entries as $1$. The Kronecker's product of two matrices $A_{m \times n}$ and $B_{p \times q}$ is defined as follows:
\begin{equation}
(A \otimes B)_{mp \times nq} = \begin{bmatrix} a_{11}B & \dots & a_{1n}B\\ \vdots & \ddots & \vdots \\ a_{m1}B & \dots & a_{mn}B\end{bmatrix}, \nonumber
\end{equation}
where $a_{ij}$ is the element at $i^{\mathrm{th}}$ row and $j^{\mathrm{th}}$ column of A. For a matrix $A$, $\mathrm{tr} (A)$ is the trace operation, $\lambda_{\min}(A)$ and $\lambda_{\max}(A)$ are the minimum and maximum eigenvalues of $A$.
\section{Problem and Preliminaries} \label{sec:nsoc_prelim}
In this section, we present the stochastic optimal control problem formulation, preliminaries on the relaxations used for chance constraints, and the generalized polynomial chaos approach that forms a basis for constructing a surrogate deterministic optimal control problem.
\subsection{Stochastic Nonlinear Optimal Control Problem}\label{subsec:stoc_opt_prob}
We consider the finite-horizon stochastic nonlinear optimal control (SNOC) problem with joint chance constraints in continuous time and continuous space. The SNOC problem minimizes an expectation cost function, which is the sum of a quadratic function in the random state variable $x(t)$ and a convex norm of the control policy $\Bar{u}(t)$. The evolution of the stochastic process $x(t)$ for all sampled paths is defined by a stochastic differential equation. The joint chance constraints guarantee constraint feasibility with a probability of $1- \epsilon$, where $\epsilon >0$ and is chosen to be a small value (e.g. $\epsilon \in [0.001,0.05]$) for better constraint satisfaction. The following optimal control problem is considered with the state distribution and control as the decision variables.
\begin{problem}
\label{prob:cc_stoc_opt_cntrl}
Chance-Constrained Stochastic Nonlinear Optimal Control.
\begin{align}
J_{\mathrm{SNOC}}^{*} = & \underset{x(t),\Bar{u}(t)}{\min}
\scalebox{0.9}{$\mathbb{E} \left[\int_{t_{0}}^{t_{f}}J(x(t),\Bar{u}(t))dt + J_{f}(x(t_{f}))\right]$}\\
\text{s.t.} \quad
& \scalebox{0.9}{$dx = f(x(t),\Bar{u}(t))dt + g(x(t),\Bar{u}(t)) dw(t)$}\\
&\scalebox{0.9}{$\mathrm{Pr} (x(t) \in \mathcal{X}_{\mathcal{F}}) \geq 1-\epsilon \ \forall t \in [t_{0},t_{f}]$}\\
&\scalebox{0.9}{$\Bar{u}(t) \in \mathcal{U} \quad \forall t \in [t_{0},t_{f}]$} \label{eq:control_limits}\\
&\scalebox{0.9}{$x(t_{0}) = x_{0} \quad x_{t_{f}} \in \mathcal{X}_{f}$} \label{eq:init_term_conditions}
\end{align}
\end{problem}
The cost functional $J$ and the terminal cost $J_{f}$ are:
\begin{align}
J(x(t),\Bar{u}(t))&= x(t)^\top Qx(t) + \|\bar{u}\|_{p},\ \text{where} \ p \in \{1,2,\infty\},\nonumber\\
J_{f}(x(t_{f}))&=x(t_f)^\top Q_{f}x(t_f).
\label{eq:cost_stopt}
\end{align}
where $Q$ and $Q_{f}$ are positive definite matrices. The $p-$norm of a vector $\bar{u}$ is defined as $\| \bar{u}\|_{p} = (\sum_{1}^{d_{u}}|\bar{u}_i|^p)^{\frac{1}{p}}$. The terminal set $\mathcal{X}_{f}$ is the set of allowed realization of the state $x$ after propagation. We apply the terminal constraint with slackness in the terminal variance to ensure feasibility of~\cref{prob:cc_stoc_opt_cntrl}.
In the following, we define each of the aforementioned elements of \cref{prob:cc_stoc_opt_cntrl} and discuss convex approximations of linear and quadratic chance constraints.
\subsubsection{Stochastic Differential Equation (SDE)~\cite{arnold1974stochastic}}
The dynamics of the system is modeled as a controlled diffusion process with $\mathrm{It\hat{o}}$ assumptions. The random variable $x(t)$ is defined on a probability space $(\Omega,\mathbb{F},\mathrm{Pr})$ where $\Omega$ is the sample space, $\mathbb{F}$ forms a $\sigma$-field with measure $\mathrm{Pr}$.
\begin{equation}
\begin{aligned}
& dx(t) = f(x(t),\Bar{u}(t))dt + g(x(t),\Bar{u}(t))dw(t),\\&
\mathrm{Pr}(|x(t_{0}) - x_{0}| = 0) = 1, \quad \forall t_{0} \leq t \leq t_{f} < \infty,
\end{aligned}\label{eq:nl_stochastic_dynamics}
\end{equation}
where: $f(.,.):\mathcal{X} \times \mathcal{U} \to \mathbb{R}^{d_{x}}$, $g(.,.): \mathcal{X} \times \mathcal{U} \to \mathbb{R}^{d_{x} \times d_{\xi}}$, and $w(t)$ is a $d_{\xi}$-dimensional Wiener process and the initial random variable $x_{0}$ is independent of $w(t)-w(t_{0})$ for $t \geq t_{0}$, and $dw(t) \sim \mathcal{N}(0,dt\mathbb{I})$. The sets $\mathcal{X} \subseteq \mathbb{R}^{d_{x}}$ and $\mathcal{U} \subseteq \mathbb{R}^{d_{u}}$ are compact sets. We make the following assumptions to ensure the existence and uniqueness of a solution to the SDE.
\begin{asmp}The functions $f(x(t),\Bar{u}(t))$ and $g(x(t),\Bar{u}(t))$ are defined and measurable on $\mathcal{X} \times \mathcal{U}$.
\end{asmp}
\begin{asmp}Equation~\cref{eq:nl_stochastic_dynamics} has a unique solution $x(t)$, which is continuous with probability 1, and $\exists$ a $K \in\mathbb{R}^{++}$ such that the following conditions are satisfied:\\
a) Lipschitz condition~\cite{arnold1974stochastic}: $\forall t \in [t_{0},t_{f}]$, $s_{1} \& s_{2} \in \mathcal{X} \times \mathcal{U}$,
\begin{align}
&\|f(s_1) - f(s_2)\| + \|g(s_1)-g(s_2)\|_{\mathrm{F}} \leq K \|s_1-s_2\|,
\label{eq:lipshitz_stoc}\end{align}
\noindent b) Restriction on growth: $\forall t \in [t_{0},t_{f}]$, $s_1\in \mathcal{X} \times \mathcal{U}$
\begin{equation}
\begin{aligned}
\|f(s_1)\|^2 + \|g(s_1)\|_{\textit{F}}^2 \leq K^2 (1 + \|s_1\|^2).
\end{aligned}\label{eq:growth_stoc}
\end{equation}
\end{asmp}
We use the following definition to study the controllability of the deterministic approximation of the SDE~\cref{eq:nl_stochastic_dynamics}.
\begin{defn}\label{defn:stoc_controllability} The SDE~\cref{eq:nl_stochastic_dynamics} is $\epsilon_{c}$-controllable~\cite{kushner1967stochastic}. For any initial state $x_{0} \in \mathcal{X}$, we can compute a sequence of control $\bar{u}(t)$ $\forall t \in [t_{0}, t_{f}]$ such that $\mathrm{Pr}\left(\|x - x(t_f)\|^2 \geq \delta \mid x(t_0) = x_0 \right) \leq \epsilon_c$, where $x(t_f)$ is the terminal state, $\delta>0$ and $\epsilon_c>0$ are small, and $t_f$ is finite.
\end{defn}
\textit{Control Policy.} We assume that the control policy $\Bar{u}(t) \in \mathcal{U} \subseteq \mathbb{R}^{d_u}$ is deterministic and the set $\mathcal{U}$ is a convex set. The deterministic control policy is motivated by a hardware implementation strategy, where a state-dependent Markov control policy defined on the compact set $\mathcal{X}$ is sampled for a value with highest probability (or) for the mean.
\noindent \textbf{Note:} We use the gPC method to project the SDE to an Ordinary Differential Equation (ODE) in a higher dimensional space for propagating the dynamics.
\subsubsection{Chance Constraints~\cite{calafiore2006distributionally}}
In order to accommodate the unbounded uncertainty model in the dynamics, the feasible region $\mathcal{X}_{\mathcal{F}}$ defined as,
\begin{equation}
\mathcal{X}_{\mathcal{F}} = \{x(t) \in \mathcal{X}: h_i(x(t)) \leq 0 \ \forall \ i\ \in\ \{1,\dots, m \}\},
\label{eq:const_set}
\end{equation}
is relaxed to a chance constraint (CC) using the risk measure $\epsilon$,
\begin{equation}
\mathcal{X}_\mathrm{CC} = \{ x(t) \in \mathcal{X}: \mathrm{Pr}(x(t) \in \mathcal{X}_{\mathcal{F}}) \geq 1- \epsilon\},
\label{eq:chance_const}
\end{equation}
with a constraint satisfaction probability of $1-\epsilon$. The constraint set $\mathcal{X}_{\mathcal{F}}$ is assumed to be the polytope $\mathcal{X}_{\mathcal{F}} = \{x \in \mathcal{X}: \land_{i =1}^{m} a_{i}^\top x + b_{i} \leq 0\}$ with $m$ flat sides, or a quadratic constraint set $\mathcal{X}_{\mathcal{F}} = \{x \in \mathcal{X}: x^\top A x \leq c\}$ for any realization $x$ of the state. The joint chance constraint formulation of the polytopic constraint is of the form, $\mathrm{Pr}(\land_{i =1}^{m} a_{i}^\top x + b_{i} \leq 0) \geq 1- \epsilon$.
A convex relaxation of the individual chance constraint for an arbitrary distribution of the state vector $x(t)$ due to the nonlinearity in the system is intractable, so an extension of the problem called Distributionally-Robust Chance Constraints (DRCC) given as follows,
\begin{equation}
\mathcal{X}_\mathrm{DRCC} = \{ x(t)\in \mathcal{X}: \inf_{x(t) \sim (\mu_x,\Sigma_{x})} \mathrm{Pr}(x(t) \in \mathcal{X}_{\mathcal{F}}) \geq 1- \epsilon \},
\label{eq:drcc}
\end{equation}
where the chance constraint is satisfied for all distributions with known mean and variance of the decision variable is used. The set defined by the DRCC in~\cref{eq:drcc} is a conservative approximation ~\cite{zymler2013distributionally} of the chance constraint i.e., $\mathcal{X}_\mathrm{DRCC} \subseteq \mathcal{X}_\mathrm{CC}$.
\noindent a) \textit{Distributionally-Robust Linear Chance Constraint (DRLCC)}~\cite{calafiore2006distributionally}: Consider a single Linear Chance Constraint (LCC) with $a \in \mathbb{R}^{d_{x}}$ and $b \in \mathbb{R}$:
\begin{equation}
\mathcal{X}_\mathrm{LCC} = \{ x(t)\in \mathcal{X}:\mathrm{Pr}(a^\top x(t)+ b \leq 0) \geq 1- \epsilon \}.
\label{eq:cc_lin}
\end{equation}
The column vector $a$, real constant $b$, and risk measure $\epsilon$ are known a priori. The state vector $x$ is the decision variable. Assuming that the mean $\mu_x$ and the covariance $\Sigma_{x}$ of $x$ are known, a distributionally-robust constraint version of \cref{eq:cc_lin} is given as follows:
\begin{equation}
\mathcal{X}_\mathrm{DRLCC}=\{x(t)\in\mathcal{X}:\inf_{x(t)\sim(\mu_x,\Sigma_{x})}\mathrm{Pr}(a^\top x(t)+ b\leq 0)\geq 1-\epsilon\}.\label{eq:drcc_lin}
\end{equation}
Equivalently, \cref{eq:drcc_lin} can be rewritten in the following deterministic form, which will be used to derive a second-order cone constraint for the DNOC in~\cref{sec:finite_dimensional_approximation}.
\begin{equation}
\mathcal{X}_\mathrm{DRLCC} = \{ x(t)\in \mathcal{X}: a^\top \mu_x(t) + b+ \sqrt{\frac{1-\epsilon}{\epsilon}}\sqrt{a^\top \Sigma_{x}a} \leq 0 \}
\label{eq:lin_socc}
\end{equation}
\begin{lemma} \label{lemma:lin_consev_approx}
The set $\mathcal{X}_\mathrm{DRLCC}$ in~\cref{eq:lin_socc} is a subset of $\mathcal{X}_\mathrm{LCC}$ defined in \cref{eq:cc_lin}.
\end{lemma}
\begin{proof}
See Theorem 3.1 in~\cite{calafiore2006distributionally}.
\end{proof}
If the dynamics are linear, the constraint in~\cref{eq:lin_socc} is replaced with a tighter equivalent deterministic constraint given by the following inequality:
\begin{equation}
a^\top \mu_x(t) + b+ \sqrt{2}\mathrm{erf}^{-1}(1-2\epsilon)\sqrt{a^\top \Sigma_{x}a} \leq 0,
\label{eq:gaussian_linear_constraint}
\end{equation}
where the function $\mathrm{erf}(.)$ is defined as $\text{erf}(\delta)=\frac{2}{\sqrt{\pi}}\int_{0}^{\delta} e^{-t^2} dt$ and $ \forall \epsilon \in (0,0.5)$. The constraint set is transformed to a second-order cone constraint in the gPC variables. \changed{In~\cref{fig:dr_vs_gaussian}, we show a comparison of the robustness provided by the distributional robustness cosntraint~\cref{eq:lin_socc} and the Gaussian constraint~\cref{eq:gaussian_linear_constraint}.}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{mstar_results/dr_vs_g.pdf}
\caption{\changed{A comparision of the distributionally robust constraint tightening bound $\sqrt{\frac{1-\epsilon}{\epsilon}}$ used in~\cref{eq:lin_socc} and the Gaussian bound $\sqrt{2}\mathrm{erf}^{-1}(1-2\epsilon)$ used in~\cref{eq:gaussian_linear_constraint}.}}
\label{fig:dr_vs_gaussian}
\end{figure}
\begin{rmrk}
The risk measure $\epsilon$ in~\cref{eq:cc_lin} is assumed to be in the range $[0.001,0.5]$. For small values of $\epsilon$ ($< 0.001$), the value $\sqrt{\frac{1-\epsilon}{\epsilon}}$ increases dramatically, as shown in~\cref{fig:dr_vs_gaussian}. This decreases the feasible space defined by the set $\mathcal{X}_\mathrm{LCC}$ drastically leading to numerical issues in the gPC-SCP method. For handling the risk of a very small value of $\epsilon$ (e.g., $1e-7$, as discussed in~\cite{cheng2021limits}), the uncertainty in the system needs to be modeled accurately such that $\Sigma_x$ is small or a newer deterministic surrogate method needs to be developed to overcome the numerical instability.
\end{rmrk}
\noindent b) \textit{Conservative Quadratic Chance Constraint (CQCC):} \cref{lemma:quad_const} presents a new conservative deterministic relaxation for the quadratic chance constraint that is used to bound the deviation of the random vector $x(t)$ from the mean $\mu_x(t)$.
\begin{lemma}
\label{lemma:quad_const}
The constraint set
\begin{equation}
\mathcal{X}_\mathrm{CQCC} = \{ x(t)\in \mathcal{X}: \frac{1}{c}\text{tr}(Q\Sigma_x) \leq \epsilon\}
\label{eq:sdp_constraint}
\end{equation}
is a conservative approximation of the original Quadratic Chance Constraint (QCC)
\begin{equation}
\mathcal{X}_\mathrm{QCC} = \{ x\in \mathcal{X}: \mathrm{Pr}((x-\mu_x)^\top Q (x-\mu_x) \geq c) \leq \epsilon\}
\label{eq:quad_chnc_cons}
\end{equation}
i.e., $\mathcal{X}_\mathrm{CQCC} \subseteq \mathcal{X}_\mathrm{QCC}$, where $Q \in \mathbb{R}^{n\times n}$ is a positive definite matrix and $c \in \mathbb{R}^{++}$ and $\Sigma_x$ is the covariance of the random variable $x$
\end{lemma}
\changed{\begin{proof}
We prove that any random vector $x \in \mathcal{X}$ that is in the set $\mathcal{X}_\mathrm{CQCC}$ is also in the set $\mathcal{X}_\mathrm{QCC}$ implying $\mathcal{X}_\mathrm{CQCC} \subseteq \mathcal{X}_\mathrm{QCC}$. The proof follows from the approach taken to prove the multivariate Chebyschev's inequality~\cite{chen2007new}. Let $F(x)$ be the Cumulative Distribution Function (CDF) of the random variable $x$ and $v = x - \mu_x$.
\begin{equation*}
\begin{aligned}
\mathcal{G} = \{ v \in \mathcal{X} : v^\top Q v \geq c \} \implies \frac{1}{c} v^\top Q v \geq 1 \ \ \forall v \in \mathcal{G}.
\end{aligned}
\end{equation*}
Using the definition of probability in terms of the CDF,
\begin{equation*}
\begin{aligned}
\mathrm{Pr}((x-\mu_x) \in \mathcal{G}) & \leq \frac{1}{c} \int_{v\in \mathcal{G}} v^\top Q v dF(v)\\ & \leq \frac{1}{c} \int_{v \in \mathcal{R}^n}v^\top Q v dF(v).
\end{aligned}
\end{equation*}
Let $q_{ij}$ denote an element of matrix $Q$ in the $i^\mathrm{th}$ row and $j^\mathrm{th}$ column, and $v_{i}$ be the $i^\mathrm{th}$ element in the vector $v$.
Using the expansion $v^\top Q v = \sum_{i = 1}^{d_{x}}\sum_{j = 1}^{d_{x}} q_{ij}v_{i}v_{j}$ in the inequality above, the integral is simplified.
\begin{equation}
\begin{aligned}
\int_{v \in \mathbb{R}^n}v^\top Q v dF(v) & = \int_{v \in \mathbb{R}^{v}} \sum_{i = 1}^{d_{x}}\sum_{j = 1}^{d_{x}} q_{ij}v_{i}v_{j} dF(v) \\
& = \sum_{i = 1}^{d_{x}}\sum_{j = 1}^{d_{x}} q_{ij} \int_{v \in \mathbb{R}^{d_{x}} }v_{i}v_{j} dF(v)\\
& = \text{tr}(Q \Sigma_{x}).\\
\end{aligned}
\end{equation}
The quadratic chance constraint holds if~(\ref{eq:sdp_constraint}) is satisfied, as $\mathrm{Pr}((x-\mu_x)\in \mathcal{G}) \leq \frac{1}{c}\text{tr}(Q\Sigma_{x})$. Therefore,~(\ref{eq:sdp_constraint}) is a conservative deterministic approximation of the quadratic chance constraint $\mathrm{Pr}((x-\mu_x)^\top Q (x-\mu_\mathbf{x}) \geq c) \leq \epsilon$ i.e., $\mathcal{X}_\mathrm{CQCC} \subseteq \mathcal{X}_\mathrm{QCC}$. Note that if $\epsilon$ is a design variable, the approximation can be made tight by solving an inner semi-definite program following the approach in~\cite{vandenberghe2007generalized}. \end{proof}}
\begin{corollary}
The constraint set $\frac{1}{c}\text{tr}(A\Sigma_x) + \frac{1}{c}(\mu_x^\top A \mu_x )\leq \epsilon$ is a conservative approximation of the quadratic chance constraint $\mathrm{Pr}(x^\top A x \geq c) \leq \epsilon\}$, where $A \in \mathbb{R}^{n\times n}$ is a positive definite matrix and $c \in \mathbb{R}^{++}$ and $\Sigma_x$ is the co-variance of the random variable $x$.
\end{corollary}
\begin{proof}
\changed{The proof follows from~\cref{lemma:quad_const} by transforming $x$ to $x+ \mu_x$ in \Cref{eq:quad_chnc_cons}.}
\end{proof}
\noindent c) \textit{Joint Chance Constraints (JCC)}~\cite{zymler2013distributionally}:
The distributionally-robust joint chance constraint (DRJCC) for a polytope set is defined as
$\inf_{x(t) \sim (\mu_x,\Sigma_{x})} \mathrm{Pr}(\land_{i = 1}^{m} a_{i}^\top x + b_{i} \leq 0) \geq 1- \epsilon$. The joint constraints are split into multiple single chance constraints using Bonferroni's inequality~\cite{zymler2013distributionally} method as follows:
\begin{equation}
\begin{aligned}
& \inf_{x(t) \sim (\mu_x,\Sigma_{x})} \mathrm{Pr}(\land_{i = 1}^{m} a_{i}^\top x + b_{i} \leq 0) \geq 1- \epsilon \\ & \iff \sup_{x(t) \sim (\mu_x,\Sigma_{x})} \mathrm{Pr}( \lor_{i = 1}^{m} a_{i}^\top x + b_{i} \geq 0) \leq \epsilon \\ & \subseteq \sum_{i = 1}^{m} \sup_{x(t) \sim (\mu_x,\Sigma_{x})} \mathrm{Pr}( a_{i}^\top x + b_{i} \geq 0) \leq \epsilon.
\end{aligned} \label{eq:benf_inq}
\end{equation}
If the probability distribution of $x$ is Gaussian, then the JCC are split using Boole's inequality~\cite{blackmore2010probabilistic}. The total risk measure $\epsilon$ is allocated between each of the chance constraints in the summation such the $\sum_{i = 1}^{m} \epsilon_{i} = \epsilon$ leading to $m$ individual DRCC of the following form.
\begin{equation}
\inf_{x(t) \sim (\mu_x,\Sigma_{x})} \mathrm{Pr}( a_{i}^\top x + b_{i} \leq 0) \geq 1-\epsilon_{i}
\label{eq:benf_inequality_indv}
\end{equation}
We follow a naive risk allocation approach by equally distributing the risk measure $\epsilon$ among the $m$ constraints such that $\epsilon_{i} = \frac{\epsilon}{m}$. Alternatively, optimal risk allocation~\cite{hiro2008} can be achieved using iterative optimization techniques. Using distributional robustness, \cref{prob:cc_stoc_opt_cntrl} is reformulated to the following~\cref{prob:dr_stoc_opt_cntrl}.
\begin{problem}
\label{prob:dr_stoc_opt_cntrl}
Distributionally-Robust Chance-Constrained Stochastic Nonlinear Optimal Control.
\begin{equation*}
\begin{aligned}
J_{\mathrm{DR-SNOC}}^{*} =& \underset{x(t),\Bar{u}(t)}{\min}
& & \scalebox{0.9}{$\mathbb{E} \left[\int_{t_{0}}^{t_{f}}J(x(t),\Bar{u}(t))dt + J_{f}(x(t_{f}))\right]$} \\
& \text{s.t.} & & \cref{eq:nl_stochastic_dynamics},~\cref{eq:lin_socc},~\cref{eq:sdp_constraint},~\cref{eq:control_limits}, and~\cref{eq:init_term_conditions}.
\end{aligned} \label{eq:dr_stoptprob}
\end{equation*}
\end{problem}
\textbf{Note:} Given a risk measure $\epsilon$, the constraints in~\cref{prob:dr_stoc_opt_cntrl} are a function of mean $\mu_x$ and covariance matrix $\Sigma_x$ of the state at any time $t$. \changed{While this enables fast computation of chance constraints, it reduces the feasible space $\mathcal{X}_{\mathcal{F}}$. We compute a distributionally robust deterministic subset of the stochastic set $\mathcal{X}_{\mathcal{F}}$ that is convex in gPC space, as discussed in the following~\cref{subsec:gPC-SCP-scp}. We present empirical evidence that using the distributional robustness approach does not lead to infeasibility in practical scenarios. This approach should be integrated with system design and modelling to ensure feasibility. We transform the SNOC problem into a DNOC problem by applying the generalized polynomial chaos expansion. This approach transforms the infinite-dimensional SNOC problem in both the state (stochastic state) and time to a DNOC problem with infinite dimension only in time.}
\subsection{Generalized Polynomial Chaos} \label{subsec:gpc}
The generalized Polynomial Chaos (gPC)~\cite{xiu2002wiener,xiu2009fast,boutselis2017stochastic} expansion theory is used to model uncertainty with finite second-order moments as a series expansion of orthogonal polynomials. The polynomials are orthogonal with respect to a known density function $\rho(\cdot)$. Consider the random vector $\xi$ with independent identically distributed (iid) random variables $\{\xi_{i}\}_{i=1}^{d_{\xi}}$ as elements. Each $\xi_{i} \sim \mathcal{N}(0,1)$ is normally distributed with zero mean and unit variance. The random vector $x(t)$, defined by the SDE in~\cref{eq:nl_stochastic_dynamics}, can be expressed as the following series\begin{equation}
\scalebox{1}{$x_{i}(t) = \sum_{j = 0}^{\infty} x_{ij}(t) \phi_{j}(\xi)$},
\label{eq:series_expan}
\end{equation}
where $x_{i}$ denote the $i^\mathrm{th}$ element in the vector $x \in \mathcal{X}$ and $x_{ij}$ is the $j^\mathrm{th}$ coefficient in the series expansion. The dimension $d_{\xi}$ is the sum of the number of random inputs in the SDE~\cref{eq:nl_stochastic_dynamics} and the number of random initial conditions. The functions $\phi_{j}(\xi)$ are constructed using the Hermite polynomial~\cite{xiu2002wiener} basis functions. The functions $\phi_{j}(\xi)$ are orthogonal with respect to the joint probability density function $\boldsymbol{\rho}(\xi) = \varrho(\xi_1)\varrho(\xi_2)\cdots\varrho(\xi_{d_{\xi}})$, where $\varrho(\xi_{k}) = \tfrac{1}{\sqrt{2\pi}} e^{\tfrac{-\xi_{k}^2}{2}}$. The choice of the orthogonal polynomials depends on the uncertainty model effecting the dynamics. We refer to~\cite{ghanem2003stochastic} for details on type and construction of the polynomials for different standard uncertainty models such as uniform, beta and Poisson distributions.
\begin{rmrk} \label{rmrk:P_gpc_and_ell}
The series expansion is truncated to a finite number $\ell+1$ as $x_{i} \approx \sum_{j = 0}^{\ell} x_{ij}(t) \phi_{j}(\xi)$ based on the maximum degree of the polynomials ($\mathrm{P}_{\mathrm{gPC}}$) required to represent the variable $x$. The minimum $\ell$ required to appropriately represent $x$ with uncertainty parameter $\xi \in \mathbb{R}^{d_{\xi}}$ is given by $\ell = \Big(\begin{smallmatrix} \mathrm{P}_{\mathrm{gPC}} + d_{\xi} \\ d_{\xi}\end{smallmatrix} \Big) - 1$.
\end{rmrk}
The coefficients $x_{ij}(t)$ are computed using the Galerkin projection given by the following equation:
\begin{equation}
x_{ij}(t) = \frac{\int_{\mathbb{D}} \boldsymbol{\rho}(\xi) x_i (t) \phi_{j}(\xi) d\xi}{\langle \phi_j(\xi),\phi_j(\xi) \rangle},\label{eq:galerkin_projec}
\end{equation}
where $\langle\phi_i(\xi),\phi_j(\xi)\rangle=\int_{\mathbb{D}}\boldsymbol{\rho}(\xi)\phi_i (\xi)\phi_j(\xi) d\xi$. For non-polynomial functions, the Galerkin projection is computed using the Stochastic Collocation~\cite{xiu2009fast} method as follows:
\begin{equation}
\int_{\mathbb{D}} \rho(\xi) x_i (t) \phi_{j}(\xi) d\xi \approx \sum_{k = 1}^{m} w_{k} x_i (t) \phi_{j}(n_{k}),
\label{eq:stoch_coll}
\end{equation}
where Gauss-Hermite quadrature is used to generate the nodes $n_{k}$ and the corresponding node weights $w_{k}$. In the following section, we derive an approximate nonlinear ordinary differential equation system for the SDE in~\cref{eq:nl_stochastic_dynamics} using gPC expansion and the Galerkin scheme. The DRCC are projected to the gPC coordinates $x_{ij}$ leading to convex constraints. \cref{lemma:conv_gpc,lemma:approximation_error} discuss the convergence of the gPC expansion to the true distribution and the error due to truncated polynomial approximation of a distribution.
\begin{figure}\begin{center}
\includegraphics[width=0.4\columnwidth]{figures_tro/Beta.pdf}\includegraphics[width=0.4\columnwidth]{figures_tro/Exponential.pdf}
\includegraphics[width=0.4\columnwidth]{figures_tro/Gaussian.pdf}\includegraphics[width=0.4\columnwidth]{figures_tro/Bimodal.pdf}
\caption{Example gPC approximation of some standard probability distribution functions (PDF) using gPC expansion. For the beta and exponential distributions, gPC expansion represents the PDF well with just second order $P_{\mathrm{gPC}} = 2$ and $d_{\xi} = 1$ approximation. For a Gaussian distribution, the gPC representation is exact.}\label{fig:gpc_examples}\end{center}
\end{figure}
\begin{lemma}\label{lemma:conv_gpc}(Cameron-Martin Theorem~\cite{cameron1947orthogonal}) The gPC series approximation in~\cref{eq:series_expan} converges to the true value $x_{i} \in \mathcal{L}_2$.
\begin{equation}
\scalebox{1}{$\|x_{i}(t) - \sum_{j = 0}^{\ell} x_{ij}(t) \phi_{j}(\xi) \|_{\mathcal{L}_2}$} \to 0,\ \text{as} \ \ell \ \to \ \infty \ \forall \ t \in [t_{0},t_{f}]
\end{equation}
\end{lemma}
\begin{rmrk} \label{rmrk:exp_var_gpc} The expectation $\mathbb{E}(x_{i})$ and variance $\Sigma_{x_{i}}$ of the random variable $x_{i}$ can be expressed in terms of the coefficients of the expansion as follows:
\begin{equation}
\mathbb{E}(x_{i}) = x_{i0}, \quad \Sigma_{x_{i}} \approx \sum_{j = 1}^{\ell} x_{ij}^2 \langle \phi_{j},\phi_{j} \rangle \ \text{as} \ \ell \to \infty.
\label{eq:exp_var}
\end{equation}
\end{rmrk}
\begin{lemma}\label{lemma:approximation_error} (Truncation Error Theorem~\cite{muhlpfordt2017comments}) If an element $x_i$ of the random variable $x$ is represented using $\ell$ polynomials, then the approximation error is given as follows:
\begin{align}
\|x_i -\sum_{j = 0}^{\ell} x_{ij}(t) \phi_{j}(\xi)\| = \|e_\ell\| \leq \sqrt{\sum_{j = \ell+1}^{\infty}x^2_{ij}\|\phi_{j}\|^2}.
\end{align}
\end{lemma}
\cref{lemma:conv_gpc,lemma:approximation_error}, and~\cref{rmrk:exp_var_gpc} will be used in studying the convergence of the gPC approximation of the cost function, the SDE, and the chance constraints. Furthermore, the higher-order moments can be expressed as a polynomial function of the coefficients.
\subsubsection*{Curse of Dimensionality} The truncated polynomial expansion is a finite-dimensional approximation of the random variable. The number of polynomials $\ell$ grow exponentially large based on the degree of polynomial used to represent the state distribution. The large dimensionality can be reduced, inducing sparsity in the gPC expansion, by using techniques like sparse gPC~\cite{blatman2011adaptive}, and data-driven gPC~\cite{oladyshkin2012data}. A cost-effective approach to estimate moments up to second order is to use gPC polynomials up to degree 2, i.e., $\mathrm{P}_{\mathrm{gPC}} = 2$~\cite{xu2018novel}. For a given $\mathrm{P}_{\mathrm{gPC}}$, we use~\cref{rmrk:P_gpc_and_ell} to compute $\ell$. The computationally complexity for $\mathrm{P}_{\mathrm{gPC}} = 2$ is equivalent to linear covariance propagation. Note that, unlike the linear covariance propagation method, the gPC method with $\mathrm{P}_{\mathrm{gPC}}=2$ accounts for the coupling between the state $x$ and the white-noise process $dw$.
\section{Deterministic surrogate of the SNOC Problem} \label{sec:det_opt_prob}
The stochastic nonlinear optimal control problem discussed in~\cref{subsec:stoc_opt_prob} is reformulated in terms of the coefficients of the gPC expansion, with decision variables as the gPC coefficients and the control $\Bar{u}$. In the following, we discuss the existence and uniqueness of a solution to the coupled Ordinary Differential Equations (ODE) obtained form gPC approximation of SDE, the cost function in the gPC space, and present the convex constraints for the gPC coefficients obtained from deterministic approximation of chance constraints. We present the convergence and feasibility theorem of the approximation at the end of this section.
\subsection{Deterministic ODE Approximation of the SDE}
The gPC expansion in~\cref{eq:series_expan} is applied for all the elements in the vector $x \in \mathcal{X} \subseteq \mathbb{R}^{d_{x}}$ and the matrix representation using Kronecker product is given in the following, where $\mathrm{X} = \begin{bmatrix}x_{10} &\cdots &x_{1\ell} & \cdots & x_{d_{x}0} & \cdots& x_{d_{x}\ell}\end{bmatrix}^\top$ are gPC states.
\begin{equation}
\Phi(\xi) = \begin{bmatrix} \phi_{0}(\xi) & \cdots & \phi_{\ell}(\xi)\end{bmatrix}^\top
\label{eq:phi_gpc_matrix}
\end{equation}
\begin{equation}
x \approx \Bar{\Phi} \mathrm{X} ;\ \text{where}\ \Bar{\Phi} = \mathbb{I}_{d_{x}\times d_{x}} \otimes \Phi(\xi)^\top
\label{eq:state_kronck_notation}
\end{equation}
Consider the following Ito's integral form of the SDE in~\cref{eq:nl_stochastic_dynamics}.
\begin{equation}
x(t) = x(t_0) + \int_{t_0}^{t} f(x,\bar{u}) dt + \int_{t_0}^{t} g(x, \bar{u})dw
\label{eq:ito_nl_sde}
\end{equation}
The gPC projection of the above SDE is given by the following ODE.
\begin{align}
\hspace{-5pt}x_{ij}(t) & = x_{ij}(t_{0}) + \int_{t_{0}}^{t}\Bar{f}_{ij}(\mathrm{X},\Bar{u})dt + \int_{t_{0}}^{t}\Bar{g}_{ij}(\mathrm{X},\Bar{u})\sqrt{dt} \label{eq:continuous_ode_form}\\
&\Bar{f}_{ij} = \frac{\int_{\mathbb{D}}\rho(\xi)\phi_{j}(\xi)f_{i}(\Bar{\Phi}\mathrm{X},\Bar{u})d\xi}{\langle \phi_j(\xi),\phi_j(\xi) \rangle}, \nonumber\\
&\Bar{g}_{ij} = \frac{\int_{\mathbb{D}}\rho(\xi)\phi_{j}(\xi)g_{i}(\Bar{\Phi}\mathrm{X},\Bar{u}) \xi d\xi}{\langle \phi_j(\xi),\phi_j(\xi) \rangle} \nonumber
\end{align}
The dynamics of the coefficients $x_{ij}$ with the above notation is given in~\cref{eq:ode_form}, where: $f_{i}$ and $g_{i}$ are the $i^\mathrm{th}$ element of the vector $f$ and $i^\mathrm{th}$ row of the matrix $g$ respectively. We use the Euler-Maruyama discretization method of the SDE for time integration. The discrete time stochastic dynamics is given as follows:
\begin{equation}
x[k+1] = x[k] + f(x[k],\bar{u}[k])\Delta t + g(x[k],\bar{u}[k]) \sqrt{\Delta t} \xi,
\label{eq:discrete_sde}
\end{equation}
where $x[k]$, $\bar{u}[k]$ are the states and controls at time step $k$, $\Delta t$ is the integration time interval, and $\xi$ is a multivariate Gaussian distribution $\mathcal{N}(0,\mathbb{I})$. The discrete stochastic system is projected to a discrete deterministic system using the gPC method.
\begin{align}
x_{ij}[k+1] & = x_{ij}[k] + \Bar{f}_{ij}(\mathrm{X}[k],\Bar{u}[k])\Delta t \nonumber\\&+ \Bar{g}_{ij}(\mathrm{X}[k],\Bar{u}[k])\sqrt{\Delta t} \label{eq:ode_form}
\end{align}
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{figures_tro/gpc_mc_v1.jpeg}
\caption{Example gPC propagation for a pendulum. The figure compares the mean and $2\sigma$ confidence computed using gPC Projection~\cref{eq:discrete_sde} ($P_{\mathrm{gPC}}=1$), linear covariance propagation, and Monte Carlo (MC) propagation of the simple pendulum dynamics $\ddot{\theta} = -\sin{\theta} -0.8\dot{\theta} + \sqrt{0.001}\xi(t)$. It is observed that the gPC approximation overestimates the variance compared to MC and the linear covariance propagation underestimates the variance. The $P_{\mathrm{gPC}}=1$ projection corresponds to a Gaussian approximation that includes the cross correlation between the state and uncertainty.}
\label{fig:gpc_propgation}
\end{figure}
The full nonlinear discrete time ODE with the stacked vector $\mathrm{X}$ is given as follows:
\begin{align}
\mathrm{X}[k+1] =& \mathrm{X}[k] + \Bar{f}(\mathrm{X}[k],\Bar{u}[k],\Delta t) \nonumber\\
&+\Bar{g}(\mathrm{X}[k],\Bar{u}[k],\sqrt{\Delta t}).\label{eq:odefull_form}
\end{align}
\Cref{fig:gpc_propgation} shows an example of propagation using~\cref{eq:odefull_form}. While not discussed in this paper, the projection is also applicable to a higher-order discretization methods~\cite{platen2010numerical}. The sequential convex programming method used for trajectory optimization involves successive linearizations~\cite{morgan2014model} of the dynamics about a given trajectory and discretization for time integration. In~\cref{prop:exist_uniq}, we present the conditions for existence and uniqueness of the solution to the projected system. The existence and uniqueness of solution to the ODE surrogate ensure convergence of any Picard iteration scheme used for integration.
\begin{prop} \label{prop:exist_uniq}
The ODE system~\cref{eq:ode_form} obtained using gPC approximation of the SDE has a solution and the solution is unique, for a given initial condition, assuming that the SDE satisfies the existence and uniqueness conditions in~\cref{eq:lipshitz_stoc},~\cref{eq:growth_stoc} and the expectation
\begin{equation}
\begin{aligned}
K_{g_{ij}} & = \frac{K}{k_{j}} \mathbb{E}(L_{g_{j}} (\xi)), K_{f_{j}} & = \frac{K}{k_{j}}\mathbb{E}(L_{f_{j}}(\xi))
\end{aligned} \label{eq:bound_lipz}
\end{equation}
in~\cref{eq:bound_lipz} are bounded for each $j = 0,1,\cdots,\ell$, where:$k_{j} = \langle \phi_{j},\phi_{j} \rangle$, $L_{f_{j}}(\xi) = |\phi_{j}(\xi)|\|\big[\begin{smallmatrix}\Bar{\Phi} & 0 \\ 0 & \mathbb{I} \end{smallmatrix}\big]\|_{2}$, $L_{g_{j}} (\xi)= L_{f_{j}}(\xi) |\phi_{1}(\xi)|$. The constants $K_{g_{ij}}$ and $K_{f_{j}}$ are the Lipschitz coefficients of the projected functions $\bar{g}_{ij}$ and $\bar{f}_{j}$ respectively.
\end{prop}
\begin{proof}See Proposition 2 in~\cite{nakka2019nsoc} for the proof.\end{proof}
While the projection operation preserves the existence and uniqueness properties of the SDE, it might not retain the controllability of the moments of the system. Following examples discuss on how the $\epsilon_c$ controllability in~\cref{defn:stoc_controllability} of the SDE effects the controllablility of the projected ODE system.
\begin{example} \label{example:linear_example_1}
Consider the linear SDE $dx = xdt + \bar{u}dt + \sqrt{dt}\xi$, where $x \in \mathbb{R}^{1}$, $\bar{u} \in \mathbb{R}^{1}$ and $\xi \sim \mathcal{N}(0,1)$. Using the random variable $\xi$ as the variable, we can construct the first order gPC expansion $x = x_{0} + x_{1}\xi$ of the state with $x_{0},x_{1} \in \mathbb{R}^{1}$. The projected dynamics using the expansion is given as follows:
\begin{equation}
\begin{bmatrix}
dx_{0} \\ dx_{1}
\end{bmatrix} = \begin{bmatrix}
1 & 0 \\ 0 & 1 \end{bmatrix} \begin{bmatrix} x_{0} \\ x_{1} \end{bmatrix} dt + \begin{bmatrix}
1 \\ 0 \end{bmatrix} \bar{u} dt + \begin{bmatrix} 0 \\ \sqrt{dt} \end{bmatrix}.
\label{eq:linear_sde_example}
\end{equation}
The dynamics of $x_{1}$ is decoupled from $x_{0}$ and the propagation is not influenced by the control $\bar{u}$. Notice that, even though the original SDE ($dx=xdt+\bar{u}dt$) is controllable, the projected system~\cref{eq:linear_sde_example} is not fully controllable. The projection operation converts a SDE to an ODE in higher dimensions. Though this operation enables for fast uncertainty propagation, the linear projected system is underactuated and not fully controllable.
\end{example}
\begin{rmrk} \label{remark:feedback_covariance_control}
Using a stochastic state feedback of the form $u = -kx$ in~\cref{example:linear_example_1}, we get the closed-loop SDE $dx=(1-k)xdt + \sqrt{dt}\xi$. The gPC projection of the system is as follows:
\begin{equation}
\begin{bmatrix}
dx_{0} \\ dx_{1}
\end{bmatrix} = \begin{bmatrix}
1-k & 0 \\ 0 & 1-k \end{bmatrix} \begin{bmatrix} x_{0} \\ x_{1} \end{bmatrix} dt + \begin{bmatrix} 0 \\ \sqrt{dt} \end{bmatrix}.
\label{eq:linear_sde_randomfeedback_example}
\end{equation}
Using a stochastic feedback, the state $x_{1}$ that corresponds to the variance of the SDE can be controlled.
\end{rmrk}
\begin{example}\label{example:nonlinear_example_2} The gPC projection of the nonlinear SDE $dx = x^2 dt + \sqrt{dt}\xi$ using the expansion $x = x_{0} + x_{1}\xi$ is given as follows:
\begin{equation}
\begin{bmatrix}
dx_{0} \\ dx_{1}
\end{bmatrix} = \begin{bmatrix}
x^2_{0} + x^2_{1}+ \bar{u}\\ 2x_{0}x_{1} \end{bmatrix}dt + \begin{bmatrix} 0 \\ \sqrt{dt} \end{bmatrix}.
\label{eq:nonlinear_sde_example}
\end{equation}
The projected system~\cref{eq:nonlinear_sde_example} is underactuated. In the case of nonlinear systems, the coupling between the dynamics of $x_0$ and $x_1$ allows for indirectly controlling the state $x_1$.
\end{example}
\begin{rmrk}\label{remark:control} The gPC projected ODE system in~\cref{eq:ode_form} might not be fully controllable as discussed in~\cref{example:linear_example_1}. We introduce a slack variable terminal constraint on the variance of the state variable to ensure the feasibility of~\cref{prob:cc_stoc_opt_cntrl} in accordance with~\cref{defn:stoc_controllability}. Based on the value of the slack variable, this would increase (or) decrease the probability of reaching the terminal set.
\end{rmrk}
With~\cref{remark:control} on the controllability of the projected system, we proceed to construct a finite-dimensional approximation of the cost functional and chance-constraints to formulate the convex-constrained nonlinear deterministic optimal control problem.
\subsection{Cost Function}
\label{sec:cost_functioj}
Using the notation in~\cref{eq:state_kronck_notation}, the expectation of the cost functional in~\cref{eq:cost_stopt} is expressed in the gPC coefficients as follows:
\begin{align}
\begin{aligned}
J_{\mathrm{gPC}}(\mathrm{X}(t),\Bar{u}(t))&= \mathrm{X}(t)^\top Q_{\mathrm{gPC}} \mathrm{X}(t) + \|\bar{u}\|_{p},\\
J_{\mathrm{gPC}_{f}}(\mathrm{X}(t_{f}))&=\mathrm{X}(t_f)^\top Q_{\mathrm{gPC}_{f}} \mathrm{X}(t_f),
\end{aligned}\label{eq:cost_odeopt}
\end{align}
where $Q_{\mathrm{gPC}} = \mathbb{E}(\Bar{\Phi}^\top Q \Bar{\Phi})$ and $Q_{\mathrm{gPC}_{f}} = \mathbb{E}(\Bar{\Phi}^\top Q_{f} \Bar{\Phi})$. Since the gPC projection is a canonical transformation, we can prove that the projected matrix $Q_{\mathrm{gPC}}$ is positive definite.
\begin{prop}\label{prop:positive_definite_cost}
The expectation matrix $\mathbb{E}(\Bar{\Phi}^\top \Bar{\Phi})$ is a positive definite matrix.
\end{prop}
\begin{proof}
We can prove the following equality by expanding the matrix multiplication.
\begin{equation}
\mathbb{E}(\Bar{\Phi}^\top \Bar{\Phi}) = \mathbb{I} \otimes \mathbb{E}(\Phi \Phi^\top) \label{eq: }
\end{equation}
The block matrix $\mathbb{E}(\Phi \Phi^\top)$ is positive definite as the functions $\phi_i$ used to construct the column vector $\Phi$ are orthogonal with respect to the density function $\rho$. Therefore, $\mathbb{E}(\bar{\Phi}^\top \bar{\Phi})$ is positive definite, since $\mathbb{E}(\Phi \Phi^\top)$ is positive definite.
\end{proof}
\begin{lemma}\label{lemma:cost_gpc}
If $Q$ is a positive definite matrix, then the expectation $Q_{\mathrm{gPC}} = \mathbb{E}(\Bar{\Phi}^\top Q \Bar{\Phi})$ is a positive definite matrix.
\end{lemma}
\begin{proof}
Since $Q$ is a positive definite matrix, we have $Q \succcurlyeq \lambda_{\min} (Q)\mathbb{I}$ where $\lambda_{\min} (Q) >0$. The expectation $\mathbb{E}(\bar{\Phi}^\top Q \Bar{\Phi})$ can be lower bounded as follows:
$
\mathbb{E}(\bar{\Phi}^\top Q \Bar{\Phi}) \succcurlyeq \mathbb{E}(\bar{\Phi}^\top \lambda_{\min} (Q) \mathbb{I} \Bar{\Phi})
\succcurlyeq \lambda_{\min} (Q). \mathbb{E}(\bar{\Phi}^\top\Bar{\Phi})$.
Using~\cref{prop:positive_definite_cost}, we conclude that $\mathbb{E}(\bar{\Phi}^\top Q \Bar{\Phi})$ is a positive definite matrix.
\end{proof}
\changed{\begin{corollary}
\label{coroll:cost_gpc}
If the polynomials $\phi_{j}$ used for gPC projection are Hermite polynomials, then $\mathbb{E}(\bar{\Phi}^\top Q \Bar{\Phi}) \succcurlyeq \lambda_{\min}(Q) \mathbb{I}$.
\end{corollary}
\begin{proof}
From~\cref{lemma:cost_gpc}, we have $\mathbb{E}(\bar{\Phi}^\top Q\bar{\Phi}) \succcurlyeq \lambda_{\min}(Q) \mathbb{E}(\bar{\Phi}^\top \bar{\Phi}).$ Using the gPC expansion in~\cref{eq:state_kronck_notation}, we have
\begin{equation}
\begin{aligned}
\lambda_{\min}(Q)& \mathbb{E}(\bar{\Phi}^\top \bar{\Phi}) \\ &= \lambda_{\min}(Q) \mathbb{E}(\left(\mathbb{I} \otimes \Phi(\xi) \right)\left(\mathbb{I} \otimes \Phi(\xi)\right)^\top) \\
& \succcurlyeq \lambda_{\min}(Q) \mathbb{I} \Big ( \min \{\mathbb{E}(\phi^2_0),\mathbb{E}(\phi^2_1),\dots,\mathbb{E}(\phi^2_{d_x})\} \Big)
\end{aligned} \nonumber
\end{equation}
For the Hermite polynomials~\cite{xiu2009fast}, $ \min \{\mathbb{E}(\phi^2_0),$ $\mathbb{E}(\phi^2_1),\dots,\mathbb{E}(\phi^2_{d_x})\} = \mathbb{E}(\phi^2_0) = 1 $. Hence, we have the inequality $\mathbb{E}(\bar{\Phi}^\top Q \Bar{\Phi}) \succcurlyeq \lambda_{\min}(Q) \mathbb{I}$.
\end{proof}
\cref{lemma:cost_gpc} and \cref{coroll:cost_gpc} prove that the projected cost functional in the gPC space is positive-definite. In the following example, we illustrate the projection of a quadratic cost function using the gPC method.}
\changed{\begin{example}\label{example:quadratic_cost}
Consider the quadratic cost, $J = \mathbb{E}(x q_0 x)$, where $x,q_0 \in \mathbb{R}^1$ and $q_0$ and $x$ is a random variable. Let us represent the the random variable $x$ using the Hermite polynomials with single uncertainty $\xi \sim \mathcal{N}(0,1)$ upto order 2 as follows:
\begin{equation}
x = x_0 + x_1\xi + +x_2 (\xi^2 - 1).\label{eq:example_x_expansion}
\end{equation}
The projected cost $J_\mathrm{gPC}$ in terms of the gPC state $X = [x_0, x_1, x_2]^\top$ is given as follows:
\begin{equation}
\begin{aligned}
\mathbb{E}(x q_0 x) & = X^\top \mathbb{E} \Big([1,\xi,\xi^2-1]^\top q_0 [1,\xi,\xi^2-1] \Big) X \\
& = X^\top \begin{bmatrix}
q_0 \mathbb{E}(1) & \mathbb{E}(\xi) & \mathbb{E}(\xi^2-1) \\ \mathbb{E}(\xi) & \mathbb{E}(\xi^2) & \mathbb{E}(\xi(\xi^2-1)) \\ \mathbb{E}(\xi^2-1) & \mathbb{E}(\xi(\xi^2-1)) & \mathbb{E}((\xi^2-1)^2)
\end{bmatrix} X
\end{aligned} \nonumber
\end{equation}
The expectation operation is with respect to a normal probability distribution. Additionally, due to the orthogonality of the Hermite polynomials we have only diagonal terms. The projected cost is of the following quadratic form.
\begin{equation}
\mathbb{E}(x q_0 x) = X^\top \Big[\begin{smallmatrix}
q_0 & 0 & 0 \\ 0 & q_0 & 0 \\ 0 & 0 & 2q_0
\end{smallmatrix} \Big] X
\end{equation}
Note that the projected cost is also quadratic and positive definite as proved in~\cref{lemma:cost_gpc} and~\cref{prop:positive_definite_cost}.
\end{example}}
\subsection{Convex Approximation of the Chance Constraint}
The deterministic approximations of the chance constraints discussed in~\cref{subsec:stoc_opt_prob} are expressed in terms of the gPC coefficients that define a feasible set for the deterministic optimal control problem with gPC coefficients as decision variables.
\begin{lemma}\label{lemma:lin_socc_gpc}
The second-order cone constraint given below
\begin{align}
(a^\top \otimes M) \mathrm{X} + b + \sqrt{\frac{1-\epsilon}{\epsilon}}\sqrt{\mathrm{X}^\top U N N^\top U^\top \mathrm{X}} \leq 0
\label{eq:socc_gpc}
\end{align}
is equivalent to the deterministic approximation of the DRLCC in~\cref{eq:drcc_lin} as $\ell \to \infty.$, where the matrices $M,U,N$ are given by
\begin{equation}
\begin{aligned}
M & = \begin{bmatrix}1 & 0 & \cdots & 0 \end{bmatrix}_{1 \times (\ell+1)} \\
U & = \begin{bmatrix}
a_{1} &0 & 0\\
0& \ddots & 0\\
0&0 & a_{d_{x}}
\end{bmatrix} \otimes \mathbb{I}_{(\ell+1)\times(\ell+1)} \\
N & = \mathds{1}_{d_{x}\times d_{x}} \otimes \mathds{H} ; \ \mathds{H} = \begin{bmatrix}
0 & \mathbb{O} \\
\mathbb{O} & \sqrt{\mathbb{E}(HH^\top)}
\end{bmatrix} \\
\text{where} \ & H = \begin{bmatrix} \phi_{1}(\xi) & \cdots & \phi_{\ell}(\xi)\end{bmatrix}^\top
\end{aligned} \label{eq:mat_lin_con_gpc}
\end{equation} and $\mathds{1}$ is a matrix with entries as $1$.
\end{lemma}
\begin{proof} It is sufficient to prove that $(a^\top \otimes M) \approx a^\top \mu_x$ and $\mathrm{X}^\top U NN^\top U^\top \mathrm{X} \approx a^\top \Sigma_{x} a $ as $\ell \to \infty$. Invoking Lemma~\ref{lemma:conv_gpc} and Remark~\ref{rmrk:exp_var_gpc}, the polynomials of gPC coefficients can be replaced by mean and variable of the variable $x$.
\begin{align}
\begin{aligned}
(a^\top \otimes M) \mathrm{X}& = \begin{bmatrix} a_{1}M & a_{2}M&\cdots &a_{d_{x}}M \end{bmatrix}\mathrm{X} \\
& = a_{1}x_{10} + a_{2}x_{20} +\cdots +a_{d_{x}}x_{dx0} \\
& \approx a^\top \mu_x
\end{aligned}\label{eq:mean_eq}
\end{align}
\noindent Equation~(\ref{eq:mean_eq}) shows the steps involved to prove $(a^\top \otimes M) \approx a^\top \mu_x$. Let us define a vector $p_{i} = \begin{bmatrix}x_{i0} & \Bar{p}^\top_{i} \end{bmatrix}^\top$ where $\Bar{p}_{i} = \begin{bmatrix} x_{i1} & \cdots & x_{i\ell} \end{bmatrix}^\top$. \begin{equation}
U^\top \mathrm{X} = \begin{bmatrix} a_{1}p^\top_{1} & a_{2}p^\top_{2} & \cdots & a_{d_{x}}p^\top_{d_{x}} \end{bmatrix}^{\top} \label{eq:expnd1}
\end{equation}
\begin{align}
NN^\top U^\top \mathrm{X} =& \begin{bmatrix} \mathds{H} a_{1}p_{1}& \mathds{H} a_{2} p_{2}& \cdots & \mathds{H} a_{d_{x}}p_{d_{x}}\end{bmatrix}\label{eq:expnd2}\\
\hspace{-30pt}\mathrm{X}^\top U NN^\top U^\top X & = \sum_{i = 1}^{d_{x}}\sum_{j = 1}^{d_{x}} a_{i}a_{j}p^\top_{i}\mathds{H}p_{j} \label{eq:expnd3}\\ &= \sum_{i = 1}^{d_{x}}\sum_{j = 1}^{d_{x}} a_{i}a_{j}\Bar{p}^\top_{i}\mathbb{E} (HH^\top)\Bar{p}_{j} \approx a^\top \Sigma_{x} a \nonumber
\end{align}
Using this notation, the matrices in~(\ref{eq:socc_gpc}) are expanded as shown in~(\ref{eq:expnd1}),~(\ref{eq:expnd2}), and~(\ref{eq:expnd3}). Therefore, the equivalence is proved by Lemma~\ref{lemma:conv_gpc} as $\ell \to \infty$. \end{proof}
In the following example, we show the projection of a one-dimensional linear constraint.
\changed{\begin{example}\label{example:}
Consider the linear chance constraint,$\mathrm{Pr}(ax+b\leq0) \geq 1-\epsilon$,
where $a,b \in \mathbb{R}^1$ and $x\in \mathbb{R}^1$ is a random variable. Let us express $x$ using Hermite polynomials upto order 2, $x= [1,\xi,\xi^2-1]X$, where $\xi \sim \mathcal{N}(0,1)$ and $X = [x_0, x_1, x_2]^\top$. We can express the above linear chance constraint as an inequality constraint using~\cref{lemma:lin_consev_approx,lemma:lin_socc_gpc}, and the matrices $M, N, U$ defined in~\cref{eq:socc_gpc} are computed as:
\begin{equation}
\begin{aligned}
& a^\top \otimes M = a \begin{bmatrix}1 & 0 &0\end{bmatrix},\ \ U = a \otimes \mathbb{I}_{3\times 3},\\& N = 1 \otimes \begin{bmatrix}
0 & 0 & 0\\ 0 & \sqrt{\mathbb{E}(\xi^2)} & 0 \\ 0 & 0 & \sqrt{\mathbb{E}((\xi^2-1)^2)}
\end{bmatrix} = \begin{bmatrix}
0 & 0 & 0\\ 0 & 1 & 0 \\ 0 & 0 & \sqrt{2}
\end{bmatrix}.
\end{aligned}\nonumber
\end{equation}
Then, the deterministic constraint \cref{eq:socc_gpc} becomes
\begin{equation}
\begin{aligned}
& [a,0,0] X + b + \sqrt{\tfrac{1-\epsilon}{\epsilon}} \sqrt{X^\top \Big[\begin{smallmatrix}
0 & 0 & 0\\ 0 & 1 & 0 \\ 0 & 0 & 2\end{smallmatrix}\Big] X } \leq 0.
\end{aligned}
\end{equation}
The second-order cone constraint enables the use of convex programming, where the deterministic constraint $a\mu_x + b + \sqrt{ a\sigma_x a} \leq 0$ is non-convex in terms of the mean $\mu_x$ and variance $\sigma_x$ of the random variable $x$.
\end{example}}
\begin{lemma} \label{lemma:sdp_con_gpc} The quadratic inequality
\begin{equation}
\sum_{i = 1}^{d_{x}} \sum_{k = 1}^{\ell} a_{i}\langle \phi_{k},\phi_{k}\rangle x_{ik}^2 \leq \epsilon c,
\label{eq:sdp_con_gpc}
\end{equation}
expressed in terms of the gPC coefficients is equivalent to the constraint in~\cref{eq:sdp_constraint} as $\ell \to \infty$, where $A$ is a diagonal matrix with $i^\mathrm{th}$ diagonal element as $a_{i}$ and $\langle\phi_k,\phi_k\rangle=\int_{\mathbb{D}}\boldsymbol{\rho}(\xi)\phi_k\phi_k d\xi$.
\end{lemma}
\begin{proof}The deterministic approximation, $\mathrm{tr}(A\Sigma_{x}) \leq c \epsilon$, of the QCC in~\cref{eq:quad_chnc_cons} can be expanded as follows.
\begin{equation}
\begin{aligned}
\mathrm{tr}(A\Sigma_{x}) \leq c \epsilon & \equiv \sum_{i = 1}^{d_{x}} a_{i} \mathbb{E}((x_{i}-\mu_{x_i})^\top (x_{i}-\mu_{x_i})) \leq c \epsilon\\
& \equiv \sum_{i = 1}^{d_{x}} \sum_{j = 1}^{\ell} a_{i} \langle\phi_{j},\phi_{j} \rangle x^2_{ij} \leq c \epsilon
\end{aligned}\label{eq:proof_sdp}
\end{equation}The equivalence is proved by directly expanding the trace and using Remark~\ref{rmrk:exp_var_gpc} as shown in~(\ref{eq:proof_sdp}).\end{proof}
\begin{figure}
\centering
\includegraphics[width=1.\columnwidth,height=2.5in]{figures_tro/tro_theorem1_v3.pdf}
\caption{Illustration of the gPC projection method. We use the gPC projection method to derive a deterministic surrogate of the chance-constrained optimal control problem. We use Lemmas 1 to 7 and Propositions 1 and 2 to prove Theorem 1.}
\label{fig:projection_lemmas}
\end{figure}
\changed{
\begin{example}
Consider the one-dimensional quadratic constraint $\mathrm{Pr}(a(x-\mu_x)^2\leq c) \geq1-\epsilon$. where $x \in \mathbb{R}^1$ is a random variable with mean $\mu_x$, $a, c \in \mathbb{R}^1$. Let us express $x$ using Hermite polynomials up to order 2, $x= [1,\xi,\xi^2-1]X$, where $\xi \sim \mathcal{N}(0,1)$ and $X = [x_0, x_1, x_2]^\top$. The quadratic chance constraint in terms of $\sigma_x$ can be expressed in $X$ as follows using~\cref{lemma:quad_const,lemma:sdp_con_gpc}:
\begin{equation}
\begin{aligned}
\mathrm{tr}(a\sigma_x) &= a \mathbb{E}( 1x^2_0 + \xi^2 x^2_1 + (\xi^2 -1)^2 x^2_2 )\\
& = a x^2_0 + a x^2_1 + 2a x^2_2 \leq c\epsilon .
\end{aligned}
\end{equation}
The semidefinite constraint $\mathrm{tr}(a \sigma_x)$ is transformed into a quadratic constraint in the gPC state $X$.
\end{example}}
Using the projected dynamics~\cref{eq:continuous_ode_form}, the cost functional in~\cref{eq:cost_odeopt}, and the linear and quadratic chance constraints~\cref{eq:socc_gpc},~\cref{eq:sdp_con_gpc} in the gPC coefficients (as shown in~\cref{fig:projection_lemmas}) we can formulate the following distributionally robust deterministic nonlinear optimal control problem with the gPC states $\mathrm{X}$ and the control $\bar{u}$ as decision variables.
\begin{problem}
\label{prob:dr_det_optimal_cntrl}
Distributionally Robust Deterministic Nonlinear Optimal Control Problem.
\begin{equation*}
\begin{aligned}
J_{\mathrm{gPC}}^{*} = & \underset{\mathrm{gPC}\mathrm{X}(t),\Bar{u}(t)}{\min}
& & \int_{t_{0}}^{t_{f}}J_{\mathrm{gPC}}(\mathrm{X}(t),\Bar{u}(t))dt + J_{\mathrm{gPC}_{f}}(\mathrm{X}(t_{f})) \\
& \text{s.t.}
& & \cref{eq:odefull_form},~\cref{eq:socc_gpc},~\cref{eq:sdp_con_gpc}\\
& & & \Bar{u}(t) \in \mathcal{U} \quad \forall t \in [t_{0},t_{f}]\\
& & & \mathrm{X}(t_{0}) = \mathrm{X}_{0} \quad \mathrm{X}(t_{f}) \in \mathcal{X}_{\mathrm{X}_{f}}
\end{aligned}
\end{equation*}
\end{problem}
\noindent where the projection of the initial condition $x_{0}$ in the gPC space is $\mathrm{X}_{0}$. The terminal set $\mathcal{X}_{\mathrm{X}_{f}}$ is constructed using a distributionally robust polytope (or) a conservative ellipsoid approximation of the set $x_{f}$ in~\cref{prob:cc_stoc_opt_cntrl} with probabilistic guarantees using \cref{lemma:lin_socc_gpc,lemma:sdp_con_gpc}. We make the following observations about the distributional robustness and gPC projection discussed above to transform~\cref{prob:cc_stoc_opt_cntrl} to~\cref{prob:dr_det_optimal_cntrl}.
\begin{itemize}
\item The infinite-dimensional optimal control problem in state space and time, as described in~\cref{prob:cc_stoc_opt_cntrl}, is projected to~\cref{prob:dr_det_optimal_cntrl} that is finite-dimensional in space and infinite-dimensional in time.
\item The ODE approximation of the SDE using the gPC projection diverges over a long-horizon problem (or) when the uncertainty affecting the SDE has large variance (or) when the uncertainty model has large gradients with respect to state and control. A multi-element gPC method can be used to overcome this divergence due to the finite-dimensional approximation. The structure of the proposed constraint reformulation is invariant to the multi-element gPC method.
\item The choice of the terminal set used in~\cref{prob:cc_stoc_opt_cntrl} is restricted due to the $\epsilon_{c}-$controllability of the SDE. We use slack variables on the terminal state variance to ensure the feasibility of both~\cref{prob:cc_stoc_opt_cntrl,prob:dr_det_optimal_cntrl}.
\item The projected cost functional preserves the positive definite property of the quadratic cost used in~\cref{prob:cc_stoc_opt_cntrl}.
\item \changed{For a given risk measure, the linear and quadratic chance constraints are second-order cone and semidefinite constraints in the gPC coefficients, respectively.}
\end{itemize}
\Cref{prob:dr_det_optimal_cntrl} (DNOC) enables the use of techniques such as the pseudospectral method and sequential convex programming for solving \cref{prob:dr_det_optimal_cntrl} (SNOC). We use sequential convex programming to solve~\cref{prob:dr_det_optimal_cntrl} and apply this technique to compute safe and optimal motion plans under uncertainty. We extend the gPC-SCP to formulate a predictor-corrector formulation for fast real-time planning under uncertainty.
\subsection{gPC-SCP: Generalized Polynomial Chaos-Based Sequential Convex Programming}\label{subsec:gPC-SCP-scp}
We formulate the gPC-SCP problem by constructing a sequential convex programming (SCP) approximation of~\cref{prob:dr_det_optimal_cntrl} with gPC state ($\mathrm{X}$) and control ($\bar{u}$) as decision variables. The convex program is then solved iteratively using an interior point method until a convergence criterion is satisfied and projected back to the probability space from the gPC space to compute a solution of~\cref{prob:dr_stoc_opt_cntrl}.
The SCP problem formulation involves two steps: 1) Discretizing the continuous-time optimal control problem into a discrete-time optimal control problem, and 2) Convexifing the non-convex constraints and cost function about a nominal initial state and control trajectory. Following this approach, the projected integral cost functional~\cref{eq:cost_odeopt}, the nonlinear dynamics~\cref{eq:odefull_form}, and the second-order cone constraint~\cref{eq:socc_gpc} and the semi-definite constraint~\cref{eq:sdp_con_gpc} are discretized using a first-order hold approach for $T$ time steps between the time horizon $[t_0,t_f]$ with gPC state and control as decision variables.
At iteration $i$, the cost functional, constraints~\cref{eq:socc_gpc} and~\cref{eq:sdp_con_gpc}, and feasible control set $\mathcal{U}$ are convex. The discretized gPC dynamics in~\cref{eq:odefull_form} is a nonlinear equality constraint at each time step. We convexify the nonlinear dynamics~\cref{eq:odefull_form} by linearizing it about the state and the control trajectory $S^{(i-1)} = \{\mathrm{X}^{(i-1)},\bar{u}^{(i-1)}\}$ computed at $(i-1)^{th}$ iteration. The linearized equations form a set of linear constraints on the state and control action as follows.
\begin{equation}
\begin{aligned}
\mathrm{X}^{(i)}[k+1] & = \mathrm{X}^{(i)}[k] + A^{(i)}[k] \mathrm{X}^{(i)}[k] + B^{(i)}[k] \Bar{u}^{(i)}[k]\\ & + Z^{(i)}[k], \quad \mathrm{where:} \ k \in \{1,\dots,T-1\}.
\end{aligned}\label{eq:linearize_discrete_dyn}
\end{equation}
\begin{align}
A^{(i)}& = \frac{\partial(\Bar{f} + \Bar{g})}{\partial \mathrm{X}} \Big|_{\scalebox{0.85}{$S^{(i-1)}$}}; \quad B^{(i)} = \frac{\partial(\Bar{f} + \Bar{g})}{\partial \Bar{u}} \Big|_{\scalebox{0.85}{$S^{(i-1)}$}} \nonumber
\end{align}
\begin{align}
Z^{(i)} & = \Bar{f}(S^{(i-1)},\Delta t)+\Bar{g}(S^{(i-1)},\sqrt{\Delta t}) \nonumber\\ & -A^{(i)} \mathrm{X}^{(i-1)}- B^{(i)}\Bar{u}^{(i-1)}\label{eq:linear_mat}
\end{align}
The gPC-SCP problem at iteration $i$, after discretization and convexification, is given below in~\cref{prob:dr_trajopt}.
\begin{problem}
\label{prob:dr_trajopt}
gPC-SCP: Generalized Polynomial Chaos-based Sequential Convex Programming.
\begin{align}
& \underset{\mathrm{X}^{(i)},\Bar{u}^{(i)}}{\min}
& &\hspace{-10pt}\sum_{k = 1}^{T-1} J_{\mathrm{gPC}}(\mathrm{X}^{(i)}[k],\Bar{u}^{(i)}[k])\Delta t + J_{\mathrm{gPC}_{f}}(\mathrm{X}^{(i)}[T]) \nonumber \\
& \textit{s.t.}& & \mathrm{Projected \ Dynamics: } \cref{eq:linearize_discrete_dyn} \nonumber\\
& & & \mathrm{Constraints:} ~\{\cref{eq:socc_gpc},~\cref{eq:sdp_con_gpc}\} \nonumber\\
& & & \Bar{u}^{(i)}[k] \in \mathcal{U} \quad \forall \ k \in \{1,\dots,T-1\} \nonumber\\
& & & \mathrm{X}^{(i)}[1] = \mathrm{X}_{0} \quad \mathrm{X}^{(i)}[T] \in \mathcal{X}_{\mathrm{X}_{f}} \nonumber\\
& & & \hspace{-30pt}\|\mathrm{X}^{(i)}[k]-\mathrm{X}^{(i-1)}[k]\|^2 _{2} \leq \alpha_{x}\beta \ \forall \ k \in \{1,\dots,T\} \label{eq:state_trust}\\
& & & \hspace{-30pt}\|\bar{u}^{(i)}[k]-\bar{u}^{(i-1)}[k]\|^2 _{2} \leq \alpha_{u}\beta \ \forall \ k \in \{1,\dots,T-1\}\label{eq:control_trust}
\end{align}
\end{problem}
\cref{prob:dr_trajopt} shows the SCP formulation at $i^{\mathrm{th}}$, given a nominal trajectory $S^{(i-1)} = \{\mathrm{X}^{(i-1)},\bar{u}^{(i-1)}\}$ computed at $(i-1)^\mathrm{th}$ iteration with the constraint set at each time step $k$ and at iteration $i$, where $\mathcal{X}_{\mathrm{X}_{f}}$ is the projected terminal constraint. The nominal trajectory $S^{0}$ at $i=1$, used to initialize gPC-SCP, is computed using a deterministic trajectory optimization for the nominal dynamics $\dot{x} = f(x,\bar{u})$, which ignores the uncertainty that affects the system. For the motion planning problem, the nominal trajectory $S^{0}$ is computed using kinodynamic motion planning algorithms such as asymptotically optimal and rapid exploration of random trees~\cite{hauser2016}.
An additional trust region constraint on the gPC state~\cref{eq:state_trust} and control~\cref{eq:control_trust} is used to ensure the convergence and feasibility of the SCP as $i \to \infty$, where $\alpha_x >0$, $\alpha_x >0$, and $\beta \in ( 0, 1)$. The choice of $\beta$ ensures the convergence of the trust region as the number of iterations increases. This acts as a convergence criterion, while ensuring that the search space is small. The trust region in the gPC state in~\cref{eq:state_trust} can be equivalently understood as a probabilistic constraint of the form $\mathrm{Pr}(\|x^{i}-x^{i-1}\| \leq \alpha_{xp}) \geq 1-\epsilon_{t}$, where $\alpha_{x}$ is a function of $\alpha_{xp}$ using the quadratic projection discussed in~\cref{lemma:sdp_con_gpc}. The SCP algorithm is known to converge to the KKT point of the DNOC problem under mild conditions. For a detailed analysis of convergence, see~\cite{morgan2016swarm,morgan2014model,morgan2012spacecraft}.
We ensure the feasibility of gPC-SCP: 1) by using stochastic reachable terminal sets, as discussed in~\cite{hewing2018stochastic}, that are constructed using the linearized approximation of the dynamics, and 2) by increasing the trust region in loop with $\beta>1$ when infeasibility occurs.
\subsection{Sub-Optimality and Convergence}
\label{sec:finite_dimensional_approximation}
In this subsection, we study the optimality of~\cref{prob:dr_det_optimal_cntrl} and show that~\cref{prob:dr_det_optimal_cntrl} computes a sub-optimal solution to~\cref{prob:cc_stoc_opt_cntrl}. We make a two-step approximation of~\cref{prob:cc_stoc_opt_cntrl} by using distributional robustness to formulate~\cref{prob:dr_stoc_opt_cntrl} with known mean and variance of the state and then use gPC propagation to construct the deterministic optimal control~\cref{prob:dr_det_optimal_cntrl} that is solved using SCP. In~\cref{lemma:sub_optimal_dr_snoc}, we prove the sub-optimality of the optimal cost $J^{*}_{\mathrm{SNOC}}$ of~\cref{prob:cc_stoc_opt_cntrl} compared to the optimal cost $J^{*}_{\mathrm{DR-SNOC}}$ of \cref{prob:dr_stoc_opt_cntrl} that is distributionally robust.
\begin{lemma}\label{lemma:sub_optimal_dr_snoc}
The optimal solution of~\cref{prob:dr_stoc_opt_cntrl} is a sub-optimal solution of~\cref{prob:cc_stoc_opt_cntrl}, i.e., $J^{*}_{\mathrm{SNOC}} \leq J^{*}_{\mathrm{DR-SNOC}}$.
\end{lemma}
\begin{proof}
The constraint set $\mathcal{X}_{\mathrm{DRLCC}}$ and $\mathcal{X}_{\mathrm{CQCC}}$ in~\cref{prob:dr_stoc_opt_cntrl} are a subset of the constraint set $\mathcal{X}_{\mathrm{LCC}}$ and $\mathcal{X}_{\mathrm{QCC}}$ of~\cref{prob:cc_stoc_opt_cntrl} respectively. Therefore, $J^{*}_{\mathrm{SNOC}} \leq J^{*}_{\mathrm{DR-SNOC}}$ as the feasible space of~\cref{prob:cc_stoc_opt_cntrl} is larger than the feasible space of~\cref{prob:dr_stoc_opt_cntrl}.
\end{proof}
\Cref{prob:dr_det_optimal_cntrl} (DNOC) computed by the gPC projection converges asymptotically to \Cref{prob:dr_stoc_opt_cntrl} (SNOC). The following theorems discuss the conditions for convergence.
\begin{theorem} \label{theorem:convergence_to_true_optimal}
The surrogate deterministic nonlinear optimal control~\cref{prob:dr_det_optimal_cntrl} with convex constraints
is a sub-optimal surrogate for the stochastic nonlinear optimal control~\cref{prob:cc_stoc_opt_cntrl} with the following being true:\\
\textbf{(a)} In the case without chance constraints, the cost $|J_{\mathrm{gPC}}^{*} - J^*| \to 0$ as $\ell \to \infty$\\
\textbf{(b)} In the case with linear and quadratic chance constraints, any feasible solution of~\cref{prob:dr_det_optimal_cntrl} is a feasible solution of~\cref{prob:cc_stoc_opt_cntrl} as $\ell \to \infty$ and $J^{*}_{\mathrm{SNOC}} \leq J^{*}_{\mathrm{gPC}}$, assuming that a feasible solution exists.
\end{theorem}
\begin{proof}\textbf{Case (a)}: It is sufficient to prove that the cost function and the dynamics are exact as $\ell \to \infty$. Using the Kronecker product notation, due to \cref{lemma:conv_gpc}, we have the following
\begin{equation}
\| x - \bar{\Phi} \mathrm{X}\|_{\mathcal{L}_2} \to 0 \ \text{as} \ \ell \to \infty
\label{eq:conv_kon_not}
\end{equation}
\begin{equation}
(\ref{eq:conv_kon_not}) \implies\ \| \dot{x} \to \bar{\Phi}\dot{\mathrm{X}} \|_{\mathcal{L}_2}\to 0 \ \text{as} \ \ell \to \infty
\label{eq:dyn_equi}
\end{equation}
\begin{equation}
\begin{aligned}
(\ref{eq:conv_kon_not}) \implies \ & |J_{\mathrm{gPC}} - J | \to 0 \ \text{as} \ \ell \to \infty \\
& |J_{\mathrm{gPC}_{f}} - J_{f} | \to 0 \ \text{as} \ \ell \to \infty
\end{aligned} \label{eq:cost_equi}
\end{equation}
From~\cref{eq:dyn_equi}, and~\cref{eq:cost_equi} we conclude that the optimal value $|J_{\mathrm{gPC}}^* - J^*| \to 0$ as $\ell \to \infty$, since the cost function, the dynamics, and the initial and terminal conditions converge to the original stochastic formulation (\cref{prob:cc_stoc_opt_cntrl}) as $\ell \to \infty$.\\
\noindent \textbf{Case (b):} Consider the sets $\mathcal{X}_\mathrm{LgPC}$, and $\mathcal{X}_\mathrm{QgPC}$ defined below.
\begin{equation}
\mathcal{X}_\mathrm{LgPC} = \{ x \in \mathcal{X}: x \approx \bar{\Phi}\mathrm{X} \ \text{where} \ \mathrm{X} \ \mathrm{satisfies}~(\ref{eq:socc_gpc}) \}
\label{eq:set_lin_gpc}
\end{equation}
\begin{equation}
\mathcal{X}_\mathrm{QgPC} = \{ x \in \mathcal{X}: x \approx \bar{\Phi}\mathrm{X} \ \text{where} \ \mathrm{X} \ \mathrm{satisfies}~(\ref{eq:sdp_con_gpc}) \}
\label{eq:set_sdp_gpc}
\end{equation}
Using \cref{lemma:lin_socc_gpc,lemma:sdp_con_gpc}, we see that the approximate convex constraints converge to the deterministic equivalent of the distributionally robust chance constraint as $\ell \to \infty$.
\begin{equation}
\begin{aligned}
\cref{lemma:lin_socc_gpc} & \implies \mathcal{X}_\mathrm{LgPC} \to \mathcal{X}_\mathrm{DRLCC} \ \text{as} \ \ell \to \infty \\
\cref{lemma:sdp_con_gpc} & \implies \mathcal{X}_\mathrm{QgPC} \to \mathcal{X}_\mathrm{CQCC} \ \text{as} \ \ell \to \infty
\end{aligned} \label{eq:pp1}
\end{equation}
Using \cref{lemma:lin_consev_approx,lemma:quad_const}, we have the following:
\begin{equation}
\begin{aligned}
\cref{lemma:lin_consev_approx} & \implies \mathcal{X}_\mathrm{DRLCC} \subseteq \mathcal{X}_\mathrm{LCC}\\
\cref{lemma:quad_const} & \implies \mathcal{X}_\mathrm{CQCC} \subseteq \mathcal{X}_\mathrm{QCC}\\
\end{aligned}\label{eq:pp2}
\end{equation}
\begin{equation}
\begin{aligned}
\{(\ref{eq:pp1}),~(\ref{eq:pp2})\} \implies
\begin{cases}
\mathcal{X}_\mathrm{LgPC} \subseteq \mathcal{X}_\mathrm{LCC} \ \text{as} \ \ell \to \infty \\
\mathcal{X}_\mathrm{QgPC} \subseteq \mathcal{X}_\mathrm{QCC} \ \text{as} \ \ell \to \infty
\end{cases}
\end{aligned} \label{eq:proof_caseb_thm1}
\end{equation}
Combining~(\ref{eq:pp1}) and~(\ref{eq:pp2}), we can conclude that~(\ref{eq:proof_caseb_thm1}) holds as $\ell \to \infty$. In \cref{fig:projection_lemmas}, we illustrate Lemmas 1, 2, 6 and 7 used in proving this theorem. This proves that if there is a feasible solution exists for~\cref{prob:dr_det_optimal_cntrl} then it is a feasible solution of~\cref{prob:cc_stoc_opt_cntrl} as $\ell \to \infty$. Using~\cref{lemma:sub_optimal_dr_snoc}, as $\ell \to \infty$ we have $J^{*}_{\mathrm{gPC}} \to J^{*}_{\mathrm{DR-SNOC}}$. This implies that $J^{*}_{\mathrm{SNOC}} \leq J^{*}_{\mathrm{\mathrm{gPC}}}.$
\end{proof}
Theorem~\ref{theorem:convergence_to_true_optimal} proves the consistency of the gPC projection method as $\ell \to \infty$. The asymptotic convergence of the cost and chance constraints is achieved with a large number of polynomials $\phi$. This leads to a deterministic optimal control problem of size $\ell d_{x}$. The choice of $\ell$ depends on the number of uncertainties in the system and the nature of the state distribution. A computationally efficient approach is to use $P_{\mathrm{gPC}} = 2$ to generate the functions $\phi$ used in the projection. This replicates the computational efficiency of linear covariance propagation techniques, while ensuring the convexity of the chance constraints in the gPC space. We study the chance constraint formulation for collision checking under uncertainty in dynamics and obstacle locations in the following section using distributional robustness and gPC projection.
\section{Motion Planning Under Uncertainty}\label{sec:motion_planning_gpc_scp}
The gPC-SCP method is applied to plan a safe trajectory under uncertain obstacles and stochastic nonlinear dynamics. We formulate the motion planning problem to incorporate uncertainty in the dynamics and then show that deterministic projections of chance constraints in the gPC space enable convex formulations of the collision constraint for obstacles with both deterministic and stochastic state models.
The motion planning problem is to compute an optimal and safe trajectory ($x \in \mathcal{X}_{\mathcal{F}}$) for the SDE in~\cref{eq:nl_stochastic_dynamics} from an initial state $x_{0}\in\mathcal{X}$ to the terminal set $\mathcal{X}_{f}\subseteq\mathcal{X}$ on a given map with static obstacles. In the following, we derive a chance constraint formulation of the collision constraint and the terminal-state constraint. The chance constraints are used to formulate an SNOC problem as described in~\cref{prob:cc_stoc_opt_cntrl}. The SNOC problem is then projected into the gPC space for solving via the SCP method. At each SCP iteration, the collision constraints are approximated as a linear chance constraint around the nominal trajectory and form a second-order cone constraint in the gPC state $\mathrm{X}$ as discussed in~\cref{lemma:lin_socc_gpc}. The terminal set is defined as a chance constraint on an ellipsoidal set and forms a semidefinite constraint in the gPC states as discussed in~\cref{lemma:sdp_con_gpc}.
In the following, we first discuss the linear chance constraint formulation for collision checking with a deterministic obstacle, and then extend it to include the uncertainty in obstacle locations for SCP. We prove that the approximation is a subset of the original nonlinear chance constraint. We then discuss the chance constraint formulation of the terminal set constraint. The chance constraint formulations for collision checking and terminal set are used to design the motion planning algorithm that integrates an asymptotically optimal sampling-based planner~\cite{hauser2016} with the gPC-SCP~\cref{prob:dr_trajopt} for computing a safe trajectory under uncertainty.
\subsection{Collision Checking with Deterministic Obstacles} \label{subsec:det_obstacle}
We derive a second-order cone constraint approximation of the circular obstacle in the gPC coordinates under uncertainty in dynamics at any point in time $t \in [t_0,t_f]$. The approximation involves two steps. We first derive a conservative linear chance constraint approximation of the nonlinear collision chance constraint. In the second step, we project the linear chance constraint into a second-order cone constraint in the gPC coordinates. Let the state of the obstacle be $\bar{p}_{\mathrm{obs}}$ at time $t$ and the radius of the obstacle be $r_{\mathrm{obs}}$. The collision chance constraint at any time $t$ for a robot with the state distribution $x$ and radius $r_{\mathrm{rob}}$ is given as follows:
\begin{equation}
\mathrm{Pr}\left( \|\mathbf{C}(x - \bar{p}_{\mathrm{obs}})\|_2 \geq r_{\mathrm{rob}} + r_{\mathrm{obs}}\right) \geq 1 - \epsilon_{\mathrm{col}},
\label{eq:chance_constraint_det_obst}
\end{equation}
where the matrix $\mathbf{C}$ is used to compute the position of the obstacle and the robot given the states $\bar{p}_{\mathrm{obs}}$ and $x$, respectively. The probability of collision is tuned using the risk measure $\epsilon_{\mathrm{col}} \in [0.001,0.1]$.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figures_tro/hyperplane_constraint.pdf}
\caption{\changed{A comparison of hyperplane approximations for deterministic and stochastic obstacle models using distributional robust~\cref{eq:lin_socc} and Gaussian approximations~\cref{eq:gaussian_linear_constraint} at three different robot locations. The obstacle position follows a multivariate Gaussian distribution $p_{\mathrm{obs}} \sim \mathcal{N} \left([1.5,0],\mathrm{diag\left([0.01,0.01]\right)} \right)$.}}
\label{fig:hyperplane_comparison}
\end{figure}
In the following theorem, we prove that, given a nominal state distribution trajectory $x_{\mathrm{nom}}$, we can compute a conservative linear chance constraint approximation of the collision constraint~\eqref{eq:chance_constraint_det_obst}. The nominal trajectory $x_{\mathrm{nom}}$ is computed by using a deterministic planner without considering the uncertainty in dynamics.
\begin{lemma}\label{theorem:deterministic_obstacles}
The linear chance constraint,
\begin{align}
\mathrm{Pr}\Big(&(\bar{x}_{\mathrm{nom}}- \bar{p}_{\mathrm{obs}})^\top \mathbf{C}^\top \mathbf{C} (x - \bar{p}_{\mathrm{obs}}) \nonumber \\ &\geq r_{\mathrm{safe}} \|\mathbf{C}(\bar{x}_{\mathrm{nom}}- \bar{p}_{\mathrm{obs}})\|_{2} \Big) \geq 1 - \epsilon_{\mathrm{col}},
\label{eq:lin_chance_constraint_det_obst}
\end{align}
in the robot state distribution, $x$ is a conservative approximation of the nonlinear collision chance constraint in~\eqref{eq:chance_constraint_det_obst} at any time $t \in [t_0, t_f]$, where $\bar{x}_{\mathrm{nom}}$ is a realization of the nominal state distribution $x_{\mathrm{nom}}$ of the robot that satisfies~\eqref{eq:chance_constraint_det_obst}, $\bar{p}_{\mathrm{obs}}$ is the state of the obstacle, and $r_{\mathrm{safe}} = r_{\mathrm{rob}} + r_{\mathrm{obs}}$.
\end{lemma}
\begin{proof}
Consider the set $\mathcal{X}_{\mathrm{free}}$, defined as $\mathcal{X}_{\mathrm{free}} = \{x: \|\mathbf{C}(x - \bar{p}_{\mathrm{obs}})\|_2 \geq r_{\mathrm{safe}}\}$. Given a nominal trajectory $x_{\mathrm{nom}}$, the set $\mathcal{X}_{s}$ defined as $\mathcal{X}_{s} = \{x:(\bar{x}_{\mathrm{nom}}- \bar{p}_{\mathrm{obs}})^\top \mathbf{C}^\top \mathbf{C} (x - \bar{p}_{\mathrm{obs}}) \geq r_{\mathrm{safe}} \|\mathbf{C}(\bar{x}_{\mathrm{nom}}- \bar{p}_{\mathrm{obs}})\|_{2}\}$ is such that $\mathcal{X}_{s} \subseteq \mathcal{X}_{\mathrm{free}}$. For a proof of $\mathcal{X}_{s} \subseteq \mathcal{X}_{\mathrm{free}}$ see~\cite{morgan2014model}. \Cref{fig:hyperplane_comparison} shows an example of the set $\mathcal{X}_{s}$ and $\mathcal{X}_{\mathrm{free}}$, where the hyperplane used to linearize the circular constraint leads to a reduced feasible space. We can construct an indicator function $I_{\mathrm{free}}(x)$ such that $I_{\mathrm{free}}(x) = 1$ if $x \in \mathcal{X}_{\mathrm{free}}$ and $I_{\mathrm{free}}(x) = 0$ otherwise. Similarly, the indicator $I_s(x)$ is such that $I_{s}(x) = 1$ if $x \in \mathcal{X}_{s}$ and $I_{s}(x) = 0$ otherwise.
\begin{align}
& \ \textrm{Since} \ \mathcal{X}_s \subseteq \mathcal{X}_{\mathrm{free}}, \ x \in \mathcal{X}_{s} \implies x \in \mathcal{X}_{\mathrm{free}}, \nonumber\\
& \ \ \textrm{and} \ I_{s}(x) = 1 \implies I_{\mathrm{free}}(x) =1. \label{eq:theorem_det_obstacle_eq1}
\end{align}
Therefore, if $\mathbb{E}(I_{s}) \geq 1 - \epsilon_{\mathrm{col}}$, then $\mathbb{E}(I_{\mathrm{free}}) \geq 1 - \epsilon_{\mathrm{col}}$ with at least $1- \epsilon_{\mathrm{col}}$ probability. Note that $\mathrm{Pr}(x \in \mathcal{X}_{\mathrm{free}}) = \mathbb{E}(I_{\mathrm{free}}(x))$ and $\mathrm{Pr}(x \in \mathcal{X}_{s}) = \mathbb{E}(I_{s}(x))$. This implies that if the chance constraint in~\eqref{eq:lin_chance_constraint_det_obst} is satisfied with the probability $1- \epsilon_{\mathrm{col}}$, then the constraint in~\eqref{eq:chance_constraint_det_obst} is satisfied with at least a probability of $1 - \epsilon_{\mathrm{col}}$. The distributional robustness approach can be visualized as a robust ball around the robot's state for collision checking using the convex feasible subset $\mathcal{X}_s$ of the non-convex feasible space $\mathcal{X}_{\mathrm{free}}$ that is consistent with the specified risk of constraint violation.\end{proof}
\begin{rmrk} \label{rmrk:socp_lin_collision_constraint}
The constraint~\eqref{eq:lin_chance_constraint_det_obst} is of the form $\mathrm{Pr}(a^\top x+b \leq 0)\geq 1- \epsilon_{\mathrm{col}}$, where $a = - (\bar{x}_{\mathrm{nom}} - \bar{p}_{\mathrm{obs}})^\top \mathbf{C}^\top \mathbf{C}$, and $b = (\bar{x}_{\mathrm{nom}} - \bar{p}_{\mathrm{obs}})^\top \mathbf{C}^\top \mathbf{C} \bar{p}_{\mathrm{obs}} + r_{\mathrm{safe}} \|\mathbf{C}(\bar{x}_{\mathrm{nom}} - \bar{p}_{\mathrm{obs}})\|_2$.
Using the Lemma~2, we formulate a second-order cone constraint that is used in the SCP problem for collision checking.
\end{rmrk}
\subsection{Collision Checking with Stochastic Obstacle} \label{subsec:stoc_obstacle}
We extend the deterministic formulation of the linear chance constraint in~\cref{eq:lin_chance_constraint_det_obst} to include uncertainty in obstacle state. Let the distribution of the obstacle state be $p_{\mathrm{obs}} \sim \mathcal{N} (\mu_p,\Sigma_p)$, where $\mu_p$ is the mean, $\Sigma_p$ is the variance matrix and the radius of the obstacle is $r_{\mathrm{obs}}$.
\begin{asmp}
\label{asmp:decorrelated_obstacle}
The obstacle state distribution $p$ is uncorrelated to the state distribution $x$ of the robot.
\end{asmp}
The collision chance constraint at any time $t$ for a state distribution $x$ and radius $r_{\mathrm{rob}}$ is given as follows:
\begin{equation}
\mathrm{Pr}\left( \|\mathbf{C}(x - p_{\mathrm{obs}})\|_2 \geq r_{\mathrm{rob}} + r_{\mathrm{obs}}\right) \geq 1 - \epsilon_{\mathrm{col}},
\label{eq:chance_constraint_stoc_obst}
\end{equation}
where both $x$ and $p_{\mathrm{obs}}$ are random variables, unlike~\eqref{eq:chance_constraint_det_obst}.
\begin{lemma}\label{theorem:stochastic_obstacles}
The linear chance constraint,
\begin{align}
\mathrm{Pr}\Big(&(\bar{x}_{\mathrm{nom}}- \bar{p}_{\mathrm{obs}})^\top \mathbf{C}^\top \mathbf{C} (x - p_{\mathrm{obs}}) \nonumber\\ &\geq r_{\mathrm{safe}} \|\mathbf{C}(\bar{x}_{\mathrm{nom}}- \bar{p}_{\mathrm{obs}})\|_{2} \Big) \geq 1 - \epsilon_{\mathrm{col}},
\label{eq:lin_chance_constraint_stoc_obst}
\end{align}
in the robot state distribution $x$ and the obstacle state distribution $p_{\mathrm{obs}}$ is a conservative approximation of the nonlinear collision chance constraint in~\eqref{eq:chance_constraint_stoc_obst} at any time $t \in [t_0, t_f]$, where $\bar{x}_{\mathrm{nom}}$ is a realization of the nominal state distribution $x_{\mathrm{nom}}$ of the robot, $\bar{p}_{\mathrm{obs}}$ is a realization of the obstacle state distribution $p_{\mathrm{obs}}$, and $r_{\mathrm{safe}} = r_{\mathrm{rob}} + r_{\mathrm{obs}}$.
\end{lemma}
\begin{proof}
Consider the sets $\mathcal{X}_{\mathrm{free}}$ and $\mathcal{X}_{s}$, defined as $\mathcal{X}_{\mathrm{free}} = \{x: \|\mathbf{C}(x - p_{\mathrm{obs}})\|_2 \geq r_{\mathrm{safe}}$ and $\mathcal{X}_s = \{x:(\bar{x}_{\mathrm{nom}}- \bar{p}_{\mathrm{obs}})^\top \mathbf{C}^\top \mathbf{C} (x - p_{\mathrm{obs}}) \geq r_{\mathrm{safe}} \|\mathbf{C}(\bar{x}_{\mathrm{nom}}- \bar{p}_{\mathrm{obs}})\|_{2}\}$ respectively, where $\bar{p}_{\mathrm{obs}}$ is a sample of the obstacle state distribution $p_{\mathrm{obs}}$. In constraint $\mathcal{X}_s$, $(x - p_{\mathrm{obs}})$ is the decision variable. Note that for any realization of the states $\bar{x}$ and $\bar{p}_{\mathrm{obs}}$ we have $\mathcal{X}_s \subseteq \mathcal{X}_{\mathrm{free}}$ (see~\cite{morgan2014model} for proof). Using the arguments in \cref{theorem:deterministic_obstacles}, constraint \cref{eq:lin_chance_constraint_stoc_obst} is a conservative approximation of constraint \cref{eq:chance_constraint_stoc_obst}.
\end{proof}
\begin{rmrk} \label{remark:socp_stoc_obstacles}
The constraint~\cref{eq:lin_chance_constraint_stoc_obst} is a linear chance constraint of the form $\mathrm{Pr}(a^\top (x - p_{obs}) + b \leq 0) \geq 1 - \epsilon_{\mathrm{col}}$, where $ a = - (\bar{x}_{\mathrm{nom}}- \bar{p}_{\mathrm{obs}})^\top \mathbf{C}^\top \mathbf{C}$, $b = r_{\mathrm{safe}} \|\mathbf{C}(\bar{x}_{\mathrm{nom}}- \bar{p}_{\mathrm{obs}})\|_{2}$, and $p_{\mathrm{obs}} \sim \mathcal{N}(\mu_p, \Sigma_p)$. In this case, the distributionally robust deterministic surrogate is computed for the stacked state $x_{c} = [x^\top p_{\mathrm{obs}}^\top]^\top$, which includes both the robot state and the obstacle state. The surrogate constraint is given as follows:
\begin{equation}
a^\top \mu_x - a^\top \mu_p + b + \sqrt{\tfrac{1-\epsilon_{\mathrm{col}}}{\epsilon_{\mathrm{col}}}} \sqrt{a^\top \Sigma_{x} a + a^\top \Sigma_p a} \leq 0.
\label{eq:det_surrogate_stoc_obstacles}
\end{equation}
Using~\cref{lemma:lin_socc_gpc}, the inequality constraint in the moment space is transformed into a second-order cone constraint in terms of the gPC states $\mathrm{X}$ of the robot dynamics.
\end{rmrk}
\begin{rmrk} \label{remark:socp_correlated_obstacles}
For a correlated obstacle state $p$ and a robot state $x$, with the cross-correlation matrix $\Sigma_{xp}$, the deterministic surrogate of~\cref{eq:chance_constraint_stoc_obst} is given as follows:
\begin{align}
& a^\top \mu_x - a^\top \mu_p + b + \nonumber\\ & \sqrt{\tfrac{1-\epsilon_{\mathrm{col}}}{\epsilon_{\mathrm{col}}}} \sqrt{a^\top \Sigma_{x} a + 2 a^\top \Sigma_{xp} a + a^\top \Sigma_p a} \leq 0.
\label{eq:det_surrogate_stoc_obstacles_decorrelated}
\end{align}
The derivation uses the stacked state $x_c$, as shown in~\cref{remark:socp_stoc_obstacles}.
\end{rmrk}
\begin{rmrk} \cref{theorem:stochastic_obstacles} can be applied for safe multi-agent reconfiguration under uncertainty by replacing the obstacle state $p_{obs}$ with the neighbouring robots state. The robots communicate the moments used in~\cref{eq:det_surrogate_stoc_obstacles_decorrelated} for collision checking with the neighbouring agents.
\end{rmrk}
\subsection{Terminal Constraint}\label{subsec:term_constraint} The terminal constraint is defined as an ellipsoidal set $(x-\bar{x}_f)^\top Q_{\mathcal{X}_f}(x-\bar{x}_{f}) \leq c_{f}$ around a terminal point $\bar{x}_{f}$, where $Q_{\mathcal{X}_f}$ is a positive definite matrix. The chance constraint formulation of the terminal set involves two steps: 1) constrain the mean of the terminal point as $\mu_{f} = \bar{x}_{f}$ and 2) formulate the quadratic chance constraint $\mathrm{Pr}((x-\bar{x}_f)^\top Q_{\mathcal{X}_f}(x-\bar{x}_{f}) \leq c_{f})\geq 1- \epsilon_{f}$ around the mean $\mu_f$ with the risk measure $\epsilon_f$ of not reaching the terminal set. We use the conservative deterministic constraint discussed in~\cref{lemma:quad_const}, that bounds the variance of the state. The terminal constraints are summarized as follows.
\begin{align}
\mu_{f} = \bar{x}_{f},\quad \frac{1}{c_f}\text{tr}(Q_{\mathcal{X}_f}\Sigma_{x_{f}}) \leq \epsilon_{f} \label{eq:terminal_set_mp}
\end{align}
where $\mu_f$ is the mean and $\Sigma_{x_{f}}$ is the variance of the terminal state. The conservative approximations we presented in this section are a trade-off between the knowledge of the moments available and the computational speed achieved by convex constraints. For a linear SDE with obstacles whose uncertainty is described by the Gaussian distribution, a tighter equivalent deterministic surrogate constraint can be derived using the covariance propagation technique for uncertainty propagation and the inverse cumulative distribution function for the Gaussian distribution.
\subsection{Motion Planning Algorithm (gPC-SCP)}
For motion planning, we integrate the deterministic approximations discussed in~\cref{subsec:det_obstacle,subsec:stoc_obstacle,subsec:term_constraint} with the asymptotically optimal rapidly exploring random trees~\cite{hauser2016} (AO-RRT). Following Algorithm~\ref{Algo:MP-gPC-SCP}, outlines the motion planning method using gPC-SCP for a dynamical system under uncertainty.
\begin{algorithm}
\DontPrintSemicolon
\caption{gPC-SCP}
\label{Algo:MP-gPC-SCP}
\KwInput{Map, obstacle location, $x_{0}$, $\mathcal{X}_{f}$, $\Delta t$, $\ell$,}
\KwInput{Uncertainty model of $g(x,\bar{u})$ in SDE~\cref{eq:nl_stochastic_dynamics}.}
\KwOutput{Optimal and safe state distribution $\mathcal{X}_{\mathrm{sol}} = \{x_{0},x_{1},...,x_{T}\}$ and control input $\mathcal{U}_{\mathrm{sol}} = \{\bar{u}_{0},\bar{u}_{1},...,\bar{u}_{T-1}\}$.}
\Comment{\emph{Stage 1: gPC Projection.}}
\cref{prob:dr_det_optimal_cntrl} $\leftarrow$ \blockgpcprojection { Formulate the collision constraint using~\cref{theorem:deterministic_obstacles,theorem:stochastic_obstacles},\;
Formulate the terminal set $\mathcal{X}_f$ using~\cref{eq:terminal_set_mp},\;
Project the SDE using~\cref{eq:continuous_ode_form},\;
Project the collision constraint using~\cref{lemma:lin_socc_gpc},\; Project the terminal set using~\cref{lemma:sdp_con_gpc},\;
Setup and project cost function using~\cref{eq:cost_odeopt},\;
\textbf{return:}~\cref{prob:dr_det_optimal_cntrl} in gPC space.\;}
\cref{prob:dr_trajopt} $\leftarrow$ \Kwlinearize
(\KwDiscretize(\cref{prob:dr_det_optimal_cntrl}))\\
Save \cref{prob:dr_trajopt}.\\
\Comment{\emph{Stage 2: Compute a nominal trajectory using AO-RRT.}}
$\{\mathcal{X}^{0}_{\mathrm{sol}},\mathcal{U}^{0}_{\mathrm{sol}},T\}$ $\leftarrow$ \Kwaorrt($x_0,\mathcal{X}_f,\Delta t,\dot{x} =f(x,\bar{u})$)
\tcc*[r]{For detailed implementation of AO-RRT see~\cite{hauser2016}.}
\Comment{\emph{Stage 3: gPC-SCP.}}
$\{\mathcal{X}_{\mathrm{sol}},\mathcal{U}_{\mathrm{sol}}\}$ $\leftarrow$\KwSolve(\cref{prob:dr_trajopt},$\{\mathcal{X}^{0}_{\mathrm{sol}},\mathcal{U}^{0}_{\mathrm{sol}},T\}$)\tcc*[r]{The sequential convex programming (SCP) approach is described in~\cref{subsec:gPC-SCP-scp}.}
\end{algorithm}
Algorithm~\ref{Algo:MP-gPC-SCP} has three stages. In Stage 1, we formulate the linear chance constraint for collision checking and the quadratic chance constraint for the terminal constraint, respectively. Using the chance constraints, we setup \cref{prob:dr_stoc_opt_cntrl} (SNOC) and project it to \cref{prob:dr_det_optimal_cntrl} (DNOC). We formulate gPC-SCP in~\cref{prob:dr_trajopt} by discretizing~\cref{prob:dr_det_optimal_cntrl}. In Stage 2, we use AO-RRT to compute an initial feasible trajectory $\{\mathcal{X}^0_{\mathrm{sol}},\mathcal{U}^0_{\mathrm{sol}}\}$ for the nominal dynamics $\dot{x} = f(x,\bar{u})$. In Stage 3, the feasible trajectory is then used to initialize the SCP iterations in gPC-SCP, which optimizes for the uncertainty in dynamics. The output of stage 3 is a safe and optimal state trajectory in the gPC space. Using the gPC polynomials in~\cref{eq:state_kronck_notation}, the gPC space trajectory is projected back to the state space distribution to output $\{\mathcal{X}_{\mathrm{sol}},\mathcal{U}_{\mathrm{sol}}\}$ in line 12 of Algorithm~\ref{Algo:MP-gPC-SCP}. Note that RRT in AO-RRT can be replaced with sparse tree~\cite{li2016asymptotically} algorithm for improved speed and with RRT$^{*}$~\cite{karaman2011sampling} for optimality. We discuss the application of Algorithm~\ref{Algo:MP-gPC-SCP} in~\cref{sec:simulations_and_experiments}.
\color{black}
\subsection{Real-Time Planning using Predictor-Corrector gPC-SCP}
In the gPC-SCP method, we compute the full trajectory distribution of the state $x$ in the gPC state $X$. However, in many applications, we might need only the mean of the state trajectory $\mu_x$. Therefore, we present an iterative algorithm, called predictor-corrector gPC-SCP (gPC-SCP$^\mathrm{PC}$), to compute the mean trajectory $\mu_x$ safe under stochastic uncertainty in dynamics and obstacles. In the gPC-SCP$^\mathrm{PC}$ algorithm, at any SCP iteration, we first predict a covariance correction term by propagating the gPC dynamics~\cref{eq:continuous_ode_form} using the nominal trajectory $\{\mathcal{X}^0_{\mathrm{sol}},\mathcal{U}^0_{\mathrm{sol}}\}$ and then solve a deterministic optimal control problem to compute the mean of the trajectory that accommodates for the uncertainty through the correction term. We describe the two main steps involved in the gPC-SCP$^\mathrm{PC}$ method and the algorithm for computing a mean trajectory.
\subsubsection{Prediction Step}\label{subsec:prediction} Given a nominal open-loop control input $\mathcal{U}_{\mathrm{sol}} = \{\bar{u}_0, \bar{u}_1,\dots,\bar{u}_{T-1}\}$ and the initial condition $x_0$, we propagate the gPC dynamics~\cref{eq:continuous_ode_form} using the control $\mathcal{U}_{\mathrm{sol}}$ and compute the corresponding trajectory of the gPC state $X(t)$.
\begin{equation}
b_{\mathrm{c}}(t) = X(t)^\top U N N^\top U^\top X(t) \label{eq:gpc_to_state_covariance_correction}
\end{equation}
Using \cref{eq:gpc_to_state_covariance_correction} and the matrices defined in~\cref{eq:mat_lin_con_gpc}, we compute the covariance correction term $b_{\mathrm{c}}$. The deterministic linear collison chance constraint~\cref{eq:det_surrogate_stoc_obstacles} in terms of $\mu_x$ and the correction term $b_{\mathrm{c}}$ for a given risk of collison $\epsilon_{\mathrm{col}}$ is given as follows:
\begin{equation}
a^\top \mu_x - a^\top \mu_p + b + \sqrt{\tfrac{1-\epsilon_{\mathrm{col}}}{\epsilon_{\mathrm{col}}}} \sqrt{b_{\mathrm{c}} + a^\top \Sigma_p a} \leq 0.
\label{eq:mean_lin_col_checking}
\end{equation}
where $\Sigma_p$ is the known (or estimated) covariance matrix of the obstacle state as defined in~\cref{remark:socp_stoc_obstacles}. Note that~\cref{eq:mean_lin_col_checking} is a linear constraint in $\mu_x$ with known $b_{\mathrm{c}}$ and risk value $\epsilon_{\mathrm{col}}$, where as~\cref{eq:det_surrogate_stoc_obstacles} is a second-order cone constraint. This reduces the complexity of the SCP problem and improves the computation speed. Using~\cref{eq:mean_lin_col_checking}, the safe set defined using the distributionally robust linear joint chance constraint $\inf_{x(t) \sim (\mu_x,\Sigma_{x})} \mathrm{Pr}(\land_{i = 1}^{m} a_{i}^\top x + b_{i} \leq 0) \geq 1- \epsilon$ is transformed into a set of linear constraints shown below.
\begin{align}
& a_{i}^\top \mu_x - a_{i}^\top \mu_p + b_{i} \nonumber \\ &+ \sqrt{\tfrac{1-\epsilon_{\mathrm{col}}}{\epsilon_{\mathrm{col}}}} \sqrt{b_{\mathrm{c}} + a_{i}^\top \Sigma_p a_{i}} \leq 0 \ \forall i \in \{1,\dots,m\} \label{eq:mean_safe_set}
\end{align}
The terminal constraint is defined using the mean of the terminal state~\cref{eq:terminal_set_mp}. The prediction step involves the propagation of the SDE using gPC approximated nonlinear dynamics. Although this is an additional step in the overall planning approach, it reduces the complexity of the optimal control problem.
\subsubsection{Correction Step}\label{subsec:correction_step}
The prediction step is followed by a correction optimization to optimize the nominal trajectory for uncertainty. We describe the optimal control problem in the following with mean $\mu_x$ and $\bar{u}$ as decision variables.
\begin{problem}\label{prob:correction_optimal_control}
Distributionally Robust DNOC with Mean State Variable.
\begin{align}
\underset{\mu_{x(t)},\Bar{u}(t)}{\min} &
\scalebox{0.9}{$\left[\int_{t_{0}}^{t_{f}}\left(\mu^\top_x Q \mu_x + \|\bar{u}\|_{p} \right)dt +\mu^\top_x(t_f) Q_{f} \mu_x(t_f)\right]$} \nonumber \\
\text{s.t.} \quad
& \scalebox{0.9}{$d\mu_{x} = f(\mu_{x}(t),\Bar{u}(t))dt$} \label{eq:mean_dynamics}\\
& \mathrm{Safe\ Set:}\ \cref{eq:mean_safe_set} \\
&\scalebox{0.9}{$\Bar{u}(t) \in \mathcal{U} \quad \forall t \in [t_{0},t_{f}]$} \label{eq:control_constraints}\\
&\scalebox{0.9}{$\mu_{x}(t_{0}) = \mathbb{E}(x_0) \quad \mu_{x}(t_{f}) = \bar{x}_f$} \label{eq:mean_init_term_constraints}
\end{align}
\end{problem}
\noindent where $Q$ and $Q_f$ are positive definite matrices, as in~\cref{eq:cost_stopt}. The dynamics of the mean state is given in~\cref{eq:mean_dynamics}. We compute the mean dynamics using only $\phi_0$ as the polynomial in the gPC expansion method. The initial and terminal constraints in~\cref{eq:mean_init_term_constraints} are the expected values of the initial and terminal distributions. The above optimal control problem is linearized and discretized to formulate a convex optimization problem as described in~\cref{subsec:gPC-SCP-scp}. We use the prediction and correction steps to formulate the gPC-SCP$^\mathrm{PC}$ algorithm by modifying the gPC-SCP algorithm. The gPC-SCP$^\mathrm{PC}$ algorithm is given in the following Algorithm~\ref{Algo:PC-gPC-SCP}. We initialize gPC-SCP$^\mathrm{PC}$ using AO-RRT similar to Algorithm~\ref{Algo:MP-gPC-SCP}.
\textbf{Convergence and Feasibility.} Similar to the gPC-SCP method, we use a trust region-based iteration scheme to ensure feasibility of~\cref{prob:correction_optimal_control}. Additionally, in~\cref{prob:correction_optimal_control}, the risk of collision $\sqrt{\frac{1-\epsilon_{\mathrm{col}}}{\epsilon_\mathrm{col}}}$ can be used as a decision variable to compute a safe trajectory with optimal risk. This also ensures feasibility by computing a high-risk trajectory when a specified risk violation trajectory is not available. The convergence of gPC-SCP$^\mathrm{PC}$ depends on the size of the uncertainty $g$ and the SCP iteration scheme similar to gPC-SCP. We refer to the work in~\cite{foust2020optimal} for the proof of convergence when the prediction step is integrated with SCP for correction.
\begin{table}[h]
\centering
\begin{tabular}{|c|c|c|}
\hline
& gPC-SCP& gPC-SCP$^\mathrm{PC}$\\
\hline
Trajectory & Full distribution ($x$) & Mean ($\mu_x$)\\
\hline
Convex Problem Type & SOCP & QP \\
\hline
Solution & $(x,\bar{u})$ & $(\mu_x,\bar{u})$\\
\hline
Covariance Minimization & Yes & No\\
\hline
Computation Time & Medium (\SI{10}{s}) & Small (\SI{0.9}{s})\\
\hline
\end{tabular} \caption{A comparison of gPC-SCP and gPC-SCP$^\mathrm{PC}$.}
\label{tab:gPC_vs_PC_SCP}
\end{table}
\textbf{gPC-SCP vs. gPC-SCP$^\mathrm{PC}$.}
In \cref{tab:gPC_vs_PC_SCP}, we make a comparison of the gPC-SCP and gPC-SCP$^\mathrm{PC}$ methods. In the gPC-SCP method, we compute the full state trajectory distribution ($x$), while in the gPC-SCP$^\mathrm{PC}$ method, we compute the mean state trajectory ($\mu_x$). The convexified collision constraint in gPC-SCP is a second-order cone, whereas, in gPC-SCP$^\mathrm{PC}$, it is a quadratic constraint. In the gPC-SCP method, we can compute a trajectory with minimum variance and design a trajectory with specified terminal state covariance.
\begin{algorithm}
\DontPrintSemicolon
\caption{gPC-SCP$^\mathrm{PC}$}
\label{Algo:PC-gPC-SCP}
\KwInput{Map, obstacle location, $\mu_{x}(t_0)$, $\mu_x(t_f)$, $\Delta t$, $\ell$,}
\KwInput{Uncertainty model of $g(x,\bar{u})$ in SDE.}
\KwOutput{Optimal and safe mean trajectory $\mathbf{\mu}_{\mathrm{sol}} = \{\mu_{x_0},\mu_{x_1},...,\mu_{x_T}\}$ and control input $\mathcal{U}_{\mathrm{sol}} = \{\bar{u}_{0},\bar{u}_{1},...,\bar{u}_{T-1}\}$.}
\Comment{\emph{Stage 1: gPC Projection}}
\cref{eq:continuous_ode_form} $\leftarrow$ $\textbf{gPC\ Projection}$~\cref{eq:nl_stochastic_dynamics}\\
\Comment{\emph{Stage 2: Formulate the Mean Variable Optimal Control Problem.}}
\textbf{Correction-SCP} $\leftarrow$ \Kwlinearize
(\KwDiscretize(\cref{prob:correction_optimal_control}))\\
\Comment{\emph{Compute Nominal Trajectory.}}
$\{\mathcal{X}^{0}_{\mathrm{sol}},\mathcal{U}^{0}_{\mathrm{sol}},T\}$ $\leftarrow$ \Kwaorrt($x_0,\mathcal{X}_f,\Delta t,\dot{x} =f(x,\bar{u})$)
\tcc*[r]{For detailed implementation of AO-RRT see~\cite{hauser2016}.}
\Comment{\emph{Stage 3: gPC-SCP$^\mathrm{PC}$ Algorithm.}}
$\mathcal{X}_{\mathrm{sol}},\mathcal{U}_{\mathrm{sol}}$ $\leftarrow$ $\mathcal{X}^{0}_{\mathrm{sol}},\mathcal{U}^{0}_{\mathrm{sol}}$\\
\blockpcscp{
\While{Not Converged}{
\Comment{Step 1: Prediction}
$X(t)$ $\leftarrow$ Propagate~\cref{eq:continuous_ode_form} using $\mathcal{U}_{\mathrm{sol}}$ as discussed in~\cref{subsec:prediction}\;
$b_{\mathrm{c}}(t)$ $\leftarrow$ $X(t)^\top U N N^\top U^\top X(t)$\;
\Comment{Step 2: Correction}
$\mathcal{X}_{\mathrm{sol}},\mathcal{U}_{\mathrm{sol}}$ $\leftarrow$ $\textbf{Correction-SCP}$ as discussed in~\cref{subsec:correction_step}\; $\left(\{\mathcal{X}_{\mathrm{sol}},\mathcal{U}_{\mathrm{sol}},T\},b_{\mathrm{c}}(t) \right)$ \;
}
}
\end{algorithm}
\changed{In Section~\ref{sec:simulations_and_experiments}, we compare the motion plans generated by the gPC-SCP$^\mathrm{PC}$ Algorithm~\ref{Algo:PC-gPC-SCP} with the plans computed using the gPC-SCP Algorithm~\ref{Algo:MP-gPC-SCP}.}
\color{black}
\section{Simulations and Experiments}
\label{sec:simulations_and_experiments}
We apply the gPC-SCP method (Algorithm~\ref{Algo:MP-gPC-SCP}) and the gPC-SCP$^\mathrm{PC}$ method (Algorithm~\ref{Algo:PC-gPC-SCP}) to design the safe and optimal motion plan distribution and a safe mean trajectory, respectively, for the dynamics of the modular robotic spacecraft simulator with three degrees of freedom configuration~\cite{nakka2018six} and six degrees of freedom configuration~\cite{nakka2018six}. For the dynamics of the spacecraft simulator, we conducted an empirical study using simulation to demonstrate the safety provided by Algorithms~\ref{Algo:MP-gPC-SCP} and~\ref{Algo:PC-gPC-SCP} compared to the Gaussian collision constraint~\cite{zhu2019chance,du2011probabilistic,blackmore2010probabilistic,blackmore2011chance,janson2018monte} given in~\cref{eq:gaussian_linear_constraint}. We ran the simulations on a Mac machine with the configuration: $7^{th}$ generation, Intel Core i7 process, and 16 GB RAM. We solve the SCP problem using CVXpy~\cite{diamond2016cvxpy} and the ECOS~\cite{domahidi2013ecos} solver. We then validate the experimental results on the spacecraft simulator hardware platform, where the motion plans are computed on an NVIDIA Jetson TX2 computer
\subsection{Robotic Spacecraft Dynamics Simulator~\cite{nakka2018six}}\label{subsec:problem_setup}
The Caltech's 3-DOF M-STAR (Multi-Spacecraft Testbed for Autonomy Research) is shown in~\cref{fig:mstar_setup}. The testbed floats on an ultraprecise epoxy floor using linear air bearings to achieve 3DOF frictionless motion. The M-STAR is equipped with eight thrusters for position $(x,y)$ and yaw angle ($\theta$) control. The dynamics of the robot is given as follows:
\begin{align}
d\mathbf{x} = f(\mathbf{x},\bar{u})dt + \sigma g(\mathbf{x},\bar{u})dw,
\end{align}
where $\mathbf{x} = [x,y,\theta,\dot{x},\dot{y},\dot{\theta}]^\top$, $\bar{u} \in \mathbb{R}^{8}$, $\sigma \in \mathbb{R}$. The functions $f(\mathbf{x},\bar{u})$ and $g(\mathbf{x},\bar{u})$ are given below:
\begin{align}
f(\mathbf{x},\bar{u}) &= \begin{bmatrix} \mathbb{I}_{3\times3}&0\\0&0\end{bmatrix}\mathbf{x} + \begin{bmatrix}0\\B(m,\mathrm{I},l,\theta)\bar{u}\end{bmatrix},\\
g(\mathbf{x},\bar{u}) &=\mathrm{blkdiag} \{0,B(m,I,l,\theta)\bar{u} \}. \label{eq:spacecraft_dynamics}
\end{align}
The control effort $\bar{u}$ is constrained to be $0 \leq \bar{u}\leq$\SI{0.45}{N}, and $B(m,\mathrm{I},l,\theta) \in \mathbb{R}^{3 \times 8}$ is the control allocation matrix (see~\cite{nakka2018six}), where $m=$ \SI{10}{kg} and $\mathrm{I}=$ \SI{1.62}{kg.m.s^{-2}} are the mass and the inertia matrix, and $l=$ \SI{0.4}{m} is the moment arm. Uncertainty $\sigma g(\mathbf{x},\bar{u})$ is due to viscous friction between the robot and the flat floor, drift due to gravity gradient, and uncertainty in thruster actuation. We choose $\sigma = 0.1$, this value encompasses all the above forms of uncertainty. With this model, we study the convergence and collision checking discussed in~\cref{theorem:convergence_to_true_optimal,theorem:deterministic_obstacles,theorem:stochastic_obstacles}.
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{figures_tro/exp_setup.pdf}
\caption{The top and side view of the Caltech's robotic spacecraft dynamics simulator.}
\label{fig:mstar_setup}
\end{figure}
\subsubsection{Simulation}\label{subsec:simulations}
Consider the map shown in~\cref{fig:motion_planning_variation_pgpc}. We design a safe and optimal trajectory, $J = \|\bar{u}\|_{2}$, $J_f = \mathbb{E}(\mathbf{x}_{f}^\top \mathbf{x}_{f})$, from the initial state $\mathbb{E}(x_{0}) = 0$ to the terminal state $\mathbb{E}(x_{f}) = [0.3,2.3,0,0,0,0]^\top$, while avoiding the obstacle located at $p_{\mathrm{obs}} = [0.3,1,0,0,0,0]^\top$ with radius $r_{\mathrm{safe}}=$ \SI{0.5}{m}. We formulate the collision constraint using~\cref{theorem:deterministic_obstacles,theorem:stochastic_obstacles} and bound the terminal variance using a slack variable to ensure feasibility.
\begin{figure}[htpb]
\centering
\includegraphics[width=\columnwidth]{figures_tro/gpc_vs_polynomials.pdf}
\caption{The figure demonstrates convergence of the mean (shown in the $(x,y)$ plane) and the variance ($\sigma_{x},\sigma_{y}$) (shown as a function of time) of the states $(x,y)$ with increasing $P_{\mathrm{gPC}}$ for $\sigma=\{0.01,0.1\}$.}
\label{fig:motion_planning_variation_pgpc}
\end{figure}
In~\cref{fig:motion_planning_variation_pgpc}, we show the mean and variance of the position of M-STAR, $(x,y)$, computed using Algorithm~\ref{Algo:MP-gPC-SCP}. We compare the mean and variance computed for the gPC polynomial degree $P_{\mathrm{gPC}}=\{1,2,3,4\}$ and the variance $\sigma = \{0.01,0.1\}$ with $g(\mathbf{x},\bar{u}) = [0,B \bar{u}]^\top$ in~\cref{fig:motion_planning_variation_pgpc}. The convergence of mean and variance with increasing $P_{\mathrm{gPC}}$, implies convergence with respect to $\ell$, validating~\cref{theorem:convergence_to_true_optimal}. Since there are no known methods to compute global optimal solution for \cref{prob:cc_stoc_opt_cntrl}, we cannot comment on the sub-optimality of the solution. For the case with low variance $\sigma=0.01$, we observe that the degree gPC polynomials $P_{\mathrm{gPC}} =1$ are sufficient to accurately calculate the mean and variance. However, for a large variance $\sigma = 0.1$, we need gPC polynomials of degree $P_{\mathrm{gPC}}=2$. We can use $P_{\mathrm{gPC}}=1$ with large variance in dynamics for the following two cases: 1) short-horizon planning and 2) iterative planning with closed-loop state information updates. We use gPC polynomials with degree $P_{\mathrm{gPC}} =2$ in the following motion planning analysis.
\begin{figure}[htpb]
\centering
\includegraphics[width=0.8\columnwidth]{mstar_results/gPC_vs_PC_gPC_SCP.pdf}\caption{\changed{We compare the state trajectories generated using gPC-SCP and gPC-SCP$^\mathrm{PC}$ for deterministic(left) and stochastic obstacle (right).}}
\label{fig:gpc_vs_pc_gpc}
\end{figure}
\begin{figure}[htpb]
\centering
\includegraphics[width=0.9\columnwidth,height=1.3in]{mstar_results/risk_vs_opt.pdf}
\caption{\changed{Integral cost of control for the open-loop trajectories computed using gPC-SCP and gPC-SCP$^\mathrm{PC}$ for various collison risk measures $\epsilon_{\mathrm{col}}=\{0.01,0.05,0.1,0.2,0.3,0.4,0.5\}$.}}
\label{fig:cost_vs_risk}
\end{figure}
\changed{\subsubsection{gPC-SCP vs. gPC-SCP$^\mathrm{PC}$}\label{subsec:results_sim_motionplan}
In~\cref{fig:gpc_vs_pc_gpc}, we show the mean of the state trajectory distribution $x$ computed using gPC-SCP and the mean trajectory $\mu_x$ computed using gPC-SCP$^\mathrm{PC}$ for deterministic (light red) and stochastic obstacle (dark red). The uncertainty in the obstacle position is assumed to be a multivariate Gaussian with covariance $\Sigma_p=\mathrm{diag}([1e-4,1e-4])$. We use the distributionally robust linear chance constraint~\cref{eq:lin_socc} with risk measure $\epsilon_{\mathrm{col}}=0.05$ to convexify the free space, as discussed in~\cref{theorem:deterministic_obstacles,theorem:stochastic_obstacles}. For an equivalent comparison of gPC-SCP and gPC-SCP$^\mathrm{PC}$, we only use the control cost $\|u\|_2$ and constrain the expected initial and terminal states $\mathbb{E}{x}_0$ and $\mathbb{E}{x}_f$, respectively.}
\changed{We observe in~\cref{fig:gpc_vs_pc_gpc} that the mean of the trajectory distribution $x$ computed using gPC-SCP and the mean trajectory $\mu_x$ computed using gPC-SCP$^\mathrm{PC}$ are equivalent for the risk measure $\epsilon_{\mathrm{col}} = 0.05$. The computation time for gPC-SCP was \SI{10.86}{s} and for gPC-SCP$^\mathrm{PC}$ was \SI{0.9}{s}. Note that using gPC-SCP, we can calculate the full trajectory distribution $x$, which is useful for applications such as safe exploration~\cite{nakka2020chance} and covariance minimization, while using gPC-SCP$^\mathrm{PC}$, we can calculate a probabilistic safe trajectory under tight computational time constraints for closed-loop feedback control. Furthermore, in~\cref{fig:cost_vs_risk}, we observe that the total optimal control cost for various collision risk measures $\epsilon_{\mathrm{col}} = \{0.01,0.05,0.1,0.2,0.3,0.4,0.5\}$ is similar for gPC-SCP and gPC-SCP$^\mathrm{PC}$ as we use the deterministic control assumption in the gPC-SCP method. For this reason, we use the mean of the state trajectory computed using the gPC-SCP method as the nominal state and control trajectory for hardware validation. Note that the optimal control cost decreases with increasing risk measures. As the risk measure increases, the robustness provided by the DRLCC~\cref{eq:lin_socc} decreases, leading to an increase in the feasible space. The larger feasible space leads to a more optimal and less safe trajectory.}
\begin{figure}[htpb]
\centering
\includegraphics[width=\columnwidth]{mstar_results/DRLCC_vs_Gaussian.pdf}\caption{\changed{We compare the trajectories computed using gPC-SCP and gPC-SCP$^\mathrm{PC}$ with distributionally-robust linear chance constraint (DRLCC)~\cref{eq:lin_socc} and Gaussian chance constraint~\cref{eq:gaussian_linear_constraint} for collision checking for deterministic (light red) and stochastic obstacles (dark red).}}
\label{fig:drlcc_vs_gaussian}
\end{figure}
\begin{figure}[htpb]
\centering
\includegraphics[width=0.85\columnwidth]{mstar_results/collision_vs_risk.pdf}\caption{\changed{We show the results of 1000 Monte Carlo feedback control executions of the trajectories computed using gPC-SCP with distributionally-robust linear chance constraint (DRLCC)~\cref{eq:lin_socc} and Gaussian chance constraintx~\cref{eq:gaussian_linear_constraint} for collision checking for various risk measures $\epsilon_{\mathrm{col}}=\{0.01,0.05,0.1,0.2,0.3,0.4,0.5\}$.}}
\label{fig:closed_loop_linear_constraint}
\end{figure}
\begin{figure}[htpb]
\centering
\includegraphics[width=0.9\columnwidth]{mstar_results/terminal_set_dr.pdf}
\caption{\changed{We compare the trajectories generated for different values $c=\{0.005,0.01,0.015\}$ of \cref{eq:quad_chnc_cons} (shown as the green circle) of the terminal set using the robust quadratic constraint~\cref{eq:sdp_constraint} with risk measure $\epsilon = 0.05$ for the terminal set. Finally, we show the terminal state of the robot (blue) when a nominal trajectory (sampled from the probabilistic trajectory) is executed using an exponentially-stabilizing controller.}}
\label{fig:dr_terminal_set}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{mstar_results/terminal_set_gau.pdf}
\caption{\changed{We compare the trajectories generated for different values $c=\{0.005,0.01,0.015\}$ of \cref{eq:quad_chnc_cons} (shown as the green circle) of the terminal set using $3\sigma$ confidence-based terminal set $3\mathrm{tr}(\Sigma_x) \leq c$ with risk measure $\epsilon = 0.05$.}}
\label{fig:gaussian_terminal_set}
\end{figure}
\changed{\subsubsection{Chance Constraints}
In~\cref{fig:drlcc_vs_gaussian}, we compare the motion plans computed using the DRLCC~\cref{eq:lin_socc} and the Gaussian linear chance constraint ~\cref{eq:gaussian_linear_constraint} for collision checking in both the gPC-SCP and the gPC-SCP$^\mathrm{PC}$ algorithms. We validate the safety of the motion plans computed using gPC-SCP by tracking a sampled trajectory with the exponentially-stabilizing controller designed in~\cite{nakka2018six} for the nominal dynamics. We sample a trajectory $\mathbf{x}$ using the projection equation $\mathbf{x} = \Phi^{\top}(\xi) \mathrm{X}$. The gPC-SCP algorithm computes the gPC coordinates $\mathrm{X}$. We compute $\mathbf{x}$ by randomly sampling the multivariate normal distribution $\xi$. Using the motion plans shown in~\cref{fig:drlcc_vs_gaussian} as input to the controller, we get the collision profile shown in~\cref{fig:closed_loop_linear_constraint} for various risk measures $\epsilon_{\mathrm{col}}=\{0.01,0.05,0.1,0.2,0.3,0.4,0.5\}$ over 1000 trials.}
\changed{ We observe in~\cref{fig:closed_loop_linear_constraint} that the DRLCC performs better than Gaussian collision checking for all the risk measures. Although the number of collisions when using Gaussian approximation is equivalent to the expected collisions for the deterministic obstacle case, it performs poorly in the stochastic obstacle case. The increased number of collisions under stochastic obstacles when using Gaussian approximation could be because the robot dynamics constraint in the SCP formulation is weakly nonlinear under deterministic constraints and strongly nonlinear under stochastic constraints. We observed that the performance of gPC-SCP and gPC-SCP$^\mathrm{PC}$ is equivalent as the trajectory computed by gPC-SCP$^\mathrm{PC}$ is a subset of the state distribution computed by gPC-SCP.}
\changed{The goal reaching of the robot for various terminal set sizes $c = \{0.005,0.01,0.015\}$m with $A = \mathrm{I}$ is shown in~\cref{fig:dr_terminal_set} for the robust quadratic chance constraint~\cref{eq:sdp_constraint} and in~\cref{fig:gaussian_terminal_set} for the $3\sigma$-confidence-bound $3\mathrm{tr}(\Sigma_x) \leq c$. We use a risk measure of $0.05$ for terminal constraint violation. The robust quadratic constraint is formulated as described in~\cref{subsec:term_constraint}. We observe in \cref{fig:dr_terminal_set} that for 1000 trials of feedback execution there are no violations of the terminal constraint using~\cref{eq:sdp_constraint}, while in \cref{fig:gaussian_terminal_set} for the $3\sigma$-confidence bound on variance the constraint violations for $c = \{0.005,0.01,0.015\}$ are $\{104, 31, 7\}$, respectively. Through this simulation, we conclude that the proposed constraint formulations provide higher degree of safety in comparison with the state-of-the-art Gaussian approximation.}
\subsubsection{Experiments}\label{subsec:experiments}
We apply Algorithm~\ref{Algo:MP-gPC-SCP} to the scenario shown in~\cref{fig:space_lab_experiment} using the closed-loop described in~\cref{fig:autonomy_loop_gnc} to design and execute safe plans for SS-1 in \cref{fig:space_lab_experiment} under uncertainty in dynamics and obstacle location. This scenario is relevant to the low-earth orbit, on-orbit, servicing application discussed in~\cite{nakka2021information}. For details on the sensing module to estimate the full state, control, and control allocation algorithm, refer to~\cite{nakka2018six}. We use the location of obstacles SS-2$=[-0.46,1.48]$, SS-3$=[-0.71,-0.57]$, SS-4$=[1.3,0.04]$, and Asteroid$=[-2.29,0.34]$ with radius \SI{0.4}{m} as shown in~\cref{fig:spacecraft_simulator_experiments}, and the uncertainty in position of obstacles $\Sigma_{p}=\mathrm{diag}([1e-4,1e-4])$ as input to Algorithms~\ref{Algo:MP-gPC-SCP},~\ref{Algo:PC-gPC-SCP}. The initial and terminal states of SS-1 are $\mathbb{E}(x_{0})=[-0.9,-2.3,0,0,0,0]^\top$ and $\mathbb{E}(x_{f})=[0,2.3,0,0,0,0]^\top$, respectively. We minimize the total control cost $\|u\|_2$ and terminal variance of the trajectory using gPC-SCP method. The uncertainty description in~\cref{subsec:simulations} is valid for the following experimental setup.
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{figures_tro/exp_planning_control_v1.pdf}
\caption{The guidance, navigation and control loop used for planning a distributionally-robust safe trajectory using gPC-SCP and controlling the 3 DOF spacecraft simulators.}
\label{fig:autonomy_loop_gnc}
\end{figure}
\changed{In \cref{fig:spacecraft_simulator_experiments}, we present the results for 10 trials of the closed-loop tracking experiment. We compute an initial anytime trajectory using AO-RRT and optimize it for nominal dynamics. We use the optimized solution as an input for the gPC-SCP method. We observe that the method is biased towards the initial trajectory. As shown, the gPC-SCP method outputs a safe trajectory. We use the mean of the output trajectory as a reference trajectory for the controller. As shown in~\cref{fig:spacecraft_simulator_experiments}, the uncertainty in the model leads to drift in the system. We use a risk measure of $0.05$ for safe collision checking. The gPC-SCP method provides a safe trajectory for control by accommodating the uncertainty in dynamics by using the distributionally robust linear chance constraints. We observe one failure out of the 10 trials of closed-loop tracking. The failure was due to a large disturbance torque from the damaged floor on SS-1. Out of the 10 trials, 7 closed-loop tracking trials reached the expected terminal set. This validates the safety provided by the gPC-SCP method under uncertainty.}
\begin{figure*}
\centering
\includegraphics[width=0.3\textwidth]{mstar_results/rrt_scp_case1.pdf}\includegraphics[width=0.3\textwidth]{mstar_results/gpc_scp_case1.pdf}\includegraphics[width=0.3\textwidth]{mstar_results/feedback_case1.pdf}
\includegraphics[width=0.3\textwidth]{mstar_results/rrt_scp_case2.pdf}\includegraphics[width=0.3\textwidth]{mstar_results/gpc_scp_case2.pdf}\includegraphics[width=0.3\textwidth]{mstar_results/feedback_case2.pdf}
\caption{We show the output of the gPC-SCP method applied to the scenario shown in~\cref{fig:space_lab_experiment} at each stage of Algorithm~\ref{Algo:MP-gPC-SCP} and 10 trials of closed-loop trajectory tracking by using an exponentially-stabilizing controller designed in~\cite{nakka2018six}. Left: We show the output of AO-RRT for 5000 nodes and the SCP for the nominal dynamics. Middle: We show the probabilistic safe trajectory (2$\sigma$-confidence level) generated using the gPC-SCP method with a risk measure of $\epsilon = 0.05$ for collision checking. Right: We observe one failure in the 10 trials of the closed-loop trajectory execution.}
\label{fig:spacecraft_simulator_experiments}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=0.4\columnwidth]{6dof_flyer/6dof_f1.pdf}\includegraphics[width=0.42\columnwidth]{6dof_flyer/6dof_mc.pdf}
\caption{\changed{Left: We show the safe and feasible trajectories computed using gPC-SCP and gPC-SCP$^\mathrm{PC}$ for the Six-DOF spacecraft under uncertainty in dynamics. Right: Monte Carlo simulations of the closed-loop tracking of the trajectory computed using gPC-SCP.}}
\label{fig:6DOF_free_flyer_simulation}
\end{figure}
\vspace{-0.5pt}
\changed{\subsection{Six-DOF Spacecraft Simulator}
We apply the gPC-SCP method and the gPC-SCP$^\mathrm{PC}$ method to a Six-DOF spacecraft simulator model~\cite{nakka2018six} and demonstrate the scalability of the methods to higher dimensions for planning a safe trajectory. The dynamics of the spacecraft simulator in states $\mathbf{x} = [\mathbf{p},\mathbf{v},\mathbf{q},\boldsymbol\omega] \in \mathbb{R}^{13}$ is given as $\dot{\mathbf{p}}=\mathbf{v}$, $m\dot{\mathbf{v}} = \bar{\mathbf{u}} + \sigma_0 \xi_0 \bar{\mathbf{u}}$, $\dot{\mathbf{q}} = \frac{1}{2}\boldsymbol{\Omega}(\boldsymbol\omega)$ and $J \dot{\boldsymbol\omega} = -S(\boldsymbol\omega)J\boldsymbol\omega + \bar{\boldsymbol\tau} + \Xi (-S(\boldsymbol\omega)\boldsymbol\omega + \bar{\boldsymbol\tau}) $,
where $m$ is the mass, $J =\mathrm{diag}([J_1,J_2,J_3])$ is the moment of inertial, $\bar{\mathbf{u}} \in \mathbb{R}^3$ and $\bar{\boldsymbol\tau} \in \mathbb{R}^3$ are force and torque input. We refer the readers to~\cite{nakka2018six} for the definitions of $S(.)$ and $\boldsymbol{\Omega}(.)$, and $\Xi = \textrm{diag}([\sigma_1\xi_1,\sigma_2\xi_2,\sigma_3\xi_3])$. The Gaussian variables $\xi_0$ and $\Xi$ are due to uncertainty in $m$ and $J$, respectively. The uncertainty is a function of both state $\boldsymbol\omega$ and control $[\bar{\mathbf{u}},\bar{\boldsymbol\omega}]$. The mass $m=$\SI{10}{kg}, moment of inertial $J_1=J_2=$\SI{0.07}{kg m^2}, $J_3=$\SI{0.1}{kg m^2}. We have $\sigma_0=0.03$, $\sigma_1=0.002$, $\sigma_1=0.002$, and $\sigma_3=0.002$. The control limits are $[\bar{\mathbf{u}}_{\max},\bar{\boldsymbol\tau}_{\max}] = [1,1,1,0.3,0.3,0.3])$, $[\bar{\mathbf{u}}_{\min},\bar{\boldsymbol\tau}_{\min}]=-[\bar{\mathbf{u}}_{\max},\bar{\boldsymbol\tau}_{\max}]$.}
\changed{We design safe motion plan for the Six-DOF spacecraft in an obstacle rich environment with initial and terminal states as $\mathbb{E}(\mathbf{x}_0)= [-1,-1,-1,0,0,0,0,0,0,1,0,0,0]$ and $\mathbb{E}(\mathbf{x}_f)=[1.2,1.2,1.0,0,0,0,-0.5,0.5,-0.5,0.5,0,0,0]$, respectively. We minimize the control cost $\|[\bar{\mathbf{u}},\bar{\boldsymbol\tau}]\|_2$ and the terminal variance of the trajectory $Q_f = \mathrm{I}$. The risk measure for collision avoidance is $\epsilon_{\mathrm{col}} =0.05$. We show the safe motion plans in~\Cref{fig:6DOF_free_flyer_simulation}. Note that the gPC-SCP and gPC-SCP$^\mathrm{PC}$ compute equivalent mean trajectories. The computation time for gPC-SCP and gPC-SCP$^\mathrm{PC}$ was \SI{40}{s} and \SI{5}{s}, respectively. We execute the designed motion plans using an exponentially-stabilizing feedback controller~\cite{nakka2018six}. Over 1000 Monte Carlo trials of feedback execution, we had 31 constraint violations for our method, while for Gaussian confidence-based collision constraint~\cref{eq:gaussian_linear_constraint}, it was 57.}
\section{Conclusion}\label{sec:conclusion}
We present a generalized polynomial chaos-based sequential convex programming method for safe and optimal motion planning and control under uncertainty in dynamics and constraints. The method uses generalized polynomial chaos projection and distributional robustness to compute a convex subset of the multimodel state-dependent chance constraints and a high-fidelity deterministic surrogate of the stochastic dynamics and the cost functional. \changed{We prove the asymptotic convergence of the surrogate problem to the stochastic optimal control problem. The asymptotic convergence property of the deterministic surrogate and distributional robustness allows for achieving a greater degree of safety. Furthermore, we derive a predictor-corrector formulation to improve the computation speed by an order of magnitude without compromising the safety and optimality of the motion trajectories.}
\changed{We validate our approach in simulations and on the robotic spacecraft simulator hardware and demonstrate a higher success rate in ensuring the safety of motion trajectories compared to a Gaussian approximation of the chance constraints. Our approach outperforms Gaussian approximation in computing safe motion plans under stochastic uncertainty in dynamics and obstacles by at least 20 percent over 1000 trials of feedback execution.}
\section*{Acknowledgement}
The authors thank Amir Rahmani, Fred Y. Hadaegh, Joel Burdick, Richard Murray, and Yisong Yue for stimulating discussions and technical help.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Analysis and control of the multi-agent system (MASs) have attracted the interest of many researchers in the control community. The MASs is generally referred to as a system composed of a set of dynamical agents that interact through a communication network to reach a coordinated behavior or operation \cite{Nguyen2020, Pham2019b}. Especially, the cooperative control problems of MASs have been extensively investigated in the past two decades because there are many practical applications such as energy systems, social networks, brain science, epidemic, etc.
The most interesting problem in the cooperative control of MASs is consensus or synchronization. Many verifiable consensus algorithms have been developed based on continuous-time models and discrete-time models in the past decade, where the close relations between the network connectivity and consensus behavior of MASs are constructed, for example in \cite{Jadbabaie2003a, Olfati-Saber2004, Ren2005}. In \cite{Olfati-Saber2004}, the result of the average consensus of continuous-time MASs with fixed and switching topologies was investigated. It has shown that the strongly connected and balanced directed graphs play a key role in solving average consensus problem. In order to asymptotically convergence analysis of these consensus protocols, the Lyapunov function of disagreement vector is introduced. Moreover, the discrete and continuous consensus algorithms are proposed under dynamically changing topology \cite{Ren2005}. These proved that the consensus can be achieved asymptotically if the union of the directed interaction graphs contains a spanning tree frequently enough. Note here that, the communication on the aforementioned results is only continuous or discrete.
The question for considering the analysis consensus in a clustered network arises from problems that would benefit from a division of a large network into subnetworks (called clusters) that are almost all the time isolated one from another, such as the energy optimization in the wireless sensor network \cite {Halgamuge2003, Kushner2013}, or the problem of formation of multiple unmanned aerial vehicles \cite{Pham2019c, Pham2020c}, where each cluster is a platoon. In each cluster, there exists an agent called a leader who can exchange information outside of its cluster at some specific discrete-time, while the interaction among agents inside each cluster happens in continuous-time.
The main distinctness of this question compared to the analysis consensus in \cite{Jadbabaie2003a, Olfati-Saber2004, Ren2005} lies on two main aspects: 1) The communication among agents in the network either continuous-time or discrete-time, meanwhile the communication in the consideration network is hybrid; 2) The network is composed/ divided of/into by several clusters. This leads to the problem that although each cluster can achieve consensus, the consensus of the overall network is not guaranteed. Thus, some works focus on finding out the sufficient conditions of stability over the clustered network. In particular, the study in \cite{Pham2019a} showed the existence of a sufficient condition to guarantee the overall network asymptotic consensus. In our recent work \cite{Pham2020d,Pham2020c}, a robust formation controller design was proposed for clustered networks of unmanned aerial vehicles. Nevertheless, there is still the loss of mathematical tools to analyze explicitly the consensus over the clustered network under hybrid communication. Moreover, the leader can communicate with any agents outside its cluster instead of with the other leaders. In addition, an external disturbance, and model uncertainty, which are usually considered in the practical networks, are missing on the aforementioned results.
The first contribution of this paper is to analyze the consensus over the clustered network. These take into account the continuous-time communications among agents within each cluster and discrete update information inter-clusters. Our results are more general than ones in \cite{Pham2020c,Pham2020d,Pham2019a, Bragagnolo2016a}, where the communication across clusters is implemented by only leaders, which expresses in \textit{Remark 1}. Thanks to results from matrix and graph theories, it shows explicitly that the consensus over the clustered network can be solved. Next, a sufficient condition will be derived for the robust $\mathcal{H}_{\infty}$ consensus over the clustered network subjected to external disturbances, which forms the second contribution of this work.
The rest of the paper is organized as follows. Section II is dedicated to the formulation of the problem under consideration. In Section III, the consensus problem and the robust $\mathcal{H}_{\infty}$ consensus problems of MASs are analyzed. Simulations are given in Section IV. A conclusion sums up this paper.
\textit{Notations and symbols}
The following notations will be use throughout this paper: $\mathbb{N}, \mathbb{R}$, and $\mathbb{R}^+$ stand for the sets of non-negative integers, real, and non-negative real numbers, respectively. $0_N$ stands for a matrix of zeros with appropriate dimensions. $\textbf{1}_N$ is an $N-$dimensional column vector of 1. We also use $x(t_k^+) = lim_{t\rightarrow t_k, t \geq t_k} x(t)$.
\section{Preliminaries}
\subsection{Graph Theory}
A directed graph $\mathcal{G}$ with its vertex set $\mathcal{V}=\{v_1,\cdots,v_N\}$ and edge set $\mathcal{E} \subseteq \mathcal{V} \times \mathcal{V}$. Each vertex refers to an agent and each edge $(i,j) \in \mathcal{E}$ corresponds to the communication among the agents $i$ and $j$. The neighboring set of the agent $i$ is represented by $\mathcal{N}_i =\{j\in \mathcal{V}: (i,j) \in \mathcal{E}\}$. Moreover, let $a_{(ij)}$ be the elements of the adjacency matrix $\mathcal A$, defined as $\mathcal A = [a_{(ij)}]$, with $a_{(ij)} >0$ if $ (v_j, v_i) \in \mathcal{E}$, and $a_{(ij)} = 0$ otherwise. The Laplacian matrix $ \mathcal{L} = [L_{(ij)}]\in \mathbb{R}^{N\times N}$ is defined as $ L_{(ii)}=\sum_{j\neq i} a_{(ij)}; L_{(ij)}=-a_{(ij)}$. The graph is called weighted whenever the elements $a_{(ij)}$ of its adjacency matrix $\mathcal A$ are other than $0-1$. A weighted directed graph has or contains a directed spanning tree if there exists a node called root such that there exists a directed path from this node to every other node. A weighted directed graph is strongly connected if there is a directed path from every node to every other node.
In the following, we consider that the network $\mathcal{G}$ is composed of $m$ weighted directed clusters $\mathcal{C}_{\tau}, \forall \tau \in \{1,\cdots,m\}$ represented by the graphs $\mathcal{G}_1, \cdots, \mathcal{G}_m$ such that $\mathcal{G}_1=(\mathcal{V}_1,\mathcal{E}_1),\cdots,\mathcal{G}_m=(\mathcal{V}_m,\mathcal{E}_m)$, where $\mathcal{V}=\cup_{\tau=1}^m \mathcal{V}_\tau$ and $V_{\tau} \cap V_g =\varnothing $ for all $\tau,g=1,\cdots m,\tau\neq g$ and $\mathcal{E}=\cup_{\tau=1}^m \mathcal{E}_\tau \cup \mathcal{E}_l$. The graph of each cluster is supposed to be a \textit{directed spanning tree graph}. The communication graph of each cluster $\mathcal{G}_{\tau}$ is represented by a Laplacian matrix $\mathcal{L}_{\tau}$. Each cluster has a specific agent called the leader, and denoted in the following by $l_{\tau } \in \mathcal{V}_{\tau}, \forall \tau \in \{1,\cdots,m\}$. The remaining agents are called followers and are denoted by $f_h$. The set of communication link activated at time $t_k \in \mathbb{N}, t_k \geq 0$ of a time sequence $\{t_k\}$ that satisfies $t_1 <t_2<\cdots, lim_{t_k \rightarrow \infty} t_k =\infty$ with respect to the leader $l_{\tau}$ is defined as $\mathcal{E}_{l_{\tau}} \neq \varnothing$, where
\begin{align*}
\mathcal{E}_{l_{\tau}}=\{(l_\tau,h) \in \mathcal{E}| \exists h, l_\tau, h \neq l_\tau, h \in \mathcal V_{l_\tau}\}
\end{align*}
and $\mathcal V_{l_\tau} =\mathcal{V} \cap \mathcal{V}_\tau \cup l_\tau$. A inter-cluster graph $\mathcal{G}_l=(\mathcal{V}_l, \mathcal{E}_l)$, where $\mathcal{E}_l$ is the set of inter-cluster link activated at time $t_k$ defined as $\mathcal{E}_l=\cup_{\tau=1}^m \mathcal{E}_{l_{\tau}}$ and $\mathcal{V}_l=\cup_{\tau=1}^m \mathcal{V}_{l_\tau}$ is supposed to be \textit{a strongly connected graph}.
\subsection{Consensus protocols}
Consider a group of $N$ agents with with integrator dynamics that interact in $m$ clusters. The dynamics of each agent $i$ is described by
\begin{equation}\label{1}
\dot x_i(t) = u_i(t) + w_i(t),
\end{equation}
where $w_i(t)$ is the external disturbance for agent $i$, which is assumed that $w_i (t)\in \mathfrak{L}_2(0, T]$, where
\begin{align}\label{27}
\|w_i(t)\|_T=\left[\int_0^T \|w_i(t)\|^2 dt\right]^{\frac{1}{2}}=\left[\int_0^T w_i^T(t)w_i(t) dt\right]^{\frac{1}{2}}
\end{align}
Our objective now is to design a distributed consensus protocol for a network divided into some clusters, where only a leader of a cluster can interact outside its cluster at some reset times $t_k$. The protocol herein designed considers that each agent has access to the relative state measurement of its neighbors, and is given by
\begin{align}\label{2}
u_i(t) =\sum_{j=1}^{N} a_{(ij)} \left[x_j(t)-x_i(t)\right],\;\; t \in (t_k,t_{k+1})
\end{align}
Moreover, let us consider that the inter-cluster exchange information at some discrete-time $t_k$ according to
\begin{align}\label{3}
x_i(t_k^+)=\sum_{j =1}^N P_{e(i,j)}x_j(t_k), \;\;t=t_k
\end{align}
where $P_e \in \mathbb{R}^{N \times N}$ is a row stochastic matrix associated to the inter-cluster graph, and therefore, the interaction between the leader $l_\tau$ and other agents outside its cluster can be represented as
\begin{align}\label{5}
x_{l_\tau}(t_k^+)=\sum_{j\in \mathcal V_{{l_{\tau}}}} P_{e(l_\tau,j)}x_j(t_k), \;\;t=t_k
\end{align}
where $\mathcal V_{l_\tau} =\mathcal{V} \cap \mathcal{V}_\tau \cup l_\tau$. In addition, according to \eqref{3}--\eqref{5}, a given some $i\neq l_\tau \in \mathcal{C}_\tau$, if $P_{e(i\neq l_\tau,j)}=0$ for all $i \neq j$ then $P_{e(i\neq l_\tau,i)}=1$. It means that there is no jump occuring on the state of the agent $i$ at discrete-time $t_k$.
Then, the collective dynamics of system \eqref{1} under the consensus protocol \eqref{2} and the interaction the inter-cluster \eqref{3} can be rewritten as
\begin{align}\label{7}
\left\{
\begin{aligned}
&\dot x_i(t)=-\sum_{j=1}^{N} L_{(ij)} x_j(t) + w_i(t),\;\; t \in (t_k,t_{k+1})\\
&x_i(t_k^+)=\sum_{j =1}^N P_{e(i,j)}x_j(t_k), \;\;t=t_k
\end{aligned}
\right.
\end{align}
where $L_{(ij)}$ are element of the Laplacian matrix $\mathcal{L}$.
\subsection{Some useful lemmas}
In the sequel, the following \textit{Lemmas} are recalled.
\begin{lemma}\label{l3}\cite{Ren2005}
Let $\Gamma$ be a compact set consisting of $n \times n$ SIA matrices with the property that for any non-negative integer $k$ and any $B_1,\cdots,B_k \in \Gamma$, the matrix product $\prod_{i=1}^k B_i$ is SIA. Then, for given any infinite sequence $B_1,B_2,\cdots$, there exits a column vector $c^\mathsf T$ such that $\lim_{k \rightarrow \infty}\prod_{i=1}^k B_i=\mathbf{1}c^\mathsf T$.
\end{lemma}
\begin{lemma}\label{l4}\cite{Jadbabaie2003a}
If $B=[b_{ij}]_{n \times n}$ is a stochastic matrix with positive diagonal elements, and the graph associated with $B$ has a spanning tree, then $B$ is SIA.
\end{lemma}
\begin{lemma}\label{l5}\cite{Ren2005}
For any $t >0$, $e^{-\mathcal{L}t}$ is a stochastic matrix with positive diagonal entries, where $\mathcal{L}$ is the Laplacian of graph $\mathcal{G}$.
\end{lemma}
\begin{lemma}\cite{Jadbabaie2003a}\label{l6}
Let $m \geq 2$ be a positive integer and let $D_1,\cdots,D_m$ be non-negative $n \times n$ matrices with positive diagonal elements, then
\begin{align*}
D_1D_2 \cdots D_m \geq \gamma (D_1 +D_2+\cdots+D_m),
\end{align*}
where $\gamma > 0$ can be specified from $D_i, i=1,\cdots,m$
\end{lemma}
\section{System Analysis}
\subsection{Consensus Problem Analysis with $w_i(t)=0$}
In this subsection, consensus problems in the clustered network of MASs are considered. Now we are in the position to present one of the main results.
\begin{theorem}
Consider the collective dynamic system \eqref{7} with $w_i(t) = 0$ and assume that impulsive intervals ($t_k, \;t_{k+1})$ for $k \in \mathbb N$ are uniformly bounded, that is, there exist positive constants $\delta_{min}$ and $\delta_{max}$ such that $\delta_{min} \leq t_{k+1} -t_k \leq \delta_{max}$ for all $k \in \mathbb N$. Then, the proposed control protocol \eqref{2} guarantees that the collective dynamics system \eqref{7} achieves consensus.
\end{theorem}
\begin{proof}
Let $x=[x_1, x_2,\cdots,x_N]^\mathsf T \in \mathbb{R}^{N}$, then the system \eqref{7} can be rewritten in a compact form
\begin{align}\label{10}
\left\{
\begin{aligned}
&\dot x(t)=-\mathcal L x(t),\;\;\;\;\; t \in (t_k,t_{k+1})\\
&x(t_k^+)=P_e x(t_k), \;\;t=t_k,
\end{aligned}
\right.
\end{align}
where $\mathcal{L} \in \mathbb{R}^{N \times N}$ is the Laplacian matrix associated with the graph $\mathcal{G}$
\begin{equation}\label{11}
\mathcal{L}=Diag\{\mathcal{L}_1,\cdots,\mathcal{L}_m\}
\end{equation}
Then, for any initial condition $x(t_0)=x_0$, the solution of \eqref{10} is obtained by
\begin{align}\label{12}
x(t)=e^{-\mathcal L (t-t_k)}x(t_k^+),
\end{align}
and $x(t_k^+)$ can be expressed as
\begin{align}\label{13}
x(t_k^+)&=\lim_{k \rightarrow \infty} \prod_{i=1}^k P_e e^{-\mathcal L (t_i -t_{i-1})}x(t_0).
\end{align}
Moreover, one has
\begin{align}
\sum_{j=1}^{N}P_{e(ij)}=1,
P_{e(ij)} >0 \label{14}.
\end{align}
Subsequently, by employing \eqref{14}, we see that the matrix $P_e \in \mathbb{R}^{N \times N}$ is a row stochastic matrix with positive diagonal elements.
Next, according to \textit{Lemma \ref{l5}}, $e^{-\mathcal L t}$ is a stochastic matrix with positive diagonal entries. Moreover, it follows from \textit{Lemma \ref{l6}} that $P_e e^{-\mathcal L t} \geq \gamma (P_e+e^{-\mathcal L t})$, where $\gamma$ is a positive constant. Therefore, the matrix $P_e e^{-\mathcal L t}$ is a positive diagonal entries.
Due to the inter-cluster communication graph $\mathcal{G}_l=(\mathcal{V}_l, \mathcal{E}_l)$ is supposed to a strongly connected directed graph. Thus, according to \eqref{14}, we can deduce that the graph of $P_e$ has at least spanning tree. Moreover, since $P_e e^{-\mathcal L t} \geq \gamma (P_e+e^{-\mathcal L t})$, and the graph of $e^{-\mathcal L t}$ have at least one spanning tree, one can remark that $P_e e^{-\mathcal L t}$ has a spanning tree.
Based on the above analysis, we showed that the matrix $P_e e^{-\mathcal L t}$ is a row stochastic matrix with positive diagonal elements and that its graph has a spanning tree. Then, according to \textit{Lemma \ref{l4}}, the matrix $P_e e^{-\mathcal L t}$ is SIA.
Therefore, from \textit{Lemma \ref{l3}}, there exits a column vector $c^\mathsf T$ such that
\begin{align}\label{15}
\lim_{k \rightarrow \infty} \prod_{i=1}^k P_e e^{-\mathcal L (t_i -t_{i-1})} =\mathbf{1}_N c^\mathsf T.
\end{align}
Moreover, $e^{-\mathcal L (t-t_k)}$ is a stochastic matrix, which means that $e^{-\mathcal L (t-t_k)}\mathbf{1}_N =\mathbf{1}_N$.
Then, according to \eqref{15}, one has
\begin{align}\label{16}
&e^{-\mathcal L (t-t_k)}\left[ \lim_{k \rightarrow \infty}\prod_{i=1}^k P_e e^{-\mathcal L (t_i -t_{i-1})}-\mathbf{1}_N c^\mathsf T\right]=0\nonumber\\
\Leftrightarrow &e^{-\mathcal L (t-t_k)} \lim_{k \rightarrow \infty}\prod_{i=1}^k P_e e^{-\mathcal L (t_i -t_{i-1})}-\mathbf{1}_N c^\mathsf T =0,
\end{align}
it follows that
\begin{align} \label{17}
e^{-\mathcal L (t-t_k)} \lim_{k \rightarrow \infty}\prod_{i=1}^k P_e e^{-\mathcal L (t_i -t_{i-1})}\rightarrow \mathbf{1}_N c^\mathsf T.
\end{align}
Finally, according to (\ref{17}), \eqref{12} and \eqref{13}, one has
\begin{align}\label{18}
\lim_{t \rightarrow \infty}(x(t)-\mathbf{1}_Nc^\mathsf Tx_0) \rightarrow 0,
\end{align}
or
\begin{align}\label{19}
\lim_{t \rightarrow \infty}(x_i(t)-\underbrace{c^\mathsf Tx_0}_{x^{\star}}) \rightarrow 0,
\end{align}
which implies that the system \eqref{7} can achieve the consensus. This completes the proof.
\end{proof}
\begin{remark}
In the particular case, the inter-cluster interaction is done by only leaders, who interact at some discrete-time $t_k$ through predefined graph $\mathcal{G}_l=(\mathcal{I}, \mathcal{E}_l)$, where $\mathcal{I}=\{l_1,\cdots,l_m \}$ and $\mathcal{E}_l \subset \mathcal{I} \times \mathcal{I}$ (more information please see in \cite{Pham2019c}) such as
\begin{equation}\label{20}
x_{l_\tau}(t_k^+)=\sum_{j=1}^{m}P_{l(l_\tau,j)}x_{l_j}(t_k), t=t_k
\end{equation}
where $P_l \in \mathbb{R}^{m \times m}$ is a row stochastic matrix associated to the graph $\mathcal{G}_l$
\begin{equation}\label{21}
\left\{
\begin{aligned}
&P_{l(l_\tau,j)}=0, \;\;\;\; if \;(l_\tau,j) \notin \mathcal{E}_l \\
&P_{l(l_\tau,j)}>0, \;\;\;\; if \;(l_\tau,j) \in \mathcal{E}_l;i\neq j \\
&\sum_{j=1}^{m}P_{l(l_\tau,j)}=1, \;\;\;\; \forall l_\tau=1,\cdots,m\\
\end{aligned}
\right.
\end{equation}
Then, we can represent the interaction of the inter-cluster by using a extended stochastic matrix $P_e$ as follows
\begin{align}\label{22}
P_e=\mathcal{M}^\mathsf T\left[ {\begin{array}{*{20}{c}}
{{P_l}}&{{0}}\\
{{0}}&{{I_{N-m}}}\\
\end{array}} \right]\mathcal{M} \in \mathbb{R}^{N \times N},
\end{align}
where $\mathcal{M}$ is a permutation matrix. Thus, the equation \eqref{20} can be expressed by
\begin{equation}
x_{i}(t_k^+)=\sum_{j=1}^{N}P_{e(ij)}x_{j}(t_k), \;\;t=t_k,
\end{equation}
Therefore, we achieve a consensus in the clustered network by using the same analysis above. Moreover, according to \eqref{15} and \eqref{19}, one sees that the column vector $c^\mathsf T$ depends on the communication between clusters $P_e$ and the graph of each cluster $\mathcal{L}$. Therefore, the final consensus value of the clustered network (see in \cite{Pham2019c} with the case of simple integrator $\dot{x}_i(t) =u_i(t)$) is
\begin{align}\label{23}
x^{\star}=\underbrace{\frac{\phi^\mathsf T \mathcal{Q}}{\sum_{\tau=1}^{m}\phi_{\tau}}}_{c^\mathsf T}x_0,
\end{align}
\end{remark}
\subsection{Consensus problem with external disturbances}
In the following, we recall the collective dynamical system under external disturbaces, which is discribed as
\begin{align}\label{28}
\left\{
\begin{aligned}
&\dot x_i(t)=-\sum_{j=1}^{N} L_{(ij)} x_j (t)+w_i(t),\; t \in (t_k,t_{k+1})\\
&x_i(t_k^+)=\sum_{j =1}^N P_{e(i,j)}x_j(t_k), \;\;t=t_k
\end{aligned}
\right.
\end{align}
where $L_{(ij)}$ are element of the Laplacian matrix $\mathcal{L}$.
According to the results in Section III.A, let's introduce
\begin{align*}
\psi_i(t)=x_i(t)-c^\mathsf{T}x_0,
\end{align*}
where $c^\mathsf{T}$ is given in \eqref{18}, and note that $\psi_i(t_k^+)=x_i(t_k^+)-c^\mathsf{T}x_0$. Then, the system \eqref{28} can be rewritten as
\begin{align}\label{29}
\left\{
\begin{aligned}
&\dot \psi_i(t)=-\sum_{j=1}^{N} L_{(ij)} \psi_j (t)+w_i(t),\; t \in (t_k,t_{k+1})\\
&\psi_{i}(t_k^+)=\sum_{j=1}^{N}P_{e(ij)}\psi_{j}(t_k), \;\;t=t_k,
\end{aligned}
\right.
\end{align}
\begin{remark}
According to $\psi_i(t)=x_i(t)-c^\mathsf{T}x_0$, it is not difficult to recognize that $\psi_i =0$ if and only if $x_1 =\cdots =x_N=c^\mathsf{T}x_0$. Therefore, we need only to prove that the system (\ref{29}) is asymptotically stable with the disturbance $w(t)$.
\end{remark}
In the following, to suppress the external disturbances in the clustered network, we define an output function $z_i(t)=\psi_i(t)$, then the system \eqref{29} can be described as
\begin{align}\label{30}
\left\{
\begin{aligned}
&\dot \psi_i(t)=-\sum_{j=1}^{N} L_{(ij)} \psi_j(t) +w_i(t),\; t \in (t_k,t_{k+1})\\
&\psi_{i}(t_k^+)=\sum_{j=1}^{N}P_{e(ij)}\psi_{j}(t_k), \;\;t=t_k,\\
&z_i(t)=\psi_i(t), \;\psi_i(t_0^+)=\psi_{i0}
\end{aligned}
\right.
\end{align}
where $z_i(t)$ is the controlled output and $t_0$ is the initial time, and let $\psi=[\psi_1,\psi_2,\cdots,\psi_N]^\mathsf T \in \mathbb{R}^N$, then the system \eqref{30} can be rewritten in a compact form
\begin{align}\label{31}
\left\{
\begin{aligned}
&\dot \psi(t)=-\mathcal{L}\psi(t) +w(t),\; t \in (t_k,t_{k+1})\\
&\psi(t_k^+)=P_{e}\psi(t_k), \;\;t=t_k,\\
&z(t)=\psi(t), \;\psi(t_0^+)=\psi_0
\end{aligned}
\right.
\end{align}
Finally, the robust $\mathcal{H}_{\infty}$ problem (\textit{the controlled output $z(t)$ satisfies $\|z(t)\|_T \leq \rho \|w(t)\|_T$ with $\rho >0$ for any nonzero $w(t) \in \mathfrak{L}_2(0, T]$}) is equivalent to the performance index
\begin{align}\label{39k}
J=\int_{0}^{T} \left(z(t)^\mathsf Tz(t) -\rho^2 w(t)^\mathsf Tw(t) \right) dt <0,
\end{align}
which can be dealt with in the following theorem.
\begin{theorem}
Consider the overall collective dynamical system \eqref{31}. If there exits positive scalars $\alpha >0, \rho >0, 0 <\beta < 1$ and positive-definite matrices $P$ such that
\begin{align}
\begin{bmatrix}\label{31t}
-\mathsf P \mathcal{L} -\mathcal{L}^\mathsf T \mathsf P +\alpha \mathsf P &\mathsf P^\mathsf T\\
\mathsf P&-\rho^2
\end{bmatrix}<0\\
\begin{bmatrix}\label{32}
- \beta \mathsf P & P_{e}^\mathsf T\\
P_{e} & -\mathsf P^{-1}
\end{bmatrix}<0
\end{align}
then
\begin{itemize}
\item [1)] The controlled output $z(t)$ satisfies $J< 0$ for zero-initial condition and for all nonzero $w_i(t) \in \mathfrak{L}_2(0, T]$.
\item[2)] When $w(t)=0$ the system \eqref{31} is stable with exponential convergence rate $\eta=\alpha -\frac{ln \beta}{T}$.
\end{itemize}
\end{theorem}
\begin{proof}
To prove the stability of the (\ref{31}), let's consider the candidate Lyapunov function
\begin{align}\label{32t}
V(t)=V(\psi(t))=\psi^\mathsf T(t) \mathsf P \psi(t)
\end{align}
It is sufficient to prove that the Lyapunov function is satisfied:
\begin{itemize}
\item [i)] At some discrete-time $t_k$, one has
\begin{align}\label{33t}
V(t_k^+)\leq \beta V(t_k) \;\text{or} \; \lim_{t \rightarrow t_k^+} V(t) \leq \beta V(t_k)
\end{align}
where $0 <\beta < 1$.
\item[ii)] Between impulses $t_k$ and $t_{k+1}$, for a prescribed scalar $\rho > 0$, the Lyapunov function satisfying
\begin{align}\label{34t}
\dot{V}(t)+\alpha V(t)+z(t)^\mathsf Tz(t) -\rho^2 w(t)^\mathsf Tw(t) <0.
\end{align}
\end{itemize}
In the sequel, we will show the conditions \eqref{33t} and \eqref{34t} are equivalent to the performance index \eqref{39k} by giving $T>0$. For any given $T \in (t_k, t_{k+1}]$, it follows from \eqref{34t} that
\begin{align}\label{35t}
\int_{t_k}^{t_{k+1}}\left(\dot{V}(t)+\alpha V(t)+z^\mathsf T(t)z(t)-\rho^2 w^\mathsf T(t)w(t \right) dt <0,
\end{align}
and using \eqref{35t} successively on each interval from $t_0$ to $T$ with $\psi(t_0)=0$, one obtains
\begin{align}\label{36t}
\begin{aligned}
&\sum_{v=1}^{k} \int_{t_{v-1}}^{t_{v}} \dot{V}(t)dt+\int_{t_{k}}^{T}\dot{V}(t)dt+
\sum_{v=1}^{k} \int_{t_{v-1}}^{t_{v}} \alpha V(t)dt\\
&+\int_{t_{k}}^{T}\alpha V(t)dt+\int_{t_{0}}^{T} \left(z^\mathsf T(t)z(t) -\rho^2 w^\mathsf T(t)w(t) \right) dt <0
\end{aligned}
\end{align}
Moreover, it follows from \eqref{33t} and $0 <\beta < 1$, one has $V(t_k^+)- V(t_k)\leq -(1-\beta)V(t_k) <0$ and yields
\begin{align}\label{37t}
\begin{aligned}
&\int_{t_{0}}^{t_{1}} \dot{V}(t)dt+\cdots+\int_{t_{k-1}}^{t_{k}} \dot{V}(t)dt+\int_{t_{k}}^{T} \dot{V}(t)dt\\
&=V(T)-\sum_{v=0}^{k}\left(V(t_v^+) -V(t_v)\right) >0.
\end{aligned}
\end{align}
Therefore,
\begin{align}\label{38t}
\begin{aligned}
&\sum_{v=1}^{k} \int_{t_{v-1}}^{t_{v}} \dot{V}(t)dt+\int_{t_{k}}^{T}\dot{V}(t)dt+\\
&+\sum_{v=1}^{k} \int_{t_{v-1}}^{t_{v}} \alpha V(t)dt +\int_{t_{k}}^{T}\alpha V(t)dt>0.
\end{aligned}
\end{align}
It follows from \eqref{36t} and \eqref{38t} that
\begin{align}\label{39t}
\int_{t_{0}}^{T} \left(z^\mathsf T(t)z(t) -\rho^2 w^\mathsf T(t)w(t) \right) dt <0.
\end{align}
In the following, with $t \in (t_k, \;t_{k+1})$ the derivative of Lyapunov function with respect to \eqref{31} is
\begin{align}\label{40t}
\begin{aligned}
\dot{V}(t)= &\psi^\mathsf T(t) (-\mathsf P \mathcal{L} -\mathcal{L}^\mathsf T \mathsf P ) \psi(t) \\
&+ \psi^\mathsf T(t) \mathsf P w(t) + \psi(t) \mathsf P w^\mathsf T(t)
\end{aligned}
\end{align}
and by using \eqref{34t}, one has
\begin{align}\label{41t}
&\psi^\mathsf T (-\mathsf P \mathcal{L} -\mathcal{L}^\mathsf T \mathsf P +\alpha \mathsf P ) \psi +\nonumber\\
&+\psi^\mathsf T \mathsf P w + \psi \mathsf P w^\mathsf T +\psi^\mathsf T \psi -\rho^2 w^\mathsf Tw <0.\\
\Leftrightarrow &-\left[\rho w -\frac{1}{\rho} \mathsf P \psi\right]^\mathsf T\left[\rho w -\frac{1}{\rho} \mathsf P \psi \right]+\nonumber\\
&+\psi^\mathsf T (\frac{1}{\rho^2} \mathsf P ^\mathsf T \mathsf P -\mathsf P \mathcal{L} -\mathcal{L}^\mathsf T \mathsf P +\alpha \mathsf P ) \psi <0.
\end{align}
In order to ensure the condition \eqref{34t}, one has
\begin{align}\label{42t}
\frac{1}{\rho^2} P^\mathsf T \mathsf P -\mathsf P \mathcal{L} -\mathcal{L}^\mathsf T \mathsf P +\alpha \mathsf P <0
\end{align}
and \eqref{31t} is obtained by using Shur complement .
On the other hand, at the reset time $t=t_k$, one has
\begin{align}\label{43t}
V(t_{k}^+)-\beta V(t_k)=\psi^\mathsf T(t_k)(P_{e}^\mathsf T\mathsf P P_{e}- \beta \mathsf P )\psi(t_k)
\end{align}
Then, to guarantee the condition \eqref{32t}, one needs
\begin{align}\label{44t}
P_{e}^\mathsf T\mathsf P P_{e}- \beta \mathsf P <0.
\end{align}
and \eqref{32} is obtained by using again Shur complement.This completes the proof of part 1.
In case of $w(t)=0$, from \eqref{34t}, one obtains
\begin{align}\label{45t}
\dot{V}(t)+\alpha V(t) <0, \; t\in (t_k, t_{k+1})
\end{align}
which obtains that
\begin{align}\label{46t}
{V}(t)<V(t_k^+) e^{- \alpha (t-t_{k+1})} , \; t\in (t_k, t_{k+1})
\end{align}
and at reset time $t=t_k^+$,
\begin{align}\label{47t}
V(t_k^+) \leq \beta V(t_k)
\end{align}
According to the results from \eqref{46t} and \eqref{47t}, one has
\begin{align}\label{48t}
V(t_1) \leq V(t_0^+)e^{- \alpha (t_1-t_0)} , \; t\in (t_0, t_{1})
\end{align}
and
\begin{align}\label{49t}
V(t_1^+) \leq \beta V(t_1) \leq V(t_0^+)\beta e^{-\alpha (t_1-t_0)} , \; t=t_1^+
\end{align}
In general, for $t \in (t_k, t_{k+1})$,
\begin{align}\label{52t}
V(t_k^+)<V(t_0^+) \beta ^k e^{- \alpha (t_{k+1}-t_0)},
\end{align}
and using \eqref{47t}, one has
\begin{align}\label{53t}
V(t)<V(t_0^+) \beta^k e^{- \alpha (t-t_0)} ,
\end{align}
Denoting $T$ as the average impulsive interval of the impulsive sequence $\{t_1,t_2,...\}$, one has $k\geq \frac{t-t_0}{T}-N_0$, where $N_0$ is a positive integer \cite{Lu2010}, then
\begin{align}\label{54t}
V(t)&<V(t_0^+) \beta^{\frac{t-t_0}{T}-N_0} e^{- \alpha (t-t_0)} ,\nonumber\\
&<V(t_0^+) \beta^{-N_0}e^{- (\alpha -\frac{ln \beta}{T})(t-t_0)}
\end{align}
It means that the system \eqref{31} is stable with $w(t)=0$. This completes the proof of part 2.
\end{proof}
\begin{remark}
$-\mathcal{L}$ has a $m$ unique zero eigenvalues and other $N-m$ eigenvalues which has negative a real parts \cite{Pham2019a}, then there always exist the solution for LMIs \eqref{31t}.
\end{remark}
\section{Simulation Results}
In this section, numerical simulation is given to validate the results reported in the previous section. The network contains 7 agents and it is partitioned into 2 clusters having 4 and 3 elements, respectively. Moreover, let us suppose that the graphs of clusters $\mathcal{C}_1$ and $\mathcal{C}_2$ are respectively represented by the Laplacian matrices $\mathcal{L}_1$ and $\mathcal{L}_2$
\begin{align*}
\mathcal{L}_1=
\begin{bmatrix}
2 &-1&0&-1\\
-1 &2&-1&0\\
-1 &-1&2&0\\
-1&-1&0&2\\
\end{bmatrix}; \mathcal{L}_2=\begin{bmatrix}
3&-3&0\\
0 &2&-2\\
-1&0&1\\
\end{bmatrix}
\end{align*}
Each cluster has only one agent, who is able to interact with agents outside its own cluster (agent 1 in the first cluster and agent 5 in the second cluster). The weights of the inter- cluster interactions are chosen as follows
\begin{align*}
P_e=\begin{bmatrix}
0.5&0&0&0&0&0.5&0\\
0&1&0&0&0&0&0\\
0&0&1&0&0&0&0\\
0&0&0&1&0&0&0\\
0&0.1&0.1&0&0.8&0&0\\
0&0&0&0&0&1&0\\
0&0&0&0&0&0&1\\
\end{bmatrix}
\end{align*}
According to \eqref{5}, the reset dynamics of the leader 1 (agent 1) and leader 2 (agent 5) are
\begin{align*}
&x_{l_1}(t_k^+)=0.5 x_{l_1}(t_k) +0.5 x_6(t_k), \;x_{l_1}=x_1\\
&x_{l_2}(t_k^+)=0.1 x_2(t_k)+0.1 x_3(t_k)+0.8 x_{l_2}(t_k), \;x_{l_2}=x_5
\end{align*}
and by by supposing that $x(0)=[0 -1 -2 -4 \;\; 2\;\; 3\;\; 4]^\mathsf T$. As a result, Fig. 1 depicts the state trajectories of agents in the clustered network and indicates that the MASs can reach consensus, which is consistent with the result of \textit{Theorem 1}.
\begin{figure}[htb]
\centering
\includegraphics[height=4.7cm]{figure/F4.eps}
\caption{Consensus of 7 agents grouped in 2 clusters}
\end{figure}
Moreover, if the inter-cluster interaction is done by only leaders, who interact at some discrete-time $t_k$ through predefined graph $\mathcal{G}_l=(\mathcal{I}, \mathcal{E}_l)$ corresponding to the matrix $P_l$
\begin{align*}
P_l=\begin{bmatrix}
0.45 &0.55\\
0.55 &0.45\\
\end{bmatrix} \Rightarrow w^T =[0.5\;\;0.5]
\end{align*}
and the eigenvectors of matrix $\mathcal{L}_1$ and $\mathcal{L}_2$ are
\begin{align*}
\mathsf{r}_1^T=[1/3\;\;1/3\;\;1/6\;\; 1/6];\mathsf{r}_2^T=[2/11\;\;3/11\;\;6/11]
\end{align*}
then by using \eqref{23} we obtain the following consensus value $x^{\star}=2.513$. Finally, we investigate the case consensus problem with external disturbance $w(t)=0.1*sin(2t)$. By choosing $\alpha =1, \rho=1$ and $\beta =0.7$, and solving LMI \eqref{31t} and \eqref{32} in \textit{Theorem 2}, one has
\begin{align*}
\small
\mathsf P=\begin{bmatrix}
2.36 &-0.69 &-0.42 &-1.25 &0&0&0\\
-0.69 &2.36 &-1.25 &-0.42&0&0&0\\
-0.42 &-1.25 & 1.97 &-0.30 &0&0&0\\
-1.25 &-0.42 &-0.30 & 1.96 &0&0&0\\
0&0&0&0&2.02 &-1.6 &-0.39\\
0&0&0&0&-1.6& 2.94& -1.30\\
0&0&0&0&-0.39 & -1.3 &1.69\\
\end{bmatrix}
\end{align*}
\begin{figure}[htb!]
\centering
\includegraphics[height=6.4cm]{figure/F3.eps}
\caption{Consensus of 7 agents grouped in 2 clusters under disturbances}
\end{figure}
The result simulation is shown in the Fig. 2 and the evolution of the Lyapunov function $V$ is also depicted in the Fig.2, which is consistent with the result of \textit{Theorem 2}.
\section{Conclusion}
In this paper, a novel approach has been proposed to design distributed consensus controllers for general linear MASs with the following features. First, the considered networks are partitioned into clusters, where the communication between agents inside each cluster is continuous, but the cluster's leader interacts outside its cluster at some reset times. Second, thanks to results from matrix theory and algebraic graph theory, the consensus problem in the clustered network is analyzed. Third, sufficient conditions for the robust $H_{\infty}$ stability of this equivalent system were derived from solutions of local convex LMIs problems, which can be solved in a distributed manner. A possible example of our proposed approaches was illustrated.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:intro}
The expected presence of physics beyond the Standard Model (BSM) is motivated by the observation of non-vanishing neutrino masses, the overwhelming evidence for dark matter (DM), and the need for an effective mechanism to explain the baryon asymmetry; in addition, theoretical issues such as the stabilization of Higgs vacuum, also point towards the presence of new physics (NP). The search for such new particles and interactions is one of the central programs at the Large Hadron Collider (LHC). Despite this effort, and excepting the discovery of the long sought-after Higgs boson in 2012 \cite{Aad:2012tfa,Chatrchyan:2012ufa}, no direct observation of new physics at the LHC (or other experiments) has been confirmed, though there are significant hints \cite{Lees:2012xj,Aaij:2017tyk,Lees:2013uzd,Aaij:2017deq,Aaij:2021vac,Bennett:2002jb,Bennett:2006fi,Aoyama:2012wk,Aoyama:2020ynm,Abi:2021gix}.
One major challenge facing the search for NP at the LHC is its large QCD background that makes the detection of possible weakly-coupled BSM physics difficult. Yet this type of NP is expected in several scenarios ({\it e.g.} many DM and neutrino mass generation paradigms), which makes the prospect of an electron-positron ($ e^+e^-$) collider, such as the International Linear Collider (ILC) \cite{Behnke:2013xla}, a very attractive possibility for probing a variety of BSM physics. It is the goal of this paper to study some of the capabilities of the ILC and determine its sensitivity to simple NP extensions of the SM: we consider first the sensitivity of the ILC to an additional heavy vector-like fermion, and then to a SM extension that includes, in addition, a viable DM candidate. The discussion here presented can be easily extended to other proposed $ e^+e^-$ colliders and to a diversity of other types of hypothesized BSM physics.
The current bounds on heavy lepton masses depend on their nature (chiral or vector-like) and dominating decay channels. For example, LEP put a bound on the mass of $101.2\,\hbox{GeV}$ (95\% CL) on the mass of heavy, unstable, singly-charged fermion \cite{Achard:2001qw} when its main decay channel is $\nu W^\pm$, while the bound is $102.6\,\hbox{GeV}$ (95\% CL) if it is stable \cite{Tanabashi:2018oca}. Searches at LHC have been so far in three main directions: {\it(i)} triplet leptons associated with type III seesaw mechanism for neutrino mass generation \cite{Sirunyan:2017qkz}; {\it(ii)} stable or long-lived charged leptons; and {\it(iii)} superpartners of the SM gauge bosons (neutralino and chargino). In the first case, CMS has put a ($3\sigma$) bound of $840\,\hbox{GeV}$ \cite{Sirunyan:2017qkz} (using $137\,\text{fb}^{-1}$ of data at $\sqrt s=13\,\hbox{TeV}$). The current bound for a long-lived singly charged fermion is $\sim$ 574 GeV \cite{Chatrchyan:2013oca} (CMS, using $18.8\,\text{fb}^{-1}$ of data at $\sqrt s=8\,\hbox{TeV}$). The limit on chargino mass in supersymmetric theories from production of chargino pairs \cite{Aad:2019vnb} is $\sim 400\,\hbox{GeV}$ when the neutralino mass is zero, and $\sim 250\,\hbox{GeV}$ from chargino-neutralino pair production \cite{Aad:2019qnd} (both obtained at $13\,\hbox{TeV}$ CM energy).
In our discussion below we will first study the detectability of a singly-charged lepton with mass of either $150\,\hbox{GeV}$ or $ 245\,\hbox{GeV}$ at the ILC, with a center-of-mass (CM) energy of $500\,\hbox{GeV}$ and determine the optimal statistical precision to which its couplings to the $Z$ boson can be measured using the optimal-observable technique (OOT) \cite{Atwood:1991ka, Davier:1992nw, Diehl:1993br, Gunion:1996vv}. Charged fermion pair production in the context of type-III seesaw framework has been studied in literature \cite{Das:2020gnt, Das:2020uer}, but no study has been done yet using the optimal observable approach. We will then consider this particle in the context of a specific NP model and provide an event-level collider simulation of its dominating decay channel; this model has the added feature of containing a viable dark matter candidate, some of whose effects at the ILC will also be considered. We will discuss the effects of beam polarization and the extent to which the conclusions drawn for these specific cases can be generalized.
The OOT has been used previous in a variety of studies, including the estimation of the uncertainty of the Higgs couplings \cite{Gunion:1996vv,Hagiwara:2000tk,Dutta:2008bh} and top-quark couplings at $e^+e^-$ colliders \cite{Grzadkowski:1996pc,Grzadkowski:1997cj,Grzadkowski:1998bh,Grzadkowski:1999kx,Grzadkowski:2000nx}, of the top-quark interactions in a $\gamma\gamma$ collider \cite{Grzadkowski:2003tf,Grzadkowski:2004iw,Grzadkowski:2005ye}, of the CP properties of Higgs boson at a muon collider \cite{Hioki:2007jc}, and of possible non-standard top-quark couplings at LHC \cite{Gunion:1998hm,Hioki:2012vn, Hioki:2014eca}; other studies using this technique include estimating the sensitivity to NP effects in flavor physics \cite{Bhattacharya:2015ida, Calcuttawala:2017usw, Calcuttawala:2018wgo} and NP searches in top-quark production at $e\gamma$ colliders \cite{Cao:2006pu}.
Our paper is organized as follows: the OOT is described in section \ref{sec:method}; the phenomenological model that we will use to study the $Z$ couplings of a heavy charged lepton is presented in \ref{sec:model}; sections \ref{sec:cross-section} and \ref{sec:analysis} discuss the relevant cross-section calculations and OOT for this model; the UV-complete model and associated collider signals are examined in section \ref{sec:DMmodel}; with section \ref{sec:conclusion} containing parting comments and conclusions.
\section{Optimal uncertainties}
\label{sec:method}
This section contains a summary of several results concerning the statistical uncertainty of experimental observables. These results have appeared previously (see, {\it e.g.}, \cite{Diehl:1993br}); they are included here for convenience and to ensure uniformity of notation.
We consider models where the SM has been complemented by some type of new physics; the (theoretical) differential cross section for any given collider process involving the production of new particles can be written in the form
\beq
\mathcal{O}=\frac{d\sigma_{\tt theo}}{d\phi} = \sum_i c_i f_i(\phi) \,,
\label{eq:expnd1}
\eeq
where $ \phi $ denotes the appropriate phase-space coordinates and the coefficients $c_i $, composed of (sums of products of) coupling and numerical constants, parametrize the process in terms of the linearly-independent functions $ f_i $. In the following, we will discuss 2 $\rightarrow$ 2 scattering process for which there is a single phase-space variable, that we take as the CM scattering angle; naturally, $\phi$ changes according to the process under consideration and experimental convenience. The separation of coefficients $c_i$ and functions $f_i $ is not unique -- we will comment on this below.
The goal is now to determine the coefficients $c_i$ as accurately as possible. If one assumes a constant event rate together with the fact that an experiment occurs over a finite time, the event number follows a Poisson distribution, then the optimal covariance matrix becomes
\beq
V_{ij} =\frac{ M_{ij}^{-1} \sigma_T}{N}= \inv{\mathfrak{L}}_{\tt int} M_{ij}^{-1}\, ,
\label{eq:covmat2}
\eeq
with
\beq
M_{ij} =\int \frac{f_i(\phi) f_j(\phi)}{\mathcal{O}(\phi)}d\phi\, ,
\label{eq:mij}
\eeq
where $\sigma_T=\int \mathcal{O}(\phi) d\phi$ and N is total number of events ($N=\sigma_T {\mathfrak{L}}_{\tt int} $). ${\mathfrak{L}}_{\tt int} $ denotes the integrated luminosity over this period. The detailed
derivation of the covariance matrix as in Eq.~\eqref{eq:covmat2} has been furnished in the Appendix \ref{sec:covmat} for the convenience of the readers. This can also be
achieved by choosing a weighting function $w_i(\phi)$ such that $c_i=\int w_i(\phi) \mathcal{O}(\phi)d\phi$ \cite{Gunion:1996vv}, where the expression of $w_i(\phi)$ is given by,
\beq
w_i(\phi)=\frac{\sum_{j}M_{ij}^{-1}f_{j}(\phi)}{\mathcal{O}(\phi)}.
\eeq
$V$ can be used to estimate the width of the distribution of the $c_i$ as follows. We {\em assume} that these parameters have average (or `seed') values $ c_i^0 $ and define ($ \epsilon$ is an efficiency factor discussed below)
\beq
\chi^2= \epsilon \sum_{\{i,j\}=1}^{n} \delta c_i \, \delta c_j \left( V_0^{-1}\right)_{ij}, \qquad \delta c_i=c_i-c_i^0, \qquad V_0 = \left. V\right|_{ c=c^0 }\,;
\label{chi2}
\eeq
In practice the optimal observable technique (OOT) consists in using this covariance matrix to determine statistical uncertainties and correlations between the coefficients $ c_i $.
Regarding $ \chi^2$ as a random variable, one can determine the probability ${\sf p}_n(\ell)$ (often termed the confidence level, C.L.) for $ \chi^2 \le \ell $ to occur. If the $c_i$ are normally distributed this is given by the usual $ \chi^2$ distribution with $n$ degrees of freedom: ${\sf p}_n(\ell) = 1- \Gamma(n/2,\ell/2)/\Gamma(n/2) $. In particular, ${\sf p}_2(1) = 39.3\%$, and ${\sf p}_3(1) = 19.8\%$ so that for $n=2,\,3$ the $ \ell=1 $ C.L. is relatively low; a $68\%$ C.L. requires $ \ell = 2.3$ for $n=2$ and $ \ell=3.5$ for $n=3$. In the discussion below we will be mainly concerned with the regions determined by $ \chi^2 \le \ell $ for a given $ \ell$, referring to them as the $ \sqrt \ell$-$\sigma$ regions; this can be contrasted to the common usage of ``1-$\sigma$ standard deviation" referred to $\chi^2=2.3$ for 2 parameter space and $\chi^2=3.5$ for 3 parameter space. An illustration on how the $ 1$-$\sigma$ regions change when we use a given C.L. is discussed in Appendix \ref{sec:68cl}.
In the following we will consider the (electron-positron) collider production of new physics (NP) which in turn decays to SM states; symbolically, $e^+ e^- \rightarrow$NP$\rightarrow$SM. We denote the `hard' cross section for $ e^+ e^- \rightarrow$NP production by $\sigma^{\tt NP}$, and by $\sigma^{\tt FS}$ the final-state cross section $ e^+ e^- \rightarrow$NP$\rightarrow$SM, including all event selection cuts aimed at reducing and SM background and enhancing the NP contribution. The efficiency factor $\epsilon$ in Eq.~\eqref{chi2} is then defined by the ratio
\beq
\epsilon=\frac{\sigma^{\tt FS}}{\sigma^{\tt NP}}~,
\label{eq:eps.def}
\eeq
We would also like to note further that the statistical analysis done in section \ref{sec:model} is based on the NP signal process without including the effects of
SM backgrounds, since this requires a specific model for a detailed characterization of the final state events (we return to this in section \ref{sec:DMmodel}).
However, the efficiency $\epsilon$ in Eq.~\eqref{chi2} includes not only the branching ratio of NP$\rightarrow$ SM final state, but also the effects of event selection cuts
that suppress the SM background contamination. The values of $ \epsilon $ must be then estimated using a complete model of NP production and decay; in the
next section we will assume $\epsilon=0.001$ and $ 0.005 $, justified by the analysis of the specific model of section \ref{sec:DMmodel}. The use of $ \epsilon$
to include these effects is, of course, an approximation; it is appropriate for the type of situations we consider: the resonant production of new particles which then
decay into a SM final state. This approximation would not be appropriate in processes where on-shell NP particle is similar in mass and spin to that of a SM particle
leading to same signal and providing large interference (for example, a new $Z^{'}$ boson having similar mass to SM $Z$ boson), or when the new particle contribution
to the signal is virtual or in narrow-width s-channel resonance. In all such cases the cross section in Eq.~\eqref{eq:expnd1} receives also a SM contribution, and the corresponding
OOT must be modified ({\it cf. e.g.} \cite{Diehl:1993br}); we will return to this issue in a future publication. However, the procedure as adopted here, will be less
conclusive, given a large irreducible SM background contribution and the estimation of $\epsilon$ will be limited in this case.
As noted earlier the choice of $c_i $ and $f_i $ is not unique; in practice one uses a separation that is convenient computationally and, if possible, has some physical motivation. The final results are independent of this choice in the sense that, if we use different functions and coefficients, $f_i = \sum u_{ij} \tilde f_j $ and $ \tilde c_i = \sum u_{ji} c_j $, where $u_{ij}$ is a constant invertible matrix, $ \chi^2 $ in Eq.~\eqref{chi2} is invariant.
The covariance matrix $V$ depends on the physical process under consideration and on the experimental parameters such as collider energy and luminosity. Therefore the above expression can also be used to determine the (minimal) collider properties that are required to obtain a given desired statistical uncertainty.
The seed coefficients $ c_i^0 $ take different values depending on the type of new physics being considered. One can then take a different approach and regard Eq.~\eqref{eq:expnd1} as a generic expansion of the cross section under consideration in a convenient basis of functions $ f_i$. If a model has parameters $ p_a $, then the $c_i = c_i(p_a)$ and $ \delta c_i = c_i(p_a^0 + \delta p_a) - c_i(p_a^0)$; from which the statistical uncertainties and correlations of the $ p_a$ can be readily extracted; an example of this procedure when is presented in the next section. The case where the $c_i$ are linear combinations of the $p_a$ is considered in appendix \ref{sec:ellipses}. The number of parameters $p_a$ can be larger than the number of coefficients $c_i $; in which case the measurements under consideration provide a consistency test of the model.
\section{Phenomenological framework}
\label{sec:model}
In this and the following sections we will use the OOT to determine the accuracy to which the parameters of a simple model of BSM physics can be measured at the projected International Linear Collider (ILC). The model we consider is a simple extension of the SM by the addition of a heavy charged fermion $ \psi^\pm $, that can be produced by $Z$ and photon exchange (Fig.~\ref{fig:productn}). We will discuss the precision to which the OOT allows the determination of the $\psi $ couplings to the $Z$ at an $e^+e^-$ collider.
This type of heavy fermion appears in various extensions of SM; {\it e.g.} those containing a fermion isodoublet $\left(\psi^0,\psi^-\right)$ with hypercharge $Y_\psi=-1$; we elaborate upon a possible model framework below (Sect. \ref{sec:DMmodel}). Here we adopt a purely phenomenological approach, allowing $ \psi^\pm $ to have general chiral couplings to the $Z$ boson\footnote {We postpone any constraints coming from chiral anomalies to our discussion of a specific model.}:
\beq
\psi^+\psi^-Z:- ~\frac{i e_0}s_{2\tt w} \gamma^\mu\left(a+b\gamma^5\right)~,
\label{eq:vertex1}
\eeq
(where $e_0=U(1)_{\tt em}$ coupling, and $s_{2\tt w} = \sin(2 \theta_{\tt w})$; $\theta_{\tt w}$ is the weak-mixing angle) assuming for simplicity~\footnote{It is worth noting that in weakly-coupled theories modifications to the photon minimal coupling are generated at 1 or higher loops and are correspondingly suppressed.} that it has the usual minimal coupling to the photon:
\beq
\psi^+\psi^-\gamma:- ~ie_0\gamma^\mu \,.
\label{eq:vertex2}
\eeq
The paramters $a,\,b$ correspond to the $ p_a $ discussed briefly at the end of section \ref{sec:method}.
\begin{figure}[htb!]
$$
\includegraphics[scale=0.27]{prodctn.png}
$$
\caption{Production of heavy charged fermions ($\psi^+\psi^-$) at $e^-e^+$ collision (ILC).}
\label{fig:productn}
\end{figure}
We will call any specific choice of $a,\,b$ a {\em hypothesis} and the corresponding parameters as {\em seed parameters}, of which we will consider the following:
\begin{itemize}
\item $a=\pm1$, $b=0$ (pure vector coupling).
\item $a=0$, $b=\pm 1$ (pure axial vector coupling).
\item $a=\pm1$, $b=\pm 1$ (chiral coupling).
\end{itemize}
We note that, for this simple model, $ b \rightarrow - b $ under a parity transformation, so we need to consider only $ b\ge 0 $.
Using the couplings in Eqs.~\eqref{eq:vertex1} and \eqref{eq:vertex2}, we can evaluate $ d\sigma_{\tt theo} $ and, upon selection of the $f_i $, extract the coefficients $ c_i = c_i(a,b) $; the hypothesis $ a=a^0,\,b=b^0$ corresponds to assuming that these coefficients have seed values $ c_i^0 = c_i(a^0,b^0)$ ({\it cf.} comments at the end of sect. \ref{sec:method}). We then use Eq.~\eqref{eq:covmat2} to compute the covariance matrix $ V$ and corresponding $ \chi^2$; the regions $ \chi^2 < $const. determine the optimal statistical uncertainties \cite{Gunion:1996vv, Calcuttawala:2017usw}, and the accuracy to which different hypotheses can be differentiated.
For the calculations below we will assume the following collider parameters:
\beq
m_{\psi^{\pm}} = 150\,\hbox{GeV}\,,~\text{or}~ 245\,\hbox{GeV} \nonumber\\[6pt] \,, \qquad \sqrt{s} = 500 \, \hbox{GeV}\,; \quad {\mathfrak{L}}_{\tt int} =567 \, \text{fb}^{-1}\,,
\label{eq:params}
\eeq
where $\sqrt{s}$ is the CM energy of the collider and integrated luminosity $ {\mathfrak{L}}_{\tt int} $, whose values were taken from the ILC design parameters~\cite{Behnke:2013xla}. The lower value of $ m_{\psi^\pm} $ is chosen above the current collider limit of $O( 100 )\,\hbox{GeV} $ \cite{Tanabashi:2018oca} (Sect. \ref{sec:intro})~\footnote{This limit is obtained using the $ \psi^\pm \rightarrow W^\pm+$neutral decay, which naturally occurs in the simplest models containing a $ \psi^\pm $; see Sect. \ref{sec:DMmodel}.}; the higher value is chosen to be close to threshold.
Our analysis is carried out for an $ e^+ e^- $ linear collider because {\it(i)} it provides a much cleaner platform where QCD processes are suppressed, and so provides much better opportunity for the precision measurements we consider here; {\it(ii)} the expected availability of (partially) polarized beams allows a better probe of the new physics we are considering; and {\it(iii)} the construction of the covariance matrix and $\chi^2$ can be done analytically, avoiding insertion of the quark distribution functions that are unavoidable in a hadron collider.
\subsection{The $\psi^+\psi^-$ production cross section at an $e^+e^-$ collider}
\label{sec:cross-section}
The amplitude for the process $ e^+e^- \rightarrow \psi^+ \psi^-$, which we denote by ${\cal M}(\lambda_{e^-},\,\lambda_{e^+},\,\lambda_\psi,\,\lambda_{\bar\psi} )$ (where $\lambda_i = \pm1$ denotes the helicity of particle $i$), is easily calculated \cite{Vega:1995cc}:
\bal
{\cal M}(\lambda_{e^-},\,- \lambda_{e^-},\,\lambda_\psi,\,-\lambda_\psi ) &= - e e_0 \left(\lambda_{e^-} \lambda_\psi + \cos\theta \right) \left[ 1 + \xi \left(a + b \lambda_\psi \beta_\psi \right) \right] \,; \quad \xi = \xi_1+\lambda_{e^-} \xi_2\,,\nonumber\\[6pt]
{\cal M}(\lambda_{e^-},\,- \lambda_{e^-},\,\lambda_\psi,\,\lambda_\psi ) &= - e e_0\left( \frac{2 m_{\psi^{\pm}} \lambda_\psi \sin\theta}{\sqrt{s}} \right) \left( 1 + \xi a \right) \,,
\label{eq:amplitude}
\end{align}
where $e$ is the electron charge, $\sqrt{s}$ the CM energy, $ \beta_\psi = \sqrt{1 - 4 m_{\psi^{\pm}}^2/s}$, and
\beq
\xi_1=\frac{C_v}{s_{2\tt w}^2 (1 - m_{\tt z}^2/s)}\,, \qquad \xi_2=\frac{C_a}{s_{2\tt w}^2 (1 - m_{\tt z}^2/s)} \,,
\eeq
with $C_v = ( 4 s_{\tt w}^2-1)/2,\, C_a=1/2$, the vector and axial couplings of the electron to the $Z$, respectively (and $ s_{\tt w} = \sin\theta_{\tt w}$). If $ \hat{\bf p}_{e^-}$ and $ \hat{\bf p}_{\psi^-} $ are unit vectors parallel to the corresponding momenta, then the scattering angle $ \theta $ is defined by $ \cos\theta = \hat{\bf p}_{e^-} \cdot \hat{\bf p}_{\psi^-} $.
\begin{figure}[htb!]
$$
\includegraphics[scale=0.36]{cs_comprsn.png}
\includegraphics[scale=0.5]{dcs_costh_comprsn.png}
$$
\caption{Left: Total spin-averaged cross-section for $e^+e^-\rightarrow \psi^+\psi^-$ as a function of the CM energy $\sqrt{s}$; right: differential spin-averaged cross-section as a function of the scattering angle for c.o.m energy ($\sqrt s$) $=500$ GeV. We took $m_{\psi^\pm} =245$ GeV and the collider parameters of Eq.~(\ref{eq:params}).}
\label{fig:cs}
\end{figure}
Using Eq.~(\ref{eq:amplitude}), the cross-section when the $ e^\pm$ beams have partial polarizations $P_{e^\pm} $ (with $ -1\le P_{e^\pm}\le1$) is given by
\bal
\label{eq:poldiffcrosssection}
\frac{d\sigma(P_{e^+},\,P_{e^-})}{d\Omega} &= \frac{(1-P_{e^-})(1+P_{e^+})}4 \left( \frac{d\sigma}{d\Omega}\right)_{\lambda_{e^-}=-1} + \frac{(1+P_ {e^-})(1-P_{e^+})}4 \left( \frac{d\sigma}{d\Omega}\right)_{\lambda_{e^-}=1}\,, \nonumber\\[6pt]
&= \sum c_i f_i\,,
\end{align}
where we choose \footnote{It is straightforward to verify that these functions are linearly independent}
\beq
\{f_1,\,f_2,\,f_3\} = \frac{ \beta_\psi }{2 s} \left\{ (2 - \beta_\psi^2),\, \beta_\psi \cos \theta,\,\beta_\psi^2 \cos^2\theta \right\}\,,
\label{fi}
\eeq
and
\begin{align}
\frac{c_1}{\alpha \alpha_0} =&\frac{1-P_{e^-}P_{e^+}}{2}\bigg[ 1+2\xi_1 a+(\xi_1^2+\xi_2^2) \left( a^2+\frac{\beta_{\psi}^2}{2-\beta_{\psi}^2} b^2 \right) -2P_{\tt eff}\bigg\{\xi_2 a+\xi_1\xi_2 a^2+\frac{\beta_{\psi}^2}{2-\beta_{\psi}^2}\xi_1\xi_2b^2\bigg\}\bigg];\nonumber\\[6pt]
\frac{c_2}{\alpha \alpha_0} =&\frac{1-P_{e^-}P_{e^+}}{2}\bigg[2\xi_2b+4\xi_1\xi_2ab-P_{\tt eff}\bigg\{2\xi_1b+(\xi_1^2+\xi_2^2)ab\bigg\}\bigg];\nonumber\\[6pt]
\frac{c_3}{\alpha \alpha_0} =&\frac{1-P_{e^-}P_{e^+}}{2}\bigg[ 1+2\xi_1a+(\xi_1^2+\xi_2^2)(a^2+b^2) -2P_{\tt eff}\bigg\{\xi_2a+\xi_1\xi_2(a^2+b^2)\bigg\}\bigg];
\label{ci}
\end{align}
We defined
\beq
P_{\tt eff} = \frac{P_{e^-}-P_{e^+}}{1-P_{e^-}P_{e^+}}.
\eeq
and $ \alpha_0 = e_0^2/(4\pi)$, while $\alpha = e^2/(4\pi) $ is the usual fine-structure constant. It is also useful to note that $ d\sigma/d\Omega $ is invariant under
$ b \rightarrow - b$ and $ \theta \rightarrow \pi - \theta $, a consequence of the invariance of Eqs.~(\ref{eq:vertex1}) and (\ref{eq:vertex2}) under CP.
The spin-averaged total and differential cross-sections (corresponding to $ P_{e^\pm} =0 $) for different seed values of $\{a,b\}$ are plotted in Fig.~\ref{fig:cs}. It is worth noting that the total cross sections exhibit the same behavior for the values of $a$ and $b$ considered, especially at large $s$; this is due to a combination of two effects. First, since $ s_{\tt w}^2 \simeq1/4 $, $ \xi_1 \sim 0 $; second, for large $s$, $ \beta_\psi \simeq 1$; it follows that the average cross section $ \sigma \propto 1+ \xi_2^2 (a^2 + b^2) $, explicitly displaying its dependence on $a$ and $b$. In contrast the unpolarized {\em differential} cross section depends on $ c_2 \simeq \xi_2 b $ and will have a very different behavior depending on the values of $b$. As expected from unitarity, the total cross section drops with increasing CM energy.
\begin{figure}[htb!]
$$
\includegraphics[scale=0.28]{totsigma_com_plot_a1b0_comprsn.png}
\includegraphics[scale=0.28]{totsigma_com_plot_a1bpm1_comprsn.png}
\includegraphics[scale=0.28]{totsigma_com_plot_a0bpm1_comprsn.png}
$$
$$
\includegraphics[scale=0.3]{totsigma_com_plot_a-1b0_comprsn.png}
\includegraphics[scale=0.3]{totsigma_com_plot_a-1bpm1_comprsn.png}
$$
\caption{Total cross section as a function of the CM energy for unpolarized ($P_{e^\pm}=0$) beams (solid line); $P_{e^\mp}=^{-50\%}_{+10\%}$ (dashed line), and $P_{e^\mp}=^{-80\%}_{+20\%}$ (dashed-dot line) for various seed values of $a$ and $b$.}
\label{fig:cscom1}
\end{figure}
According to the design report \cite{Behnke:2013xla}, the ILC will produce highly polarized electron beam and moderately polarized positron beam; we will choose $ P_{e^-}=-0.8,\,P_{e^+}=+0.2 $ when considering this option. Also, for polarized beams we have
\bal
a>0: & \quad \sigma\left(P_{e^-},\,P_{e^+} \right) \ge \sigma\left( P_{e^\pm} =0 \right)\,, \nonumber\\[6pt]
a<0: & \quad \sigma\left(P_{e^-},\,P_{e^+} \right) \le \sigma\left( P_{e^\pm} =0 \right)\,,
\end{align}
whence it follows that polarization will enhance detectability. We illustrate these features in figures \ref{fig:cscom1} where we plot the total cross section for various seed values of $a,b$ and choices of $ P_{e^\pm}$.
\subsection{Optimal statistical analysis at $\sqrt{s}=500~\rm GeV$}
\label{sec:analysis}
We now apply the optimal observable method described in Sect.~\ref{sec:method} to the case of $ \psi^\pm $ production at the ILC, using the parameters of Eq.~\eqref{eq:params}; the cases of $250\,\hbox{GeV}$ and $2\,\hbox{TeV}$ CM energy collider are briefly discussed in appendix \ref{sec:1tev}.
\subsubsection{$\chi^2=1$ surfaces in the $a-b$ plane}
As a first step, we use the above expressions to obtain the coefficients $ c_i $ and functions $ f_i $; for example, for unpolarized beams ($P_{e^\pm}=0$) Eqs.~(\ref{fi}) and (\ref{ci}) give:
\beq
\begin{array}{|c|l|l|}
\hline
\multicolumn{3}{|c|}{\sqrt{s} = 500\, \hbox{GeV}\,, \quad P_{e^\pm}=0 } \cr \hline
m_{\psi^\pm}\, (\hbox{GeV}) & \qquad\qquad\qquad c_i/(\alpha\alpha_0)
& f_i \times 10^{-8}\, \hbox{GeV}^2 \cr
\hline\hline
& i=1:~ \frac{1}{2}(1-0.086 a +0.522 a^2 +0.245b^2) & i=1:~ 217.60 \\
150 & i=2:~ \frac{1}{2}(1.442b - 0.124ab) & i=2:~ 128.00 \cos \theta \\
& i=3:~ \frac{1}{2}(1- 0.086a +0.552 (a^2 + b^2)) & i=3:~ 102.40 \cos^2\theta \\
\hline
& i=1:~ \frac{1}{2}(1-0.086 a +0.522 a^2 +0.011b^2) & i=1:~ 78.02 \\
245 & i=2:~ \frac{1}{2}(1.442b - 0.124ab) & i=2:~ 7.92 \cos \theta \\
& i=3:~ \frac{1}{2}(1- 0.086a +0.552 (a^2 + b^2)) & i=3:~ 1.58 \cos^2 \theta .\\
\hline
\end{array}
\label{eq:c-a-b}
\eeq
\begin{figure}[htb!]
\begin{align*}
\includegraphics[scale=0.254]{S_500p2_a1b1_150GeV.png} \quad&\quad \includegraphics[scale=0.25]{S_500p2_a1b1_245GeV.png} \cr
\includegraphics[scale=0.264]{S_500p2_a1b0_150GeV.png} \quad&\quad \includegraphics[scale=0.25]{S_500p2_a1b0_245GeV.png} \cr
\includegraphics[scale=0.253]{S_500p2_a0b1_150GeV.png} \quad&\quad \includegraphics[scale=0.25]{S_500p2_a0b1_245GeV.png} \cr
\end{align*}
\caption{ $\chi^2=1$ surfaces for hypotheses with $ a^0 \ge 0 $, $P_{e^\pm} =0 $ and $ P_{e^\pm} = ^{+20\%}_{-80\%} $, and $ \epsilon = 0.001$. Left (right) column: $ m_{\psi^\pm} = 150\, (245)\, \hbox{GeV}$. Note: the scales in the graphs are not all equal.}
\label{fig:poldiff1}
\end{figure}
\begin{figure}[htb!]
\begin{align*}
\includegraphics[scale=0.265]{S_500p2_a-1b0_150GeV.png}\quad&\quad \includegraphics[scale=0.25]{S_500p2_a-1b0_245GeV.png} \cr
\includegraphics[scale=0.25]{S_500p2_a-1b1_150GeV.png}\quad&\quad \includegraphics[scale=0.25]{S_500p2_a-1b1_245GeV.png} \cr
\end{align*}
\caption{ Same as Fig. \ref{fig:poldiff1} for $a^0<0 $.}
\label{fig:poldiff2}
\end{figure}
Next, the optimal 1-$\sigma$ statistical uncertainties in the NP parameters $a,\,b$ are obtained from the $ \chi^2 \le1 $ regions using Eqs.~\eqref{eq:covmat2} and \eqref{chi2} for the parameters in Eq.~\eqref{eq:params}. As illustrative examples we manifest the seed values listed at the beginning of this section, and take~\footnote{It is clear from Eq.~\eqref{chi2} that if $ \chi^2$ is held fixed, the $ \delta c $ will scale as $ 1/\sqrt{\epsilon} $, but the dependence of $ \delta a,\,\delta b $ on $ \epsilon$ is more complicated, see Eq.~\eqref{eq:c-a-b}; in the calculations below we use $ \delta c = c(a^0 + \delta a,\,b^0 + \delta b) - c^0 $ in Eq.~\eqref{chi2}; see also section \ref{sec:method}.} $\epsilon=0.001,\,0.005$ as reasonable estimates of the efficiency of signal identification (see Sect. \ref{sec:collider} for a discussion); we consider both unpolarized ($P_{e^\pm}=0$) and polarized ($P_{e^\pm} = ^{+20\%}_{-80\%}$) beams. The results are presented in figure \ref{fig:poldiff1} (for $a^0\ge 0$), figure \ref{fig:poldiff2} (for $a^0=-1$) and Table \ref{table:abe1}. These results illustrate the advantages that polarization provides in the determination of the couplings of these new particles.
From these results we can see that of the cases considered, the $a^0=0,\, b^0=\pm1$ (pure axial coupling) hypothesis has the largest statistical errors and is therefore the most challenging. Also worth noting is that, while the magnitude of the total cross section strongly affects the statistical uncertainties, it is not the only factor; this is illustrated by considering the $ a^0=-1$ case where the unpolarized cross section is larger than the polarized one (Fig. \ref{fig:cscom1}) but the uncertainties are larger (Fig. \ref{fig:poldiff2}). We also note that 1$\sigma$ regions for the lower mass ($150$ GeV) are smaller than those for higher mass ($245$ GeV), making the determination of the NP couplings for the latter case more difficult.
\begin{center}
\begin{table}
\begin{tabular}{| c | c | c c | c c | c c | c c | }
\hline
\multicolumn{2}{|c|}{Seed parameters} &
\multicolumn{4}{|c|}{$P_{e^\pm}=0$} &
\multicolumn{4}{|c|}{$P_{e^\pm}=^{+20\%}_{-80\%}$} \\
\cline{3-10}
\multicolumn{2}{|c|}{} &
\multicolumn{2}{|c|}{$\epsilon=0.005$} &
\multicolumn{2}{|c|}{$\epsilon=0.001$}&
\multicolumn{2}{|c|}{$\epsilon=0.005$} &
\multicolumn{2}{|c|}{$\epsilon=0.001$}\\
\cline{1-10}
model &$m_{\psi^\pm}$(GeV) & $\pm\Delta a$ & $\pm\Delta b$ & $\pm\Delta a$ & $\pm\Delta b$ & $\pm\Delta a$ & $\pm\Delta b$ & $\pm \Delta a$ & $\pm\Delta b$\\
\hline
\multirow{4}*{\minitab[l]{$a=1$ \\ $b=0$}}
& \multirow{2}*{150}$ $ & $+0.04$ & $+0.11$ &$+0.08$ & $+0.24$ &$+0.02$&$+0.10$&$+0.05$&$+0.23$ \\
& &$-0.04$ & $-0.11$ & $-0.09$ & $-0.24$&$+0.02$&$+0.10$&$-0.05$ & $-0.23$ \\
\cline{2-10}
&$\multirow{2}*{245}$ & $+0.07$ & $+0.87$ & $+0.15$ & $+1.95$ &$+0.04$&$+0.05$ &$+0.09$ & $+0.91$ \\
& &$-0.07$ & $-0.87$ &$-0.17$ & $-1.95$&$+0.04$&$+0.05$&$-0.09$ & $-0.91$ \\
\hline
\multirow{4}*{\minitab[l]{$a=1$\\ $b=\pm1$}}
& $ \multirow{2}*{150} $ &$+0.06$ & $+0.10$ & $+0.13$ & $+0.23$&$+0.03$&$+0.11$ & $+0.07$ & $+0.25$\\
& &$-0.06$ & $-0.10$ & $-0.18$ & $-0.24$&$-0.03$&$-0.11$& $-0.08$ & $-0.24$ \\
\cline{2-10}
& $\multirow{2}*{245}$ &$+0.07$ & $+0.87$ & $+0.15$ & $+1.96$&$+0.04$&$+0.86$&$+0.09$ & $+1.93$ \\
& &$-0.07$ & $-0.87$ & $-0.21$ & $-1.96$&$-0.04$&$-0.86$&$-0.10$ & $-1.92$ \\
\hline
\multirow{4}*{\minitab[l]{$a=0$\\$b=\pm1$}}
& $ \multirow{2}*{150} $ &$+0.35$ & $+0.06$ & $+0.51$ & $+0.13$&$+0.04$&$+0.09$&$+0.07$ & $+0.22$ \\
& &$-0.27$ & $-0.11$ & $-0.41$ & $-0.22$&$-0.05$&$-0.09$&$-0.11$ & $-0.22$ \\
\cline{2-10}
&$\multirow{2}*{245}$ &$+0.43$ & $+0.66$ & $+0.59$ & $+1.39$&$+0.04$&$+0.66$&$+0.10$ & $+1.50$\\
& & $-0.26$ & $-0.67$ & $-0.13$ & $-1.51$&$-0.04$&$-0.66$&$-0.11$ & $-1.47$ \\
\hline
\multirow{4}*{\minitab[l]{$a=-1$\\$b=0$}}
& \multirow{2}*{150}$ $ & $+0.04$ & $+0.09$& $+0.07$ & $+0.21$&$+0.06$&$+0.06$& $+0.12$ & $-0.16$\\
& &$-0.04$ & $-0.09$ & $-0.07$ & $-0.21$ &$-0.08$&$-0.06$&$-0.63$ & $-0.16$ \\
\cline{2-10}
&$\multirow{2}*{245}$ & $+0.06$ & $+0.76$ & $+0.15$ & $+1.75$&$+0.10$&$+0.48$& $+0.20$ & $+1.23$ \\
& &$-0.06$ & $-0.76$ &$-0.15$ & $-1.75$&$-0.21$&$-0.48$&$-0.71$ & $-1.23$ \\
\hline
\multirow{4}*{\minitab[l]{$a=-1$\\$b=\pm1$}}
& $ \multirow{2}*{150} $ &$+0.06$ & $+0.10$ & $+0.15$ & $+0.22$&$+0.17$&$+0.11$&$+0.26$ & $+0.14$ \\
& &$-0.05$ & $-0.10$ & $-0.11$ & $-0.22$&$-0.34$&$-0.11$& $-0.48$ & $-0.22$ \\
\cline{2-10}
& $\multirow{2}*{245}$ &$+0.07$ & $+0.78$ & $+0.17$ & $+1.79$&$+0.11$&$+0.71$&$+0.20$ & $+1.31$\\
& &$-0.06$ & $-0.78$ &$-0.14$ & $-1.74$&$-0.60$&$-0.49$& $-0.70$ & $-1.11$ \\
\hline
\end{tabular}
\caption{Optimal 1$\sigma$ statistical uncertainty in the $a,b$ couplings for both unpolarized and polarized ($P_{e^\pm} = ^{+20\%}_{-80\%}$) beams and two values of $ \epsilon$; we used the parameters in Eq.~(\ref{eq:params}).}
\label{table:abe1}
\end{table}
\end{center}
\subsubsection{Differentiation of models}
One of the most important uses of the OOT is the ability to estimate the extent to which different hypotheses can be distinguished. Specifically, we consider a ``base'' hypothesis $ a = a^0,\,b=b^0$ and, using Eq.~(\ref{eq:covmat2}), define
\beq
\left[\Delta\sigma(a^0,b^0;\,\bar a,\bar b)\right]^2 = \epsilon\sum_{i,j} \left(c^0_i - \bar c_i \right) \left( c^0_j - \bar c_j \right) \left( V^{-1}_0 \right)_{ij}\,, \quad V_0 = V(c=c^0)\,;
\label{eq:signif-def}
\eeq
(where $ c^0_i= c_i(a^0,\,b^0),\,\bar c_i= c_i(\bar a,\,\bar b)$) which we take as a measure of the degree to which the $ a = \bar a,\,b = \bar b $ hypothesis can be distinguished from the base hypothesis; we refer to $ \Delta \sigma $
as the statistical significance of the $ \bar a,\,\bar b$ hypothesis (which depends on the base model chosen).
We will use $ \Delta\sigma $ as a measure of the separation of an alternate model from the base one. The distribution of $ \Delta\sigma $ can also be used to determine the probability that $ \Delta\sigma \le \ell $ occurs; in general this distribution is not simple, but for the cases where $c_0$ and $ \bar c$ have normal distributions with averages ${\mathbbm c}_0,\,\bar{\mathbbm c}$, and covariance matrices $V_0 $ and $ \bar V$ that are approximately proportional to the unit matrix, $ \sigma_0^2 {\mathbbm1},\, \bar\sigma^2{\mathbbm1}$ respectively, then $ \sqrt{\Delta\sigma} $ is approximately normally distributed with average $ |{\mathbbm c}_0-\bar{\mathbbm c}|/\sigma_0$ and variance $ 1 + (\bar\sigma/\sigma_0 )^2 $; in practice this means that the values quoted for $\Delta\sigma $ will have errors $ \sim \pm \sqrt{1 + (\bar\sigma/\sigma_0 )^2}$. Similar results hold when $ \Delta\sigma $ is written in terms of $a$ and $b$ provided they also are normally distributed.
We now consider a few examples\footnote{For these choices $V_0,\,\bar V$ are approximately proportional to ${\mathbbm1}$.} corresponding to some of the cases presented in Table~\ref{table:discovery1} or in Figs.~\ref{fig:dscvry1}, \ref{fig:dscvry2}, \ref{fig:absig1pol2}. If $a^0=1,\,b^0=1$ and $\bar a=-1,\,\bar b=0$, and we choose $m_{\psi} = 150\,\hbox{GeV}$, unpolarized beams, $ \epsilon = 0.005 $, and ${\mathfrak{L}}_{\tt int} =567 \, \text{fb}^{-1}$, we find $ \Delta\sigma \simeq 9 $ with a $\pm1.43$ uncertainty. Assuming now $m_{\psi} = 245\,\hbox{GeV}$, polarized beam ($P_{e^\pm} = ^{+20\%}_{-80\%}$), $ \epsilon = 0.001 $, and ${\mathfrak{L}}_{\tt int} =567 \, \text{fb}^{-1}$ and taking $a^0=0,\,b^0=0$ as the base model, we find that when $\bar a=1,\,\bar b=0$ (the purely vector-like case) $ \Delta\sigma = 13.96 $ with $\pm1.05$ uncertainty; w if $\bar a=1,\,\bar b=1$ we find $ \Delta\sigma = 14.09$ with an uncertainty of $\pm1.07$. We do not consider the $a^0=0, b^0=\pm 1$ cases since the $ \Delta\sigma$ distribution is not normal, and a full analysis statistical analysis of the $\Delta\sigma$ statistics lies beyond the scope of this paper; however, we expect that the uncertainties in these cases will continue to be $O(\lesssim10\%)$.
It is worth noting that, as expected, larger efficiency $ \epsilon$ and luminosity $ {\mathfrak{L}}_{\tt int} $ increases the significance, while larger masses reduce it. It is also important to note that though the significance $ \Delta\sigma $ depends on the magnitude of the cross section of the base model, this is not the only factor. For the example considered, the unpolarized cross section of the base model $a^0=b^0=0$ is smaller than the one for polarized beams by about 30\%, yet the significance of the $ \bar a=0,\,|\bar b|=1 $ models is the same, while that of $ \bar a=1,\,\bar b=0 $ is larger than expected from the cross-section alone.
\begin{figure}[htb!]
$$
\includegraphics[scale=0.25]{S_500p2_mpsi_245GeV_a1b0_001_dscvry_unpolarized.png}
~~\includegraphics[scale=0.25]{S_500p2_mpsi_245GeV_a1bpm1_001_dscvry_unpolarized.png}
~~\includegraphics[scale=0.25]{S_500p2_mpsi_245GeV_a0bpm1_001_dscvry_unpolarized.png}
$$
$$
\includegraphics[scale=0.25]{S_500p2_mpsi_245GeV_a1b0_005_dscvry_unpolarized.png}
~~\includegraphics[scale=0.25]{S_500p2_mpsi_245GeV_a1bpm1_005_dscvry_unpolarized.png}
~~\includegraphics[scale=0.25]{S_500p2_mpsi_245GeV_a0bpm1_005_dscvry_unpolarized.png}
$$
$$
\includegraphics[scale=0.25]{S_500p2_mpsi_245GeV_a1b0_001_2000fbinv_dscvry_unpolarized.png}
~~\includegraphics[scale=0.25]{S_500p2_mpsi_245GeV_a1bpm1_001_2000fbinv_dscvry_unpolarized.png}
~~\includegraphics[scale=0.25]{S_500p2_mpsi_245GeV_a0bpm1_001_2000fbinv_dscvry_unpolarized.png}
$$
\caption{$ 2\sigma,\,3\sigma$ and $5\sigma$ regions (red, blue and green, respectively) when $ a^0=b^0=0$, indicated by a star. The yellow area denotes the significance, Eq.~(\ref{eq:signif-def}), of alternate hypotheses $ a=\bar a,\,b=\bar b$ (indicated by crosses) for various choices of luminosity ($ {\cal L}$) and efficiency ($\epsilon$): $\bar a=1,\, \bar b=0$ (left column), $\bar a=1, \,\bar b=\pm 1$ (middle column) and $\bar a=0,\,\bar b=\pm 1$ (right column). We assumed $ m_{\psi^\pm} = 245\,\hbox{GeV}$ and unpolarized beams.}
\label{fig:dscvry1}
\end{figure}
\begin{figure}[htb!]
$$
\includegraphics[scale=0.27]{S_500p2_mpsi_245GeV_a-1b0_001_dscvry_unpolarized.png}
\includegraphics[scale=0.27]{S_500p2_mpsi_245GeV_a-1bpm1_001_dscvry_unpolarized.png}
$$
$$
\includegraphics[scale=0.27]{S_500p2_mpsi_245GeV_a-1b0_005_dscvry_unpolarized.png} \includegraphics[scale=0.27]{S_500p2_mpsi_245GeV_a-1bpm1_005_dscvry_unpolarized.png}
$$
$$
\includegraphics[scale=0.27]{S_500p2_mpsi_245GeV_a-1b0_001_2000fbinv_dscvry_unpolarized.png}
\includegraphics[scale=0.27]{S_500p2_mpsi_245GeV_a-1bpm1_001_2000fbinv_dscvry_unpolarized.png}
$$
\caption{Same as figure \ref{fig:dscvry1} for $\bar a = - 1,\,\bar b=0$ (left column) and $ \bar a = -1,\,\bar b=\pm1$ (right column).}
\label{fig:dscvry2}
\end{figure}
\begin{figure}[htb!]
$$
\includegraphics[scale=0.27]{S_500p2_mpsi_245GeV_a1b0_001_dscvry_polarized.png}
\includegraphics[scale=0.27]{S_500p2_mpsi_245GeV_a1bpm1_001_dscvry_polarized.png}
\includegraphics[scale=0.27]{S_500p2_mpsi_245GeV_a0bpm1_001_dscvry_polarized.png}
$$
$$
\includegraphics[scale=0.27]{S_500p2_mpsi_245GeV_a-1b0_001_dscvry_polarized.png}
\includegraphics[scale=0.27]{S_500p2_mpsi_245GeV_a-1bpm1_001_dscvry_polarized.png}
$$
\caption{Same as figure \ref{fig:dscvry1} for polarized beams ($P_{e^\pm} = ^{+20\%}_{-80\%}$).}
\label{fig:absig1pol2}
\end{figure}
\begin{center}
\begin{table}
$$ \begin{array}{| c | c | c | c | c | c | c | c | c | c| c|}
\hline
\multicolumn{1}{|c|}{}&
\multicolumn{1}{|c|}{}&
\multicolumn{1}{|c|}{} &
\multicolumn{4}{|c|}{\text{significance($\Delta\sigma$)\,}} \\
\cline{4-7}
\multicolumn{1}{|c|}{\text{model}}&
\multicolumn{1}{|c|}{\epsilon}&
\multicolumn{1}{|c|}{{\cal L}_{\tt int}~[\text{fb}^{-1}]}&
\multicolumn{2}{|c|}{\text{$m_{\psi^\pm}=150$ GeV}}&
\multicolumn{2}{|c|}{\text{$m_{\psi^\pm}=245$ GeV}}\\
\cline{4-7}
&& & P_{e^\pm}=0 &P_{e^\pm}=^{+20\%}_{-80\%} & P_{e^\pm}=0 &P_{e^\pm}=^{+20\%}_{-80\%}\\
\hline
\multirow{3}*{\minitab[l]{$\bar a=1$ \\ $\bar b=0$}}&0.001&567 &6.13&25.03 &3.42&13.96\\
&0.001&2000&11.51&47.03 &6.62&26.22\\
&0.005&567 &13.72&55.98 &7.65&31.21\\
\hline
\multirow{3}*{\minitab[l]{$\bar a=1$ \\ $\bar b=\pm1$}}&0.001&567 &11.46&29.81 &3.58&14.09\\
&0.001&2000&21.52&56.00 &6.73&26.46\\
&0.005&567 &25.62&66.67 &8.01&31.50\\
\hline
\multirow{3}*{\minitab[l]{$\bar a=0$ \\ $\bar b=\pm1$}}&0.001&567 &7.07&7.09 &0.68&0.69\\
&0.001&2000&13.28&13.32 &1.28&1.29\\
&0.005&567 &15.81&15.86 &1.52&1.54\\
\hline
\multirow{3}*{\minitab[l]{$\bar a=-1$ \\ $\bar b=0$}}&0.001&567 &8.65&10.81 &4.82&6.01\\
&0.001&2000&16.24&20.30 &9.05&11.28\\
&0.005&567 &19.33&24.17&10.78&13.44\\
\hline
\multirow{3}*{\minitab[l]{$\bar a=-1$ \\ $\bar b=\pm1$}}&0.001&567 &14.15&14.92 &4.99&5.92\\
&0.001&2000&26.57&28.02 &9.36&11.12\\
&0.005&567 &31.63&33.36&11.15&13.24\\
\hline
\end{array} $$
\caption{Statistical significance $\Delta \sigma $ (see Eq. (\ref{eq:signif-def})) of hypotheses $\bar a,\,\bar b$ with respect to the base hypothesis $a^0=b^0=0$.}
\label{table:discovery1}
\end{table}
\end{center}
\section{Model example}
\label{sec:DMmodel}
The above analysis focused on the application of the OOT to the study and detectability of the properties of a hypothetical new heavy lepton.
In this section we turn to a possible underlying economical and UV complete model that contains such a particle. This model provides a viable theoretical underpinning of the previous discussion, a framework for
studying other aspects of its detectability at the ILC, and can be used to obtain an estimate of the efficiency $ \epsilon $ ({\it cf.} Eq. (\ref{chi2})).
In addition, the study of this model using event-level simulation allows for a comparison of the expected ILC sensitivity to the optimal statistical
uncertainties derived above. Finally, we will see that the model proposed contains a viable dark matter candidate, satisfying the relic-density,
direct-search and electroweak constraints in a large region of parameter space.
The model consists of an extension of the SM by two vector-like leptons: a weak iso-doublet, $\psi=(\psi^0 ,\, \psi^-\, )$ of hypercharge $-1$, and an iso-singlet $\chi$ of zero hypercharge; both $ \psi$ and $ \chi $ are odd under an exact ${\mathbbm Z}_2$ symmetry under which all the SM fields are even \cite{Bhattacharya:2015qpa, Bhattacharya:2018fus}. Upon electroweak symmetry breaking (EWSB) the $\psi\chi H$ Yukawa coupling (see Eq.~(\ref{lag:lagVF}) below) generates a mixing between the neutral component $\psi^0$ and $\chi$, the resulting lighter mass eigenstate will be odd under ${\mathbbm Z}_2$ and therefore stable, and serves as a DM candidate. The quantum numbers under the SM$\times {\mathbbm Z}_2$ symmetry are summarized in Table \ref{tab:tab5}.
\begin{table}[htb]
\begin{center}
\begin{tabular}{|c||c|c|c|c|}
\hline
field & $\su3_{\tt C}$ & $\su2_{\tt L}$ & $U(1)_{\tt Y}$ & ${\mathbbm Z}_2$ \cr \hline
$\psi$ & $1$ & $2$ & $-1$ & odd \cr
$\chi$ & $1$ & $1$ & $0 $ & odd \cr \hline
\end{tabular}
\caption{Quantum numbers of the additional dark-sector fermions under $ \su3_{\tt C} \times \su2_{\tt L} \times U(1)_{\tt Y} \times {\mathbbm Z}_2 $. SM fields have the usual gauge quantum numbers and are even under $ {\mathbbm Z}_2$.}
\label{tab:tab5}
\end{center}
\end{table}
The Lagrangian of the model is
\beq
\label{lag:lagVF}
\mathcal{L}^{\tt VF} = \bar\psi \left[ i \left(\slashed\partial - i \frac g2 {\bm\sigma}} % \def\sigbf{{\pmb{$\sigma$}}\cdot\slashed{\bf W} - i \frac{g'}2 \slashed B\right)-m_{\psi^\pm} \right]\psi + \bar\chi \left(i \slashed\partial-m_{\chi} \right)\chi - \left(Y_1\bar\psi \widetilde{H}\chi + \text{H.c} \right)\,,
\eeq
(plus the usual SM terms); $H$ denotes the SM Higgs isodoublet, $ {\bf W}_\mu$ and $ B$ the $\su2_L$ and $U(1)_Y$ gauge fields, respectively, and $g,\,g'$ the corresponding gauge couplings.
After electroweak symmetry breaking $H$ acquires a vacuum expectation value\ $v/\sqrt{2}$:
\beq
H \rightarrow \frac{ v + h }{\sqrt{2}} \bpm1 \cr 0 \epm,
\label{eq:hssb}
\eeq
and, as noted above, the $ \chi$ and $ \psi_0 $ will mix through the Yukawa interaction $ \propto Y_1 $. The mass Lagrangian then becomes
\beq
-\mathcal{L}_{\tt mass} = \left( \bar\chi , \bar\psi^0 \right) \bpm m_\chi & \mu \cr \mu & m_{\psi^\pm} \epm \bpm \chi \cr \psi^0 \epm +m_{\psi^\pm}{\psi^+}\psi^- ; \qquad \mu = \frac{ Y_1 v}{\sqrt{2}}.
\eeq
The mass eigenstates $ \psi_{1,2} $ are then given by
\beq
\bpm \chi \cr \psi^0 \epm = \bpm \cos\gamma & - \sin\gamma \cr \sin\gamma & \cos\gamma \epm \bpm \psi_1 \cr \psi_2 \epm \,; \qquad \tan{2\gamma} = \frac{2\mu}{m_\chi - m_{\psi^\pm}}.
\label{ref:mixang}
\eeq
We will assume~\footnote{The case $ |\mu| < m_{\chi} \ll m_{\psi^\pm} $ is excluded by DM direct-detection and relic abundance constraints.} $ |\mu| \ll m_{\chi} < m_{\psi^\pm} $ so that $2\mu \ll |m_{\chi} - m_{\psi^\pm}|$; in this case $ \gamma $ is small and
\beq
m_{\psi_1} \simeq m_\chi - \frac{\mu^2}{m_{\psi^\pm} - m_\chi}\,, \qquad m_{\psi_2} \simeq m_{\psi^\pm} + \frac{\mu^2}{m_{\psi^\pm} - m_\chi};
\eeq
so that $ m_{\psi_2} > m_{\psi^\pm}> m_{\psi_1} $ and $ \psi_1 $ is the DM candidate. Note that we also have
\beq
Y_1 = - \sin(2\gamma) \frac{\Delta m}{\sqrt{2}\, v}\,, \qquad \Delta m = m_{\psi_2} - m_{\psi_1} > 0.
\label{ref:reltn}
\eeq
In the mass-eigenstate basis the interaction Lagrangian becomes
\bal
\mathcal{L}^{\tt VF}_{int} =& \frac{e_0}{s_{2\tt w}}\left[ s^2_\gamma (\bar {\psi_1} \gamma^\mu \psi_1) + c^2_\gamma (\bar {\psi_2}\gamma^\mu \psi_2) + s_\gamma c_\gamma (\bar {\psi_1} \gamma^\mu \psi_2 + \bar {\psi_2} \gamma^\mu \psi_1) - c_{2\tt w} ({\psi^+}\gamma^{\mu}\psi^-)\right] Z_{\mu} \cr
& - e_0 ({\psi^+}\gamma^{\mu}\psi^-)A_{\mu} +\frac{e_0}{\sqrt2s_{\tt w}}\left\{ \left[ s_\gamma (\bar {\psi_1}\gamma^\mu \psi^-) + c_\gamma (\bar {\psi_2}\gamma^\mu \psi^-) \right]W_\mu^+ + \text{H.c.} \right\} \cr
& -\frac{Y_1}{\sqrt2}h \left[s_{2\gamma} (\bar {\psi_1}{\psi_1}-\bar {\psi_2}\psi_2)+ c_{2\gamma}(\bar {\psi_1}\psi_2+\bar {\psi_2}\psi_1)\right],
\label{eq:interaction}
\end{align}
where $ s_\gamma = \sin\gamma$, etc., and $h$ is defined in Eq. (\ref{eq:hssb}).
We see that the charged heavy fermions (${\psi^\pm}$) have vector-like interactions with $Z$ boson (corresponding $a^0=c_{2\tt w}\sim1/2,\,b^0=0$ in Eq.~\eqref{eq:vertex1}); the $W$ couplings are also vector-like. Comparing with Table \ref{table:abe1} we that for $ \epsilon =0.005$ and polarized beams we expect~\footnote{The case at hand is similar to $ a^0=1,\,b^0=0$.} the ILC to be able to measure $a^0,\,b^0 $ to within $ \lesssim 10\% $ at $ 1 \sigma $ (ignoring systematic uncertainties).
The strongest limits on the model parameters come from dark matter constraints. The interactions in Eq.~\eqref{eq:interaction} show that the DM relic density is determined by the $h$ and $Z$-mediated annihilation and co-annihilation channels, while nuclear scattering, probed by direct-search experiments, is dominated by the $Z$ exchange process alone. The experimental constraint on the spin-independent cross section $ \sigma^{\tt SI}_{\tt dir.\,det.} \lesssim 10^{-47} \rm{cm}^2 $ (XENON1T collaboration, \cite{Aprile:2018dbl}), and the fact that this cross section is $ \propto \sin^4\gamma $ gives
\beq
\sin\gamma \lesssim 0.05 ~;
\eeq
with a weak dependence on the DM mass. This limit on $\sin\gamma$ sharply reduces DM annihilation cross-section via $Z$ mediation, and also via Higgs portal interactions since $Y_1 \propto \sin(\gamma) $. Though the SM$\rightarrow$DM annihilation channels are suppressed, the relic-abundance restriction~\footnote{Here {\sf h} denotes the Hubble parameter in units of $100\,$km\,s$^{-1}$\,Mpc$^{-1}$.} $\Omega_{\tt DM} {\sf h}^2=0.11933\pm 0.00091$ (PLANCK collaboration \cite{Ade:2013zuv, Aghanim:2018eyx}) can still be met through co-annihilation channels involving $ \psi^\pm $, provided $ |\Delta m| \ll m_{\psi^\pm} $ \cite{Bhattacharya:2018cgx}.
Fig.~\ref{fig:relicdd} displays various regions allowed by the direct-detection and relic-density constraints. The top panel displays the spin-independent direct-detection cross-section $ \sigma^{\tt SI}_{\tt DD}$ a function of DM mass ($m_{\psi_1}$) for various ranges of $\sin\gamma$. The allowed region in the $ m_{\psi_1} - \Delta m $ plane is displayed in the bottom left panel, while the allowed region in the $ m_{\psi_1} - \sin\gamma $ plane for several ranges of $ \Delta m$ is displayed on the right bottom panel of that figure. The parabola-like region in the bottom left panel is responsible for having two allowed values of $ m_{\psi_1} $ for each choice of $ \Delta m $ range on the right bottom panel; this paraboloid shape can be traced to the contribution from co-annihilation channels $\psi^\pm \psi_1\rightarrow$SM to the DM annihilation cross section:
\beq
\vevof{ \sigma v}_{\tt tot} \simeq \vevof{ \sigma v}_{\psi_1\bar{\psi_1}\rightarrow \rm SM} + \vevof{ \sigma v}_{\psi_1\psi^\pm\rightarrow \rm SM} \left(1+\frac{\Delta m}{m_{\psi_1}} \right)^{3/2}e^{-\Delta m/T} +\cdots
\label{eq:coann}
\eeq
where $T$ denotes the temperature of the bath, and the ellipses indicate other co-annihilation channels ({\it e.g.} $ \psi^+ \psi^-\rightarrow$SM, $\psi_1\bar{\psi_2} \rightarrow$SM) with a stronger exponential suppression; for details, see \cite{Bhattacharya:2018fus}. The relic density is then
\beq
\Omega_{\tt DM} {\sf h}^2 =\left. \frac{1.09 \times 10^9 ~{\rm GeV}^{-1} (m_{\psi_1}/T)}{{g_*}^{1/2}M_{\tt Pl}}\frac{1}{\vevof{ \sigma v}_{\tt tot}} \right|_{T=T_f};
\eeq
where $x = T/m_{\psi_1}$, $g_*$ denotes the effective relativistic degrees of freedom, $M_{\tt Pl}$ the Planck mass, and $ T_f $ the value of $T$ at freeze-out. From these expressions it follows that for small $ \Delta m $ the allowed values increase with $m_{\psi_1}$, but only up to a point beyond which $ \Delta m $ must drop to balance the the exponential suppression in Eq.~\eqref{eq:coann}.
\begin{figure}[htb!]
$$
\includegraphics[height=5.1cm]{ddsigma_mass_plot.png}
$$
$$
\includegraphics[height=5.1cm]{deltam_mDM.png} \quad
\includegraphics[height=5.1cm]{singamma_mass_plot.png}
$$
\caption{Regions allowed by the direct-detection and relic density constraints. Top: in the SI direct-detection cross-section ($\sigma^{\tt SI}_{\tt DD}$) vs DM mass ($m_{\psi_1}$) plane for various ranges of $ \sin \gamma$. Bottom left: in the $\Delta m -m_{\psi_1}$ plane for $0.01\le \sin\gamma \le0.05$; bottom right: in the $\sin \gamma - m_{\psi_1}$ plane for different ranges of $\Delta m$.}
\label{fig:relicdd}
\end{figure}
Collider data also impose constraints on this model, with the strongest limits from those on production of chargino pairs \cite{Aad:2019vnb}, or chargino and second neutralino production \cite{Aad:2019qnd}, in supersymmetric theories. Chargino pair production is the exact parallel of the one we study below (see Fig. \ref{fig:signal} with $ \psi^\pm $ replaced by charginos, and $ \psi_1 $ by neutralinos), in the limit where the chargino is wino-dominated and the sneutrinos are heavy~\footnote{The sneutrinos generate a $t$-channel graph not present in our model, and other contributions to the charginos generate chiral couplings to the $W$.}, and assuming the on-shell production of charginos dominates the cross section. These SUSY limits give
\beq
m_{\psi_2} + 115 \,\hbox{GeV} \gtrsim m_{\psi^\pm} \quad \text{for}~ 250\,\hbox{GeV} > m_{\psi^\pm} > 150 \,\hbox{GeV}
\eeq
which, for $ m_{\psi^\pm} = 150 \, \hbox{GeV} \, (245\, \hbox{GeV}) $, requires $ m_{\psi_1} > 35\, \hbox{GeV} \, (111\, \hbox{GeV}) $. The model is also consistent with electroweak precision observables, and with the invisible decay widths for the Higgs and $Z$ boson whenever $m_{\psi_1}>m_h/2$, which we assume.
With these constraints in mind, we select several benchmark points, listed in Table \ref{tab:BP}, where all constraints are obeyed and which we will use in our study of the model at the ILC; for these we also assumed $\Delta m <m_W$, so that the decay of heavy fermion occurs via an off-shell $W$. We will show that for such relatively small mass splitting there is better segregation of the signal from the SM background at the ILC. The benchmark points are compared with limits from ATLAS \cite{Aad:2019vnb,Aad:2019qnd}~\footnote{The limits from CMS \cite{CMS:2020bfa} agree well with Fig.~\ref{fig:masslim} and also allow the chosen benchmark points.} in Fig.~\ref{fig:masslim}.
\begin{figure}[htb!]
$$
\includegraphics[scale=0.405]{mass_limit_chargino_pp.png}
\includegraphics[scale=0.4]{mass_limit_com_susy.png}
$$
\caption{Current experimental limits from the LHC for supersymmetric chargino-neutralino production from the dilepton plus missing energy channel \cite{Aad:2019vnb} (left) and dilepton plus dijet plus missing energy channel \cite{Aad:2019qnd} (right). The benchmark points for our model (Table.~\ref{tab:BP}) are included for comparison.}
\label{fig:masslim}
\end{figure}
\begin{table}[htb]
\begin{center}
\begin{tabular}{| c | c | c | c | c | }
\hline
Benchmark Points&$m_{\psi^\pm}$ (GeV)& $m_{\psi_1}$ (GeV) & $\Delta m$ (GeV) \\
\hline
BP1&\multirow{2}*{\minitab[l]{$245$}}&215 & 30 \\
\cline{1-1}\cline{3-4}
BP2&&207 & 38 \\
\hline
BP3&\multirow{2}*{\minitab[l]{$150$}}&117 & 33 \\
\cline{1-1}\cline{3-4}
BP4&&110 & 40 \\
\hline
\end{tabular}
\caption{Benchmark points chosen for collider analysis for singlet-doublet fermion model; in all cases we took $\sin\gamma=0.05$.}
\label{tab:BP}
\end{center}
\end{table}
\subsection{Simulation of collider events}
\label{sec:collider}
We now turn to the ILC collider signatures for this model for the chosen benchmark points (table \ref{tab:BP}) using the simplest signal: $ \psi^\pm$ on-shell pair production with their subsequent decay into DM + opposite-sign leptons ({\tt OSL}) via off-shell $W$ bosons (Fig. \ref{fig:signal}); we adopt the above mass hierarchy, $ m_{\psi_2} > m_{\psi^\pm}> m_{\psi_1} $. We note that $ \psi_1 \, \psi_2 $ production, followed by $ \psi_2 \rightarrow Z \psi _1 $ generates a similar final state ({\tt OSL} plus missing energy), but the cross is $ \propto \sin^4\gamma $ and significantly smaller. It is also possible to pair produce $ \psi_1,\,\psi_2 $ but the final state signature is different.
\begin{figure}[htb!]
$$
\includegraphics[height=5.2cm]{ilc_sig.png}
$$
\caption{Production and decay of the heavy charged fermions at the ILC for the model described in section \ref{sec:DMmodel}.}
\label{fig:signal}
\end{figure}
We simulated {\tt OSL} events at the ILC with $\sqrt s=500$ GeV as follows: the model was implemented in {\sf Feynrules} \cite{Alloul:2013bka}, and parton-level signal events were generated using {\sf CalcHEP} \cite{Belyaev:2012qa}, and then showered and analyzed using {\sf Pythia }\cite{Sjostrand:2006za}; SM background events were generated using {\sf MadGraph} \cite{Alwall:2014hca} and showered using {\sf Pythia}. For event reconstruction, we use the following criteria:
\begin{itemize}
\item Leptons are required to have at least transverse momentum $p_T> 10$ GeV; we consider only electrons and muons with pseudorapidity $|\eta|< 2.4$ -- we do not consider $ \tau $ signals.
Two leptons are assumed isolated if $\Delta R_{\ell\ell}=\sqrt{(\Delta \eta)^2+(\Delta \phi)^2} \ge 0.2$, while a lepton and a
jet are assumed isolated if $\Delta R_{\ell j}\ge 0.4$.
\item We impose a zero-jet requirement, where jets are reconstructed using the cone jet algorithm around initiating parton. We further require $p_T>20$ GeV and $|\eta|< 3.0$.
\item Background signal was minimized by imposing cuts at $20$ and $30$ GeV (see below) on the missing transverse energy, which is defined by
\beq
E_{\tt T}^{\tt miss} = \left| {\bf p}\up{\tt vis}_\perp \right|^2
\eeq
where ${\bf p}\up{\tt vis}_\perp$ is the total visible momentum perpendicular to the beam direction.
\end{itemize}
\begin{figure}[htb!]
$$
\includegraphics[scale=0.4]{E_T1.png}
\includegraphics[scale=0.4]{E_T2.png}
$$
$$
\includegraphics[scale=0.422]{E_T1_pol.png}
\includegraphics[scale=0.4]{E_T2_pol.png}
$$
\caption{Normalized $E_{\tt T}^{\tt miss}$ distribution for the {\tt OSL} final state events at the ILC with $\sqrt s$= 500 GeV. Top: results for unpolarized beams; left (right), benchmark points BP1, BP2 (BP3, BP4) ({\it cf.} Table \ref{tab:BP}). Bottom: same for polarized beams ($P_{e^\pm} = ^{+20\%}_{-80\%}$). The SM background distributions from $WW,\, WWZ,\, ZZ$ production are also shown. The cuts $E_{\tt T}^{\tt miss}<20,\,30$ GeV cut used in the analysis are also indicated.}
\label{fig:c10c21}
\end{figure}
\begin{table}[htb]
\begin{center}
\begin{tabular}{| c | c | c | c | c | c | c | }
\hline
\multicolumn{1}{|c|}{\multirow{2}*{\minitab[l]{Background}}}&
\multicolumn{2}{c|}{$\sigma^{\psi^+ \psi^-}_{\tt SM}$ [pb]} &
\multicolumn{1}{c|}{\multirow{2}*{\minitab[l]{$E_{\tt T}^{\tt miss}$ [GeV]}} } &
\multicolumn{2}{c|}{$\sigma^{\tt OSL}_{\tt SM} $ [fb]} \\
\cline{2-3}
\cline{5-6}
\multicolumn{1}{|c|}{}&
\multicolumn{1}{c|}{$ P_{e^\pm}=0$}&
\multicolumn{1}{c|}{$P_{e^\pm}=^{+20\%}_{-80\%}$}&
\multicolumn{1}{c|}{}&
\multicolumn{1}{c|}{$ P_{e^\pm}=0$}&
\multicolumn{1}{c|}{$P_{e^\pm}=^{+20\%}_{-80\%}$}\\
\hline
\multirow{2}*{\minitab[l]{$WW$}}& \multirow{2}*{\minitab[l]{$7.4 $}}&\multirow{2}*{\minitab[l]{$15.5$}} & $<20$ &0.90 &2.73 \\
& & &$<30$ & 0.90&2.73 \\
\hline
\multirow{2}*{\minitab[l]{$WWZ$}} & \multirow{2}*{\minitab[l]{$0.04$}}& \multirow{2}*{\minitab[l]{$0.085$}}&$<20$ &$5.3\times10^{-4}$ &$1.5\times10^{-3}$ \\
& & &$<$30 & $1.6\times10^{-3}$ &$4.5\times10^{-4}$ \\
\hline
\multirow{2}*{\minitab[l]{$ZZ$}} & \multirow{2}*{\minitab[l]{$0.41$}} & \multirow{2}*{\minitab[l]{$0.66$}}&$<20$ &$3.1\times10^{-3}$ &$9.4\times10^{-3}$ \\
& & &$<30$ & $4.7\times10^{-3}$&$1.4\times10^{-2}$ \\
\hline
\end{tabular}
\caption{SM background cross sections for the $ \psi^\pm$ and {\tt OSL} final states at the ILC with $\sqrt{s}$=500 GeV with the selection cuts adopted (see text), and for unpolarized and polarized beams ($P_{e^\pm} = ^{+20\%}_{-80\%}$).}
\label{bg1}
\end{center}
\end{table}
\begin{table}[htb]
\begin{center}
\begin{tabular}{| c | c | c | c|c |c|c|c|c| }
\hline
\multicolumn{2}{|c|}{$\sigma^{\psi^+ \psi^-}$[pb]} &
\multicolumn{1}{c|}{\multirow{2}*{\minitab[l]{BPs}}} &
\multicolumn{1}{c|}{\multirow{2}*{\minitab[l]{$E_{\tt T}^{\tt miss}$ [GeV]}}}&
\multicolumn{2}{c|}{$\sigma^{\tt OSL}$[fb]} &
\multicolumn{2}{c|}{Efficiency ($\epsilon$)}\\
\cline{1-2}
\cline{5-8}
\multicolumn{1}{|c|}{$ P_{e^\pm}=0$}&
\multicolumn{1}{c|}{$P_{e^\pm}=^{+20\%}_{-80\%}$}&
\multicolumn{1}{c|}{}&
\multicolumn{1}{c|}{}&
\multicolumn{1}{c|}{$ P_{e^\pm}=0$}&
\multicolumn{1}{c|}{$P_{e^\pm}=^{+20\%}_{-80\%}$}&
\multicolumn{1}{c|}{$ P_{e^\pm}=0$}&
\multicolumn{1}{c|}{$P_{e^\pm}=^{+20\%}_{-80\%}$}\\
\hline
\multirow{4}*{\minitab[l]{$0.14$}} & \multirow{4}*{\minitab[l]{$0.32$}} & \multirow{2}*{\minitab[l]{BP1}} & $<20$&0.93 & 2.35 & $6.43\times10^{-3}$&$6.70\times10^{-3}$ \\
& & & $<30$ & 0.98 &2.50 &$6.98\times10^{-3}$&$7.11\times10^{-3}$\\
\cline{3-8}
& & \multirow{2}*{\minitab[l]{BP2}} &$<20$ & 0.70 &1.80 & $5.00\times10^{-3}$&$5.62\times10^{-3}$ \\
& & & $<30$ & 0.73 & 1.88&$5.21\times10^{-3}$ &$5.87\times10^{-3}$\\
\hline
\multirow{4}*{\minitab[l]{$0.45$}} & \multirow{4}*{\minitab[l]{$1.13$}} &\multirow{2}*{\minitab[l]{BP3}} & $<20$ &2.97 & 7.92& $6.60\times10^{-3}$ &$7.00\times10^{-3}$ \\
& & & $<30$ &3.32 & 8.45 & $7.37\times10^{-3}$ &$7.47\times10^{-3}$ \\
\cline{3-8}
&& \multirow{2}*{\minitab[l]{BP4}} &$<20$ & 2.30 & 6.05 & $5.09\times10^{-3}$ &$5.36\times10^{-3}$ \\
& & & $<30$ & 2.40 & 6.32 &$5.30\times10^{-3}$&$5.60\times10^{-3}$ \\
\hline
\end{tabular}
\caption{Signal ({\tt OSL}) and $ \psi^\pm$ pair production cross sections and associated signal efficiency $\epsilon$ Eq.~(\ref{eq:eff.def}) at the ILC for $\sqrt{s}=500$ GeV with two missing-energy selection cuts and other selection cuts (see text) for unpolarized and polarized beams ($P_{e^\pm} = ^{+20\%}_{-80\%}$).}
\label{tab:efficiency}
\end{center}
\end{table}
\begin{table}[htb]
\begin{center}
\begin{tabular}{| c | c | c | c | c | c | c | c|c | }
\hline
\multicolumn{1}{|c|}{\multirow{1}*{\minitab[l]{Benchmark}}}&
\multicolumn{1}{c|}{\multirow{1}*{\minitab[l]{$E_{\tt T}^{\tt miss}$}}}&
\multicolumn{6}{c|}{$\sigma^{\tt signal}$ [fb]}\\
\cline{3-8}
\multicolumn{1}{|c|}{ Points}&
\multicolumn{1}{|c|}{[GeV]}&
\multicolumn{2}{c|}{OSL + 0 photon }&
\multicolumn{2}{c|}{OSL+ $\le$ 1 photon}&
\multicolumn{2}{c|}{OSL+ $\le$ 2 photon} \\
\cline{3-8}
\multicolumn{1}{|c|}{}&
\multicolumn{1}{|c|}{}&
\multicolumn{1}{|c|}{$ P_{e^\pm}=0$}&
\multicolumn{1}{c|}{$P_{e^\pm}=^{+20\%}_{-80\%}$}&
\multicolumn{1}{|c|}{$ P_{e^\pm}=0$}&
\multicolumn{1}{c|}{$P_{e^\pm}=^{+20\%}_{-80\%}$}&
\multicolumn{1}{|c|}{$ P_{e^\pm}=0$}&
\multicolumn{1}{c|}{$P_{e^\pm}=^{+20\%}_{-80\%}$}\\
\hline
\multirow{2}*{\minitab[l]{BP1}} & $<20$ & 0.76& 2.17 & 0.84& 2.26 & 0.93 &2.35 \\
& $<30$ & 0.80& 2.31 & 0.87& 2.39 & 0.98 &2.50 \\
\hline
\multirow{2}*{\minitab[l]{BP2}} & $<20$ & 0.55& 1.79 & 0.62& 1.76 & 0.70 &1.80 \\
& $<30$ & 0.60& 1.71 & 0.67 & 1.79 & 0.73 &1.88 \\
\hline
\multirow{2}*{\minitab[l]{BP3}} & $<20$ & 2.82& 7.70 & 2.90& 7.81 & 2.97 &7.92 \\
& $<30$ & 2.15& 8.21 & 2.24 & 8.33 & 3.32 &8.45 \\
\hline
\multirow{2}*{\minitab[l]{BP4}} & $<20$ & 2.17& 5.87 & 2.23& 5.95 & 2.30 & 6.05 \\
& $<30$ & 2.26& 6.16 & 2.34 & 6.23 & 2.40 & 6.32 \\
\hline
\end{tabular}
\caption{Signal cross-section including ISR and FSR photon count.}
\label{tab:sig}
\end{center}
\end{table}
We present in Fig.~\ref{fig:c10c21} the missing transverse energy ($E_{\tt T}^{\tt miss}$) distributions (normalized to one event) at the benchmark points for both the signal and
the dominant SM background processes events ($WW,\,WWZ,\,ZZ$); and for both polarized and unpolarized beams. Since the intermediate $W$ bosons are off
shell, the peak of the missing energy distribution for the signal is at a much smaller value than those SM background, where $W$ production is on-shell.
Based on these distributions we choose upper cuts $E_{\tt T}^{\tt miss} < 20,\,30\,\hbox{GeV}$\footnote{ILC projections indicate this collider will be able to measure missing energy very accurately, so that the cut used in our analysis is viable \cite{Behnke:2013xla}.}, which retain a significant part of the signal and eliminate most background events, as illustrated in Table \ref{bg1}.
The values of the $ \psi^\pm$ pair production and signal event cross sections ($\sigma^{\psi^+\psi^-}$ and $\sigma^{\tt OSL}$, respectively) after imposing the above selection criteria and $E_{\tt T}^{\tt miss}$ cuts are given in Table~\ref{tab:efficiency} for both unpolarized and polarized ($P_{e^\pm}=^{+20\%}_{-80\%}$) beams. Following the discussion in Sect. \ref{sec:method} we define
\bea
\epsilon=\frac{\sigma^{\tt OSL}}{\sigma^{\psi^+\psi^-}}~,
\label{eq:eff.def}
\eea
(see Eq. \eqref{eq:eps.def}) whose values are listed in Table~\ref{tab:efficiency}. It is clear that $\epsilon\sim 0.001$ used throughout the previous OOT analysis is a conservative choice ($\epsilon\sim 0.005$ more closely corresponds to the results derived in this section).
We also note that signal events are often accompanied with initial state radiation (ISR) and final state radiation (FSR) photons simulated with inbuilt functions in {\tt Pythia} event generator. Using the photon selection criteria $p^{\tt photon}_T>5\,\hbox{GeV}$ and $|\eta_{\tt photon}|<0.24$, the signal cross-section with zero photons, and the inclusive $\le1,\,2$ photon cross sections are listed in table \ref{tab:sig}. From this we can see that the inclusive diphoton cross sections match quite accurately the signal cross-section without photon tagging listed in Table \ref{tab:efficiency}\footnote{With three photon and four photon events being very rare, inclusive di-photon event counts match quite accurately to signal cross-section without photon tagging as in Table~\ref{tab:efficiency}.}.
We determine the discovery potential of the {\tt OSL} signal at the ILC by plotting signal significance ($S/\sqrt{S+B}$, where $S$ and $B$ denote, respectively the number of signal and SM background events), as a function of luminosity ${\cal L} $. The results are presented in Fig.~\ref{fig:sig1}; of particular interest is the advantage provided by using polarized beams, which require a lower luminosity for either the discovery or exclusion of the selected signal; in either case the design luminosity ({\it cf.} Eq. (\ref{eq:params})) will be sufficient to exclude or detect the NP signal here investigated.
\begin{figure}[htb!]
$$
\includegraphics[scale=0.4]{discovery3s5s_1_unpol.png}
\includegraphics[scale=0.4]{discovery3s5s_1_pol.png}
$$
\caption{ Signal significance ($S/\sqrt{S+B}$) for the benchmark points in Table \ref{tab:BP} for unpolarized (left) and polarized ($P_{e^\pm}=^{+20\%}_{-80\%}$) beams (right). Blue and green lines correspond to $3\sigma$ and $5\sigma$ exclusion and discovery limits, respectively.}
\label{fig:sig1}
\end{figure}
It is intriguing to investigate whether the optimal uncertainty of the NP parameters obtained in the preceding section can be realized in a collider environment given the SM background contribution as analyzed in model specific scenario here. The experimental determination of model parameters will depend on the choice of signatures, the selection criteria and the corresponding significance; the authors are unaware of such a study, but the required dedicated analysis lies beyond the scope of the present paper. We will return to this issue in a future publication.
\section{Summary and Conclusions}
\label{sec:conclusion}
In this paper, we have analyzed the optimal statistical determination of the parameters of physics beyond the Standard Model, using as a specific example the production of a new heavy charged fermion $ \psi^\pm$ that can couple to the $Z$ boson and photon. We assumed for simplicity that the photon coupling is known, and allowing for both vector and axial coupling to the $Z$ with couplings $a$ and $b$ respectively. The optimal observable technique generates the minimal statistic uncertainty to which the couplings $a,\,b$ can be determined. We find that uncertainties for the case where $|a| \sim 1 $ are roughly independent of the value of $b$ and smaller than those for the quasi-axial case $ |a| \ll1 $. Find find, in addition, that, as expected, beam polarization allows for a different distinction of these couplings.
We also studied a sample model where the vector-like case ($b=0$) is realized; the model consists of an extension of the SM by a fermion isodoublet and a fermion singlet, both assumed odd under a ${\mathbbm Z}_2$ symmetry. In addition to providing a realization of the more general optimal observable analysis, this model contains a viable DM candidate. the presence of which can be probed at the ILC though $ \psi^\pm$ pair production followed by their decays into DM and $W$ bosons. The analysis shows that given the expected ILC luminosity (Eq.~\eqref{eq:params}) this collider will provide early evidence (or provide an early exclusion) of the model here proposed, and that this collider will be able to measure the model parameters with an accuracy very close to the optimal one. It is worth noting that the analysis is applicable to supersymmetric chargino pair production in the limit of heavy sneutrino, including our results on the optimal statistical uncertainties for the charged-lepton couplings to the $Z$ in the $a\neq0,b=0$ case.
\acknowledgments
SB acknowledges grant CRG/2019/004078 from SERB, Govt. of India and Mr. Abdur Rouf with whom the project was initiated. SJ would like to acknowledge Dr. Basabendu Barman and Dr. Sunando Patra for technical help.
|
Subsets and Splits